Merge branch 'develop' into mi300

example/26_contraction/CMakeLists.txt

 add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp)
 add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
+
+add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
+add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)

example/26_contraction/contraction_bilinear_xdl_fp64.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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_contraction_multiple_d_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/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DDataType        = F64;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceKNNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceMKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceMNNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
+struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_ms_ks,
+                 const Tensor<BDataType>& b_ns_ks,
+                 Tensor<EDataType>& e_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_ms_ks_{a_ms_ks},
+              b_ns_ks_{b_ns_ks},
+              e_ms_ns_{e_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_ms_ks_;
+        const Tensor<BDataType>& b_ns_ks_;
+        Tensor<EDataType>& e_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_M2_N2_K2::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
+                const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    for(int k1 = 0; k1 < K1; ++k1)
+                    {
+                        AccDataType v_a;
+                        AccDataType v_b;
+
+                        arg.a_element_op_(
+                            v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
+                        arg.b_element_op_(
+                            v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
+
+                        v_acc += v_a * v_b;
+                    }
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
+                             const Tensor<BDataType>& b_ns_ks,
+                             Tensor<EDataType>& e_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{a_ms_ks, b_ns_ks, e_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M2_N2_K2"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    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 == 28)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t M0 = std::stoi(argv[4]);
+        const ck::index_t M1 = std::stoi(argv[5]);
+
+        const ck::index_t N0 = std::stoi(argv[6]);
+        const ck::index_t N1 = std::stoi(argv[7]);
+
+        const ck::index_t K0 = std::stoi(argv[8]);
+        const ck::index_t K1 = std::stoi(argv[9]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[22]), std::stoi(argv[23]), std::stoi(argv[24]), std::stoi(argv[25])};
+
+        alpha = std::stof(argv[26]);
+        beta  = std::stof(argv[27]);
+    }
+    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 7: M0, M1, N0, N1, K0, K1\n");
+        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg18 to 21: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg22 to 25: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg26 to 27: alpha, beta\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+    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, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                  NumDimN,
+                                                                  NumDimK,
+                                                                  ADataType,
+                                                                  BDataType,
+                                                                  CShuffleDataType,
+                                                                  AccDataType,
+                                                                  AElementOp,
+                                                                  BElementOp,
+                                                                  PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1),
+                                       d_ms_ns(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/26_contraction/contraction_scale_xdl_fp64.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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_contraction_multiple_d_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/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceOpInstanceKKN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceMKN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceMNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
+struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_ms_ks,
+                 const Tensor<BDataType>& b_ns_ks,
+                 Tensor<EDataType>& e_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_ms_ks_{a_ms_ks},
+              b_ns_ks_{b_ns_ks},
+              e_ms_ns_{e_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_ms_ks_;
+        const Tensor<BDataType>& b_ns_ks_;
+        Tensor<EDataType>& e_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_M2_N2_K2::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
+                const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    for(int k1 = 0; k1 < K1; ++k1)
+                    {
+                        AccDataType v_a;
+                        AccDataType v_b;
+
+                        arg.a_element_op_(
+                            v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
+                        arg.b_element_op_(
+                            v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
+
+                        v_acc += v_a * v_b;
+                    }
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
+                             const Tensor<BDataType>& b_ns_ks,
+                             Tensor<EDataType>& e_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{a_ms_ks, b_ns_ks, e_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M2_N2_K2"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float scale = 1.f;
+
+    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 == 23)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t M0 = std::stoi(argv[4]);
+        const ck::index_t M1 = std::stoi(argv[5]);
+
+        const ck::index_t N0 = std::stoi(argv[6]);
+        const ck::index_t N1 = std::stoi(argv[7]);
+
+        const ck::index_t K0 = std::stoi(argv[8]);
+        const ck::index_t K1 = std::stoi(argv[9]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
+
+        scale = std::stof(argv[22]);
+    }
+    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 9: M0, M1, N0, N1, K0, K1\n");
+        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg18 to 21: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg22: scale\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{scale};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 0>{},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 0>{},
+                                    std::array<std::vector<ck::index_t>, 0>{},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * M * N;
+
+    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, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                  NumDimN,
+                                                                  NumDimK,
+                                                                  ADataType,
+                                                                  BDataType,
+                                                                  CShuffleDataType,
+                                                                  AccDataType,
+                                                                  AElementOp,
+                                                                  BElementOp,
+                                                                  PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/27_layernorm/layernorm_blockwise.cpp

@@ -20,12 +20,12 @@
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
 
-using XDataType     = ck::half_t;
-using GammaDataType = ck::half_t;
-using BetaDataType  = ck::half_t;
-using YDataType     = ck::half_t;
-using AccDataType   = float;
-using PassThrough   = ck::tensor_operation::element_wise::PassThrough;
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+using PassThrough     = ck::tensor_operation::element_wise::PassThrough;
 
 constexpr int Rank         = 2;
 constexpr int NumReduceDim = 1;
@@ -34,7 +34,7 @@ using DeviceInstance =
     ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
                                                           GammaDataType,
                                                           BetaDataType,
-                                                          AccDataType,
+                                                          ComputeDataType,
                                                           YDataType,
                                                           PassThrough,
                                                           Rank,
@@ -121,7 +121,7 @@ int main()
                                                                                  GammaDataType,
                                                                                  BetaDataType,
                                                                                  YDataType,
-                                                                                 AccDataType,
+                                                                                 ComputeDataType,
                                                                                  PassThrough,
                                                                                  Rank,
                                                                                  NumReduceDim>;

