Merge branch 'develop' into mi300_time_measurement

.gitignore

@@ -64,3 +64,5 @@ build*/
 # Python virtualenv
 .venv/
 
+# Python cache
+__pycache__/

CMakeLists.txt

@@ -26,6 +26,8 @@ set(version 1.1.0)
 project(composable_kernel VERSION ${version} LANGUAGES CXX)
 include(CTest)
 
+find_package(Python3 3.8 COMPONENTS Interpreter REQUIRED)
+
 list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
 
 if (DTYPES)

client_example/30_gemm_multi_abd/CMakeLists.txt

+if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "int8" AND DTYPES MATCHES "bf16") OR NOT DEFINED DTYPES))
+	add_executable(client_gemm_bias_fastgelu_bf16_i8_bf16 gemm_bias_fastgelu_xdl_bf16_i8.cpp)
+	target_link_libraries(client_gemm_bias_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_gemm_bias_bf16_i8_bf16 gemm_bias_xdl_bf16_i8.cpp)
+	target_link_libraries(client_gemm_bias_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_gemm_gelu_bf16_i8_bf16 gemm_xdl_gelu_bf16_i8.cpp)
+	target_link_libraries(client_gemm_gelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_gemm_bf16_i8_bf16 gemm_xdl_bf16_i8.cpp)
+	target_link_libraries(client_gemm_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/30_gemm_multi_abd/gemm_bias_fastgelu_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Col;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using Scales      = ck::tensor_operation::element_wise::Scales;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Scales;
+using CDEElementOp = AddFastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 64;
+    ck::index_t N = 1024;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideD = N;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 8)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideD = std::stoi(argv[6]);
+        StrideE = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, Row>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
+                                  f_matrix_space_size(M, K, StrideA, A0Layout{}));
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
+                                  f_matrix_space_size(K, N, StrideB, B0Layout{}));
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD, ELayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                                         BsLayout,
+                                                                         DsLayout,
+                                                                         Row,
+                                                                         AsDataType,
+                                                                         BsDataType,
+                                                                         DsDataType,
+                                                                         BF16,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CDEElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    if(found)
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/30_gemm_multi_abd/gemm_bias_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Col;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using Scales      = ck::tensor_operation::element_wise::Scales;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Scales;
+using CDEElementOp = Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 64;
+    ck::index_t N = 1024;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = M;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = N;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 8)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideD = std::stoi(argv[6]);
+        StrideE = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, Row>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
+                                  f_matrix_space_size(M, K, StrideA, A0Layout{}));
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
+                                  f_matrix_space_size(K, N, StrideB, B0Layout{}));
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD, ELayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                                         BsLayout,
+                                                                         DsLayout,
+                                                                         Row,
+                                                                         AsDataType,
+                                                                         BsDataType,
+                                                                         DsDataType,
+                                                                         BF16,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CDEElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    if(found)
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/30_gemm_multi_abd/gemm_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Col;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using Scales      = ck::tensor_operation::element_wise::Scales;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Scales;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 64;
+    ck::index_t N = 1024;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideE = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, Row>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
+                                  f_matrix_space_size(M, K, StrideA, A0Layout{}));
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
+                                  f_matrix_space_size(K, N, StrideB, B0Layout{}));
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 0;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                                         BsLayout,
+                                                                         DsLayout,
+                                                                         Row,
+                                                                         AsDataType,
+                                                                         BsDataType,
+                                                                         DsDataType,
+                                                                         BF16,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CDEElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    if(found)
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/30_gemm_multi_abd/gemm_xdl_gelu_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Col;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using Scales      = ck::tensor_operation::element_wise::Scales;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Scales;
+using CDEElementOp = FastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 64;
+    ck::index_t N = 1024;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideE = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, Row>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
+                                  f_matrix_space_size(M, K, StrideA, A0Layout{}));
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
+                                  f_matrix_space_size(K, N, StrideB, B0Layout{}));
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 0;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                                         BsLayout,
+                                                                         DsLayout,
+                                                                         Row,
+                                                                         AsDataType,
+                                                                         BsDataType,
+                                                                         DsDataType,
+                                                                         BF16,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CDEElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    if(found)
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                b1_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB, 0},
+            std::array<ck::index_t, NumDTensor>{},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/31_grouped_gemm_multi_abd/CMakeLists.txt

+if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "int8" AND DTYPES MATCHES "bf16") OR NOT DEFINED DTYPES))
+	add_executable(client_grouped_gemm_bias_fastgelu_bf16_i8_bf16 grouped_gemm_bias_fastgelu_xdl_bf16_i8.cpp)
+	target_link_libraries(client_grouped_gemm_bias_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_grouped_gemm_fastgelu_bf16_i8_bf16 grouped_gemm_fastgelu_xdl_bf16_i8.cpp)
+	target_link_libraries(client_grouped_gemm_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/31_grouped_gemm_multi_abd/grouped_gemm_bias_fastgelu_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_multi_abd_fixed_nk.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Col;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using Scales      = ck::tensor_operation::element_wise::Scales;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Scales;
+using CDEElementOp = AddFastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
+
+    std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
+        d0_tensors_device, c_tensors_device;
+
+    a0_tensors_device.reserve(group_count);
+    b0_tensors_device.reserve(group_count);
+    b1_tensors_device.reserve(group_count);
+    d0_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+    }
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::
+        GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a0_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
+
+        b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
+            sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        b1_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
+
+        d0_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        gemm_descs.push_back(
+            {sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {0}, 1});
+
+        grouped_gemm_kernel_args_.push_back(
+            {std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
+                                                 b1_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumDTensor>{d0_tensors_device[i]->GetDeviceBuffer()},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
+             std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
+             std::array<ck::index_t, NumDTensor>{0},
+             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmMultiABDFixedNK<AsLayout,
+                                                                                    BsLayout,
+                                                                                    DsLayout,
+                                                                                    Row,
+                                                                                    AsDataType,
+                                                                                    BsDataType,
+                                                                                    DsDataType,
+                                                                                    BF16,
+                                                                                    AElementOp,
+                                                                                    BElementOp,
+                                                                                    CDEElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        std::vector<std::array<const void*, NumATensor>> p_As = {};
+        std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
+        std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
+        std::vector<void*> p_Cs                               = {};
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            SimpleDeviceMem gemm_kernel_args_dev(
+                op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+            hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                                      grouped_gemm_kernel_args_.data(),
+                                      op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                      hipMemcpyHostToDevice));
+
+            op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
+
+            op_ptr->SetElementwiseOps(
+                argument_ptr.get(), a_element_op, b_element_op, cde_element_op);
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
+
+            std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
+                                    sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
+                                    sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return true;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(32 + rand() % 32);
+        problem_size.Ns.push_back(1024);
+        problem_size.Ks.push_back(512);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ns[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

client_example/31_grouped_gemm_multi_abd/grouped_gemm_fastgelu_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_multi_abd_fixed_nk.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using A0Layout = Col;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using Scales      = ck::tensor_operation::element_wise::Scales;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Scales;
+using CDEElementOp = FastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
+
+    std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
+        c_tensors_device;
+
+    a0_tensors_device.reserve(group_count);
+    b0_tensors_device.reserve(group_count);
+    b1_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+    }
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 0;
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::
+        GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a0_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
+
+        b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
+            sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        b1_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        gemm_descs.push_back(
+            {sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {}, 1});
+
+        grouped_gemm_kernel_args_.push_back(
+            {std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
+                                                 b1_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumDTensor>{},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
+             std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
+             std::array<ck::index_t, NumDTensor>{},
+             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmMultiABDFixedNK<AsLayout,
+                                                                                    BsLayout,
+                                                                                    DsLayout,
+                                                                                    Row,
+                                                                                    AsDataType,
+                                                                                    BsDataType,
+                                                                                    DsDataType,
+                                                                                    BF16,
+                                                                                    AElementOp,
+                                                                                    BElementOp,
+                                                                                    CDEElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        std::vector<std::array<const void*, NumATensor>> p_As = {};
+        std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
+        std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
+        std::vector<void*> p_Cs                               = {};
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            SimpleDeviceMem gemm_kernel_args_dev(
+                op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+            hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                                      grouped_gemm_kernel_args_.data(),
+                                      op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                      hipMemcpyHostToDevice));
+
+            op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
+
+            op_ptr->SetElementwiseOps(
+                argument_ptr.get(), a_element_op, b_element_op, cde_element_op);
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
+
+            std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
+                                    sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
+                                    sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return true;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(32 + rand() % 32);
+        problem_size.Ns.push_back(1024);
+        problem_size.Ks.push_back(512);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ns[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

cmake/EnableCompilerWarnings.cmake

@@ -95,6 +95,7 @@ else()
                 -Wno-weak-vtables
                 -Wno-covered-switch-default
                 -Wno-unsafe-buffer-usage
+                -Wno-unused-lambda-capture
             )
         else()
             if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "GNU" AND ${COMPILER} MATCHES "CXX")

docs/index.rst

@@ -33,6 +33,6 @@ The CK documentation is structured as follows:
 
     * :ref:`hello-world`
 
-To contribute to the documentation refer to `Contributing to ROCm  <https://rocm.docs.amd.com/en/latest/contribute/index.html>`_.
+To contribute to the documentation refer to `Contributing to ROCm  <https://rocm.docs.amd.com/en/latest/contribute/contributing.html>`_.
 
 You can find licensing information on the `Licensing <https://rocm.docs.amd.com/en/latest/about/license.html>`_ page.

example/59_grouped_gemm_multi_ABD/CMakeLists.txt

+add_custom_target(example_grouped_gemm_xdl_multi_abd)
+
+add_example_executable(example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16 grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16)
+
+add_example_executable(example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8 grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8)

example/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multi_abd_xdl_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
+#include "ck/tensor_operation/gpu/element/unary_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/utility/literals.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 I8   = int8_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 A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Col;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using DsLayout = ck::Tuple<Row>;
+using ELayout  = Row;
+
+using Scales      = ck::tensor_operation::element_wise::Scales;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Scales;
+using CDEElementOp = AddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Multi_ABD_Fixed_NK
+    // 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|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   128,    16,   128,    32,   8,   8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               1>;
+
+// clang-format on
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::vector<Tensor<A0DataType>> a0_tensors;
+    std::vector<Tensor<B1DataType>> b_tensors;
+    std::vector<Tensor<B0DataType>> b0_tensors;
+    std::vector<Tensor<B1DataType>> b1_tensors;
+    std::vector<Tensor<D0DataType>> d0_tensors;
+    std::vector<Tensor<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> c_device_tensors;
+
+    a0_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    b0_tensors.reserve(group_count);
+    b1_tensors.reserve(group_count);
+    d0_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> a0_tensors_device, b0_tensors_device, b1_tensors_device,
+        d0_tensors_device, c_tensors_device;
+
+    a0_tensors_device.reserve(group_count);
+    b0_tensors_device.reserve(group_count);
+    b1_tensors_device.reserve(group_count);
+    d0_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+
+        a0_tensors.push_back(Tensor<A0DataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], A0Layout{})));
+
+        b_tensors.push_back(Tensor<B1DataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], B0Layout{})));
+        b0_tensors.push_back(Tensor<B0DataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], B0Layout{})));
+        b1_tensors.push_back(Tensor<B1DataType>(
+            f_host_tensor_descriptor(problem_size.Ks[i], problem_size.Ns[i], 0, B1Layout{})));
+
+        d0_tensors.push_back(Tensor<D0DataType>(
+            f_host_tensor_descriptor(problem_size.Ms[i], problem_size.Ns[i], 0, ELayout{})));
+
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+
+        std::cout << "gemm[" << i << "] a_m_k: " << a0_tensors[i].mDesc
+                  << " b_k_n: " << b0_tensors[i].mDesc << " d_m_n: " << d0_tensors[i].mDesc
+                  << " c_m_n: " << c_device_tensors[i].mDesc << std::endl;
+
+        flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
+        num_btype += sizeof(A0DataType) * a0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(B0DataType) * b0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(B1DataType) * b1_tensors[i].mDesc.GetElementSize() +
+                     sizeof(D0DataType) * d0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+            b0_tensors[i].GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+            b1_tensors[i].GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
+            break;
+        case 2:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+            b0_tensors[i].GenerateTensorValue(GeneratorTensor_3<B0DataType>{-5, 5});
+            b1_tensors[i].GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+            break;
+        default:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+            b1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+
+        d0_tensors[i].GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::
+        GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a0_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
+
+        b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        b1_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
+
+        d0_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        a0_tensors_device[i]->ToDevice(a0_tensors[i].mData.data(),
+                                       a0_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(A0DataType));
+
+        b0_tensors_device[i]->ToDevice(b0_tensors[i].mData.data(),
+                                       b0_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(B0DataType));
+
+        b1_tensors_device[i]->ToDevice(b1_tensors[i].mData.data(),
+                                       b1_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(B1DataType));
+
+        d0_tensors_device[i]->ToDevice(d0_tensors[i].mData.data());
+        c_tensors_device[i]->SetZero();
+
+        gemm_descs.push_back(
+            {sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {0}, 1});
+
+        grouped_gemm_kernel_args_.push_back(
+            {std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
+                                                 b1_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumDTensor>{d0_tensors_device[i]->GetDeviceBuffer()},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
+             std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
+             std::array<ck::index_t, NumDTensor>{0},
+             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, NumATensor>> p_As = {};
+    std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
+    std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
+    std::vector<void*> p_Cs                               = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
+
+    DeviceMem gemm_kernel_args_dev(gemm.GetDeviceKernelArgSize(&argument));
+    hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                              grouped_gemm_kernel_args_.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+
+    gemm.SetDeviceKernelArgs(argument, gemm_kernel_args_dev.GetDeviceBuffer());
+    gemm.SetKBatch(argument, config.k_batch);
+
+    gemm.SetElementwiseOps(argument, a_element_op, b_element_op, cde_element_op);
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    if(config.time_kernel)
+    {
+        float ave_time   = invoker.Run(argument, StreamConfig{nullptr, config.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(config.do_verification)
+    {
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B1DataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            for(int n = 0; n < problem_size.Ns[i]; ++n)
+            {
+                for(int k = 0; k < problem_size.Ks[i]; ++k)
+                {
+                    b_element_op(b_tensors[i](k, n), b0_tensors[i](k, n), b1_tensors[i](k, n));
+                }
+            }
+
+            c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data(),
+                                            c_device_tensors[i].mDesc.GetElementSize() *
+                                                sizeof(EDataType));
+
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a0_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      PassThrough{},
+                                                      PassThrough{},
+                                                      PassThrough{});
+
+            ref_invoker.Run(ref_argument);
+
+            for(int m = 0; m < problem_size.Ms[i]; ++m)
+            {
+                for(int n = 0; n < problem_size.Ns[i]; ++n)
+                {
+                    cde_element_op(
+                        c_host_tensors[i](m, n), c_host_tensors[i](m, n), d0_tensors[i](m, n));
+                }
+            }
+
+            pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
+        }
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(32 + rand() % 32);
+        problem_size.Ns.push_back(1024);
+        problem_size.Ks.push_back(512);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    if(argc == 5)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        config.k_batch         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4: k_batch (>0)\n");
+        exit(0);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

example/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, 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/impl/device_grouped_gemm_multi_abd_xdl_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
+#include "ck/tensor_operation/gpu/element/combined_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/utility/literals.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 Add         = ck::tensor_operation::element_wise::Add;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+using AddScale = ck::tensor_operation::element_wise::BinaryWithUnaryCombinedOp<Add, Scale, Scale>;
+
+using A0DataType       = F16;
+using A1DataType       = F32;
+using AsDataType       = ck::Tuple<A0DataType, A1DataType>;
+using B0DataType       = F16;
+using BsDataType       = ck::Tuple<B0DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = F32;
+
+using A0Layout = Row;
+using A1Layout = Row;
+using AsLayout = ck::Tuple<A0Layout, A1Layout>;
+using B0Layout = Col;
+using BsLayout = ck::Tuple<B0Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using AElementOp = AddScale;
+using BElementOp = PassThrough;
+
+using CDEElementOp = Add;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Multi_ABD_Fixed_NK
+    // 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|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   128,    16,   128,    32,   8,   8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               1,  ck::half_t>;
+
+// clang-format on
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::vector<Tensor<A0DataType>> a0_tensors;
+    std::vector<Tensor<A1DataType>> a1_tensors;
+    std::vector<Tensor<B0DataType>> b_tensors;
+    std::vector<Tensor<D0DataType>> d0_tensors;
+    std::vector<Tensor<EDataType>> e_host_tensors;
+    std::vector<Tensor<EDataType>> e_device_tensors;
+
+    a0_tensors.reserve(group_count);
+    a1_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    d0_tensors.reserve(group_count);
+    e_host_tensors.reserve(group_count);
+    e_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a0_tensors_device, a1_tensors_device, b_tensors_device,
+        d0_tensors_device, c_tensors_device;
+
+    a0_tensors_device.reserve(group_count);
+    a1_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    d0_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+        a0_tensors.push_back(Tensor<A0DataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], A0Layout{})));
+        a1_tensors.push_back(Tensor<A1DataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], A1Layout{})));
+        b_tensors.push_back(Tensor<B0DataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], B0Layout{})));
+        d0_tensors.push_back(Tensor<D0DataType>(
+            f_host_tensor_descriptor(problem_size.Ms[i], problem_size.Ns[i], 0, ELayout{})));
+        e_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        e_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        std::cout << "gemm[" << i << "] a_m_k: " << a0_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " d_m_n: " << d0_tensors[i].mDesc
+                  << " c_m_n: " << e_device_tensors[i].mDesc << std::endl;
+
+        flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
+        num_btype += sizeof(A0DataType) * a0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(A1DataType) * a1_tensors[i].mDesc.GetElementSize() +
+                     sizeof(B0DataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(D0DataType) * d0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+            a1_tensors[i].GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+            break;
+        case 2:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+            a1_tensors[i].GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+            break;
+        default:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            a1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+
+        d0_tensors[i].GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    constexpr ck::index_t NumATensor = 2;
+    constexpr ck::index_t NumBTensor = 1;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::
+        GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a0_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
+
+        a1_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(A1DataType) * sum_of_m * problem_size.Ks[i]));
+
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        d0_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        a0_tensors_device[i]->ToDevice(a0_tensors[i].mData.data(),
+                                       a0_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(A0DataType));
+
+        a1_tensors_device[i]->ToDevice(a1_tensors[i].mData.data(),
+                                       a1_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(A1DataType));
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data(),
+                                      b_tensors[i].mDesc.GetElementSpaceSize() *
+                                          sizeof(B0DataType));
+        d0_tensors_device[i]->ToDevice(d0_tensors[i].mData.data());
+        c_tensors_device[i]->SetZero();
+
+        gemm_descs.push_back({sum_of_m,
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              {1, 1},
+                              {problem_size.stride_Bs[i]},
+                              {0},
+                              1});
+
+        grouped_gemm_kernel_args_.push_back(
+            {std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer(),
+                                                 a1_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumBTensor>{b_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumDTensor>{d0_tensors_device[i]->GetDeviceBuffer()},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             std::array<ck::index_t, NumATensor>{problem_size.stride_As[i],
+                                                 problem_size.stride_As[i]},
+             std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i]},
+             std::array<ck::index_t, NumDTensor>{0},
+             problem_size.stride_Cs[i]});
+    }
+
+    constexpr float scale = 1.f;
+    auto a_element_op     = AElementOp{Add{}, Scale{scale}, Scale{scale}};
+    auto b_element_op     = BElementOp{};
+    auto cde_element_op   = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, NumATensor>> p_As = {};
+    std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
+    std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
+    std::vector<void*> p_Cs                               = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
+
+    DeviceMem gemm_kernel_args_dev(gemm.GetDeviceKernelArgSize(&argument));
+    hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                              grouped_gemm_kernel_args_.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+
+    gemm.SetDeviceKernelArgs(argument, gemm_kernel_args_dev.GetDeviceBuffer());
+    gemm.SetKBatch(argument, config.k_batch);
+
+    gemm.SetElementwiseOps(argument, a_element_op, b_element_op, cde_element_op);
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    if(config.time_kernel)
+    {
+        float ave_time   = invoker.Run(argument, StreamConfig{nullptr, config.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(config.do_verification)
+    {
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B0DataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            for(int m = 0; m < problem_size.Ms[i]; ++m)
+            {
+                for(int k = 0; k < problem_size.Ks[i]; ++k)
+                {
+                    a_element_op(a0_tensors[i](m, k), a0_tensors[i](m, k), a1_tensors[i](m, k));
+                }
+            }
+
+            c_tensors_device[i]->FromDevice(e_device_tensors[i].mData.data(),
+                                            e_device_tensors[i].mDesc.GetElementSize() *
+                                                sizeof(EDataType));
+
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a0_tensors[i],
+                                                      b_tensors[i],
+                                                      e_host_tensors[i],
+                                                      PassThrough{},
+                                                      b_element_op,
+                                                      PassThrough{});
+
+            ref_invoker.Run(ref_argument);
+
+            for(int m = 0; m < problem_size.Ms[i]; ++m)
+            {
+                for(int n = 0; n < problem_size.Ns[i]; ++n)
+                {
+                    cde_element_op(
+                        e_host_tensors[i](m, n), e_host_tensors[i](m, n), d0_tensors[i](m, n));
+                }
+            }
+
+            pass &= ck::utils::check_err(e_device_tensors[i], e_host_tensors[i]);
+        }
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(32 + rand() % 32);
+        problem_size.Ns.push_back(1024);
+        problem_size.Ks.push_back(512);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    if(argc == 5)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        config.k_batch         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4: k_batch (>0)\n");
+        exit(0);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

example/60_gemm_multi_ABD/CMakeLists.txt

 add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
+add_example_executable(example_gemm_multi_ABD_xdl_bf16_i8 gemm_multi_ABD_xdl_bf16_i8.cpp)
\ No newline at end of file

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Col;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using Scales      = ck::tensor_operation::element_wise::Scales;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Scales;
+using CDEElementOp = AddFastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
+    // 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|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,   128,    32,   8,   8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               1>;
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 64;
+    ck::index_t N = 1024;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideD = N;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
+    Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    b1_device_buf.ToDevice(b1_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 1;
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                                   b1_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumDTensor>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               std::array<ck::index_t, NumATensor>{StrideA},
+                               std::array<ck::index_t, NumBTensor>{StrideB, 0},
+                               std::array<ck::index_t, NumDTensor>{StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        Tensor<A0DataType> a_m_k({M, K});
+
+        Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
+            }
+        }
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B1DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp

@@ -37,7 +37,7 @@ using DDataType        = F16;
 using EDataType        = F16;
 
 using ALayout = Row;
-using BLayout = Col;
+using BLayout = Row;
 using DLayout = Row;
 using ELayout = Row;
 
@@ -141,9 +141,9 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl
     S<4, 64, 1>,
     S<1, 0, 2>,
     S<1, 0, 2>,
+    1,
     2,
     8,
-    8,
     1,
     1,
     1,
@@ -161,10 +161,10 @@ int main(int argc, char* argv[])
     ck::index_t N = 4096;
     ck::index_t K = 4096;
 
-    ck::index_t StrideA = 4096;
-    ck::index_t StrideB = 4096;
-    ck::index_t StrideD = 4096;
-    ck::index_t StrideE = 4096;
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = N;
+    ck::index_t StrideE = N;
 
     float alpha = 1.0f;
     float beta  = 1.0f;

example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp

@@ -102,7 +102,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceContractionMultiple
     S<1, 0, 2>,
     S<1, 0, 2>,
     2,
-    8,
+    1,
     8,
     1,
     S<4, 64, 1>,
@@ -131,7 +131,7 @@ int main(int argc, char* argv[])
     std::vector<ck::index_t> a0_ms_ks_strides{128 * 32 * 64, 32 * 64, 64, 1};
     // A1[M1, K1] -> A1[M0, M1, K0, K1]
     std::vector<ck::index_t> a1_ms_ks_lengths{30, 128, 32, 64};
-    std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 0, 1};
+    std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 1, 0};
     // B[N0, N1, K0, K1]
     std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
     std::vector<ck::index_t> b_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};

example/ck_tile/01_fmha/CMakeLists.txt

+# generate a list of kernels, but not actually emit files at config stage
+execute_process(
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/blob_list.txt
+)
+
+# NOTE: for cmake, the FMHA_FWD_GEN_BLOBS files must be in the same directory
+#       as current cmake list, otherwise will not figure out the dependency properly
+file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/blob_list.txt FMHA_FWD_GEN_BLOBS)
+
+add_custom_command(
+  OUTPUT ${FMHA_FWD_GEN_BLOBS}
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --output_dir ${CMAKE_CURRENT_BINARY_DIR}
+)
+
+set(EXAMPLE_FMHA_FWD "tile_example_fmha_fwd")
+# not using add_example_executable() to add this target, since we don't want this to have
+# to be included in "make all/install/check"
+message("adding tile_example ${EXAMPLE_NAME}")
+add_executable(${EXAMPLE_FMHA_FWD} EXCLUDE_FROM_ALL fmha_fwd.cpp)
+target_include_directories(${EXAMPLE_FMHA_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+target_sources(${EXAMPLE_FMHA_FWD} PRIVATE ${FMHA_FWD_GEN_BLOBS})
+
+# NOTE: this is dangerous since will change the whole kernel to flush denormals
+#       WIP with compiler team for an exp2 intrinsic..., then remove this
+if(NOT DEFINED FMHA_FWD_FAST_EXP2)
+    set(FMHA_FWD_FAST_EXP2 true)
+endif()
+
+set(EXAMPLE_FMHA_FWD_COMPILE_OPTIONS)
+
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+#       ... because they are auto-generated
+if(FMHA_FWD_FAST_EXP2)
+	list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero)
+else()
+	list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=0)
+endif()
+
+# Allow comparing floating points directly in order to check sentinel values
+list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-float-equal)
+
+target_compile_options(${EXAMPLE_FMHA_FWD} PRIVATE ${EXAMPLE_FMHA_FWD_COMPILE_OPTIONS})