提交FT和CK交叉编译代码

3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }

3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using B0DataType       = BF16;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = BF16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmSpec,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    64,          // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    2,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<16, 16, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+    false>;
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }

3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmSpec,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    64,          // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    2,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<16, 16, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+    false>;
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }

3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }

3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }

3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M             = 1020;
+    ck::index_t N             = 1020;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA       = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideC       = -1;
+    ck::index_t BatchStrideA  = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideC  = -1;
+    float alpha               = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 9)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+    }
+    else if(argc == 18)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+
+        StrideA  = std::stoi(argv[9]);
+        StrideB0 = std::stoi(argv[10]);
+        StrideB1 = std::stoi(argv[11]);
+        StrideC  = std::stoi(argv[12]);
+
+        BatchStrideA  = std::stoi(argv[13]);
+        BatchStrideB0 = std::stoi(argv[14]);
+        BatchStrideB1 = std::stoi(argv[15]);
+        BatchStrideC  = std::stoi(argv[16]);
+
+        alpha = std::stof(argv[17]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
+               "BatchStrideB0, BatchStrideB1, BatchStrideC\n");
+        printf("arg17: scale (alpha)\n");
+        exit(0);
+    }
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
+                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    auto argument =
+        gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+                          static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+                          M,
+                          N,
+                          K,
+                          O,
+                          BatchCount,
+                          StrideA,
+                          StrideB0,
+                          StrideB1,
+                          StrideC,
+                          BatchStrideA,
+                          BatchStrideB0,
+                          BatchStrideB1,
+                          BatchStrideC,
+                          a_element_op,
+                          b0_element_op,
+                          acc0_element_op,
+                          b1_element_op,
+                          c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+    if(do_verification)
+    {
+        // Output of Gemm0 is input A of Gemm1
+        Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData) ? 0 : 1;
+    }
+
+    return 0;
+}

3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape for A/B0/B1/C
+    // C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
+    ck::index_t M = 120;
+    ck::index_t N = 1000;
+    ck::index_t K = 64;
+    ck::index_t O = 128;
+
+    // Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
+    // C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
+    // C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
+    ck::index_t G0 = 7;
+    ck::index_t G1 = 13;
+
+    float alpha = 1;
+
+    bool input_permute  = false;
+    bool output_permute = true;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 13)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M  = std::stoi(argv[4]);
+        N  = std::stoi(argv[5]);
+        K  = std::stoi(argv[6]);
+        O  = std::stoi(argv[7]);
+        G0 = std::stoi(argv[8]);
+        G1 = std::stoi(argv[9]);
+
+        alpha = std::stof(argv[10]);
+
+        input_permute  = std::stoi(argv[11]);
+        output_permute = std::stoi(argv[12]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 11: M, N, K, O, G0, G1\n");
+        printf("arg10: scale (alpha)\n");
+        printf("arg11 to 12: input / output permute\n");
+        exit(0);
+    }
+
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
+            : std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
+
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
+            : std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
+
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
+            : std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
+
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides =
+        output_permute
+            ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
+            : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    // TODO ANT: replace array with vector?
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(
+        static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+        static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+        static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        {}, // std::array<void*, 1> p_acc0_biases;
+        {}, // std::array<void*, 1> p_acc1_biases;
+        a_gs_ms_ks_lengths,
+        a_gs_ms_ks_strides,
+        b0_gs_ns_ks_lengths,
+        b0_gs_ns_ks_strides,
+        b1_gs_os_ns_lengths,
+        b1_gs_os_ns_strides,
+        c_gs_ms_os_lengths,
+        c_gs_ms_os_strides,
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+        a_element_op,
+        b0_element_op,
+        acc0_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    ck::index_t BatchCount = G0 * G1;
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+
+        // gemm 0
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // masking
+        const auto mask = DeviceGemmInstance::C0MatrixMask(N);
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(mask.IsMaskedElement(idx[1], idx[2]))
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        // softmax
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        // gemm1
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+
+        // default absolute error and relative error is 0.001
+        double rtol = 1e-3;
+        double atol = 1e-3;
+
+        // when BF16 is taken, set absolute error and relative error to 0.01
+        if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
+           std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
+        {
+            rtol = 1e-2;
+            atol = 1e-2;
+        }
+
+        return ck::utils::check_err(c_gs_ms_os_device_result.mData,
+                                    c_gs_ms_os_host_result.mData,
+                                    "Error: Incorrect results!",
+                                    rtol,
+                                    atol)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}

