Merge branch 'develop' into improve_pipeline_v3

example/01_gemm/CMakeLists.txt

@@ -1,3 +1,8 @@
+add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)
+add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
+add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
 add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
 add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
 add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
+# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
+add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)

example/01_gemm/gemm_dl_fp16.cpp

@@ -0,0 +1,209 @@
+#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_dl.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 = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+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;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>;
+// clang-format on
+
+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;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    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 == 10)
+    {
+        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]);
+    }
+    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(1);
+    }
+
+    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>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        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{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    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);
+
+    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());
+
+    bool pass = true;
+
+    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);
+
+        pass = ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+    }
+
+    return pass ? 0 : 1;
+}

example/01_gemm/gemm_dl_fp32.cpp

@@ -0,0 +1,208 @@
+#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_dl.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 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   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+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;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>,                5,                  4>;
+// clang-format on
+
+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;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    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 == 10)
+    {
+        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]);
+    }
+    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(1);
+    }
+
+    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>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        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{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    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);
+
+    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());
+
+    bool pass = true;
+
+    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);
+
+        pass = ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+    }
+
+    return pass ? 0 : 1;
+}

example/01_gemm/gemm_dl_int8.cpp

@@ -0,0 +1,206 @@
+#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_dl.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 Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = int8_t;
+using BDataType   = int8_t;
+using CDataType   = int8_t;
+using AccDataType = int32_t;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+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;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+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;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    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 == 10)
+    {
+        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]);
+    }
+    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(1);
+    }
+
+    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>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        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{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    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);
+
+    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());
+
+    bool pass = true;
+
+    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);
+
+        pass = ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+    }
+
+    return pass ? 0 : 1;
+}

example/01_gemm/gemm_xdl_bf16.cpp

@@ -84,13 +84,13 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
 // clang-format on
 
 using ReferenceGemmInstance = ck::tensor_operation::host::
-    ReferenceGemm<float, float, float, PassThrough, PassThrough, PassThrough>;
+    ReferenceGemm<float, float, float, float, PassThrough, PassThrough, PassThrough>;
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     // GEMM shape
     ck::index_t M = 3840;
@@ -105,13 +105,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 10)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -125,7 +125,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\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);
     }
@@ -193,12 +193,12 @@ int main(int argc, char* argv[])
 
     if(!gemm.IsSupportedArgument(argument))
     {
-        throw std::runtime_error(
-            "wrong! device_gemm with the specified compilation parameters does "
-            "not support this GEMM problem");
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    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 =
@@ -232,7 +232,7 @@ int main(int argc, char* argv[])
 
         ref_invoker.Run(ref_argument);
 
-        ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData);
+        return ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData) ? 0 : 1;
     }
 
     return 0;

example/01_gemm/gemm_xdl_fp16.cpp

@@ -4,7 +4,6 @@
 #include <cstdlib>
 #include <stdlib.h>
 #include <half.hpp>
-
 #include "check_err.hpp"
 #include "config.hpp"
 #include "device.hpp"
@@ -29,29 +28,30 @@ 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 ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
 
-using ALayout = ck::tensor_layout::gemm::RowMajor;
-using BLayout = ck::tensor_layout::gemm::ColumnMajor;
-using CLayout = ck::tensor_layout::gemm::RowMajor;
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
 
-using AElementOp = ck::tensor_operation::element_wise::PassThrough;
-using BElementOp = ck::tensor_operation::element_wise::PassThrough;
-using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
 
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 
 // clang-format off
 #if 1
 using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
-//######| 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,     Col,     Row,   F16,   F16,   F16,     F32,      F32,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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>;
+//######| 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|
+//######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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>;
 #elif 0
 using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_ProducerConsumer_CShuffle
 //######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| ABBlockTransfer|       BlockGemm|  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|
@@ -70,13 +70,13 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
 // clang-format on
 
 using ReferenceGemmInstance = ck::tensor_operation::host::
-    ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     // GEMM shape
     ck::index_t M = 3840;
@@ -87,17 +87,21 @@ int main(int argc, char* argv[])
     ck::index_t StrideB = 4096;
     ck::index_t StrideC = 4096;
 
-    if(argc == 4)
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 10)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -111,7 +115,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
         printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
         exit(0);
     }
@@ -184,12 +188,12 @@ int main(int argc, char* argv[])
 
     if(!gemm.IsSupportedArgument(argument))
     {
-        throw std::runtime_error(
-            "wrong! device_gemm with the specified compilation parameters does "
-            "not support this GEMM problem");
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    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 =
@@ -214,7 +218,7 @@ int main(int argc, char* argv[])
 
         ref_invoker.Run(ref_argument);
 
-        ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+        return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
     }
 
     return 0;

example/01_gemm/gemm_xdl_fp64.cpp

@@ -0,0 +1,238 @@
+#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_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 F64 = double;
+
+using ADataType   = double;
+using BDataType   = double;
+using CDataType   = double;
+using AccDataType = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+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;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
+//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+//##########|  Type|  Type|  Type|    Type|        |        |        | 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| SrcDstVectorDim|       DstScalar|
+//##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+//##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+#if 0
+             <  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   64,    32,    32,     4,  1,   16,   16,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,               7,               1>;
+#else
+             <  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,  256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>;
+#endif
+    // clang-format on
+
+    using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                            BDataType,
+                                                                            CDataType,
+                                                                            AccDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CElementOp>;
+
+template <typename DataType>
+std::ostream& show_2d_matrix(std::ostream& os, Tensor<DataType>& matrix)
+{
+    os << "[" << std::endl;
+    for(int x = 0; x < matrix.mDesc.GetLengths()[0]; x++)
+    {
+        os << "[";
+        for(int y = 0; y < matrix.mDesc.GetLengths()[1]; y++)
+        {
+            os << std::setw(4) << static_cast<float>(matrix(x, y));
+        }
+        os << "]" << std::endl;
+    }
+    os << "]";
+    return os;
+}
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = 0;
+    int init_method      = 0;
+    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;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        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]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: run kernel # of times (>1)\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 << "data type: " << typeid(ADataType{}).name() << std::endl;
+    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>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        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_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{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{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    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);
+
+    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);
+
+#if 0
+        {
+            show_2d_matrix(std::cout << "a : ", a_m_k) << std::endl;
+            show_2d_matrix(std::cout << "b: ", b_k_n) << std::endl;
+            show_2d_matrix(std::cout << "c_device: ", c_m_n_device_result) << std::endl;
+            show_2d_matrix(std::cout << "c_host  :", c_m_n_host_result) << std::endl;
+        }
+#endif
+        return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
+    }
+
+    return 0;
+}

example/01_gemm/gemm_xdl_int8.cpp

@@ -78,14 +78,19 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
      16>;                        // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
 // clang-format on
 
-using ReferenceGemmInstance = ck::tensor_operation::host::
-    ReferenceGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        PassThrough>;
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     // GEMM shape
     ck::index_t M = 3840;
@@ -100,13 +105,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 10)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -120,7 +125,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\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);
     }
@@ -189,12 +194,12 @@ int main(int argc, char* argv[])
 
     if(!gemm.IsSupportedArgument(argument))
     {
-        throw std::runtime_error(
-            "wrong! device_gemm with the specified compilation parameters does "
-            "not support this GEMM problem");
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    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 =
@@ -219,7 +224,7 @@ int main(int argc, char* argv[])
 
         ref_invoker.Run(ref_argument);
 
-        ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+        return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
     }
 
     return 0;

example/02_gemm_alpha_beta/gemm_xdl_alpha_beta.cpp

@@ -86,9 +86,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBias2D<AD
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     // GEMM shape
     ck::index_t M = 3840;
@@ -106,13 +106,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 6)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         alpha = std::stof(argv[4]);
         beta  = std::stof(argv[5]);
@@ -121,7 +121,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -138,7 +138,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, alpha, beta\n");
         exit(0);
     }
@@ -216,7 +216,7 @@ int main(int argc, char* argv[])
             "not support this GEMM problem");
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    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 =
@@ -246,6 +246,8 @@ int main(int argc, char* argv[])
 
         ref_invoker.Run(ref_argument);
 
-        ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+        return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
     }
+
+    return 0;
 }

example/03_gemm_bias_relu/gemm_xdl_bias_relu.cpp

@@ -3,89 +3,109 @@
 #include <initializer_list>
 #include <cstdlib>
 #include <stdlib.h>
-#include <half.hpp>
 
 #include "check_err.hpp"
 #include "config.hpp"
-#include "print.hpp"
 #include "device.hpp"
 #include "host_tensor.hpp"
 #include "host_tensor_generator.hpp"
-#include "host_gemm.hpp"
 #include "device_tensor.hpp"
 #include "element_wise_operation.hpp"
-#include "device_gemm_xdl_c_shuffle_bias_activation.hpp"
-#include "reference_gemm_bias_activation.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
 
-using ADataType   = ck::half_t;
-using BDataType   = ck::half_t;
-using CDataType   = ck::half_t;
-using AccDataType = float;
+using F16 = ck::half_t;
+using F32 = float;
 
-using ALayout = ck::tensor_layout::gemm::RowMajor;
-using BLayout = ck::tensor_layout::gemm::ColumnMajor;
-using CLayout = ck::tensor_layout::gemm::RowMajor;
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
 
-using AElementOp = ck::tensor_operation::element_wise::PassThrough;
-using BElementOp = ck::tensor_operation::element_wise::PassThrough;
-using CElementOp = ck::tensor_operation::element_wise::AddRelu;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
-// clang-format off
-using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_Activation<
-    ADataType,              // ADataType
-    BDataType,              // BDataType
-    CDataType,              // CDataType
-    AccDataType,            // AccDataType
-    ALayout,                // ALayout
-    BLayout,                // BLayout
-    CLayout,                // CLayout
-    AElementOp,             // AElementwiseOperation
-    BElementOp,             // BElementwiseOperation
-    CElementOp,             // CElementwiseOperation
-    256,                    // BlockSize
-    256,                    // MPerBlock
-    128,                    // NPerBlock
-    4,                      // K0PerBlock
-    8,                      // K1
-    32,                     // MPerXDL
-    32,                     // NPerXDL
-    4,                      // MXdlPerWave
-    2,                      // NXdlPerWave
-    S<4, 64, 1>,            // ABlockTransferThreadClusterLengths_K0_M_K1
-    S<1, 0, 2>,             // ABlockTransferThreadClusterArrangeOrder
-    S<1, 0, 2>,             // ABlockTransferSrcAccessOrder
-    2,                      // ABlockTransferSrcVectorDim
-    8,                      // ABlockTransferSrcScalarPerVector
-    8,                      // ABlockTransferDstScalarPerVector_K1
-    true,                   // ABlockLdsAddExtraM
-    S<4, 64, 1>,            // BBlockTransferThreadClusterLengths_K0_N_K1
-    S<1, 0, 2>,             // BBlockTransferThreadClusterArrangeOrder
-    S<1, 0, 2>,             // BBlockTransferSrcAccessOrder
-    2,                      // BBlockTransferSrcVectorDim
-    8,                      // BBlockTransferSrcScalarPerVector
-    8,                      // BBlockTransferDstScalarPerVector_K1
-    true,                   // BBlockLdsAddExtraN
-    1,                      // CShuffleMXdlPerWavePerShuffle
-    1,                      // CShuffleNXdlPerWavePerShuffle
-    S<1, 1, 32, 1, 1, 8>,   // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
-    8>;                     // CBlockTransferScalarPerVector_NWaveNPerXdl
-// clang-format on
+// C = A * B
+// E = Relu(C + D);
+struct AddRelu
+{
+    __host__ __device__ void
+    operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d) const
+    {
+        const ck::half_t x = c + d;
 
-using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBiasActivation<ADataType,
-                                                                                      BDataType,
-                                                                                      CDataType,
-                                                                                      AElementOp,
-                                                                                      BElementOp,
-                                                                                      CElementOp>;
+        e = x > 0 ? x : 0;
+    }
+};
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddRelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
+                                                                   BLayout,
+                                                                   ELayout,
+                                                                   ADataType,
+                                                                   BDataType,
+                                                                   AccDataType,
+                                                                   CShuffleDataType,
+                                                                   DsDataType,
+                                                                   EDataType,
+                                                                   AElementOp,
+                                                                   BElementOp,
+                                                                   CDEElementOp,
+                                                                   GemmDefault,
+                                                                   1,
+                                                                   256,
+                                                                   256,
+                                                                   128,
+                                                                   32,
+                                                                   8,
+                                                                   8,
+                                                                   32,
+                                                                   32,
+                                                                   4,
+                                                                   2,
+                                                                   S<4, 64, 1>,
+                                                                   S<1, 0, 2>,
+                                                                   S<1, 0, 2>,
+                                                                   2,
+                                                                   8,
+                                                                   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 = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     // GEMM shape
     ck::index_t M = 3840;
@@ -94,19 +114,23 @@ int main(int argc, char* argv[])
 
     ck::index_t StrideA = 4096;
     ck::index_t StrideB = 4096;
-    ck::index_t StrideC = 4096;
+    ck::index_t StrideE = 4096;
 
-    if(argc == 4)
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 10)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -114,14 +138,14 @@ int main(int argc, char* argv[])
 
         StrideA = std::stoi(argv[7]);
         StrideB = std::stoi(argv[8]);
-        StrideC = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[9]);
     }
     else
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
-        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
         exit(0);
     }
 
@@ -141,17 +165,14 @@ int main(int argc, char* argv[])
 
     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{}));
-
-    // c0_n[n]
-    Tensor<CDataType> c0_n(HostTensorDescriptor(
-        std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, 0, ELayout{}));
+    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 << "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;
-    std::cout << "c0_n: " << c0_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
 
     switch(init_method)
     {
@@ -159,59 +180,59 @@ int main(int argc, char* argv[])
     case 1:
         a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
         b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-        c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
         break;
     default:
         a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
         b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-        c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
     }
 
     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());
-    DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
+    DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpace());
+    DeviceMem e_m_n_device_buf(sizeof(EDataType) * e_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());
-    c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
-    c0_n_device_buf.ToDevice(c0_n.mData.data());
+    d_m_n_device_buf.ToDevice(d_m_n.mData.data());
 
-    auto a_element_op = AElementOp{};
-    auto b_element_op = BElementOp{};
-    auto c_element_op = CElementOp{};
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
 
     // do GEMM
-    auto gemm = DeviceGemmInstance{};
+    auto device_op = DeviceOpInstance{};
 
-    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()),
-                                      static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
-                                      M,
-                                      N,
-                                      K,
-                                      StrideA,
-                                      StrideB,
-                                      StrideC,
-                                      a_element_op,
-                                      b_element_op,
-                                      c_element_op);
+    auto invoker = device_op.MakeInvoker();
 
-    if(!gemm.IsSupportedArgument(argument))
+    auto argument =
+        device_op.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(),
+                               b_k_n_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, 1>{d_m_n_device_buf.GetDeviceBuffer()},
+                               e_m_n_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, 1>{0},
+                               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");
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    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 * M +
-                            sizeof(CDataType) * M * N + sizeof(CDataType) * N;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(EDataType) * M * N + sizeof(EDataType) * N;
 
     float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
 
@@ -220,18 +241,38 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
               << std::endl;
 
-    c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
-
     if(do_verification)
     {
+        e_m_n_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        Tensor<AccDataType> c_m_n(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
         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, c0_n, a_element_op, b_element_op, c_element_op);
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
 
         ref_invoker.Run(ref_argument);
 
-        ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+        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), d_m_n(m, n));
+            }
+        }
+
+        return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
     }
+
+    return 0;
 }

example/04_gemm_add_add_fastgelu/CMakeLists.txt

@@ -0,0 +1 @@
+add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)

example/04_gemm_add_add_fastgelu/README.md

@@ -0,0 +1,23 @@
+# Instructions for ```example_gemm_add_add_fastgelu_xdl_fp16```
+
+## Run ```example_gemm_add_add_fastgelu_xdl_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
+./bin/example_gemm_add_add_fastgelu_xdl_fp16 1 1 1
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
+d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 1.26914 ms, 101.525 TFlops, 100.804 GB/s, DeviceGemmMultipleD_Xdl_CShuffle<256, 256, 128, 32, 8, 8>
+```

example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp

@@ -0,0 +1,245 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+#include "device_gemm_multiple_d_xdl_cshuffle.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 AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F16;
+using D1DataType       = F16;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| ELayout|     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|
+//######|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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>;
+// 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  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    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 == 12)
+    {
+        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]);
+        StrideD0 = std::stoi(argv[9]);
+        StrideD1 = std::stoi(argv[10]);
+        StrideE  = std::stoi(argv[11]);
+    }
+    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 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, "
+               "StrideE\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<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, 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 << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_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:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    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 d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpace());
+    DeviceMem e_m_n_device_buf(sizeof(EDataType) * e_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());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(),
+                               b_k_n_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                                          d1_m_n_device_buf.GetDeviceBuffer()},
+                               e_m_n_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, 2>{StrideD0, StrideD1},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_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});
+
+    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(D0DataType) * N + sizeof(D1DataType) * 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, "
+              << device_op.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n(HostTensorDescriptor(
+            std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        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, a_element_op, b_element_op, 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_m_n_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
+    }
+
+    return 0;
+}

example/04_gemm_bias_relu_add/CMakeLists.txt

@@ -1 +0,0 @@
-add_example_executable(example_gemm_xdl_bias_relu_add gemm_xdl_bias_relu_add.cpp)

example/04_gemm_bias_relu_add/README.md

@@ -1,28 +0,0 @@
-# Instructions for ```example_gemm_xdl_bias_relu_add```
-
-## Run ```example_gemm_xdl_bias_relu_add```
-```bash
-#arg1: verification (0=no, 1=yes)
-#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
-#arg3: run kernel # of times (>1)
-#arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC
-./bin/example_gemm_xdl_bias_relu_add 0 1 5 3840 4096 4096 4096 4096 4096
-```
-
-Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
-```
-a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
-b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
-c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
-c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
-c1_m_n: dim 2, lengths {3840, 4096}, strides {1, 0}
-arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
-arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
-arg.c_grid_desc_m_n_{ 3840, 4096}
-arg.c0_grid_desc_m_n_{ 3840, 4096}
-arg.c1_grid_desc_m_n_{ 3840, 4096}
-launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
-Warm up
-Start running 5 times...
-Perf: 1.27583 ms, 100.992 TFlops, 73.9688 GB/s
-```

example/04_gemm_bias_relu_add/gemm_xdl_bias_relu_add.cpp

@@ -1,255 +0,0 @@
-#include <iostream>
-#include <numeric>
-#include <initializer_list>
-#include <cstdlib>
-#include <stdlib.h>
-#include <half.hpp>
-
-#include "check_err.hpp"
-#include "config.hpp"
-#include "print.hpp"
-#include "device.hpp"
-#include "host_tensor.hpp"
-#include "host_tensor_generator.hpp"
-#include "host_gemm.hpp"
-#include "device_tensor.hpp"
-#include "element_wise_operation.hpp"
-#include "device_gemm_xdl_c_shuffle_bias_activation_add.hpp"
-#include "reference_gemm_bias_activation_add.hpp"
-
-template <ck::index_t... Is>
-using S = ck::Sequence<Is...>;
-
-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::ColumnMajor;
-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::AddReluAdd;
-
-// clang-format off
-using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_Activation_Add<
-    ADataType,              // ADataType
-    BDataType,              // BDataType
-    CDataType,              // CDataType
-    AccDataType,            // AccDataType
-    ALayout,                // ALayout
-    BLayout,                // BLayout
-    CLayout,                // CLayout
-    AElementOp,             // AElementwiseOperation
-    BElementOp,             // BElementwiseOperation
-    CElementOp,             // CElementwiseOperation
-    256,                    // BlockSize
-    256,                    // MPerBlock
-    128,                    // NPerBlock
-    4,                      // K0PerBlock
-    8,                      // K1
-    32,                     // MPerXDL
-    32,                     // NPerXDL
-    4,                      // MXdlPerWave
-    2,                      // NXdlPerWave
-    S<4, 64, 1>,            // ABlockTransferThreadClusterLengths_K0_M_K1
-    S<1, 0, 2>,             // ABlockTransferThreadClusterArrangeOrder
-    S<1, 0, 2>,             // ABlockTransferSrcAccessOrder
-    2,                      // ABlockTransferSrcVectorDim
-    8,                      // ABlockTransferSrcScalarPerVector
-    8,                      // ABlockTransferDstScalarPerVector_K1
-    true,                   // ABlockLdsAddExtraM
-    S<4, 64, 1>,            // BBlockTransferThreadClusterLengths_K0_N_K1
-    S<1, 0, 2>,             // BBlockTransferThreadClusterArrangeOrder
-    S<1, 0, 2>,             // BBlockTransferSrcAccessOrder
-    2,                      // BBlockTransferSrcVectorDim
-    8,                      // BBlockTransferSrcScalarPerVector
-    8,                      // BBlockTransferDstScalarPerVector_K1
-    true,                   // BBlockLdsAddExtraN
-    1,                      // CShuffleMXdlPerWavePerShuffle
-    1,                      // CShuffleNXdlPerWavePerShuffle
-    S<1, 1, 32, 1, 1, 8>,   // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
-    8>;                     // CBlockTransferScalarPerVector_NWaveNPerXdl
-// clang-format on
-
-using ReferenceGemmInstance =
-    ck::tensor_operation::host::ReferenceGemmBiasActivationAdd<ADataType,
-                                                               BDataType,
-                                                               CDataType,
-                                                               AElementOp,
-                                                               BElementOp,
-                                                               CElementOp>;
-int main(int argc, char* argv[])
-{
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
-
-    // 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 StrideC1 = 4096;
-
-    if(argc == 4)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
-    }
-    else if(argc == 11)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        nrepeat         = 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]);
-        StrideC1 = 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: run kernel # of times (>1)\n");
-        printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, StrideC1\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{}));
-
-    // c0_n[n]
-    Tensor<CDataType> c0_n(HostTensorDescriptor(
-        std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
-
-    // c1_m_n[m ,n]
-    Tensor<CDataType> c1_m_n(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;
-    std::cout << "c0_n: " << c0_n.mDesc << std::endl;
-    std::cout << "c1_m_n: " << c1_m_n.mDesc << std::endl;
-
-    switch(init_method)
-    {
-    case 0: break;
-    case 1:
-        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-        c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
-        c1_m_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
-        break;
-    default:
-        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-        c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
-        c1_m_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
-    }
-
-    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());
-    DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
-    DeviceMem c1_m_n_device_buf(sizeof(CDataType) * c1_m_n.mDesc.GetElementSpace());
-
-    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
-    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
-    c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
-    c0_n_device_buf.ToDevice(c0_n.mData.data());
-    c1_m_n_device_buf.ToDevice(c1_m_n.mData.data());
-
-    auto a_element_op = AElementOp{};
-    auto b_element_op = BElementOp{};
-    auto c_element_op = CElementOp{};
-
-    // do GEMM
-    auto gemm = DeviceGemmInstance{};
-
-    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()),
-                                      static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
-                                      static_cast<CDataType*>(c1_m_n_device_buf.GetDeviceBuffer()),
-                                      M,
-                                      N,
-                                      K,
-                                      StrideA,
-                                      StrideB,
-                                      StrideC,
-                                      StrideC1,
-                                      a_element_op,
-                                      b_element_op,
-                                      c_element_op);
-
-    if(!gemm.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, nrepeat);
-
-    std::size_t flop      = std::size_t(2) * M * N * K;
-    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
-                            sizeof(CDataType) * M * N + sizeof(CDataType) * 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"
-              << 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,
-                                                  c0_n,
-                                                  c1_m_n,
-                                                  a_element_op,
-                                                  b_element_op,
-                                                  c_element_op);
-
-        ref_invoker.Run(ref_argument);
-
-        ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
-    }
-}

example/06_conv2d_fwd_bias_relu/CMakeLists.txt

@@ -1,2 +1,2 @@
 add_example_executable(example_conv2d_fwd_xdl_bias_relu conv2d_fwd_xdl_bias_relu.cpp)
-target_link_libraries(example_conv2d_fwd_xdl_bias_relu PRIVATE conv_fwd_util)
+target_link_libraries(example_conv2d_fwd_xdl_bias_relu PRIVATE conv_util)

example/06_conv2d_fwd_bias_relu/conv2d_fwd_xdl_bias_relu.cpp

@@ -7,7 +7,7 @@
 
 #include "check_err.hpp"
 #include "config.hpp"
-#include "conv_fwd_util.hpp"
+#include "conv_util.hpp"
 #include "device.hpp"
 #include "device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp"
 #include "device_tensor.hpp"
@@ -93,7 +93,7 @@ void PrintUseMsg()
 {
     std::cout << "arg1: verification (0=no, 1=yes)\n"
               << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
-              << "arg3: run kernel # of times (>1)\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
               << "Following arguments:\n"
               << " N, K, C, \n"
               << " <filter spatial dimensions>, (ie Y, X for 2D)\n"
@@ -120,40 +120,40 @@ ck::utils::conv::ConvParams ParseConvParams(int argc, char* argv[])
     ck::utils::conv::ConvParams params;
     int arg_idx = 4;
 