example/29_batched_gemm_bias_e_permute/CMakeLists.txt

 add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
+
+if(GPU_TARGETS MATCHES "gfx1100")
+    add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
+endif()

example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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_contraction_multiple_d_wmma_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/numeric.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 Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceOpInstanceKKNN =
+    ck::tensor_operation::device::DeviceBatchedContractionMultipleD_Wmma_CShuffle<NumDimG,
+                                                                                  NumDimM,
+                                                                                  NumDimN,
+                                                                                  NumDimK,
+                                                                                  ADataType,
+                                                                                  BDataType,
+                                                                                  DsDataType,
+                                                                                  EDataType,
+                                                                                  AccDataType,
+                                                                                  CShuffleDataType,
+                                                                                  AElementOp,
+                                                                                  BElementOp,
+                                                                                  CDEElementOp,
+                                                                                  GemmSpec,
+                                                                                  ABSpec,
+                                                                                  ABSpec,
+                                                                                  DESpec,
+                                                                                  256,
+                                                                                  128,
+                                                                                  256,
+                                                                                  8,
+                                                                                  8,
+                                                                                  16,
+                                                                                  16,
+                                                                                  4,
+                                                                                  4,
+                                                                                  S<4, 64, 1>,
+                                                                                  S<1, 0, 2>,
+                                                                                  S<1, 0, 2>,
+                                                                                  2,
+                                                                                  8,
+                                                                                  8,
+                                                                                  true,
+                                                                                  S<4, 64, 1>,
+                                                                                  S<1, 0, 2>,
+                                                                                  S<1, 0, 2>,
+                                                                                  2,
+                                                                                  8,
+                                                                                  8,
+                                                                                  true,
+                                                                                  1,
+                                                                                  1,
+                                                                                  S<1, 32, 1, 8>,
+                                                                                  8>;
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G2_M2_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    ck::index_t G0 = 1;
+    ck::index_t G1 = 2;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 128;
+
+    ck::index_t N0 = 16;
+    ck::index_t N1 = 256;
+
+    ck::index_t K0 = 2048;
+
+    // A[G0, G1, M0, M1, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[G0, G1, N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
+    // E[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 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
+    {
+        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");
+        exit(0);
+    }
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+    Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+        break;
+    }
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t G =
+        ck::accumulate_n<ck::index_t>(e_gs_ms_ns_lengths.begin(), NumDimG, 1, std::multiplies<>{});
+
+    ck::index_t M = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
+    std::cout << "GMNK=" << G << ", " << M << ", " << N << ", " << K << std::endl;
+    std::size_t flop      = std::size_t(2) * G * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
+                            sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * M * N;
+
+    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, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t g0 = 0; g0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[0]; ++g0)
+        {
+            for(size_t g1 = 0; g1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[1]; ++g1)
+            {
+                for(size_t m0 = 0; m0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[2]; ++m0)
+                {
+                    for(size_t m1 = 0; m1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[3]; ++m1)
+                    {
+                        for(size_t n0 = 0; n0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[4]; ++n0)
+                        {
+                            for(size_t n1 = 0; n1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[5];
+                                ++n1)
+                            {
+                                cde_element_op(e_gs_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
+                                               c_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
+                                               d_gs_ms_ns(g0, g1, m0, m1, n0, n1));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt

@@ -16,6 +16,9 @@ if(USE_BITINT_EXTENSION_INT4)
   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
 endif() # USE_BITINT_EXTENSION_INT4
 
+if(GPU_TARGETS MATCHES "gfx1100")
+  add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
+endif()
 
 add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
 

example/30_grouped_conv_fwd_multiple_d/common.hpp

@@ -137,7 +137,7 @@ inline bool parse_cmd_args(int argc,
 
         const ck::index_t num_dim_spatial = std::stoi(argv[4]);
         conv_param                        = ck::utils::conv::parse_conv_param(
-            num_dim_spatial, threshold_to_catch_partial_args, argv);
+            num_dim_spatial, threshold_to_catch_partial_args + 1, argv);
     }
     else
     {

