Skip lds of b matrix (#326)

* start

* read for gridwise gemm

* add MakeBGridDescriptor_K0_N0_N1_N2_N3_K1

* add thread  copy desc and register buffer

* add K0PerBlock dim

* add read global data

* finish gridwise gemm

* finish blockwise gemm

* add print data

* add smallest config

* add compare code for gridwis gemm

* fix NXdlPerWave

* fix k0perthread and gridewis gemm main loop

* remove b matrix lds alloc

* fix name

* add test code

* create b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3 from parameter

* add double register

* modify b_thread_desc_

* add float

* fp16 tag

* add tail for pipeline

* finish main loop

* optimize main loop

* start clear gridwise gemm

* clear code

* clear redundant code

* change file name

* change file name

* fix bug after merge develop

* fix input parameters

* using MultiK0 control b load data loop

* fix some config

* 4 buffer

* fix bug

* one can use

* change read order

* change buffer array to tuple

* change to 8 buffer

* interleave buffer load

* change to 16

* read 8 buffer

* add data buffer to template

* fix after merge develop(head file)

* format

* change to 4 buffer

* remove unnecessary lambda fun

example/01_gemm/CMakeLists.txt

@@ -4,5 +4,6 @@ add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
 add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
 add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
 add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
+add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
 # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
 add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)

example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp

+#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/device_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_xdl_skip_b_lds.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/host_tensor/device_memory.hpp"
+#include "ck/library/host_tensor/host_tensor.hpp"
+#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+#define USING_SKIP_LDS 1
+
+// clang-format off
+#if USING_SKIP_LDS
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSkipBLds
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BThreadTransfer|  BBlock| CThreadTransfer| CThreadTransfer|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|       SrcScalar|  buffer| SrcDstVectorDim|       DstScalar|
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |       PerVector|  size  |                |       PerVector|
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |        |                |                |
+#if 0       
+                    <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    16,   64,     4,  8,   16,   16,    1,    1,     S<16, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,            8,       8,        7,               1>;
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+#else  
+                    <   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    16,   64,     4,  4,   16,   16,    1,    1,     S<16, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,             4,      4,        7,               1>;
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+#endif
+
+#else
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+                   <   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    16,   64,     4,  4,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,              1,      2>;
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+
+#endif
+    // clang-format on
+
+    using ReferenceGemmInstance = ck::tensor_operation::host::
+        ReferenceGemm<ADataType, BDataType, CDataType, float, AElementOp, BElementOp, CElementOp>;
+
+template <typename DataType>
+std::ostream& show_2d_matrix(std::ostream& os, Tensor<DataType>& matrix)
+{
+    os << "[" << std::endl;
+    for(size_t x = 0; x < matrix.mDesc.GetLengths()[0]; x++)
+    {
+        os << "[";
+        for(size_t y = 0; y < matrix.mDesc.GetLengths()[1]; y++)
+        {
+            os << std::setw(5) << static_cast<float>(matrix(x, y));
+        }
+        os << "]" << std::endl;
+    }
+    os << "]";
+    return os;
+}
+int main(int argc, char* argv[])
+{
+    bool do_verification = 0;
+    int init_method      = 0;
+    bool time_kernel     = false;
+
+    // GEMM shape
+#if 1
+    ck::index_t M = 16;
+    ck::index_t N = 64 * 120;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideC = N;
+#else
+    ck::index_t M = 16;
+    ck::index_t N = 16;
+    ck::index_t K = 32;
+
+    ck::index_t StrideA = 8;
+    ck::index_t StrideB = 8;
+    ck::index_t StrideC = 16;
+#endif
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        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]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1, stride}));
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        // a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+    }
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_element_op);
+
+    if(!gemm.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(CDataType) * 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, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        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_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+#if 0
+        {
+            show_2d_matrix(std::cout << "a : ", a_m_k) << std::endl;
+            show_2d_matrix(std::cout << "b: ", b_k_n) << std::endl;
+            show_2d_matrix(std::cout << "c_device: ", c_m_n_device_result) << std::endl;
+            show_2d_matrix(std::cout << "c_host  :", c_m_n_host_result) << std::endl;
+        }
+#endif
+        ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+    }
+
+    return 0;
+}

