Add multiple A/B support (#906)

* 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

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

example/60_gemm_multiABD/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_multiABD_xdl_fp16 gemm_multiABD_xdl_fp16.cpp)
+   set(target 1)
+ endif()
+endforeach()
+endif()

example/60_gemm_multiABD/gemm_multiABD_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);
+    }
+
+    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;
+}

include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp"
+#include "ck/utility/is_detected.hpp"
+
+namespace ck {
+
+// Thread-group level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//
+// Does following things to avoid scratch memory issue
+//   1. Pass tensor descritpors by reference (or tuple of references)
+//   2. Does not keep reference to tensor descriptor
+//   3. Does not construct new tensor coordinate when call Run()
+template <typename ThreadGroup,
+          typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          typename ThreadTransferSrcResetCoordinateAfterRunFlags,
+          typename ThreadTransferDstResetCoordinateAfterRunFlags>
+struct ThreadGroupTensorSliceTransfer_v7r2
+{
+    static constexpr index_t nDim =
+        remove_cvref_t<tuple_element_t<0, SrcDescs>>::GetNumOfDimension();
+
+    static constexpr index_t nSrc = remove_cvref_t<SrcDescs>::Size();
+    static constexpr index_t nDst = remove_cvref_t<DstDescs>::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v7r2(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_block_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_block_slice_origins,
+        const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src_descs,
+                               StaticallyIndexedArray<Index, nSrc>{},
+                               dst_descs,
+                               StaticallyIndexedArray<Index, nDst>{},
+                               element_op)
+    {
+        static_assert(nSrc == SrcDatas::Size() && nSrc == SrcDescs::Size() &&
+                          nSrc == ThreadTransferSrcResetCoordinateAfterRunFlags::Size() &&
+                          nDst == DstDatas::Size() && nDst == DstDescs::Size() &&
+                          nDst == ThreadTransferDstResetCoordinateAfterRunFlags::Size(),
+                      "wrong!");
+
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            static_assert(
+                nDim == remove_cvref_t<tuple_element_t<i.value, SrcDescs>>::GetNumOfDimension(),
+                "wrong!");
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            static_assert(
+                nDim == remove_cvref_t<tuple_element_t<i.value, DstDescs>>::GetNumOfDimension(),
+                "wrong!");
+        });
+
+        static_assert(nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(get_thread_local_1d_id()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            const auto src_thread_slice_origins = generate_tuple(
+                [&](auto i) { return src_block_slice_origins[i] + thread_data_idx_begin; },
+                Number<nSrc>{});
+
+            const auto dst_thread_slice_origins = generate_tuple(
+                [&](auto i) { return dst_block_slice_origins[i] + thread_data_idx_begin; },
+                Number<nDst>{});
+
+            threadwise_transfer_.SetSrcSliceOrigins(src_descs, src_thread_slice_origins);
+            threadwise_transfer_.SetDstSliceOrigins(dst_descs, dst_thread_slice_origins);
+        }
+    }
+
+    template <typename SrcBuffers>
+    __device__ void RunRead(const SrcDescs& src_descs, const SrcBuffers& src_bufs)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunRead(src_descs, src_bufs);
+        }
+    }
+
+    template <typename T>
+    using is_tuple = decltype(std::declval<T&>().IsTuple());
+
+    template <typename DstBuffers>
+    __device__ void RunWrite(const DstDescs& dst_descs, DstBuffers dst_bufs)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            if constexpr(is_detected<is_tuple, decltype(dst_bufs)>::value)
+                threadwise_transfer_.RunWrite(dst_descs, dst_bufs);
+            else
+                threadwise_transfer_.RunWrite(dst_descs, tie(dst_bufs));
+        }
+    }
+
+    template <typename SrcBuffers, typename DstBuffers>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        RunRead(src_descs, src_bufs);
+        RunWrite(dst_descs, dst_bufs);
+    }
+
+    template <index_t ISrc>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDescs& src_descs, Number<ISrc> iSrc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_descs, iSrc, step);
+        }
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs, const Index& step)
+    {
+        static_for<0, SrcDescs::Size(), 1>{}(
+            [&](auto i) { MoveSrcSliceWindow(src_descs, i, step); });
+    }
+
+    template <index_t IDst>
+    __device__ void
+    MoveDstSliceWindow(const DstDescs& dst_descs, Number<IDst> iDst, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_descs, iDst, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs, const Index& step)
+    {
+        static_for<0, DstDescs::Size(), 1>{}(
+            [&](auto i) { MoveDstSliceWindow(dst_descs, i, step); });
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v7r2<SrcDatas,
+                                           DstDatas,
+                                           SrcDescs,
+                                           DstDescs,
+                                           ElementwiseOperation,
+                                           DstInMemOps,
+                                           decltype(thread_slice_lengths),
+                                           SrcDimAccessOrder,
+                                           DstDimAccessOrder,
+                                           SrcVectorDim,
+                                           DstVectorDim,
+                                           SrcScalarPerVector,
+                                           DstScalarPerVector,
+                                           ThreadTransferSrcResetCoordinateAfterRunFlags,
+                                           ThreadTransferDstResetCoordinateAfterRunFlags>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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[M, K], B0[K, N],
+//   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 <typename AsLayout,
+          typename BsLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename AsDataType,
+          typename BsDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGemmMultipleABD : 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,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        std::array<ck::index_t, NumATensor> StrideAs,