3rdparty/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    bool input_permute  = false;
+    bool output_permute = true;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        input_permute  = std::stoi(argv[4]);
+        output_permute = std::stoi(argv[5]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 5: input / output permute\n");
+        exit(0);
+    }
+
+    float alpha = 1; // scaling after 1st gemm
+
+    std::size_t group_count = 7;
+
+    // Problem descs
+    std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
+    std::vector<const void*> p_a;
+    std::vector<const void*> p_b0;
+    std::vector<const void*> p_b1;
+    std::vector<void*> p_c;
+    std::vector<std::vector<int>> g0_g1_m_n_k_o;
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<B0DataType>> b0_tensors;
+    std::vector<Tensor<B1DataType>> b1_tensors;
+    std::vector<Tensor<CDataType>> c_tensors;
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+    std::vector<DeviceMemPtr> a_tensors_device;
+    std::vector<DeviceMemPtr> b0_tensors_device;
+    std::vector<DeviceMemPtr> b1_tensors_device;
+    std::vector<DeviceMemPtr> c_tensors_device;
+
+    std::size_t flop = 0, num_byte = 0;
+
+    std::cout << "group count " << group_count << ". printing first 4 groups\n";
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        int M  = 128 * (rand() % 8 + 1);
+        int N  = 128 * (rand() % 8 + 1);
+        int K  = 40;
+        int O  = 40 * (rand() % 2 + 1);
+        int G0 = rand() % 3 + 1;
+        int G1 = rand() % 5 + 1;
+
+        g0_g1_m_n_k_o.push_back({G0, G1, M, N, K, O});
+
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides =
+            input_permute
+                ? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
+                : std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
+
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides =
+            input_permute
+                ? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
+                : std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
+
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides =
+            input_permute
+                ? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
+                : std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
+
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides =
+            output_permute
+                ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
+                : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
+
+        problem_descs.push_back({a_gs_ms_ks_lengths,
+                                 a_gs_ms_ks_strides,
+                                 b0_gs_ns_ks_lengths,
+                                 b0_gs_ns_ks_strides,
+                                 b1_gs_os_ns_lengths,
+                                 b1_gs_os_ns_strides,
+                                 c_gs_ms_os_lengths,
+                                 c_gs_ms_os_strides,
+                                 {},   // acc0_biases_gs_ms_ns_lengths
+                                 {},   // acc0_biases_gs_ms_ns_strides
+                                 {},   // acc1_biases_gs_ms_os_lengths
+                                 {}}); // acc1_biases_gs_ms_os_strides
+
+        // C_m_o = A_m_k * B0_k_n * B1_n_o
+        Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+        Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+        Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+        Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+        int Batch = G0 * G1;
+        flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
+        num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                    Batch;
+
+        if(i < 4)
+        {
+            std::cout << "a_gs_ms_ks[" << i << "]: " << a_gs_ms_ks.mDesc << ", "
+                      << "b0_gs_ns_ks[" << i << "]: " << b0_gs_ns_ks.mDesc << ", "
+                      << "b1_gs_os_ns[" << i << "]: " << b1_gs_os_ns.mDesc << ", "
+                      << "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
+        }
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+            break;
+        case 2:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+            break;
+        case 3:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+            break;
+        default:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        }
+
+        a_tensors.push_back(a_gs_ms_ks);
+        b0_tensors.push_back(b0_gs_ns_ks);
+        b1_tensors.push_back(b1_gs_os_ns);
+        c_tensors.push_back(c_gs_ms_os_device_result);
+
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize()));
+        b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize()));
+        b1_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
+
+        a_tensors_device[i]->ToDevice(a_gs_ms_ks.mData.data());
+        b0_tensors_device[i]->ToDevice(b0_gs_ns_ks.mData.data());
+        b1_tensors_device[i]->ToDevice(b1_gs_os_ns.mData.data());
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
+        p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
+        p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(p_a,
+                                      p_b0,
+                                      p_b1,
+                                      p_c,
+                                      {}, // p_acc0_biases
+                                      {}, // p_acc1_biases
+                                      problem_descs,
+                                      a_element_op,
+                                      b0_element_op,
+                                      acc0_element_op,
+                                      b1_element_op,
+                                      c_element_op);
+
+    // specify workspace for problem_desc
+    DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < group_count; i++)
+        {
+            const int& G0 = g0_g1_m_n_k_o[i][0];
+            const int& G1 = g0_g1_m_n_k_o[i][1];
+            const int& M  = g0_g1_m_n_k_o[i][2];
+            const int& N  = g0_g1_m_n_k_o[i][3];
+            const int& K  = g0_g1_m_n_k_o[i][4];
+            const int& O  = g0_g1_m_n_k_o[i][5];
+
+            const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
+            const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
+
+            const auto& a_gs_ms_ks         = a_tensors[i];
+            const auto& b0_gs_ns_ks        = b0_tensors[i];
+            const auto& b1_gs_os_ns        = b1_tensors[i];
+            auto& c_gs_ms_os_device_result = c_tensors[i];
+            auto& c_gs_ms_os_device_buf    = *c_tensors_device[i];
+
+            c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+            Tensor<ADataType> a_g_m_k({G0 * G1, M, K});
+            Tensor<B0DataType> b0_g_k_n({G0 * G1, K, N});
+            Tensor<B1DataType> b1_g_n_o({G0 * G1, N, O});
+            Tensor<AccDataType> acc0_g_m_n({G0 * G1, M, N});        // scratch object after gemm0
+            Tensor<ADataType> a1_g_m_n({G0 * G1, M, N});            // scratch object after softmax
+            Tensor<CDataType> c_g_m_o_host_result({G0 * G1, M, O}); // scratch object after gemm1
+            Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+            // permute
+            a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+                a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+            });
+            b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+                b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+            });
+            b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+                b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+            });
+
+            // gemm 0
+            auto ref_gemm0          = ReferenceGemm0Instance{};
+            auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+            auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+                a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+            ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+            // masking
+            const auto mask = DeviceGemmInstance::C0MatrixMask(N);
+            acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+                if(mask.IsMaskedElement(idx[1], idx[2]))
+                    self(idx) = -ck::NumericLimits<float>::Infinity();
+            });
+
+            // softmax
+            auto ref_softmax          = ReferenceSoftmaxInstance{};
+            auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+            auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+            ref_softmax_invoker.Run(ref_softmax_argument);
+
+            // gemm 1
+            auto ref_gemm1          = ReferenceGemm1Instance{};
+            auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+            auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
+                                                             b1_g_n_o,
+                                                             c_g_m_o_host_result,
+                                                             PassThrough{},
+                                                             b1_element_op,
+                                                             c_element_op);
+
+            ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+            // permute
+            c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+                const size_t& g0 = idx[0];
+                const size_t& g1 = idx[1];
+
+                const size_t g = g0 * G1 + g1;
+
+                self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+            });
+
+            bool pass_ =
+                ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
+            pass &= pass_;
+        }
+    }
+
+    return pass ? 0 : 1;
+}

3rdparty/composable_kernel/example/33_multiple_reduce/CMakeLists.txt

+add_example_executable(example_dual_reduce_multiblock dual_reduce_multiblock.cpp)
+add_example_executable(example_dual_reduce_threadwise dual_reduce_threadwise.cpp)

3rdparty/composable_kernel/example/33_multiple_reduce/README.md

+# Instructions for ```example_dual_reduce```
+
+## Run ```example_dual_reduce_multiblock```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg2: time kernel (0=no, 1=yes) 
+./bin/example_dual_reduce_multiblock -D 600,28,28,256 -v 1 2 1
+```
+
+Result
+```
+./bin/example_dual_reduce_multiblock -D 600,28,28,256 -v 1 2 1                        
+launch_and_time_kernel: grid_dim {150, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.19529 ms, 201.499 GB/s, DeviceMultipleReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_1_InSrcVectorSize_1,OutDstVectorSize_1_1>
+```
+
+## Run ```example_dual_reduce_threadwise```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg2: time kernel (0=no, 1=yes)
+./bin/example_dual_reduce_multiblock -D 8000,4,4,4 -v 1 2 1
+```
+
+Result
+```
+./bin/example_dual_reduce_threadwise -D 8000,4,4,4 -v 1 2 1
+launch_and_time_kernel: grid_dim {32, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.01512 ms, 71.9577 GB/s, DeviceMultipleReduceThreadwise<256,M_C256_S1,K_C1_S4,InSrcVectorDim_1_InSrcVectorSize_2,OutDstVectorSize_1_1>
+```

