Merge branch 'gemm_bf16_sk_muozturk' of...

Merge branch 'gemm_bf16_sk_muozturk' of https://github.com/ROCm/composable_kernel into gemm_bf16_sk_muozturk

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_tp_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1,  256, 8,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1,  256, 4,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1, 1024, 2,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1, 1024, 1,  true,  true, true>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n512_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  4,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  8,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n64_n128_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn    pd     x      3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n768_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  3,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  6,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1, 12,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1024_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                                rm  rn  tm  tn  vn  pd      x     3p
+#if 0
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  8,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 16,  4,  64, 1,  true , true, false>>(const S&, A);
+
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  1, 256, 4,  true , true, false>>(const S&, A);
+#endif
+
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 1, 2,  128, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 2,  128, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1536_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 4,   64, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 6, 1,  256, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n2048_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 1, 1,  256, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 8, 1,  256, 1,  true,  true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n256_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n3072_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  128, 8,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  256, 4,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 6, 1,  256, 2,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1, 1024, 1,  true, true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 8,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 4,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1, 1024, 2,  true, true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1, 1024, 1,  true, true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n4096_tp_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 8,  true, true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 4,  true, true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1, 1024, 2,  true, true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1, 1024, 1,  true, true, true>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n512_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  8,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n64_n128_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n768_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  3,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  6,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 12,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_instance_common.hpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "add_rmsnorm2d_rdquant_fwd.hpp"
+#include <iostream>
+
+#pragma once
+
+using S = ck_tile::stream_config;
+using A = add_rmsnorm2d_rdquant_fwd_args;
+
+template <typename DataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kSaveInvRms_,
+          bool kTwoPass_>
+using trait_ = add_rmsnorm2d_rdquant_fwd_traits_<DataType_,
+                                                 Repeat_M_,
+                                                 Repeat_N_,
+                                                 ThreadPerBlock_M_,
+                                                 ThreadPerBlock_N_,
+                                                 Vector_N_,
+                                                 kPadN_,
+                                                 kSaveInvRms_,
+                                                 kTwoPass_>;
+
+template <typename Traits_>
+float add_rmsnorm2d_rdquant_fwd_(const S& s, A a)
+{
+    using DataType = typename Traits_::DataType;
+
+    using PipelineProblem = ck_tile::AddRmsnorm2dRdquantFwdPipelineProblem<
+        typename AddRmsnormRdquantTypeConfig<DataType>::ADataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::BDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::GammaDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::ComputeDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::XDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::YScaleDataType,
+        typename AddRmsnormRdquantTypeConfig<DataType>::QYDataType,
+        typename Traits_::Shape,
+        Traits_::kPadN,
+        Traits_::kSaveX,
+        Traits_::kThreePass>;
+
+    using OnePassPipeline   = ck_tile::AddRmsnorm2dRdquantFwdPipelineOnePass<PipelineProblem>;
+    using ThreePassPipeline = ck_tile::AddRmsnorm2dRdquantFwdPipelineThreePass<PipelineProblem>;
+    using Pipeline = std::conditional_t<Traits_::kThreePass, ThreePassPipeline, OnePassPipeline>;
+
+    using Kernel = ck_tile::AddRmsnorm2dRdquantFwd<Pipeline>;
+
+    const dim3 grids                       = Kernel::GridSize(a);
+    constexpr dim3 blocks                  = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+
+    auto kargs = Kernel::MakeKargs(a);
+    if(s.log_level_ > 0)
+        std::cout << ", " << Kernel::GetName() << std::flush;
+
+    return ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+}

example/ck_tile/11_add_rmsnorm2d_rdquant/script/perf_test.sh

+#!/bin/sh
+EXE="$(find . -name tile_add_rmsnorm2d_rdquant_fwd -type f | head -n 1)"
+
+$EXE -m=1 -n=1 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+
+$EXE -m=700 -n=80 -e=1e-12 -v=1  -prec=fp16 -repeat=1000
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
\ No newline at end of file

example/ck_tile/11_add_rmsnorm2d_rdquant/script/smoke_test.sh

+#!/bin/sh
+EXE="$(find . -name tile_add_rmsnorm2d_rdquant_fwd -type f | head -n 1)"
+
+for pr_i in "fp16" "bf16" ; do
+$EXE -prec=$pr_i -m=99  -n=13
+$EXE -prec=$pr_i -m=17  -n=16
+$EXE -prec=$pr_i -m=1   -n=100
+$EXE -prec=$pr_i -m=4   -n=128
+$EXE -prec=$pr_i -m=80  -n=127
+$EXE -prec=$pr_i -m=22  -n=255 -stride=256
+$EXE -prec=$pr_i -m=7   -n=599
+$EXE -prec=$pr_i -m=19  -n=512
+$EXE -prec=$pr_i -m=33  -n=313 -stride=1000
+$EXE -prec=$pr_i -m=11  -n=510
+$EXE -prec=$pr_i -m=171 -n=676 -stride=818
+$EXE -prec=$pr_i -m=91  -n=636
+$EXE -prec=$pr_i -m=12  -n=768 -stride=800
+$EXE -prec=$pr_i -m=100 -n=766 -stride=812
+$EXE -prec=$pr_i -m=31  -n=1024
+$EXE -prec=$pr_i -m=64  -n=1000 -stride=1004
+$EXE -prec=$pr_i -m=8   -n=1501
+$EXE -prec=$pr_i -m=3   -n=1826
+$EXE -prec=$pr_i -m=5   -n=2040
+$EXE -prec=$pr_i -m=7   -n=2734
+$EXE -prec=$pr_i -m=1   -n=3182
+$EXE -prec=$pr_i -m=9   -n=4096
+$EXE -prec=$pr_i -m=3   -n=8192
+$EXE -prec=$pr_i -m=1   -n=10547
+$EXE -prec=$pr_i -m=3   -n=17134
+done

