gemm_xdl_fp16_splitk.cpp

#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_xdl_splitk_c_shuffle_small_gemm.hpp"
#include "device_gemm_xdl_splitk_c_shuffle_static.hpp"
#include "device_gemm_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"

template <ck::index_t... Is>
using S = ck::Sequence<Is...>;

using F16 = ck::half_t;
using F32 = float;

using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;

using PassThrough = ck::tensor_operation::element_wise::PassThrough;

using ADataType   = ck::half_t;
using BDataType   = ck::half_t;
using CDataType   = ck::half_t;
using AccDataType = float;

using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::RowMajor;
//using BLayout = Col;
using CLayout = ck::tensor_layout::gemm::RowMajor;

using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;

static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;


#if USEING_STATIC_KERNEL
#if MNKB_1_4
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffleStatic
//######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           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|    Type| DataType| 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|
//######|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
            <Row,      Row,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,    16,   128,    32,   8,   2,   16,   16,    1,    2,  S<1, 4, 16, 4>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              2,              2,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         8,           1,           2,              S<1, 4, 1, 64>,               2>;

#else
using DeviceGemmInstance_0 = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffleStatic
//######|     M,    N,    K, K_batch| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           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|
//######|     M,    N,    K, K_batch|        |        |        |  Type|  Type|  Type|    Type| DataType| 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|
//######|     M,    N,    K, K_batch|        |        |        |      |      |      |        |         |   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|
//######|     M,    N,    K, K_batch|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
            <16, 1152, 5120,       8, Row,      Row,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        3,   256,    16,   128,    32,   8,   2,   16,   16,    1,    2,  S<1, 4, 16, 4>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              2,              2,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         8,           1,           2,              S<1, 4, 1, 64>,               2>;
            //<Row,      Row,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        2,   256,    16,   256,    32,   8,   2,   16,   16,    1,    4,  S<1, 4, 16, 4>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              2,              2,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         8,           1,           2,              S<1, 4, 1, 64>,               2>;
            //<Row,      Col,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,    16,   128,    128,   8,   8,   16,   16,    1,    2,  S<1, 16, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           2,              S<1, 4, 1, 64>,               2>;
            //<Row,      Row,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        4,   256,    16,   128,    32,   8,   2,   16,   16,    1,    2,  S<1, 4, 16, 4>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              2,              2,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         8,           1,           2,              S<1, 4, 1, 64>,               2>;
using DeviceGemmInstance_1 = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffleStatic
//######|     M,    N,    K, K_batch| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           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|
//######|     M,    N,    K, K_batch|        |        |        |  Type|  Type|  Type|    Type| DataType| 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|
//######|     M,    N,    K, K_batch|        |        |        |      |      |      |        |         |   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|
//######|     M,    N,    K, K_batch|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
            <16, 5120,  384,       4, Row,      Row,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,    16,   128,    32,   8,   2,   16,   16,    1,    2,  S<1, 4, 16, 4>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              2,              2,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         8,           1,           2,              S<1, 4, 1, 64>,               2>;

using DeviceGemmInstance_2 = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffleStatic
//######|     M,    N,    K, K_batch| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           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|
//######|     M,    N,    K, K_batch|        |        |        |  Type|  Type|  Type|    Type| DataType| 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|
//######|     M,    N,    K, K_batch|        |        |        |      |      |      |        |         |   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|
//######|     M,    N,    K, K_batch|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
            <16, 1280, 5120,       8, Row,      Row,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        3,   256,    16,   128,    32,   8,   2,   16,   16,    1,    2,  S<1, 4, 16, 4>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              2,              2,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         8,           1,           2,              S<1, 4, 1, 64>,               2>;

using DeviceGemmInstance_3 = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffleStatic
//######|     M,    N,    K, K_batch| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           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|
//######|     M,    N,    K, K_batch|        |        |        |  Type|  Type|  Type|    Type| DataType| 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|
//######|     M,    N,    K, K_batch|        |        |        |      |      |      |        |         |   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|
//######|     M,    N,    K, K_batch|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
            <16, 5120, 1280,       5, Row,      Row,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        3,   256,    16,   128,    32,   8,   2,   16,   16,    1,    2,  S<1, 4, 16, 4>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              2,              2,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         8,           1,           2,              S<1, 4, 1, 64>,               2>;


#endif
#else
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffleSmallGemm
//######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           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|    Type| DataType| 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|
//######|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
            <Row,      Row,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        3,   256,    16,   128,    32,   8,   2,   16,   16,    1,    2,  S<1, 4, 16, 4>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              2,              2,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         8,           1,           1,              S<1, 16, 1, 16>,               2>;
            //<Row,      Row,     Row,   F16,   F16,   F16,     F32,      F16,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        3,   256,    16,   256,    32,   8,   2,   16,   16,    1,    4,  S<1, 4, 16, 4>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              2,              2,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         8,           1,           2,              S<1, 4, 1, 64>,               2>;
                     
// clang-format on

#endif  

using ReferenceGemmInstance = ck::tensor_operation::host::
    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;

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 = 4096;
    ck::index_t StrideB = 4096;
    ck::index_t StrideC = 4096;

    ck::index_t splitk = 2;