3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {600, 28, 28, 256};
+    size_t n, h, w, c;
+
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    SimpleAppArgs()
+    {
+        n = inLengths[0];
+        h = inLengths[1];
+        w = inLengths[2];
+        c = inLengths[3];
+    };
+
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg2 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                if(inLengths.size() != 4)
+                    throw std::runtime_error(
+                        "Invalid option format! The number of integers is incorrect!");
+
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        n = inLengths[0];
+        h = inLengths[1];
+        w = inLengths[2];
+        c = inLengths[3];
+
+        return (0);
+    };
+};
+
+template <typename InDataType, typename OutDataType1, typename OutDataType2, typename AccDataType>
+static void mean_meansquare_host(const Tensor<InDataType>& in,
+                                 Tensor<OutDataType1>& mean_ref,
+                                 Tensor<OutDataType2>& meansquare_ref,
+                                 size_t n,
+                                 size_t h,
+                                 size_t w,
+                                 size_t c)
+
+{
+    auto thread_reduce_func = [&](auto iN) {
+        AccDataType mean       = ck::type_convert<AccDataType>(0.0f);
+        AccDataType meansquare = ck::type_convert<AccDataType>(0.0f);
+
+        // compute mean, meanquare, variance, invVariance
+        for(std::size_t iH = 0; iH < h; iH++)
+        {
+            for(std::size_t iW = 0; iW < w; iW++)
+            {
+                for(std::size_t iC = 0; iC < c; iC++)
+                {
+                    AccDataType curr_value = ck::type_convert<AccDataType>(in(iN, iH, iW, iC));
+
+                    mean += curr_value;
+                    meansquare += curr_value * curr_value;
+                };
+            }
+        };
+
+        mean       = mean / (h * w * c);
+        meansquare = meansquare / (h * w * c);
+
+        mean_ref(iN)       = ck::type_convert<OutDataType1>(mean);
+        meansquare_ref(iN) = ck::type_convert<OutDataType2>(meansquare);
+    };
+
+    std::size_t num_thread      = std::thread::hardware_concurrency();
+    std::size_t work_per_thread = (n + num_thread - 1) / num_thread;
+
+    std::vector<joinable_thread> threads(num_thread);
+
+    for(std::size_t it = 0; it < num_thread; it++)
+    {
+        std::size_t iN_begin = it * work_per_thread;
+        std::size_t iN_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread), n);
+
+        auto f = [=] {
+            for(std::size_t iN = iN_begin; iN < iN_end; iN++)
+            {
+                thread_reduce_func(iN);
+            }
+        };
+
+        threads[it] = joinable_thread(f);
+    }
+};
+
+using ReduceOperation = ck::reduce::Add;
+
+using InElementwiseOperation_Mean  = ck::tensor_operation::element_wise::PassThrough;
+using AccElementwiseOperation_Mean = ck::tensor_operation::element_wise::UnaryDivide;
+
+using InElementwiseOperation_Meansquare  = ck::tensor_operation::element_wise::UnarySquare;
+using AccElementwiseOperation_Meansquare = ck::tensor_operation::element_wise::UnaryDivide;
+
+using InElementwiseOperationTuple =
+    ck::Tuple<InElementwiseOperation_Mean, InElementwiseOperation_Meansquare>;
+using AccElementwiseOperationTuple =
+    ck::Tuple<AccElementwiseOperation_Mean, AccElementwiseOperation_Meansquare>;
+
+template <typename DeviceDualReduce,
+          typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          int Rank,
+          int NumReduceDim>
+int mean_meansquare_dual_reduce_test(size_t n,
+                                     size_t h,
+                                     size_t w,
+                                     size_t c,
+                                     bool do_verification,
+                                     int init_method,
+                                     bool time_kernel,
+                                     const std::array<int, NumReduceDim> reduceDims)
+{
+    const std::vector<size_t> inLengths = {n, h, w, c};
+
+    Tensor<InDataType> in(inLengths);
+
+    std::vector<size_t> outLengths{n};
+
+    Tensor<OutDataType> mean_ref(outLengths);
+    Tensor<OutDataType> mean(outLengths);
+    Tensor<OutDataType> meansquare_ref(outLengths);
+    Tensor<OutDataType> meansquare(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = mean.mDesc.GetStrides();
+
+    size_t invariant_total_length = n;
+    size_t reduce_total_length    = h * w * c;
+
+    const AccDataType alpha = ck::type_convert<AccDataType>(1.0f);
+    const AccDataType beta  = ck::type_convert<AccDataType>(0.0f);
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1: in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread); break;
+        case 2: in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread); break;
+        default: in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem mean_dev(sizeof(OutDataType) * mean.mDesc.GetElementSpaceSize());
+    DeviceMem meansquare_dev(sizeof(OutDataType) * meansquare.mDesc.GetElementSpaceSize());
+
+    in_dev.ToDevice(in.mData.data());
+
+    if(do_verification)
+    {
+        mean_meansquare_host<InDataType, OutDataType, OutDataType, AccDataType>(
+            in, mean_ref, meansquare_ref, n, h, w, c);
+    };
+
+    constexpr ck::index_t NumInputDim  = Rank;
+    constexpr ck::index_t NumOutputDim = (Rank - NumReduceDim > 1) ? Rank - NumReduceDim : 1;
+
+    std::array<ck::index_t, NumInputDim> i_inLengths;
+    std::array<ck::index_t, NumInputDim> i_inStrides;
+    std::array<ck::index_t, NumOutputDim> i_outLengths;
+    std::array<ck::index_t, NumOutputDim> i_outStrides;
+
+    ck::ranges::copy(inLengths, i_inLengths.begin());
+    ck::ranges::copy(inStrides, i_inStrides.begin());
+    ck::ranges::copy(outLengths, i_outLengths.begin());
+    ck::ranges::copy(outStrides, i_outStrides.begin());
+
+    auto dual_reduce_op = DeviceDualReduce{};
+
+    auto argument_ptr = dual_reduce_op.MakeArgumentPointer(
+        i_inLengths,
+        i_inStrides,
+        i_outLengths,
+        {i_outStrides, i_outStrides},
+        reduceDims,
+        {&alpha, &alpha},
+        {&beta, &beta},
+        in_dev.GetDeviceBuffer(),
+        {mean_dev.GetDeviceBuffer(), meansquare_dev.GetDeviceBuffer()},
+        ck::make_tuple(InElementwiseOperation_Mean{}, InElementwiseOperation_Meansquare{}),
+        ck::make_tuple(
+            AccElementwiseOperation_Mean{static_cast<int32_t>(reduce_total_length)},
+            AccElementwiseOperation_Meansquare{static_cast<int32_t>(reduce_total_length)}));
+
+    if(!dual_reduce_op.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+        return (-1);
+    };
+
+    std::string reduce_name = dual_reduce_op.GetTypeString();
+
+    auto invoker_ptr = dual_reduce_op.MakeInvokerPointer();
+
+    float avg_time = 0.0f;
+
+    avg_time += invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InDataType) +
+                            2 * invariant_total_length * sizeof(OutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        mean_dev.FromDevice(mean.mData.data());
+        meansquare_dev.FromDevice(meansquare.mData.data());
+        pass = pass && ck::utils::check_err(mean, mean_ref);
+        pass = pass && ck::utils::check_err(meansquare, meansquare_ref);
+    };
+
+    return (pass ? 0 : 1);
+}

