create several new files for int4/int8 dequantization

example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_b_scale.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <initializer_list>
+#include <iostream>
+#include <numeric>
+
+#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_scale.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/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.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/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using FP16 = ck::half_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 A0DataType = FP16;
+// using A1DataType       = F32;
+// using B0DataType       = FP8;
+// using B1DataType       = F32;
+using B0DataType       = uint8_t;
+using B1DataType       = FP16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = FP16;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using D0Layout = Row;
+using D1Layout = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// static constexpr ck::index_t Scale_Block_M = 128;
+static constexpr ck::index_t Scale_Block_N = 128;
+static constexpr ck::index_t Scale_Block_K = 128;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_BScale_Xdl_CShuffle_V3
+    // clang-format off
+         <Row, Col, DsLayout, ELayout,
+          A0DataType, B0DataType, B1DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, 
+          AElementOp,  BElementOp, CDEElementOp, GemmSpec,
+          256, Scale_Block_N, Scale_Block_K,
+          128, 
+          128, 128,
+        //   16, 16,
+          8, 8,
+          16,   16,
+          4,    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,
+          S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0,
+          S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0,
+          1,    2,  S<1, 32, 1, 8>,  S<8, 8, 1>,
+          ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, 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 StrideE = N;
+
+    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 == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    // ck::index_t Scale_Stride_AM = (K + Scale_Block_K - 1) / Scale_Block_K;
+    ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
+
+    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<A1DataType> a1_m_k(f_host_tensor_descriptor((M + Scale_Block_M - 1) / Scale_Block_M,
+    //                                                    (K + Scale_Block_K - 1) / Scale_Block_K,
+    //                                                    Scale_Stride_AM,
+    //                                                    A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor((K + Scale_Block_K - 1) / Scale_Block_K,
+                                                       (N + Scale_Block_N - 1) / Scale_Block_N,
+                                                       Scale_Stride_BN,
+                                                       B0Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    // std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+#if 1
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        // a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
+        break;
+    case 2:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        // a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
+        break;
+    case 3:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        // a1_m_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
+        break;
+    case 4:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        // a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        // a1_m_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
+    }
+#endif
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    // DeviceMem a1_device_buf(sizeof(A1DataType) * a1_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    // a1_device_buf.ToDevice(a1_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    b1_device_buf.ToDevice(b1_k_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();
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument  = device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(),
+                                           b0_device_buf.GetDeviceBuffer(),
+                                           std::array<const void*, NumDTensor>{},
+                                           e_device_buf.GetDeviceBuffer(),
+                                           M,
+                                           N,
+                                           K,
+                                           StrideA,
+                                           StrideB,
+                                           std::array<ck::index_t, NumDTensor>{},
+                                           StrideE,
+                                           //    a1_device_buf.GetDeviceBuffer(),
+                                           b1_device_buf.GetDeviceBuffer(),
+                                           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, 20, 50});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+        Tensor<float> a_m_k({M, K});
+        Tensor<float> b_k_n({K, N});
+
+        for(int m = 0; m < M; m++)
+        {
+            for(int k = 0; k < K; k++)
+            {
+                a_m_k(m, k) = ck::type_convert<float>(a0_m_k(m, k));
+            }
+        }
+
+        for(int n = 0; n < N; n++)
+        {
+            for(int k = 0; k < K; k++)
+            {
+                b_k_n(k, n) = ck::type_convert<float>(b0_k_n(k, n)) *
+                              b1_k_n(k / Scale_Block_K, n / Scale_Block_N);
+            }
+        }
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<float,
+                                                                                float,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+#if 1
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                e_m_n_host_result(m, n) = ck::type_convert<EDataType>(c_m_n(m, n));
+            }
+        }
+#endif
+
+        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, "Error: Incorrect results!", 5e-2, 5e-2)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+// #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp"
+// #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_b_scale.hpp"
+
+namespace ck {
+
+enum struct BlockGemmPipelineVersion
+{
+    v1, // Naive
+    v2, // Mem
+    v3, // Comp
+};
+
+template <BlockGemmPipelineVersion BlkGemmPipelineVer,
+          BlockGemmPipelineScheduler BlkGemmPipeSche,
+          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>
+constexpr auto BlockGemmBScalePipeline_Selector()
+{
+    printf("I'm Here\n");
+    return BlockwiseGemmXdlops_pipeline_v3_b_scale<BlkGemmPipeSche,
+                                                   BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>{};
+}
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_b_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 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: 1
+// LocalPreFetchStages: 1
+// LocalSharedMemoryBuffer: 1
+
+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_v3_b_scale
+{
+};
+
+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_v3_b_scale<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::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::b_block_desc_n0_n1_n2_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__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        ignore = num_loop;
+        return TailNumber::Full;
+    }
+
+    __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_read_inst_b =
+            HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16
+                ? HotLoopInstList::B_LDS_Read_Inst_Num
+                : HotLoopInstList::B_LDS_Read_Inst_Num / 2;
+
+        constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
+        constexpr auto num_ds_write_inst_b = HotLoopInstList::B_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 = 4;
+        // HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
+        constexpr auto ds_read_b_issue_cycle = 4;
+        // HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 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 ds_read_b_mfma_rate =
+            (mfma_cycle - 4 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
+
+        constexpr auto num_dsread_a_mfma =
+            (num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
+        constexpr auto num_dsread_b_mfma =
+            (num_ds_read_inst_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
+
+        // stage 1
+        // Separate this part?