example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <iostream>
+#include <string>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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/convolution_parameter.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using BF16 = ck::bhalf_t;
+using FP16 = ck::half_t;
+using FP32 = float;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using I4 = ck::int4_t;
+#endif
+using I8  = std::int8_t;
+using I32 = std::int32_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <typename InputLay, typename WeightLay, typename OutputLay>
+struct CommonLayoutSetting
+{
+    using InputLayout  = InputLay;
+    using WeightLayout = WeightLay;
+    using OutputLayout = OutputLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct CommonLayoutSettingSelector;
+
+namespace ctl = ck::tensor_layout::convolution;
+
+template <>
+struct CommonLayoutSettingSelector<1> final
+    : CommonLayoutSetting<ctl::G_NW_C, ctl::G_K_X_C, ctl::G_NW_K>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<2> final
+    : CommonLayoutSetting<ctl::G_NHW_C, ctl::G_K_YX_C, ctl::G_NHW_K>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<3> final
+    : CommonLayoutSetting<ctl::G_NDHW_C, ctl::G_K_ZYX_C, ctl::G_NDHW_K>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using InputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::InputLayout;
+
+template <ck::index_t NDimSpatial>
+using WeightLayout = typename CommonLayoutSettingSelector<NDimSpatial>::WeightLayout;
+
+template <ck::index_t NDimSpatial>
+using OutputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::OutputLayout;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+};
+
+#define DefaultConvParam                                                       \
+    ck::utils::conv::ConvParam                                                 \
+    {                                                                          \
+        2, 32, 2, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, { 1, 1 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_param)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_param                        = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args + 1, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
+
+inline HostTensorDescriptor make_input_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.C_, conv_param.input_spatial_lengths_[0]},
+            {
+                conv_param.C_,                                                        // g
+                conv_param.input_spatial_lengths_[0] * conv_param.G_ * conv_param.C_, // n
+                1,                                                                    // c
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.C_,
+             conv_param.input_spatial_lengths_[0],
+             conv_param.input_spatial_lengths_[1]},
+            {
+                conv_param.C_, // g
+                conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
+                    conv_param.G_ * conv_param.C_,                                    // n
+                1,                                                                    // c
+                conv_param.input_spatial_lengths_[1] * conv_param.G_ * conv_param.C_, // hi
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.C_,
+             conv_param.input_spatial_lengths_[0],
+             conv_param.input_spatial_lengths_[1],
+             conv_param.input_spatial_lengths_[2]},
+            {
+                conv_param.C_, // g
+                conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
+                    conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // n
+                1,                                                                        // c
+                conv_param.input_spatial_lengths_[1] * conv_param.input_spatial_lengths_[2] *
+                    conv_param.G_ * conv_param.C_,                                    // di
+                conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // hi
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_weight_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.K_, conv_param.C_, conv_param.filter_spatial_lengths_[0]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] * conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.C_,                 // k
+                1,                                                                     // c
+                conv_param.C_                                                          // x
+            });
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.K_,
+             conv_param.C_,
+             conv_param.filter_spatial_lengths_[0],
+             conv_param.filter_spatial_lengths_[1]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
+                    conv_param.filter_spatial_lengths_[1] * conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
+                    conv_param.C_,                                     // k
+                1,                                                     // c
+                conv_param.filter_spatial_lengths_[1] * conv_param.C_, // y
+                conv_param.C_                                          // x
+            });
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.K_,
+             conv_param.C_,
+             conv_param.filter_spatial_lengths_[0],
+             conv_param.filter_spatial_lengths_[1],
+             conv_param.filter_spatial_lengths_[2]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
+                    conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
+                    conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
+                    conv_param.filter_spatial_lengths_[2] * conv_param.C_, // k
+                1,                                                         // c
+                conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
+                    conv_param.C_,                                     // z
+                conv_param.filter_spatial_lengths_[2] * conv_param.C_, // y
+                conv_param.C_                                          // x
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_bias_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
+            {
+                conv_param.K_, // g
+                0,             // k
+                1,             // c
+                0              // x
+            });
+    case 2:
+        return HostTensorDescriptor({conv_param.G_,
+                                     conv_param.N_,
+                                     conv_param.K_,
+                                     conv_param.output_spatial_lengths_[0],
+                                     conv_param.output_spatial_lengths_[1]},
+                                    {
+                                        conv_param.K_, // g
+                                        0,             // n
+                                        1,             // k
+                                        0,             // ho
+                                        0              // wo
+                                    });
+    case 3:
+        return HostTensorDescriptor({conv_param.G_,
+                                     conv_param.N_,
+                                     conv_param.K_,
+                                     conv_param.output_spatial_lengths_[0],
+                                     conv_param.output_spatial_lengths_[1],
+                                     conv_param.output_spatial_lengths_[2]},
+                                    {
+                                        conv_param.K_, // g
+                                        0,             // n
+                                        1,             // k
+                                        0,             // z
+                                        0,             // y
+                                        0              // x
+                                    });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_output_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
+            {
+                conv_param.K_,                                                         // g
+                conv_param.output_spatial_lengths_[0] * conv_param.G_ * conv_param.K_, // n
+                1,                                                                     // k
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.K_,
+             conv_param.output_spatial_lengths_[0],
+             conv_param.output_spatial_lengths_[1]},
+            {
+                conv_param.K_, // g
+                conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
+                    conv_param.G_ * conv_param.K_,                                     // n
+                1,                                                                     // k
+                conv_param.output_spatial_lengths_[1] * conv_param.G_ * conv_param.K_, // ho
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.K_,
+             conv_param.output_spatial_lengths_[0],
+             conv_param.output_spatial_lengths_[1],
+             conv_param.output_spatial_lengths_[2]},
+            {
+                conv_param.K_, // g
+                conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
+                    conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // n
+                1,                                                                         // k
+                conv_param.output_spatial_lengths_[1] * conv_param.output_spatial_lengths_[2] *
+                    conv_param.G_ * conv_param.K_,                                     // do
+                conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // ho
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}

example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common_wmma.hpp"
+
+// kernel data types
+using InKernelDataType       = FP16;
+using WeiKernelDataType      = FP16;
+using AccDataType            = FP32;
+using CShuffleDataType       = FP16;
+using BiasKernelDataType     = FP16;
+using ResidualKernelDataType = FP16;
+using OutKernelDataType      = FP16;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_wmma_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }

example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+template <typename BiasLay, typename ResidualLay>
+struct LayoutSetting
+{
+    using BiasLayout     = BiasLay;
+    using ResidualLayout = ResidualLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct LayoutSettingSelector;
+
+template <>
+struct LayoutSettingSelector<1> final : LayoutSetting<ctl::G_K, ctl::G_NW_K>
+{
+};
+
+template <>
+struct LayoutSettingSelector<2> final : LayoutSetting<ctl::G_K, ctl::G_NHW_K>
+{
+};
+
+template <>
+struct LayoutSettingSelector<3> final : LayoutSetting<ctl::G_K, ctl::G_NDHW_K>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using BiasLayout = typename LayoutSettingSelector<NDimSpatial>::BiasLayout;
+
+template <ck::index_t NDimSpatial>
+using ResidualLayout = typename LayoutSettingSelector<NDimSpatial>::ResidualLayout;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
+        NDimSpatial,
+        InputLayout<NDimSpatial>,
+        WeightLayout<NDimSpatial>,
+        ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>,
+        OutputLayout<NDimSpatial>,
+        InKernelDataType,
+        WeiKernelDataType,
+        ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
+        OutKernelDataType,
+        AccDataType,
+        CShuffleDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        256,         // BlockSize
+        128,         // MPerBlock
+        128,         // NPerBlock
+        4,           // K0PerBlock
+        8,           // K1
+        16,          // MPerWMMA
+        16,          // NPerWMMA
+        4,           // MRepeat
+        2,           // NRepeat
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        true,        // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        true,        // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8>;
+
+template <ck::index_t NDimSpatial>
+using HostConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                         InUserDataType,
+                                                                         WeiUserDataType,
+                                                                         CShuffleDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         PassThrough>;
+
+template <ck::index_t NDimSpatial>
+bool run_grouped_conv_fwd_bias_relu_add(const ExecutionConfig& config,
+                                        const ck::utils::conv::ConvParam& conv_param)
+{
+    static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
+
+    const auto in_g_n_c_wis_desc   = make_input_descriptor(conv_param);
+    const auto wei_g_k_c_xs_desc   = make_weight_descriptor(conv_param);
+    const auto bias_g_n_k_wos_desc = make_bias_descriptor(conv_param);
+    const auto out_g_n_k_wos_desc  = make_output_descriptor(conv_param);
+
+    Tensor<InUserDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiUserDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutUserDataType> bias(bias_g_n_k_wos_desc);
+    Tensor<OutUserDataType> residual(bias_g_n_k_wos_desc);
+    Tensor<OutUserDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutKernelDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "residual: " << residual.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InUserDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiUserDataType>{-5, 5});
+        bias.GenerateTensorValue(GeneratorTensor_2<OutUserDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InUserDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiUserDataType>{-0.5, 0.5});
+        bias.GenerateTensorValue(GeneratorTensor_3<OutUserDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InKernelDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiKernelDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutKernelDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem residual_device_buf(sizeof(OutKernelDataType) * residual.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutKernelDataType) * out_device.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<InKernelDataType> in_converted(in);
+    const Tensor<WeiKernelDataType> wei_converted(wei);
+    const Tensor<OutKernelDataType> bias_converted(bias);
+    const Tensor<OutKernelDataType> residual_converted(residual);
+
+    in_device_buf.ToDevice(in_converted.mData.data());
+    wei_device_buf.ToDevice(wei_converted.mData.data());
+    bias_device_buf.ToDevice(bias_converted.mData.data());
+    residual_device_buf.ToDevice(residual_converted.mData.data());
+#else
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+    residual_device_buf.ToDevice(residual.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_n_k_wos_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_n_k_wos_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(bias_g_n_k_wos_desc.GetLengths(), d1_g_n_k_wos_lengths);
+    copy(bias_g_n_k_wos_desc.GetStrides(), d1_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv    = DeviceConvFwdInstance<NDimSpatial>{};
+    auto invoker = conv.MakeInvoker();
+    auto argument =
+        conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                          wei_device_buf.GetDeviceBuffer(),
+                          std::array<const void*, 2>{bias_device_buf.GetDeviceBuffer(),
+                                                     residual_device_buf.GetDeviceBuffer()},
+                          out_device_buf.GetDeviceBuffer(),
+                          a_g_n_c_wis_lengths,
+                          a_g_n_c_wis_strides,
+                          b_g_k_c_xs_lengths,
+                          b_g_k_c_xs_strides,
+                          std::array<std::array<ck::index_t, NDimSpatial + 3>, 2>{
+                              {d0_g_n_k_wos_lengths, d1_g_n_k_wos_lengths}},
+                          std::array<std::array<ck::index_t, NDimSpatial + 3>, 2>{
+                              {d0_g_n_k_wos_strides, d1_g_n_k_wos_strides}},
+                          e_g_n_k_wos_lengths,
+                          e_g_n_k_wos_strides,
+                          conv_filter_strides,
+                          conv_filter_dilations,
+                          input_left_pads,
+                          input_right_pads,
+                          InElementOp{},
+                          WeiElementOp{},
+                          OutElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InUserDataType, WeiUserDataType, OutUserDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv     = HostConvFwdInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach([&](auto&, auto idx) {
+            OutElementOp{}(out_host(idx), c_host(idx), bias(idx), residual(idx));
+        });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<OutUserDataType> out_device_converted(out_device);
+
+        return ck::utils::check_err(
+            out_device_converted, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+#else
+        return ck::utils::check_err(
+            out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+#endif
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_fwd_bias_relu_add_example(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return false;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return run_grouped_conv_fwd_bias_relu_add<1>(config, conv_param);
+    case 2: return run_grouped_conv_fwd_bias_relu_add<2>(config, conv_param);
+    case 3: return run_grouped_conv_fwd_bias_relu_add<3>(config, conv_param);
+    }
+
+    return false;
+}

example/42_groupnorm/groupnorm_sigmoid_fp16.cpp

@@ -23,11 +23,11 @@
 constexpr int Rank         = 5;
 constexpr int NumReduceDim = 3;
 
