Contraction multi abd (#957)

* add gridwise_multi_abd

* move element_op into RunRead

* merge element_wise op with data read

* add multiABD example

* allow packed elementwise_op

* changed example

* clean

* clean

* add is_detected

* fix

* minor fix

* add scaleAdd_vec4 example

* init commit for contraction_multi_ABD

* add examples

* add examples of multiA and broadcast

* update example

* fixed comments

* Update cmake-ck-dev.sh

* Update cmake-ck-dev.sh

* Add comments into the example

---------
Co-authored-by: Jing Zhang <jizha@amd.com>

example/60_gemm_multi_ABD/CMakeLists.txt

+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
+     add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
+   set(target 1)
+ endif()
+endforeach()
+endif()

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F16;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using DLayout = Row;
+using ELayout = Row;
+
+struct AddScale
+{
+    static constexpr auto I0 = ck::Number<0>{};
+    static constexpr auto I1 = ck::Number<1>{};
+    static constexpr auto I2 = ck::Number<2>{};
+    static constexpr auto I3 = ck::Number<3>{};
+
+    __host__ __device__ constexpr void
+    operator()(ck::half4_t& a, const ck::half4_t& a0, const ck::half4_t& a1) const
+    {
+        const auto a0_v_t = ck::vector_type<ck::half_t, 4>{a0};
+        const auto a1_v_t = ck::vector_type<ck::half_t, 4>{a1};
+
+        auto r_v_t = ck::vector_type<ck::half_t, 4>{};
+
+        r_v_t.AsType<ck::half_t>()(I0) =
+            scale * (a0_v_t.AsType<ck::half_t>()[I0] + a1_v_t.AsType<ck::half_t>()[I0]);
+        r_v_t.AsType<ck::half_t>()(I1) =
+            scale * (a0_v_t.AsType<ck::half_t>()[I1] + a1_v_t.AsType<ck::half_t>()[I1]);
+        r_v_t.AsType<ck::half_t>()(I2) =
+            scale * (a0_v_t.AsType<ck::half_t>()[I2] + a1_v_t.AsType<ck::half_t>()[I2]);
+        r_v_t.AsType<ck::half_t>()(I3) =
+            scale * (a0_v_t.AsType<ck::half_t>()[I3] + a1_v_t.AsType<ck::half_t>()[I3]);
+
+        a = r_v_t.AsType<ck::half4_t>()[I0];
+    }
+
+    __host__ __device__ constexpr void
+    operator()(ck::half_t& a, const ck::half_t& a0, const ck::half_t& a1) const
+    {
+        a = scale * (a0 + a1);
+    }
+
+    //this attribute will force copy_function applying element_wise with vector_type
+    static constexpr ck::index_t vec_len = 4;
+
+    float scale = 1.0;
+};
+
+struct AlphaBetaAdd
+{
+    AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
+        ck::half_t& e, const float& c, const ck::half_t& d) const
+    {
+        e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
+    };
+
+    float alpha_;
+    float beta_;
+};
+
+using AElementOp   = AddScale;
+using BElementOp   = PassThrough;
+using CDEElementOp = AlphaBetaAdd;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle<
+    ck::Tuple<ALayout, ALayout>,
+    ck::Tuple<BLayout>,
+    ck::Tuple<DLayout>,
+    ELayout,
+    ck::Tuple<ADataType, ADataType>,
+    ck::Tuple<BDataType>,
+    AccDataType,
+    CShuffleDataType,
+    ck::Tuple<DDataType>,
+    EDataType,
+    AElementOp,
+    BElementOp,
+    CDEElementOp,
+    GemmSpec,
+    1,
+    256,
+    256,
+    128,
+    32,
+    8,
+    8,
+    32,
+    32,
+    4,
+    2,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    1,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    1,
+    1,
+    1,
+    S<1, 32, 1, 8>,
+    8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideD = 4096;
+    ck::index_t StrideE = 4096;
+
+    float alpha = 1.0f;
+    float beta  = 1.0f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        alpha = std::stof(argv[4]);
+        beta  = std::stof(argv[5]);
+    }
+    else if(argc == 13)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+
+        alpha = std::stof(argv[11]);
+        beta  = std::stof(argv[12]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
+               "beta\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<ADataType> a1_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        a1_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        a1_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(ADataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem a1_device_buf(sizeof(ADataType) * a1_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    a1_device_buf.ToDevice(a1_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{0.2};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
+                                                          a1_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, 1>{b_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               std::array<ck::index_t, 2>{StrideA, StrideA},
+                               std::array<ck::index_t, 1>{StrideB},
+                               std::array<ck::index_t, 1>{StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        Tensor<ADataType> a_m_k({M, K});
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                a_element_op(a_m_k(m, k), a0_m_k(m, k), a1_m_k(m, k));
+            }
+        }
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/61_contraction_multi_ABD/CMakeLists.txt