3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_multiblock.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "dual_reduce_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InDataType       = ck::half_t;
+using OutDataType      = float;
+using OutDataTypeTuple = Tuple<OutDataType, OutDataType>;
+using AccDataType      = float;
+
+// for NHWC layer-norm calculation of mean and meansquare
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+constexpr bool PropagateNan = false;
+
+constexpr InMemoryDataOperationEnum OutMemoryDataOperation = InMemoryDataOperationEnum::Set;
+
+using DeviceDualReduce = DeviceMultipleReduceMultiBlock<2,
+                                                        InDataType,
+                                                        AccDataType,
+                                                        OutDataTypeTuple,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperationTuple,
+                                                        AccElementwiseOperationTuple,
+                                                        OutMemoryDataOperation,
+                                                        PropagateNan,
+                                                        256,
+                                                        4,
+                                                        64,
+                                                        1,
+                                                        1,
+                                                        1, // InSrcVectorDim
+                                                        1,
+                                                        ck::Sequence<1, 1>>;
+
+int main(int argc, char* argv[])
+{
+    int retval = 0;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(arg.n,
+                                                                arg.h,
+                                                                arg.w,
+                                                                arg.c,
+                                                                arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                reduceDims);
+    }
+    else
+    {
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(
+            600, 28, 28, 256, true, 2, true, reduceDims);
+    };
+
+    return (retval);
+}

3rdparty/composable_kernel/example/33_multiple_reduce/dual_reduce_threadwise.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "dual_reduce_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InDataType       = ck::half_t;
+using OutDataType      = float;
+using OutDataTypeTuple = Tuple<OutDataType, OutDataType>;
+using AccDataType      = float;
+
+// for NHWC layer-norm calculation of mean and meansquare
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+constexpr bool PropagateNan = false;
+
+using DeviceDualReduce = DeviceMultipleReduceThreadWise<2,
+                                                        InDataType,
+                                                        AccDataType,
+                                                        OutDataTypeTuple,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperationTuple,
+                                                        AccElementwiseOperationTuple,
+                                                        PropagateNan,
+                                                        256,
+                                                        1,
+                                                        4,
+                                                        1, // InSrcVectorDim
+                                                        2,
+                                                        ck::Sequence<1, 1>>;
+
+int main(int argc, char* argv[])
+{
+    int retval = 0;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(arg.n,
+                                                                arg.h,
+                                                                arg.w,
+                                                                arg.c,
+                                                                arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                reduceDims);
+    }
+    else
+    {
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(
+            8000, 4, 4, 4, true, 2, true, reduceDims);
+    };
+
+    return (retval);
+}

3rdparty/composable_kernel/example/34_batchnorm/CMakeLists.txt

+add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
+add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)
+add_example_executable(example_batchnorm_backward batchnorm_backward_nhwc.cpp)

3rdparty/composable_kernel/example/34_batchnorm/README.md

+# Instructions for ```batchnorm nhwc``` Example
+
+## Run ```batchnorm forward nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1:  data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+#arg2: 1/0 to indicate whether to update the moving average and variance (0=no, 1=yes)
+#arg3: 1/0 to indicate whether to save result mean/invVariance (0=no, 1=yes)
+#arg4: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg5: time kernel (0=no, 1=yes) 
+./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 0 1 2 1
+```
+
+Result 
+```
+./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 0 1 2 1
+launch_and_time_kernel: grid_dim {64, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 2.08231 ms, 354.519 GB/s
+```
+
+Result
+```
+./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 1 0 2 0
+echo $?
+0
+```
+
+## Run ```batchnorm infer nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1:  data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_batchnorm_infer -D 128,16,16,1024 -v 1 0 2 1
+```
+
+Result
+```
+./bin/example_batchnorm_infer -D 128,16,16,1024 -v 1 0 2 1
+launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.28235 ms, 523.329 GB/s
+```
+
+## Run ```batchnorm backward nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+Arg2 -- 1/0 to indicate whether to use saved mean and invVariance
+Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+Arg4 -- time kernel (0=no, 1=yes)
+Arg5: use multi-block welford (0=n0, 1=yes)
+./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
+```
+
+Result 
+```
+./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.411026 ms, 91.8702 GB/s
+```