-using XDataType     = ck::half_t;
-using GammaDataType = ck::half_t;
-using BetaDataType  = ck::half_t;
-using YDataType     = ck::half_t;
-using AccDataType   = float;
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
 
 struct YElementOp
 {
@@ -50,7 +50,7 @@ using DeviceInstance =
     ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
                                                           GammaDataType,
                                                           BetaDataType,
-                                                          AccDataType,
+                                                          ComputeDataType,
                                                           YDataType,
                                                           YElementOp,
                                                           Rank,
@@ -157,7 +157,7 @@ int main(int argc, char* argv[])
                                                                                  GammaDataType,
                                                                                  BetaDataType,
                                                                                  YDataType,
-                                                                                 AccDataType,
+                                                                                 ComputeDataType,
                                                                                  YElementOp>;
 
         ReferenceInstance ref;

example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc

@@ -53,7 +53,6 @@ bool run_gemm_add_multiply(const ProblemSize& problem_size, const ExecutionConfi
     DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
     DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
 
-
     a_device_buf.ToDevice(a_m_k.mData.data());
     b_device_buf.ToDevice(b_k_n.mData.data());
     d0_device_buf.ToDevice(d0_m_n.mData.data());
@@ -84,8 +83,8 @@ bool run_gemm_add_multiply(const ProblemSize& problem_size, const ExecutionConfi
 
     if(!device_op.IsSupportedArgument(argument))
     {
-	    std::cout << "wrong! this device_op instance does not support this problem" << std::endl;
-	    return true;
+        std::cout << "wrong! this device_op instance does not support this problem" << std::endl;
+        return true;
     }
 
     float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});

example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt

+add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute.cpp)

example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.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"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using C0DEElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+using Acc0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+constexpr static 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 F16              = ck::half_t;
+using F32              = float;
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using D0DataType       = F16;
+using Acc0BiasDataType = ck::Tuple<D0DataType>;
+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 DeviceOpInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        C0DEElementOp,
+        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>;
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    int G0      = 3;
+    int G1      = 2;
+    int M       = 1024;
+    int N       = 1024;
+    int K       = 64;
+    int O       = 64;
+    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 == 11)
+    {
+        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]);
+    }
+    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");
+        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{
+        M * G1 * K, K, G1 * K, 1}; // A layout [G0, M, G1, 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{
+        N * G1 * K, K, G1 * K, 1}; // B0 layout [G0, N, G1, 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{
+        N * G1 * O, O, 1, G1 * O}; // B1 layout [G0, N, G1, 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{
+        M * G1 * O, O, G1 * O, 1}; // C layout [G0, M, G1, O]
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    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<D0DataType> d0_gs_ms_ns(d0_gs_ms_ns_lengths, d0_gs_ms_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});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-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});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-1, 1});
+        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>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        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>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    DeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    DeviceMem d0_device_buf(sizeof(D0DataType) * G0 * G1 * M * N);
+    DeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    DeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    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());
+    d0_device_buf.ToDevice(d0_gs_ms_ns.mData.data());
+
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto c0de_element_op = C0DEElementOp{alpha};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    auto argument = device_op.MakeArgument(
+        static_cast<const ADataType*>(a_device_buf.GetDeviceBuffer()),
+        static_cast<const B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+        static_cast<const B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+        {},                                                    // 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>{
+            d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+        std::array<std::vector<ck::index_t>, 1>{
+            d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+        {},                       // acc1_biases_gs_ms_os_lengths
+        {},                       // acc1_biases_gs_ms_os_strides
+        a_element_op,
+        b0_element_op,
+        c0de_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t BatchCount = G0 * G1;
+    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 + sizeof(D0DataType) * M * N) *
+        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, "
+              << 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
+        Tensor<D0DataType> d0_g_m_n({BatchCount, M, N});
+
+        // 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);
+        });
+        d0_gs_ms_ns.ForEach([&](auto& self, auto idx) {
+            d0_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = 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);
+
+        acc0_g_m_n.ForEach([&](auto&, auto idx) {
+            c0de_element_op(acc0_g_m_n(idx), acc0_g_m_n(idx), d0_g_m_n(idx));
+        });
+        // masking
+        const auto mask = DeviceOpInstance::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;
+
+        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;
+}

include/ck/ck.hpp

@@ -172,13 +172,6 @@
 // tuning parameter
 #define CK_WORKAROUND_SWDEV_325164 0
 
-// workaround: a BF16 attention kernel for gfx908 is likely affected by a compiler issue
-#ifdef __gfx908__
-#define CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE 1
-#else // __gfx90a__, ...
-#define CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE 0
-#endif // __gfx908__
-
 // flag to enable (1) or disable (0) the debugging output in some kernels
 #define DEBUG_LOG 0
 

include/ck/host_utility/kernel_launch.hpp

@@ -20,6 +20,7 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
 #if CK_TIME_KERNEL
     if(stream_config.time_kernel_)
     {
+#if DEBUG_LOG
         printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
                __func__,
                grid_dim.x,
@@ -29,15 +30,15 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
                block_dim.y,
                block_dim.z);
 