+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
+     add_example_executable(example_contraction_multi_ABD_xdl_fp16 contraction_multi_ABD_xdl_fp16.cpp)
+   set(target 1)
+ endif()
+endforeach()
+endif()

example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using A0DataType       = F16;
+using A1DataType       = F32;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F16;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct AlphaBetaAdd
+{
+    AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
+        ck::half_t& e, const float& c, const ck::half_t& d) const
+    {
+        e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
+    };
+
+    float alpha_;
+    float beta_;
+};
+
+struct Multiply
+{
+    __host__ __device__ constexpr void
+    operator()(ck::half_t& a, const ck::half_t& a0, const float& a1) const
+    {
+        a = ck::type_convert<ck::half_t>(ck::type_convert<float>(a0) * a1);
+    }
+};
+
+using AElementOp   = Multiply;
+using BElementOp   = PassThrough;
+using CDEElementOp = AlphaBetaAdd;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceContractionMultipleABD_Xdl_CShuffle<
+    NumDimM,
+    NumDimN,
+    NumDimK,
+    ck::Tuple<A0DataType, A1DataType>,
+    ck::Tuple<BDataType>,
+    AccDataType,
+    CShuffleDataType,
+    ck::Tuple<DDataType>,
+    EDataType,
+    AElementOp,
+    BElementOp,
+    CDEElementOp,
+    GemmSpec,
+    1,
+    256,
+    256,
+    128,
+    32,
+    8,
+    8,
+    32,
+    32,
+    4,
+    2,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    1,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    1,
+    1,
+    1,
+    S<1, 32, 1, 8>,
+    8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    float alpha = 1.0f;
+    float beta  = 1.0f;
+
+    // A0[M0, M1, K0, K1]
+    std::vector<ck::index_t> a0_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a0_ms_ks_strides{128 * 32 * 64, 32 * 64, 64, 1};
+    // A1[M1, K1] -> A1[M0, M1, K0, K1]
+    std::vector<ck::index_t> a1_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 0, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<A0DataType> a0_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+    Tensor<A1DataType> a1_ms_ks(a1_ms_ks_lengths, a1_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a0_ms_ks: " << a0_ms_ks.mDesc << std::endl;
+    std::cout << "a1_ms_ks: " << a1_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem a1_device_buf(sizeof(A1DataType) * a1_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_ms_ks.mData.data());
+    a1_device_buf.ToDevice(a1_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument  = device_op.MakeArgument(
+        std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
+                                   a1_device_buf.GetDeviceBuffer()},
+        std::array<const void*, 1>{b_device_buf.GetDeviceBuffer()},
+        std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+        e_device_buf.GetDeviceBuffer(),
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_lengths, a1_ms_ks_lengths},
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_strides, a1_ms_ks_strides},
+        std::array<std::vector<ck::index_t>, 1>{b_ns_ks_lengths},
+        std::array<std::vector<ck::index_t>, 1>{b_ns_ks_strides},
+        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+        e_ms_ns_lengths,
+        e_ms_ns_strides,
+        a_element_op,
+        b_element_op,
+        cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_contraction with the specified compilation parameters does "
+            "not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    if(time_kernel)
+    {
+        ck::index_t M =
+            ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+        ck::index_t N = ck::accumulate_n<ck::index_t>(
+            e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+        ck::index_t K = ck::accumulate_n<ck::index_t>(
+            a0_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+        std::size_t flop = std::size_t(2) * M * N * K;
+        std::size_t num_btype =
+            sizeof(A0DataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * M * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s" << std::endl;
+    }
+
+    if(do_verification)
+    {
+
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        Tensor<A0DataType> a_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+
+        for(size_t m0 = 0; m0 < a_ms_ks.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < a_ms_ks.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t k0 = 0; k0 < a_ms_ks.mDesc.GetLengths()[2]; ++k0)
+                {
+                    for(size_t k1 = 0; k1 < a_ms_ks.mDesc.GetLengths()[3]; ++k1)
+                    {
+                        a_element_op(a_ms_ks(m0, m1, k0, k1),
+                                     a0_ms_ks(m0, m1, k0, k1),
+                                     a1_ms_ks(m0, m1, k0, k1));
+                    }
+                }
+            }
+        }
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      A0DataType,
+                                                                      BDataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      PassThrough,
+                                                                      BElementOp>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
+        auto ref_argument =
+            ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, PassThrough{}, b_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1),
+                                       d_ms_ns(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A0[M0, M1, ... K0, K1, ...], ...
+//   input : B0[N0, N1, ... K0, K1, ...], ...
+//   input : D0[M0, M1, ... N0, N1, ...], D1[M0, M1, ... N0, N1, ...], ...
+//   output : E[M0, M1, ... N0, N1, ...]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename AsDataType,
+          typename BsDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceContractionMultipleABD : public BaseOperator
+{
+    static constexpr index_t NumATensor = AsDataType::Size();
+    static constexpr index_t NumBTensor = BsDataType::Size();
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::array<const void*, NumATensor> p_as,
+                        std::array<const void*, NumBTensor> p_bs,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_lengths,
+                        const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_strides,
+                        const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_lengths,
+                        const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_strides,
+                        const std::vector<index_t>& e_ms_ns_length,
+                        const std::vector<index_t>& e_ms_ns_stride,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename AsPointer,
+          typename BsPointer,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AsGridDesc_AK0_M_AK1,
+          typename BsGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_multiple_abd_xdl_cshuffle(
+            AsPointer p_as_grid,
+            BsPointer p_bs_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AsGridDesc_AK0_M_AK1 as_grid_desc_ak0_m_ak1,
+            const BsGridDesc_BK0_N_BK1 bs_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_as_grid,
+                                                  p_bs_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  as_grid_desc_ak0_m_ak1,
+                                                  bs_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_as_grid;
+    ignore = p_bs_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = as_grid_desc_ak0_m_ak1;
+    ignore = bs_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename AsDataType,
+          typename BsDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct DeviceContractionMultipleABD_Xdl_CShuffle
+    : public DeviceContractionMultipleABD<NumDimM,
+                                          NumDimN,
+                                          NumDimK,
+                                          AsDataType,
+                                          BsDataType,
+                                          DsDataType,
+                                          EDataType,
+                                          AElementwiseOperation,
+                                          BElementwiseOperation,
+                                          CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceContractionMultipleABD_Xdl_CShuffle;
+
+    static constexpr index_t NumATensor = AsDataType::Size();
+    static constexpr index_t NumBTensor = BsDataType::Size();
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ComputeDataType = EDataType;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleABD_xdl_cshuffle<
+        AsDataType,
+        BsDataType,
+        ComputeDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        PipelineVer>;
+
+    static constexpr auto matrix_padder =
+        ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_ms_ks_lengths_,
+                                        const std::vector<index_t>& a_ms_ks_strides_)
+    {
+        assert(a_ms_ks_lengths_.size() == NumDimM + NumDimK &&
+               a_ms_ks_strides_.size() == NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(a_ms_ks_lengths_, Number<NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(a_ms_ks_strides_, Number<NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+        const auto a_grid_desc_ms_ks =
+            make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+        // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+        const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+            a_grid_desc_ms_ks,
+            make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+            make_tuple(mDimIds, kDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    __host__ __device__ static auto
+    MakeAsGridDescriptor_M_K(const std::array<std::vector<index_t>, NumATensor>& as_ms_ks_lengths,