+                        std::array<ck::index_t, NumBTensor> StrideBs,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        ck::index_t StrideE,
+                        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/thread/threadwise_tensor_slice_transfer_v7r2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+#include "ck/utility/is_detected.hpp"
+
+namespace ck {
+
+// Thread-level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//   6. Does not need to know src_descs and dst_descs at compile-time
+//   7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
+//
+// Does following things to avoid scratch memory issue
+//   1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
+//   2. Pass tensor descritpors by reference (or tuple of references)
+//   3. Does not keep reference to tensor descriptor
+//   4. Does not construct new tensor coordinate when call Run()
+template <typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags> // Sequence<bool ...>
+struct ThreadwiseTensorSliceTransfer_v7r2
+{
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    // scalar per access on each dim
+    // FIXME: don't use lambda_scalar_per_access
+    static constexpr auto src_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+    using SrcSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   SrcDimAccessOrder,
+                                                   remove_cv_t<decltype(src_scalar_per_access)>>;
+
+    static constexpr auto dst_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+    using DstSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   DstDimAccessOrder,
+                                                   remove_cv_t<decltype(dst_scalar_per_access)>>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v7r2(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto i) {
+            dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename DataTypes, index_t ScalarPerVector>
+    __device__ static auto generate_vectors()
+    {
+        auto data_types = DataTypes{};
+
+        constexpr index_t num = data_types.Size();
+
+        return generate_tuple(
+            [&](auto i) {
+                using DataType = remove_cvref_t<decltype(data_types[i])>;
+
+                return vector_type_maker_t<DataType, ScalarPerVector>{};
+            },
+            Number<num>{});
+    }
+
+    template <typename T>
+    using has_vec_len = decltype(std::declval<T&>().vec_len);
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    template <typename SrcBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size(), bool> = false>
+    __device__ void RunRead(const SrcDescs& src_descs, const SrcBuffers& src_bufs)
+    {
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto iAccess) {
+            auto src_vectors = generate_vectors<SrcDatas, SrcScalarPerVector>();
+            auto dst_vectors = generate_vectors<DstDatas, DstScalarPerVector>();
+
+            // copy data from src_bufs into src_vectors
+            static_for<0, nSrc, 1>{}([&](auto i) {
+                using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
+
+                const bool is_src_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
+                                                                                src_coords_[i]);
+
+                src_vectors(i).template AsType<src_vector_t>()(I0) =
+                    src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(),
+                                                           is_src_valid);
+            });
+
+            if constexpr(is_detected<has_vec_len, decltype(element_op_)>::value)
+            {
+                constexpr auto elem_op_vec_len = decltype(element_op_)::vec_len;
+
+                static_assert(is_same<remove_cvref_t<decltype(elem_op_vec_len)>, index_t>::value,
+                              "vec_len in element_op_ type is not index_t");
+
+                static_assert(elem_op_vec_len == 1 || elem_op_vec_len == 2 ||
+                                  elem_op_vec_len == 4 || elem_op_vec_len == 8,
+                              "vec_len in element_op_ must be 1, 2, 4, 8");
+
+                static_assert(SrcScalarPerVector % elem_op_vec_len == 0,
+                              "vec_len in element_op_ cannot be divided by SrcScalarPerVector!");
+
+                // apply pointwise function
+                static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
+                    // get reference to src data
+                    const auto src_data_refs = generate_tie(
+                        // return type should be lvalue
+                        [&](auto iSrc) -> const auto& {
+                            using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                            using elem_op_vec_t =
+                                typename vector_type<SrcData, elem_op_vec_len>::type;
+
+                            return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
+                        },
+                        Number<nSrc>{});
+
+                    // get reference to dst data
+                    auto dst_data_refs = generate_tie(
+                        // return type should be lvalue
+                        [&](auto iDst) -> auto& {
+                            using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                            using elem_op_vec_t =
+                                typename vector_type<DstData, elem_op_vec_len>::type;
+
+                            return dst_vectors(iDst).template AsType<elem_op_vec_t>()(i);
+                        },
+                        Number<nDst>{});
+
+                    // apply pointwise function
+                    // pointwise function signature:
+                    // element_op_(dst_data_refs[I0],
+                    //             dst_data_refs[I1],
+                    //             ...,
+                    //             src_data_refs[I0],
+                    //             src_data_refs[I1],
+                    //             ...)
+                    unpack2(element_op_, dst_data_refs, src_data_refs);
+                });
+            }
+            else
+            {
+                // apply pointwise function
+                static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                    // get reference to src data
+                    const auto src_data_refs = generate_tie(
+                        // return type should be lvalue
+                        [&](auto iSrc) -> const auto& {
+                            using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                            return src_vectors[iSrc].template AsType<SrcData>()[i];
+                        },
+                        Number<nSrc>{});
+
+                    // get reference to dst data
+                    auto dst_data_refs = generate_tie(
+                        // return type should be lvalue
+                        [&](auto iDst) -> auto& {
+                            using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                            return dst_vectors(iDst).template AsType<DstData>()(i);
+                        },
+                        Number<nDst>{});
+
+                    // apply pointwise function
+                    // pointwise function signature:
+                    // element_op_(dst_data_refs[I0],
+                    //             dst_data_refs[I1],
+                    //             ...,
+                    //             src_data_refs[I0],
+                    //             src_data_refs[I1],
+                    //             ...)
+                    unpack2(element_op_, dst_data_refs, src_data_refs);
+                });
+            }
+
+            dst_vectors_tuple_(iAccess) = dst_vectors;
+
+            // move coordinate
+            if constexpr(iAccess.value != num_access - 1)
+            {
+                constexpr auto forward_step = SrcSpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nSrc, 1>{}([&](auto i) {
+                    move_tensor_coordinate(src_descs[i],
+                                           src_coords_(i),
+                                           make_tensor_coordinate_step(src_descs[i], forward_step));
+                });
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto src_reset_step =
+                    make_tensor_coordinate_step(src_descs[i], GetSrcCoordinateResetStep());
+
+                move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
+            }
+        });
+    }
+
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename DstBuffers,
+              enable_if_t<DstDescs::Size() == DstBuffers::Size(), bool> = false>
+    __device__ void RunWrite(const DstDescs& dst_descs, DstBuffers dst_bufs)
+    {
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto iAccess) {
+            auto dst_vectors = dst_vectors_tuple_[iAccess];
+
+            // copy data from buf_vectors into dst_bufs
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
+
+                const bool is_dst_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
+                                                                                dst_coords_[i]);
+
+                constexpr InMemoryDataOperationEnum DstInMemOp =
+                    static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
+
+                dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
+                    dst_coords_[i].GetOffset(),
+                    is_dst_valid,
+                    dst_vectors[i].template AsType<dst_vector_t>()[I0]);
+            });
+
+            // move coordinate
+            if constexpr(iAccess.value != num_access - 1)
+            {
+                constexpr auto forward_step = DstSpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nDst, 1>{}([&](auto i) {
+                    move_tensor_coordinate(dst_descs[i],
+                                           dst_coords_(i),
+                                           make_tensor_coordinate_step(dst_descs[i], forward_step));
+                });
+            }
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            if constexpr(DstResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto dst_reset_step =
+                    make_tensor_coordinate_step(dst_descs[i], GetDstCoordinateResetStep());
+
+                move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
+            }
+        });
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename SrcBuffers,
+              typename DstBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
+                              DstDescs::Size() == DstBuffers::Size(),
+                          bool> = false>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        RunRead(src_descs, src_bufs);
+        RunWrite(dst_descs, dst_bufs);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        if constexpr(num_access == 0)
+        {
+            return typename SrcSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            return SrcSpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+        }
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        if constexpr(num_access == 0)
+        {
+            return typename DstSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            return DstSpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t ISrc>
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       Number<ISrc> iSrc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRunFlags::At(iSrc)
+                ? src_slice_origin_step_idx
+                : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
+
+        move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t IDst>
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       Number<IDst> iDst,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRunFlags::At(iDst)
+                ? dst_slice_origin_step_idx
+                : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx);
+
+        move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
+    }
+
+    private:
+    using SrcVectorsType = decltype(generate_vectors<SrcDatas, SrcScalarPerVector>());
+    using DstVectorsType = decltype(generate_vectors<DstDatas, DstScalarPerVector>());
+
+    static constexpr auto num_access = SrcSpaceFillingCurve::GetNumOfAccess();
+
+    StaticallyIndexedArray<DstVectorsType, num_access> dst_vectors_tuple_;
+
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/utility/is_detected.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+namespace detail {
+template <class Default, class AlwaysVoid, template <class...> class Op, class... Args>
+struct detector
+{
+    using value_t = std::false_type;
+    using type    = Default;
+};
+
+template <class Default, template <class...> class Op, class... Args>
+struct detector<Default, std::void_t<Op<Args...>>, Op, Args...>
+{
+    using value_t = std::true_type;
+    using type    = Op<Args...>;
+};
+} // namespace detail
+
+struct nonesuch
+{
+    ~nonesuch()               = delete;
+    nonesuch(nonesuch const&) = delete;
+    void operator=(nonesuch const&) = delete;
+};
+
+template <template <class...> class Op, class... Args>
+using is_detected = typename detail::detector<nonesuch, void, Op, Args...>::value_t;
+
+} // namespace ck

include/ck/utility/tuple.hpp

@@ -177,6 +177,8 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X
     }
 
     __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+
+    __host__ __device__ static constexpr bool IsTuple() { return true; }
 };
 
 template <>