-    params.num_dim_spatial = num_dim_spatial;
-    params.N               = std::stoi(argv[arg_idx++]);
-    params.K               = std::stoi(argv[arg_idx++]);
-    params.C               = std::stoi(argv[arg_idx++]);
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
 
-    params.filter_spatial_lengths.resize(num_dim_spatial);
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_spatial_lengths.resize(num_dim_spatial);
+    params.input_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_strides.resize(num_dim_spatial);
+    params.conv_filter_strides_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_dilations.resize(num_dim_spatial);
+    params.conv_filter_dilations_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_left_pads.resize(num_dim_spatial);
+    params.input_left_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_right_pads.resize(num_dim_spatial);
+    params.input_right_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
     }
 
     return params;
@@ -165,9 +165,9 @@ int main(int argc, char* argv[])
 {
     using namespace ck::utils::conv;
 
-    bool do_verification      = 0;
-    int init_method           = 0;
-    int nrepeat               = 5;
+    bool do_verification      = true;
+    int init_method           = 1;
+    bool time_kernel          = false;
     const int num_dim_spatial = 2;
 
     ck::utils::conv::ConvParams params;
@@ -176,7 +176,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
 
     if(argc >= 5)
@@ -184,21 +184,21 @@ int main(int argc, char* argv[])
         params = ParseConvParams(argc, argv);
     }
 
-    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
-                                        static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
     input_dims.insert(std::end(input_dims),
-                      std::begin(params.input_spatial_lengths),
-                      std::end(params.input_spatial_lengths));
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
 
-    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
-                                         static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
     filter_dims.insert(std::end(filter_dims),
-                       std::begin(params.filter_spatial_lengths),
-                       std::end(params.filter_spatial_lengths));
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
 
     const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
-    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
-                                         static_cast<std::size_t>(params.K)};
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
     output_dims.insert(std::end(output_dims),
                        std::begin(output_spatial_lengths),
                        std::end(output_spatial_lengths));
@@ -211,7 +211,7 @@ int main(int argc, char* argv[])
         get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
     // bias: assume contiguous 1d vector
     Tensor<OutDataType> bias(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K)})));
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K_)})));
 
     std::cout << "input: " << input.mDesc << std::endl;
     std::cout << "weights: " << weights.mDesc << std::endl;
@@ -248,16 +248,16 @@ int main(int argc, char* argv[])
                           static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                           static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
                           static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
-                          params.N,
-                          params.K,
-                          params.C,
-                          params.input_spatial_lengths,
-                          params.filter_spatial_lengths,
+                          params.N_,
+                          params.K_,
+                          params.C_,
+                          params.input_spatial_lengths_,
+                          params.filter_spatial_lengths_,
                           output_spatial_lengths,
-                          params.conv_filter_strides,
-                          params.conv_filter_dilations,
-                          params.input_left_pads,
-                          params.input_right_pads,
+                          params.conv_filter_strides_,
+                          params.conv_filter_dilations_,
+                          params.input_left_pads_,
+                          params.input_right_pads_,
                           InElementOp{},
                           WeiElementOp{},
                           OutElementOp{});
@@ -269,18 +269,18 @@ int main(int argc, char* argv[])
             "not support this problem");
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = get_flops(
-        params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
     std::size_t num_btype =
-        get_btype<InDataType, WeiDataType, OutDataType>(params.N,
-                                                        params.C,
-                                                        params.K,
-                                                        params.input_spatial_lengths,
-                                                        params.filter_spatial_lengths,
+        get_btype<InDataType, WeiDataType, OutDataType>(params.N_,
+                                                        params.C_,
+                                                        params.K_,
+                                                        params.input_spatial_lengths_,
+                                                        params.filter_spatial_lengths_,
                                                         output_spatial_lengths) +
-        sizeof(OutDataType) * (params.K);
+        sizeof(OutDataType) * (params.K_);
 
     float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
     float gb_per_sec = num_btype / 1.E6 / ave_time;
@@ -296,16 +296,17 @@ int main(int argc, char* argv[])
                                                   weights,
                                                   host_output,
                                                   bias,
-                                                  params.conv_filter_strides,
-                                                  params.conv_filter_dilations,
-                                                  params.input_left_pads,
-                                                  params.input_right_pads,
+                                                  params.conv_filter_strides_,
+                                                  params.conv_filter_dilations_,
+                                                  params.input_left_pads_,
+                                                  params.input_right_pads_,
                                                   InElementOp{},
                                                   WeiElementOp{},
                                                   OutElementOp{});
         ref_invoker.Run(ref_argument);
         out_device_buf.FromDevice(device_output.mData.data());
-        ck::utils::check_err(
-            host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+        return ck::utils::check_err(device_output.mData, host_output.mData) ? 0 : 1;
     }
+
+    return 0;
 }

example/07_conv2d_fwd_bias_relu_add/CMakeLists.txt

@@ -1,2 +1,3 @@
-add_example_executable(example_conv2d_fwd_xdl_bias_relu_add conv2d_fwd_xdl_bias_relu_add.cpp)
-target_link_libraries(example_conv2d_fwd_xdl_bias_relu_add PRIVATE conv_fwd_util)
+# FIXME: should fix validation failure
+add_example_executable_no_testing(example_conv2d_fwd_xdl_bias_relu_add conv2d_fwd_xdl_bias_relu_add.cpp)
+target_link_libraries(example_conv2d_fwd_xdl_bias_relu_add PRIVATE conv_util)

example/07_conv2d_fwd_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp

@@ -7,7 +7,7 @@
 
 #include "check_err.hpp"
 #include "config.hpp"
-#include "conv_fwd_util.hpp"
+#include "conv_util.hpp"
 #include "device.hpp"
 #include "device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp"
 #include "device_tensor.hpp"
@@ -90,7 +90,7 @@ void PrintUseMsg()
 {
     std::cout << "arg1: verification (0=no, 1=yes)\n"
               << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
-              << "arg3: run kernel # of times (>1)\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
               << "Following arguments:\n"
               << " N, K, C, \n"
               << " <filter spatial dimensions>, (ie Y, X for 2D)\n"
@@ -117,40 +117,40 @@ ck::utils::conv::ConvParams ParseConvParams(int argc, char* argv[])
     ck::utils::conv::ConvParams params;
     int arg_idx = 4;
 
-    params.num_dim_spatial = num_dim_spatial;
-    params.N               = std::stoi(argv[arg_idx++]);
-    params.K               = std::stoi(argv[arg_idx++]);
-    params.C               = std::stoi(argv[arg_idx++]);
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
 
-    params.filter_spatial_lengths.resize(num_dim_spatial);
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_spatial_lengths.resize(num_dim_spatial);
+    params.input_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_strides.resize(num_dim_spatial);
+    params.conv_filter_strides_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_dilations.resize(num_dim_spatial);
+    params.conv_filter_dilations_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_left_pads.resize(num_dim_spatial);
+    params.input_left_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_right_pads.resize(num_dim_spatial);
+    params.input_right_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
     }
 
     return params;
@@ -162,9 +162,9 @@ int main(int argc, char* argv[])
 {
     using namespace ck::utils::conv;
 
-    bool do_verification      = 0;
-    int init_method           = 0;
-    int nrepeat               = 5;
+    bool do_verification      = true;
+    int init_method           = 1;
+    bool time_kernel          = false;
     const int num_dim_spatial = 2;
 
     ck::utils::conv::ConvParams params;
@@ -173,7 +173,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
 
     if(argc >= 5)
@@ -181,21 +181,21 @@ int main(int argc, char* argv[])
         params = ParseConvParams(argc, argv);
     }
 
-    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
-                                        static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
     input_dims.insert(std::end(input_dims),
-                      std::begin(params.input_spatial_lengths),
-                      std::end(params.input_spatial_lengths));
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
 
-    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
-                                         static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
     filter_dims.insert(std::end(filter_dims),
-                       std::begin(params.filter_spatial_lengths),
-                       std::end(params.filter_spatial_lengths));
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
 
     const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
-    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
-                                         static_cast<std::size_t>(params.K)};
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
     output_dims.insert(std::end(output_dims),
                        std::begin(output_spatial_lengths),
                        std::end(output_spatial_lengths));
@@ -209,7 +209,7 @@ int main(int argc, char* argv[])
 
     // bias: assume contiguous 1d vector
     Tensor<OutDataType> bias(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K)})));
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K_)})));
 
     // residual: assume same layout as output tensor
     Tensor<OutDataType> residual(get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
@@ -224,10 +224,10 @@ int main(int argc, char* argv[])
     {
     case 0: break;
     case 1:
-        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
-        weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
-        bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
-        residual.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
+        weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
+        bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
+        residual.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
         break;
     default:
         input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
@@ -259,16 +259,16 @@ int main(int argc, char* argv[])
                           static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
                           static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
                           static_cast<const OutDataType*>(resi_device_buf.GetDeviceBuffer()),
-                          params.N,
-                          params.K,
-                          params.C,
-                          params.input_spatial_lengths,
-                          params.filter_spatial_lengths,
+                          params.N_,
+                          params.K_,
+                          params.C_,
+                          params.input_spatial_lengths_,
+                          params.filter_spatial_lengths_,
                           output_spatial_lengths,
-                          params.conv_filter_strides,
-                          params.conv_filter_dilations,
-                          params.input_left_pads,
-                          params.input_right_pads,
+                          params.conv_filter_strides_,
+                          params.conv_filter_dilations_,
+                          params.input_left_pads_,
+                          params.input_right_pads_,
                           in_element_op,
                           wei_element_op,
                           out_element_op);
@@ -280,20 +280,20 @@ int main(int argc, char* argv[])
             "not support this problem");
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = get_flops(
-        params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
     std::size_t num_btype =
-        get_btype<InDataType, WeiDataType, OutDataType>(params.N,
-                                                        params.C,
-                                                        params.K,
-                                                        params.input_spatial_lengths,
-                                                        params.filter_spatial_lengths,
+        get_btype<InDataType, WeiDataType, OutDataType>(params.N_,
+                                                        params.C_,
+                                                        params.K_,
+                                                        params.input_spatial_lengths_,
+                                                        params.filter_spatial_lengths_,
                                                         output_spatial_lengths) +
-        sizeof(OutDataType) * (params.K) +
+        sizeof(OutDataType) * (params.K_) +
         sizeof(OutDataType) *
-            (params.N * params.K * output_spatial_lengths[0] * output_spatial_lengths[1]);
+            (params.N_ * params.K_ * output_spatial_lengths[0] * output_spatial_lengths[1]);
 
     float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
     float gb_per_sec = num_btype / 1.E6 / ave_time;
@@ -310,17 +310,18 @@ int main(int argc, char* argv[])
                                                   host_output,
                                                   bias,
                                                   residual,
-                                                  params.conv_filter_strides,
-                                                  params.conv_filter_dilations,
-                                                  params.input_left_pads,
-                                                  params.input_right_pads,
+                                                  params.conv_filter_strides_,
+                                                  params.conv_filter_dilations_,
+                                                  params.input_left_pads_,
+                                                  params.input_right_pads_,
                                                   in_element_op,
                                                   wei_element_op,
                                                   out_element_op);
 
         ref_invoker.Run(ref_argument);
         out_device_buf.FromDevice(device_output.mData.data());
-        ck::utils::check_err(
-            host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+        return ck::utils::check_err(device_output.mData, host_output.mData) ? 0 : 1;
     }
+
+    return 0;
 }

example/09_convnd_fwd/CMakeLists.txt

@@ -1,6 +1,9 @@
-add_example_executable(example_convnd_fwd_xdl convnd_fwd_xdl.cpp)
-target_link_libraries(example_convnd_fwd_xdl PRIVATE conv_fwd_util)
+add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
 add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
-target_link_libraries(example_convnd_fwd_xdl_int8 PRIVATE conv_fwd_util)
 add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
-target_link_libraries(example_convnd_fwd_xdl_fp16 PRIVATE conv_fwd_util)
+# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
+add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
+target_link_libraries(example_convnd_fwd_xdl_fp64 PRIVATE conv_util)
+target_link_libraries(example_convnd_fwd_xdl_fp32 PRIVATE conv_util)
+target_link_libraries(example_convnd_fwd_xdl_int8 PRIVATE conv_util)
+target_link_libraries(example_convnd_fwd_xdl_fp16 PRIVATE conv_util)

example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp

@@ -5,7 +5,7 @@
 
 #include "check_err.hpp"
 #include "config.hpp"
-#include "conv_fwd_util.hpp"
+#include "conv_util.hpp"
 #include "device.hpp"
 #include "device_tensor.hpp"
 #include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
@@ -43,10 +43,10 @@ template <ck::index_t NumDimSpatial>
 using DeviceConvNDFwdInstance = ck::tensor_operation::device::
     DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
         // clang-format off
-        InDataType,         // 
+        InDataType,         //
         WeiDataType,        //
         OutDataType,        //
-        AccDataType,        // 
+        AccDataType,        //
         InElementOp,        // Input Elementwise Operation
         WeiElementOp,       // Weights Elementwise Operation
         OutElementOp,       // Output Elementwise Operation
@@ -110,7 +110,7 @@ void print_use_msg()
 {
     std::cout << "arg1: verification (0=no, 1=yes)\n"
               << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
-              << "arg3: run kernel # of times (>1)\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
               << "arg4: N spatial dimensions (default 2)\n"
               << "Following arguments (depending on number of spatial dims):\n"
               << " N, K, C, \n"
@@ -137,40 +137,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, cha
     ck::utils::conv::ConvParams params;
     int arg_idx = 5;
 
-    params.num_dim_spatial = num_dim_spatial;
-    params.N               = std::stoi(argv[arg_idx++]);
-    params.K               = std::stoi(argv[arg_idx++]);
-    params.C               = std::stoi(argv[arg_idx++]);
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
 
-    params.filter_spatial_lengths.resize(num_dim_spatial);
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_spatial_lengths.resize(num_dim_spatial);
+    params.input_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_strides.resize(num_dim_spatial);
+    params.conv_filter_strides_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_dilations.resize(num_dim_spatial);
+    params.conv_filter_dilations_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_left_pads.resize(num_dim_spatial);
+    params.input_left_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_right_pads.resize(num_dim_spatial);
+    params.input_right_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
     }
 
     return params;
@@ -182,9 +182,9 @@ int main(int argc, char* argv[])
 {
     using namespace ck::utils::conv;
 
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
     int num_dim_spatial  = 2;
 
     ck::utils::conv::ConvParams params;
@@ -193,7 +193,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
         num_dim_spatial = std::stoi(argv[4]);
     }
 
@@ -202,21 +202,21 @@ int main(int argc, char* argv[])
         params = parse_conv_params(num_dim_spatial, argc, argv);
     }
 
-    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
-                                        static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
     input_dims.insert(std::end(input_dims),
-                      std::begin(params.input_spatial_lengths),
-                      std::end(params.input_spatial_lengths));
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
 
-    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
-                                         static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
     filter_dims.insert(std::end(filter_dims),
-                       std::begin(params.filter_spatial_lengths),
-                       std::end(params.filter_spatial_lengths));
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
 
     const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
-    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
-                                         static_cast<std::size_t>(params.K)};
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
     output_dims.insert(std::end(output_dims),
                        std::begin(output_spatial_lengths),
                        std::end(output_spatial_lengths));
@@ -256,16 +256,16 @@ int main(int argc, char* argv[])
         conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                                   static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                                   static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-                                  params.N,
-                                  params.K,
-                                  params.C,
-                                  params.input_spatial_lengths,
-                                  params.filter_spatial_lengths,
+                                  params.N_,
+                                  params.K_,
+                                  params.C_,
+                                  params.input_spatial_lengths_,
+                                  params.filter_spatial_lengths_,
                                   output_spatial_lengths,
-                                  params.conv_filter_strides,
-                                  params.conv_filter_dilations,
-                                  params.input_left_pads,
-                                  params.input_right_pads,
+                                  params.conv_filter_strides_,
+                                  params.conv_filter_dilations_,
+                                  params.input_left_pads_,
+                                  params.input_right_pads_,
                                   InElementOp{},
                                   WeiElementOp{},
                                   OutElementOp{});
@@ -277,22 +277,22 @@ int main(int argc, char* argv[])
             "not support this Conv problem");
     }
 
-    float ave_time = invoker->Run(argument.get(), nrepeat);
+    float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = get_flops(
-        params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
     std::size_t num_btype = get_btype<InDataType, WeiDataType, OutDataType>(
-        params.N,
-        params.C,
-        params.K,
-        params.input_spatial_lengths,
-        params.filter_spatial_lengths,
+        params.N_,
+        params.C_,
+        params.K_,
+        params.input_spatial_lengths_,
+        params.filter_spatial_lengths_,
         output_spatial_lengths);
 
     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;
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv->GetTypeString() << std::endl;
 
     if(do_verification)
     {
@@ -302,40 +302,38 @@ int main(int argc, char* argv[])
             auto ref_argument = ref_conv.MakeArgument(input,
                                                       weights,
                                                       host_output,
-                                                      params.conv_filter_strides,
-                                                      params.conv_filter_dilations,
-                                                      params.input_left_pads,
-                                                      params.input_right_pads,
+                                                      params.conv_filter_strides_,
+                                                      params.conv_filter_dilations_,
+                                                      params.input_left_pads_,
+                                                      params.input_right_pads_,
                                                       InElementOp{},
                                                       WeiElementOp{},
                                                       OutElementOp{});
 
             ref_invoker.Run(ref_argument);
             out_device_buf.FromDevice(device_output.mData.data());
-            ck::utils::check_err(
-                host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+            return ck::utils::check_err(
+                host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f) ? 0 : 1;
         };
 
         switch(num_dim_spatial)
         {
         case 3: {
             auto ref_conv = ReferenceConvNDFwdInstance<3>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 2: {
             auto ref_conv = ReferenceConvNDFwdInstance<2>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 1: {
             auto ref_conv = ReferenceConvNDFwdInstance<1>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         default: {
             throw std::runtime_error("Unsupported number of spatial dimensions provided!");
         }
         }
     }
+    return 0;
 }

example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp

@@ -5,7 +5,7 @@
 
 #include "check_err.hpp"
 #include "config.hpp"
-#include "conv_fwd_util.hpp"
+#include "conv_util.hpp"
 #include "device.hpp"
 #include "device_tensor.hpp"
 #include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
@@ -39,10 +39,10 @@ template <ck::index_t NumDimSpatial>
 using DeviceConvNDFwdInstance = ck::tensor_operation::device::
     DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
         // clang-format off
-        InDataType,         // 
+        InDataType,         //
         WeiDataType,        //
         OutDataType,        //
-        AccDataType,        // 
+        AccDataType,        //
         InElementOp,        // Input Elementwise Operation
         WeiElementOp,       // Weights Elementwise Operation
         OutElementOp,       // Output Elementwise Operation
@@ -107,7 +107,7 @@ void print_use_msg()
 {
     std::cout << "arg1: verification (0=no, 1=yes)\n"
               << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
-              << "arg3: run kernel # of times (>1)\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
               << "arg4: N spatial dimensions (default 2)\n"
               << "Following arguments (depending on number of spatial dims):\n"
               << " N, K, C, \n"
@@ -134,40 +134,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, cha
     ck::utils::conv::ConvParams params;
     int arg_idx = 5;
 
-    params.num_dim_spatial = num_dim_spatial;
-    params.N               = std::stoi(argv[arg_idx++]);
-    params.K               = std::stoi(argv[arg_idx++]);
-    params.C               = std::stoi(argv[arg_idx++]);
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
 
-    params.filter_spatial_lengths.resize(num_dim_spatial);
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_spatial_lengths.resize(num_dim_spatial);
+    params.input_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_strides.resize(num_dim_spatial);
+    params.conv_filter_strides_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_dilations.resize(num_dim_spatial);
+    params.conv_filter_dilations_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_left_pads.resize(num_dim_spatial);
+    params.input_left_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_right_pads.resize(num_dim_spatial);
+    params.input_right_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
     }
 
     return params;
@@ -179,9 +179,9 @@ int main(int argc, char* argv[])
 {
     using namespace ck::utils::conv;
 
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
     int num_dim_spatial  = 2;
 
     ck::utils::conv::ConvParams params;
@@ -190,7 +190,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
         num_dim_spatial = std::stoi(argv[4]);
     }
 
@@ -199,21 +199,21 @@ int main(int argc, char* argv[])
         params = parse_conv_params(num_dim_spatial, argc, argv);
     }
 
-    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
-                                        static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
     input_dims.insert(std::end(input_dims),
-                      std::begin(params.input_spatial_lengths),
-                      std::end(params.input_spatial_lengths));
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
 
-    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
-                                         static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
     filter_dims.insert(std::end(filter_dims),
-                       std::begin(params.filter_spatial_lengths),
-                       std::end(params.filter_spatial_lengths));
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
 
     const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
-    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
-                                         static_cast<std::size_t>(params.K)};
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
     output_dims.insert(std::end(output_dims),
                        std::begin(output_spatial_lengths),
                        std::end(output_spatial_lengths));
@@ -255,16 +255,16 @@ int main(int argc, char* argv[])
         conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                                   static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                                   static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-                                  params.N,
-                                  params.K,
-                                  params.C,
-                                  params.input_spatial_lengths,
-                                  params.filter_spatial_lengths,
+                                  params.N_,
+                                  params.K_,
+                                  params.C_,
+                                  params.input_spatial_lengths_,
+                                  params.filter_spatial_lengths_,
                                   output_spatial_lengths,
-                                  params.conv_filter_strides,
-                                  params.conv_filter_dilations,
-                                  params.input_left_pads,
-                                  params.input_right_pads,
+                                  params.conv_filter_strides_,
+                                  params.conv_filter_dilations_,
+                                  params.input_left_pads_,
+                                  params.input_right_pads_,
                                   InElementOp{},
                                   WeiElementOp{},
                                   OutElementOp{});
@@ -276,16 +276,16 @@ int main(int argc, char* argv[])
             "not support this Conv problem");
     }
 
