temp save

example/01_gemm/CMakeLists.txt

@@ -32,6 +32,7 @@ add_example_executable(example_gemm_xdl_fp16_fp8_v3 gemm_xdl_fp16_fp8_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_v3)
 add_example_executable(example_gemm_xdl_bf16_v3 gemm_xdl_bf16_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_v3)
+target_compile_options(example_gemm_xdl_bf16_v3 PRIVATE -mllvm -greedy-reverse-local-assignment=1 -save-temps=$PWD -Wno-gnu-line-marker)
 
 add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)

example/01_gemm/gemm_xdl_bf16_v3.cpp

@@ -12,7 +12,7 @@ using CShuffleDataType = ck::bhalf_t;
 using CDataType        = ck::bhalf_t;
 
 using ALayout = Row;
-using BLayout = Col;
+using BLayout = Row;
 using CLayout = Row;
 
 using AElementOp = PassThrough;
@@ -29,13 +29,13 @@ using DeviceGemmV2Instance =
         PassThrough, PassThrough, PassThrough, GemmDefault, 
         256,
         128, 128, 
-        64, 8, 8,
+        64, 8, 1,
         16,   16,
         4,    4,
         S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
         2, 8, 8, 0,
-        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
-        2, 8, 8, 0,
+        S<16, 16, 1>,  S<0, 2, 1>,  S<0, 2, 1>, 
+        1, 8, 8, 1,
         1, 2, S<1, 32, 1, 8>, 8,
         ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
 // clang-format on

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp

@@ -67,8 +67,8 @@ struct BlockwiseGemmXdlops_pipeline_base
                                                       KPerBlock,
                                                       ABlockTransferSrcScalarPerVector,
                                                       BBlockTransferSrcScalarPerVector,
-                                                      A_K1,
-                                                      B_K1,
+                                                      ABlockTransferSrcScalarPerVector,
+                                                      BBlockTransferSrcScalarPerVector,
                                                       A_K1,
                                                       B_K1,
                                                       MRepeat,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp

@@ -756,7 +756,8 @@ struct GridwiseGemm_xdl_cshuffle_v3
         {
             return make_naive_tensor_descriptor(
                 make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
-                make_tuple(BK1Number, Number<KPerBlock + BBlockLdsExtraN>{}, I1));
+                // make_tuple(BK1Number, Number<KPerBlock + BBlockLdsExtraN>{}, I1));
+                make_tuple(BK1Number, Number<KPerBlock>{}, I1));
         }
         else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
         {
@@ -1286,9 +1287,9 @@ struct GridwiseGemm_xdl_cshuffle_v3
                                                 decltype(b_grid_desc_bk0_n_bk1),
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
-                                                Sequence<0, 1, 2>,
+                                                BBlockTransferSrcAccessOrder,
+                                                BBlockTransferSrcVectorDim,
                                                 BBlockTransferSrcVectorDim,
-                                                2,
                                                 BBlockTransferSrcScalarPerVector,
                                                 BBlockTransferDstScalarPerVector_BK1,
                                                 1,