Merge branch 'add_a8w8_preshuffle_ckprofiler' of...

Merge branch 'add_a8w8_preshuffle_ckprofiler' of https://github.com/ROCm/composable_kernel into update_cka8w8_uc

example/65_gemm_multiply_multiply/CMakeLists.txt

 add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp)
 # target_compile_options(example_gemm_multiply_multiply_xdl_fp8 PRIVATE -mllvm -greedy-reverse-local-assignment=1 -save-temps=$PWD -Wno-gnu-line-marker)
 add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp)
+add_example_executable(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle gemm_multiply_multiply_xdl_fp8.cpp)
 add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp)
 add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp)

example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, 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_d_xdl_cshuffle_v3_b_preshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_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"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using FP8  = ck::f8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = FP8;
+using B0DataType       = FP8;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using D0Layout = Row;
+using D1Layout = Col;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+struct MultiplyMultiply
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<F16, float, float, float>(F16& e,
+                                                                            const float& c,
+                                                                            const float& d0,
+                                                                            const float& d1) const
+    {
+        const float x0_f = c * d0 * d1;
+
+        e = ck::type_convert<F16>(x0_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<BF16, float, float, float>(BF16& e,
+                                                                             const float& c,
+                                                                             const float& d0,
+                                                                             const float& d1) const
+    {
+        const float x0_f = c * d0 * d1;
+
+        e = ck::type_convert<BF16>(x0_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, int, float, float>(
+        ck::half_t& e, const int& c, const float& d0, const float& d1) const
+    {
+        const float x0_f =
+            ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
+
+        e = ck::type_convert<ck::half_t>(x0_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::bhalf_t, int, float, float>(
+        ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const
+    {
+        const float x0_f =
+            ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
+
+        e = ck::type_convert<ck::bhalf_t>(x0_f);
+    }
+};
+
+void preShuffleBuffer(const FP8* src,
+                      FP8* dst,
+                      int N,
+                      int K,
+                      int NRepeat,
+                      int KRepeat,
+                      int NWave,
+                      int KLane,
+                      int NLane,
+                      int KPack)
+{
+    int K0 = K / (KRepeat * KLane * KPack);
+    // K -> src: K0 KLane KRepeat KPack  -> dst: K0 KRpeat KLane KPack, move klane inner to make all
+    // lanes contiguous N -> N0 NRepeat NWave NLane  // todo : is NRepeat outer or inner? now it's 1
+    int tempn, tempk;
+    for(int n = 0; n < N; ++n)
+    {
+        for(int k = 0; k < K; ++k)
+        {
+            int n0 = n / (NRepeat * NLane * NWave);
+            int k0 = k / (KRepeat * KLane * KPack);
+            tempn  = n % (NRepeat * NLane * NWave);
+            tempk  = k % (KRepeat * KLane * KPack);
+
+            int n1 = tempn / (NLane * NWave);
+            int k1 = tempk / (KRepeat * KPack); // Klane
+            tempn  = tempn % (NLane * NWave);
+            tempk  = tempk % (KRepeat * KPack);
+            int n2 = tempn / NLane;
+            int k2 = tempk / KPack; // KRepeat
+            int n3 = tempn % NLane;
+            int k3 = tempk % KPack; // Kpack
+
+            int outputIndex = n0 * KPack * NLane * KLane * NWave * KRepeat * K0 * NRepeat +
+                              n1 * KPack * NLane * KLane * NWave * KRepeat * K0 +
+                              k0 * KPack * NLane * KLane * NWave * KRepeat +
+                              k2 * KPack * NLane * KLane * NWave + n2 * KPack * NLane * KLane +
+                              k1 * KPack * NLane + n3 * KPack + k3;
+#if 0
+            int k1 = tempk / (KLane * KPack); //KRepeat
+            int n1 = tempn / (NLane * NWave); //NRepeat
+            tempn  = tempn % (NLane * NWave);
+            tempk  = tempk % (KLane * KPack);
+            int n2 = tempn / NLane; // NWave
+            int k2 = tempk / KPack; // KLane
+            int n3 = tempn % NLane; // NLane
+            int k3 = tempk % KPack; // Kpack
+
+            int outputIndex = n0 * KPack * NLane * KLane * NWave * NRepeat * KRepeat * K0 +
+                              k0 * KPack * NLane * KLane * NWave * NRepeat * KRepeat +
+                              k1 * KPack * NLane * KLane * NWave * NRepeat +
+                              n1 * KPack * NLane * KLane * NWave +
+                              n2 * KPack * NLane * KLane +
+                              k2 * KPack * NLane +
+                              n3 * KPack +
+                              k3;
+#endif
+            dst[outputIndex] = src[n * K + k];
+        }
+    }
+}
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MultiplyMultiply;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
+    // clang-format off
+///######|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     DsData|     EData|     AccData|         CShuffle|           A|           B|          CDE|           GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+///######|         |         |         |        |       Type|       Type|       Type|      Type|        Type|         DataType| 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| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+///######|         |         |         |        |           |           |           |          |            |                 |   Operation|   Operation|    Operation|               |      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+///######|         |         |         |        |           |           |           |          |            |                 |            |            |             |               |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |    S<C, D0, D1>|
+///###### RCR
+        // kernel 1: 256->32x128x128 
+        // <      Row,      Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   32,   128,    128,  16,  16,  32,   32,    1,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,     S<8, 32, 1>,    S<1, 0, 2>,     S<1, 0, 2>,             2,              16,             16,          0,          1,           1,               S<1, 32, 1, 8>,      S<8, 8, 1>,  ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>;
+        // <      Row,      Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   32,   128,    256,  16,  16,  32,   32,    1,    1,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,     S<16, 16, 1>,    S<1, 0, 2>,     S<1, 0, 2>,             2,              16,             16,          0,          1,           1,               S<1, 32, 1, 8>,      S<8, 8, 1>,  ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
+        <      Row,      Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
+               AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,
+               256,   256,    128,
+               16,   16,
+               32,   32, 
+               8,    2,
+               S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
+               S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
+               1,    1,   S<1, 32, 1, 8>, S<8, 8, 1>,
+               ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
+        // kernel 2: 128->32x128x128
+        //  <      Row,      Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   128,   32,   128,    128,  16,  16,  32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,     S<8, 16, 1>,    S<1, 0, 2>,     S<1, 0, 2>,             2,              16,             16,          0,          1,           1,               S<1, 16, 1, 8>,      S<8, 8, 1>,  ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>;
+
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideD = 0;
+    ck::index_t StrideE = N;
+
+    ck::index_t KBatch = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 12)
+    {
+        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]);
+
+        KBatch = std::stoi(argv[11]);
+    }
+    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, KBatch\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<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B0DataType> b0_preshuffled(
+        f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); // use laout only for size
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD, D1Layout{}));
+    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 << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_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<A0DataType>{-2, 2});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{0, 2});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-2, 2});
+        break;
+    case 2:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_1<D1DataType>{});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_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());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+
+    constexpr auto I0 = ck::Number<0>{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+
+    auto preshuffle_params = device_op.GetPreShuffleParameters();
+
+    preShuffleBuffer(b0_k_n.mData.data(),
+                     b0_preshuffled.mData.data(),
+                     N,
+                     K,
+                     preshuffle_params[0],
+                     preshuffle_params[1],
+                     preshuffle_params[2],
+                     preshuffle_params[3],
+                     preshuffle_params[4],
+                     preshuffle_params[5]);
+
+    b0_device_buf.ToDevice(b0_preshuffled.mData.data());
+
+    auto invoker = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(),
+                               b0_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer(),
+                                                                   d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, NumDTensor>{I0, I0},
+                               StrideE,
+                               KBatch,
+                               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, 0, 50, 50, true, 50});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(do_verification)
+    {
+        invoker.Run(argument, StreamConfig{nullptr, false});
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B0DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, 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), d0_m_n(m, n), d1_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/blockwise_gemm_pipeline_xdlops_b_preshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Compute optimized pipeline
+// GlobalPrefetchStages: 2
+// LocalPreFillStages: 2
+// LocalPreFetchStages: 2
+// LocalSharedMemoryBuffer: 2
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPacks>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle
+{
+};
+
+template <index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack
+          // ,bool TransposeC //disable transposec right now...
+          >
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intrawave,
+                                                BlockSize,
+                                                ADataType,
+                                                BDataType,
+                                                ComputeDataType,
+                                                AccDataType,
+                                                ATileDesc,
+                                                BTileDesc,
+                                                AMmaTileDesc,
+                                                BMmaTileDesc,
+                                                ABlockTransferSrcScalarPerVector,
+                                                BBlockTransferSrcScalarPerVector,
+                                                MPerBlock,
+                                                NPerBlock,
+                                                KPerBlock,
+                                                MPerXDL,
+                                                NPerXDL,
+                                                MRepeat,
+                                                NRepeat,
+                                                KPack>
+    : BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                        ADataType,
+                                        BDataType,
+                                        ComputeDataType,
+                                        AccDataType,
+                                        ATileDesc,
+                                        BTileDesc,
+                                        AMmaTileDesc,
+                                        BMmaTileDesc,
+                                        ABlockTransferSrcScalarPerVector,
+                                        BBlockTransferSrcScalarPerVector,
+                                        MPerBlock,
+                                        NPerBlock,
+                                        KPerBlock,
+                                        MPerXDL,
+                                        NPerXDL,
+                                        MRepeat,
+                                        NRepeat,
+                                        KPack>
+
+{
+    using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>;
+    using Base::I0;
+    using Base::I1;
+    using Base::KRepeat;
+    using Base::xdlops_gemm;
+    using typename Base::HotLoopInstList;
+
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::CalculateCThreadOriginDataIndex8D;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+
+    static constexpr index_t PrefetchStages  = 2;
+    static constexpr index_t PrefillStages   = 1;
+    static constexpr index_t GlobalBufferNum = 1;
+
+    __host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+
+        return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
+    }
+
+    __device__ static constexpr auto HotLoopScheduler()
+    {
+        // A/B split schedule
+        // compiler is likely to use ds_read2 when instruction width smaller than 16bytes
+        constexpr auto num_ds_read_inst_a =
+            HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
+                ? HotLoopInstList::A_LDS_Read_Inst_Num
+                : HotLoopInstList::A_LDS_Read_Inst_Num / 2;
+
+        constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
+
+        constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
+        constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
+
+        constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
+
+        constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
+        constexpr auto ds_read_a_issue_cycle =
+            HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
+        constexpr auto ds_read_a_mfma_rate =
+            (mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
+
+        constexpr auto num_dsread_a_mfma =
+            (num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
+
+        // stage 1
+        constexpr auto num_mfma_stage1 = num_mfma_inst - num_dsread_a_mfma;
+        constexpr auto num_mfma_per_issue =
+            num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b);
+        constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a;
+
+        static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) {
+            ignore = i;
+            static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
+                ignore = idswrite;
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            });
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+            __builtin_amdgcn_sched_group_barrier(
+                0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA
+        });
+
+        static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0);                  // VMEM read
+            __builtin_amdgcn_sched_group_barrier(0x008, num_mfma_per_issue, 0); // MFMA
+        });
+
+        // stage 2
+        static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) {
+            if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >=
+                         ds_read_a_mfma_rate)
+            {
+                __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
+            }
+            else
+            {
+                __builtin_amdgcn_sched_group_barrier(0x100,
+                                                     num_ds_read_inst_a - (num_dsread_a_mfma - 1) *
+                                                                              ds_read_a_mfma_rate,
+                                                     0); // DS read
+            }
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+        });
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf0,
+                        ABlockBuffer& a_block_buf1,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        const BGridBuffer& b_grid_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        index_t num_loop) const
+    {
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+
+        // Global prefetch 1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, Number<0>{});
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // // Local prefill 1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf0);
+
+        // // Global prefetch 2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // Local prefetch 1
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k0) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
+                                   a_block_buf0,
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, k0, I0),
+                                   a_thread_buf);
+            });
+        });
+        __builtin_amdgcn_sched_barrier(0);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf1);
+
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, Number<1>{});
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<ComputeDataType, KPack> a_thread_vec;
+                            vector_type<ComputeDataType, KPack> b_thread_vec =
+                                b_blockwise_copy
+                                    .template GetSrcThreadScratchIdx<Sequence<n0, k0, 0>,
+                                                                     Number<0>{}>();
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, I0, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<ComputeDataType,
+                                                     xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+                });
+
+                block_sync_lds();
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
+                                           a_block_buf1,
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k0, I0),
+                                           a_thread_buf);
+                    });
+                });
+
+                HotLoopScheduler();
+                __builtin_amdgcn_sched_barrier(0);
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf0);
+
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, Number<0>{});
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<ComputeDataType, KPack> a_thread_vec;
+                            vector_type<ComputeDataType, KPack> b_thread_vec =
+                                //  b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<n0,
+                                //  k0, 0>,
+                                b_blockwise_copy
+                                    .template GetSrcThreadScratchIdx<Sequence<n0, k0, 0>,
+                                                                     Number<1>{}>();
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, I0, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<ComputeDataType,
+                                                     xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+                });
+
+                block_sync_lds();
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
+                                           a_block_buf0,
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k0, I0),
+                                           a_thread_buf);
+                    });
+                });
+                HotLoopScheduler();
+                __builtin_amdgcn_sched_barrier(0);
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+        // tail
+        if constexpr(TailNum == TailNumber::Even)
+        {
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf1);
+            b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, Number<1>{});
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec =
+                            b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<n0, k0, 0>,
+                                                                             Number<0>{}>();
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            block_sync_lds();
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
+                                       a_block_buf1,
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k0, I0),
+                                       a_thread_buf);
+                });
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec =
+                            b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<n0, k0, 0>,
+                                                                             Number<1>{}>();
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+            // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle
+            // latency
+            // __builtin_amdgcn_sched_barrier(0);
+        }
+        else
+        {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec =
+                            b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<n0, k0, 0>,
+                                                                             Number<0>{}>();
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+        }
+    }
+
+    protected:
+    using Base::a_thread_copy_;
+    using Base::a_thread_desc_;
+    using Base::b_thread_desc_;
+    using Base::c_thread_desc_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp

@@ -109,6 +109,12 @@ struct ThreadGroupTensorSliceTransfer_v4r1
         }
     }
 
+    template <typename SeqIdx, index_t ThreadScratchId = 0>
+    __device__ constexpr auto GetSrcThreadScratchIdx()
+    {
+        return threadwise_transfer_.template GetSrcThreadScratchIdx<SeqIdx, ThreadScratchId>();
+    }
+
     template <typename SrcBuffer, index_t ThreadScratchId = 0>
     __device__ void RunRead(const SrcDesc& src_desc,
                             const SrcBuffer& src_buf,