-    float ave_time = invoker->Run(argument.get(), nrepeat);
+    float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = get_flops(
-        params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
     std::size_t num_btype =
-        get_btype<InDataType, WeiDataType, OutDataType>(params.N,
-                                                        params.C,
-                                                        params.K,
-                                                        params.input_spatial_lengths,
-                                                        params.filter_spatial_lengths,
+        get_btype<InDataType, WeiDataType, OutDataType>(params.N_,
+                                                        params.C_,
+                                                        params.K_,
+                                                        params.input_spatial_lengths_,
+                                                        params.filter_spatial_lengths_,
                                                         output_spatial_lengths);
 
     float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -301,40 +301,43 @@ int main(int argc, char* argv[])
             auto ref_argument = ref_conv.MakeArgument(input,
                                                       weights,
                                                       host_output,
-                                                      params.conv_filter_strides,
-                                                      params.conv_filter_dilations,
-                                                      params.input_left_pads,
-                                                      params.input_right_pads,
+                                                      params.conv_filter_strides_,
+                                                      params.conv_filter_dilations_,
+                                                      params.input_left_pads_,
+                                                      params.input_right_pads_,
                                                       InElementOp{},
                                                       WeiElementOp{},
                                                       OutElementOp{});
 
             ref_invoker.Run(ref_argument);
             out_device_buf.FromDevice(device_output.mData.data());
-            ck::utils::check_err(
-                host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+            return ck::utils::check_err(device_output.mData,
+                                        host_output.mData,
+                                        "Error: incorrect results!",
+                                        1e-5f,
+                                        1e-4f)
+                       ? 0
+                       : 1;
         };
 
         switch(num_dim_spatial)
         {
         case 3: {
             auto ref_conv = ReferenceConvNDFwdInstance<3>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 2: {
             auto ref_conv = ReferenceConvNDFwdInstance<2>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 1: {
             auto ref_conv = ReferenceConvNDFwdInstance<1>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         default: {
             throw std::runtime_error("Unsupported number of spatial dimensions provided!");
         }
         }
     }
+    return 0;
 }

example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp

@@ -0,0 +1,344 @@
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "check_err.hpp"
+#include "config.hpp"
+#include "conv_util.hpp"
+#include "device.hpp"
+#include "device_tensor.hpp"
+#include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
+#include "element_wise_operation.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "reference_conv_fwd.hpp"
+#include "tensor_layout.hpp"
+
+namespace {
+
+using InDataType  = double;
+using WeiDataType = double;
+using OutDataType = double;
+using AccDataType = double;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+using DeviceConvFwdBasePtr =
+    ck::tensor_operation::device::DeviceConvFwdPtr<InElementOp, WeiElementOp, OutElementOp>;
+
+template <ck::index_t NumDimSpatial>
+using DeviceConvNDFwdInstance = ck::tensor_operation::device::
+    DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
+        // clang-format off
+        InDataType,         // 
+        WeiDataType,        //
+        OutDataType,        //
+        AccDataType,        // 
+        InElementOp,        // Input Elementwise Operation
+        WeiElementOp,       // Weights Elementwise Operation
+        OutElementOp,       // Output Elementwise Operation
+        ConvFwdDefault,     // ConvForwardSpecialization
+        NumDimSpatial,      // NumDimSpatial
+        256,                // BlockSize
+        128,                // MPerBlock
+        128,                // NPerBlock
+        4,                  // K0PerBlock
+        2,                  // K1
+        16,                 // MPerXDL
+        16,                 // NPerXDL
+        4,                  // MXdlPerWave
+        4,                  // NXdlPerWave
+        S<4, 64, 1>,        // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<1, 0, 2>,         // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,         // ABlockTransferSrcAccessOrder
+        2,                  // ABlockTransferSrcVectorDim
+        2,                  // ABlockTransferSrcScalarPerVector
+        2,                  // ABlockTransferDstScalarPerVector_K1
+        true,               // ABlockLdsAddExtraM
+        S<4, 64, 1>,        // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<1, 0, 2>,         // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,         // BBlockTransferSrcAccessOrder
+        2,                  // BBlockTransferSrcVectorDim
+        2,                  // BBlockTransferSrcScalarPerVector
+        2,                  // BBlockTransferDstScalarPerVector_K1
+        true,               // BBlockTransferAddExtraN
+        7,                  // CThreadTransferSrcDstVectorDim
+        1>;                 // CThreadTransferDstScalarPerVector
+// clang-format on
+
+template <ck::index_t NumDimSpatial>
+using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
+                                                                                WeiDataType,
+                                                                                OutDataType,
+                                                                                InElementOp,
+                                                                                WeiElementOp,
+                                                                                OutElementOp,
+                                                                                NumDimSpatial>;
+
+DeviceConvFwdBasePtr get_conv_instance(int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return std::make_unique<DeviceConvNDFwdInstance<3>>();
+    }
+    case 2: {
+        return std::make_unique<DeviceConvNDFwdInstance<2>>();
+    }
+    case 1: {
+        return std::make_unique<DeviceConvNDFwdInstance<1>>();
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+
+void print_use_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: run kernel # of times (>1)\n"
+              << "arg4: N spatial dimensions (default 2)\n"
+              << "Following arguments (depending on number of spatial dims):\n"
+              << " N, K, C, \n"
+              << " <filter spatial dimensions>, (ie Y, X for 2D)\n"
+              << " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
+              << " <strides>, (ie Sy, Sx for 2D)\n"
+              << " <dilations>, (ie Dy, Dx for 2D)\n"
+              << " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
+              << " <right padding>, (ie RightPy, RightPx for 2D)\n"
+              << std::endl;
+}
+
+ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, char* argv[])
+{
+    // (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
+    int conv_args     = 3 + num_dim_spatial * 6;
+    int cmdline_nargs = conv_args + 5;
+    if(cmdline_nargs != argc)
+    {
+        print_use_msg();
+        exit(0);
+    }
+
+    ck::utils::conv::ConvParams params;
+    int arg_idx = 5;
+
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
+
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_spatial_lengths_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.conv_filter_strides_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.conv_filter_dilations_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_left_pads_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_right_pads_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    return params;
+}
+
+} // anonymous namespace
+
+int main(int argc, char* argv[])
+{
+    using namespace ck::utils::conv;
+
+    bool do_verification = 0;
+    int init_method      = 0;
+    bool time_kernel     = false;
+    int num_dim_spatial  = 2;
+
+    ck::utils::conv::ConvParams params;
+
+    if(argc >= 5)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        num_dim_spatial = std::stoi(argv[4]);
+    }
+
+    if(argc >= 6)
+    {
+        params = parse_conv_params(num_dim_spatial, argc, argv);
+    }
+
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
+    input_dims.insert(std::end(input_dims),
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
+
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
+    filter_dims.insert(std::end(filter_dims),
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
+
+    const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
+    output_dims.insert(std::end(output_dims),
+                       std::begin(output_spatial_lengths),
+                       std::end(output_spatial_lengths));
+
+    Tensor<InDataType> input(get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
+    Tensor<WeiDataType> weights(get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
+    Tensor<OutDataType> host_output(
+        get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
+    Tensor<OutDataType> device_output(
+        get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weights: " << weights.mDesc << std::endl;
+    std::cout << "output: " << host_output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    case 2:
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        weights.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
+        weights.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpace());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weights.mDesc.GetElementSpace());
+    DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpace());
+
+    in_device_buf.ToDevice(input.mData.data());
+    wei_device_buf.ToDevice(weights.mData.data());
+
+    // do GEMM
+    auto conv    = get_conv_instance(num_dim_spatial);
+    auto invoker = conv->MakeInvokerPointer();
+    auto argument =
+        conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                  static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                  static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                  params.N_,
+                                  params.K_,
+                                  params.C_,
+                                  params.input_spatial_lengths_,
+                                  params.filter_spatial_lengths_,
+                                  output_spatial_lengths,
+                                  params.conv_filter_strides_,
+                                  params.conv_filter_dilations_,
+                                  params.input_left_pads_,
+                                  params.input_right_pads_,
+                                  InElementOp{},
+                                  WeiElementOp{},
+                                  OutElementOp{});
+
+    if(!conv->IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = get_flops(
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
+    std::size_t num_btype =
+        get_btype<InDataType, WeiDataType, OutDataType>(params.N_,
+                                                        params.C_,
+                                                        params.K_,
+                                                        params.input_spatial_lengths_,
+                                                        params.filter_spatial_lengths_,
+                                                        output_spatial_lengths);
+
+    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)
+    {
+        auto verify_f = [&input, &weights, &host_output, &params, &out_device_buf, &device_output](
+                            const auto& ref_conv) {
+            auto ref_invoker  = ref_conv.MakeInvoker();
+            auto ref_argument = ref_conv.MakeArgument(input,
+                                                      weights,
+                                                      host_output,
+                                                      params.conv_filter_strides_,
+                                                      params.conv_filter_dilations_,
+                                                      params.input_left_pads_,
+                                                      params.input_right_pads_,
+                                                      InElementOp{},
+                                                      WeiElementOp{},
+                                                      OutElementOp{});
+
+            ref_invoker.Run(ref_argument);
+            out_device_buf.FromDevice(device_output.mData.data());
+            ck::utils::check_err(
+                host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+        };
+
+        switch(num_dim_spatial)
+        {
+        case 3: {
+            auto ref_conv = ReferenceConvNDFwdInstance<3>();
+            verify_f(ref_conv);
+            break;
+        }
+        case 2: {
+            auto ref_conv = ReferenceConvNDFwdInstance<2>();
+            verify_f(ref_conv);
+            break;
+        }
+        case 1: {
+            auto ref_conv = ReferenceConvNDFwdInstance<1>();
+            verify_f(ref_conv);
+            break;
+        }
+        default: {
+            throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+        }
+        }
+    }
+}

example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp

@@ -5,7 +5,7 @@
 
 #include "check_err.hpp"
 #include "config.hpp"
-#include "conv_fwd_util.hpp"
+#include "conv_util.hpp"
 #include "device.hpp"
 #include "device_tensor.hpp"
 #include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
@@ -45,10 +45,10 @@ template <ck::index_t NumDimSpatial>
 using DeviceConvNDFwdInstance = ck::tensor_operation::device::
     DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
         // clang-format off
-        InDataType,         // 
+        InDataType,         //
         WeiDataType,        //
         OutDataType,        //
-        AccDataType,        // 
+        AccDataType,        //
         InElementOp,        // Input Elementwise Operation
         WeiElementOp,       // Weights Elementwise Operation
         OutElementOp,       // Output Elementwise Operation
@@ -112,7 +112,7 @@ void print_use_msg()
 {
     std::cout << "arg1: verification (0=no, 1=yes)\n"
               << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
-              << "arg3: run kernel # of times (>1)\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
               << "arg4: N spatial dimensions (default 2)\n"
               << "Following arguments (depending on number of spatial dims):\n"
               << " N, K, C, \n"
@@ -139,40 +139,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, cha
     ck::utils::conv::ConvParams params;
     int arg_idx = 5;
 
-    params.num_dim_spatial = num_dim_spatial;
-    params.N               = std::stoi(argv[arg_idx++]);
-    params.K               = std::stoi(argv[arg_idx++]);
-    params.C               = std::stoi(argv[arg_idx++]);
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
 
-    params.filter_spatial_lengths.resize(num_dim_spatial);
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_spatial_lengths.resize(num_dim_spatial);
+    params.input_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_strides.resize(num_dim_spatial);
+    params.conv_filter_strides_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_dilations.resize(num_dim_spatial);
+    params.conv_filter_dilations_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_left_pads.resize(num_dim_spatial);
+    params.input_left_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_right_pads.resize(num_dim_spatial);
+    params.input_right_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
     }
 
     return params;
@@ -184,9 +184,9 @@ int main(int argc, char* argv[])
 {
     using namespace ck::utils::conv;
 
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
     int num_dim_spatial  = 2;
 
     ck::utils::conv::ConvParams params;
@@ -195,7 +195,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
         num_dim_spatial = std::stoi(argv[4]);
     }
 
@@ -204,21 +204,21 @@ int main(int argc, char* argv[])
         params = parse_conv_params(num_dim_spatial, argc, argv);
     }
 
-    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
-                                        static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
     input_dims.insert(std::end(input_dims),
-                      std::begin(params.input_spatial_lengths),
-                      std::end(params.input_spatial_lengths));
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
 
-    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
-                                         static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
     filter_dims.insert(std::end(filter_dims),
-                       std::begin(params.filter_spatial_lengths),
-                       std::end(params.filter_spatial_lengths));
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
 
     const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
-    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
-                                         static_cast<std::size_t>(params.K)};
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
     output_dims.insert(std::end(output_dims),
                        std::begin(output_spatial_lengths),
                        std::end(output_spatial_lengths));
@@ -258,16 +258,16 @@ int main(int argc, char* argv[])
         conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                                   static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                                   static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-                                  params.N,
-                                  params.K,
-                                  params.C,
-                                  params.input_spatial_lengths,
-                                  params.filter_spatial_lengths,
+                                  params.N_,
+                                  params.K_,
+                                  params.C_,
+                                  params.input_spatial_lengths_,
+                                  params.filter_spatial_lengths_,
                                   output_spatial_lengths,
-                                  params.conv_filter_strides,
-                                  params.conv_filter_dilations,
-                                  params.input_left_pads,
-                                  params.input_right_pads,
+                                  params.conv_filter_strides_,
+                                  params.conv_filter_dilations_,
+                                  params.input_left_pads_,
+                                  params.input_right_pads_,
                                   InElementOp{},
                                   WeiElementOp{},
                                   OutElementOp{});
@@ -279,16 +279,16 @@ int main(int argc, char* argv[])
             "not support this Conv problem");
     }
 
-    float ave_time = invoker->Run(argument.get(), nrepeat);
+    float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = get_flops(
-        params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
     std::size_t num_btype = get_btype<InDataType, WeiDataType, OutDataType>(
-        params.N,
-        params.C,
-        params.K,
-        params.input_spatial_lengths,
-        params.filter_spatial_lengths,
+        params.N_,
+        params.C_,
+        params.K_,
+        params.input_spatial_lengths_,
+        params.filter_spatial_lengths_,
         output_spatial_lengths);
 
     float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -304,40 +304,38 @@ int main(int argc, char* argv[])
             auto ref_argument = ref_conv.MakeArgument(input,
                                                       weights,
                                                       host_output,
-                                                      params.conv_filter_strides,
-                                                      params.conv_filter_dilations,
-                                                      params.input_left_pads,
-                                                      params.input_right_pads,
+                                                      params.conv_filter_strides_,
+                                                      params.conv_filter_dilations_,
+                                                      params.input_left_pads_,
+                                                      params.input_right_pads_,
                                                       InElementOp{},
                                                       WeiElementOp{},
                                                       OutElementOp{});
 
             ref_invoker.Run(ref_argument);
             out_device_buf.FromDevice(device_output.mData.data());
-            ck::utils::check_err(
-                host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+            return ck::utils::check_err(
+                host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f) ? 0 : 1;
         };
 
         switch(num_dim_spatial)
         {
         case 3: {
             auto ref_conv = ReferenceConvNDFwdInstance<3>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 2: {
             auto ref_conv = ReferenceConvNDFwdInstance<2>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 1: {
             auto ref_conv = ReferenceConvNDFwdInstance<1>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         default: {
             throw std::runtime_error("Unsupported number of spatial dimensions provided!");
         }
         }
     }
+    return 0;
 }

example/10_conv2d_bwd_data/CMakeLists.txt

@@ -1,2 +1,2 @@
 add_example_executable(example_conv2d_bwd_data_xdl conv2d_bwd_data_xdl.cpp)
-target_link_libraries(example_conv2d_bwd_data_xdl PRIVATE conv_fwd_util)
+target_link_libraries(example_conv2d_bwd_data_xdl PRIVATE conv_util)

example/10_conv2d_bwd_data/conv2d_bwd_data_xdl.cpp

@@ -77,9 +77,9 @@ using ReferenceConvBwdInstance = ck::tensor_operation::host::ReferenceConvBwdDat
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     // Conv shape
     ck::index_t N               = 128;
@@ -102,13 +102,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 19)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         N               = std::stoi(argv[4]);
         K               = std::stoi(argv[5]);
@@ -130,7 +130,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         printf("arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
                "RightPx\n");
         exit(0);
@@ -214,7 +214,7 @@ int main(int argc, char* argv[])
             "not support this Conv problem");
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
 
@@ -249,6 +249,10 @@ int main(int argc, char* argv[])
 
         in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
 
-        ck::utils::check_err(in_n_c_hi_wi_device_result.mData, in_n_c_hi_wi_host_result.mData);
+        return ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
+                                    in_n_c_hi_wi_host_result.mData)
+                   ? 0
+                   : 1;
     }
+    return 0;
 }

example/11_conv2d_bwd_weight/CMakeLists.txt

@@ -1,2 +1,2 @@
 add_example_executable(example_conv2d_bwd_weight_xdl conv2d_bwd_weight_xdl.cpp)
-target_link_libraries(example_conv2d_bwd_weight_xdl PRIVATE conv_fwd_util)
+target_link_libraries(example_conv2d_bwd_weight_xdl PRIVATE conv_util)

example/11_conv2d_bwd_weight/conv2d_bwd_weight_xdl.cpp

@@ -82,9 +82,9 @@ using ReferenceConvBwdWeightInstance =
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
     int do_log           = 0;
     int split_k          = 4;
 
@@ -109,7 +109,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
         do_log          = std::stoi(argv[4]);
         split_k         = std::stoi(argv[5]);
     }
@@ -117,7 +117,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
         do_log          = std::stoi(argv[4]);
         split_k         = std::stoi(argv[5]);
 
@@ -141,7 +141,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         printf("arg4: is show log (0=no, 1=yes)\n");
         printf("arg5: split-k \n");
         printf("arg6 to 19: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
@@ -246,7 +246,7 @@ int main(int argc, char* argv[])
         return 1;
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
 
@@ -291,6 +291,9 @@ int main(int argc, char* argv[])
             LogRangeAsType<float>(std::cout << "wei_host  : ", wei_k_c_y_x_host_result.mData, ",")
                 << std::endl;
         }
-        ck::utils::check_err(wei_k_c_y_x_device_result.mData, wei_k_c_y_x_host_result.mData);
+        return ck::utils::check_err(wei_k_c_y_x_device_result.mData, wei_k_c_y_x_host_result.mData)
+                   ? 0
+                   : 1;
     }
+    return 0;
 }

example/12_reduce/CMakeLists.txt

@@ -1 +1,2 @@
 add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
+add_example_executable(example_reduce_blockwise_two_call reduce_blockwise_two_call.cpp)