example/ck_tile/12_smoothquant/CMakeLists.txt

+function (add_smoothquant_example TARGET_NAME MAIN_SRC)
+    message("adding ${TARGET_NAME}")
+    # not using add_example_executable() to add target, since we don't want this to have
+    # to be included in "make all/install/check"
+    add_executable(${TARGET_NAME} EXCLUDE_FROM_ALL ${MAIN_SRC})
+    target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+
+    foreach(source IN LISTS ARGN)
+        list(APPEND INSTANCE_SRCS ${source})
+    endforeach()
+
+    target_sources(${TARGET_NAME} PRIVATE ${INSTANCE_SRCS})
+
+    set(COMPILE_OPTIONS)
+    # NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+    list(APPEND COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
+
+    target_compile_options(${TARGET_NAME} PRIVATE ${COMPILE_OPTIONS})
+endfunction(add_smoothquant_example TARGET_NAME MAIN_SRC)
+
+
+add_smoothquant_example(tile_example_smoothquant example_smoothquant.cpp)
+file(GLOB INSTANCE_SRCS instances/*.cpp)
+add_smoothquant_example(tile_smoothquant smoothquant.cpp ${INSTANCE_SRCS})

example/ck_tile/12_smoothquant/README.md

+# smoothquant
+
+This folder contains example for smoothquant using ck_tile tile-programming implementation.
+
+## build
+```
+# in the root of ck_tile
+mkdir build && cd build
+sh ../script/cmake-ck-dev.sh  ../ <arch>  # you can replace this <arch> to gfx90a, gfx942...
+make tile_smoothquant -j
+```
+This will result in an executable `build/bin/tile_smoothquant`
+
+## cmdline
+```
+args:
+          -m    m dimension (default:3328)
+          -n    m dimension (default:4096)
+          -v    cpu validation or not (default:1)
+       -prec    precision (default:fp16)
+```