3rdparty/composable_kernel/example/34_batchnorm/batchnorm_backward_nhwc.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormBwdArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    bool haveSavedMeanInvVar;
+
+    int data_type               = 0;
+    int init_method             = 3;
+    bool time_kernel            = false;
+    bool use_multiblock_welford = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
+        std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg5: use multi-block welford (0=n0, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 5 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        haveSavedMeanInvVar    = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+        use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename XDataType, typename AccDataType, bool UseMultiblockInK>
+bool bnorm_bwd_nhwc_test(bool do_verification,
+                         int init_method,
+                         bool time_kernel,
+                         const std::vector<size_t> inOutLengths,
+                         bool haveSavedMeanInvVar,
+                         double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr index_t Rank         = 4;
+    constexpr index_t NumReduceDim = 3;
+
+    using ScaleDataType = XDataType;
+
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm backward algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<AccDataType> dy(inOutLengths);
+
+    Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> savedMean(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> savedInvVar(scaleBiasMeanVarLengths);
+    // savedVariance is only used for initializing savedInvVar
+    Tensor<AccDataType> savedVariance(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm backward algorithm
+    Tensor<AccDataType> dx_ref(inOutLengths);
+    Tensor<AccDataType> dx(inOutLengths);
+
+    Tensor<AccDataType> dscale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> dbias(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> dscale_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> dbias_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = dy.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = dscale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(haveSavedMeanInvVar)
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        savedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                      num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        savedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+
+        auto it_src       = savedVariance.mData.begin();
+        auto it_dst       = savedInvVar.mData.begin();
+        float tmp_epsilon = std::numeric_limits<float>::epsilon();
+
+        while(it_src != savedVariance.mData.end())
+        {
+            *it_dst = type_convert<AccDataType>(
+                1.0f / std::sqrtf(type_convert<float>(*it_src) + tmp_epsilon));
+
+            it_src++;
+            it_dst++;
+        };
+    }
+    else
+    {
+        const float x_mean   = 0.0f;
+        const float x_stddev = 1.0f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            dy.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            break;
+        case 1:
+            dy.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            break;
+        case 2:
+            dy.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-2, 2}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            dy.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-0.2f, 0.2f}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-0.5f, 0.5f}, num_thread);
+        }
+    };
+
+    // input data of the batchnorm backward algorithm
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem dy_dev(sizeof(AccDataType) * dy.mDesc.GetElementSpaceSize());
+
+    DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
+
+    DeviceMem savedMean_dev(sizeof(AccDataType) * savedMean.mDesc.GetElementSpaceSize());
+    DeviceMem savedInvVar_dev(sizeof(AccDataType) * savedInvVar.mDesc.GetElementSpaceSize());
+
+    // output data of the batchnorm backward algorithm
+    DeviceMem dx_dev(sizeof(AccDataType) * dx.mDesc.GetElementSpaceSize());
+
+    DeviceMem dscale_dev(sizeof(AccDataType) * dscale.mDesc.GetElementSpaceSize());
+    DeviceMem dbias_dev(sizeof(AccDataType) * dbias.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    dy_dev.ToDevice(dy.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+
+    if(haveSavedMeanInvVar)
+    {
+        savedMean_dev.ToDevice(savedMean.mData.data());
+        savedInvVar_dev.ToDevice(savedInvVar.mData.data());
+    };
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), i_inOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), i_inOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              i_scaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              i_scaleBiasMeanVarStrides.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    using DeviceBatchNormBwdInstance =
+        ck::tensor_operation::device::DeviceBatchNormBwdImpl<XDataType,
+                                                             AccDataType,
+                                                             AccDataType,
+                                                             AccDataType,
+                                                             ScaleDataType, // ScaleDataType
+                                                             AccDataType,   // DscaleDbiasDataType
+                                                             AccDataType,   // MeanVarDataType
+                                                             PassThroughOp,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             UseMultiblockInK,
+                                                             256,
+                                                             16,
+                                                             16,
+                                                             1,
+                                                             2,
+                                                             0,
+                                                             1,  // XSrcVectorSize
+                                                             1,  // DySrcVectorSize
+                                                             1,  // DxDstVectorSize
+                                                             1,  // ScaleSrcVectorSize
+                                                             1,  // DscaleDbiasDstVectorSize
+                                                             1>; // MeanVarSrcVectorSize
+
+    auto batchnorm_bwd = DeviceBatchNormBwdInstance{};
+
+    auto argument_ptr = batchnorm_bwd.MakeArgumentPointer(
+        i_inOutLengths,
+        i_inOutStrides,
+        i_inOutStrides,
+        i_inOutStrides,
+        {0, 1, 2},
+        i_scaleBiasMeanVarLengths,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        x_dev.GetDeviceBuffer(),
+        dy_dev.GetDeviceBuffer(),
+        bnScale_dev.GetDeviceBuffer(),
+        haveSavedMeanInvVar ? savedMean_dev.GetDeviceBuffer() : nullptr,
+        haveSavedMeanInvVar ? savedInvVar_dev.GetDeviceBuffer() : nullptr,
+        epsilon,
+        PassThroughOp{},
+        dx_dev.GetDeviceBuffer(),
+        dscale_dev.GetDeviceBuffer(),
+        dbias_dev.GetDeviceBuffer());
+
+    if(!batchnorm_bwd.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
+                     "exiting!"
+                  << std::endl;
+        return (false);
+    };
+
+    size_t workspace_sz = batchnorm_bwd.GetWorkSpaceSize(argument_ptr.get());
+
+    DeviceMem workspace_dev(workspace_sz);
+
+    batchnorm_bwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = batchnorm_bwd.MakeInvokerPointer();
+
+    if(time_kernel)
+    {
+        float avg_time   = 0.0f;
+        size_t num_bytes = 0;
+
+        size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
+        size_t invariant_length = inOutLengths[3];
+
+        avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        // inputing of x, dy, scale, outputing of dx, dscale, dbias
+        num_bytes +=
+            total_length * sizeof(XDataType) * 3 + invariant_length * sizeof(AccDataType) * 3;
+
+        // outputing of mean, inv-variance
+        num_bytes += haveSavedMeanInvVar ? invariant_length * sizeof(AccDataType) * 2 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    }
+    else
+        (void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using ReferenceBatchNormBwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormBwd<XDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              ScaleDataType, // ScaleDataType
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumReduceDim>;
+
+        auto batchNormBwd_ref = ReferenceBatchNormBwdInstance{};
+
+        auto argument_ptr_ref = batchNormBwd_ref.MakeArgumentPointer(
+            i_inOutLengths,
+            i_inOutStrides,
+            i_inOutStrides,
+            i_inOutStrides,
+            {0, 1, 2},
+            i_scaleBiasMeanVarLengths,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            x.mData.data(),
+            dy.mData.data(),
+            bnScale.mData.data(),
+            haveSavedMeanInvVar ? savedMean.mData.data() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar.mData.data() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_ref.mData.data(),
+            dscale_ref.mData.data(),
+            dbias_ref.mData.data());
+
+        if(!batchNormBwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout
+                << "The runtime parameters seems not supported by the device instance, exiting!"
+                << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormBwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        dx_dev.FromDevice(dx.mData.data());
+        dscale_dev.FromDevice(dscale.data());
+        dbias_dev.FromDevice(dbias.data());
+
+        // clang-format off
+        pass = pass && ck::utils::check_err(dbias.mData, dbias_ref.mData, "dBias result:", 2e-4, 2e-4);
+        pass = pass && ck::utils::check_err(dscale.mData, dscale_ref.mData, "dScale result:", 2e-4, 2e-4);
+        pass = pass && ck::utils::check_err(dx.mData, dx_ref.mData, "dx result:");
+        // clang-format on
+    };
+
+    return (pass);
+};
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormBwdArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
+                                                                    arg.init_method,
+                                                                    arg.time_kernel,
+                                                                    arg.inOutLengths,
+                                                                    arg.haveSavedMeanInvVar,
+                                                                    epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.haveSavedMeanInvVar,
+                                                                     epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<float, float, true>(arg.do_verification,
+                                                               arg.init_method,
+                                                               arg.time_kernel,
+                                                               arg.inOutLengths,
+                                                               arg.haveSavedMeanInvVar,
+                                                               epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<float, float, false>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.haveSavedMeanInvVar,
+                                                                epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.haveSavedMeanInvVar,
+                                                                     epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
+                                                                      arg.init_method,
+                                                                      arg.time_kernel,
+                                                                      arg.inOutLengths,
+                                                                      arg.haveSavedMeanInvVar,
+                                                                      epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<double, double, true>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.haveSavedMeanInvVar,
+                                                                 epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<double, double, false>(arg.do_verification,
+                                                                  arg.init_method,
+                                                                  arg.time_kernel,
+                                                                  arg.inOutLengths,
+                                                                  arg.haveSavedMeanInvVar,
+                                                                  epsilon);
+        }
+    }
+    else
+    {
+        pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(true,
+                                                            3,
+                                                            false, // don't time kernel
+                                                            {128, 16, 6, 512},
+                                                            false,
+                                                            epsilon);
+
+        pass = pass && bnorm_bwd_nhwc_test<ck::half_t, float, false>(true,
+                                                                     3,
+                                                                     false, // don't time kernel
+                                                                     {128, 16, 3, 1024},
+                                                                     false,
+                                                                     epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}