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

@@ -96,6 +96,51 @@ struct DeviceGemmMultipleDSplitK : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
+// GEMM:
+//   input : A[M, K], B[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 ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGemmMultipleDSplitKBPreShuffle : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        ck::index_t StrideE,
+                        ck::index_t KBatch,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+
+    virtual std::array<int, 6> GetPreShuffleParameters() = 0;
+};
+
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/flush_cache.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename CDEShuffleBlockTransferScalarPerVectors,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA,
+          typename LDSTypeA                           = ComputeTypeA,
+          typename LDSTypeB                           = ComputeTypeB>
+struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
+    : public DeviceGemmMultipleDSplitKBPreShuffle<ALayout,
+                                                  BLayout,
+                                                  DsLayout,
+                                                  CLayout,
+                                                  ADataType,
+                                                  BDataType,
+                                                  DsDataType,
+                                                  CDataType,
+                                                  AElementwiseOperation,
+                                                  BElementwiseOperation,
+                                                  CElementwiseOperation>
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle<
+        ALayout,
+        BLayout,
+        DsLayout,
+        CLayout,
+        ADataType,
+        BDataType,
+        GemmAccDataType,
+        CShuffleDataType,
+        DsDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEShuffleBlockTransferScalarPerVectors,
+        BlkGemmPipeSched,
+        BlkGemmPipelineVer,
+        ComputeTypeA,
+        ComputeTypeB,
+        LDSTypeA,
+        LDSTypeB>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    std::array<int, 6> GetPreShuffleParameters() override
+    {
+        std::array<int, 6> preshuffle_params{NXdlPerWave,
+                                             GridwiseGemm::KRepeat,
+                                             GridwiseGemm::NWave,
+                                             GridwiseGemm::KLane,
+                                             GridwiseGemm::NLane,
+                                             GridwiseGemm::KPack};
+
+        return preshuffle_params;
+    }
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            index_t gdx, gdy, gdz;
+            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch);
+
+            float ave_time = 0;
+
+            index_t k_grain = arg.KBatch * KPerBlock;
+            index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
+
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
+
+            const auto Run = [&](const auto& kernel) {
+                if(stream_config.flush_cache)
+                {
+
+                    std::array<std::size_t, NumDTensor> DsSize;
+
+                    Argument arg_ = arg;
+
+                    const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
+                        arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
+                    const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
+                        arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
+
+                    auto size_a_buffer =
+                        a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
+                    auto size_b_buffer =
+                        b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
+
+                    const auto ds_grid_desc_m_n = GridwiseGemm::MakeDsGridDescriptor_M_N(
+                        arg_.M, arg_.MPadded, arg_.N, arg_.NPadded, arg_.StrideDs);
+
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+                        DsSize[i] = ds_grid_desc_m_n[i].GetElementSpaceSize() * sizeof(DDataType);
+                    });
+                    ck::utility::RotatingMemWrapperMultiD<Argument, DsDataType> rotating_mem(
+                        arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer, DsSize);
+                    rotating_mem.Print();
+
+                    auto run_flush_cache = [&]() {
+                        // flush icache
+                        ck::utility::flush_icache();
+                        // rotating mem
+                        rotating_mem.Next();
+                        // clear c mem
+                        if(arg_.KBatch > 1)
+                            hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
+                                                             0,
+                                                             arg_.M * arg_.N * sizeof(CDataType),
+                                                             stream_config.stream_id_));
+                    };
+
+                    ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
+                        stream_config,
+                        run_flush_cache,
+                        kernel,
+                        dim3(gdx, gdy, gdz),
+                        dim3(BlockSize),
+                        0,
+                        arg_);
+                }
+                else
+                {
+                    if(arg.KBatch > 1)
+                        hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
+                                                         0,
+                                                         arg.M * arg.N * sizeof(CDataType),
+                                                         stream_config.stream_id_));
+
+                    ave_time = launch_and_time_kernel(
+                        stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+                }
+            };
+
+            constexpr index_t minimum_occupancy =
+                BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave ? 1 : 2;
+
+            // static_assert(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3 &&
+            // has_main_k_block_loop, "only impl BlockGemmPipelineVersion::v3 and has mainloop right
+            // now");
+            if(has_main_k_block_loop)
+            {
+                // Tail number always full
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+                    if(arg.KBatch > 1)
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::AtomicAdd,
+                                minimum_occupancy,
+                                TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::AtomicAdd,
+                                minimum_occupancy,
+                                TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                    else
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::Set,
+                                minimum_occupancy,
+                                TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::Set,
+                                minimum_occupancy,
+                                TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                }
+                else
+                {
+                    throw std::runtime_error("todo: only v3 support now");
+                }
+            }
+            else
+            {
+                if(arg.KBatch > 1)
+                {
+                    const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
+                        GridwiseGemm,
+                        false,
+                        InMemoryDataOperationEnum::AtomicAdd,
+                        minimum_occupancy>;
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
+                        GridwiseGemm,
+                        false,
+                        InMemoryDataOperationEnum::Set,
+                        minimum_occupancy>;
+                    Run(kernel);
+                }
+            }
+
+            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)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
+        {
+            return false;
+        }
+
+        if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
+                                                       GemmSpec == GemmSpecialization::NKPadding ||
+                                                       GemmSpec == GemmSpecialization::MNKPadding ||
+                                                       GemmSpec == GemmSpecialization::KPadding))
+        {
+            return false;
+        }
+
+        if(arg.N % NPerBlock != 0 || arg.K % KPerBlock != 0)
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<index_t, NumDTensor> StrideDs,
+                             index_t StrideC,
+                             index_t KBatch,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{static_cast<const ADataType*>(p_a),
+                        static_cast<const BDataType*>(p_b),
+                        p_ds,
+                        static_cast<CDataType*>(p_c),
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideC,
+                        KBatch,
+                        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,
+                                                      std::array<const void*, NumDTensor> p_ds,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      std::array<ck::index_t, NumDTensor> StrideDs,
+                                                      index_t StrideC,
+                                                      index_t KBatch,
+                                                      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),
+                                          p_ds,
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideC,
+                                          KBatch,
+                                          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();
+
+        std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
+            {BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
+            {BlockGemmPipelineScheduler::Interwave, "Interwave"}};
+
+        std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
+            {BlockGemmPipelineVersion::v1, "v1"},
+            {BlockGemmPipelineVersion::v2, "v2"},
+            {BlockGemmPipelineVersion::v3, "v3"},
+            {BlockGemmPipelineVersion::v4, "v4"},
+            {BlockGemmPipelineVersion::v5, "v5"}};
+
+        // clang-format off
+        str << "DeviceGemmXdlUniversal"
+            << "<"
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << ">"
+            << " BlkSize: "
+            << BlockSize << ", "
+            << "BlkTile: "
+            << MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
+            << "WaveTile: "
+            << MPerXDL<<"x"<<NPerXDL << ", "
+            << "WaveMap: "
+            << MXdlPerWave<<"x" << NXdlPerWave<<", "
+            << "VmemReadVec: "
+            << ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
+            << "BlkGemmPipelineScheduler: "
+            << BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
+            << "BlkGemmPipelineVersion: "
+            << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
+            << "BlkGemmPipelinePrefetchStages: "
+            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -21,6 +21,12 @@ struct ColumnMajor : public BaseTensorLayout
 {
     static constexpr const char* name = "ColumnMajor";
 };
+
+struct MFMA : public BaseTensorLayout
+{
+    static constexpr const char* name = "MFMA";
+};
+
 } // namespace gemm
 
 namespace convolution {

include/ck/tensor_operation/gpu/element/element_wise_operation.hpp

@@ -283,6 +283,16 @@ struct MultiplyMultiply
         e = ck::type_convert<ck::half_t>(x0_f);
     }
 