example/ck_tile/12_smoothquant/example_smoothquant.cpp

+#include "ck_tile/host.hpp"
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/kernel_launch.hpp"
+#include "ck_tile/ops/smoothquant.hpp"
+#include <cstring>
+
+// different threshold for different dtype
+template <typename DataType>
+auto get_elimit()
+{
+    double rtol = 1e-5;
+    double atol = 1e-5;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::bf16_t>()
+{
+    double rtol = 1e-5;
+    double atol = 1e-5;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::int8_t>()
+{
+    // due to rounding, int8 quantization might have 1 abs error
+    double rtol = 1;
+    double atol = 1;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("m", "3328", "m dimension")
+        .insert("n", "4096", "n dimension")
+        .insert("x_stride", "-1", "input stride per row, if -1 then equal to n")
+        .insert("y_stride", "-1", "output stride per row, if -1 then equal to n")
+        .insert("e", "1e-5", "epsilon")
+        .insert("v", "1", "cpu validation or not")
+        .insert("prec", "fp16", "precision")
+        .insert("warmup", "0", "cold iter")
+        .insert("repeat", "1", "hot iter");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename DataType>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    ck_tile::index_t m        = arg_parser.get_int("m");
+    ck_tile::index_t n        = arg_parser.get_int("n");
+    ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
+    if(x_stride < 0)
+        x_stride = n;
+    ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
+    if(y_stride < 0)
+        y_stride = n;
+    std::string data_type = arg_parser.get_str("prec");
+    int do_validation     = arg_parser.get_int("v");
+    int warmup            = arg_parser.get_int("warmup");
+    int repeat            = arg_parser.get_int("repeat");
+
+    assert(stride >= n);
+
+    using XDataType       = DataType;
+    using XScaleDataType  = float;
+    using YScaleDataType  = float;
+    using QYDataType      = ck_tile::int8_t;
+    using ComputeDataType = float;
+
+    // host verify
+    ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
+    ck_tile::HostTensor<XScaleDataType> xscale_host({n});
+
+    ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
+    ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
+
+    ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
+
+    ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
+    ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
+
+    ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem xscale_buf(xscale_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
+
+    x_buf.ToDevice(x_host.data());
+    xscale_buf.ToDevice(xscale_host.data());
+
+    constexpr bool kTwoPass = true;
+
+    using BlockWarps = ck_tile::sequence<2, 2>;
+    using BlockTile  = ck_tile::sequence<2, 128>;
+    using WarpTile   = ck_tile::sequence<1, 64>;
+    using Vector     = ck_tile::sequence<1, 1>;
+
+    using Shape   = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
+    using Problem = ck_tile::SmoothquantPipelineProblem<XDataType,
+                                                        XScaleDataType,
+                                                        ComputeDataType,
+                                                        YScaleDataType,
+                                                        QYDataType,
+                                                        Shape,
+                                                        true,
+                                                        kTwoPass>;
+
+    using OnePassPipeline = ck_tile::SmoothquantPipelineOnePass<Problem>;
+    using TwoPassPipeline = ck_tile::SmoothquantPipelineTwoPass<Problem>;
+    using Pipeline        = std::conditional_t<kTwoPass, TwoPassPipeline, OnePassPipeline>;
+    using Kernel          = ck_tile::Smoothquant<Pipeline>;
+
+    ck_tile::SmoothquantHostArgs args{x_buf.GetDeviceBuffer(),
+                                      xscale_buf.GetDeviceBuffer(),
+                                      yscale_buf.GetDeviceBuffer(),
+                                      qy_buf.GetDeviceBuffer(),
+                                      m,
+                                      n,
+                                      x_stride,
+                                      y_stride};
+
+    auto kargs = Kernel::MakeKargs(args);
+
+    const dim3 grids                       = Kernel::GridSize(args);
+    constexpr dim3 blocks                  = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+    auto s = ck_tile::stream_config{nullptr, true, 1, warmup, repeat};
+
+    ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+
+    bool pass = true;
+
+    if(do_validation)
+    {
+        using YDataType = ComputeDataType;
+        ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {y_stride, 1});
+        // smooth outlier
+        {
+            auto f = [&](auto n_) {
+                auto v_xscale = ck_tile::type_convert<ComputeDataType>(xscale_host(n_));
+
+                for(int m_ = 0; m_ < m; ++m_)
+                {
+                    auto v_x       = ck_tile::type_convert<ComputeDataType>(x_host(m_, n_));
+                    y_host(m_, n_) = v_x * v_xscale;
+                }
+            };
+
+            ck_tile::make_ParallelTensorFunctor(f, xscale_host.get_element_space_size())(
+                std::thread::hardware_concurrency());
+        }
+
+        // yscale
+        {
+            ck_tile::HostTensor<YDataType> y_rowwise_amax_host({m});
+
+            using ReduceAmax = ck_tile::ReduceOp::AbsMax;
+            ck_tile::reference_reduce<ComputeDataType, ComputeDataType, YDataType>(
+                y_host, y_rowwise_amax_host, ReduceAmax{});
+
+            auto op = [](const auto& v0) {
+                return v0 /
+                       ck_tile::type_convert<ComputeDataType>(ck_tile::numeric<QYDataType>::max());
+            };
+            ck_tile::reference_unary_elementwise<YDataType, YScaleDataType, ComputeDataType>(
+                y_rowwise_amax_host, yscale_host_ref, op);
+
+            yscale_buf.FromDevice(yscale_host_dev.mData.data());
+
+            auto [rtol, atol] = get_elimit<YScaleDataType>();
+            pass &= ck_tile::check_err(yscale_host_dev,
+                                       yscale_host_ref,
+                                       std::string("yscale Error: Incorrect results!"),
+                                       rtol,
+                                       atol);
+        }
+
+        // rowwise quantization
+        {
+            ck_tile::reference_rowwise_quantization2d<YDataType, YScaleDataType, QYDataType>(
+                y_host, yscale_host_ref, qy_host_ref);
+
+            qy_buf.FromDevice(qy_host_dev.data());
+            auto [rtol, atol] = get_elimit<QYDataType>();
+
+            if(y_stride == n)
+            {
+                pass = ck_tile::check_err(qy_host_dev,
+                                          qy_host_ref,
+                                          std::string("qy Error: Incorrect results!"),
+                                          rtol,
+                                          atol);
+            }
+            else
+            {
+                for(int i_r = 0; i_r < m; i_r++)
+                {
+                    std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * y_stride,
+                                                            qy_host_dev.begin() + i_r * y_stride +
+                                                                n);
+                    std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * y_stride,
+                                                            qy_host_ref.begin() + i_r * y_stride +
+                                                                n);
+                    pass &= ck_tile::check_err(qy_host_dev_row,
+                                               qy_host_ref_row,
+                                               std::string("qy[") + std::to_string(i_r) +
+                                                   std::string("] Error: Incorrect results!"),
+                                               rtol,
+                                               atol);
+                }
+            }
+        }
+
+        std::cout << "[" << data_type << "]"
+                  << " m:" << m << ", n:" << n << ", x_stride:" << x_stride
+                  << ", y_stride:" << y_stride << ", valid:" << (pass ? "y" : "n") << std::flush
+                  << std::endl;
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    const std::string data_type = arg_parser.get_str("prec");
+    if(data_type == "fp16")
+    {
+        return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
+    }
+    /*else if(data_type == "bf16")
+    {
+        return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2;
+    }*/
+
+    return -3;
+}