    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]);
        StrideC = std::stoi(argv[9]);

        splitk = 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=n0, 1=yes)\n");
        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
        exit(0);
    }

    auto f_host_tensor_descriptor =
        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
            {
                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
                                            std::vector<std::size_t>({stride, 1}));
            }
            else
            {
                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
                                            std::vector<std::size_t>({1, stride}));
            }
        };

    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));

    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;

    switch(init_method)
    {
    case 0: break;
    case 1:
        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
        break;
    case 2:
        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
        break;
    default:
        a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
        b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
    }

    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());

    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
    b_k_n_device_buf.ToDevice(b_k_n.mData.data());

    auto a_element_op = AElementOp{};
    auto b_element_op = BElementOp{};
    auto c_element_op = CElementOp{};

    std::cout << "a device buf: " << a_m_k_device_buf.GetDeviceBuffer() << std::endl;
    std::cout << "b device buf: " << b_k_n_device_buf.GetDeviceBuffer() << std::endl;
    std::cout << "c device buf: " << c_m_n_device_buf.GetDeviceBuffer() << std::endl;

    // do GEMM
    if(M == 16 && N == 1152 && K == 5120 && splitk == 8)
    {
        auto gemm = DeviceGemmInstance_0{};
        auto invoker  = gemm.MakeInvoker();
        auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
                                          static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
                                          static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
                                          M,
                                          N,
                                          K,
                                          StrideA,
                                          StrideB,
                                          StrideC,
                                          a_element_op,
                                          b_element_op,
                                          c_element_op,
                                          splitk);

        if(!gemm.IsSupportedArgument(argument))
        {
            std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;

            return 0;
        }

        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});

        std::size_t flop = std::size_t(2) * M * N * K;
        std::size_t num_btype =
            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * 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, "
                  << gemm.GetTypeString() << std::endl;

        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());

        if(do_verification)
        {
            auto ref_gemm    = ReferenceGemmInstance{};
            auto ref_invoker = ref_gemm.MakeInvoker();

            auto ref_argument = ref_gemm.MakeArgument(
                a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);

            ref_invoker.Run(ref_argument);

            return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
        }
    }
    else if(M == 16 && N == 5120 && K == 384 && splitk == 4)
    {
        auto gemm = DeviceGemmInstance_1{};
        auto invoker  = gemm.MakeInvoker();
        auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
                                          static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
                                          static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
                                          M,
                                          N,
                                          K,
                                          StrideA,
                                          StrideB,
                                          StrideC,
                                          a_element_op,
                                          b_element_op,
                                          c_element_op,
                                          splitk);

        if(!gemm.IsSupportedArgument(argument))
        {
            std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;

            return 0;
        }

        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});

        std::size_t flop = std::size_t(2) * M * N * K;
        std::size_t num_btype =
            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * 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, "
                  << gemm.GetTypeString() << std::endl;

        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());

        if(do_verification)
        {
            auto ref_gemm    = ReferenceGemmInstance{};
            auto ref_invoker = ref_gemm.MakeInvoker();

            auto ref_argument = ref_gemm.MakeArgument(
                a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);

            ref_invoker.Run(ref_argument);

            return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
        }
    }
    else if(M == 16 && N == 1280 && K == 5120 && splitk == 8)
    {
        auto gemm = DeviceGemmInstance_2{};
        auto invoker  = gemm.MakeInvoker();
        auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
                                          static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
                                          static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
                                          M,
                                          N,
                                          K,
                                          StrideA,
                                          StrideB,
                                          StrideC,
                                          a_element_op,
                                          b_element_op,
                                          c_element_op,
                                          splitk);

        if(!gemm.IsSupportedArgument(argument))
        {
            std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;

            return 0;
        }

        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});

        std::size_t flop = std::size_t(2) * M * N * K;
        std::size_t num_btype =
            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * 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, "
                  << gemm.GetTypeString() << std::endl;

        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());

        if(do_verification)
        {
            auto ref_gemm    = ReferenceGemmInstance{};
            auto ref_invoker = ref_gemm.MakeInvoker();

            auto ref_argument = ref_gemm.MakeArgument(
                a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);

            ref_invoker.Run(ref_argument);

            return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
        }
    }
    else if(M == 16 && N == 5120 && K == 1280 && splitk == 5)
    {
        auto gemm = DeviceGemmInstance_3{};
        auto invoker  = gemm.MakeInvoker();
        auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
                                          static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
                                          static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
                                          M,
                                          N,
                                          K,
                                          StrideA,
                                          StrideB,
                                          StrideC,
                                          a_element_op,
                                          b_element_op,
                                          c_element_op,
                                          splitk);

        if(!gemm.IsSupportedArgument(argument))
        {
            std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;

            return 0;
        }

        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});

        std::size_t flop = std::size_t(2) * M * N * K;
        std::size_t num_btype =
            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * 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, "
                  << gemm.GetTypeString() << std::endl;

        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());

        if(do_verification)
        {
            auto ref_gemm    = ReferenceGemmInstance{};
            auto ref_invoker = ref_gemm.MakeInvoker();

            auto ref_argument = ref_gemm.MakeArgument(
                a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);

            ref_invoker.Run(ref_argument);

            return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
        }
    }

    return 0;
}