3rdparty/composable_kernel/example/34_batchnorm/batchnorm_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cassert>
+#include <vector>
+#include <array>
+#include <type_traits>
+
+#include "ck/utility/data_type.hpp"
+
+struct NormalizeInInfer
+{
+    NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
+
+    template <typename T1, typename T2, typename T3, typename T4>
+    __host__ __device__ constexpr void operator()(T1& y,
+                                                  const T1& x,
+                                                  const T2& mean,
+                                                  const T2& variance,
+                                                  const T3& gamma,
+                                                  const T4& beta) const
+    {
+        static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
+                      "Data type is not supported by this operation!");
+
+        using ck::type_convert;
+        using ck::math::sqrt;
+
+        T2 tmp_x, tmp_y;
+
+        tmp_x = type_convert<T2>(x);
+
+        tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
+                    type_convert<T2>(gamma) +
+                type_convert<T2>(beta);
+        y = type_convert<T1>(tmp_y);
+    };
+
+    double epsilon_;
+};
+
+template <int Rank, int NumReduceDim>
+static inline std::array<int, Rank - NumReduceDim>
+get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
+{
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    std::array<int, Rank - NumReduceDim> invariantDims;
+
+    // collect invariant dimensions
+    int dim = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            invariantDims[dim] = i;
+            dim++;
+        };
+
+    return invariantDims;
+};

3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp"
+
+#include "batchnorm_infer_impl.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormInferArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
+                     "lengths, must have 4 integers for nhwc"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
+                     "result by "
+                     "comparing with the host-based batch-normalization"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: init method used for bnScale and bnBias (0=no init, 1=single integer "
+                     "value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg3: time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 3 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename InOutDataType, typename AccDataType>
+bool bnorm_infer_nhwc_test(bool do_verification,
+                           int init_method,
+                           bool time_kernel,
+                           const std::vector<size_t> inOutLengths,
+                           double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr int Rank         = 4;
+    constexpr int NumReduceDim = 3;
+
+    // when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
+    // eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm forward algorithm
+    Tensor<InOutDataType> x(inOutLengths);
+    Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm forward algorithm
+    Tensor<InOutDataType> y_ref(inOutLengths);
+    Tensor<InOutDataType> y(inOutLengths);
+
+    Tensor<AccDataType> estimatedMean(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> estimatedVariance(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if constexpr(std::is_same<InOutDataType, int8_t>::value)
+    {
+        x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 2.5f;
+        const float noise_stddev = 0.0001f;
+
+        estimatedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                          num_thread);
+
+        estimatedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+    }
+    else
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        estimatedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                          num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        estimatedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            break;
+        case 1:
+            bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
+            break;
+        case 2:
+            bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
+    DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
+
+    // mean_dev or resultSaveMean_dev
+    DeviceMem estimatedMean_dev(sizeof(AccDataType) * estimatedMean.mDesc.GetElementSpaceSize());
+    // meansquare_dev or resultSaveInvVariance_dev
+    DeviceMem estimatedVariance_dev(sizeof(AccDataType) *
+                                    estimatedVariance.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+    bnBias_dev.ToDevice(bnBias.mData.data());
+    estimatedMean_dev.ToDevice(estimatedMean.mData.data());
+    estimatedVariance_dev.ToDevice(estimatedVariance.mData.data());
+
+    using ck::index_t;
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
+    ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
+    ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
+    ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
+
+    int result = 0;
+
+    result = bnorm_infer<InOutDataType,
+                         InOutDataType,
+                         AccDataType,
+                         AccDataType,
+                         AccDataType,
+                         AccDataType,
+                         Rank,
+                         NumReduceDim,
+                         false>(time_kernel,
+                                {0, 1, 2},
+                                i_inOutLengths,
+                                i_inOutStrides,
+                                i_inOutStrides,
+                                i_scaleBiasMeanVarLengths,
+                                i_scaleBiasMeanVarStrides,
+                                i_scaleBiasMeanVarStrides,
+                                i_scaleBiasMeanVarStrides,
+                                x_dev.GetDeviceBuffer(),
+                                bnScale_dev.GetDeviceBuffer(),
+                                bnBias_dev.GetDeviceBuffer(),
+                                epsilon,
+                                estimatedMean_dev.GetDeviceBuffer(),
+                                estimatedVariance_dev.GetDeviceBuffer(),
+                                y_dev.GetDeviceBuffer());
+
+    if(result < 0)
+        return (false);
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+        using ReferenceBatchNormInferInstance =
+            ck::tensor_operation::host::ReferenceBatchNormInfer<InOutDataType,
+                                                                InOutDataType,
+                                                                AccDataType,
+                                                                AccDataType,
+                                                                AccDataType,
+                                                                AccDataType,
+                                                                PassThroughOp,
+                                                                Rank,
+                                                                NumReduceDim>;
+        auto batchNormInfer_ref = ReferenceBatchNormInferInstance{};
+
+        auto argument_ptr_ref =
+            batchNormInfer_ref.MakeArgumentPointer(i_inOutLengths,
+                                                   i_inOutStrides,
+                                                   i_inOutStrides,
+                                                   {0, 1, 2},
+                                                   i_scaleBiasMeanVarLengths,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   x.mData.data(),
+                                                   bnScale.mData.data(),
+                                                   bnBias.mData.data(),
+                                                   epsilon,
+                                                   PassThroughOp{},
+                                                   estimatedMean.mData.data(),
+                                                   estimatedVariance.mData.data(),
+                                                   y_ref.mData.data());
+
+        if(!batchNormInfer_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout
+                << "The runtime parameters seems not supported by the BatchNorm instance, exiting!"
+                << std::endl;
+            return (-2);
+        };
+
+        auto invoker_ptr_ref = batchNormInfer_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        y_dev.FromDevice(y.mData.data());
+        pass = pass && ck::utils::check_err(y, y_ref);
+    };
+
+    return (pass);
+};
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormInferArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            pass = bnorm_infer_nhwc_test<ck::half_t, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            pass = bnorm_infer_nhwc_test<float, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 3)
+        {
+            pass = bnorm_infer_nhwc_test<int8_t, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            pass = bnorm_infer_nhwc_test<ck::bhalf_t, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            pass = bnorm_infer_nhwc_test<double, double>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        };
+    }
+    else
+    {
+        pass = bnorm_infer_nhwc_test<ck::half_t, float>(true,
+                                                        2,
+                                                        false, // don't time kernel
+                                                        {128, 16, 16, 1024},
+                                                        epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}