example/12_reduce/README.md

@@ -5,23 +5,37 @@
 # -D <xxx> : input 4-d tensor lengths
 # -v <x> :   verification (0=no, 1=yes)
 #arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
-#arg2: run kernel # of times (>1)
-./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 10
+#arg2: time kernel (0=no, 1=yes)
+./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 1
 ```
 
 Result
 ```
-launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1} 
-Warm up
-Start running 3 times...
-Perf: 0.23536 ms, 267.32 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
-error: 0
-max_diff: 0, 529, 529
-root@dc-smc-18:/data/composable_kernel/Build3# bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 10
-launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1} 
-Warm up
+./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 1
+launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
 Start running 10 times...
-Perf: 0.23392 ms, 268.966 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
-error: 0
-max_diff: 0, 528, 528
+Perf: 0.282592 ms, 222.641 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
+```
+
+# Instructions for ```example_reduce_blockwise_two_call```
+
+## Run ```example_reduce_blockwise_two_call```
+```bash
+#arg1:  verification (0=no, 1=yes(
+#arg2:  initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3:  time kernel (0=no, 1=yes)
+./bin/example_reduce_blockwise_two_call 1 2 1
+```
+
+Result
+```
+./bin/example_reduce_blockwise_two_call 1 2 1
+launch_and_time_kernel: grid_dim {204800, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6400, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 2.1791 ms, 771.42 GB/s, DeviceReduceBlockWise<256,M_C32_S1,K_C8_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1> => DeviceReduceBlockWise<256,M_C256_S1,K_C1_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1>
 ```

example/12_reduce/reduce_blockwise.cpp

@@ -12,8 +12,8 @@
 #include "host_tensor_generator.hpp"
 #include "device_tensor.hpp"
 #include "device_base.hpp"
-#include "device_reduce_blockwise.hpp"
-#include "host_reduce_util.hpp"
+#include "device_reduce_multiblock.hpp"
+#include "host_common_util.hpp"
 #include "host_reduction.hpp"
 
 #include "reduction_enums.hpp"
@@ -30,96 +30,53 @@ constexpr int Rank         = 4;
 constexpr int NumReduceDim = 3;
 
 constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
-constexpr NanPropagation NanOpt     = NanPropagation::PROPAGATE_NAN;
-constexpr bool PropagateNan         = (NanOpt == NanPropagation::NOT_PROPAGATE_NAN) ? false : true;
-constexpr ReduceTensorIndices IndicesOpt = ReduceTensorIndices::NO_INDICES;
+constexpr bool PropagateNan         = true;
+constexpr bool OutputIndex          = false;
 
-using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
+using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
 using InElementwiseOperation =
-    typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
+    typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
 using AccElementwiseOperation =
-    typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation;
+    typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
 
-using DeviceReduceInstance = DeviceReduceBlockWise<InDataType,
-                                                   AccDataType,
-                                                   OutDataType,
-                                                   Rank,
-                                                   NumReduceDim,
-                                                   ReduceOperation,
-                                                   InElementwiseOperation,
-                                                   AccElementwiseOperation,
-                                                   PropagateNan,
-                                                   false,
-                                                   256,
-                                                   4,
-                                                   64,
-                                                   1,
-                                                   1,
-                                                   0,
-                                                   1,
-                                                   1>;
+using DeviceReduceInstance = DeviceReduceMultiBlock<InDataType,
+                                                    AccDataType,
+                                                    OutDataType,
+                                                    Rank,
+                                                    NumReduceDim,
+                                                    ReduceOperation,
+                                                    InElementwiseOperation,
+                                                    AccElementwiseOperation,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    PropagateNan,
+                                                    OutputIndex,
+                                                    false, // HaveIndexInputIfOutputIndex
+                                                    256,
+                                                    4,
+                                                    64,
+                                                    1,
+                                                    1,
+                                                    0,
+                                                    1,
+                                                    1>;
 
 static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
-                                       {"scales", required_argument, nullptr, 'S'},
                                        {"verify", required_argument, nullptr, 'v'},
                                        {"help", no_argument, nullptr, '?'},
                                        {nullptr, 0, nullptr, 0}};
 
 class SimpleAppArgs
 {
-    template <typename T>
-    static T getSingleValueFromString(const std::string& valueStr)
-    {
-        std::istringstream iss(valueStr);
-
-        T ret;
-
-        iss >> ret;
-
-        return (ret);
-    };
-
-    template <typename T>
-    static std::vector<T> getTypeValuesFromString(const char* cstr_values)
-    {
-        std::string valuesStr(cstr_values);
-
-        std::vector<T> values;
-        std::size_t pos = 0;
-        std::size_t new_pos;
-
-        new_pos = valuesStr.find(',', pos);
-        while(new_pos != std::string::npos)
-        {
-            const std::string sliceStr = valuesStr.substr(pos, new_pos - pos);
-
-            T val = getSingleValueFromString<T>(sliceStr);
-
-            values.push_back(val);
-
-            pos     = new_pos + 1;
-            new_pos = valuesStr.find(',', pos);
-        };
-
-        std::string sliceStr = valuesStr.substr(pos);
-        T val                = getSingleValueFromString<T>(sliceStr);
-
-        values.push_back(val);
-
-        return (values);
-    };
-
     private:
     int option_index = 0;
 
     public:
-    std::vector<size_t> inLengths;
-    std::vector<float> scales;
+    std::vector<size_t> inLengths = {16, 64, 32, 960};
+    std::vector<float> scales     = {1.0f, 0.0f};
 
-    bool do_verification = false;
-
-    int init_method = 1;
-    int nrepeat     = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
 
     public:
     void show_usage(const char* cmd)
@@ -127,24 +84,24 @@ class SimpleAppArgs
         std::cout << "Usage of " << cmd << std::endl;
         std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
                   << std::endl;
-        std::cout << "--scales or -S, comma separated two float values for alpha and beta"
-                  << std::endl;
         std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
                      "comparing with the host-based reduction"
                   << std::endl;
         std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
                      "value, 3=decimal value)"
                   << std::endl;
-        std::cout << "Arg2 -- number of repeats to run the kernel" << std::endl;
+        std::cout << "Arg2 -- time kernel (0=no, 1=yes)" << std::endl;
     };
 
     int processArgs(int argc, char* argv[])
     {
-        unsigned int ch;
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
 
         while(1)
         {
-            ch = getopt_long(argc, argv, "D:S:v:l:", long_options, &option_index);
+            ch = getopt_long(argc, argv, "D:v:l:", long_options, &option_index);
             if(ch == -1)
                 break;
             switch(ch)
@@ -155,12 +112,6 @@ class SimpleAppArgs
 
                 inLengths = getTypeValuesFromString<size_t>(optarg);
                 break;
-            case 'S':
-                if(!optarg)
-                    throw std::runtime_error("Invalid option format!");
-
-                scales = getTypeValuesFromString<float>(optarg);
-                break;
             case 'v':
                 if(!optarg)
                     throw std::runtime_error("Invalid option format!");
@@ -182,7 +133,7 @@ class SimpleAppArgs
             throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
 
         init_method = std::atoi(argv[optind++]);
-        nrepeat     = std::atoi(argv[optind]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
 
         if(scales.empty())
         {
@@ -196,23 +147,21 @@ class SimpleAppArgs
 
 int main(int argc, char* argv[])
 {
-    using namespace ck::host_reduce;
-
     const std::vector<int> reduceDims{0, 1, 2};
     const std::vector<int> invariantDims{3};
 
     SimpleAppArgs args;
 
-    if(args.processArgs(argc, argv) < 0)
-        return (-1);
+    if(argc > 1)
+    {
+        if(args.processArgs(argc, argv) < 0)
+            return (-1);
+    };
 
     constexpr bool op_support_indices =
         (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
          ReduceOpId == ReduceTensorOp::AMAX);
 
-    constexpr bool NeedIndices =
-        (op_support_indices && (IndicesOpt != ReduceTensorIndices::NO_INDICES));
-
     // if input is half type, no reason to use float for indiced reduction operation and must use
     // float for non-indiced reduction operation for accuracy
     constexpr bool invalid_reduce_1 =
@@ -226,8 +175,7 @@ int main(int argc, char* argv[])
         (op_support_indices && !std::is_same<AccDataType, float>::value);
 
     // indices option can only be used when it is really needed
-    constexpr bool invalid_reduce_3 =
-        (!op_support_indices && IndicesOpt != ReduceTensorIndices::NO_INDICES);
+    constexpr bool invalid_reduce_3 = (!op_support_indices && OutputIndex);
 
     constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3);
 
@@ -295,51 +243,65 @@ int main(int argc, char* argv[])
     if(beta != 0.0f)
         out_dev.ToDevice(out.mData.data());
 
-    size_t indicesSizeInBytes = NeedIndices ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
+    size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
 
-    DeviceMem out_indices_dev(indicesSizeInBytes);
+    DeviceMem out_index_dev(indicesSizeInBytes);
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
 
     if(args.do_verification)
     {
         ReductionHost<InDataType,
                       AccDataType,
                       OutDataType,
-                      ReduceOpId,
+                      ReduceOperation,
+                      InElementwiseOperation,
+                      AccElementwiseOperation,
                       Rank,
                       NumReduceDim,
                       PropagateNan,
-                      NeedIndices>
+                      OutputIndex>
             hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
 
-        hostReduce.Run(
-            alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data());
+        hostReduce.Run(alpha,
+                       in.mData.data(),
+                       beta,
+                       out_ref.mData.data(),
+                       out_indices_ref.mData.data(),
+                       in_elementwise_op,
+                       acc_elementwise_op);
     };
 
-    const auto i_inLengths  = to_int_vector(args.inLengths);
-    const auto i_inStrides  = to_int_vector(inStrides);
-    const auto i_outLengths = to_int_vector(outLengths);
-    const auto i_outStrides = to_int_vector(outStrides);
+    std::vector<ck::index_t> i_inLengths;
+    std::vector<ck::index_t> i_inStrides;
+    std::vector<ck::index_t> i_outLengths;
+    std::vector<ck::index_t> i_outStrides;
+
+    i_inLengths.assign(args.inLengths.begin(), args.inLengths.end());
+    i_inStrides.assign(inStrides.begin(), inStrides.end());
+    i_outLengths.assign(outLengths.begin(), outLengths.end());
+    i_outStrides.assign(outStrides.begin(), outStrides.end());
 
     auto reduce = DeviceReduceInstance{};
 
-    auto wsSizeInBytes = reduce.GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
-
-    DeviceMem ws_dev(wsSizeInBytes);
-
-    auto argument_ptr =
-        reduce.MakeArgumentPointer(i_inLengths,
-                                   i_inStrides,
-                                   i_outLengths,
-                                   i_outStrides,
-                                   reduceDims,
-                                   alpha,
-                                   beta,
-                                   in_dev.GetDeviceBuffer(),
-                                   out_dev.GetDeviceBuffer(),
-                                   out_indices_dev.GetDeviceBuffer(),
-                                   ws_dev.GetDeviceBuffer(),
-                                   InElementwiseOperation{static_cast<int>(reduce_total_length)},
-                                   AccElementwiseOperation{static_cast<int>(reduce_total_length)});
+    auto argument_ptr = reduce.MakeArgumentPointer(i_inLengths,
+                                                   i_inStrides,
+                                                   i_outLengths,
+                                                   i_outStrides,
+                                                   reduceDims,
+                                                   alpha,
+                                                   beta,
+                                                   in_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   out_dev.GetDeviceBuffer(),
+                                                   out_index_dev.GetDeviceBuffer(),
+                                                   in_elementwise_op,
+                                                   acc_elementwise_op);
 
     if(!reduce.IsSupportedArgument(argument_ptr.get()))
     {
@@ -352,7 +314,7 @@ int main(int argc, char* argv[])
 
     auto invoker_ptr = reduce.MakeInvokerPointer();
 
-    float avg_time = invoker_ptr->Run(argument_ptr.get(), args.nrepeat);
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, args.time_kernel});
 
     std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InDataType) +
                             invariant_total_length * sizeof(OutDataType);
@@ -362,16 +324,19 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
               << std::endl;
 
+    bool pass = true;
+
     if(args.do_verification)
     {
         out_dev.FromDevice(out.mData.data());
-        ck::utils::check_err(out.mData, out_ref.mData);
+        pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
 
-        if(NeedIndices)
+        if(OutputIndex)
         {
-            out_indices_dev.FromDevice(out_indices.mData.data());
-            ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
-            ;
+            out_index_dev.FromDevice(out_indices.mData.data());
+            pass = pass && ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
         };
     };
+
+    return (pass ? 0 : 1);
 }

example/12_reduce/reduce_blockwise_two_call.cpp

@@ -0,0 +1,301 @@
+#include <iostream>
+#include <numeric>
+#include <sstream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "check_err.hpp"
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "device_base.hpp"
+#include "device_reduce_multiblock.hpp"
+#include "host_common_util.hpp"
+#include "host_reduction.hpp"
+
+#include "reduction_enums.hpp"
+#include "reduction_operator_mapping.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InOutDataType = ck::half_t;
+using InOutDataType = ck::half_t;
+using AccDataType   = float;
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
+constexpr bool PropagateNan         = true;
+constexpr bool OutputIndex          = false;
+
+using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+using InElementwiseOperation =
+    typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+using AccElementwiseOperation =
+    typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+using DeviceReduceInstance_1 = DeviceReduceMultiBlock<InOutDataType,
+                                                      AccDataType,
+                                                      InOutDataType,
+                                                      5, // Rank
+                                                      1, // NumReduceDim
+                                                      ReduceOperation,
+                                                      InElementwiseOperation,
+                                                      PassThroughOp,
+                                                      InMemoryDataOperationEnum::Set,
+                                                      PropagateNan,
+                                                      OutputIndex,
+                                                      false, // HaveIndexInputIfOutputIndex
+                                                      256,
+                                                      32,
+                                                      8,
+                                                      1,
+                                                      1,
+                                                      1, // vector dim
+                                                      1,
+                                                      1>;
+
+using DeviceReduceInstance_2 = DeviceReduceMultiBlock<InOutDataType,
+                                                      AccDataType,
+                                                      InOutDataType,
+                                                      4, // Rank
+                                                      1, // NumReduceDim
+                                                      ReduceOperation,
+                                                      PassThroughOp,
+                                                      AccElementwiseOperation,
+                                                      InMemoryDataOperationEnum::Set,
+                                                      PropagateNan,
+                                                      OutputIndex,
+                                                      false, // HaveIndexInputIfOutputIndex
+                                                      256,
+                                                      128,
+                                                      2,
+                                                      1,
+                                                      1,
+                                                      1, // vector dim
+                                                      1,
+                                                      1>;
+
+static bool do_verify;
+static int init_method;
+static float alpha;
+static float beta;
+static bool time_kernel;
+
+int main(int argc, char* argv[])
+{
+    // used by the device reduction
+    const std::vector<int> reduceDims_1    = {4};
+    const std::vector<int> invariantDims_1 = {0, 1, 2, 3};
+
+    const std::vector<int> reduceDims_2    = {3};
+    const std::vector<int> invariantDims_2 = {0, 1, 2};
+
+    // used by the host reduction
+    const std::vector<int> reduceDims    = {3, 4};
+    const std::vector<int> invariantDims = {0, 1, 2};
+
+    const std::vector<size_t> inLengths_1 = {64, 320, 80, 4, 128};
+
+    // input lengths of the second reduction, which is also the output lengths of the first
+    // reduction
+    const std::vector<size_t> inLengths_2 = {64, 320, 80, 4};
+
+    const std::vector<size_t> outLengths = {64, 320, 80};
+
+    if(argc == 1)
+    {
+        do_verify   = true;
+        init_method = 2;
+        time_kernel = true;
+    }
+    else if(argc == 4)
+    {
+        do_verify   = static_cast<bool>(argv[1]);
+        init_method = atoi(argv[2]);
+        time_kernel = static_cast<bool>(atoi(argv[3]));
+    }
+    else
+    {
+        std::ostringstream ostr;
+
+        ostr << "Wrong parameter! " << std::endl
+             << "Usage: " << argv[0] << "[verify 0/1] init_method time_kernel" << std::endl;
+
+        throw std::runtime_error(ostr.str());
+    };
+
+    alpha = 1.0f;
+    beta  = 0.0f;
+
+    Tensor<InOutDataType> in_1(inLengths_1);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> in_2(inLengths_2); // also the output tensor of the first reduction
+    Tensor<InOutDataType> out(outLengths);
+
+    auto inStrides_1 = in_1.mDesc.GetStrides();
+    auto inStrides_2 = in_2.mDesc.GetStrides();
+    auto outStrides  = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in_1.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verify)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in_1.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in_1.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in_1.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpace(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    DeviceMem in_1_dev(sizeof(InOutDataType) * in_1.mDesc.GetElementSpace());
+    DeviceMem in_2_dev(sizeof(InOutDataType) * in_2.mDesc.GetElementSpace());
+    DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpace());
+
+    in_1_dev.ToDevice(in_1.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    if(do_verify)
+    {
+        ReductionHost<InOutDataType,
+                      AccDataType,
+                      InOutDataType,
+                      ReduceOperation,
+                      InElementwiseOperation,
+                      AccElementwiseOperation,
+                      5, // Rank
+                      2, // NumReduceDim
+                      PropagateNan,
+                      OutputIndex>
+            hostReduce(in_1.mDesc, out_ref.mDesc, invariantDims, reduceDims);
+
+        hostReduce.Run(alpha,
+                       in_1.mData.data(),
+                       beta,
+                       out_ref.mData.data(),
+                       nullptr,
+                       in_elementwise_op,
+                       acc_elementwise_op);
+    };
+
+    std::vector<ck::index_t> i_inLengths_1;
+    std::vector<ck::index_t> i_inStrides_1;
+    std::vector<ck::index_t> i_inLengths_2;
+    std::vector<ck::index_t> i_inStrides_2;
+    std::vector<ck::index_t> i_outLengths;
+    std::vector<ck::index_t> i_outStrides;
+
+    i_inLengths_1.assign(inLengths_1.begin(), inLengths_1.end());
+    i_inStrides_1.assign(inStrides_1.begin(), inStrides_1.end());
+    i_inLengths_2.assign(inLengths_2.begin(), inLengths_2.end());
+    i_inStrides_2.assign(inStrides_2.begin(), inStrides_2.end());
+    i_outLengths.assign(outLengths.begin(), outLengths.end());
+    i_outStrides.assign(outStrides.begin(), outStrides.end());
+
+    auto reduce_1 = DeviceReduceInstance_1{};
+
+    auto argument_ptr_1 = reduce_1.MakeArgumentPointer(i_inLengths_1,
+                                                       i_inStrides_1,
+                                                       i_inLengths_2,
+                                                       i_inStrides_2,
+                                                       reduceDims_1,
+                                                       1.0f,
+                                                       0.0f,
+                                                       in_1_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       in_2_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       in_elementwise_op,
+                                                       PassThroughOp{});
+
+    if(!reduce_1.IsSupportedArgument(argument_ptr_1.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+    };
+
+    auto invoker_ptr_1 = reduce_1.MakeInvokerPointer();
+
+    auto reduce_2 = DeviceReduceInstance_2{};
+
+    auto argument_ptr_2 = reduce_2.MakeArgumentPointer(i_inLengths_2,
+                                                       i_inStrides_2,
+                                                       i_outLengths,
+                                                       i_outStrides,
+                                                       reduceDims_2,
+                                                       alpha,
+                                                       beta,
+                                                       in_2_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       out_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       PassThroughOp{},
+                                                       acc_elementwise_op);
+
+    if(!reduce_2.IsSupportedArgument(argument_ptr_2.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+    };
+
+    auto invoker_ptr_2 = reduce_2.MakeInvokerPointer();
+
+    float avg_time_1 = invoker_ptr_1->Run(argument_ptr_1.get(), StreamConfig{nullptr, time_kernel});
+    float avg_time_2 = invoker_ptr_2->Run(argument_ptr_2.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / (avg_time_1 + avg_time_2);
+
+    std::cout << "Perf: " << avg_time_1 + avg_time_2 << " ms, " << gb_per_sec << " GB/s, "
+              << reduce_1.GetTypeString() << " => " << reduce_2.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verify)
+    {
+        out_dev.FromDevice(out.mData.data());
+        pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
+    };
+
+    return (pass ? 0 : 1);
+}

example/13_pool2d_fwd/CMakeLists.txt

@@ -1 +1,3 @@
-add_example_executable(example_pool2d_fwd pool2d_fwd.cpp)
+add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
+add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
+

example/13_pool2d_fwd/README.md

@@ -1,12 +1,12 @@
-# Instructions for ```example_pool2d_fwd``` Example
+# Instructions for ```example_pool2d_fwd``` Examples
 
-## Run ```example_pool2d_fwd```
+## Run ```example_pool2d_fwd_fp16```
 ```bash
 #arg1: verification (0=no, 1=yes)
 #arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
-#arg3: run kernel # of times (>1)
+#arg3: time kernel (0=no, 1=yes)
 #arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
-./bin/example_pool2d_fwd 1 1 10
+./bin/example_pool2d_fwd_fp16 1 1 1
 ```
 
 Result 
@@ -14,9 +14,28 @@ Result
 in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
 out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
 launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1} 
-Warm up
+Warm up 1 time
 Start running 10 times...
-Perf: 0.415453 ms, 1.37996 TFlops, 749.726 GB/s
-error: 0
-max_diff: 0, 1, 1
+Perf: 0.397436 ms, 1.44252 TFlops, 783.713 GB/s
+```
+
+## Run ```example_pool2d_fwd_fp32```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
+./bin/example_pool2d_fwd_fp32 1 1 1
+```
+
+
+Result 
+```
+./bin/example_pool2d_fwd_fp32 1 1 1
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
+launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.01823 ms, 0.563045 TFlops, 611.8 GB/s
 ```

example/13_pool2d_fwd/pool2d_fwd.cpp

@@ -1,315 +0,0 @@
-#include <iostream>
-#include <numeric>
-#include <initializer_list>
-#include <cstdlib>
-#include <stdlib.h>
-
-#include "check_err.hpp"
-#include "config.hpp"
-#include "print.hpp"
-#include "device.hpp"
-#include "host_tensor.hpp"
-#include "host_tensor_generator.hpp"
-#include "host_reduce_util.hpp"
-#include "device_tensor.hpp"
-#include "tensor_layout.hpp"
-#include "reduction_operator.hpp"
-#include "device_pool2d_fwd_nhwc_nhwc.hpp"
-
-using InDataType  = ck::half_t;
-using OutDataType = ck::half_t;
-using AccDataType = float;
-
-using InLayout  = ck::tensor_layout::convolution::NHWC;
-using OutLayout = ck::tensor_layout::convolution::NHWC;
-
-#if 1
-static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
-#else
-static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
-#endif
-
-static constexpr bool NeedIndices  = false;
-static constexpr bool PropagateNan = false;
-
-using DevicePoolFwdInstance =
-    ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
-        InDataType,  // InDataType
-        OutDataType, // OutDataType
-        AccDataType, // AccDataType
-        ReduceOpId,
-        NeedIndices,
-        64, // BlockSize
-        64, // ReduceMThreadClusterSize
-        1,  // ReduceKThreadClusterSize
-        4,  // ReduceMThreadSliceSize
-        1,  // ReduceKThreadSliceSize
-        4>; // InSrcOutDstVectorSize
-
-template <typename InDataType,
-          typename OutDataType,
-          typename AccDataType,
-          ck::ReduceTensorOp ReduceOpId,
-          bool PropagateNan,
-          bool NeedIndices>
-static void pool_host_verify(const Tensor<InDataType>& in,
-                             Tensor<OutDataType>& out,
-                             Tensor<int>& out_indices,
-                             const std::array<ck::index_t, 2>& window_spatial_lengths,
-                             const std::array<ck::index_t, 2>& window_strides,
-                             const std::array<ck::index_t, 2>& in_left_pads,
-                             const std::array<ck::index_t, 2>& /*in_right_pads*/)
-{
-    using namespace ck::host_reduce;
-
-    const int divider = window_spatial_lengths[0] * window_spatial_lengths[1];
-
-    const auto PreUnaryOp = PreUnaryOpFn<AccDataType, ReduceOpId>(divider);
-    const auto PosUnaryOp = PosUnaryOpFn<AccDataType, ReduceOpId>(divider);
-
-    if constexpr(!NeedIndices)
-    {
-        auto opReduce = ReduceOpFn<AccDataType, ReduceOpId>();
-
-        auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
-            auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
-
-            for(int y = 0; y < window_spatial_lengths[0]; ++y)
-            {
-                int hi = ho * window_strides[0] + y - in_left_pads[0];
-                for(int x = 0; x < window_spatial_lengths[1]; ++x)
-                {
-                    int wi = wo * window_strides[1] + x - in_left_pads[1];
-                    if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
-                       wi < in.mDesc.GetLengths()[3])
-                    {
-                        AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
-
-                        PreUnaryOp(currVal);
-
-                        binop_with_nan_check<AccDataType, PropagateNan>(opReduce, accuVal, currVal);
-                    }
-                }
-            }
-
-            PosUnaryOp(accuVal);
-
-            out(n, c, ho, wo) = accuVal;
-        };
-
-        make_ParallelTensorFunctor(f_nchw,
-                                   out.mDesc.GetLengths()[0],
-                                   out.mDesc.GetLengths()[1],
-                                   out.mDesc.GetLengths()[2],
-                                   out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
-    }
-    else
-    {
-        auto opReduce = ReduceOpFn2<AccDataType, ReduceOpId>();
-
-        auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
-            auto accuVal  = ReduceOpZeroVal<AccDataType, ReduceOpId>();
-            int accuIndex = 0;
-
-            for(int y = 0; y < window_spatial_lengths[0]; ++y)
-            {
-                int hi = ho * window_strides[0] + y - in_left_pads[0];
-                for(int x = 0; x < window_spatial_lengths[1]; ++x)
-                {
-                    int wi = wo * window_strides[1] + x - in_left_pads[1];
-                    if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
-                       wi < in.mDesc.GetLengths()[3])
-                    {
-                        AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
-                        int currIndex       = y * window_spatial_lengths[1] + x;
-
-                        PreUnaryOp(currVal);
-
-                        binop_with_nan_check2<AccDataType, PropagateNan>(
-                            opReduce, accuVal, currVal, accuIndex, currIndex);
-                    }
-                }
-            }
-
-            PosUnaryOp(accuVal);
-
-            out(n, c, ho, wo)         = accuVal;
-            out_indices(n, c, ho, wo) = accuIndex;
-        };
-
-        make_ParallelTensorFunctor(f_nchw,
-                                   out.mDesc.GetLengths()[0],
-                                   out.mDesc.GetLengths()[1],
-                                   out.mDesc.GetLengths()[2],
-                                   out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
-    };
-}
-
-int main(int argc, char* argv[])
-{
-    using namespace ck::host_reduce;
-
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
-
-    // Pool shape
-    ck::index_t N               = 128;
-    ck::index_t C               = 192;
-    ck::index_t Y               = 3;
-    ck::index_t X               = 3;
-    ck::index_t Hi              = 71;
-    ck::index_t Wi              = 71;
-    ck::index_t window_stride_h = 2;
-    ck::index_t window_stride_w = 2;
-    ck::index_t in_left_pad_h   = 1;
-    ck::index_t in_left_pad_w   = 1;
-    ck::index_t in_right_pad_h  = 1;
-    ck::index_t in_right_pad_w  = 1;
-
-    if(argc == 4)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
-    }
-    else if(argc == 16)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
-
-        N               = std::stoi(argv[4]);
-        C               = std::stoi(argv[5]);
-        Y               = std::stoi(argv[6]);
-        X               = std::stoi(argv[7]);
-        Hi              = std::stoi(argv[8]);
-        Wi              = std::stoi(argv[9]);
-        window_stride_h = std::stoi(argv[10]);
-        window_stride_w = std::stoi(argv[11]);
-        in_left_pad_h   = std::stoi(argv[12]);
-        in_left_pad_w   = std::stoi(argv[13]);
-        in_right_pad_h  = std::stoi(argv[14]);
-        in_right_pad_w  = std::stoi(argv[15]);
-    }
-    else
-    {
-        printf("arg1: verification (0=no, 1=yes)\n");
-        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
-        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
-               "RightPx\n");
-        exit(0);
-    }
-
-    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
-    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
-
-    const std::array<ck::index_t, 2> window_spatial_lengths{{Y, X}};
-    const std::array<ck::index_t, 2> window_strides{{window_stride_h, window_stride_w}};
-    const std::array<ck::index_t, 2> input_left_pads{{in_left_pad_h, in_left_pad_w}};
-    const std::array<ck::index_t, 2> input_right_pads{{in_right_pad_h, in_right_pad_w}};
-
-    // tensor layout
-    auto f_host_tensor_descriptor =
-        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
-            if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value)
-            {
-                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
-                                            std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
-            }
-            else if constexpr(ck::is_same<decltype(layout),
-                                          ck::tensor_layout::convolution::NHWC>::value)
-            {
-                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
-                                            std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
-            }
-        };
-
-    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
-    Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
-    Tensor<int> out_indices_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
-    Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
-    Tensor<int> out_indices_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
-
-    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
-    std::cout << "out_n_c_ho_wo: " << out_n_c_ho_wo_host.mDesc << std::endl;
-
-    switch(init_method)
-    {
-    case 0: break;
-    case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}); break;
-    case 2: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
-    default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0});
-    }
-
-    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
-    DeviceMem out_device_buf(sizeof(OutDataType) * out_n_c_ho_wo_device.mDesc.GetElementSpace());
-    DeviceMem out_indices_device_buf(sizeof(int) *
-                                     out_indices_n_c_ho_wo_device.mDesc.GetElementSpace());
-
-    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
-
-    auto pool        = DevicePoolFwdInstance{};
-    auto invoker_ptr = pool.MakeInvokerPointer();
-    auto argument_ptr =
-        pool.MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
-                                 static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-                                 static_cast<int*>(out_indices_device_buf.GetDeviceBuffer()),
-                                 N,
-                                 C,
-                                 std::array<ck::index_t, 2>{{Hi, Wi}},
-                                 std::array<ck::index_t, 2>{{Y, X}},
-                                 std::array<ck::index_t, 2>{{Ho, Wo}},
-                                 window_strides,
-                                 input_left_pads,
-                                 input_right_pads);
-
-    if(!pool.IsSupportedArgument(argument_ptr.get()))
-    {
-        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
-                                 "not support this problem");
-    }
-
-    float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
-
-    std::size_t flop = std::size_t(2) * N * C * Ho * Wo * Y * X;
-
-    std::size_t num_btype =
-        sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(OutDataType) * (N * C * Ho * Wo);
-
-    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)
-    {
-        pool_host_verify<InDataType,
-                         OutDataType,
-                         AccDataType,
-                         ReduceOpId,
-                         PropagateNan,
-                         NeedIndices>(in_n_c_hi_wi,
-                                      out_n_c_ho_wo_host,
-                                      out_indices_n_c_ho_wo_host,
-                                      window_spatial_lengths,
-                                      window_strides,
-                                      input_left_pads,
-                                      input_right_pads);
-
-        out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
-
-        ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
-
-        if constexpr(NeedIndices)
-        {
-            out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
-
-            //          ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
-            //          out_indices_n_c_ho_wo_host.mData);;
-        };
-    }
-}