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, int, float, float>(
+        ck::half_t& e, const int& c, const float& d0, const float& d1) const
+    {
+        const float x0_f =
+            ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
+
+        e = ck::type_convert<ck::half_t>(x0_f);
+    }
+
     template <>
     __host__ __device__ constexpr void operator()<ck::bhalf_t, int, float, float>(
         ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#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/block/blockwise_gemm_pipeline_xdlops_b_preshuffle.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_selector.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r3.hpp"
+
+#define DEBUG_LOG 0
+
+namespace ck {
+
+// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same
+// kernel function Blockers:
+// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on
+// two lds chunks.
+// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds
+// buffer when we declare __shared__ inside blkgemmpipe
+template <typename GridwiseGemm,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t MinimumOccupancy = 1,
+          TailNumber TailNum       = TailNumber::Even>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
+#endif
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    __shared__ char p_shared1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
+        karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
+        karg.p_ds_grid,
+        karg.p_c_grid,
+        p_shared,
+        p_shared1,
+        karg,
+        karg.a_element_op,
+        karg.b_element_op,
+        karg.c_element_op);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx9__))
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename CDEShuffleBlockTransferScalarPerVectors,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v4,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA,
+          typename LDSTypeA                           = ADataType,
+          typename LDSTypeB                           = BDataType>
+struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
+{
+    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>{};
+
+    static constexpr auto CShuffleBlockTransferScalarPerVector_NPerBlock =
+        CDEShuffleBlockTransferScalarPerVectors{}[I0];
+    // K1 should be Number<...>
+    static constexpr auto AK0Number       = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0Number       = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1Number       = Number<AK1Value>{};
+    static constexpr auto BK1Number       = Number<BK1Value>{};
+    static constexpr auto BlockSizeNumber = Number<BlockSize>{};
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using mfma_selector = MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>;
+    static constexpr index_t KPack =
+        math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk);
+    static constexpr index_t KLane =
+        mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops();
+    static constexpr index_t KRepeat = KPerBlock / KLane / KPack;
+    static constexpr index_t NLane   = NPerXdl;
+    static constexpr index_t NWave   = NPerBlock / NPerXdl / NXdlPerWave;
+    static_assert(NLane * NWave * KLane == BlockSize);
+    // static_assert(NXdlPerWave == 1, "only 1 validated now, tbd next week");
+
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch)
+    {
+        return std::make_tuple(Block2CTileMapDefault::CalculateGridSize(M, N), 1, KBatch);
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateBN0Shuffled(index_t N)
+    {
+        return math::integer_divide_ceil(N, NLane * NWave);
+    }
+    __host__ __device__ static auto CalculateBK0Shuffled(index_t K)
+    {
+        return math::integer_divide_ceil(K, KLane * KPack);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ __device__ static auto CalculateAK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / AK1Value);
+    }
+
+    __host__ __device__ static auto CalculateBK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / BK1Value);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * KPerBlock;
+    }
+
+    __host__ __device__ static auto CalculateKRead(index_t K, index_t K_Batch = 1)
+    {
+        constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+        auto K_t                = K_Batch * KReadVec;
+        return (K + K_t - 1) / K_t * KReadVec;
+    }
+
+    __host__ __device__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNBlock(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock);
+    }
+
+    template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+    __host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&)
+    {
+        constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
+        constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
+
+        return transform_tensor_descriptor(
+            TileDesc_K0_MN_K1{},
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
+                       make_unmerge_transform(make_tuple(
+                           Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
+        index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(M, MPad - M),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(MPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    __host__ __device__ static auto MakeBGridDescriptor_Preshuffled(index_t N0, index_t K0)
+    {
+        constexpr index_t NkSwizzleNumber = Number<BlockSize * KPack>{};
+        return make_naive_tensor_descriptor(make_tuple(N0, K0, NkSwizzleNumber),
+                                            make_tuple(K0 * NkSwizzleNumber, NkSwizzleNumber, I1));
+    }
+
+    __host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
+        index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(N, NPad - N),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(NPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor<MXdlPerWave, MWaves, MPerXdl>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        return MakeGemmMmaTileDescriptor<NXdlPerWave, NWave, NPerXdl>(BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ELayout>
+    __host__ __device__ static auto
+    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        // pad M and N
+        return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                           make_tuple(make_right_pad_transform(M, MPad - M),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}));
+#if 0
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(M, MPad - M),
+                                                          make_right_pad_transform(N, NPad - N)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+#endif
+    }
+
+    __host__ __device__ static auto MakeDsGridDescriptor_M_N(
+        index_t M, index_t MPad, index_t N, index_t NPad, std::array<index_t, NumDTensor> StrideDs)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                return MakeCGridDescriptor_M_N<DLayout>(M, MPad, N, NPad, StrideDs[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    template <typename DsGridDesc>
+    __device__ static constexpr auto MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDesc& ds_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n[i], MBlock, NBlock);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N(0, 0, 0, 0, {}))>;
+
+    struct Problem
+    {
+        __host__ __device__ Problem(index_t M_,
+                                    index_t N_,
+                                    index_t K_,
+                                    index_t StrideA_,
+                                    index_t StrideB_,
+                                    std::array<index_t, NumDTensor> StrideDs_,
+                                    index_t StrideC_,
+                                    index_t KBatch_)
+            : M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideDs{StrideDs_},
+              StrideC{StrideC_},
+              KBatch{KBatch_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              KRead{CalculateKRead(K_, KBatch_)},
+              KPadded{CalculateKPadded(K_, KBatch_)},
+              AK0{CalculateAK0Padded(K_, KBatch_)},
+              BK0{CalculateBK0Padded(K_, KBatch_)},
+              MBlock{CalculateMBlock(M_)},
+              NBlock{CalculateNBlock(N_)},
+              BN0Shuffled{CalculateBN0Shuffled(N_)},
+              BK0Shuffled{CalculateBK0Shuffled(K_)}
+        {
+        }
+
+        __host__ void Print() const
+        {
+            std::cout << "problem {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "KRead:" << KRead << ", "
+                      << "KP:" << KPadded << ", "
+                      << "AK0:" << AK0 << ", "
+                      << "BK0:" << BK0 << ", "
+                      << "MBlock: " << MBlock << ", "
+                      << "NBlock: " << NBlock << "}" << std::endl;
+        }
+
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        std::array<index_t, NumDTensor> StrideDs;
+        index_t StrideC;
+        index_t KBatch;
+        index_t MPadded;
+        index_t NPadded;
+        index_t KRead;
+        index_t KPadded;
+        index_t AK0;
+        index_t BK0;
+        index_t MBlock;
+        index_t NBlock;
+        // FOR PRESHUFFLE ONLY
+        index_t BN0Shuffled;
+        index_t BK0Shuffled;
+    };
+
+    // Argument
+    struct Argument : public tensor_operation::device::BaseArgument, public Problem
+    {
+        __host__ Argument(const ADataType* p_a_grid_,
+                          const BDataType* p_b_grid_,
+                          std::array<const void*, NumDTensor> p_ds_grid_,
+                          CDataType* p_c_grid_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          std::array<index_t, NumDTensor> StrideDs_,
+                          index_t StrideC_,
+                          index_t k_batch_,
+                          AElementwiseOperation a_element_op_,
+                          BElementwiseOperation b_element_op_,
+                          CElementwiseOperation c_element_op_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideDs_, StrideC_, k_batch_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_ds_grid{},
+              p_c_grid{p_c_grid_},
+              a_element_op{a_element_op_},
+              b_element_op{b_element_op_},
+              c_element_op{c_element_op_}
+        {
+
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType_ = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid(i) = static_cast<const DDataType_*>(p_ds_grid_[i]);
+            });
+        }
+
+        const ADataType* p_a_grid;
+        const BDataType* p_b_grid;
+        DsGridPointer p_ds_grid;
+        CDataType* p_c_grid;
+
+        const AElementwiseOperation a_element_op;
+        const BElementwiseOperation b_element_op;
+        const CElementwiseOperation c_element_op;
+    };
+
+    struct SplitKBatchOffset
+    {
+        __device__ SplitKBatchOffset(Argument& karg, index_t k_id)
+        {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                a_k_split_offset = k_id * karg.KRead;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                a_k_split_offset = k_id * karg.KRead * karg.StrideA;
+            }