3rdparty/composable_kernel/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormFwdArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    bool updateMovingAverage;
+    bool saveMeanAndInvVariance;
+
+    int data_type               = 0;
+    int init_method             = 2;
+    bool time_kernel            = false;
+    bool use_multiblock_welford = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
+                     "lengths, must have 4 integers for nhwc"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
+                     "result by "
+                     "comparing with the host-based batch-normalization"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance "
+                     "(0=no, 1=yes)"
+                  << std::endl;
+        std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance "
+                     "(0=no, 1=yes)"
+                  << std::endl;
+        std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer "
+                     "value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg6: use multi-block welford (0=n0, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 6 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        updateMovingAverage    = std::atoi(argv[optind++]);
+        saveMeanAndInvVariance = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+        use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename InOutDataType, typename AccDataType, bool UseMultiblockInK>
+bool bnorm_fwd_nhwc_test(bool do_verification,
+                         int init_method,
+                         bool time_kernel,
+                         const std::vector<size_t> inOutLengths,
+                         bool updateMovingAverage,
+                         bool saveMeanAndInvVariance,
+                         double averageFactor,
+                         double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr int Rank         = 4;
+    constexpr int NumReduceDim = 3;
+
+    // when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
+    // eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm forward algorithm
+    Tensor<InOutDataType> x(inOutLengths);
+    Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm forward algorithm
+    Tensor<InOutDataType> y_ref(inOutLengths);
+    Tensor<InOutDataType> y(inOutLengths);
+
+    Tensor<AccDataType> resultSaveMean_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> resultSaveInvVariance_ref(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> resultRunningMean_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> resultRunningVariance_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(updateMovingAverage)
+    {
+        if constexpr(std::is_same<InOutDataType, int8_t>::value)
+        {
+            x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+
+            const float x_mean       = 0.0f;
+            const float x_stddev     = 2.5f;
+            const float noise_stddev = 0.04f;
+
+            resultRunningMean_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
+
+            resultRunningVariance_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+        }
+        else
+        {
+            const float x_mean       = 0.0f;
+            const float x_stddev     = 1.0f;
+            const float noise_stddev = 0.04f;
+
+            // input data in normal distribution
+            x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
+
+            // initialize the runningMean to be values with tiny variation to the mean of the x
+            // values
+            resultRunningMean_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
+
+            // initialize the runningVariance to be values with tiny variation to the variance of
+            // the x values
+            resultRunningVariance_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+        };
+    }
+    else
+    {
+        if constexpr(std::is_same<InOutDataType, int8_t>::value)
+            x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+        else
+            x.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0f, 5.0f}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            break;
+        case 1:
+            bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
+            break;
+        case 2:
+            bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
+    DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
+
+    // mean_dev or resultSaveMean_dev
+    DeviceMem resultSaveMean_dev(sizeof(AccDataType) *
+                                 resultSaveMean_ref.mDesc.GetElementSpaceSize());
+    // meansquare_dev or resultSaveInvVariance_dev
+    DeviceMem resultSaveInvVariance_dev(sizeof(AccDataType) *
+                                        resultSaveInvVariance_ref.mDesc.GetElementSpaceSize());
+    // resultRunningMean_dev
+    DeviceMem resultRunningMean_dev(sizeof(AccDataType) *
+                                    resultRunningMean_ref.mDesc.GetElementSpaceSize());
+    // resultRunningVariance_dev
+    DeviceMem resultRunningVariance_dev(sizeof(AccDataType) *
+                                        resultRunningVariance_ref.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+    bnBias_dev.ToDevice(bnBias.mData.data());
+
+    if(updateMovingAverage)
+    {
+        resultRunningMean_dev.ToDevice(resultRunningMean_ref.mData.data());
+        resultRunningVariance_dev.ToDevice(resultRunningVariance_ref.mData.data());
+    };
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
+    ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
+    ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
+    ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    using DeviceBatchNormFwdInstance =
+        ck::tensor_operation::device::DeviceBatchNormFwdImpl<InOutDataType,
+                                                             InOutDataType,
+                                                             AccDataType,
+                                                             AccDataType,   // ScaleDataType
+                                                             AccDataType,   // BiasDataType
+                                                             AccDataType,   // MeanVarDataType
+                                                             PassThroughOp, // YElementwiseOp
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             UseMultiblockInK,
+                                                             256,
+                                                             16,
+                                                             16,
+                                                             1,
+                                                             2,
+                                                             0,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1>;
+
+    auto batchnorm_fwd = DeviceBatchNormFwdInstance{};
+
+    auto argument_ptr = batchnorm_fwd.MakeArgumentPointer(
+        i_inOutLengths,
+        i_inOutStrides,
+        i_inOutStrides,
+        {0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
+        i_scaleBiasMeanVarLengths,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        x_dev.GetDeviceBuffer(),
+        bnScale_dev.GetDeviceBuffer(),
+        bnBias_dev.GetDeviceBuffer(),
+        epsilon,
+        PassThroughOp{},
+        y_dev.GetDeviceBuffer(),
+        saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
+        saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
+        averageFactor,
+        updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
+        updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
+
+    if(!batchnorm_fwd.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
+                     "exiting!"
+                  << std::endl;
+        return (false);
+    };
+
+    size_t workspace_sz = batchnorm_fwd.GetWorkSpaceSize(argument_ptr.get());
+
+    DeviceMem workspace_dev(workspace_sz);
+
+    batchnorm_fwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = batchnorm_fwd.MakeInvokerPointer();
+
+    if(time_kernel)
+    {
+        float avg_time   = 0.0f;
+        size_t num_bytes = 0;
+
+        size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
+        size_t invariant_length = inOutLengths[3];
+
+        avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        // inputing of x, scale, bias, outputing of y
+        num_bytes +=
+            total_length * sizeof(InOutDataType) * 2 + invariant_length * sizeof(AccDataType) * 2;
+
+        // outputing of mean, inv-variance
+        num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(AccDataType) * 2 : 0;
+
+        // updating of moving mean, variance
+        num_bytes += updateMovingAverage ? invariant_length * sizeof(AccDataType) * 4 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    }
+    else
+        (void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+
+        using ReferenceBatchNormFwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormFwd<InOutDataType,
+                                                              InOutDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumReduceDim>;
+
+        auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
+
+        auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
+            i_inOutLengths,
+            i_inOutStrides,
+            i_inOutStrides,
+            {0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
+            i_scaleBiasMeanVarLengths,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            x.mData.data(),
+            bnScale.mData.data(),
+            bnBias.mData.data(),
+            epsilon,
+            PassThroughOp{},
+            y_ref.mData.data(),
+            saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
+            saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
+            averageFactor,
+            updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
+            updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
+
+        if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters seems not supported by the BatchNorm reference "
+                         "instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        y_dev.FromDevice(y.mData.data());
+        pass = pass && ck::utils::check_err(y, y_ref);
+
+        if(updateMovingAverage)
+        {
+            Tensor<AccDataType> resultRunningMean(scaleBiasMeanVarLengths);
+            Tensor<AccDataType> resultRunningVariance(scaleBiasMeanVarLengths);
+
+            resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
+            resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
+
+            pass = pass && ck::utils::check_err(resultRunningMean, resultRunningMean_ref);
+            pass = pass && ck::utils::check_err(resultRunningVariance, resultRunningVariance_ref);
+        };
+
+        if(saveMeanAndInvVariance)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            Tensor<AccDataType> resultSaveMean(scaleBiasMeanVarLengths);
+            Tensor<AccDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
+
+            resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
+            resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
+
+            pass = pass && ck::utils::check_err(resultSaveMean, resultSaveMean_ref);
+            pass = pass && ck::utils::check_err(resultSaveInvVariance, resultSaveInvVariance_ref);
+        };
+    };
+
+    return (pass);
+};
+
+const double epsilon              = std::numeric_limits<float>::epsilon();
+static const double averageFactor = 0.1;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormFwdArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
+                                                                    arg.init_method,
+                                                                    arg.time_kernel,
+                                                                    arg.inOutLengths,
+                                                                    arg.updateMovingAverage,
+                                                                    arg.saveMeanAndInvVariance,
+                                                                    averageFactor,
+                                                                    epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<float, float, true>(arg.do_verification,
+                                                               arg.init_method,
+                                                               arg.time_kernel,
+                                                               arg.inOutLengths,
+                                                               arg.updateMovingAverage,
+                                                               arg.saveMeanAndInvVariance,
+                                                               averageFactor,
+                                                               epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<float, float, false>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
+        }
+        else if(arg.data_type == 3)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<int8_t, float, true>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<int8_t, float, false>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
+                                                                      arg.init_method,
+                                                                      arg.time_kernel,
+                                                                      arg.inOutLengths,
+                                                                      arg.updateMovingAverage,
+                                                                      arg.saveMeanAndInvVariance,
+                                                                      averageFactor,
+                                                                      epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<double, double, true>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<double, double, false>(arg.do_verification,
+                                                                  arg.init_method,
+                                                                  arg.time_kernel,
+                                                                  arg.inOutLengths,
+                                                                  arg.updateMovingAverage,
+                                                                  arg.saveMeanAndInvVariance,
+                                                                  averageFactor,
+                                                                  epsilon);
+        }
+    }
+    else
+    {
+        pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(true,
+                                                            2,
+                                                            false, // don't time kernel
+                                                            {128, 16, 6, 512},
+                                                            true,
+                                                            true,
+                                                            averageFactor,
+                                                            epsilon);
+
+        pass = pass && bnorm_fwd_nhwc_test<ck::half_t, float, false>(true,
+                                                                     2,
+                                                                     false, // don't time kernel
+                                                                     {128, 16, 3, 1024},
+                                                                     true,
+                                                                     true,
+                                                                     averageFactor,
+                                                                     epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}

