Merge branch 'amd-develop-0605' into amd-master

client_example/24_grouped_conv_activation/grouped_convnd_bwd_weight_scale/grouped_conv_bwd_weight_scale_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_scale.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout   = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout   = ck::tensor_layout::convolution::NDHWGK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 32; // batch size
+static constexpr ck::index_t K             = 32; // output channel
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Z             = 3;  // filter D
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Di            = 14; // input D
+static constexpr ck::index_t Hi            = 14; // input H
+static constexpr ck::index_t Wi            = 14; // input W
+static constexpr ck::index_t Do            = 14; // output D
+static constexpr ck::index_t Ho            = 14; // output H
+static constexpr ck::index_t Wo            = 14; // output W
+
+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_;
+};
+
+int execute_conv_bwd_weight_scale()
+{
+    constexpr ck::index_t split_k = 2;
+
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{
+        C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
+        K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{
+        K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvBwdWeightMultipleD<NumDimSpatial,
+                                                                          InLayout,
+                                                                          WeiLayout,
+                                                                          OutLayout,
+                                                                          ck::Tuple<>,
+                                                                          InDataType,
+                                                                          WeiDataType,
+                                                                          OutDataType,
+                                                                          ck::Tuple<>,
+                                                                          PassThrough,
+                                                                          Scale,
+                                                                          PassThrough>;
+
+    // get device op instances
+    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;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 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(static_cast<InDataType*>(in.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out.GetDeviceBuffer()),
+                                        {},
+                                        in_lengths,
+                                        in_strides,
+                                        wei_lengths,
+                                        wei_strides,
+                                        out_lengths,
+                                        out_strides,
+                                        {},
+                                        {},
+                                        filter_strides,
+                                        filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        PassThrough{},
+                                        Scale{2.f},
+                                        PassThrough{},
+                                        split_k);
+
+        SimpleDeviceMem workspace_buf(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_buf.GetDeviceBuffer());
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop =
+                std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X + G * K * Z * Y * X * C;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Di * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Z * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        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(static_cast<InDataType*>(in.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out.GetDeviceBuffer()),
+                                        {},
+                                        in_lengths,
+                                        in_strides,
+                                        wei_lengths,
+                                        wei_strides,
+                                        out_lengths,
+                                        out_strides,
+                                        {},
+                                        {},
+                                        filter_strides,
+                                        filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        PassThrough{},
+                                        Scale{2.f},
+                                        PassThrough{},
+                                        split_k);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+        SimpleDeviceMem workspace_buf(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_buf.GetDeviceBuffer());
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return 0;
+}
+
+int main() { return execute_conv_bwd_weight_scale(); }

client_example/25_wrapper/wrapper_img2col.cpp

@@ -181,4 +181,3 @@ int main(int argc, char* argv[])
                              {1, 1, 1} /*filter_dilations*/);
     return 0;
 }
-// MI100 Perf:   0.255178 ms, 1698.9 GB/s,

client_example/30_gemm_bf16Aint8B/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)
+
+	add_executable(client_gemm_multiply_bf16_i8_bf16 gemm_xdl_multiply_bf16_i8.cpp)
+	target_link_libraries(client_gemm_multiply_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/30_gemm_bf16Aint8B/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 = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+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_bf16Aint8B/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 = Row;
+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 Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+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_bf16Aint8B/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 = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+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 = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    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_bf16Aint8B/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 = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+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/30_gemm_bf16Aint8B/gemm_xdl_multiply_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>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<B1DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<B1Layout>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Multiply;
+
+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 = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    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 = 1;
+    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()},
+            std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB},
+            std::array<ck::index_t, NumDTensor>{0},
+            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;
+
+    return 0;
+}

client_example/31_grouped_gemm_bf16Aint8B/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)
+
+	add_executable(client_grouped_gemm_multiply_bf16_i8_bf16 grouped_gemm_multiply_xdl_bf16_i8.cpp)
+	target_link_libraries(client_grouped_gemm_multiply_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_grouped_gemm_multiply_bias_fastgelu_bf16_i8_bf16 grouped_gemm_multiply_bias_fastgelu_xdl_bf16_i8.cpp)
+	target_link_libraries(client_grouped_gemm_multiply_bias_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_grouped_gemm_bf16_i8_bf16 grouped_gemm_xdl_bf16_i8.cpp)
+	target_link_libraries(client_grouped_gemm_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+endif()

client_example/31_grouped_gemm_bf16Aint8B/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 = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+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_bf16Aint8B/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"
+
+#include "ck/host_utility/hip_check_error.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 = Row;
+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 Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+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);
+}

client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_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_tile_loop_multply.hpp"
+
+#include "ck/host_utility/hip_check_error.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 B0DataType       = I8;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<B1DataType, D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<B0Layout, D0Layout>;
+using ELayout  = Row;
+
+using PassThrough         = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MultiplyAddFastGelu;
+
+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::GemmDesc> 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 NumDTensor = 2;
+
+    using GroupedGemmKernelArgument =
+        ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<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) * problem_size.Ms[i] * 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) * problem_size.Ms[i] * problem_size.Ns[i]));
+
+        d0_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
+
+        gemm_descs.push_back({problem_size.Ms[i],
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              problem_size.stride_As[i],
+                              problem_size.stride_Bs[i],
+                              problem_size.stride_Cs[i],
+                              {0, 0}});
+
+        grouped_gemm_kernel_args_.push_back(
+            {a0_tensors_device[i]->GetDeviceBuffer(),
+             b0_tensors_device[i]->GetDeviceBuffer(),
+             {b1_tensors_device[i]->GetDeviceBuffer(), d0_tensors_device[i]->GetDeviceBuffer()},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             problem_size.stride_As[i],
+             problem_size.stride_Bs[i],
+             {0, 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::DeviceGroupedGemmTileLoop<A0Layout,
+                                                                             B0Layout,
+                                                                             DsLayout,
+                                                                             ELayout,
+                                                                             A0DataType,
+                                                                             B0DataType,
+                                                                             DsDataType,
+                                                                             EDataType,
+                                                                             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<const void*> p_As                         = {};
+        std::vector<const void*> 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, 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()))
+        {
+            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());
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
+
+            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(1 + rand() % 1024);
+        problem_size.Ns.push_back(6144);
+        problem_size.Ks.push_back(4096);
+
+        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]);
+
+        std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
+                  << problem_size.Ks[i] << std::endl;
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_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_tile_loop_multply.hpp"
+
+#include "ck/host_utility/hip_check_error.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 B0DataType       = I8;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<B1DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<B1Layout>;
+using ELayout  = Row;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Multiply;
+
+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::GemmDesc> 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 NumDTensor = 1;
+
+    using GroupedGemmKernelArgument =
+        ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<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) * problem_size.Ms[i] * 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) * problem_size.Ms[i] * problem_size.Ns[i]));
+
+        gemm_descs.push_back({problem_size.Ms[i],
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              problem_size.stride_As[i],
+                              problem_size.stride_Bs[i],
+                              problem_size.stride_Cs[i],
+                              {0}});
+
+        grouped_gemm_kernel_args_.push_back({a0_tensors_device[i]->GetDeviceBuffer(),
+                                             b0_tensors_device[i]->GetDeviceBuffer(),
+                                             {b1_tensors_device[i]->GetDeviceBuffer()},
+                                             c_tensors_device[i]->GetDeviceBuffer(),
+                                             problem_size.Ms[i],
+                                             problem_size.Ns[i],
+                                             problem_size.Ks[i],
+                                             problem_size.stride_As[i],
+                                             problem_size.stride_Bs[i],
+                                             {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::DeviceGroupedGemmTileLoop<A0Layout,
+                                                                             B0Layout,
+                                                                             DsLayout,
+                                                                             ELayout,
+                                                                             A0DataType,
+                                                                             B0DataType,
+                                                                             DsDataType,
+                                                                             EDataType,
+                                                                             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<const void*> p_As                         = {};
+        std::vector<const void*> 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, 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()))
+        {
+            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());
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
+
+            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(1 + rand() % 1024);
+        problem_size.Ns.push_back(4096);
+        problem_size.Ks.push_back(4096);
+
+        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]);
+
+        std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
+                  << problem_size.Ks[i] << std::endl;
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_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"
+
+#include "ck/host_utility/hip_check_error.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 = Row;
+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 Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+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_;
+};
+
+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, 0, 20, 50});
+
+            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(1 + rand() % 1024);
+        problem_size.Ns.push_back(4096);
+        problem_size.Ks.push_back(4096);
+
+        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]);
+
+        std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
+                  << problem_size.Ks[i] << std::endl;
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

client_example/CMakeLists.txt

@@ -48,6 +48,21 @@ else()
     endif()
 endif()
 
+if (GPU_TARGETS)
+    if (GPU_TARGETS MATCHES "gfx9")
+        add_definitions(-DCK_USE_XDL)
+        set(CK_USE_XDL "ON")
+    endif()
+    if (GPU_TARGETS MATCHES "gfx11")
+        add_definitions(-DCK_USE_WMMA)
+        set(CK_USE_WMMA "ON")
+    endif()
+else()
+    add_definitions(-DCK_USE_WMMA -DCK_USE_XDL)
+    set(CK_USE_XDL "ON")
+    set(CK_USE_WMMA "ON")
+endif()
+
 find_package(composable_kernel COMPONENTS device_other_operations device_gemm_operations device_conv_operations  device_reduction_operations)
 if(GPU_TARGETS MATCHES "gfx9")
     find_package(composable_kernel COMPONENTS device_contraction_operations)

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.

docs/sphinx/requirements.in

-rocm-docs-core==0.38.1
+rocm-docs-core==1.1.2
 sphinxcontrib-bibtex==2.6.2

docs/sphinx/requirements.txt

 #
-# This file is autogenerated by pip-compile with Python 3.8
+# This file is autogenerated by pip-compile with Python 3.10
 # by the following command:
 #
 #    pip-compile requirements.in
@@ -48,12 +48,6 @@ idna==3.4
     # via requests
 imagesize==1.4.1
     # via sphinx
-importlib-metadata==6.8.0
-    # via
-    #   sphinx
-    #   sphinxcontrib-bibtex
-importlib-resources==6.1.0
-    # via rocm-docs-core
 jinja2==3.1.2
     # via
     #   myst-parser
@@ -99,8 +93,6 @@ pyjwt[crypto]==2.6.0
     # via pygithub
 pynacl==1.5.0
     # via pygithub
-pytz==2023.3.post1
-    # via babel
 pyyaml==6.0
     # via
     #   myst-parser
@@ -111,7 +103,7 @@ requests==2.31.0
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core==0.38.1
+rocm-docs-core==1.1.2
     # via -r requirements.in
 six==1.16.0
     # via
@@ -165,7 +157,3 @@ urllib3==1.26.18
     # via requests
 wrapt==1.15.0
     # via deprecated
-zipp==3.17.0
-    # via
-    #   importlib-metadata
-    #   importlib-resources

example/01_gemm/CMakeLists.txt

@@ -22,6 +22,15 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
 add_example_executable(example_gemm_xdl_fp16_v2 gemm_xdl_fp16_v2.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v2)
 
+add_example_executable(example_gemm_xdl_fp16_v3 gemm_xdl_fp16_v3.cpp)
+add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v3)
+add_example_executable(example_gemm_xdl_fp8_v3 gemm_xdl_fp8_v3.cpp)
+add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_v3)
+add_example_executable(example_gemm_xdl_fp16_fp8_v3 gemm_xdl_fp16_fp8_v3.cpp)
+add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_v3)
+add_example_executable(example_gemm_xdl_bf16_v3 gemm_xdl_bf16_v3.cpp)
+add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_v3)
+
 add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)