-        const int nrepeat = 10;
-
         printf("Warm up 1 time\n");
-
+#endif
         // warm up
         kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
 
+        const int nrepeat = 10;
+#if DEBUG_LOG
         printf("Start running %d times...\n", nrepeat);
-
+#endif
         hipEvent_t start, stop;
 
         hip_check_error(hipEventCreate(&start));

include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp

@@ -26,9 +26,9 @@ template <index_t NumDimG,
           typename Acc1BiasDataType,
           typename AElementwiseOperation,
           typename B0ElementwiseOperation,
-          typename Acc0ElementwiseOperation,
+          typename C0DEElementwiseOperation,
           typename B1ElementwiseOperation,
-          typename CElementwiseOperation,
+          typename C1DEElementwiseOperation,
           MaskingSpecialization MaskingSpec>
 struct DeviceBatchedGemmSoftmaxGemmPermute : public BaseOperator
 {
@@ -58,9 +58,9 @@ struct DeviceBatchedGemmSoftmaxGemmPermute : public BaseOperator
             acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
         AElementwiseOperation a_element_op,
         B0ElementwiseOperation b0_element_op,
-        Acc0ElementwiseOperation acc0_element_op,
+        C0DEElementwiseOperation c0de_element_op,
         B1ElementwiseOperation b1_element_op,
-        CElementwiseOperation c_element_op) = 0;
+        C1DEElementwiseOperation c1de_element_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };

include/ck/tensor_operation/gpu/device/device_normalization.hpp

@@ -14,9 +14,9 @@ namespace device {
 template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
-          typename AccDataType,
+          typename ComputeDataType,
           typename YDataType,
-          typename AccElementwiseOperation,
+          typename YElementwiseOperation,
           index_t Rank,
           index_t NumReduceDim>
 struct DeviceNormalization : public BaseOperator
@@ -35,7 +35,7 @@ struct DeviceNormalization : public BaseOperator
                         void* p_y,
                         void* p_savedMean,
                         void* p_savedInvVar,
-                        AccElementwiseOperation acc_elementwise_op) = 0;
+                        YElementwiseOperation y_elementwise_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
@@ -43,17 +43,17 @@ struct DeviceNormalization : public BaseOperator
 template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
-          typename AccDataType,
+          typename ComputeDataType,
           typename YDataType,
-          typename AccElementwiseOperation,
+          typename YElementwiseOperation,
           index_t Rank,
           index_t NumReduceDim>
 using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization<XDataType,
                                                                    GammaDataType,
                                                                    BetaDataType,
-                                                                   AccDataType,
+                                                                   ComputeDataType,
                                                                    YDataType,
-                                                                   AccElementwiseOperation,
+                                                                   YElementwiseOperation,
                                                                    Rank,
                                                                    NumReduceDim>>;
 