3rdparty/composable_kernel/example/34_batchnorm/batchnorm_infer_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cassert>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+
+#include "batchnorm_common.hpp"
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          ck::index_t Rank,
+          ck::index_t NumBatchNormReduceDim,
+          bool fastest_dim_is_reduced = false>
+int bnorm_infer(
+    bool time_kernel,
+    const std::array<int, NumBatchNormReduceDim> reduceDims,
+    const std::array<ck::index_t, Rank> xyLengths,
+    const std::array<ck::index_t, Rank> xStrides,
+    const std::array<ck::index_t, Rank> yStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+    const void* p_x,
+    const void* p_scale,
+    const void* p_bias,
+    double epsilon,
+    const void* p_estimatedMean,
+    const void* p_estimatedVariance,
+    void* p_y)
+{
+    (void)bnScaleBiasMeanVarLengths;
+
+    static_assert(NumBatchNormReduceDim < Rank,
+                  "Invalid number of reduced dimensions for batchnorm!");
+
+    using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<XDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
+                                                                                  // variance,
+                                                                                  // scale,
+                                                                                  // bias,
+        ck::Tuple<YDataType>,                                                     // y
+        NormalizeInInfer,
+        Rank,
+        2,                           // MPerthread
+        ck::Sequence<1, 1, 1, 1, 1>, // x, mean, variance, scale, bias
+        ck::Sequence<1>>;            // scalarPerVector: y
+
+    auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims);
+    std::array<ck::index_t, Rank> aligned_bnScaleStrides{0};
+    std::array<ck::index_t, Rank> aligned_bnBiasStrides{0};
+    std::array<ck::index_t, Rank> aligned_bnMeanVarStrides{0};
+
+    int i = 0;
+    for(auto dim : invariantDims)
+    {
+        assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]);
+
+        aligned_bnScaleStrides[dim]   = bnScaleStrides[i];
+        aligned_bnBiasStrides[dim]    = bnBiasStrides[i];
+        aligned_bnMeanVarStrides[dim] = bnMeanVarStrides[i];
+        i++;
+    };
+
+    int32_t reduceLength = 1;
+
+    for(auto dim : reduceDims)
+        reduceLength *= xyLengths[dim];
+
+    int32_t invariantLength = 1;
+
+    for(auto dim : invariantDims)
+        invariantLength *= xyLengths[dim];
+
+    size_t total_length = static_cast<size_t>(invariantLength) * reduceLength;
+
+    float avg_time        = 0.0f;
+    std::size_t num_bytes = 0;
+
+    auto dev_normalize = DeviceNormalizeInstance{};
+
+    auto argument_ptr1 = dev_normalize.MakeArgumentPointer(
+        xyLengths,
+        {xStrides,
+         aligned_bnMeanVarStrides,
+         aligned_bnMeanVarStrides,
+         aligned_bnScaleStrides,
+         aligned_bnBiasStrides},
+        {yStrides},
+        {p_x, p_estimatedMean, p_estimatedVariance, p_scale, p_bias},
+        {p_y},
+        NormalizeInInfer{epsilon});
+
+    if(!dev_normalize.IsSupportedArgument(argument_ptr1.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the Devic, exiting!"
+                  << std::endl;
+
+        return (-1);
+    };
+
+    auto invoker_ptr1 = dev_normalize.MakeInvokerPointer();
+
+    avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel});
+
+    num_bytes += total_length * sizeof(XDataType) +
+                 invariantLength *
+                     (sizeof(ScaleDataType) + sizeof(BiasDataType) + 2 * sizeof(MeanVarDataType)) +
+                 total_length * sizeof(YDataType);
+
+    if(time_kernel)
+    {
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    };
+
+    return (0);
+};