+
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
+            {
+                b_k_split_offset = k_id * karg.KRead * karg.StrideB;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
+            {
+                // KPack * NLane * KLane * NWave * KRepeat * K0* NRepeat *  N0
+                b_k_split_offset = k_id * karg.KRead * NLane * NWave;
+            }
+
+            if(k_id < karg.KBatch - 1)
+            {
+                karg.K = karg.KRead;
+            }
+            else
+            {
+                karg.K = karg.K - karg.KRead * (karg.KBatch - 1);
+            }
+        }
+
+        index_t a_k_split_offset;
+        index_t b_k_split_offset;
+    };
+
+    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        if constexpr(ABlockLdsExtraM)
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock + ABlockLdsExtraM>{}, I1));
+        }
+        // xor tensor transformation request more unnecessary vgpr usage, would cause register spill
+        // in some cases.
+        else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            constexpr auto MLdsLayer        = 32 * 4 / KPerBlock / sizeof(LDSTypeA) < 1
+                                                  ? 1
+                                                  : 32 * 4 / KPerBlock / sizeof(LDSTypeA);
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
+                make_tuple(
+                    AK0Number * Number<MLdsLayer>{}, Number<MPerBlock / MLdsLayer>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock * MLdsLayer>{}, I1));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(make_xor_with_modulo_transform(make_tuple(
+                               Number<MPerBlock / MLdsLayer>{}, Number<AK0Number * MLdsLayer>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
+
+            constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number<MLdsLayer>{})),
+                           make_pass_through_transform(Number<MPerBlock / MLdsLayer>{}),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
+
+            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_ak0_mldslayer_m_ak1,
+                make_tuple(make_pass_through_transform(AK0Number),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<MPerBlock / MLdsLayer>{}, Number<MLdsLayer>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_ak0_m_ak1;
+        }
+        else // ColumnMajor A
+        {
+            // kfold and mpair dimension is not always required.
+            // more dimension in merge_transform increase the difficulty of generating immarg offset
+            // for compiler.
+            constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
+            constexpr auto M1 = MPerBlock / M0;
+
+            constexpr auto KThreadWrite     = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0);
+            constexpr auto K0PerThreadWrite = AK0Number / KThreadWrite;
+            constexpr auto KThreadRead      = 64 / MPerXdl;
+            constexpr auto K0PerThreadRead  = AK0Number / KThreadRead;
+
+            constexpr auto kfold = (AK1Number * M0 * sizeof(LDSTypeA) > 128)
+                                       ? 1
+                                       : 128 / (AK1Number * M0 * sizeof(LDSTypeA));
+            constexpr auto KThreadReadPerm =
+                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
+                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
+                    : KThreadRead;
+
+            // 1<=mpair<=n0
+            constexpr auto mpair = (AK1Number * MPerXdl * sizeof(LDSTypeA) > 128)
+                                       ? 1
+                                       : ((128 / (AK1Number * MPerXdl * sizeof(LDSTypeA))) > M0
+                                              ? M0
+                                              : 128 / (AK1Number * MPerXdl * sizeof(LDSTypeA)));
+
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                           Number<K0PerThreadWrite>{},
+                           Number<KThreadReadPerm * M1>{},
+                           Number<kfold * M0 / mpair>{},
+                           Number<mpair>{},
+                           AK1Number));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_xor_with_modulo_transform(
+                        make_tuple(Number<KThreadReadPerm * M1>{}, Number<kfold * M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
+
+            constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<M1>{})),
+                    make_unmerge_transform(make_tuple(Number<kfold>{}, Number<M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<0, 3>{},
+                           Sequence<4, 5>{},
+                           Sequence<6>{},
+                           Sequence<7>{}));
+
+            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_unmerged,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(Number<KThreadReadPerm>{},
+                                          Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                                          Number<kfold>{},
+                                          Number<K0PerThreadWrite>{})),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<M0 / mpair>{}, Number<mpair>{}, Number<M1>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_ak0_m_ak1;
+        }
+    }
+
+    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        if constexpr(BBlockLdsExtraN)
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
+                make_tuple(BK1Number, Number<KPerBlock + BBlockLdsExtraN>{}, I1));
+        }
+        else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+        {
+            // NLdsLayer * K0 as logical Bank
+            constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeB) < 1
+                                           ? 1
+                                           : 32 * 4 / KPerBlock / sizeof(LDSTypeB);
+            ;
+            constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
+                make_tuple(
+                    BK0Number * Number<NLdsLayer>{}, Number<NPerBlock / NLdsLayer>{}, BK1Number),
+                make_tuple(BK1Number, Number<KPerBlock * NLdsLayer>{}, I1));
+
+            constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
+                b_lds_block_desc,
+                make_tuple(make_xor_with_modulo_transform(make_tuple(
+                               Number<NPerBlock / NLdsLayer>{}, Number<BK0Number * NLdsLayer>{})),
+                           make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
+
+            constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor(
+                b_lds_block_desc_permuted,
+                make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number<NLdsLayer>{})),
+                           make_pass_through_transform(Number<NPerBlock / NLdsLayer>{}),
+                           make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
+
+            constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_lds_block_desc_bk0_nldslayer_n_bk1,
+                make_tuple(make_pass_through_transform(BK0Number),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<NPerBlock / NLdsLayer>{}, Number<NLdsLayer>{})),
+                           make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return b_lds_block_desc_bk0_n_bk1;
+        }
+        else // RowMajor B
+        {
+            constexpr auto N0 = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I1);
+            constexpr auto N1 = NPerBlock / N0;
+
+            constexpr auto KThreadWrite     = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I0);
+            constexpr auto K0PerThreadWrite = BK0Number / KThreadWrite;
+            constexpr auto KThreadRead      = 64 / NPerXdl;
+            constexpr auto K0PerThreadRead  = BK0Number / KThreadRead;
+
+            constexpr auto kfold = (BK1Number * N0 * sizeof(LDSTypeB) > 128)
+                                       ? 1
+                                       : 128 / (BK1Number * N0 * sizeof(LDSTypeB));
+            constexpr auto KThreadReadPerm =
+                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
+                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
+                    : KThreadRead;
+
+            // 1<=npair<=n0
+            constexpr auto npair = (BK1Number * NPerXdl * sizeof(LDSTypeB) > 128)
+                                       ? 1
+                                       : ((128 / (BK1Number * NPerXdl * sizeof(LDSTypeB))) > N0
+                                              ? N0
+                                              : 128 / (BK1Number * NPerXdl * sizeof(LDSTypeB)));
+
+            constexpr auto b_lds_block_desc = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                           Number<K0PerThreadWrite>{},
+                           Number<KThreadReadPerm * N1>{},
+                           Number<kfold * N0 / npair>{},
+                           Number<npair>{},
+                           BK1Number));
+
+            constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
+                b_lds_block_desc,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_xor_with_modulo_transform(
+                        make_tuple(Number<KThreadReadPerm * N1>{}, Number<kfold * N0 / npair>{})),
+                    make_pass_through_transform(Number<npair>{}),
+                    make_pass_through_transform(BK1Number)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
+
+            constexpr auto b_lds_block_desc_unmerged = transform_tensor_descriptor(
+                b_lds_block_desc_permuted,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<N1>{})),
+                    make_unmerge_transform(make_tuple(Number<kfold>{}, Number<N0 / npair>{})),
+                    make_pass_through_transform(Number<npair>{}),
+                    make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<0, 3>{},
+                           Sequence<4, 5>{},
+                           Sequence<6>{},
+                           Sequence<7>{}));
+
+            constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_lds_block_desc_unmerged,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(Number<KThreadReadPerm>{},
+                                          Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                                          Number<kfold>{},
+                                          Number<K0PerThreadWrite>{})),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<N0 / npair>{}, Number<npair>{}, Number<N1>{})),
+                           make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return b_lds_block_desc_bk0_n_bk1;
+        }
+    }
+
+    __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    using BlockwiseGemmPipe =
+        remove_cvref_t<decltype(BlockwiseGemmXdlops_pipeline_bpreshuffle<
+                                BlkGemmPipeSched,
+                                BlockSize,
+                                LDSTypeA,
+                                LDSTypeB,
+                                ComputeTypeA,
+                                AccDataType,
+                                decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()),