+                             const std::array<std::vector<index_t>, NumATensor>& as_ms_ks_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeAGridDescriptor_M_K(as_ms_ks_lengths[i], as_ms_ks_strides[i]);
+            },
+            Number<NumATensor>{});
+    }
+
+    // Assume: B[N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_ns_ks_lengths_,
+                                        const std::vector<index_t>& b_ns_ks_strides_)
+    {
+        assert(b_ns_ks_lengths_.size() == NumDimN + NumDimK &&
+               b_ns_ks_strides_.size() == NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(b_ns_ks_lengths_, Number<NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(b_ns_ks_strides_, Number<NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+        const auto b_grid_desc_ns_ks =
+            make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+        // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+        const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+            b_grid_desc_ns_ks,
+            make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+            make_tuple(nDimIds, kDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    __host__ __device__ static auto
+    MakeBsGridDescriptor_N_K(const std::array<std::vector<index_t>, NumBTensor>& bs_ns_ks_lengths,
+                             const std::array<std::vector<index_t>, NumBTensor>& bs_ns_ks_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeBGridDescriptor_N_K(bs_ns_ks_lengths[i], bs_ns_ks_strides[i]);
+            },
+            Number<NumBTensor>{});
+    }
+
+    // assume E[M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_ms_ns_lengths_,
+                                        const std::vector<index_t>& e_ms_ns_strides_)
+    {
+        assert(e_ms_ns_lengths_.size() == NumDimM + NumDimN &&
+               e_ms_ns_strides_.size() == NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(e_ms_ns_lengths_, Number<NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(e_ms_ns_strides_, Number<NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+        const auto e_grid_desc_ms_ns =
+            make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+        // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+        const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+            e_grid_desc_ms_ns,
+            make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+            make_tuple(mDimIds, nDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto
+    MakeDsGridDescriptor_M_N(const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                             const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_M_N(ds_ms_ns_lengths[i], ds_ms_ns_strides[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AsGridDesc_M_K = remove_cvref_t<decltype(MakeAsGridDescriptor_M_K({}, {}))>;
+    using BsGridDesc_N_K = remove_cvref_t<decltype(MakeBsGridDescriptor_N_K({}, {}))>;
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N({}, {}))>;
+
+    // desc for blockwise copy
+    using AsGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeAsGridDescriptor_AK0_M_AK1(AsGridDesc_M_K{}))>;
+    using BsGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeBsGridDescriptor_BK0_N_BK1(BsGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(std::array<const void*, NumATensor> p_as_grid,
+                 std::array<const void*, NumBTensor> p_bs_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_lengths,
+                 const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_strides,
+                 const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_lengths,
+                 const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_strides,
+                 const std::vector<index_t>& e_ms_ns_length,
+                 const std::vector<index_t>& e_ms_ns_stride,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_as_grid_{},
+              p_bs_grid_{},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              as_grid_desc_m_k_{},
+              bs_grid_desc_n_k_{},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{MakeEGridDescriptor_M_N(e_ms_ns_length, e_ms_ns_stride)},
+              as_grid_desc_ak0_m_ak1_{},
+              bs_grid_desc_bk0_n_bk1_{},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            // populate pointer, desc for As
+            static_for<0, NumATensor, 1>{}([&](auto i) {
+                // using ALayout   = remove_cvref_t<tuple_element_t<i.value, AsLayout>>;
+                using ADataType = remove_cvref_t<tuple_element_t<i.value, AsDataType>>;
+
+                // A pointer
+                p_as_grid_(i) = static_cast<const ADataType*>(p_as_grid[i]);
+
+                // A desc
+                as_grid_desc_m_k_(i) =
+                    MakeAGridDescriptor_M_K(a_ms_ks_lengths[i], a_ms_ks_strides[i]);
+            });
+
+            // populate pointer, desc for Bs
+            static_for<0, NumBTensor, 1>{}([&](auto i) {
+                // using BLayout   = remove_cvref_t<tuple_element_t<i.value, BsLayout>>;
+                using BDataType = remove_cvref_t<tuple_element_t<i.value, BsDataType>>;
+
+                // B pointer
+                p_bs_grid_(i) = static_cast<const BDataType*>(p_bs_grid[i]);
+
+                // B desc
+                bs_grid_desc_n_k_(i) =
+                    MakeBGridDescriptor_N_K(b_ns_ks_lengths[i], b_ns_ks_strides[i]);
+            });
+
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                // using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    MakeEGridDescriptor_M_N(d_ms_ns_lengths[i], d_ms_ns_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(as_grid_desc_m_k_,
+                                           bs_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                as_grid_desc_ak0_m_ak1_ =
+                    GridwiseGemm::MakeAsGridDescriptor_AK0_M_AK1(as_grid_desc_m_k_);
+
+                bs_grid_desc_bk0_n_bk1_ =
+                    GridwiseGemm::MakeBsGridDescriptor_BK0_N_BK1(bs_grid_desc_n_k_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+            }
+
+            // for sanity check of vector memory access
+            for(index_t i = 0; i < NumATensor; ++i)
+            {
+                a_mz_stride_[i] = a_ms_ks_strides[i][NumDimM - 1];
+                a_kz_stride_[i] = a_ms_ks_strides[i][NumDimM + NumDimK - 1];
+            }
+
+            for(index_t i = 0; i < NumBTensor; ++i)
+            {
+                b_nz_stride_[i] = b_ns_ks_strides[i][NumDimN - 1];
+                b_kz_stride_[i] = b_ns_ks_strides[i][NumDimN + NumDimK - 1];
+            }
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = d_ms_ns_strides[i][NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_ms_ns_stride[NumDimM + NumDimN - 1];
+        }
+
+        // pointers
+        typename GridwiseGemm::AsGridPointer p_as_grid_;
+        typename GridwiseGemm::BsGridPointer p_bs_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AsGridDesc_M_K as_grid_desc_m_k_;
+        BsGridDesc_N_K bs_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AsGridDesc_AK0_M_AK1 as_grid_desc_ak0_m_ak1_;
+        BsGridDesc_BK0_N_BK1 bs_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        std::array<index_t, NumATensor> a_mz_stride_;
+        std::array<index_t, NumATensor> a_kz_stride_;
+
+        std::array<index_t, NumBTensor> b_nz_stride_;
+        std::array<index_t, NumBTensor> b_kz_stride_;
+
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        index_t e_nz_stride_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.as_grid_desc_m_k_,
+                                            arg.bs_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_abd_xdl_cshuffle<
+                    GridwiseGemm,
+                    typename GridwiseGemm::AsGridPointer,
+                    typename GridwiseGemm::BsGridPointer,
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AsGridDesc_AK0_M_AK1,
+                    DeviceOp::BsGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::Block2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_as_grid_,
+                                              arg.p_bs_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.as_grid_desc_ak0_m_ak1_,
+                                              arg.bs_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_);
+            };
+
+            const auto K = arg.as_grid_desc_m_k_[I0].GetLength(I1);
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        // check vector load/store
+        {
+            bool all_valid = true;
+
+            static_for<0, NumATensor, 1>{}([&](auto i) {
+                // vector memory access of A: could be on M or AK1 dimension
+                if constexpr(ABlockTransferSrcVectorDim == 1)
+                {
+                    if(!(arg.a_mz_stride_[i] == 1 && arg.as_grid_desc_ak0_m_ak1_[i].GetLength(I1) %
+                                                             ABlockTransferSrcScalarPerVector ==
+                                                         0))
+                    {
+                        all_valid = false;
+                    }
+                }
+                else
+                {
+                    if(!(arg.a_kz_stride_[i] == 1 && arg.as_grid_desc_ak0_m_ak1_[i].GetLength(I2) %
+                                                             ABlockTransferSrcScalarPerVector ==
+                                                         0))
+                    {
+                        all_valid = false;
+                    }
+                }
+            });
+
+            // vector memory access of B: could be on N or BK1 dimension
+            static_for<0, NumBTensor, 1>{}([&](auto i) {
+                if constexpr(BBlockTransferSrcVectorDim == 1)
+                {
+                    if(!(arg.b_nz_stride_[i] == 1 && arg.bs_grid_desc_bk0_n_bk1_[i].GetLength(I1) %
+                                                             BBlockTransferSrcScalarPerVector ==
+                                                         0))
+                    {
+                        all_valid = false;
+                    }
+                }
+                else
+                {
+                    if(!(arg.b_kz_stride_[i] == 1 && arg.bs_grid_desc_bk0_n_bk1_[i].GetLength(I2) %
+                                                             BBlockTransferSrcScalarPerVector ==
+                                                         0))
+                    {
+                        all_valid = false;
+                    }
+                }
+            });
+
+            // check vector load of Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                if(!(arg.ds_nz_stride_[i] == 1 &&
+                     arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
+                             CDEBlockTransferScalarPerVector_NPerBlock ==
+                         0))
+                {
+                    all_valid = false;
+                }
+            });
+
+            // vector memory access of E: always on NPerBlock dimension
+            if(!(arg.e_nz_stride_ == 1 &&
+                 arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                         CDEBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                all_valid = false;
+            }
+
+            if(!all_valid)
+            {
+                return false;
+            }
+        }
+
+        return GridwiseGemm::CheckValidity(arg.as_grid_desc_m_k_,
+                                           arg.bs_grid_desc_n_k_,
+                                           arg.ds_grid_desc_m_n_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::array<const void*, NumATensor> p_as,
+                             std::array<const void*, NumBTensor> p_bs,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_lengths,
+                             const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_strides,
+                             const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_lengths,
+                             const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_strides,
+                             const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_lengths,
+                             const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_strides,
+                             const std::vector<index_t>& e_ms_ns_length,
+                             const std::vector<index_t>& e_ms_ns_stride,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_as,
+                        p_bs,
+                        p_ds,
+                        p_e,
+                        a_ms_ks_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        d_ms_ns_lengths,
+                        d_ms_ns_strides,
+                        e_ms_ns_length,
+                        e_ms_ns_stride,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::array<const void*, NumATensor> p_as,
+                        std::array<const void*, NumBTensor> p_bs,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::array<std::vector<index_t>, NumATensor>& as_ms_ks_lengths,
+                        const std::array<std::vector<index_t>, NumATensor>& as_ms_ks_strides,
+                        const std::array<std::vector<index_t>, NumBTensor>& bs_ns_ks_lengths,
+                        const std::array<std::vector<index_t>, NumBTensor>& bs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides,
+                        const std::vector<index_t>& e_ms_ns_length,
+                        const std::vector<index_t>& e_ms_ns_stride,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_as,
+                                          p_bs,
+                                          p_ds,
+                                          p_e,
+                                          as_ms_ks_lengths,
+                                          as_ms_ks_strides,
+                                          bs_ns_ks_lengths,
+                                          bs_ns_ks_strides,
+                                          ds_ms_ns_lengths,
+                                          ds_ms_ns_strides,
+                                          e_ms_ns_length,
+                                          e_ms_ns_stride,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceContractionMultipleABD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">"
+            << " LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp

@@ -428,7 +428,7 @@ struct GridwiseGemmMultipleABD_xdl_cshuffle
             [&](auto i) {
                 using ALayout = remove_cvref_t<tuple_element_t<i.value, AsLayout>>;
 
-                return MakeEGridDescriptor_M_N<ALayout, GemmSpec>(MRaws[i], KRaws[i], AsStride[i]);
+                return MakeAGridDescriptor_M_K<ALayout, GemmSpec>(MRaws[i], KRaws[i], AsStride[i]);
             },
             Number<NumATensor>{});
     }