Merge branch 'develop' into jd/dev_pkg

example/15_grouped_gemm/CMakeLists.txt

+add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)

example/15_grouped_gemm/README.md

+# Instructions for ```example_grouped_gemm_xdl```
+
+## Run ```example_grouped_gemm_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+./bin/example_grouped_gemm_xdl_fp16 0 1 5
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+gemm[0] a_m_k: dim 2, lengths {256, 64}, strides {64, 1} b_k_n: dim 2, lengths {64, 128}, strides {1, 64} c_m_n: dim 2, lengths {256, 128}, strides {128, 1}
+gemm[1] a_m_k: dim 2, lengths {512, 128}, strides {128, 1} b_k_n: dim 2, lengths {128, 256}, strides {1, 128} c_m_n: dim 2, lengths {512, 256}, strides {256, 1}
+gemm[2] a_m_k: dim 2, lengths {768, 192}, strides {192, 1} b_k_n: dim 2, lengths {192, 384}, strides {1, 192} c_m_n: dim 2, lengths {768, 384}, strides {384, 1}
+gemm[3] a_m_k: dim 2, lengths {1024, 256}, strides {256, 1} b_k_n: dim 2, lengths {256, 512}, strides {1, 256} c_m_n: dim 2, lengths {1024, 512}, strides {512, 1}
+group: 0 arg.a_grid_desc_k0_m_k1_{8, 256, 8}, arg.b_grid_desc_k0_n_k1_{8, 128, 8}, arg.c_grid_desc_m_n_{ 256, 128}
+group: 1 arg.a_grid_desc_k0_m_k1_{16, 512, 8}, arg.b_grid_desc_k0_n_k1_{16, 256, 8}, arg.c_grid_desc_m_n_{ 512, 256}
+group: 2 arg.a_grid_desc_k0_m_k1_{24, 768, 8}, arg.b_grid_desc_k0_n_k1_{24, 384, 8}, arg.c_grid_desc_m_n_{ 768, 384}
+group: 3 arg.a_grid_desc_k0_m_k1_{32, 1024, 8}, arg.b_grid_desc_k0_n_k1_{32, 512, 8}, arg.c_grid_desc_m_n_{ 1024, 512}
+launch_and_time_kernel: grid_dim {30, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 5 times...
+Perf: 0.037887 ms, 11.0706 TFlops, 90.8132 GB/s, DeviceGroupedGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2>
+```

example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp

+#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 "device_grouped_gemm_xdl.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::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;
+// static constexpr auto GemmMNPadding =
+// ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdl
+//######| 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|      Num|
+//######|  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| Prefetch|
+//######|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|         |
+//######|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |
+        <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        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 = 0;
+    int init_method      = 0;
+    int nrepeat          = 5;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        nrepeat         = 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");
+        exit(0);
+    }
+
+    int group_count = 4;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmShape> gemm_shapes;
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    gemm_shapes.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        int M = 256 + 256 * i;
+        int N = 128 + 128 * i;
+        int K = 64 + 64 * i;
+
+        gemm_shapes.push_back({M, N, K, K, K, N});
+    }
+
+    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}));
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    ;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<CDataType>> c_host_tensors;
+    std::vector<Tensor<CDataType>> c_device_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    c_host_tensors.reserve(group_count);
+    c_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int 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{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].K, gemm_shapes[i].N, gemm_shapes[i].StrideB, BLayout{})));
+        c_host_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
+        c_device_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
+
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
+                  << std::endl;
+
+        flop += std::size_t(2) * gemm_shapes[i].M * gemm_shapes[i].K * gemm_shapes[i].N;
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(CDataType) * c_device_tensors[i].mDesc.GetElementSize();
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+    }
+
+    for(int i = 0; i < gemm_shapes.size(); i++)
+    {
+        a_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpace()));
+        b_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpace()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_device_tensors[i].mDesc.GetElementSpace()));
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
+        p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    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(p_a, p_b, p_c, gemm_shapes, 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);
+
+    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)
+    {
+        for(int i = 0; i < gemm_shapes.size(); i++)
+        {
+            c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+            ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
+        }
+    }
+
+    return 0;
+}

example/16_gemm_reduce/CMakeLists.txt

+add_example_executable(example_gemm_reduce_xdl_fp16 gemm_reduce_xdl_fp16.cpp)

example/16_gemm_reduce/gemm_reduce_xdl_fp16.cpp

+#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 "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+#include "element_wise_reduce_operation.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::tensor_operation::element_wise::ReduceSum;
+using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum;
+
+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|               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, 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 d0_reduce_op = D0ReduceOp{};
+    auto d1_reduce_op = D1ReduceOp{};
+
+    // 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,
+                                      d0_reduce_op,
+                                      d1_reduce_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);
+
+        for(int m = 0; m < M; ++m)
+        {
+            float d0_acc = d0_reduce_op.GetReduceZeroValue();
+            float d1_acc = d1_reduce_op.GetReduceZeroValue();
+
+            for(int n = 0; n < N; ++n)
+            {
+                d0_reduce_op.Reduce(d0_acc, c_m_n_host_result(m, n));
+                d1_reduce_op.Reduce(d1_acc, c_m_n_host_result(m, n));
+            }
+
+            d0_m_host_result(m) = d0_acc;
+            d1_m_host_result(m) = 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/17_convnd_bwd_data_xdl/CMakeLists.txt

+add_example_executable(example_convnd_bwd_data_xdl convnd_bwd_data_xdl.cpp)

example/17_convnd_bwd_data_xdl/README.md

+# Instructions for ```example_convnd_bwd_data_xdl```
+
+## Run ```example_example_convnd_bwd_data_xdl```
+```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: num_dim_spatial(1|2|3)
+#arg5 to ...: N, K, C, [Z,] [Y,] X, [Di,] [Hi,] Wi, S[z,] [Sy,] Sx, [Dz,] [Dy,] Dx, [LeftPz,] [LeftPy,] LeftPx, [RightPy,] [RightPy,] RightPx
+./bin/example_convnd_bwd_data_xdl 0 1 5 
+```
+
+Result
+```
+in_n_c_hi_wi: dim 4, lengths {128, 128, 71, 71}, strides {645248, 1, 9088, 128}
+wei_k_c_y_x: dim 4, lengths {256, 128, 3, 3}, strides {1152, 1, 384, 128}
+out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
+arg.a_grid_desc_k0_m_k1_container_{128, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{128, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+arg.a_grid_desc_k0_m_k1_container_{64, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{64, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+arg.a_grid_desc_k0_m_k1_container_{64, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{64, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+arg.a_grid_desc_k0_m_k1_container_{32, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{32, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+Perf: 1.40031 ms, 69.8734 TFlops, 179.037 GB/s
+```

example/18_batched_gemm_reduce/CMakeLists.txt

+add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp)
+

example/CMakeLists.txt

@@ -13,6 +13,7 @@ include_directories(BEFORE
     ${PROJECT_SOURCE_DIR}/library/include/ck/library/host_tensor
     ${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/cpu
     ${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/gpu
+    ${PROJECT_SOURCE_DIR}/library/include/ck/library/utility
     ${PROJECT_SOURCE_DIR}/external/include/half
 )
 
@@ -30,12 +31,15 @@ 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(05_conv2d_fwd)
 add_subdirectory(06_conv2d_fwd_bias_relu)
 add_subdirectory(07_conv2d_fwd_bias_relu_add)
-add_subdirectory(08_conv3d_fwd)
 add_subdirectory(09_convnd_fwd)
 add_subdirectory(10_conv2d_bwd_data)
-add_subdirectory(11_conv2d_bwd_wgt)
+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(18_batched_gemm_reduce)

include/ck/tensor/static_tensor.hpp

@@ -4,7 +4,7 @@
 namespace ck {
 
 // StaticTensor for Scalar
-template <AddressSpaceEnum_t AddressSpace,
+template <AddressSpaceEnum AddressSpace,
           typename T,
           typename TensorDesc,
           bool InvalidElementUseNumericalZeroValue,
@@ -80,7 +80,7 @@ struct StaticTensor
 };
 
 // StaticTensor for vector
-template <AddressSpaceEnum_t AddressSpace,
+template <AddressSpaceEnum AddressSpace,
           typename S,
           index_t ScalarPerVector,
           typename TensorDesc,
@@ -245,7 +245,7 @@ struct StaticTensorTupleOfVectorBuffer
     S ignored_element_scalar_;
 };
 
-template <AddressSpaceEnum_t AddressSpace,
+template <AddressSpaceEnum AddressSpace,
           typename T,
           typename TensorDesc,
           typename enable_if<TensorDesc::IsKnownAtCompileTime(), bool>::type = false>
@@ -255,7 +255,7 @@ __host__ __device__ constexpr auto make_static_tensor(TensorDesc)
 }
 
 template <
-    AddressSpaceEnum_t AddressSpace,
+    AddressSpaceEnum AddressSpace,
     typename T,
     typename TensorDesc,
     typename X,

include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp

@@ -207,9 +207,9 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2
                           CM0M1N0N1ThreadDesc{}.GetLength(I2) == N0,
                       "wrong");
 
-        auto a_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatA>(
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
             a_k_m0_m1_thread_desc_.GetElementSpaceSize());
-        auto b_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatB>(
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
             b_k_n0_n1_thread_desc_.GetElementSpaceSize());
 
         constexpr auto threadwise_gemm =

include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r3.hpp

@@ -220,9 +220,9 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
                           CThreadDesc_BM0_BM11_BN0_BN11{}.GetLength(I2) == BN0,
                       "wrong");
 
-        auto a_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatA>(
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
             a_thread_desc_bk0_bm0_bm1_bk1_.GetElementSpaceSize());
-        auto b_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatB>(
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
             b_thread_desc_bk0_bn0_bn1_bk1_.GetElementSpaceSize());
 
         constexpr auto threadwise_contraction =

include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp

@@ -119,7 +119,7 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
         constexpr auto a_block_mtx = ABlockDesc_E1_K1_E2{};
 
         // thread A buffer for GEMM
-        StaticBuffer<AddressSpaceEnum_t::Vgpr, FloatA, a_thread_mtx_.GetElementSpaceSize(), true>
+        StaticBuffer<AddressSpaceEnum::Vgpr, FloatA, a_thread_mtx_.GetElementSpaceSize(), true>
             a_thread_buf;
 
         constexpr auto threadwise_gemm = ThreadwiseGemmDlops_km_kn_mn_v3<FloatA,

include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp

@@ -42,7 +42,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
     static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
 
-    StaticBufferTupleOfVector<AddressSpaceEnum_t::Vgpr,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
                               FloatAcc,
                               MRepeat * NRepeat,
                               xdlops_gemm.GetRegSizePerXdlops(),
@@ -250,9 +250,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                         const BBlockBuffer& b_block_buf,
                         CThreadBuffer& c_thread_buf) const
     {
-        auto a_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAB>(
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
             a_thread_desc_.GetElementSpaceSize());
-        auto b_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAB>(
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
             b_thread_desc_.GetElementSpaceSize());
 
         static_for<0, MRepeat, 1>{}([&](auto m0) {

include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v4r1.hpp

@@ -16,7 +16,7 @@ namespace ck {
 template <index_t BlockSize,
           typename SrcElementwiseOperation,
           typename DstElementwiseOperation,
-          InMemoryDataOperationEnum_t DstInMemOp,
+          InMemoryDataOperationEnum DstInMemOp,
           typename BlockSliceLengths,
           typename ThreadClusterLengths,
           typename ThreadClusterArrangeOrder,