+                                decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()),
+                                decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(
+                                    GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())),
+                                decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(
+                                    GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())),
+                                ABlockTransferSrcScalarPerVector,
+                                BBlockTransferSrcScalarPerVector,
+                                MPerBlock,
+                                NPerBlock,
+                                KPerBlock,
+                                MPerXdl,
+                                NPerXdl,
+                                MXdlPerWave,
+                                NXdlPerWave,
+                                KPack>{})>;
+    __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max(a_block_space_size_aligned * sizeof(LDSTypeA),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ static constexpr bool CheckValidity(const Argument& karg)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
+                     !(is_same<tensor_layout::gemm::RowMajor, ALayout>::value))
+        {
+            if(!(karg.M % MPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
+                     (is_same<tensor_layout::gemm::RowMajor, BLayout>::value))
+        {
+            if(!(karg.N % NPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+
+            auto K_t = karg.KBatch * KPerBlock;
+            if(!(karg.K % K_t == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
+                          << karg.K << " " << __FILE__ << ":" << __LINE__
+                          << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+            auto K_t                = karg.KBatch * KReadVec;
+            auto KReadPadSplited    = math::integer_divide_ceil(karg.K, K_t) * KReadVec;
+            if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(karg.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M (" << karg.M
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(karg.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N (" << karg.N
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N (" << karg.N
+                          << ") value is not a multiple of "
+                             "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                          << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
+                          << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M (" << karg.M
+                          << ") value is not a multiple of "
+                             "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                          << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
+                          << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value);
+
+        if constexpr(BlkGemmPipelineVer != BlockGemmPipelineVersion::v1)
+        {
+            if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages)
+            {
+                return false;
+            }
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockHasHotloop(num_loop);
+    }
+
+    __host__ __device__ static constexpr TailNumber CalculateKBlockLoopTailNum(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockLoopTailNum(num_loop);
+    }
+
+    template <typename CGridDesc>
+    __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    // if arch = gfx942
+    using Block2CTileMapDefault = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run(const ADataType* p_a_grid,
+                               const BDataType* p_b_grid,
+                               DsGridPointer& p_ds_grid,
+                               CDataType* p_c_grid,
+                               void* p_shared,
+                               void* p_shared1,
+                               const Problem& problem,
+                               AElementwiseOperation a_element_op,
+                               BElementwiseOperation b_element_op,
+                               CElementwiseOperation c_element_op)
+    {
+        const auto block_2_ctile_map = Block2CTileMapDefault{problem.M, problem.N, 4};
+        Run<Block2CTileMapDefault, HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+            p_a_grid,
+            p_b_grid,
+            p_ds_grid,
+            p_c_grid,
+            p_shared,
+            p_shared1,
+            problem,
+            a_element_op,
+            b_element_op,
+            c_element_op,
+            block_2_ctile_map);
+    }
+
+    template <typename Block2CTileMap,
+              bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run(const ADataType* p_a_grid,
+                               const BDataType* p_b_grid,
+                               DsGridPointer& p_ds_grid,
+                               CDataType* p_c_grid,
+                               void* p_shared,
+                               void* p_shared1,
+                               const Problem& problem,
+                               AElementwiseOperation a_element_op,
+                               BElementwiseOperation b_element_op,
+                               CElementwiseOperation c_element_op,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+            problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
+        const auto b_grid_desc_bpreshuffled =
+            MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N<CLayout>(
+            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                c_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bpreshuffled.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+        const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
+
+        // 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_m_id * MPerBlock);
+
+        // N0, K0, Blocksize*KPack
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0Number, MPerBlock, AK1Number>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                LDSTypeA,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                BlockwiseGemmPipe::GlobalBufferNum>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // using BThreadClusterLengths = Sequence<1, 1, BlockSize>;
+        // using BBlockTransferClusterArrangeOrder = Sequence<0, 1, 2>;
+        // B matrix blockwise copy
+        auto b_blockwise_copy = ThreadGroupTensorSliceTransfer_v4r1<
+            ThisThreadBlock,
+            BElementwiseOperation,
+            ck::tensor_operation::element_wise::PassThrough,
+            InMemoryDataOperationEnum::Set,
+            Sequence<NXdlPerWave, KRepeat, KPack * BlockSize>,
+            Sequence<1, 1, BlockSize>, // BThreadClusterLengths,
+            Sequence<0, 1, 2>,         // BBlockTransferClusterArrangeOrder,
+            BDataType,
+            LDSTypeB,
+            decltype(b_grid_desc_bpreshuffled),
+            decltype(b_block_desc_bk0_n_bk1),
+            Sequence<0, 1, 2>, // BBlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2>,
+            BBlockTransferSrcVectorDim,
+            2,
+            BBlockTransferSrcScalarPerVector,
+            BBlockTransferDstScalarPerVector_BK1,
+            1,
+            1,
+            BThreadTransferSrcResetCoordinateAfterRun,
+            true,
+            2>(b_grid_desc_bpreshuffled,
+               make_multi_index(n_block_data_idx_on_grid, 0, 0),
+               b_element_op,
+               b_block_desc_bk0_n_bk1,
+               make_multi_index(0, 0, 0),
+               ck::tensor_operation::element_wise::PassThrough{});
+
+        // LDS allocation for A and B: be careful of alignment
+        // Cast after lds
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+        auto a_block_buf1 = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeA*>(p_shared1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KRepeat, 0);
+
+        // Blockwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
+        auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
+        auto c_thread_buf            = blockwise_gemm_pipeline.GetCThreadBuffer();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
+                                                                         a_block_desc_ak0_m_ak1,
+                                                                         a_blockwise_copy,
+                                                                         a_grid_buf,
+                                                                         a_block_buf,
+                                                                         a_block_buf1,
+                                                                         a_block_slice_copy_step,
+                                                                         b_grid_desc_bpreshuffled,
+                                                                         b_blockwise_copy,
+                                                                         b_grid_buf,
+                                                                         b_block_slice_copy_step,
+                                                                         c_thread_buf,
+                                                                         num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_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_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_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_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            using EDataType = CDataType;
+
+            const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N(
+                problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs);
+
+            const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+            const auto ds_grid_buf = generate_tuple(
+                [&](auto i) {
+                    return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                        p_ds_grid[i], ds_grid_desc_m_n[i].GetElementSpaceSize());
+                },
+                Number<NumDTensor>{});
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                c_grid_desc_mblock_mperblock_nblock_nperblock;
+
+            using CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock =
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock;
+            const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
+
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                            // support arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
+                Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
+                3,                    // index_t SrcVectorDim,
+                3,                    // index_t DstVectorDim,
+                CDEShuffleBlockTransferScalarPerVectors,
+                CShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_m_id, 0, block_n_id, 0)),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(c_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp

@@ -267,6 +267,14 @@ struct ThreadwiseTensorSliceTransfer_v3r1
         }
     }
 
+    template <typename SeqIdx, index_t ThreadScratchId = 0>
+    __device__ constexpr auto
+    GetSrcThreadScratchIdx(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        using vector_t = typename vector_type_maker<SrcData, SrcScalarPerVector>::type::type;
+        return src_thread_scratch_tuple_(thread_scratch_id).template GetAsType<vector_t>(SeqIdx{});
+    }
+
     template <index_t ThreadScratchId>
     __device__ void
     TransferDataFromSrcThreadScratchToDstThreadScratch(Number<ThreadScratchId> thread_scratch_id)

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#if(defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     F16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>&
+        instances);
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     F16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>&
+        instances);
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     F16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>&
+        instances);
+
+#if 0
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          F8,
+                                                          F8,
+                                                          Tuple<F32, F32>,
+                                                          F16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances);
+
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          F8,
+                                                          F8,
+                                                          Tuple<F32, F32>,
+                                                          F16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances);
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          F8,
+                                                          F8,
+                                                          Tuple<F32, F32>,
+                                                          F16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances);
+#endif
+#endif
+
+#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>&
+        instances);
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>&
+        instances);
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>&
+        instances);
+#if 0
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>&
+        instances);
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>&
+        instances);
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>&
+        instances);
+#endif
+#endif
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmMultipleDSplitKBPreShuffle<
+        ALayout,
+        BLayout,
+        Tuple<Row, Col>,
+        CLayout,
+        ADataType,
+        BDataType,
+        Tuple<F32, F32>,
+        CDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::MultiplyMultiply>>
+{
+    using DeviceOp =
+        DeviceGemmMultipleDSplitKBPreShuffle<ALayout,
+                                             BLayout,
+                                             Tuple<Row, Col>,
+                                             CLayout,
+                                             ADataType,
+                                             BDataType,
+                                             Tuple<F32, F32>,
+                                             CDataType,
+                                             ck::tensor_operation::element_wise::PassThrough,
+                                             ck::tensor_operation::element_wise::PassThrough,
+                                             ck::tensor_operation::element_wise::MultiplyMultiply>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+// TODO: Add MFMA layout into tensor layout
+#if(defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
+        if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
+                     is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances(
+                    op_ptrs);
+#if 0
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_padding_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_padding_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_padding_instances(
+                    op_ptrs);
+#endif
+            }
+        }
+#endif
+
+#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
+        if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
+                     is_same_v<CDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances(
+                    op_ptrs);
+#if 0
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_padding_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_padding_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_padding_instances(
+                    op_ptrs);
+#endif
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+# ONLY XDL_KERNELS
+set(GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES)
+
+list(APPEND GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES 
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp
+        # device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_padding_instance.cpp
+        # device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_padding_instance.cpp
+        # device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_padding_instance.cpp
+
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp
+        # device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_padding_instance.cpp
+        # device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_padding_instance.cpp
+        # device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_padding_instance.cpp
+        )
+
+set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+# set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_padding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+# set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_padding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+# set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_padding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+
+set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+# set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_padding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+# set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_padding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+# set_source_files_properties(device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_padding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+
+add_instance_library(device_gemm_multiply_multiply_weight_preshuffle_instance ${GEMM_MULTIPLY_MULTIPLY_WEIGHT_PRESHUFFLE_INSTANCES})

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8   = f8_t;
+using BF16 = bhalf_t;
+using F32  = float;
+
+using Row = tensor_layout::gemm::RowMajor;
+using Col = tensor_layout::gemm::ColumnMajor;
+
+template <index_t... Is>
+using S = Sequence<Is...>;
+
+using PassThrough      = element_wise::PassThrough;
+using MultiplyMultiply = element_wise::MultiplyMultiply;
+
+static constexpr auto GemmDefault    = GemmSpecialization::Default;
+static constexpr auto GemmKPadding   = GemmSpecialization::KPadding;
+static constexpr auto GemmMNPadding  = GemmSpecialization::MNPadding;
+static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
+
+static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
+static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
+
+template <GemmSpecialization GemmSpec>
+using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_instances =
+    std::tuple<
+// clang-format off
+        //################################| ALayout| BLayout|         DsLayout| ELayout|AData| BData|          DsData| EData| AccData| Cshuffle|           A|           B|              C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
+        //################################|        |        |                 |        | Type|  Type|            Type|  Type|    Type|     Type| Elementwise| Elementwise|    Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
+        //################################|        |        |                 |        |     |      |                |      |        |         |   Operation|   Operation|      Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
+        //################################|        |        |                 |        |     |      |                |      |        |         |            |            |               |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
+#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   256,    128,   128,  16,  16,  32,   32,    8,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,    128,   128,  16,  16,  32,   32,    4,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    128,   128,  16,  16,  32,   32,    2,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    128,   128,  16,  16,  32,   32,    1,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        // N 256
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   256,    256,   128,  16,  16,  32,   32,    8,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,    256,   128,  16,  16,  32,   32,    4,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    256,   128,  16,  16,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    256,   128,  16,  16,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        // N 512
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    512,   128,  16,  16,  32,   32,    2,    4,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    512,   128,  16,  16,  32,   32,    1,    4,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
+#endif
+        // clang-format on
+        >;
+
+template <GemmSpecialization GemmSpec>
+using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_instances =
+    std::tuple<
+// clang-format off
+        //################################| ALayout| BLayout|         DsLayout| ELayout|AData| BData|          DsData| EData| AccData| Cshuffle|           A|           B|              C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
+        //################################|        |        |                 |        | Type|  Type|            Type|  Type|    Type|     Type| Elementwise| Elementwise|    Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
+        //################################|        |        |                 |        |     |      |                |      |        |         |   Operation|   Operation|      Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
+        //################################|        |        |                 |        |     |      |                |      |        |         |            |            |               |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
+#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,    128,   256,  16,  16,  32,   32,    4,    1,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    128,   256,  16,  16,  32,   32,    2,    1,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    128,   256,  16,  16,  32,   32,    1,    1,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    128,   512,  16,  16,  32,   32,    2,    1,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    128,   512,  16,  16,  32,   32,    1,    1,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        // N 256
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,    256,   256,  16,  16,  32,   32,    4,    2,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    256,   256,  16,  16,  32,   32,    2,    2,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    256,   256,  16,  16,  32,   32,    1,    2,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    256,   512,  16,  16,  32,   32,    2,    2,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    256,   512,  16,  16,  32,   32,    1,    2,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        // N 512
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    512,   256,  16,  16,  32,   32,    2,    4,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    512,   256,  16,  16,  32,   32,    1,    4,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
+#endif
+        // clang-format on
+        >;
+
+template <GemmSpecialization GemmSpec>
+using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances =
+    std::tuple<
+// clang-format off
+        //################################| ALayout| BLayout|         DsLayout| ELayout|AData| BData|          DsData| EData| AccData| Cshuffle|           A|           B|              C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
+        //################################|        |        |                 |        | Type|  Type|            Type|  Type|    Type|     Type| Elementwise| Elementwise|    Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
+        //################################|        |        |                 |        |     |      |                |      |        |         |   Operation|   Operation|      Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
+        //################################|        |        |                 |        |     |      |                |      |        |         |            |            |               |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
+#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    64,   256,  16,  16,  16,   16,    1,    1,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    128,   256,  16,  16,  16,   16,    1,    2,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    256,   256,  16,  16,  16,   16,    1,    4,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    512,   256,  16,  16,  16,   16,    1,    8,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    64,   512,  16,  16,  16,   16,    1,    1,     S<32, 8, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    128,   512,  16,  16,  16,   16,    1,    2,     S<32, 8, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, BF16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    256,   512,  16,  16,  16,   16,    1,    4,     S<32, 8, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
+#endif
+        // clang-format on
+        >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_instances<
+            GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_padding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p1_instances<
+            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_instances<
+            GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_padding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p2_instances<
+            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances<
+            GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_padding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     BF16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_bf16_mk_mfma_mn_p3_instances<
+            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8  = f8_t;
+using F16 = half_t;
+using F32 = float;
+
+using Row = tensor_layout::gemm::RowMajor;
+using Col = tensor_layout::gemm::ColumnMajor;
+
+template <index_t... Is>
+using S = Sequence<Is...>;
+
+using PassThrough      = element_wise::PassThrough;
+using MultiplyMultiply = element_wise::MultiplyMultiply;
+
+static constexpr auto GemmDefault    = GemmSpecialization::Default;
+static constexpr auto GemmKPadding   = GemmSpecialization::KPadding;
+static constexpr auto GemmMNPadding  = GemmSpecialization::MNPadding;
+static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
+
+static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
+static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
+
+template <GemmSpecialization GemmSpec>
+using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_instances =
+    std::tuple<
+// clang-format off
+        //################################| ALayout| BLayout|         DsLayout| ELayout|AData| BData|          DsData| EData| AccData| Cshuffle|           A|           B|              C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
+        //################################|        |        |                 |        | Type|  Type|            Type|  Type|    Type|     Type| Elementwise| Elementwise|    Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
+        //################################|        |        |                 |        |     |      |                |      |        |         |   Operation|   Operation|      Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
+        //################################|        |        |                 |        |     |      |                |      |        |         |            |            |               |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
+#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   256,    128,   128,  16,  16,  32,   32,    8,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,    128,   128,  16,  16,  32,   32,    4,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    128,   128,  16,  16,  32,   32,    2,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    128,   128,  16,  16,  32,   32,    1,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        // N 256
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   256,    256,   128,  16,  16,  32,   32,    8,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,    256,   128,  16,  16,  32,   32,    4,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    256,   128,  16,  16,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    256,   128,  16,  16,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        // N 512
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    512,   128,  16,  16,  32,   32,    2,    4,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    512,   128,  16,  16,  32,   32,    1,    4,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
+#endif
+        // clang-format on
+        >;
+
+template <GemmSpecialization GemmSpec>
+using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_instances =
+    std::tuple<
+// clang-format off
+        //################################| ALayout| BLayout|         DsLayout| ELayout|AData| BData|          DsData| EData| AccData| Cshuffle|           A|           B|              C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
+        //################################|        |        |                 |        | Type|  Type|            Type|  Type|    Type|     Type| Elementwise| Elementwise|    Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
+        //################################|        |        |                 |        |     |      |                |      |        |         |   Operation|   Operation|      Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
+        //################################|        |        |                 |        |     |      |                |      |        |         |            |            |               |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
+#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,    128,   256,  16,  16,  32,   32,    4,    1,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    128,   256,  16,  16,  32,   32,    2,    1,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    128,   256,  16,  16,  32,   32,    1,    1,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    128,   512,  16,  16,  32,   32,    2,    1,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    128,   512,  16,  16,  32,   32,    1,    1,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        // N 256
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,   128,    256,   256,  16,  16,  32,   32,    4,    2,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    256,   256,  16,  16,  32,   32,    2,    2,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    256,   256,  16,  16,  32,   32,    1,    2,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    256,   512,  16,  16,  32,   32,    2,    2,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    256,   512,  16,  16,  32,   32,    1,    2,     S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        // N 512
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    64,    512,   256,  16,  16,  32,   32,    2,    4,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,    32,    512,   256,  16,  16,  32,   32,    1,    4,     S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,     S<8, 8, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
+#endif
+        // clang-format on
+        >;
+
+template <GemmSpecialization GemmSpec>
+using device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_instances =
+    std::tuple<
+// clang-format off
+        //################################| ALayout| BLayout|         DsLayout| ELayout|AData| BData|          DsData| EData| AccData| Cshuffle|           A|           B|              C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1|MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|                         Block-wiseGemm|               Block-wiseGemm|
+        //################################|        |        |                 |        | Type|  Type|            Type|  Type|    Type|     Type| Elementwise| Elementwise|    Elementwise|Specialization|  Size| Block| Block| Block|    |    | XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
+        //################################|        |        |                 |        |     |      |                |      |        |         |   Operation|   Operation|      Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
+        //################################|        |        |                 |        |     |      |                |      |        |         |            |            |               |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
+#if defined(__gfx94__) || defined(CK_USE_GFX94) || defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH)
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    64,   256,  16,  16,  16,   16,    1,    1,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    128,   256,  16,  16,  16,   16,    1,    2,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    256,   256,  16,  16,  16,   16,    1,    4,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    512,   256,  16,  16,  16,   16,    1,    8,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    64,   512,  16,  16,  16,   16,    1,    1,     S<32, 8, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    128,   512,  16,  16,  16,   16,    1,    2,     S<32, 8, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
+        DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle<  Row,     Col,     Tuple<Row, Col>,  Row,    F8,    F8,    Tuple<F32, F32>, F16,  F32,     F32,     PassThrough, PassThrough, MultiplyMultiply,    GemmSpec,   256,     16,    256,   512,  16,  16,  16,   16,    1,    4,     S<32, 8, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<32, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,    S<4, 4, 1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
+#endif
+        // clang-format on
+        >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck