add examples into grouped/batched_gemm

example/15_grouped_gemm/CMakeLists.txt

+add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
 add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
+add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
+add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)

example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_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        = BF16;
+using BDataType        = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| 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, DsLayout, 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
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        EDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    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]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    int group_count = rand() % 16 + 1;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    gemm_descs.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;
+
+        int stride_A = K;
+        int stride_B = K;
+        int stride_C = N;
+
+        gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, {}});
+    }
+
+    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<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> 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(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+
+        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_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * 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(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSpaceSize()));
+
+        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 = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, 0>> p_Ds = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_a, p_b, p_Ds, p_c, gemm_descs, 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(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    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;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < gemm_descs.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);
+            pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
+        }
+    }
+
+    return pass ? 0 : 1;
+}

example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.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        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| 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, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        EDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    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]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    int group_count = rand() % 16 + 1;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    gemm_descs.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;
+
+        int stride_A = K;
+        int stride_B = K;
+        int stride_C = N;
+
+        gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, {}});
+    }
+
+    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<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> 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(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+
+        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_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * 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(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSpaceSize()));
+
+        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 = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, 0>> p_Ds = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_a, p_b, p_Ds, p_c, gemm_descs, 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(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    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;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < gemm_descs.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);
+            pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
+        }
+    }
+
+    return pass ? 0 : 1;
+}

example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.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 AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| 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, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        EDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    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]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    int group_count = rand() % 16 + 1;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    gemm_descs.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        int M = 256 + 256 * i;
+        int N = 128 + 128 * i;
+        int K = 128 + 128 * i;
+
+        int stride_A = K;
+        int stride_B = K;
+        int stride_C = N;
+
+        gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, {}});
+    }
+
+    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<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> 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(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+
+        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_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * 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(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSpaceSize()));
+
+        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 = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, 0>> p_Ds = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_a, p_b, p_Ds, p_c, gemm_descs, 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(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    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;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < gemm_descs.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);
+            pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
+        }
+    }
+
+    return pass ? 0 : 1;
+}

example/24_batched_gemm/CMakeLists.txt

+add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
+add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
+add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
+add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)

example/29_batched_gemm_multi_d/batched_gemm_bias_xdl_fp16.cpp → example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp

@@ -7,7 +7,7 @@
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
-#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
@@ -18,36 +18,31 @@
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
 
-using F16 = ck::half_t;
+using BF16 = ck::bhalf_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 Add         = ck::tensor_operation::element_wise::Add;
 
-using ADataType        = F16;
-using BDataType        = F16;
+using ADataType        = BF16;
+using BDataType        = BF16;
 using AccDataType      = F32;
-using CShuffleDataType = F16;
-using DDataType        = F16;
-using DsDataType       = ck::Tuple<DDataType>;
-using EDataType        = F16;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
 
 using ALayout  = Row;
 using BLayout  = Col;
-using DLayout  = Row;
-using DsLayout = ck::Tuple<DLayout>;
+using DsLayout = ck::Tuple<>;
 using ELayout  = Row;
 
 using AElementOp   = PassThrough;
 using BElementOp   = PassThrough;
-using CDEElementOp = Add;
+using CDEElementOp = PassThrough;
 
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
-// static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
-// static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
 
 // clang-format off
 using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
@@ -58,6 +53,9 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD
         < ALayout, BLayout, DsLayout, 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
 
+using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
+    ReferenceBatchedGemm<ADataType, BDataType, EDataType, AElementOp, BElementOp, CDEElementOp>;
+
 int main(int argc, char* argv[])
 {
     bool do_verification = true;
@@ -70,13 +68,11 @@ int main(int argc, char* argv[])
 
     const int stride_A = K;
     const int stride_B = K;
-    const int stride_D = 0;
-    const int stride_E = N;
+    const int stride_C = N;
 
     const int batch_stride_A = M * K;
     const int batch_stride_B = K * N;
-    const int batch_stride_D = N;
-    const int batch_stride_E = M * N;
+    const int batch_stride_C = M * N;
 
     const int batch_count = 16;
 
@@ -118,15 +114,11 @@ int main(int argc, char* argv[])
     Tensor<BDataType> b_g_k_n(
         f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
 
-    Tensor<DDataType> d_g_m_n(
-        f_host_tensor_descriptor(batch_count, M, N, stride_D, batch_stride_D, DLayout{}));
-
     Tensor<EDataType> e_g_m_n_device_result(
-        f_host_tensor_descriptor(batch_count, M, N, stride_E, batch_stride_E, ELayout{}));
+        f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
 
     std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
     std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
-    std::cout << "d_g_m_n: " << d_g_m_n.mDesc << std::endl;
     std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
 
     switch(init_method)
@@ -135,23 +127,19 @@ int main(int argc, char* argv[])
     case 1:
         a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
         b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-        d_g_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
         break;
     default:
         a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
         b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-        d_g_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
         break;
     }
 
     DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
     DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
-    DeviceMem d_device_buf(sizeof(DDataType) * d_g_m_n.mDesc.GetElementSpaceSize());
     DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
 
     a_device_buf.ToDevice(a_g_m_k.mData.data());
     b_device_buf.ToDevice(b_g_k_n.mData.data());
-    d_device_buf.ToDevice(d_g_m_n.mData.data());
 
     auto a_element_op   = AElementOp{};
     auto b_element_op   = BElementOp{};
@@ -163,7 +151,7 @@ int main(int argc, char* argv[])
     // do GEMM
     auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
                                       b_device_buf.GetDeviceBuffer(),
-                                      {d_device_buf.GetDeviceBuffer()},
+                                      {},
                                       c_device_buf.GetDeviceBuffer(),
                                       M,
                                       N,
@@ -171,12 +159,12 @@ int main(int argc, char* argv[])
                                       batch_count,
                                       stride_A,
                                       stride_B,
-                                      {stride_D},
-                                      stride_E,
+                                      {},
+                                      stride_C,
                                       batch_stride_A,
                                       batch_stride_B,
-                                      {batch_stride_D},
-                                      batch_stride_E,
+                                      {},
+                                      batch_stride_C,
                                       a_element_op,
                                       b_element_op,
                                       cde_element_op);
@@ -208,38 +196,17 @@ int main(int argc, char* argv[])
     {
         c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
 
-        using ReferenceBatchedGemmInstance =
-            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
-                                                             BDataType,
-                                                             EDataType,
-                                                             AElementOp,
-                                                             BElementOp,
-                                                             PassThrough>;
-
         auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
         auto ref_invoker      = ref_batched_gemm.MakeInvoker();
 
         Tensor<EDataType> e_g_m_n_host_result(
-            f_host_tensor_descriptor(batch_count, M, N, stride_E, batch_stride_E, ELayout{}));
+            f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
 
         auto ref_argument = ref_batched_gemm.MakeArgument(
-            a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, PassThrough{});
+            a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
 
         ref_invoker.Run(ref_argument);
 
-        for(int g = 0; g < batch_count; g++)
-        {
-            for(int m = 0; m < M; ++m)
-            {
-                for(int n = 0; n < N; ++n)
-                {
-                    cde_element_op(e_g_m_n_host_result(g, m, n),
-                                   e_g_m_n_host_result(g, m, n),
-                                   d_g_m_n(g, m, n));
-                }
-            }
-        }
-
         pass = ck::utils::check_err(
             e_g_m_n_host_result.mData, e_g_m_n_device_result.mData, "Error: Incorrect results c");
     }

example/29_batched_gemm_multi_d/batched_gemm_xdl_fp16.cpp → example/24_batched_gemm/batched_gemm_xdl_fp16.cpp

@@ -43,8 +43,6 @@ using BElementOp   = PassThrough;
 using CDEElementOp = PassThrough;
 
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
-// static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
-// static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
 
 // clang-format off
 using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl

example/24_batched_gemm_e_permute/batched_gemm_e_permute_xdl_fp16.cpp → example/24_batched_gemm/batched_gemm_xdl_fp32.cpp

@@ -6,7 +6,7 @@
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
-#include "ck/tensor_operation/gpu/device/device_batched_gemm_e_permute_xdl.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
@@ -18,7 +18,6 @@
 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;
@@ -26,14 +25,16 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
 
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
-using ADataType        = F16;
-using BDataType        = F16;
+using ADataType        = F32;
+using BDataType        = F32;
 using AccDataType      = F32;
-using CShuffleDataType = F16;
-using EDataType        = F16;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
 
 using ALayout  = Row;
 using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
 using ELayout  = Row;
 
 using AElementOp   = PassThrough;
@@ -42,13 +43,13 @@ using CDEElementOp = PassThrough;
 
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 
-using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmEPermuteXdl
-    // clang-format off
-//######| ALayout| BLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     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| 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, 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 off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| 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, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 32, 1, 8>,               4>;
 // clang-format on
 
 using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
@@ -60,26 +61,19 @@ int main(int argc, char* argv[])
     int init_method      = 1;
     bool time_kernel     = false;
 
-    const int M = 256;
-    const int N = 128;
-    const int K = 64;
+    const int M = 256 * (rand() % 16 + 1);
+    const int N = 128 * (rand() % 16 + 1);
+    const int K = 64 * (rand() % 16 + 1);
 
     const int stride_A = K;
     const int stride_B = K;
+    const int stride_C = N;
 
     const int batch_stride_A = M * K;
     const int batch_stride_B = K * N;
+    const int batch_stride_C = M * N;
 
-    const int G0 = 16;
-    const int G1 = 8;
-
-    const int batch_count = G0 * G1;
-
-    // output layout - [G0, M, G1, N]
-    const int stride_G0 = M * G1 * N;
-    const int stride_G1 = N;
-    const int stride_M  = G1 * N;
-    const int stride_N  = 1;
+    const int batch_count = 16;
 
     if(argc == 4)
     {
@@ -96,9 +90,6 @@ int main(int argc, char* argv[])
     }
 
     // GEMM shape
-    ck::tensor_operation::device::BatchedGemmEPermuteDesc batched_gemm_e_permute_desc{
-        G0, G1, M, N, stride_G0, stride_G1, stride_M, stride_N};
-
     auto f_host_tensor_descriptor = [](std::size_t batch_count_,
                                        std::size_t row,
                                        std::size_t col,
@@ -122,28 +113,12 @@ int main(int argc, char* argv[])
     Tensor<BDataType> b_g_k_n(
         f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
 
-    auto f_host_e_tensor_descriptor = [](std::size_t G0_,
-                                         std::size_t G1_,
-                                         std::size_t M_,
-                                         std::size_t N_,
-                                         std::size_t stride_G0_,
-                                         std::size_t stride_G1_,
-                                         std::size_t stride_M_,
-                                         std::size_t stride_N_) {
-        return HostTensorDescriptor(
-            std::vector<std::size_t>({G0_, G1_, M_, N_}),
-            std::vector<std::size_t>({stride_G0_, stride_G1_, stride_M_, stride_N_}));
-    };
-
-    Tensor<EDataType> e_g0_g1_m_n_host_result(
-        f_host_e_tensor_descriptor(G0, G1, M, N, stride_G0, stride_G1, stride_M, stride_N));
-
-    Tensor<EDataType> e_g0_g1_m_n_device_result(
-        f_host_e_tensor_descriptor(G0, G1, M, N, stride_G0, stride_G1, stride_M, stride_N));
+    Tensor<EDataType> e_g_m_n_device_result(
+        f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
 
     std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
     std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
-    std::cout << "e_g0_g1_m_n: " << e_g0_g1_m_n_host_result.mDesc << std::endl;
+    std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
 
     switch(init_method)
     {
@@ -160,8 +135,7 @@ int main(int argc, char* argv[])
 
     DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
     DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
-    DeviceMem e_device_buf(sizeof(EDataType) *
-                           e_g0_g1_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
 
     a_device_buf.ToDevice(a_g_m_k.mData.data());
     b_device_buf.ToDevice(b_g_k_n.mData.data());
@@ -173,19 +147,23 @@ int main(int argc, char* argv[])
     auto gemm    = DeviceGemmInstance{};
     auto invoker = gemm.MakeInvoker();
 
-    // do GEM
-    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
-                                      static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
-                                      static_cast<EDataType*>(e_device_buf.GetDeviceBuffer()),
+    // do GEMM
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      c_device_buf.GetDeviceBuffer(),
                                       M,
                                       N,
                                       K,
+                                      batch_count,
                                       stride_A,
                                       stride_B,
+                                      {},
+                                      stride_C,
                                       batch_stride_A,
                                       batch_stride_B,
-                                      batched_gemm_e_permute_desc,
-                                      batch_count,
+                                      {},
+                                      batch_stride_C,
                                       a_element_op,
                                       b_element_op,
                                       cde_element_op);
@@ -215,38 +193,21 @@ int main(int argc, char* argv[])
 
     if(do_verification)
     {
-        e_device_buf.FromDevice(e_g0_g1_m_n_device_result.mData.data());
+        c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
 
         auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
         auto ref_invoker      = ref_batched_gemm.MakeInvoker();
 
-        Tensor<EDataType> c_g_m_n_host_result = HostTensorDescriptor(
-            std::vector<std::size_t>({batch_count, M, N}), std::vector<std::size_t>({M * N, N, 1}));
+        Tensor<EDataType> e_g_m_n_host_result(
+            f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
 
         auto ref_argument = ref_batched_gemm.MakeArgument(
-            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
+            a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
 
         ref_invoker.Run(ref_argument);
 
-        for(int g0 = 0; g0 < G0; g0++)
-        {
-            for(int g1 = 0; g1 < G1; g1++)
-            {
-                for(int m = 0; m < M; m++)
-                {
-                    for(int n = 0; n < N; n++)
-                    {
-                        int g = g0 * G1 + g1;
-
-                        e_g0_g1_m_n_host_result(g0, g1, m, n) = c_g_m_n_host_result(g, m, n);
-                    }
-                }
-            }
-        }
-
-        pass = ck::utils::check_err(e_g0_g1_m_n_host_result.mData,
-                                    e_g0_g1_m_n_device_result.mData,
-                                    "Error: Incorrect results c");
+        pass = ck::utils::check_err(
+            e_g_m_n_host_result.mData, e_g_m_n_device_result.mData, "Error: Incorrect results c");
     }
 
     return pass ? 0 : 1;

example/24_batched_gemm/batched_gemm_xdl_int8.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.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 AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| 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, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
+    ReferenceBatchedGemm<ADataType, BDataType, EDataType, AElementOp, BElementOp, CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    const int M = 256 * (rand() % 16 + 1);
+    const int N = 128 * (rand() % 16 + 1);
+    const int K = 64 * (rand() % 16 + 1);
+
+    const int stride_A = K;
+    const int stride_B = K;
+    const int stride_C = N;
+
+    const int batch_stride_A = M * K;
+    const int batch_stride_B = K * N;
+    const int batch_stride_C = M * N;
+
+    const int batch_count = 16;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    // GEMM shape
+    auto f_host_tensor_descriptor = [](std::size_t batch_count_,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
+    Tensor<BDataType> b_g_k_n(
+        f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
+
+    Tensor<EDataType> e_g_m_n_device_result(
+        f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    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 cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      c_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      batch_count,
+                                      stride_A,
+                                      stride_B,
+                                      {},
+                                      stride_C,
+                                      batch_stride_A,
+                                      batch_stride_B,
+                                      {},
+                                      batch_stride_C,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_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, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = std::size_t(2) * batch_count * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
+                            sizeof(BDataType) * batch_count * K * N +
+                            sizeof(EDataType) * batch_count * 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;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        Tensor<EDataType> e_g_m_n_host_result(
+            f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        pass = ck::utils::check_err(
+            e_g_m_n_host_result.mData, e_g_m_n_device_result.mData, "Error: Incorrect results c");
+    }
+
+    return pass ? 0 : 1;
+}

example/24_batched_gemm_e_permute/CMakeLists.txt

-add_example_executable(example_batched_gemm_e_permute_xdl_fp16 batched_gemm_e_permute_xdl_fp16.cpp)
-

example/29_batched_gemm_multi_d/CMakeLists.txt

-add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
-add_example_executable(example_batched_gemm_bias_xdl_fp16 batched_gemm_bias_xdl_fp16.cpp)
-

example/CMakeLists.txt

@@ -38,10 +38,9 @@ add_subdirectory(20_convnd_bwd_weight)
 add_subdirectory(21_gemm_layernorm)
 add_subdirectory(22_cgemm)
 add_subdirectory(23_softmax)
-add_subdirectory(24_batched_gemm_e_permute)
+add_subdirectory(24_batched_gemm)
 add_subdirectory(25_gemm_bias_e_permute)
 add_subdirectory(26_contraction)
 add_subdirectory(27_layernorm)
 add_subdirectory(28_grouped_gemm_bias)
-add_subdirectory(29_batched_gemm_multi_d)
 add_subdirectory(30_grouped_convnd_fwd_bias_relu)