include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.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/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+
+// NOTE: TensorSpecialization::Packed specialized tensor is "packed" in a sense that each inner
+// dimension in a dimension group (eg [G0, G1] in Gs, [M0, M1, M2] in Ms, etc.) are contiguous and
+// ordered. Not in a sense that the tensor [G0, G1, ..., M0, M1, ..., N0, N1...] can be permuted
+// while still being a contiguous, unpadded tensor. In other words, it merely degenerates into
+// TensorSpecialization::Default with NumDimG/M/N/K = 1
+//
+// Detail- Packed tensor satisfies
+//   stride_0 = 1
+//   stride_i = stride_{i - 1} * extent_{i - 1}
+// So tensor
+//   [G0, G1, G2, M, N]
+// transposed into tensor
+//   [G0, G2, G1, M, N]
+// with strides
+//   [G2 * G1 * M * N, G1 * M * N, M * N, N, 1]
+// is again a packed tensor. MakeGridDescriptor() currently just merges dimensions and ignores some
+// strides from input tensor extents so finer dimension information is lost. Merging dimensions is
+// essentially a degenerated case of TensorSpecialization::Default with NumDimG/M/N/K = 1.
+//
+// Might need to expose dimension order to the interface to fully support
+// TensorSpecialization::Packed in a traditional sense of "packed" tensor
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerWMMA,
+          ck::index_t NPerWMMA,
+          ck::index_t MRepeat,
+          ck::index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          ck::index_t NumPrefetch         = 1,
+          ck::LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
+struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
+    : public DeviceBatchedContractionMultipleD<NumDimG,
+                                               NumDimM,
+                                               NumDimN,
+                                               NumDimK,
+                                               ADataType,
+                                               BDataType,
+                                               DsDataType,
+                                               EDataType,
+                                               AElementwiseOperation,
+                                               BElementwiseOperation,
+                                               CDEElementwiseOperation>
+{
+    using DeviceOp                      = DeviceBatchedContractionMultipleD_Wmma_CShuffle;
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    // K1 = Max Vector Access Pixels
+    static constexpr auto K1Number = Number<K1>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock* K1};
+
+    // Assume: A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        assert(a_gs_ms_ks_lengths_vec.size() == NumDimG + NumDimM + NumDimK &&
+               a_gs_ms_ks_strides_vec.size() == NumDimG + NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(
+            a_gs_ms_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(
+            a_gs_ms_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        if constexpr(ASpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(M, K),
+                make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
+                           a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+        else
+        {
+            // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+            const auto a_grid_desc_ms_ks =
+                make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+            // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+            const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+                a_grid_desc_ms_ks,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+                make_tuple(mDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+    }
+
+    // Assume: B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        assert(b_gs_ns_ks_lengths_vec.size() == NumDimG + NumDimN + NumDimK &&
+               b_gs_ns_ks_strides_vec.size() == NumDimG + NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(
+            b_gs_ns_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(
+            b_gs_ns_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        if constexpr(BSpec == TensorSpecialization::Packed)
+        {
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(N, K),
+                make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
+                           b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+        else
+        {
+            // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+            const auto b_grid_desc_ns_ks =
+                make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+            // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+            const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+                b_grid_desc_ns_ks,
+                make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+                make_tuple(nDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(
+            e_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(
+            e_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(M, N),
+                make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
+                           e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+        else
+        {
+            // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_ms_ns =
+                make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+            // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+            const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_ms_ns,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_G_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                          const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_gs_ms_ns_lengths =
+            to_tuple(e_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_gs_ms_ns_strides =
+            to_tuple(e_gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for G0, G1, ...
+        constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds =
+            typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<NumDimG + NumDimM,
+                                                                  NumDimG + NumDimM + NumDimN,
+                                                                  1>::type{};
+
+        // lengths for G0, G1, ...
+        const auto gLengths = get_container_subset(e_gs_ms_ns_lengths, gDimIds);
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_gs_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_gs_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(G, M, N),
+                make_tuple(e_gs_ms_ns_strides[Number<NumDimG - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+        else
+        {
+            // naive tensor E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_gs_ms_ns =
+                make_naive_tensor_descriptor(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            // transformed tensor E[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+            // N2 * ...]
+            const auto e_grid_desc_g_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_gs_ms_ns,
+                make_tuple(make_merge_transform(gLengths),
+                           make_merge_transform(mLengths),
+                           make_merge_transform(nLengths)),
+                make_tuple(gDimIds, mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                         ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    static auto MakeDsGridDescriptor_G_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_G_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                           ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // Gridwise descriptor, mapping to whole given provblem.
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    using DsGridDesc_G_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_G_M_N({}, {}))>;
+    using EGridDesc_G_M_N  = decltype(MakeEGridDescriptor_G_M_N({}, {}));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t batch_stride_A,
+                                       index_t batch_stride_B,
+                                       DsGridDesc_G_M_N ds_grid_desc_g_m_n,
+                                       EGridDesc_G_M_N e_grid_desc_g_m_n)
+            : batch_stride_A_(batch_stride_A),
+              batch_stride_B_(batch_stride_B),
+              ds_grid_desc_g_m_n_(ds_grid_desc_g_m_n),
+              e_grid_desc_g_m_n_(e_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_A_;
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_B_;
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = static_cast<long_index_t>(g_idx) *
+                               ds_grid_desc_g_m_n_[i].CalculateOffset(make_multi_index(1, 0, 0));
+            });
+
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) *
+                   e_grid_desc_g_m_n_.CalculateOffset(make_multi_index(1, 0, 0));
+        }
+
+        private:
+        index_t batch_stride_A_;
+        index_t batch_stride_B_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+    };
+
+    // A desc for source in blockwise copy
+    template <typename AGridDesc_M_K>