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

+#ifndef CK_BLOCKWISE_GEMM_XDLOPS_B_REGISTER_HPP
+#define CK_BLOCKWISE_GEMM_XDLOPS_B_REGISTER_HPP
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0K0BN0N1N2N3K1BlockDesc,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr index_t WaveSize = 64;
+
+    static constexpr index_t KPerBlock = K0PerBlock * KPack;
+
+    static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack>{};
+
+    static constexpr index_t KPerThread  = KPerBlock / xdlops_gemm.K0PerXdlops;
+    static constexpr index_t K0PerThread = K0PerBlock / xdlops_gemm.K0PerXdlops;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPerThread * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPerThread * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    __host__ __device__ BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1()
+    {
+        static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
+                          BK0K0BN0N1N2N3K1BlockDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(BlockSize == MWaves * NWaves * WaveSize,
+                      "BlockSize != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
+    {
+        return transform_tensor_descriptor(
+            AK0MK1BlockDesc{},
+            make_tuple(
+                make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
+                make_unmerge_transform(
+                    make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __device__ void MoveABlockSliceWindow()
+    {
+        a_thread_copy_.MoveSrcSliceWindow(a_block_desc_m0_m1_m2_k,
+                                          make_multi_index(0, 0, 0, K0PerBlock * KPack));
+    }
+    __device__ void ResetABlockStartWindow()
+    {
+        a_thread_copy_.SetSrcCoord(CalculateAThreadOriginDataIndex());
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_thread_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            // read A
+            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                               make_tuple(m0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_,
+                               make_tuple(I0, I0, I0, I0),
+                               a_thread_buf);
+
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                static_for<0, KPerThread, KPack>{}([&](auto k) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+                    constexpr index_t k0 = k / KPack;
+                    static_for<0, KPack, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(k0, n0, i))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+
+    private:
+    // A[M0, M1, M2, KPerThread]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+
+    // B[N0, N1, N2, KPerThread]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<K0PerThread>{}, // KPerThread
+                                                       Number<NRepeat>{},     // repeat
+                                                       Number<KPack>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
+};
+
+} // namespace ck
+#endif

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

+#ifndef DEVICE_GEMM_XDL_SKIP_B_LDS_HPP
+#define DEVICE_GEMM_XDL_SKIP_B_LDS_HPP
+
+#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_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp"
+#include "ck/device_utility/device_prop.hpp"
+#include "ck/device_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockBufferSize,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceGemmXdlSkipBLds : public DeviceGemm<ALayout,
+                                                 BLayout,
+                                                 CLayout,
+                                                 ADataType,
+                                                 BDataType,
+                                                 CDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+    static_assert(BBlockBufferSize >= 2);
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_skip_b_lds_v1<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferSrcScalarPerVector,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockBufferSize,
+        Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
+        CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+            a_grid_desc_k0_m_k1_ =
+                DeviceGemmXdlSkipBLds::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ =
+                DeviceGemmXdlSkipBLds::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            c_grid_desc_m_n_ = DeviceGemmXdlSkipBLds::MakeCGridDescriptor_M_N(M, N, StrideC);
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ =
+                    GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+                b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(b_grid_desc_k0_n_k1_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3
+            b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_;
+        typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGemmXdlSkipBLds::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.M01_,
+                                            arg.N01_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            if(has_main_k0_block_loop)
+            {
+                const auto kernel = kernel_gemm_xdlops_skip_b_lds_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_skip_b_lds_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.M01_,
+                                           arg.N01_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_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();
+
+        // clang-format off
+        str << "DeviceGemmXdlSkipBLds"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp

+#ifndef CK_GRIDWISE_GEMM_XDLOPS_SKIP_B_LDS_V1_HPP
+#define CK_GRIDWISE_GEMM_XDLOPS_SKIP_B_LDS_V1_HPP
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3,
+          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainK0BlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_skip_b_lds_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
+                                                   p_b_grid,
+                                                   p_c_grid,
+                                                   p_shared,
+                                                   a_grid_desc_k0_m_k1,
+                                                   b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                                   c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op,
+                                                   block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3;
+    ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          index_t BBlockTransferSrcScalarPerVector,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockBufferSize,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_skip_b_lds_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    static constexpr index_t WaveSize = 64;
+    static constexpr index_t MWaves   = MPerBlock / (MXdlPerWave * MPerXDL);
+    static constexpr index_t NWaves   = NPerBlock / (NXdlPerWave * NPerXDL);
+
+    static constexpr auto xdlops_gemm    = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, K1>{};
+    static constexpr index_t K0PerThread = K0PerBlock / xdlops_gemm.K0PerXdlops;
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock * BBlockBufferSize>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock * BBlockBufferSize>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size_aligned) * sizeof(FloatAB);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  index_t M01,
+                  index_t N01)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // 2-stage prefetch currently only support even number of K0 loop
+        // TODO: add support for odd number of K0 loop
+        if(!((K0 / K0PerBlock) % BBlockBufferSize == 0))
+        {
+            return false;
+        }
+
+        // check M01, N01
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        if(!(M0 % M01 == 0 && N0 % N01 == 0))
+            return false;
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    // TODO move this function into GEMM-pipeline class
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = (K0 / (BBlockBufferSize * K0PerBlock)) > 1;
+
+        return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1)
+    {
+        const auto K0 = b_grid_desc_k0_n_k1.GetLength(I0);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+
+        const auto b_griddesc_k0_nblockid_nrepeat_waves_nperxdlops_k1 = transform_tensor_descriptor(
+            b_grid_desc_k0_n_k1,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(K0 / K0PerBlock, xdlops_gemm.K0PerXdlops, K0PerThread)),
+                       make_unmerge_transform(make_tuple(
+                           N / (NXdlPerWave * NWaves * NPerXDL), NXdlPerWave, NWaves, NPerXDL)),
+                       make_pass_through_transform(K1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5, 6>{}, Sequence<7>{}));
+        return b_griddesc_k0_nblockid_nrepeat_waves_nperxdlops_k1;
+    }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetWaveKNIdx(const index_t thread_id)
+    {
+        constexpr auto wave_threadid_to_nk_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(xdlops_gemm.K0PerXdlops, NPerXDL))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_nk_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix threadwise copy
+        constexpr auto b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           I1,
+                                                           Number<K0PerThread>{}, // K0PerThread
+                                                           I1,                    // NBlockId
+                                                           Number<NXdlPerWave>{}, // repeat
+                                                           I1,                    // waves
+                                                           I1,                    // NPerXdlops
+                                                           Number<K1>{}));
+
+        using BlockwiseGemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1<
+            BlockSize,
+            FloatAB,
+            FloatAcc,
+            decltype(a_block_desc_k0_m_k1),
+            decltype(b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+            MPerBlock,
+            NPerBlock,
+            K0PerBlock,
+            MPerXDL,
+            NPerXDL,
+            MXdlPerWave,
+            NXdlPerWave,
+            K1>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01, index_t N01)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        const auto M00 = M0 / M01;
+        const auto N00 = N0 / N01;
+
+        const auto m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}));
+
+        const auto cblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
+    using BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 =
+        decltype(MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(BGridDesc_K0_N_K1{}));
+
+    template <bool HasMainK0BlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+        const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2& c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy = ThreadGroupTensorSliceTransfer_v4r1<
+            ThisThreadBlock,
+            AElementwiseOperation,
+            ck::tensor_operation::element_wise::PassThrough,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock * BBlockBufferSize, MPerBlock, K1>,
+            ABlockTransferThreadClusterLengths_K0_M_K1,
+            ABlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            decltype(a_grid_desc_k0_m_k1),
+            decltype(a_block_desc_k0_m_k1),
+            ABlockTransferSrcAccessOrder,
+            Sequence<1, 0, 2>,
+            ABlockTransferSrcVectorDim,
+            2,
+            ABlockTransferSrcScalarPerVector,
+            ABlockTransferDstScalarPerVector_K1,
+            1,
+            1,
+            AThreadTransferSrcResetCoordinateAfterRun,
+            true,
+            1>(a_grid_desc_k0_m_k1,
+               make_multi_index(0, m_block_data_idx_on_grid, 0),
+               a_element_op,
+               a_block_desc_k0_m_k1,
+               make_multi_index(0, 0, 0),
+               ck::tensor_operation::element_wise::PassThrough{});
+
+        ignore = b_element_op;
+        // B matrix threadwise copy
+        constexpr auto b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           I1,
+                                                           Number<K0PerThread>{}, // K0PerThread
+                                                           I1,                    // NBlockId
+                                                           Number<NXdlPerWave>{}, // repeat
+                                                           I1,                    // waves
+                                                           I1,                    // NPerXdlops
+                                                           Number<K1>{}));
+
+        auto b_thread_buf = generate_tuple(
+            [&](auto i) {
+                ignore = i;
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    FloatAB,
+                                    b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3.GetElementSpaceSize(),
+                                    true>{};
+            },
+            Number<BBlockBufferSize>{});
+
+        const auto wave_id     = GetWaveIdx();
+        const auto wave_k_n_id = GetWaveKNIdx(wave_id[I2]);
+
+#if 0
+        const index_t block_id  = get_block_1d_id();
+        const index_t thread_id = get_thread_local_1d_id();
+        printf("block id: %d  m blockid: %d n block id: %d ,thread id: %d, wave id :{%d %d %d} "
+               "kn id: {%d %d}\n",
+               block_id,
+               block_work_idx[I0],
+               block_work_idx[I1],
+               thread_id,
+               wave_id[I0],
+               wave_id[I1],
+               wave_id[I2],
+               wave_k_n_id[I0],
+               wave_k_n_id[I1]);
+        printf("mfma thread k per xdlops: %d K0PerThread: %d HasMainK0BlockLoop: %d K0: %d  \t", 
+                xdlops_gemm.K0PerXdlops, K0PerThread, HasMainK0BlockLoop, b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3.GetLength(I0));
+#endif
+
+        auto b_threadwise_copy =
+            ThreadwiseTensorSliceTransfer_v2<FloatAB,
+                                             FloatAB,
+                                             decltype(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+                                             decltype(b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+                                             Sequence<I1,
+                                                      I1,
+                                                      Number<K0PerThread>{},
+                                                      I1,
+                                                      Number<NXdlPerWave>{},
+                                                      I1,
+                                                      I1,
+                                                      Number<K1>{}>,
+                                             Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                             7,
+                                             BBlockTransferSrcScalarPerVector,
+                                             BThreadTransferSrcResetCoordinateAfterRun,
+                                             true>(
+                b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                make_multi_index(
+                    0, wave_k_n_id[I0], 0, block_work_idx[I1], 0, wave_id[I1], wave_k_n_id[I1], 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1<
+            BlockSize,
+            FloatAB,
+            FloatAcc,
+            decltype(a_block_desc_k0_m_k1),
+            decltype(b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+            MPerBlock,
+            NPerBlock,
+            K0PerBlock,
+            MPerXDL,
+            NPerXDL,
+            MXdlPerWave,
+            NXdlPerWave,
+            K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        // gridwise GEMM pipeline
+        constexpr auto a_block_slice_copy_step =
+            make_multi_index(K0PerBlock * BBlockBufferSize, 0, 0);
+        constexpr auto b_thread_slice_copy_step = make_multi_index(1, 0, 0, 0, 0, 0, 0, 0);
+        // preload data to regiester and LDS
+        {
+            // Read
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf);
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1, a_block_slice_copy_step);
+
+            static_for<0, BBlockBufferSize, 1>{}([&](auto ii) {
+                b_threadwise_copy.Run(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                      b_grid_buf,
+                                      b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                      make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                                      b_thread_buf(Number<ii>{}));
+                b_threadwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                                     b_thread_slice_copy_step);
+            });
+
+            // Initialize C
+            c_thread_buf.Clear();
+            // a data write to lds
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m_k1, a_block_buf);
+            // main body
+            if constexpr(HasMainK0BlockLoop)
+            {
+                index_t K0BlockMainLoop =
+                    __builtin_amdgcn_readfirstlane(K0 / (BBlockBufferSize * K0PerBlock));
+                index_t i = 0;
+                do
+                {
+                    a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf);
+                    blockwise_gemm.ResetABlockStartWindow();
+                    block_sync_lds();
+
+                    static_for<0, BBlockBufferSize, 1>{}([&](auto ii) {
+                        blockwise_gemm.Run(a_block_buf, b_thread_buf(Number<ii>{}), c_thread_buf);
+                        blockwise_gemm.MoveABlockSliceWindow();
+                        s_nop();
+
+                        b_threadwise_copy.Run(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                              b_grid_buf,
+                                              b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                                              b_thread_buf(Number<ii>{}));
+                        b_threadwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                                             b_thread_slice_copy_step);
+                    });
+
+                    block_sync_lds();
+                    a_blockwise_copy.RunWrite(a_block_desc_k0_m_k1, a_block_buf);
+                    // move a and b window
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1,
+                                                        a_block_slice_copy_step);
+
+                    i += 1;
+                } while(i < (K0BlockMainLoop - 1));
+            }
+
+            // tail
+            {
+                block_sync_lds();
+
+                blockwise_gemm.ResetABlockStartWindow();
+
+                static_for<0, BBlockBufferSize, 1>{}([&](auto ii) {
+                    blockwise_gemm.Run(a_block_buf, b_thread_buf(Number<ii>{}), c_thread_buf);
+                    blockwise_gemm.MoveABlockSliceWindow();
+                });
+            }
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I7);
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+                    c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
+#endif

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp

@@ -1192,6 +1192,10 @@ struct ThreadwiseTensorSliceTransfer_v4
 
         move_tensor_coordinate(SrcDesc{}, src_ref_coord_, src_slice_move_step_iter);
     }
+    __device__ void SetSrcCoord(const Index& src_ref_idx)
+    {
+        src_ref_coord_ = make_tensor_coordinate(SrcDesc{}, src_ref_idx);
+    }
 
     private:
     SrcCoord src_ref_coord_;

include/ck/utility/synchronization.hpp

@@ -18,5 +18,15 @@ __device__ void block_sync_lds()
     __syncthreads();
 #endif
 }
+__device__ void s_nop()
+{
+#if 1
+    asm volatile("\
+    s_nop 0 \n \
+    " ::);
+#else
+    __builtin_amdgcn_sched_barrier(0);
+#endif
+}
 
 } // namespace ck