example/13_pool2d_fwd/pool2d_fwd_common.hpp

@@ -0,0 +1,280 @@
+#pragma once
+
+#include <iostream>
+
+#include "check_err.hpp"
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "tensor_layout.hpp"
+#include "reduction_enums.hpp"
+#include "reduction_operator_mapping.hpp"
+#include "reduction_functions_accumulate.hpp"
+
+#include "device_pool2d_fwd_nhwc_nhwc.hpp"
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+static void pool_host_verify(const Tensor<InDataType>& in,
+                             Tensor<OutDataType>& out,
+                             Tensor<IndexDataType>& out_indices,
+                             const std::array<ck::index_t, 2>& window_spatial_lengths,
+                             const std::array<ck::index_t, 2>& window_strides,
+                             const std::array<ck::index_t, 2>& in_left_pads,
+                             const std::array<ck::index_t, 2>& /*in_right_pads*/)
+{
+    const int32_t reduceLength = window_spatial_lengths[0] * window_spatial_lengths[1];
+
+    using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
+
+    auto elementwise_ops =
+        ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
+
+    auto in_elementwise_op  = std::get<0>(elementwise_ops);
+    auto acc_elementwise_op = std::get<1>(elementwise_ops);
+
+    if constexpr(!OutputIndex)
+    {
+        using Accumulation =
+            ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+        auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
+            auto accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+            for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
+            {
+                ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
+                for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
+                {
+                    ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
+                    if(hi >= 0 && hi < static_cast<ck::index_t>(in.mDesc.GetLengths()[2]) &&
+                       wi >= 0 && wi < static_cast<ck::index_t>(in.mDesc.GetLengths()[3]))
+                    {
+                        AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
+
+                        in_elementwise_op(currVal, currVal);
+
+                        Accumulation::Calculate(accuVal, currVal);
+                    }
+                }
+            }
+
+            acc_elementwise_op(accuVal, accuVal);
+
+            out(n, c, ho, wo) = accuVal;
+        };
+
+        make_ParallelTensorFunctor(f_nchw,
+                                   out.mDesc.GetLengths()[0],
+                                   out.mDesc.GetLengths()[1],
+                                   out.mDesc.GetLengths()[2],
+                                   out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
+    }
+    else
+    {
+        using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                        ReduceOperation,
+                                                                        AccDataType,
+                                                                        IndexDataType>;
+        auto f_nchw        = [&](auto n, auto c, auto ho, auto wo) {
+            auto accuVal            = ReduceOperation::template GetIdentityValue<AccDataType>();
+            IndexDataType accuIndex = 0;
+
+            for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
+            {
+                ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
+                for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
+                {
+                    ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
+                    if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
+                       wi < in.mDesc.GetLengths()[3])
+                    {
+                        AccDataType currVal     = static_cast<AccDataType>(in(n, c, hi, wi));
+                        IndexDataType currIndex = y * window_spatial_lengths[1] + x;
+
+                        in_elementwise_op(currVal, currVal);
+
+                        Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                    }
+                }
+            }
+
+            acc_elementwise_op(accuVal, accuVal);
+
+            out(n, c, ho, wo)         = accuVal;
+            out_indices(n, c, ho, wo) = accuIndex;
+        };
+
+        make_ParallelTensorFunctor(f_nchw,
+                                   out.mDesc.GetLengths()[0],
+                                   out.mDesc.GetLengths()[1],
+                                   out.mDesc.GetLengths()[2],
+                                   out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
+    };
+}
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+bool pool_test(bool do_verification,
+               int init_method,
+               bool time_kernel,
+               ck::index_t N,
+               ck::index_t C,
+               ck::index_t Y,
+               ck::index_t X,
+               ck::index_t Hi,
+               ck::index_t Wi,
+               ck::index_t window_stride_h,
+               ck::index_t window_stride_w,
+               ck::index_t in_left_pad_h,
+               ck::index_t in_left_pad_w,
+               ck::index_t in_right_pad_h,
+               ck::index_t in_right_pad_w)
+{
+    using DevicePoolFwdInstance =
+        ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
+            InDataType,  // InDataType
+            OutDataType, // OutDataType
+            AccDataType, // AccDataType
+            ReduceOpId,
+            OutputIndex,
+            64, // BlockSize
+            64, // ReduceMThreadClusterSize
+            1,  // ReduceKThreadClusterSize
+            4,  // ReduceMThreadSliceSize
+            1,  // ReduceKThreadSliceSize
+            4>; // InSrcOutDstVectorSize
+
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
+
+    const std::array<ck::index_t, 2> window_spatial_lengths{{Y, X}};
+    const std::array<ck::index_t, 2> window_strides{{window_stride_h, window_stride_w}};
+    const std::array<ck::index_t, 2> input_left_pads{{in_left_pad_h, in_left_pad_w}};
+    const std::array<ck::index_t, 2> input_right_pads{{in_right_pad_h, in_right_pad_w}};
+
+    // tensor layout
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
+                                            std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
+            }
+            else if constexpr(ck::is_same<decltype(layout),
+                                          ck::tensor_layout::convolution::NHWC>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
+                                            std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
+            }
+        };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
+    Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_host(
+        f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_device(
+        f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "out_n_c_ho_wo: " << out_n_c_ho_wo_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}); break;
+    case 2: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
+    default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_n_c_ho_wo_device.mDesc.GetElementSpace());
+    DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
+                                     out_indices_n_c_ho_wo_device.mDesc.GetElementSpace());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+
+    auto pool         = DevicePoolFwdInstance{};
+    auto invoker_ptr  = pool.MakeInvokerPointer();
+    auto argument_ptr = pool.MakeArgumentPointer(
+        static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+        N,
+        C,
+        std::array<ck::index_t, 2>{{Hi, Wi}},
+        std::array<ck::index_t, 2>{{Y, X}},
+        std::array<ck::index_t, 2>{{Ho, Wo}},
+        window_strides,
+        input_left_pads,
+        input_right_pads);
+
+    if(!pool.IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * N * C * Ho * Wo * Y * X;
+
+    std::size_t num_btype =
+        sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(OutDataType) * (N * C * Ho * Wo);
+
+    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;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        pool_host_verify<InDataType,
+                         OutDataType,
+                         AccDataType,
+                         IndexDataType,
+                         ReduceOpId,
+                         PropagateNan,
+                         OutputIndex>(in_n_c_hi_wi,
+                                      out_n_c_ho_wo_host,
+                                      out_indices_n_c_ho_wo_host,
+                                      window_spatial_lengths,
+                                      window_strides,
+                                      input_left_pads,
+                                      input_right_pads);
+
+        out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
+
+        pass = pass && ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
+
+        if constexpr(OutputIndex)
+        {
+            out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
+
+            pass = pass && ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
+                                                out_indices_n_c_ho_wo_host.mData);
+        };
+    }
+
+    return (pass);
+};

example/13_pool2d_fwd/pool2d_fwd_fp16.cpp

@@ -0,0 +1,114 @@
+#include <iostream>
+#include <cstdlib>
+
+#include "config.hpp"
+#include "tensor_layout.hpp"
+#include "reduction_enums.hpp"
+
+#include "pool2d_fwd_common.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using OutLayout = ck::tensor_layout::convolution::NHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification;
+    int init_method;
+    bool time_kernel;
+
+    // Pool shape
+    ck::index_t N               = 128;
+    ck::index_t C               = 192;
+    ck::index_t Y               = 3;
+    ck::index_t X               = 3;
+    ck::index_t Hi              = 71;
+    ck::index_t Wi              = 71;
+    ck::index_t window_stride_h = 2;
+    ck::index_t window_stride_w = 2;
+    ck::index_t in_left_pad_h   = 1;
+    ck::index_t in_left_pad_w   = 1;
+    ck::index_t in_right_pad_h  = 1;
+    ck::index_t in_right_pad_w  = 1;
+
+    if(argc == 1)
+    {
+        do_verification = true;
+        init_method     = 1;
+        time_kernel     = true;
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+    }
+    else if(argc == 16)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+
+        N               = std::stoi(argv[4]);
+        C               = std::stoi(argv[5]);
+        Y               = std::stoi(argv[6]);
+        X               = std::stoi(argv[7]);
+        Hi              = std::stoi(argv[8]);
+        Wi              = std::stoi(argv[9]);
+        window_stride_h = std::stoi(argv[10]);
+        window_stride_w = std::stoi(argv[11]);
+        in_left_pad_h   = std::stoi(argv[12]);
+        in_left_pad_w   = std::stoi(argv[13]);
+        in_right_pad_h  = std::stoi(argv[14]);
+        in_right_pad_w  = std::stoi(argv[15]);
+    }
+    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 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(0);
+    }
+
+    bool pass = pool_test<InDataType,
+                          OutDataType,
+                          AccDataType,
+                          IndexDataType,
+                          InLayout,
+                          OutLayout,
+                          ReduceOpId,
+                          PropagateNan,
+                          OutputIndex>(do_verification,
+                                       init_method,
+                                       time_kernel,
+                                       N,
+                                       C,
+                                       Y,
+                                       X,
+                                       Hi,
+                                       Wi,
+                                       window_stride_h,
+                                       window_stride_w,
+                                       in_left_pad_h,
+                                       in_left_pad_w,
+                                       in_right_pad_h,
+                                       in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}

example/13_pool2d_fwd/pool2d_fwd_fp32.cpp

@@ -0,0 +1,114 @@
+#include <iostream>
+#include <cstdlib>
+
+#include "config.hpp"
+#include "tensor_layout.hpp"
+#include "reduction_enums.hpp"
+
+#include "pool2d_fwd_common.hpp"
+
+using InDataType  = float;
+using OutDataType = float;
+using AccDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using OutLayout = ck::tensor_layout::convolution::NHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification;
+    int init_method;
+    bool time_kernel;
+
+    // Pool shape
+    ck::index_t N               = 128;
+    ck::index_t C               = 192;
+    ck::index_t Y               = 3;
+    ck::index_t X               = 3;
+    ck::index_t Hi              = 71;
+    ck::index_t Wi              = 71;
+    ck::index_t window_stride_h = 2;
+    ck::index_t window_stride_w = 2;
+    ck::index_t in_left_pad_h   = 1;
+    ck::index_t in_left_pad_w   = 1;
+    ck::index_t in_right_pad_h  = 1;
+    ck::index_t in_right_pad_w  = 1;
+
+    if(argc == 1)
+    {
+        do_verification = true;
+        init_method     = 1;
+        time_kernel     = true;
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+    }
+    else if(argc == 16)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+
+        N               = std::stoi(argv[4]);
+        C               = std::stoi(argv[5]);
+        Y               = std::stoi(argv[6]);
+        X               = std::stoi(argv[7]);
+        Hi              = std::stoi(argv[8]);
+        Wi              = std::stoi(argv[9]);
+        window_stride_h = std::stoi(argv[10]);
+        window_stride_w = std::stoi(argv[11]);
+        in_left_pad_h   = std::stoi(argv[12]);
+        in_left_pad_w   = std::stoi(argv[13]);
+        in_right_pad_h  = std::stoi(argv[14]);
+        in_right_pad_w  = std::stoi(argv[15]);
+    }
+    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 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(0);
+    }
+
+    bool pass = pool_test<InDataType,
+                          OutDataType,
+                          AccDataType,
+                          IndexDataType,
+                          InLayout,
+                          OutLayout,
+                          ReduceOpId,
+                          PropagateNan,
+                          OutputIndex>(do_verification,
+                                       init_method,
+                                       time_kernel,
+                                       N,
+                                       C,
+                                       Y,
+                                       X,
+                                       Hi,
+                                       Wi,
+                                       window_stride_h,
+                                       window_stride_w,
+                                       in_left_pad_h,
+                                       in_left_pad_w,
+                                       in_right_pad_h,
+                                       in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}

example/14_gemm_xdl_requant_relu_requant/gemm_xdl_requant_relu_requant_int8.cpp

@@ -100,14 +100,19 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
      16>;                        // index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
 // clang-format on
 
-using ReferenceGemmInstance = ck::tensor_operation::host::
-    ReferenceGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, RequantReluRequant>;
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        float,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        RequantReluRequant>;
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     // GEMM shape
     ck::index_t M = 3840;
@@ -125,13 +130,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 10)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -145,7 +150,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\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);
     }
@@ -219,7 +224,7 @@ int main(int argc, char* argv[])
             "not support this GEMM problem");
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    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 =
@@ -244,7 +249,7 @@ int main(int argc, char* argv[])
 
         ref_invoker.Run(ref_argument);
 
-        ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+        return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
     }
 
     return 0;

example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp

@@ -56,29 +56,29 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdl
 // clang-format on
 
 using ReferenceGemmInstance = ck::tensor_operation::host::
-    ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     if(argc == 4)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        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: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         exit(0);
     }
 
-    int group_count = 4;
+    int group_count = rand() % 16 + 1;
 
     // GEMM shape
     std::vector<ck::tensor_operation::device::GemmShape> gemm_shapes;
@@ -131,7 +131,7 @@ int main(int argc, char* argv[])
 
     std::size_t flop = 0, num_btype = 0;
 
-    for(int i = 0; i < gemm_shapes.size(); i++)
+    for(std::size_t i = 0; i < gemm_shapes.size(); i++)
     {
         a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
             gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
@@ -168,7 +168,7 @@ int main(int argc, char* argv[])
         }
     }
 
-    for(int i = 0; i < gemm_shapes.size(); i++)
+    for(std::size_t i = 0; i < gemm_shapes.size(); i++)
     {
         a_tensors_device.emplace_back(
             std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpace()));
@@ -189,12 +189,17 @@ int main(int argc, char* argv[])
     auto b_element_op = BElementOp{};
     auto c_element_op = CElementOp{};
 
-    // do GEMM
     auto gemm    = DeviceGemmInstance{};
     auto invoker = gemm.MakeInvoker();
+
+    // do GEMM
     auto argument =
         gemm.MakeArgument(p_a, p_b, p_c, gemm_shapes, a_element_op, b_element_op, c_element_op);
 
+    DeviceMem gemm_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
+
     if(!gemm.IsSupportedArgument(argument))
     {
         throw std::runtime_error(
@@ -202,7 +207,7 @@ int main(int argc, char* argv[])
             "not support this GEMM problem");
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
 
@@ -211,9 +216,10 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
               << gemm.GetTypeString() << std::endl;
 
+    bool pass = true;
     if(do_verification)
     {
-        for(int i = 0; i < gemm_shapes.size(); i++)
+        for(std::size_t i = 0; i < gemm_shapes.size(); i++)
         {
             c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
             auto ref_gemm    = ReferenceGemmInstance{};
@@ -227,9 +233,9 @@ int main(int argc, char* argv[])
                                                       c_element_op);
 
             ref_invoker.Run(ref_argument);
-            ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
+            pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
         }
     }
 
-    return 0;
+    return pass ? 0 : 1;
 }

example/16_gemm_reduce/CMakeLists.txt

@@ -1 +1,2 @@
-add_example_executable(example_gemm_reduce_xdl_fp16 gemm_reduce_xdl_fp16.cpp)
+add_example_executable(example_gemm_reduce_xdl_max_fp16 gemm_reduce_xdl_max_fp16.cpp)
+add_example_executable(example_gemm_reduce_xdl_mean_squaremean_fp16 gemm_reduce_xdl_mean_squaremean_fp16.cpp)

example/16_gemm_reduce/gemm_reduce_xdl_fp16.cpp

@@ -1,273 +0,0 @@
-#include <iostream>
-#include <numeric>
-#include <initializer_list>
-#include <cstdlib>
-#include <stdlib.h>
-#include <half.hpp>
-#include "config.hpp"
-#include "device.hpp"
-#include "host_tensor.hpp"
-#include "host_tensor_generator.hpp"
-#include "device_tensor.hpp"
-#include "device_gemm_reduce_xdl_cshuffle.hpp"
-#include "element_wise_operation.hpp"
-#include "reduction_operator.hpp"
-#include "reference_gemm.hpp"
-#include "gemm_specialization.hpp"
-#include "reduction_operator.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 ADataType = F16;
-using BDataType = F16;
-using CDataType = F16;
-using DDataType = F32;
-
-using ALayout = ck::tensor_layout::gemm::RowMajor;
-using BLayout = ck::tensor_layout::gemm::ColumnMajor;
-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;
-using D0ReduceOp  = ck::reduce::Add<float>;
-using D1ReduceOp  = ck::reduce::Add<float>;
-using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
-
-static constexpr auto GemmSpecialization =
-    ck::tensor_operation::device::GemmSpecialization::Default;
-
-// clang-format off
-using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
-//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc| DData|           A|           B|           C|         D0|         D1|    D1EleOp|    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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
-//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|  Type| Elementwise| Elementwise| Elementwise|     Reduce|     Reduce|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
-//######|        |        |        |     |      |      |         |         |          |      |   Operation|   Operation|   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_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
-//######|        |        |        |     |      |      |         |         |          |      |            |            |            |           |           |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
-        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   F32,  AElementOp,  BElementOp,  CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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,             S<64, 4>,                         4,                            1>;
-// clang-format on
-
-using ReferenceGemmInstance = ck::tensor_operation::host::
-    ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
-
-int main(int argc, char* argv[])
-{
-    bool do_verification = 1;
-    int init_method      = 1;
-    int nrepeat          = 5;
-
-    // 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;
-
-    if(argc == 1)
-    {
-        // do nothing
-    }
-    else if(argc == 4)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
-    }
-    else if(argc == 10)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        nrepeat         = 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]);
-    }
-    else
-    {
-        printf("arg1: verification (0=no, 1=yes)\n");
-        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\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<DDataType> d0_m_host_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-    Tensor<DDataType> d1_m_host_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-
-    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-    Tensor<DDataType> d0_m_device_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-    Tensor<DDataType> d1_m_device_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-
-    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;
-    std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
-    std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
-
-    switch(init_method)
-    {
-    case 0: break;
-    case 1:
-        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-        break;
-    default:
-        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-        break;
-    }
-
-    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
-    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
-    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
-    DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
-    DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
-
-    a_device_buf.ToDevice(a_m_k.mData.data());
-    b_device_buf.ToDevice(b_k_n.mData.data());
-
-    auto a_element_op  = AElementOp{};
-    auto b_element_op  = BElementOp{};
-    auto c_element_op  = CElementOp{};
-    auto d1_element_op = D1ElementOp{};
-
-    // do GEMM
-    auto gemm     = DeviceGemmReduceInstance{};
-    auto invoker  = gemm.MakeInvoker();
-    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
-                                      static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
-                                      static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
-                                      static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
-                                      static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
-                                      M,
-                                      N,
-                                      K,
-                                      StrideA,
-                                      StrideB,
-                                      StrideC,
-                                      a_element_op,
-                                      b_element_op,
-                                      c_element_op,
-                                      d1_element_op);
-
-    if(!gemm.IsSupportedArgument(argument))
-    {
-        throw std::runtime_error(
-            "wrong! device_gemm with the specified compilation parameters does "
-            "not support this GEMM problem");
-    }
-
-    // warm up
-    invoker.Run(argument);
-
-    // timing
-    float total_time = 0;
-
-    for(int i = 0; i < nrepeat; ++i)
-    {
-        // init DO, D1 to 0
-        d0_device_buf.SetZero();
-        d1_device_buf.SetZero();
-
-        KernelTimer timer;
-
-        timer.Start();
-
-        invoker.Run(argument);
-
-        timer.End();
-
-        total_time += timer.GetElapsedTime();
-    }
-
-    float ave_time = total_time / nrepeat;
-
-    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;
-
-    if(do_verification)
-    {
-        c_device_buf.FromDevice(c_m_n_device_result.mData.data());
-        d0_device_buf.FromDevice(d0_m_device_result.mData.data());
-        d1_device_buf.FromDevice(d1_m_device_result.mData.data());
-
-        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);
-
-        auto d0_reduce_op = D0ReduceOp{};
-        auto d1_reduce_op = D1ReduceOp{};
-
-        for(int m = 0; m < M; ++m)
-        {
-            float d0_acc = d0_reduce_op.GetReductionZeroVal();
-            float d1_acc = d1_reduce_op.GetReductionZeroVal();
-
-            for(int n = 0; n < N; ++n)
-            {
-                float d0_val = ck::type_convert<float>(c_m_n_host_result(m, n));
-                float d1_val;
-
-                d1_element_op(d1_val, d0_val);
-                d0_reduce_op(d0_acc, d0_val);
-                d1_reduce_op(d1_acc, d1_val);
-            }
-
-            d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
-            d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
-        }
-
-        check_error(c_m_n_host_result, c_m_n_device_result);
-        check_error(d0_m_host_result, d0_m_device_result);
-        check_error(d1_m_host_result, d1_m_device_result);
-    }
-
-    return 0;
-}

example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp

@@ -0,0 +1,268 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+
+#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_reduce_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 F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using ADataType         = F16;
+using BDataType         = F16;
+using CDataType         = F16;
+using GemmAccDataType   = F32;
+using ReduceAccDataType = F32;
+using DDataType         = F64;
+using DPtrsGlobal       = ck::Tuple<DDataType*>;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+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;
+using DsReduceOp  = ck::Tuple<ck::reduce::Max>;
+using DsElementOp = ck::Tuple<ck::tensor_operation::element_wise::PassThrough>;
+using DGlobalMemOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmSpecialization =
+    ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
+//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle|         ReduceAcc|         DData|           A|           B|           C|         Dxs|   DxsInEleOp|  DxsAccEleOp|             D|               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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type| DataType| DataType|          DataType|    Type Tuple| Elementwise| Elementwise| Elementwise|      Reduce|             |             |    MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |         |         |                  |              |   Operation|   Operation|   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_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |         |         |                  |              |            |            |            |            |             |             |              |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32, ReduceAccDataType,   DPtrsGlobal,  AElementOp,  BElementOp,  CElementOp, DsReduceOp,   DsElementOp,  DsElementOp,  DGlobalMemOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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,             S<64, 4>,                         4,                            1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CElementOp>;
+
+template <typename ADataType, typename BDataType, typename CDataType, typename DDataType>
+void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
+{
+    std::size_t gemm_flop     = std::size_t(2) * M * N * K;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(CDataType) * M * N + sizeof(DDataType) * M;
+
+    float tflops          = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
+    float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
+
+    std::cout << "gemm + reduceMax Perf: " << gemm_reduce_time << " ms, " << tflops << " TFlops, "
+              << gemm_gb_per_sec << " GB/s, " << std::endl;
+}
+
+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;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    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 == 10)
+    {
+        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]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: run kernel # of times (>1)\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<DDataType> d_m_host_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> d_m_device_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+
+    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;
+    std::cout << "d_m: " << d_m_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_m_device_result.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op  = AElementOp{};
+    auto b_element_op  = BElementOp{};
+    auto c_element_op  = CElementOp{};
+    auto ds_element_op = DsElementOp{};
+    auto p_ds_global   = ck::make_tuple(static_cast<DDataType*>(d_device_buf.GetDeviceBuffer()));
+
+    // do GEMM
+    auto gemm     = DeviceGemmReduceInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                      p_ds_global,
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_element_op,
+                                      ds_element_op,
+                                      ds_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    // [CAUSION]: launch_and_time_kernel will not initialize D.
+    // If we evaluate kernel multiple time but without initialize D. Verification will fail
+    d_device_buf.SetValue(ck::NumericLimits<DDataType>::Lowest());
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+        d_device_buf.FromDevice(d_m_device_result.mData.data());
+
+        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);
+
+        auto d_reduce_op = DsReduceOp{}[ck::Number<0>{}];
+
+        for(int m = 0; m < M; ++m)
+        {
+            ReduceAccDataType d_acc = d_reduce_op.GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType curr_val =
+                    ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
+                d_reduce_op(d_acc, curr_val);
+            };
+
+            d_m_host_result(m) = d_acc;
+        }
+
+        pass = ck::utils::check_err(c_m_n_device_result.mData,
+                                    c_m_n_host_result.mData,
+                                    "Error: Incorrect results c") &&
+               ck::utils::check_err(d_m_device_result.mData,
+                                    d_m_host_result.mData,
+                                    "Error: Incorrect results d",
+                                    1e-3,
+                                    1e-3);
+    }
+
+    if(time_kernel)
+    {
+        float gemm_reduceMax_ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
+
+        DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, DDataType>(
+            gemm_reduceMax_ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}

example/16_gemm_reduce/gemm_reduce_xdl_mean_squaremean_fp16.cpp

@@ -0,0 +1,305 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+
+#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_reduce_xdl_cshuffle.hpp"
+#include "element_wise_operation.hpp"
+#include "reduction_operator.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+#include "reduction_operator.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 ADataType         = F16;
+using BDataType         = F16;
+using CDataType         = F16;
+using GemmAccDataType   = F32;
+using ReduceAccDataType = F32;
+using DDataType         = F32;
+using DPtrsGlobal       = ck::Tuple<DDataType*, DDataType*>;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+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;
+using D0ReduceOp  = ck::reduce::Add;
+using D1ReduceOp  = ck::reduce::Add;
+using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
+
+using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
+using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+using DxsInElementOps       = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+using DxsOutElementOps      = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
+
+using DGlobalMemOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmSpecialization =
+    ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
+//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|         Dxs|     DxsInEleOp|     DxsAccEleOp|             D|               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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise|      Reduce|               |                |    MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |         |         |          |              |   Operation|   Operation|   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_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |         |         |          |              |            |            |            |            |               |                |              |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   DPtrsGlobal,  AElementOp,  BElementOp,  CElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps,  DGlobalMemOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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,             S<64, 4>,                         4,                            1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CElementOp>;
+
+template <typename ADataType, typename BDataType, typename CDataType, typename DDataType>
+void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
+{
+    std::size_t gemm_flop     = std::size_t(2) * M * N * K;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(CDataType) * M * N + sizeof(DDataType) * M +
+                                sizeof(DDataType) * M;
+
+    float tflops          = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
+    float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
+
+    std::cout << "gemm + reduce_mean + reduce_mean_square Perf: " << gemm_reduce_time << " ms, "
+              << tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
+}
+
+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;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    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 == 10)
+    {
+        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]);
+    }
+    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<DDataType> d0_m_host_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<DDataType> d1_m_host_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> d0_m_device_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<DDataType> d1_m_device_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+
+    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;
+    std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
+    std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
+    DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+    auto dxs_global   = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
+                                     static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
+
+    auto dxs_in_element_op  = DxsInElementOps{};
+    auto dxs_out_element_op = DxsOutElementOps{N, N};
+
+    // do GEMM
+    auto gemm     = DeviceGemmReduceInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                      dxs_global,
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_element_op,
+                                      dxs_in_element_op,
+                                      dxs_out_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    // init DO, D1 to 0
+    d0_device_buf.SetZero();
+    d1_device_buf.SetZero();
+
+    // if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
+    // will not be correct. need to set time_kernel = false for correctness test
+    invoker.Run(argument, StreamConfig{nullptr, false});
+    bool pass = true;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+        d0_device_buf.FromDevice(d0_m_device_result.mData.data());
+        d1_device_buf.FromDevice(d1_m_device_result.mData.data());
+
+        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);
+
+        auto d0_reduce_op = D0ReduceOp{};
+        auto d1_reduce_op = D1ReduceOp{};
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
+            auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                auto c_val = ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
+                ReduceAccDataType d0_val;
+                ReduceAccDataType d1_val;
+
+                dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
+                dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
+                d0_reduce_op(d0_acc, d0_val);
+                d1_reduce_op(d1_acc, d1_val);
+            }
+
+            dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
+            dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
+            d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
+            d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
+        }
+
+        pass = ck::utils::check_err(c_m_n_device_result.mData,
+                                    c_m_n_host_result.mData,
+                                    "Error: Incorrect results c") &&
+               ck::utils::check_err(d0_m_device_result.mData,
+                                    d0_m_host_result.mData,
+                                    "Error: Incorrect results d0",
+                                    1e-4,
+                                    1e-5) &&
+               ck::utils::check_err(d1_m_device_result.mData,
+                                    d1_m_host_result.mData,
+                                    "Error: Incorrect results d1",
+                                    1e-3,
+                                    1e-5);
+    }
+
+    if(time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
+
+        DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, DDataType>(ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}

example/17_convnd_bwd_data_xdl/CMakeLists.txt

@@ -1,2 +1,2 @@
 add_example_executable(example_convnd_bwd_data_xdl convnd_bwd_data_xdl.cpp)
-target_link_libraries(example_convnd_bwd_data_xdl PRIVATE conv_fwd_util)
+target_link_libraries(example_convnd_bwd_data_xdl PRIVATE conv_util)

example/17_convnd_bwd_data_xdl/convnd_bwd_data_xdl.cpp

@@ -6,7 +6,7 @@
 #include <half.hpp>
 
 #include "config.hpp"
-#include "conv_fwd_util.hpp"
+#include "conv_util.hpp"
 #include "print.hpp"
 #include "device.hpp"
 #include "host_tensor.hpp"
@@ -87,7 +87,7 @@ void print_use_msg()
 {
     std::cout << "arg1: verification (0=no, 1=yes)\n"
               << "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
-              << "arg3: run kernel # of times (>1)\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
               << "arg4: N spatial dimensions (default 2)\n"
               << "Following arguments (depending on number of spatial dims):\n"
               << " N, K, C, \n"
@@ -105,40 +105,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
     ck::utils::conv::ConvParams params;
     int arg_idx = 5;
 
-    params.num_dim_spatial = num_dim_spatial;
-    params.N               = std::stoi(argv[arg_idx++]);
-    params.K               = std::stoi(argv[arg_idx++]);
-    params.C               = std::stoi(argv[arg_idx++]);
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
 
-    params.filter_spatial_lengths.resize(num_dim_spatial);
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_spatial_lengths.resize(num_dim_spatial);
+    params.input_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_strides.resize(num_dim_spatial);
+    params.conv_filter_strides_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_dilations.resize(num_dim_spatial);
+    params.conv_filter_dilations_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_left_pads.resize(num_dim_spatial);
+    params.input_left_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_right_pads.resize(num_dim_spatial);
+    params.input_right_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
     }
 
     return params;
@@ -165,25 +165,25 @@ DeviceConvBwdDataBasePtr get_conv_instance(int num_dim_spatial)
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
     int num_dim_spatial  = 2;
 
     ck::utils::conv::ConvParams params;
-    params.C = 128;
+    params.C_ = 128;
 
     if(argc == 4)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc > 4)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
         num_dim_spatial = std::stoi(argv[4]);
         // check args number
         int conv_args     = 3 + num_dim_spatial * 6;
@@ -202,21 +202,21 @@ int main(int argc, char* argv[])
         exit(1);
     }
 
-    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
-                                        static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
     input_dims.insert(std::end(input_dims),
-                      std::begin(params.input_spatial_lengths),
-                      std::end(params.input_spatial_lengths));
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
 
-    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
-                                         static_cast<std::size_t>(params.C)};
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
     filter_dims.insert(std::end(filter_dims),
-                       std::begin(params.filter_spatial_lengths),
-                       std::end(params.filter_spatial_lengths));
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
 
     const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
-    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
-                                         static_cast<std::size_t>(params.K)};
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
     output_dims.insert(std::end(output_dims),
                        std::begin(output_spatial_lengths),
                        std::end(output_spatial_lengths));
@@ -263,16 +263,16 @@ int main(int argc, char* argv[])
         conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                                   static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                                   static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-                                  params.N,
-                                  params.K,
-                                  params.C,
-                                  params.input_spatial_lengths,
-                                  params.filter_spatial_lengths,
+                                  params.N_,
+                                  params.K_,
+                                  params.C_,
+                                  params.input_spatial_lengths_,
+                                  params.filter_spatial_lengths_,
                                   output_spatial_lengths,
-                                  params.conv_filter_strides,
-                                  params.conv_filter_dilations,
-                                  params.input_left_pads,
-                                  params.input_right_pads,
+                                  params.conv_filter_strides_,
+                                  params.conv_filter_dilations_,
+                                  params.input_left_pads_,
+                                  params.input_right_pads_,
                                   InElementOp{},
                                   WeiElementOp{},
                                   OutElementOp{});
@@ -284,16 +284,16 @@ int main(int argc, char* argv[])
             "not support this Conv problem");
     }
 
-    float ave_time = invoker->Run(argument.get(), nrepeat);
+    float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = ck::utils::conv::get_flops(
-        params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
     std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
-        params.N,
-        params.C,
-        params.K,
-        params.input_spatial_lengths,
-        params.filter_spatial_lengths,
+        params.N_,
+        params.C_,
+        params.K_,
+        params.input_spatial_lengths_,
+        params.filter_spatial_lengths_,
         output_spatial_lengths);
 
     float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -310,10 +310,10 @@ int main(int argc, char* argv[])
             auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result,
                                                       wei_k_c_y_x,
                                                       out_n_k_ho_wo,
-                                                      params.conv_filter_strides,
-                                                      params.conv_filter_dilations,
-                                                      params.input_left_pads,
-                                                      params.input_right_pads,
+                                                      params.conv_filter_strides_,
+                                                      params.conv_filter_dilations_,
+                                                      params.input_left_pads_,
+                                                      params.input_right_pads_,
                                                       InElementOp{},
                                                       WeiElementOp{},
                                                       OutElementOp{});
@@ -322,29 +322,30 @@ int main(int argc, char* argv[])
 
             in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
 
-            check_error(in_n_c_hi_wi_host_result, in_n_c_hi_wi_device_result);
+            return ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
+                                        in_n_c_hi_wi_host_result.mData)
+                       ? 0
+                       : 1;
         };
 
         switch(num_dim_spatial)
         {
         case 3: {
             auto ref_conv = ReferenceConvBwdDataInstance<3>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 2: {
             auto ref_conv = ReferenceConvBwdDataInstance<2>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 1: {
             auto ref_conv = ReferenceConvBwdDataInstance<1>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         default: {
             throw std::runtime_error("Unsupported number of spatial dimensions provided!");
         }
         }
     }
+    return 0;
 }

example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp

@@ -4,6 +4,7 @@
 #include <cstdlib>
 #include <stdlib.h>
 #include <half.hpp>
+#include "check_err.hpp"
 #include "config.hpp"
 #include "device.hpp"
 #include "host_tensor.hpp"
@@ -24,10 +25,12 @@ using F32 = float;
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
 
-using ADataType = F16;
-using BDataType = F16;
-using CDataType = F16;
-using DDataType = F32;
+using ADataType         = F16;
+using BDataType         = F16;
+using CDataType         = F16;
+using ReduceAccDataType = F32;
+using DDataType         = F32;
+using DPtrsGlobal       = ck::Tuple<DDataType*, DDataType*>;
 
 using ALayout = ck::tensor_layout::gemm::RowMajor;
 using BLayout = ck::tensor_layout::gemm::ColumnMajor;
@@ -36,20 +39,29 @@ 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;
-using D0ReduceOp  = ck::reduce::Add<float>;
-using D1ReduceOp  = ck::reduce::Add<float>;
-using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
+using D0ReduceOp  = ck::reduce::Add;
+using D1ReduceOp  = ck::reduce::Add;
+using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
+
+using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+using DxsInElementOps       = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+using DxsOutElementOps      = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
+
+using DGlobalMemOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
 
 static constexpr auto GemmSpecialization =
     ck::tensor_operation::device::GemmSpecialization::Default;
 
 // clang-format off
 using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatchedGemmReduce_Xdl_CShuffle
-//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc| DData|           A|           B|           C|         D0|         D1|     D1EleOp|      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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
-//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|  Type| Elementwise| Elementwise| Elementwise|     Reduce|     Reduce|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
-//######|        |        |        |     |      |      |         |         |          |      |   Operation|   Operation|   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_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
-//######|        |        |        |     |      |      |         |         |          |      |            |            |            |           |           |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
-        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   F32,  AElementOp,  BElementOp,  CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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,             S<64, 4>,                         4,                            1>;
+//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|         Dxs|     DxsInEleOp|     DxsAccEleOp|            D|               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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise|      Reduce|               |                |   MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |         |         |          |              |   Operation|   Operation|   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_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |         |         |          |              |            |            |            |            |               |                |             |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   DPtrsGlobal,  AElementOp,  BElementOp,  CElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps, DGlobalMemOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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,             S<64, 4>,                         4,                            1>;
 // clang-format on
 
 using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
@@ -57,18 +69,18 @@ using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 1;
+    bool do_verification = true;
     int init_method      = 1;
-    int nrepeat          = 5;
+    bool time_kernel     = false;
 
     // GEMM shape
-    ck::index_t M = 3840;
-    ck::index_t N = 4096;
-    ck::index_t K = 4096;
+    ck::index_t M = 2048;
+    ck::index_t N = 1920;
+    ck::index_t K = 2048;
 
-    ck::index_t StrideA = 4096;
-    ck::index_t StrideB = 4096;
-    ck::index_t StrideC = 4096;
+    ck::index_t StrideA = 2048;
+    ck::index_t StrideB = 2048;
+    ck::index_t StrideC = 1920;
 
     ck::index_t BatchCount = 4;
 
@@ -80,13 +92,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 11)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -96,13 +108,13 @@ int main(int argc, char* argv[])
         StrideB = std::stoi(argv[8]);
         StrideC = std::stoi(argv[9]);
 
-        BatchCount = std::stoi(argv[9]);
+        BatchCount = 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: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, BatchCount\n");
         exit(0);
     }
@@ -169,12 +181,11 @@ int main(int argc, char* argv[])
     a_device_buf.ToDevice(a_g_m_k.mData.data());
     b_device_buf.ToDevice(b_g_k_n.mData.data());
 
-    auto a_element_op  = AElementOp{};
-    auto b_element_op  = BElementOp{};
-    auto c_element_op  = CElementOp{};
-    auto d0_reduce_op  = D0ReduceOp{};
-    auto d1_reduce_op  = D1ReduceOp{};
-    auto d1_element_op = D1ElementOp{};
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+    auto dxs_global   = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
+                                     static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
 
     // do GEMM
     auto batched_gemm = DeviceBatchedGemmReduceInstance{};
@@ -183,8 +194,7 @@ int main(int argc, char* argv[])
         batched_gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
                                   static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
                                   static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
-                                  static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
-                                  static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
+                                  dxs_global,
                                   M,
                                   N,
                                   K,
@@ -194,7 +204,8 @@ int main(int argc, char* argv[])
                                   a_element_op,
                                   b_element_op,
                                   c_element_op,
-                                  d1_element_op,
+                                  DxsInElementOps{},
+                                  DxsOutElementOps{},
                                   BatchCount);
 
     if(!batched_gemm.IsSupportedArgument(argument))
@@ -204,30 +215,13 @@ int main(int argc, char* argv[])
             "not support this GEMM problem");
     }
 
-    // warm up
-    invoker.Run(argument);
+    // init DO, D1 to 0
+    d0_device_buf.SetZero();
+    d1_device_buf.SetZero();
 
-    // timing
-    float total_time = 0;
-
-    for(int i = 0; i < nrepeat; ++i)
-    {
-        // init DO, D1 to 0
-        d0_device_buf.SetZero();
-        d1_device_buf.SetZero();
-
-        KernelTimer timer;
-
-        timer.Start();
-
-        invoker.Run(argument);
-
-        timer.End();
-
-        total_time += timer.GetElapsedTime();
-    }
-
-    float ave_time = total_time / nrepeat;
+    // if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
+    // will not be correct. need to set time_kernel = false for correctness test
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     std::size_t flop      = std::size_t(2) * BatchCount * M * N * K;
     std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
@@ -241,6 +235,7 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
               << batched_gemm.GetTypeString() << std::endl;
 
+    bool pass = true;
     if(do_verification)
     {
         c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
@@ -255,19 +250,25 @@ int main(int argc, char* argv[])
 
         ref_invoker.Run(ref_argument);
 
+        auto d0_reduce_op = D0ReduceOp{};
+        auto d1_reduce_op = D1ReduceOp{};
+
         for(int batch = 0; batch < BatchCount; ++batch)
         {
             for(int m = 0; m < M; ++m)
             {
-                float d0_acc = d0_reduce_op.GetReductionZeroVal();
-                float d1_acc = d1_reduce_op.GetReductionZeroVal();
+                auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
+                auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
 
                 for(int n = 0; n < N; ++n)
                 {
-                    float d0_val = ck::type_convert<float>(c_g_m_n_host_result(m, n));
-                    float d1_val;
+                    auto c_val =
+                        ck::type_convert<ReduceAccDataType>(c_g_m_n_host_result(batch, m, n));
+                    ReduceAccDataType d0_val;
+                    ReduceAccDataType d1_val;
 
-                    d1_element_op(d1_val, d0_val);
+                    UnaryIdenticElementOp{}(d0_val, c_val);
+                    UnarySquareElementOp{}(d1_val, c_val);
                     d0_reduce_op(d0_acc, d0_val);
                     d1_reduce_op(d1_acc, d1_val);
                 }
@@ -277,10 +278,20 @@ int main(int argc, char* argv[])
             }
         }
 
-        check_error(c_g_m_n_host_result, c_g_m_n_device_result);
-        check_error(d0_g_m_host_result, d0_g_m_device_result);
-        check_error(d1_g_m_host_result, d1_g_m_device_result);
+        pass = ck::utils::check_err(c_g_m_n_host_result.mData,
+                                    c_g_m_n_device_result.mData,
+                                    "Error: Incorrect results c") &&
+               ck::utils::check_err(d0_g_m_device_result.mData,
+                                    d0_g_m_host_result.mData,
+                                    "Error: Incorrect results! D0",
+                                    1e-4,
+                                    1e-5) &&
+               ck::utils::check_err(d1_g_m_device_result.mData,
+                                    d1_g_m_host_result.mData,
+                                    "Error: Incorrect results! D1",
+                                    1e-3,
+                                    1e-5);
     }
 
-    return 0;
+    return pass ? 0 : 1;
 }

example/19_binary_elementwise/CMakeLists.txt

@@ -0,0 +1,4 @@
+add_example_executable(example_broadcast_add_2d_amn_bn broadcast_add_2d_amn_bn.cpp)
+add_example_executable(example_broadcast_add_3d_am_bmnk broadcast_add_3d_am_bmnk.cpp)
+add_example_executable(example_elementwise_add_1d elementwise_add_1d.cpp)
+add_example_executable(example_elementwise_add_4d elementwise_add_4d.cpp)

example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp

@@ -0,0 +1,164 @@
+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
+                                                          ABDataType,
+                                                          CDataType,
+                                                          EltwiseComputeDataType,
+                                                          Add,
+                                                          2,
+                                                          8,
+                                                          8,
+                                                          8,
+                                                          8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor,
+          int broadcastDim>
+void host_broadcast2D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, int N, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            ComputeDataType Amn = ck::type_convert<ComputeDataType>(A(m, n));
+            ComputeDataType Cmn = 0;
+            if constexpr(broadcastDim == 0)
+            {
+                ComputeDataType Bn = ck::type_convert<ComputeDataType>(B(n));
+                functor(Cmn, Amn, Bn);
+            }
+            else
+            {
+                ComputeDataType Bm = ck::type_convert<ComputeDataType>(B(m));
+                functor(Cmn, Amn, Bm);
+            }
+            C(m, n) = ck::type_convert<ctype>(Cmn);
+        }
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = 1024;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                    std::vector<std::size_t>({stride, 1}));
+    };
+
+    Tensor<ABDataType> a_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<ABDataType> b_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+    a_m_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_n_device_buf(sizeof(ABDataType) * a_m_n.mDesc.GetElementSpace());
+    DeviceMem b_n_device_buf(sizeof(ABDataType) * b_n.mDesc.GetElementSpace());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpace());
+
+    a_m_n_device_buf.ToDevice(a_m_n.mData.data());
+    b_n_device_buf.ToDevice(b_n.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(a_m_n_device_buf.GetDeviceBuffer(),
+                                                     b_n_device_buf.GetDeviceBuffer(),
+                                                     c_m_n_device_buf.GetDeviceBuffer(),
+                                                     {M, N},
+                                                     {Stride, 1},
+                                                     {0, 1}, // broadcast in first dimension
+                                                     {Stride, 1},
+                                                     Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_n_device_buf.FromDevice(c_m_n.mData.data());
+        Tensor<CDataType> host_c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+        host_broadcast2D<Tensor<ABDataType>,
+                         Tensor<ABDataType>,
+                         Tensor<CDataType>,
+                         EltwiseComputeDataType,
+                         Add,
+                         0>(host_c_m_n, a_m_n, b_n, M, N, Add{});
+
+        pass &= ck::utils::check_err(
+            c_m_n.mData, host_c_m_n.mData, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp

@@ -0,0 +1,123 @@
+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
+                                                          ABDataType,
+                                                          CDataType,
+                                                          EltwiseComputeDataType,
+                                                          Add,
+                                                          3,
+                                                          8,
+                                                          1,
+                                                          8,
+                                                          8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor>
+void host_broadcast3D_am_bmnk(HostTensorC& C,
+                              const HostTensorA& A,
+                              const HostTensorB& B,
+                              const std::vector<std::size_t>& shape,
+                              Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(std::size_t m = 0; m < shape[0]; ++m)
+        for(std::size_t n = 0; n < shape[1]; ++n)
+            for(std::size_t k = 0; k < shape[2]; ++k)
+            {
+                ComputeDataType a_val = ck::type_convert<ComputeDataType>(A(m));
+                ComputeDataType b_val = ck::type_convert<ComputeDataType>(B(m, n, k));
+                ComputeDataType c_val = 0;
+                functor(c_val, a_val, b_val);
+                C(m, n, k) = ck::type_convert<ctype>(c_val);
+            }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    std::vector<std::size_t> mnk = {4, 16, 32};
+    ck::index_t M                = mnk[0];
+
+    Tensor<ABDataType> a_m({M});
+    Tensor<ABDataType> b_m_n_k(mnk);
+    Tensor<CDataType> c_m_n_k(mnk);
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m_n_k.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpace());
+    DeviceMem b_m_n_k_device_buf(sizeof(ABDataType) * b_m_n_k.mDesc.GetElementSpace());
+    DeviceMem c_m_n_k_device_buf(sizeof(CDataType) * c_m_n_k.mDesc.GetElementSpace());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_n_k_device_buf.ToDevice(b_m_n_k.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        a_m_device_buf.GetDeviceBuffer(),
+        b_m_n_k_device_buf.GetDeviceBuffer(),
+        c_m_n_k_device_buf.GetDeviceBuffer(),
+        std::vector<ck::index_t>{mnk.begin(), mnk.end()},
+        {1, 0, 0}, // broadcast A on second and third dimension
+        std::vector<ck::index_t>{b_m_n_k.mDesc.GetStrides().begin(),
+                                 b_m_n_k.mDesc.GetStrides().end()},
+        std::vector<ck::index_t>{c_m_n_k.mDesc.GetStrides().begin(),
+                                 c_m_n_k.mDesc.GetStrides().end()},
+        Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_n_k_device_buf.FromDevice(c_m_n_k.mData.data());
+        Tensor<CDataType> host_c_m_n_k(mnk);
+
+        host_broadcast3D_am_bmnk<Tensor<ABDataType>,
+                                 Tensor<ABDataType>,
+                                 Tensor<CDataType>,
+                                 EltwiseComputeDataType,
+                                 Add>(host_c_m_n_k, a_m, b_m_n_k, mnk, Add{});
+
+        pass &= ck::utils::check_err(
+            c_m_n_k.mData, host_c_m_n_k.mData, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/19_binary_elementwise/elementwise_add_1d.cpp

@@ -0,0 +1,144 @@
+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
+                                                          ABDataType,
+                                                          CDataType,
+                                                          EltwiseComputeDataType,
+                                                          Add,
+                                                          1,
+                                                          8,
+                                                          8,
+                                                          8,
+                                                          8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor>
+void host_elementwise1D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        ComputeDataType Am = ck::type_convert<ComputeDataType>(A(m));
+        ComputeDataType Bm = ck::type_convert<ComputeDataType>(B(m));
+        ComputeDataType Cm = 0;
+        functor(Cm, Am, Bm);
+        C(m) = ck::type_convert<ctype>(Cm);
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M = 1024;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    Tensor<ABDataType> a_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<ABDataType> b_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<CDataType> c_m(f_host_tensor_descriptor1d(M, 1));
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpace());
+    DeviceMem b_m_device_buf(sizeof(ABDataType) * b_m.mDesc.GetElementSpace());
+    DeviceMem c_m_device_buf(sizeof(CDataType) * c_m.mDesc.GetElementSpace());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_device_buf.ToDevice(b_m.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(a_m_device_buf.GetDeviceBuffer(),
+                                                     b_m_device_buf.GetDeviceBuffer(),
+                                                     c_m_device_buf.GetDeviceBuffer(),
+                                                     {M},
+                                                     {1},
+                                                     {1},
+                                                     {1},
+                                                     Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_device_buf.FromDevice(c_m.mData.data());
+        Tensor<CDataType> host_c_m(f_host_tensor_descriptor1d(M, 1));
+
+        host_elementwise1D<Tensor<ABDataType>,
+                           Tensor<ABDataType>,
+                           Tensor<CDataType>,
+                           EltwiseComputeDataType,
+                           Add>(host_c_m, a_m, b_m, M, Add{});
+
+        pass &= ck::utils::check_err(
+            c_m.mData, host_c_m.mData, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/19_binary_elementwise/elementwise_add_4d.cpp

@@ -0,0 +1,146 @@
+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
+                                                          ABDataType,
+                                                          CDataType,
+                                                          EltwiseComputeDataType,
+                                                          Add,
+                                                          4,
+                                                          8,
+                                                          8,
+                                                          8,
+                                                          8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor>
+void host_elementwise4D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0, 0, 0))>;
+
+    for(std::size_t n = 0; n < shape[0]; ++n)
+        for(std::size_t c = 0; c < shape[1]; ++c)
+            for(std::size_t h = 0; h < shape[2]; ++h)
+                for(std::size_t w = 0; w < shape[3]; ++w)
+                {
+                    ComputeDataType a_val = ck::type_convert<ComputeDataType>(A(n, c, h, w));
+                    ComputeDataType b_val = ck::type_convert<ComputeDataType>(B(n, c, h, w));
+                    ComputeDataType c_val = 0;
+                    functor(c_val, a_val, b_val);
+                    C(n, c, h, w) = ck::type_convert<ctype>(c_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    std::vector<std::size_t> nchw = {4, 16, 32, 32};
+
+    Tensor<ABDataType> a(nchw);
+    Tensor<ABDataType> b(nchw);
+    Tensor<CDataType> c(nchw);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ABDataType) * a.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(ABDataType) * b.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a.mData.data());
+    b_device_buf.ToDevice(b.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        a_device_buf.GetDeviceBuffer(),
+        b_device_buf.GetDeviceBuffer(),
+        c_device_buf.GetDeviceBuffer(),
+        std::vector<ck::index_t>{nchw.begin(), nchw.end()},
+        std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
+        std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
+        std::vector<ck::index_t>{c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end()},
+        Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c.mData.data());
+        Tensor<CDataType> host_c(nchw);
+
+        host_elementwise4D<Tensor<ABDataType>,
+                           Tensor<ABDataType>,
+                           Tensor<CDataType>,
+                           EltwiseComputeDataType,
+                           Add>(host_c, a, b, nchw, Add{});
+
+        pass &=
+            ck::utils::check_err(c.mData, host_c.mData, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/20_convnd_bwd_weight_xdl/CMakeLists.txt

@@ -0,0 +1,4 @@
+add_example_executable(example_convnd_bwd_weight_xdl convnd_bwd_weight_xdl.cpp)
+add_example_executable(example_convnd_bwd_weight_xdl_bf16_splitk convnd_bwd_weight_xdl_bf16_splitk.cpp)
+target_link_libraries(example_convnd_bwd_weight_xdl PRIVATE conv_util)
+target_link_libraries(example_convnd_bwd_weight_xdl_bf16_splitk PRIVATE conv_util)

example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl.cpp

@@ -0,0 +1,385 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+
+#include "check_err.hpp"
+#include "conv_util.hpp"
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "tensor_layout.hpp"
+#include "element_wise_operation.hpp"
+#include "device_convnd_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
+#include "reference_conv_backward_weight.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+using DeviceConvBwdWeightBasePtr =
+    ck::tensor_operation::device::DeviceConvBwdWeightPtr<InElementOp, WeiElementOp, OutElementOp>;
+
+// clang-format off
+template <ck::index_t NumDimSpatial>
+using DeviceConvndBwdWeightInstance = ck::tensor_operation::device::
+    DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
+        InDataType,                       // InDataType
+        WeiDataType,                      // WeiDataType
+        OutDataType,                      // OutDataType
+        AccDataType,                      // AccDataType
+        InElementOp,                      // InElementwiseOperation
+        WeiElementOp,                     // WeiElementwiseOperation
+        OutElementOp,                     // OutElementwiseOperation
+        ConvBwdWeightDefault,             // ConvolutionBackwardWeightSpecialization
+        NumDimSpatial,                    // NumDimSpatial
+        256,                              // BlockSize
+        128,                              // MPerBlock
+        128,                              // NPerBlock
+        4,                                // K0PerBlock
+        8,                                // K1
+        32,                               // MPerXdl
+        32,                               // NPerXdl
+        2,                                // MXdlPerWave
+        2,                                // NXdlPerWave
+        S<1, 4, 16, 4>,                   // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<0, 3, 1, 2>,                    // ABlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,                    // ABlockTransferSrcAccessOrder
+        2,                                // ABlockTransferSrcVectorDim
+        8,                                // ABlockTransferSrcScalarPerVector
+        2,                                // ABlockTransferDstScalarPerVector_K1
+        true,                             // ABlockLdsAddExtraM
+        S<1, 4, 16, 4>,                   // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<0, 3, 1, 2>,                    // BBlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,                    // BBlockTransferSrcAccessOrder
+        2,                                // BBlockTransferSrcVectorDim
+        8,                                // BBlockTransferSrcScalarPerVector
+        2,                                // BBlockTransferDstScalarPerVector_K1
+        true,                             // BBlockLdsAddExtraN
+        1,                                // CShuffleMXdlPerWavePerShuffle
+        1,                                // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 4>,                   // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;                               // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+template <ck::index_t NumDimSpatial>
+using ReferenceConvBwdWeightInstance =
+    ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
+                                                       WeiDataType,
+                                                       OutDataType,
+                                                       InElementOp,
+                                                       WeiElementOp,
+                                                       OutElementOp,
+                                                       NumDimSpatial>;
+
+void print_use_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
+              << "arg4: is show log (0=no, 1=yes)\n"
+              << "arg5: split-k \n"
+              << "arg6: N spatial dimensions (default 2)\n"
+              << "Following arguments (depending on number of spatial dims):\n"
+              << " N, K, C, \n"
+              << " <filter spatial dimensions>, (ie Y, X for 2D)\n"
+              << " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
+              << " <strides>, (ie Sy, Sx for 2D)\n"
+              << " <dilations>, (ie Dy, Dx for 2D)\n"
+              << " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
+              << " <right padding>, (ie RightPy, RightPx for 2D)\n"
+              << std::endl;
+}
+
+ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
+{
+    // (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
+    ck::utils::conv::ConvParams params;
+    int arg_idx = 7;
+
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
+
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_spatial_lengths_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.conv_filter_strides_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.conv_filter_dilations_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_left_pads_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_right_pads_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    return params;
+}
+
+DeviceConvBwdWeightBasePtr get_conv_instance(int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return std::make_unique<DeviceConvndBwdWeightInstance<3>>();
+    }
+    case 2: {
+        return std::make_unique<DeviceConvndBwdWeightInstance<2>>();
+    }
+    case 1: {
+        return std::make_unique<DeviceConvndBwdWeightInstance<1>>();
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int num_dim_spatial  = 2;
+    int do_log           = 0;
+    int split_k          = 1;
+
+    ck::utils::conv::ConvParams params;
+    params.C_ = 128;
+
+    if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        do_log          = std::stoi(argv[4]);
+        split_k         = std::stoi(argv[5]);
+    }
+    else if(argc > 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        do_log          = std::stoi(argv[4]);
+        split_k         = std::stoi(argv[5]);
+        num_dim_spatial = std::stoi(argv[6]);
+        // check args number
+        int conv_args     = 3 + num_dim_spatial * 6;
+        int cmdline_nargs = conv_args + 7;
+        if(cmdline_nargs != argc)
+        {
+            print_use_msg();
+            exit(1);
+        }
+
+        params = parse_conv_params(num_dim_spatial, argv);
+    }
+    else if(argc != 1)
+    {
+        print_use_msg();
+        exit(1);
+    }
+
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
+    input_dims.insert(std::end(input_dims),
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
+
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
+    filter_dims.insert(std::end(filter_dims),
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
+
+    const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
+    output_dims.insert(std::end(output_dims),
+                       std::begin(output_spatial_lengths),
+                       std::end(output_spatial_lengths));
+
+    Tensor<InDataType> in_n_c_hi_wi(
+        ck::utils::conv::get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
+    Tensor<WeiDataType> wei_k_c_y_x_host_result(
+        ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
+    Tensor<WeiDataType> wei_k_c_y_x_device_result(
+        ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
+    Tensor<OutDataType> out_n_k_ho_wo(
+        ck::utils::conv::get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_device_result.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
+        break;
+    default:
+        out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) *
+                             wei_k_c_y_x_device_result.mDesc.GetElementSpace());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
+    // reset input to zero
+    wei_device_buf.SetZero();
+
+    // do GEMM
+    auto conv    = get_conv_instance(num_dim_spatial);
+    auto invoker = conv->MakeInvokerPointer();
+    auto argument =
+        conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                  static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                  static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                  params.N_,
+                                  params.K_,
+                                  params.C_,
+                                  params.input_spatial_lengths_,
+                                  params.filter_spatial_lengths_,
+                                  output_spatial_lengths,
+                                  params.conv_filter_strides_,
+                                  params.conv_filter_dilations_,
+                                  params.input_left_pads_,
+                                  params.input_right_pads_,
+                                  InElementOp{},
+                                  WeiElementOp{},
+                                  OutElementOp{},
+                                  split_k);
+
+    // alloc work space
+    float ave_time = 0.f;
+    if(!conv->IsSupportedArgument(argument.get()))
+    {
+        std::cout << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+        return 1;
+    }
+    ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = ck::utils::conv::get_flops(
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
+    std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
+        params.N_,
+        params.C_,
+        params.K_,
+        params.input_spatial_lengths_,
+        params.filter_spatial_lengths_,
+        output_spatial_lengths);
+
+    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)
+    {
+        auto verify_f = [&](const auto& ref_conv) {
+            auto ref_invoker = ref_conv.MakeInvoker();
+
+            auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
+                                                      wei_k_c_y_x_host_result,
+                                                      out_n_k_ho_wo,
+                                                      params.conv_filter_strides_,
+                                                      params.conv_filter_dilations_,
+                                                      params.input_left_pads_,
+                                                      params.input_right_pads_,
+                                                      InElementOp{},
+                                                      WeiElementOp{},
+                                                      OutElementOp{});
+
+            ref_invoker.Run(ref_argument);
+
+            wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",") << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "wei_device(after): ", wei_k_c_y_x_device_result.mData, ",")
+                    << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "wei_host  : ", wei_k_c_y_x_host_result.mData, ",")
+                    << std::endl;
+            }
+
+            return ck::utils::check_err(wei_k_c_y_x_device_result.mData,
+                                        wei_k_c_y_x_host_result.mData)
+                       ? 0
+                       : 1;
+        };
+
+        switch(num_dim_spatial)
+        {
+        case 3: {
+            auto ref_conv = ReferenceConvBwdWeightInstance<3>();
+            return verify_f(ref_conv);
+        }
+        case 2: {
+            auto ref_conv = ReferenceConvBwdWeightInstance<2>();
+            return verify_f(ref_conv);
+        }
+        case 1: {
+            auto ref_conv = ReferenceConvBwdWeightInstance<1>();
+            return verify_f(ref_conv);
+        }
+        default: {
+            throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+        }
+        }
+    }
+    return 0;
+}

example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl_bf16_splitk.cpp

@@ -0,0 +1,427 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+
+#include "check_err.hpp"
+#include "conv_util.hpp"
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "tensor_layout.hpp"
+#include "element_wise_operation.hpp"
+#include "device_unary_elementwise.hpp"
+#include "device_convnd_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
+#include "reference_conv_backward_weight.hpp"
+
+using InDataType  = ck::bhalf_t;
+using WeiDataType = ck::bhalf_t;
+using OutDataType = ck::bhalf_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+using UnaryTypeConvert = ck::tensor_operation::element_wise::UnaryTypeConvert<ck::bhalf_t, float>;
+
+using DeviceUnaryElementwiseTypeConvertInstance = ck::tensor_operation::device::
+    DeviceUnaryElementwise<AccDataType, WeiDataType, UnaryTypeConvert, 1, 4>;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+using DeviceConvBwdWeightBasePtr =
+    ck::tensor_operation::device::DeviceConvBwdWeightPtr<InElementOp, WeiElementOp, OutElementOp>;
+
+// clang-format off
+template <ck::index_t NumDimSpatial>
+using DeviceConvndBwdWeightInstance_bf16_splitk = ck::tensor_operation::device::
+    DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
+        InDataType,                       // InDataType
+        AccDataType,                      // WeiDataType
+        OutDataType,                      // OutDataType
+        AccDataType,                      // AccDataType
+        InElementOp,                      // InElementwiseOperation
+        WeiElementOp,                     // WeiElementwiseOperation
+        OutElementOp,                     // OutElementwiseOperation
+        ConvBwdWeightDefault,             // ConvolutionBackwardWeightSpecialization
+        NumDimSpatial,                    // NumDimSpatial
+        256,                              // BlockSize
+        128,                              // MPerBlock
+        128,                              // NPerBlock
+        4,                                // K0PerBlock
+        8,                                // K1
+        32,                               // MPerXdl
+        32,                               // NPerXdl
+        2,                                // MXdlPerWave
+        2,                                // NXdlPerWave
+        S<1, 4, 16, 4>,                   // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<0, 3, 1, 2>,                    // ABlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,                    // ABlockTransferSrcAccessOrder
+        2,                                // ABlockTransferSrcVectorDim
+        8,                                // ABlockTransferSrcScalarPerVector
+        2,                                // ABlockTransferDstScalarPerVector_K1
+        true,                             // ABlockLdsAddExtraM
+        S<1, 4, 16, 4>,                   // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<0, 3, 1, 2>,                    // BBlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,                    // BBlockTransferSrcAccessOrder
+        2,                                // BBlockTransferSrcVectorDim
+        8,                                // BBlockTransferSrcScalarPerVector
+        2,                                // BBlockTransferDstScalarPerVector_K1
+        true,                             // BBlockLdsAddExtraN
+        1,                                // CShuffleMXdlPerWavePerShuffle
+        1,                                // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 4>,                   // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        4>;                               // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+template <ck::index_t NumDimSpatial>
+using ReferenceConvBwdWeightInstance =
+    ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
+                                                       WeiDataType,
+                                                       OutDataType,
+                                                       InElementOp,
+                                                       WeiElementOp,
+                                                       OutElementOp,
+                                                       NumDimSpatial>;
+
+template <typename HostTensorB, typename HostTensorA, typename Functor>
+void host_elementwise(HostTensorB& B,
+                      const HostTensorA& A,
+                      const std::vector<std::size_t>& shape,
+                      Functor functor)
+{
+    size_t tensor_size = std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>{});
+    std::cout << __LINE__ << ":" << tensor_size << ", " << A.mData[0] << std::endl;
+    for(std::size_t n = 0; n < tensor_size; ++n)
+    {
+        B.mData[n] = functor(A.mData[n]);
+    }
+}
+
+void print_use_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
+              << "arg4: is show log (0=no, 1=yes)\n"
+              << "arg5: split-k : in this example split-k must be larger than 1\n"
+              << "arg6: N spatial dimensions (default 2)\n"
+              << "Following arguments (depending on number of spatial dims):\n"
+              << " N, K, C, \n"
+              << " <filter spatial dimensions>, (ie Y, X for 2D)\n"
+              << " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
+              << " <strides>, (ie Sy, Sx for 2D)\n"
+              << " <dilations>, (ie Dy, Dx for 2D)\n"
+              << " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
+              << " <right padding>, (ie RightPy, RightPx for 2D)\n"
+              << std::endl;
+}
+
+ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
+{
+    // (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
+    ck::utils::conv::ConvParams params;
+    int arg_idx = 7;
+
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
+
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_spatial_lengths_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.conv_filter_strides_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.conv_filter_dilations_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_left_pads_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_right_pads_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    return params;
+}
+
+DeviceConvBwdWeightBasePtr get_conv_instance(int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return std::make_unique<DeviceConvndBwdWeightInstance_bf16_splitk<3>>();
+    }
+    case 2: {
+        return std::make_unique<DeviceConvndBwdWeightInstance_bf16_splitk<2>>();
+    }
+    case 1: {
+        return std::make_unique<DeviceConvndBwdWeightInstance_bf16_splitk<1>>();
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int num_dim_spatial  = 2;
+    int do_log           = 0;
+    int split_k          = 2;
+
+    ck::utils::conv::ConvParams params;
+    params.C_ = 128;
+
+    if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        do_log          = std::stoi(argv[4]);
+        split_k         = std::stoi(argv[5]);
+    }
+    else if(argc > 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        do_log          = std::stoi(argv[4]);
+        split_k         = std::stoi(argv[5]);
+        num_dim_spatial = std::stoi(argv[6]);
+        // check args number
+        int conv_args     = 3 + num_dim_spatial * 6;
+        int cmdline_nargs = conv_args + 7;
+        if(cmdline_nargs != argc)
+        {
+            print_use_msg();
+            exit(1);
+        }
+
+        params = parse_conv_params(num_dim_spatial, argv);
+    }
+    else if(argc != 1)
+    {
+        print_use_msg();
+        exit(1);
+    }
+
+    if(split_k <= 1)
+    {
+        print_use_msg();
+        exit(1);
+    }
+
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
+    input_dims.insert(std::end(input_dims),
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
+
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
+    filter_dims.insert(std::end(filter_dims),
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
+
+    const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
+    output_dims.insert(std::end(output_dims),
+                       std::begin(output_spatial_lengths),
+                       std::end(output_spatial_lengths));
+
+    Tensor<InDataType> in_n_c_hi_wi(
+        ck::utils::conv::get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
+    Tensor<WeiDataType> wei_k_c_y_x_host_result(
+        ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
+    Tensor<WeiDataType> wei_k_c_y_x_device_result(
+        ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
+    Tensor<OutDataType> out_n_k_ho_wo(
+        ck::utils::conv::get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_device_result.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
+        break;
+    default:
+        out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) *
+                             wei_k_c_y_x_device_result.mDesc.GetElementSpace());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
+    // reset input to zero
+    wei_device_buf.SetZero();
+
+    // do GEMM
+    auto conv    = get_conv_instance(num_dim_spatial);
+    auto invoker = conv->MakeInvokerPointer();
+    auto argument =
+        conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                  static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                  static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                  params.N_,
+                                  params.K_,
+                                  params.C_,
+                                  params.input_spatial_lengths_,
+                                  params.filter_spatial_lengths_,
+                                  output_spatial_lengths,
+                                  params.conv_filter_strides_,
+                                  params.conv_filter_dilations_,
+                                  params.input_left_pads_,
+                                  params.input_right_pads_,
+                                  InElementOp{},
+                                  WeiElementOp{},
+                                  OutElementOp{},
+                                  split_k);
+
+    // alloc work space
+    size_t bwd_weight_workspace_size = conv->GetWorkSpaceSize(argument.get());
+    if(bwd_weight_workspace_size <= 0)
+    {
+        print_use_msg();
+        exit(1);
+    }
+
+    float conv_ave_time = 0.f;
+
+    DeviceMem wei_work_space_device_buf(bwd_weight_workspace_size);
+    wei_work_space_device_buf.SetZero();
+    conv->SetWorkSpacePointer(argument.get(), wei_work_space_device_buf.GetDeviceBuffer());
+
+    if(!conv->IsSupportedArgument(argument.get()))
+    {
+        std::cout << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+        return 1;
+    }
+
+    conv_ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = ck::utils::conv::get_flops(
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
+    std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
+        params.N_,
+        params.C_,
+        params.K_,
+        params.input_spatial_lengths_,
+        params.filter_spatial_lengths_,
+        output_spatial_lengths);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / conv_ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / conv_ave_time;
+
+    std::cout << "Perf: conv: " << conv_ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+              << " GB/s" << std::endl;
+
+    if(do_verification)
+    {
+        auto verify_f = [&](const auto& ref_conv) {
+            auto ref_invoker = ref_conv.MakeInvoker();
+
+            auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
+                                                      wei_k_c_y_x_host_result,
+                                                      out_n_k_ho_wo,
+                                                      params.conv_filter_strides_,
+                                                      params.conv_filter_dilations_,
+                                                      params.input_left_pads_,
+                                                      params.input_right_pads_,
+                                                      InElementOp{},
+                                                      WeiElementOp{},
+                                                      OutElementOp{});
+
+            ref_invoker.Run(ref_argument);
+
+            wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",") << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "wei_device(after): ", wei_k_c_y_x_device_result.mData, ",")
+                    << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "wei_host  : ", wei_k_c_y_x_host_result.mData, ",")
+                    << std::endl;
+            }
+
+            return ck::utils::check_err(wei_k_c_y_x_device_result.mData,
+                                        wei_k_c_y_x_host_result.mData)
+                       ? 0
+                       : 1;
+        };
+
+        switch(num_dim_spatial)
+        {
+        case 3: {
+            auto ref_conv = ReferenceConvBwdWeightInstance<3>();
+            verify_f(ref_conv);
+            break;
+        }
+        case 2: {
+            auto ref_conv = ReferenceConvBwdWeightInstance<2>();
+            verify_f(ref_conv);
+            break;
+        }
+        case 1: {
+            auto ref_conv = ReferenceConvBwdWeightInstance<1>();
+            verify_f(ref_conv);
+            break;
+        }
+        default: {
+            throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+        }
+        }
+    }
+    return 0;
+}

example/21_gemm_layernorm/CMakeLists.txt

@@ -0,0 +1,2 @@
+add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_fp16 gemm_bias_relu_add_layernorm_xdl_fp16.cpp)
+add_example_executable(example_gemm_layernorm_xdl_fp16 gemm_layernorm_xdl_fp16.cpp)

example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp

@@ -0,0 +1,425 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#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_5ary_elementwise.hpp"
+#include "device_gemm_bias_add_reduce_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 ADataType                = F16;
+using BDataType                = F16;
+using CDataType                = F16;
+using C0DataType               = F32;
+using C1DataType               = F16;
+using GemmAccDataType          = F32;
+using ReduceAccDataType        = F32;
+using DDataType                = F32;
+using DPtrsGlobal              = ck::Tuple<DDataType*, DDataType*>;
+using GammaDataType            = F16;
+using BetaDataType             = F16;
+using LayerNormOutDataType     = F16;
+using NormalizeComputeDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AElementOp  = PassThrough;
+using BElementOp  = PassThrough;
+using CElementOp  = ck::tensor_operation::element_wise::Relu;
+using C1ElementOp = PassThrough;
+using ReduceSumOp = ck::reduce::Add;
+using DxsReduceOp = ck::Tuple<ReduceSumOp, ReduceSumOp>;
+
+using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
+using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+using DxsInElementOps       = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+using DxsOutElementOps      = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
+
+using DxsGlobalMemOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmSpecialization =
+    ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmBiasAddReduceInstance = ck::tensor_operation::device::DeviceGemmBiasAddReduce_Xdl_CShuffle
+//######| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|          C1|         Dxs|     DxsInEleOp|     DxsAccEleOp|               D|               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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise|      Reduce|               |                |      MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |      |      |         |         |          |              |   Operation|   Operation|   Operation|   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_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |      |      |         |         |          |              |            |            |            |            |            |               |                |                |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,   F32,   F16,      F32,      F32,       F32,   DPtrsGlobal,  AElementOp,  BElementOp,  CElementOp, C1ElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps,  DxsGlobalMemOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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,             S<64, 4>,                         4,                            1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
+
+// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
+using DeviceNormalizeInstance =
+    ck::tensor_operation::device::Device5AryElementwise<CDataType,
+                                                        DDataType,
+                                                        DDataType,
+                                                        GammaDataType,
+                                                        BetaDataType,
+                                                        LayerNormOutDataType,
+                                                        NormalizeComputeDataType,
+                                                        NormalizeFunctor,
+                                                        2,
+                                                        8,
+                                                        8,  // scalarPerVector: gemm_out
+                                                        1,  // scalarPerVector: reduce_mean
+                                                        1,  // scalarPerVector: reduce_mean_square
+                                                        8,  // scalarPerVector: Gamma
+                                                        8,  // scalarPerVector: Beta
+                                                        8>; // scalarPerVector: LayerNorm_out
+
+auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+    return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                std::vector<std::size_t>({stride}));
+};
+
+auto f_host_tensor_descriptor2d =
+    [](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}));
+        }
+    };
+
+template <typename CDataType,
+          typename DDataType,
+          typename AccDataType,
+          typename C0DataType,
+          typename C1DataType,
+          typename A_functor,
+          typename B_functor,
+          typename C_functor,
+          typename C1_functor>
+void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<ADataType>& b_k_n,
+                         const Tensor<C0DataType>& bias_n,
+                         const Tensor<C1DataType>& c1_m_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<GammaDataType>& beta_n,
+                         A_functor a_element_op,
+                         B_functor b_element_op,
+                         C_functor c_element_op,
+                         C1_functor c1_element_op,
+                         int M,
+                         int N)
+{
+
+    int StrideC = N;
+    Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<DDataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    auto averageOpInst = UnaryDivElementOp{N};
+
+    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, a_element_op, b_element_op, PassThrough{});
+
+    ref_invoker.Run(ref_argument);
+
+    // c = activation(c + bias) + c1_functor(c1)
+    for(int m = 0; m < M; ++m)
+        for(int n = 0; n < N; ++n)
+        {
+            AccDataType acc =
+                static_cast<AccDataType>(c_m_n(m, n)) + static_cast<AccDataType>(bias_n(n));
+
+            AccDataType c1 = static_cast<AccDataType>(c1_m_n(m, n));
+
+            c_element_op(acc, acc);
+            c1_element_op(c1, c1);
+            acc += c1;
+            c_m_n(m, n) = static_cast<CDataType>(acc);
+        }
+
+    // reduce_mean and reduce_square_mean
+    auto reduceSumOpInst = ReduceSumOp{};
+    for(int m = 0; m < M; ++m)
+    {
+        auto mean_acc        = reduceSumOpInst.GetIdentityValue<AccDataType>();
+        auto square_mean_acc = reduceSumOpInst.GetIdentityValue<AccDataType>();
+
+        for(int n = 0; n < N; ++n)
+        {
+            AccDataType c_val        = ck::type_convert<AccDataType>(c_m_n(m, n));
+            AccDataType square_c_val = 0;
+            UnarySquareElementOp{}(square_c_val, c_val);
+
+            reduceSumOpInst(mean_acc, c_val);
+            reduceSumOpInst(square_mean_acc, square_c_val);
+        }
+
+        averageOpInst(mean_acc, mean_acc);
+        averageOpInst(square_mean_acc, square_mean_acc);
+        mean_m(m)       = ck::type_convert<DDataType>(mean_acc);
+        meanSquare_m(m) = ck::type_convert<DDataType>(square_mean_acc);
+    }
+
+    // LayerNorm
+    auto layerNormInst = NormalizeFunctor{};
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            AccDataType out_acc = 0;
+            layerNormInst(out_acc,
+                          static_cast<AccDataType>(c_m_n(m, n)),
+                          static_cast<AccDataType>(mean_m(m)),
+                          static_cast<AccDataType>(meanSquare_m(m)),
+                          static_cast<AccDataType>(gamma_n(n)),
+                          static_cast<AccDataType>(beta_n(n)));
+            out_m_n(m, n) = static_cast<DDataType>(out_acc);
+        }
+    }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename C0DataType,
+          typename C1DataType,
+          typename DDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename NormalizeDataType>
+void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M, int N, int K)
+{
+    std::size_t gemm_flop     = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(CDataType) * M * N + sizeof(C0DataType) * M * N +
+                                sizeof(C1DataType) * M * N + sizeof(DDataType) * M +
+                                sizeof(DDataType) * M;
+
+    std::size_t normalize_num_byte = sizeof(CDataType) * M * N + sizeof(DDataType) * M +
+                                     sizeof(DDataType) * M + sizeof(GammaDataType) * N +
+                                     sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
+
+    float tflops               = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
+    float gemm_gb_per_sec      = gemm_num_byte / 1.E6 / gemm_reduce_time;
+    float normalize_gb_per_sec = normalize_num_byte / 1.E6 / normalize_time;
+
+    std::cout << "gemm + reduce_mean + reduce_square_mean Perf: " << gemm_reduce_time << " ms, "
+              << tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
+
+    std::cout << "5-ary elementwise Perf: " << normalize_time << " ms, " << normalize_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+int main()
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = 1024;
+    ck::index_t StrideB  = 1024;
+    ck::index_t StrideC  = 1024;
+    ck::index_t StrideC1 = 1024;
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<C0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<C1DataType> c1_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> reduceMean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<DDataType> reduceMeanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<LayerNormOutDataType> layerNorm_m_n(
+        f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+    bias_n.GenerateTensorValue(GeneratorTensor_3<C0DataType>{-1, 1});
+    c1_m_n.GenerateTensorValue(GeneratorTensor_3<C1DataType>{-5, 5});
+    gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+    beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpace());
+    DeviceMem bias_device_buf(sizeof(C0DataType) * bias_n.mDesc.GetElementSpace());
+    DeviceMem c1_device_buf(sizeof(C1DataType) * c1_m_n.mDesc.GetElementSpace());
+    DeviceMem reduceMean_device_buf(sizeof(DDataType) * reduceMean_m.mDesc.GetElementSpace());
+    DeviceMem reduceMeanSquare_device_buf(sizeof(DDataType) *
+                                          reduceMeanSquare_m.mDesc.GetElementSpace());
+    DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpace());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpace());
+    DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
+                                   layerNorm_m_n.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    bias_device_buf.ToDevice(bias_n.mData.data());
+    c1_device_buf.ToDevice(c1_m_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    auto a_element_op  = AElementOp{};
+    auto b_element_op  = BElementOp{};
+    auto c_element_op  = CElementOp{};
+    auto c1_element_op = C1ElementOp{};
+    auto dxs_global =
+        ck::make_tuple(static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
+                       static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()));
+
+    auto dxs_in_element_op  = DxsInElementOps{};
+    auto dxs_out_element_op = DxsOutElementOps{N, N};
+
+    // Prepare GEMM, reduce_mean, reduce_mean_square
+    auto gemmReduce         = DeviceGemmBiasAddReduceInstance{};
+    auto gemmReduce_invoker = gemmReduce.MakeInvoker();
+    auto gemmReduce_argument =
+        gemmReduce.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                static_cast<C0DataType*>(bias_device_buf.GetDeviceBuffer()),
+                                static_cast<C1DataType*>(c1_device_buf.GetDeviceBuffer()),
+                                dxs_global,
+                                M,
+                                N,
+                                K,
+                                StrideA,
+                                StrideB,
+                                StrideC,
+                                StrideC1,
+                                a_element_op,
+                                b_element_op,
+                                c_element_op,
+                                c1_element_op,
+                                dxs_in_element_op,
+                                dxs_out_element_op);
+
+    if(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    reduceMean_device_buf.SetZero();
+    reduceMeanSquare_device_buf.SetZero();
+
+    // Prepare LayerNorm
+    auto normalize          = DeviceNormalizeInstance{};
+    auto normalize_invoker  = normalize.MakeInvoker();
+    auto normalize_argument = normalize.MakeArgument(
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
+        static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()),
+        static_cast<GammaDataType*>(gamma_device_buf.GetDeviceBuffer()),
+        static_cast<BetaDataType*>(beta_device_buf.GetDeviceBuffer()),
+        static_cast<LayerNormOutDataType*>(layerNorm_device_buf.GetDeviceBuffer()),
+        {M, N},
+        {StrideC, 1},
+        {1, 0},
+        {1, 0},
+        {0, 1},
+        {0, 1},
+        {StrideC, 1},
+        NormalizeFunctor{});
+
+    if(!normalize.IsSupportedArgument(normalize_argument))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "Device5AryElementwise instance, exiting!");
+    }
+
+    // run kernel
+    gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, false});
+    normalize_invoker.Run(normalize_argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+    {
+        // verification
+        Tensor<LayerNormOutDataType> host_layerNorm_m_n(
+            f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+
+        host_gemm_layernorm<CDataType, DDataType, ReduceAccDataType>(host_layerNorm_m_n,
+                                                                     a_m_k,
+                                                                     b_k_n,
+                                                                     bias_n,
+                                                                     c1_m_n,
+                                                                     gamma_n,
+                                                                     beta_n,
+                                                                     a_element_op,
+                                                                     b_element_op,
+                                                                     c_element_op,
+                                                                     c1_element_op,
+                                                                     M,
+                                                                     N);
+
+        layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
+        pass &= ck::utils::check_err(layerNorm_m_n.mData,
+                                     host_layerNorm_m_n.mData,
+                                     "Error: Incorrect results layerNorm_m_n",
+                                     1e-2,
+                                     1e-2);
+    }
+
+    {
+        // evaluate kernel perf
+        bool time_kernel = true;
+
+        float gemm_reduce_mean_reduce_square_mean_ave_time =
+            gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
+        float normalize_ave_time =
+            normalize_invoker.Run(normalize_argument, StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            DumpGemmLayerNormPerf<ADataType,
+                                  BDataType,
+                                  CDataType,
+                                  C0DataType,
+                                  C1DataType,
+                                  DDataType,
+                                  GammaDataType,
+                                  BetaDataType,
+                                  LayerNormOutDataType>(
+                gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}

example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp

@@ -0,0 +1,379 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#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_5ary_elementwise.hpp"
+#include "device_gemm_reduce_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 ADataType                = F16;
+using BDataType                = F16;
+using CDataType                = F16;
+using GemmAccDataType          = F32;
+using ReduceAccDataType        = F32;
+using DDataType                = F32;
+using DPtrsGlobal              = ck::Tuple<DDataType*, DDataType*>;
+using GammaDataType            = F16;
+using BetaDataType             = F16;
+using LayerNormOutDataType     = F16;
+using NormalizeComputeDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+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;
+using ReduceSumOp = ck::reduce::Add;
+using DxsReduceOp = ck::Tuple<ReduceSumOp, ReduceSumOp>;
+
+using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
+using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+using DxsInElementOps       = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+using DxsOutElementOps      = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
+
+using DxsGlobalMemOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmSpecialization =
+    ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
+//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|         Dxs|     DxsInEleOp|     DxsAccEleOp|             D|               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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise|      Reduce|               |                |    MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |         |         |          |              |   Operation|   Operation|   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_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |         |         |          |              |            |            |            |            |               |                |              |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   DPtrsGlobal,  AElementOp,  BElementOp,  CElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps,  DxsGlobalMemOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              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,             S<64, 4>,                         4,                            1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CElementOp>;
+
+using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
+
+// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
+using DeviceNormalizeInstance =
+    ck::tensor_operation::device::Device5AryElementwise<CDataType,
+                                                        DDataType,
+                                                        DDataType,
+                                                        GammaDataType,
+                                                        BetaDataType,
+                                                        LayerNormOutDataType,
+                                                        NormalizeComputeDataType,
+                                                        NormalizeFunctor,
+                                                        2,
+                                                        8,
+                                                        8,  // scalarPerVector: gemm_out
+                                                        1,  // scalarPerVector: reduce_mean
+                                                        1,  // scalarPerVector: reduce_mean_square
+                                                        8,  // scalarPerVector: Gamma
+                                                        8,  // scalarPerVector: Beta
+                                                        8>; // scalarPerVector: LayerNorm_out
+
+auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+    return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                std::vector<std::size_t>({stride}));
+};
+
+auto f_host_tensor_descriptor2d =
+    [](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}));
+        }
+    };
+
+template <typename CDataType,
+          typename DDataType,
+          typename A_functor,
+          typename B_functor,
+          typename C_functor>
+void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<ADataType>& b_k_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<GammaDataType>& beta_n,
+                         A_functor a_element_op,
+                         B_functor b_element_op,
+                         C_functor c_element_op,
+                         int M,
+                         int N)
+{
+    using out_type = ck::remove_reference_t<decltype(out_m_n(0, 0))>;
+
+    int StrideC = N;
+    Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<DDataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    auto averageOpInst = UnaryDivElementOp{N};
+
+    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, a_element_op, b_element_op, c_element_op);
+
+    ref_invoker.Run(ref_argument);
+
+    // reduce_mean and reduce_square_mean
+    auto reduceSumOpInst = ReduceSumOp{};
+    for(int m = 0; m < M; ++m)
+    {
+        auto mean_acc        = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
+        auto square_mean_acc = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
+
+        for(int n = 0; n < N; ++n)
+        {
+            auto c_val        = ck::type_convert<ReduceAccDataType>(c_m_n(m, n));
+            auto square_c_val = reduceSumOpInst.GetIdentityValue<ReduceAccDataType>();
+
+            UnarySquareElementOp{}(square_c_val, c_val);
+
+            reduceSumOpInst(mean_acc, c_val);
+            reduceSumOpInst(square_mean_acc, square_c_val);
+        }
+
+        averageOpInst(mean_acc, mean_acc);
+        averageOpInst(square_mean_acc, square_mean_acc);
+        mean_m(m)       = ck::type_convert<DDataType>(mean_acc);
+        meanSquare_m(m) = ck::type_convert<DDataType>(square_mean_acc);
+    }
+
+    // LayerNorm
+    auto layerNormInst = NormalizeFunctor{};
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            float out_f32 = 0;
+            layerNormInst(out_f32,
+                          static_cast<float>(c_m_n(m, n)),
+                          static_cast<float>(mean_m(m)),
+                          static_cast<float>(meanSquare_m(m)),
+                          static_cast<float>(gamma_n(n)),
+                          static_cast<float>(beta_n(n)));
+            out_m_n(m, n) = static_cast<out_type>(out_f32);
+        }
+    }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename DDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename NormalizeDataType>
+void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M, int N, int K)
+{
+    std::size_t gemm_flop     = std::size_t(2) * M * N * K;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(CDataType) * M * N + sizeof(DDataType) * M +
+                                sizeof(DDataType) * M;
+
+    std::size_t normalize_num_btye = sizeof(CDataType) * M * N + sizeof(DDataType) * M +
+                                     sizeof(DDataType) * M + sizeof(GammaDataType) * N +
+                                     sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
+
+    float tflops               = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
+    float gemm_gb_per_sec      = gemm_num_byte / 1.E6 / gemm_reduce_time;
+    float normalize_gb_per_sec = normalize_num_btye / 1.E6 / normalize_time;
+
+    std::cout << "gemm + reduce_mean + reduce_square_mean Perf: " << gemm_reduce_time << " ms, "
+              << tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
+
+    std::cout << "5-ary elementwise Perf: " << normalize_time << " ms, " << normalize_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+int main()
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideC = 1024;
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> reduceMean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<DDataType> reduceMeanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<LayerNormOutDataType> layerNorm_m_n(
+        f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+    gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+    beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpace());
+    DeviceMem reduceMean_device_buf(sizeof(DDataType) * reduceMean_m.mDesc.GetElementSpace());
+    DeviceMem reduceMeanSquare_device_buf(sizeof(DDataType) *
+                                          reduceMeanSquare_m.mDesc.GetElementSpace());
+    DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpace());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpace());
+    DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
+                                   layerNorm_m_n.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+    auto dxs_global =
+        ck::make_tuple(static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
+                       static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()));
+
+    auto dxs_in_element_op  = DxsInElementOps{};
+    auto dxs_out_element_op = DxsOutElementOps{N, N};
+
+    // Prepare GEMM, reduce_mean, reduce_mean_square
+    auto gemmReduce         = DeviceGemmReduceInstance{};
+    auto gemmReduce_invoker = gemmReduce.MakeInvoker();
+    auto gemmReduce_argument =
+        gemmReduce.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                dxs_global,
+                                M,
+                                N,
+                                K,
+                                StrideA,
+                                StrideB,
+                                StrideC,
+                                a_element_op,
+                                b_element_op,
+                                c_element_op,
+                                dxs_in_element_op,
+                                dxs_out_element_op);
+
+    if(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    reduceMean_device_buf.SetZero();
+    reduceMeanSquare_device_buf.SetZero();
+
+    // Prepare LayerNorm
+    auto normalize          = DeviceNormalizeInstance{};
+    auto normalize_invoker  = normalize.MakeInvoker();
+    auto normalize_argument = normalize.MakeArgument(
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        static_cast<DDataType*>(reduceMean_device_buf.GetDeviceBuffer()),
+        static_cast<DDataType*>(reduceMeanSquare_device_buf.GetDeviceBuffer()),
+        static_cast<GammaDataType*>(gamma_device_buf.GetDeviceBuffer()),
+        static_cast<BetaDataType*>(beta_device_buf.GetDeviceBuffer()),
+        static_cast<LayerNormOutDataType*>(layerNorm_device_buf.GetDeviceBuffer()),
+        {M, N},
+        {StrideC, 1},
+        {1, 0},
+        {1, 0},
+        {0, 1},
+        {0, 1},
+        {StrideC, 1},
+        NormalizeFunctor{});
+
+    if(!normalize.IsSupportedArgument(normalize_argument))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "Device5AryElementwise instance, exiting!");
+    }
+
+    // run kernel
+    gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, false});
+    normalize_invoker.Run(normalize_argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+    {
+        // verification
+        Tensor<LayerNormOutDataType> host_layerNorm_m_n(
+            f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+
+        host_gemm_layernorm<CDataType, DDataType>(host_layerNorm_m_n,
+                                                  a_m_k,
+                                                  b_k_n,
+                                                  gamma_n,
+                                                  beta_n,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  M,
+                                                  N);
+
+        layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
+        pass &= ck::utils::check_err(layerNorm_m_n.mData,
+                                     host_layerNorm_m_n.mData,
+                                     "Error: Incorrect results d1",
+                                     1e-3,
+                                     1e-3);
+    }
+
+    {
+        // evaluate kernel perf
+        bool time_kernel = true;
+
+        float gemm_reduce_mean_reduce_square_mean_ave_time =
+            gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
+        float normalize_ave_time =
+            normalize_invoker.Run(normalize_argument, StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            DumpGemmLayerNormPerf<ADataType,
+                                  BDataType,
+                                  CDataType,
+                                  DDataType,
+                                  GammaDataType,
+                                  BetaDataType,
+                                  LayerNormOutDataType>(
+                gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}

example/22_cgemm/CMakeLists.txt

@@ -0,0 +1 @@
+add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)

example/22_cgemm/cgemm_xdl_fp16.cpp

@@ -0,0 +1,302 @@
+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#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_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_cgemm.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   = F16;
+using BDataType   = F16;
+using CDataType   = F16;
+using AccDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     32,                         // index_t KPerBlock
+     8,                          // index_t AK1
+     8,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     8,                          // index_t ABlockTransferSrcScalarPerVector
+     8,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 32, 1, 8>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM 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;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        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]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: run kernel # of times (>1)\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_real(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<ADataType> a_m_k_imag(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n_real(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<BDataType> b_k_n_imag(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_real_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_imag_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k_real: " << a_m_k_real.mDesc << std::endl;
+    std::cout << "a_m_k_imag: " << a_m_k_imag.mDesc << std::endl;
+    std::cout << "b_k_n_real: " << b_k_n_real.mDesc << std::endl;
+    std::cout << "b_k_n_imag: " << b_k_n_imag.mDesc << std::endl;
+    std::cout << "c_m_n_real: " << c_m_n_real_device_result.mDesc << std::endl;
+    std::cout << "c_m_n_imag: " << c_m_n_imag_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    default:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    auto cgemm = DeviceCGemmInstance{};
+
+    DeviceMem a_m_k_real_device_buf(sizeof(ADataType) * a_m_k_real.mDesc.GetElementSpace());
+    DeviceMem a_m_k_imag_device_buf(sizeof(ADataType) * a_m_k_imag.mDesc.GetElementSpace());
+    DeviceMem b_k_n_real_device_buf(sizeof(BDataType) * b_k_n_real.mDesc.GetElementSpace());
+    DeviceMem b_k_n_imag_device_buf(sizeof(BDataType) * b_k_n_imag.mDesc.GetElementSpace());
+    DeviceMem c_m_n_real_device_buf(sizeof(CDataType) *
+                                    c_m_n_real_device_result.mDesc.GetElementSpace());
+    DeviceMem c_m_n_imag_device_buf(sizeof(CDataType) *
+                                    c_m_n_imag_device_result.mDesc.GetElementSpace());
+    DeviceMem workspace_device_buf(cgemm.GetWorkspaceSize(M, N, K, StrideA, StrideB, StrideC));
+
+    a_m_k_real_device_buf.ToDevice(a_m_k_real.mData.data());
+    a_m_k_imag_device_buf.ToDevice(a_m_k_imag.mData.data());
+    b_k_n_real_device_buf.ToDevice(b_k_n_real.mData.data());
+    b_k_n_imag_device_buf.ToDevice(b_k_n_imag.mData.data());
+
+    auto a_element_op = PassThrough{};
+    auto b_element_op = PassThrough{};
+    auto c_element_op = PassThrough{};
+
+    // do GEMM
+    auto invoker = cgemm.MakeInvoker();
+    auto argument =
+        cgemm.MakeArgument(static_cast<ADataType*>(a_m_k_real_device_buf.GetDeviceBuffer()),
+                           static_cast<ADataType*>(a_m_k_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<BDataType*>(b_k_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<BDataType*>(b_k_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<CDataType*>(c_m_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<CDataType*>(c_m_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<CDataType*>(workspace_device_buf.GetDeviceBuffer()),
+                           M,
+                           N,
+                           K,
+                           StrideA,
+                           StrideB,
+                           StrideC,
+                           a_element_op,
+                           b_element_op,
+                           c_element_op);
+
+    if(!cgemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_cgemm with the specified compilation parameters does "
+            "not support this CGEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(8) * M * N * K;
+    std::size_t num_btype =
+        std::size_t(2) *
+        (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, "
+              << cgemm.GetTypeString() << std::endl;
+
+    c_m_n_real_device_buf.FromDevice(c_m_n_real_device_result.mData.data());
+    c_m_n_imag_device_buf.FromDevice(c_m_n_imag_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CDataType> c_m_n_real_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+        Tensor<CDataType> c_m_n_imag_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+        auto ref_cgemm   = ReferenceCGemmInstance{};
+        auto ref_invoker = ref_cgemm.MakeInvoker();
+
+        auto ref_argument = ref_cgemm.MakeArgument(a_m_k_real,
+                                                   a_m_k_imag,
+                                                   b_k_n_real,
+                                                   b_k_n_imag,
+                                                   c_m_n_real_host_result,
+                                                   c_m_n_imag_host_result,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        ck::utils::check_err(c_m_n_real_device_result.mData,
+                             c_m_n_real_host_result.mData,
+                             "Verification error: incorrect results in real part!",
+                             1e-2f,
+                             1e-1f);
+        ck::utils::check_err(c_m_n_imag_device_result.mData,
+                             c_m_n_imag_host_result.mData,
+                             "Verification error: incorrect results in imaginary part!",
+                             1e-2f,
+                             1e-1f);
+    }
+
+    return 0;
+}

example/23_softmax/CMakeLists.txt

@@ -0,0 +1 @@
+add_example_executable(example_softmax_blockwise softmax_blockwise.cpp)

example/23_softmax/README.md

@@ -0,0 +1,18 @@
+# Instructions for ```example_softmax_blockwise```
+
+## Run ```example_softmax_blockwise```
+```bash
+# -D <xxx> : input 3-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg2: time kernel (0=no, 1=yes)
+example_softmax_blockwise -D 4,128,2048 -v 1 1 1
+```
+
+Result
+```
+launch_and_time_kernel: grid_dim {64, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 0.0242877 ms, 259.039 GB/s, DeviceReduceSoftmax<256,M_C8_S1,K_C32_S8,InSrcVectorDim_1_InSrcVectorSize_8_OutDstVectorSize_8>
+```

example/23_softmax/softmax_blockwise.cpp

@@ -0,0 +1,255 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "check_err.hpp"
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "device_base.hpp"
+#include "device_softmax.hpp"
+#include "host_common_util.hpp"
+#include "reference_softmax.hpp"
+
+#include "reduction_enums.hpp"
+#include "reduction_operator_mapping.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+constexpr int Rank         = 3;
+constexpr int NumReduceDim = 1;
+
+using DeviceInstance = DeviceSoftmax<InDataType,
+                                     AccDataType,
+                                     OutDataType,
+                                     Rank,
+                                     NumReduceDim,
+                                     256, // BlockSize
+                                     8,   // ClusterM
+                                     32,  // ClusterK
+                                     1,   // SliceM
+                                     8,   // SliceK
+                                     1,   // SrcVecDim (0=M, 1=K)
+                                     8,   // SrcScalarPerVector
+                                     8>;  // OutScalarPerVector
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths   = {8, 128, 2048};
+    std::vector<AccDataType> scales = {2.0f, 2.0f};
+
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg2 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        return (0);
+    };
+};
+
+int main(int argc, char* argv[])
+{
+    // Example: batched gemm C[G, M, N] applies max/sum reduction along N internally
+    const std::vector<int> invariantDims{0, 1};
+    const std::vector<int> reduceDims{2};
+
+    SimpleAppArgs args;
+
+    if(argc > 1)
+    {
+        if(args.processArgs(argc, argv) < 0)
+            return (-1);
+    };
+
+    Tensor<InDataType> in(args.inLengths);
+    Tensor<OutDataType> out_ref(args.inLengths);
+    Tensor<OutDataType> out(args.inLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    AccDataType alpha = args.scales[0];
+    AccDataType beta  = args.scales[1];
+
+    std::size_t num_thread = 1;
+
+    if(args.do_verification)
+    {
+        switch(args.init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-5.0, 5.0}, num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpace(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+    // std::cout << "beta = " << beta << std::endl;
+    // LogRangeAsType<float>(std::cout << "tensor in: " , in.mData, ",") << std::endl;
+    // LogRangeAsType<float>(std::cout << "tensor prior out: " , out.mData, ",") << std::endl;
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
+    DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
+
+    in_dev.ToDevice(in.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    if(args.do_verification)
+    {
+        using ReferenceInstance =
+            tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
+        ReferenceInstance ref;
+        auto ref_arg = ref.MakeArgument(in, out_ref, alpha, beta, Rank, reduceDims);
+        auto invoker = ref.MakeInvoker();
+        invoker.Run(ref_arg);
+        // LogRangeAsType<float>(std::cout << "tensor out_ref: ", out_ref.mData, ",") << std::endl;
+    };
+
+    std::vector<ck::index_t> i_inLengths;
+    std::vector<ck::index_t> i_inStrides;
+
+    i_inLengths.assign(args.inLengths.begin(), args.inLengths.end());
+    i_inStrides.assign(inStrides.begin(), inStrides.end());
+
+    auto device_instance = DeviceInstance{};
+
+    auto argument_ptr = device_instance.MakeArgumentPointer(i_inLengths,
+                                                            i_inStrides,
+                                                            reduceDims,
+                                                            alpha,
+                                                            beta,
+                                                            in_dev.GetDeviceBuffer(),
+                                                            out_dev.GetDeviceBuffer());
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+        return 1;
+    };
+
+    std::string instance_name = device_instance.GetTypeString();
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+
+    bool pass = true;
+    if(args.do_verification)
+    {
+        invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        out_dev.FromDevice(out.mData.data());
+        // LogRangeAsType<float>(std::cout << "tensor out: " , out.mData, ",") << std::endl;
+        pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
+    };
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, args.time_kernel});
+
+    std::size_t num_bytes =
+        in.mDesc.GetElementSize() * sizeof(InDataType) +
+        (beta == 0.0f ? 1 : 2) * out.mDesc.GetElementSize() * sizeof(OutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << instance_name
+              << std::endl;
+
+    return (pass ? 0 : 1);
+}

example/CMakeLists.txt

@@ -1,6 +1,7 @@
 include_directories(BEFORE
     ${PROJECT_SOURCE_DIR}/include/ck
     ${PROJECT_SOURCE_DIR}/include/ck/utility
+    ${PROJECT_SOURCE_DIR}/include/ck/host_utility
     ${PROJECT_SOURCE_DIR}/include/ck/tensor_description
     ${PROJECT_SOURCE_DIR}/include/ck/tensor
     ${PROJECT_SOURCE_DIR}/include/ck/problem_transform
@@ -19,17 +20,26 @@ include_directories(BEFORE
 
 add_custom_target(examples)
 
-function(add_example_executable EXAMPLE_NAME)
+function(add_example_executable EXAMPLE_NAME FILE_NAME)
     message("adding example ${EXAMPLE_NAME}")
-    add_executable(${EXAMPLE_NAME} ${ARGN})
+    add_executable(${EXAMPLE_NAME} ${FILE_NAME})
+    target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor)
+    add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN})
+    add_dependencies(examples ${EXAMPLE_NAME})
+    add_dependencies(check ${EXAMPLE_NAME})
+endfunction(add_example_executable EXAMPLE_NAME)
+
+function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
+    message("adding example ${EXAMPLE_NAME}")
+    add_executable(${EXAMPLE_NAME} ${FILE_NAME})
     target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor)
     add_dependencies(examples ${EXAMPLE_NAME})
-endfunction(add_example_executable EXAMPLE_NAME)
+endfunction(add_example_executable_no_testing EXAMPLE_NAME)
 
 add_subdirectory(01_gemm)
 add_subdirectory(02_gemm_alpha_beta)
 add_subdirectory(03_gemm_bias_relu)
-add_subdirectory(04_gemm_bias_relu_add)
+add_subdirectory(04_gemm_add_add_fastgelu)
 add_subdirectory(06_conv2d_fwd_bias_relu)
 add_subdirectory(07_conv2d_fwd_bias_relu_add)
 add_subdirectory(09_convnd_fwd)
@@ -38,7 +48,12 @@ add_subdirectory(11_conv2d_bwd_weight)
 add_subdirectory(12_reduce)
 add_subdirectory(13_pool2d_fwd)
 add_subdirectory(14_gemm_xdl_requant_relu_requant)
-add_subdirectory(17_convnd_bwd_data_xdl)
 add_subdirectory(15_grouped_gemm)
 add_subdirectory(16_gemm_reduce)
+add_subdirectory(17_convnd_bwd_data_xdl)
 add_subdirectory(18_batched_gemm_reduce)
+add_subdirectory(19_binary_elementwise)
+add_subdirectory(20_convnd_bwd_weight_xdl)
+add_subdirectory(21_gemm_layernorm)
+add_subdirectory(22_cgemm)
+add_subdirectory(23_softmax)