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_K0_M_K1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    template <typename BGridDesc_N_K>
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_N_K1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            b_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(DeviceOp::MakeAGridDescriptor_K0_M_K1(AGridDesc_M_K{}));
+    using BGridDesc_K0_N_K1 = decltype(DeviceOp::MakeBGridDescriptor_K0_N_K1(BGridDesc_N_K{}));
+
+    // GridwiseOp
+    using GridwiseOp = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle<
+        // DataType Family
+        ADataType,
+        BDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        // InMemory Data Descriptor
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        // ElementwiseOp Family
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        // Tiling Family
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerWMMA,
+        NPerWMMA,
+        K1,
+        MRepeat,
+        NRepeat,
+        // ThreadCluster Family
+        BlockSize,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+        NumPrefetch,
+        LoopSched,
+        PipelineVer>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::vector<index_t>& a_gs_ms_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{},
+              b_grid_desc_n_k_{},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{},
+              ds_grid_desc_g_m_n_{
+                  DeviceOp::MakeDsGridDescriptor_G_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides)},
+              e_grid_desc_g_m_n_{
+                  DeviceOp::MakeEGridDescriptor_G_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_mz_stride_{},
+              a_kz_stride_{},
+              b_nz_stride_{},
+              b_kz_stride_{},
+              ds_nz_stride_{},
+              e_nz_stride_{},
+              a_batch_stride_{a_gs_ms_ks_strides[NumDimG - 1]},
+              b_batch_stride_{b_gs_ns_ks_strides[NumDimG - 1]},
+              compute_ptr_offset_of_batch_{
+                  a_batch_stride_, b_batch_stride_, ds_grid_desc_g_m_n_, e_grid_desc_g_m_n_}
+        {
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+            });
+
+            a_grid_desc_m_k_ =
+                DeviceOp::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+            b_grid_desc_n_k_ =
+                DeviceOp::MakeBGridDescriptor_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+
+            ds_grid_desc_m_n_ =
+                DeviceOp::MakeDsGridDescriptor_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides);
+
+            e_grid_desc_m_n_ =
+                DeviceOp::MakeEGridDescriptor_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            a_grid_desc_k0_m_k1_ = DeviceOp::MakeAGridDescriptor_K0_M_K1(a_grid_desc_m_k_);
+            b_grid_desc_k0_n_k1_ = DeviceOp::MakeBGridDescriptor_K0_N_K1(b_grid_desc_n_k_);
+
+            block_2_ctile_map_ = GridwiseOp::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_, M01, N01);
+
+            ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                GridwiseOp::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n_);
+
+            e_grid_desc_mblock_mperblock_nblock_nperblock =
+                GridwiseOp::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n_);
+
+            // for sanity check of vector memory access
+            a_mz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM - 1];
+            a_kz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1];
+            b_nz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN - 1];
+            b_kz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1];
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = ds_gs_ms_ns_strides[i][NumDimG + NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_gs_ms_ns_strides[NumDimG + NumDimM + NumDimN - 1];
+        }
+
+        // Pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseOp::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // Tensor Descriptors
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+
+        typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock;
+        typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        // Block to Tile mapping
+        typename GridwiseOp::DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // Idle
+        index_t M01_;
+        index_t N01_;
+
+        // ElementwiseOp
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        index_t e_mz_stride_;
+        index_t e_nz_stride_;
+
+        index_t a_batch_stride_;
+        index_t b_batch_stride_;
+
+        // Batch Offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const index_t G = arg.e_grid_desc_g_m_n_.GetLength(I0);
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * G;
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_wmma_cshuffle<
+                    GridwiseOp,
+                    ADataType,
+                    BDataType,
+                    typename GridwiseOp::DsGridPointer,
+                    EDataType,
+                    DeviceOp::AGridDesc_K0_M_K1,
+                    DeviceOp::BGridDesc_K0_N_K1,
+                    typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    ComputePtrOffsetOfStridedBatch,
+                    typename GridwiseOp::DefaultBlock2CTileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              G,
+                                              arg.a_grid_desc_k0_m_k1_,
+                                              arg.b_grid_desc_k0_n_k1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_ctile_map_);
+            };
+
+            if(GridwiseOp::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::get_device_name() == "gfx1100")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        if(!GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                      arg.b_grid_desc_k0_n_k1_,
+                                      arg.ds_grid_desc_m_n_,
+                                      arg.e_grid_desc_m_n_,
+                                      arg.block_2_ctile_map_))
+        {
+            return false;
+        }
+
+        // check vector access
+        static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
+                          (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
+                      "wrong!");
+
+        // vector memory access of A: could be on M or AK1 dimension
+        if constexpr(ABlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.a_mz_stride_ == 1 &&
+                 arg.a_grid_desc_k0_m_k1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.a_kz_stride_ == 1 &&
+                 arg.a_grid_desc_k0_m_k1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of B: could be on N or BK1 dimension
+        if constexpr(BBlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.b_nz_stride_ == 1 &&
+                 arg.b_grid_desc_k0_n_k1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.b_kz_stride_ == 1 &&
+                 arg.b_grid_desc_k0_n_k1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of Ds: always on NPerBlock dimension
+        bool valid_d_access = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            if(!(arg.ds_nz_stride_[i] == 1 &&
+                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetLength(I3) %
+                         CDEShuffleBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                valid_d_access = false;
+            }
+        });
+
+        if(valid_d_access == false)
+        {
+            return false;
+        }
+
+        // vector memory access of E: always on NPerBlock dimension
+        if(!((arg.e_nz_stride_ == 1 &&
+              arg.e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I3) %
+                      CDEShuffleBlockTransferScalarPerVector_NPerBlock ==
+                  0) ||
+             CDEShuffleBlockTransferScalarPerVector_NPerBlock == 1))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const void* p_a,
+                 const void* p_b,
+                 std::array<const void*, NumDTensor> p_ds,
+                 void* p_e,
+                 const std::vector<index_t>& a_gs_ms_ks_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_gs_ms_ks_lengths,
+                        b_gs_ns_ks_lengths,
+                        ds_gs_ms_ns_lengths,
+                        e_gs_ms_ns_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_strides,
+                        ds_gs_ms_ns_strides,
+                        e_gs_ms_ns_strides,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ks_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_gs_ms_ks_lengths,
+                                          b_gs_ns_ks_lengths,
+                                          ds_gs_ms_ns_lengths,
+                                          e_gs_ms_ns_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_strides,
+                                          ds_gs_ms_ns_strides,
+                                          e_gs_ms_ns_strides,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceBatchedContractionMultipleD_Wmma_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerWMMA << ", "
+            << NPerWMMA << ", "
+            << MRepeat << ", "
+            << NRepeat
+            << ">"
+            << " NumPrefetch: "
+            << NumPrefetch << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck