add

client_example/30_gemm_bf16Aint8B_add_fastgelu/CMakeLists.txt

@@ -5,6 +5,9 @@ if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "int8" AND DTYPES MATCHES "bf
 	add_executable(client_gemm_multiply_add_fastgelu_xdl_bf16_i8 gemm_multiply_add_fastgelu_xdl_bf16_i8.cpp)
 	target_link_libraries(client_gemm_multiply_add_fastgelu_xdl_bf16_i8 PRIVATE composable_kernel::device_gemm_operations)
 
+	add_executable(client_gemm_multiply_xdl_bf16_i8 gemm_multiply_xdl_bf16_i8.cpp)
+	target_link_libraries(client_gemm_multiply_xdl_bf16_i8 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)
 

client_example/30_gemm_bf16Aint8B_add_fastgelu/gemm_multiply_add_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_multiply_add_fastgelu.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, D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using B0Layout = Row;
+using B1Layout = Row;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<B1Layout, 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_;
+};
+
+// 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 = 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 NumDTensor = 2;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<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];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a0_device_buf.GetDeviceBuffer(),
+            b0_device_buf.GetDeviceBuffer(),
+            std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer(),
+                                                d0_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, NumDTensor>{0, 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;
+
+    return 0;
+}

client_example/30_gemm_bf16Aint8B_add_fastgelu/gemm_multiply_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_multiply.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 DsDataType       = ck::Tuple<B1DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using B0Layout = Row;
+using B1Layout = Row;
+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_;
+};
+
+// 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 = 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 NumDTensor = 1;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<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];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a0_device_buf.GetDeviceBuffer(),
+            b0_device_buf.GetDeviceBuffer(),
+            std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            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, 0, 20, 50});
+
+            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;
+}

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp

@@ -65,7 +65,9 @@ template <typename ALayout,
           BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
           BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
           typename ComputeTypeA                       = CDataType,
-          typename ComputeTypeB                       = ComputeTypeA>
+          typename ComputeTypeB                       = ComputeTypeA,
+          typename LDSTypeA                           = ComputeTypeA,
+          typename LDSTypeB                           = ComputeTypeB>
 struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD<ALayout,
                                                                      BLayout,
                                                                      DsLayout,
@@ -128,7 +130,9 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD<ALayout,
         BlkGemmPipeSched,
         BlkGemmPipelineVer,
         ComputeTypeA,
-        ComputeTypeB>;
+        ComputeTypeB,
+        LDSTypeA,
+        LDSTypeB>;
 
     using Argument = typename GridwiseGemm::Argument;
 

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -32,6 +32,7 @@ using F16_F16_Tuple = ck::Tuple<F16, F16>;
 
 using F64_Tuple     = ck::Tuple<F64>;
 using F32_Tuple     = ck::Tuple<F32>;
+using BF16_Tuple    = ck::Tuple<BF16>;
 using I32_Tuple     = ck::Tuple<I32>;
 using I32_F32_Tuple = ck::Tuple<I32, F32>;
 using I8_Tuple      = ck::Tuple<I8>;
@@ -99,6 +100,7 @@ using Scale               = ck::tensor_operation::element_wise::Scale;
 using Bilinear            = ck::tensor_operation::element_wise::Bilinear;
 using AddAddFastGelu      = ck::tensor_operation::element_wise::AddAddFastGelu;
 using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
+using Multiply            = ck::tensor_operation::element_wise::Multiply;
 using AddFastGelu         = ck::tensor_operation::element_wise::AddFastGelu;
 using AddRelu             = ck::tensor_operation::element_wise::AddRelu;
 using AddSilu             = ck::tensor_operation::element_wise::AddSilu;

library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_multiply_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    BF16,
+                                                    I8,
+                                                    BF16_Tuple,
+                                                    BF16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Multiply>>>&);
+
+// GEMM + Multiply + Add + FastGelu
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::Multiply>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::Multiply>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
+                     is_same_v<D0DataType, bhalf_t> && is_same_v<EDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_multiply_mnkpadding_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt

@@ -66,6 +66,7 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES
         device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
         device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instance.cpp
 
+	device_gemm_xdl_universal_bf16_i8_bf16_multi_d/device_gemm_xdl_universal_multiply_bf16_i8_bf16_mk_kn_mn_mnkpadding_instance.cpp
 	device_gemm_xdl_universal_bf16_i8_bf16_multi_d/device_gemm_xdl_universal_multiply_add_fastgelu_bf16_i8_bf16_mk_kn_mn_mnkpadding_instance.cpp
         )
 

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_i8_bf16_multi_d/device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn.hpp

@@ -26,6 +26,7 @@ using S = Sequence<Is...>;
 
 using PassThrough         = element_wise::PassThrough;
 using MultiplyAddFastGelu = element_wise::MultiplyAddFastGelu;
+using Multiply            = element_wise::Multiply;
 
 static constexpr auto GemmDefault    = GemmSpecialization::Default;
 static constexpr auto GemmKPadding   = GemmSpecialization::KPadding;
@@ -35,9 +36,7 @@ static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
 static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
 static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
 
-using DsLayout = ck::Tuple<Row, Row>;
-
-template <typename DsDType, typename CElementwiseOp, GemmSpecialization GemmSpec>
+template <typename DsLayout, typename DsDType, typename CElementwiseOp, GemmSpecialization GemmSpec>
 using device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_comp_instances = std::tuple<
     // clang-format off
         //#########################| ALayout| BLayout| CLayout|AData| BData|  DsData| CData| AccData| Cshuffle|           A|           B|               C|           GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
@@ -56,7 +55,8 @@ using device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_comp_instances = s
     // clang-format on
     >;
 
-template <typename DsDType,
+template <typename DsLayout,
+          typename DsDType,
           typename CElementwiseOp,
           GemmSpecialization GemmSpec,
           BlockGemmPipelineScheduler BlkGemmPipeSched>
@@ -68,17 +68,17 @@ using device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_mem_instances = st
         //#########################|        |        |        |     |      |        |      |        |         |            |            |               |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                  |                             |
 
    // Latency friendly
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,    64,    16,   16,   256,   8,   4,  16,   16,    1,    1,     S<32, 2, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 1, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    16,   32,   256,   8,   4,  16,   16,    1,    1,     S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 2, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,    64,    16,   16,   256,   8,   4,  16,   16,    1,    1,     S<32, 2, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 1, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, BF16, BF16, BF16, I8>,
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    16,   32,   256,   8,   4,  16,   16,    1,    1,     S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 2, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1, BF16, BF16, BF16, I8>,
    // Memory friendly
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,    64,    16,   16,   256,   8,   4,  16,   16,    1,    1,     S<32, 2, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 1, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    16,   32,   256,   8,   4,  16,   16,    1,    1,     S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 2, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    16,   64,   128,   8,   4,  16,   16,    1,    2,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    32,   64,   128,   8,   4,  32,   32,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    16,  128,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    32,  128,    64,   8,   4,  32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   256,    16,  256,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
-   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   256,    32,  256,    64,   8,   4,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,    64,    16,   16,   256,   8,   4,  16,   16,    1,    1,     S<32, 2, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 1, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, BF16, BF16, BF16, I8>,
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    16,   32,   256,   8,   4,  16,   16,    1,    1,     S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 2, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, BF16, BF16, BF16, I8>,
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    16,   64,   128,   8,   4,  16,   16,    1,    2,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, BF16, BF16, BF16, I8>,
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    32,   64,   128,   8,   4,  32,   32,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, BF16, BF16, BF16, I8>,
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    16,  128,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, BF16, BF16, BF16, I8>,
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   128,    32,  128,    64,   8,   4,  32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, BF16, BF16, BF16, I8>,
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   256,    16,  256,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, BF16, BF16, BF16, I8>,
+   DeviceGemmMultiD_Xdl_CShuffle_V3<  Row,     Row,  DsLayout,   Row,     BF16,   I8,    DsDType, BF16,   F32,     F32,      PassThrough, PassThrough, CElementwiseOp,       GemmSpec,   256,    32,  256,    64,   8,   4,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2, BF16, BF16, BF16, I8>
     // clang-format on
     >;
 } // namespace instance

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_i8_bf16_multi_d/device_gemm_xdl_universal_multiply_add_fastgelu_bf16_i8_bf16_mk_kn_mn_mnkpadding_instance.cpp

@@ -24,13 +24,15 @@ void add_device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_multiply_add_fa
     add_device_operation_instances(
         instances,
         device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_comp_instances<
+            ck::Tuple<Row, Row>,
             ck::Tuple<BF16, BF16>,
             MultiplyAddFastGelu,
             GemmMNKPadding>{});
 
     add_device_operation_instances(
         instances,
-        device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<BF16, BF16>,
+        device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row, Row>,
+                                                                              ck::Tuple<BF16, BF16>,
                                                                               MultiplyAddFastGelu,
                                                                               GemmMNKPadding,
                                                                               Intrawave>{});

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_i8_bf16_multi_d/device_gemm_xdl_universal_multiply_bf16_i8_bf16_mk_kn_mn_mnkpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_multiply_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    ck::Tuple<Row>,
+                                                    Row,
+                                                    BF16,
+                                                    I8,
+                                                    ck::Tuple<BF16>,
+                                                    BF16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Multiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_comp_instances<ck::Tuple<Row>,
+                                                                               ck::Tuple<BF16>,
+                                                                               Multiply,
+                                                                               GemmMNKPadding>{});
+
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_multi_d_bf16_i8_bf16_mk_kn_mn_mem_instances<ck::Tuple<Row>,
+                                                                              ck::Tuple<BF16>,
+                                                                              Multiply,
+                                                                              GemmMNKPadding,
+                                                                              Intrawave>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck