add gemm_bias_add example (#1361)

* add gemm_bias_add example

* changed strideD

* clang-format

---------
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>

example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, 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_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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_contraction.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using A0DataType       = F8;
+using A1DataType       = F32;
+using B0DataType       = F8;
+using B1DataType       = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using EDataType        = F16;
+using ComputeDataType  = F8;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct Multiply
+{
+    __host__ __device__ constexpr void
+    operator()(ck::f8_t& a, const ck::f8_t& a0, const float& a1) const
+    {
+        a = ck::type_convert<ck::half_t>(ck::type_convert<float>(a0) * a1);
+    }
+};
+
+using AElementOp   = Multiply;
+using BElementOp   = Multiply;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceContractionMultipleABD_Xdl_CShuffle<
+    NumDimM,
+    NumDimN,
+    NumDimK,
+    ck::Tuple<A0DataType, A1DataType>,
+    ck::Tuple<B0DataType, B1DataType>,
+    AccDataType,
+    CShuffleDataType,
+    ck::Tuple<>,
+    EDataType,
+    AElementOp,
+    BElementOp,
+    CDEElementOp,
+    GemmSpec,
+    1,
+    256,
+    256,
+    128,
+    32,
+    8,
+    8,
+    32,
+    32,
+    4,
+    2,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    1,
+    8,
+    1,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    1,
+    1,
+    1,
+    S<1, 32, 1, 8>,
+    8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A0[M0, M1, K0, K1]
+    std::vector<ck::index_t> a0_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a0_ms_ks_strides{128 * 32 * 64, 32 * 64, 64, 1};
+    // A1[M1, K1] -> A1[M0, M1, K0, K1]
+    std::vector<ck::index_t> a1_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 1, 0};
+    // B0[N0, N1, K0, K1]
+    std::vector<ck::index_t> b0_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b0_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
+    // B1[N0, N1, K0, K1]
+    std::vector<ck::index_t> b1_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b1_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 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{128 * 32 * 64, 32 * 64, 64, 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<A0DataType> a0_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+    Tensor<A1DataType> a1_ms_ks(a1_ms_ks_lengths, a1_ms_ks_strides);
+    Tensor<B0DataType> b0_ns_ks(b0_ns_ks_lengths, b0_ns_ks_strides);
+    Tensor<B1DataType> b1_ns_ks(b1_ns_ks_lengths, b1_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 << "a0_ms_ks: " << a0_ms_ks.mDesc << std::endl;
+    std::cout << "a1_ms_ks: " << a1_ms_ks.mDesc << std::endl;
+
+    std::cout << "b0_ns_ks: " << b0_ns_ks.mDesc << std::endl;
+    std::cout << "b1_ns_ks: " << b1_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:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
+        b0_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_ns_ks.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        break;
+    default:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
+        b0_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        b1_ns_ks.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem a1_device_buf(sizeof(A1DataType) * a1_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_ms_ks.mData.data());
+    a1_device_buf.ToDevice(a1_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_ns_ks.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument  = device_op.MakeArgument(
+        std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
+                                   a1_device_buf.GetDeviceBuffer()},
+        std::array<const void*, 2>{b0_device_buf.GetDeviceBuffer(),
+                                   b1_device_buf.GetDeviceBuffer()},
+        std::array<const void*, 0>{},
+        e_device_buf.GetDeviceBuffer(),
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_lengths, a1_ms_ks_lengths},
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_strides, a1_ms_ks_strides},
+        std::array<std::vector<ck::index_t>, 2>{b0_ns_ks_lengths, b1_ns_ks_lengths},
+        std::array<std::vector<ck::index_t>, 2>{b0_ns_ks_strides, b1_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,
+        PassThrough{});
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_contraction with the specified compilation parameters does "
+            "not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    if(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>(
+            a0_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(A0DataType) * M * K + sizeof(B0DataType) * 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" << std::endl;
+    }
+
+    if(do_verification)
+    {
+
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        Tensor<A0DataType> a_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+
+        for(size_t m0 = 0; m0 < a_ms_ks.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < a_ms_ks.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t k0 = 0; k0 < a_ms_ks.mDesc.GetLengths()[2]; ++k0)
+                {
+                    for(size_t k1 = 0; k1 < a_ms_ks.mDesc.GetLengths()[3]; ++k1)
+                    {
+                        a_element_op(a_ms_ks(m0, m1, k0, k1),
+                                     a0_ms_ks(m0, m1, k0, k1),
+                                     a1_ms_ks(m0, m1, k0, k1));
+                    }
+                }
+            }
+        }
+
+        Tensor<B0DataType> b_ns_ks(b0_ns_ks_lengths, b0_ns_ks_strides);
+
+        for(size_t n0 = 0; n0 < b_ns_ks.mDesc.GetLengths()[0]; ++n0)
+        {
+            for(size_t n1 = 0; n1 < b_ns_ks.mDesc.GetLengths()[1]; ++n1)
+            {
+                for(size_t k0 = 0; k0 < b_ns_ks.mDesc.GetLengths()[2]; ++k0)
+                {
+                    for(size_t k1 = 0; k1 < b_ns_ks.mDesc.GetLengths()[3]; ++k1)
+                    {
+                        b_element_op(b_ns_ks(n0, n1, k0, k1),
+                                     b0_ns_ks(n0, n1, k0, k1),
+                                     b1_ns_ks(n0, n1, k0, k1));
+                    }
+                }
+            }
+        }
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      A0DataType,
+                                                                      B0DataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      ComputeDataType,
+                                                                      PassThrough,
+                                                                      PassThrough>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
+        auto ref_argument = ref_op.MakeArgument(
+            a_ms_ks, b_ns_ks, c_ms_ns_host_result, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/65_gemm_multiply_multiply/CMakeLists.txt

-add_example_executable(example_gemm_multiply_multiply_xdl_fp16 gemm_multiply_multiply_xdl_fp16.cpp)
+add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp)
+add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp)

example/65_gemm_multiply_multiply/gemm_add_add_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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_gemm_multiple_d_xdl_cshuffle_v3.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.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_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using FP8 = ck::f8_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = F16;
+using B0DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = F16;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+struct AddAdd
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, float, float>(
+        ck::half_t& e, const float& c, const float& d0, const float& d1) const
+    {
+        const float x0_f = c + d0 + d1;
+
+        e = ck::type_convert<ck::half_t>(x0_f);
+    }
+};
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAdd;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3
+    // clang-format off
+///######|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     DsData|     EData|     AccData|         CShuffle|           A|           B|          CDE|           GEMM| 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|      Type|        Type|         DataType| Elementwise| Elementwise|  Elementwise| Spacialization|  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|               |      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+///######|         |         |         |        |           |           |           |          |            |                 |            |            |             |               |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |    S<C, D0, D1>|
+///###### RCR
+         <      Row,      Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   256,   128,   128,  16,  16,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,               8,          0,     S<4, 64, 1>,    S<1, 0, 2>,     S<1, 0, 2>,             2,               8,             8,          0,          1,           1,               S<1, 32, 1, 8>,      S<8, 8, 8>,  ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideD = K;
+    ck::index_t StrideE = N;
+
+    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]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(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 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{0, 2});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{0, 2});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{0, 2});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(),
+                               b0_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer(),
+                                                                   d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, NumDTensor>{StrideD, StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 20, 50});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * 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"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B0DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}