+        // constexpr auto num_mfma_per_ds_read = sizeof(ComputeDataType) / sizeof(ADataType) >
+        //                                               sizeof(ComputeDataType) / sizeof(BDataType)
+        //                                           ? sizeof(ComputeDataType) / sizeof(ADataType)
+        //                                           : sizeof(ComputeDataType) / sizeof(BDataType);
+        constexpr auto num_mfma_stage1 = num_mfma_inst - (num_dsread_a_mfma + num_dsread_b_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;
+        constexpr auto num_dswrite_per_issue_b = num_ds_write_inst_b / num_buffer_load_inst_b;
+
+        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;
+            static_for<0, num_dswrite_per_issue_b, 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_b, 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
+        });
+
+        static_for<0, num_dsread_b_mfma, 1>{}([&](auto i) {
+            if constexpr((num_ds_read_inst_b - (i + 1) * ds_read_b_mfma_rate) >=
+                         ds_read_b_mfma_rate)
+            {
+                __builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read
+            }
+            else
+            {
+                __builtin_amdgcn_sched_group_barrier(0x100,
+                                                     num_ds_read_inst_b - (num_dsread_b_mfma - 1) *
+                                                                              ds_read_b_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 BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer,
+            //   typename AScaleGridBuffer,
+            //   typename AScaleGridDesc,
+            //   typename AScaleThreadDesc,
+            //   typename AScaleThreadTransfer,
+            //   typename AScaleThreadTransferStep,
+              typename BScaleGridBuffer,
+              typename BScaleGridDesc,
+              typename BScaleThreadDesc,
+              typename BScaleThreadTransfer,
+              typename BScaleThreadTransferStep>
+    __device__ void Run(
+        // ABlockCopy
+        const AGridDesc& a_grid_desc,
+        const ABlockDesc& a_block_desc,
+        ABlockTransfer& a_blockwise_copy,
+        const AGridBuffer& a_grid_buf,
+        ABlockBuffer& a_block_buf,
+        const ABlockTransferStep& a_block_copy_step,
+        // BBlockCopy
+        const BGridDesc& b_grid_desc,
+        const BBlockDesc& b_block_desc,
+        BBlockTransfer& b_blockwise_copy,
+        const BGridBuffer& b_grid_buf,
+        BBlockBuffer& b_block_buf,
+        const BBlockTransferStep& b_block_copy_step,
+        // CThread
+        CThreadBuffer& c_thread_buf,
+        // // AScaleThreadCopy
+        // const AScaleGridDesc& a_scale_grid_desc,
+        // const AScaleThreadDesc& a_scale_thread_desc,
+        // AScaleThreadTransfer& a_scale_thread_copy,
+        // const AScaleGridBuffer& a_scale_grid_buf,
+        // const AScaleThreadTransferStep& a_scale_thread_copy_step,
+        // BScaleThreadCopy
+        const BScaleGridDesc& b_scale_grid_desc,
+        const BScaleThreadDesc& b_scale_thread_desc,
+        BScaleThreadTransfer& b_scale_thread_copy,
+        const BScaleGridBuffer& b_scale_grid_buf,
+        const BScaleThreadTransferStep& b_scale_thread_copy_step,
+        // num_loop
+        index_t num_loop,
+        index_t num_loop_per_scale) const
+    {
+        __builtin_amdgcn_sched_barrier(0);
+        // assume kperblock = scaleblockk
+        ignore            = num_loop_per_scale;
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+        // auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+        //     a_scale_thread_desc.GetElementSpaceSize());
+        // auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+        //     b_scale_thread_desc.GetElementSpaceSize());
+        auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ck::half_t>(
+            b_scale_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);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // a_scale_thread_copy.Run(a_scale_grid_desc,
+        //                         a_scale_grid_buf,
+        //                         a_scale_thread_desc,
+        //                         make_tuple(I0, I0),
+        //                         a_scale_thread_buf);
+
+        b_scale_thread_copy.Run(b_scale_grid_desc,
+                                b_scale_grid_buf,
+                                b_scale_thread_desc,
+                                make_tuple(I0, I0),
+                                b_scale_thread_buf);
+
+        // a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
+        // Local prefill 1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        // Global prefetch 2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
+
+        // 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_buf,
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, k0, I0),
+                                   a_thread_buf);
+            });
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                   make_tuple(n0, I0, I0, Number<k0 * BMmaKStride>{}),
+                                   b_block_buf,
+                                   b_thread_desc_,
+                                   make_tuple(n0, I0, k0, I0),
+                                   b_thread_buf);
+            });
+        });
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                block_sync_lds();
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                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, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        c_thread_buf_per_scale.Clear();
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            vector_type<ComputeDataType, KPack> a_thread_vec;
+                            vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                            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))>{}];
+                                b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, I0, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<ComputeDataType,
+                                                     xdlops_gemm.K1PerXdlops>::type;
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                        });
+                        static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
+                            c_thread_buf(Number<c_offset>{}) +=
+                                c_thread_buf_per_scale[Number<t>{}] *
+                                // type_convert<AccDataType>(a_scale_thread_buf[I0]) *
+                                type_convert<AccDataType>(b_scale_thread_buf[I0]);
+                        });
+                    });
+                });
+
+                block_sync_lds();
+                static_for<0, KRepeat, 1>{}([&](auto k) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k, I0),
+                                           a_thread_buf);
+                    });
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf,
+                                           b_thread_desc_,
+                                           make_tuple(n0, I0, k, I0),
+                                           b_thread_buf);
+                    });
+                });
+                // a_scale_thread_copy.Run(a_scale_grid_desc,
+                //                         a_scale_grid_buf,
+                //                         a_scale_thread_desc,
+                //                         make_tuple(I0, I0),
+                //                         a_scale_thread_buf);
+
+                b_scale_thread_copy.Run(b_scale_grid_desc,
+                                        b_scale_grid_buf,
+                                        b_scale_thread_desc,
+                                        make_tuple(I0, I0),
+                                        b_scale_thread_buf);
+
+                // a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc, a_scale_thread_copy_step);
+                b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc, b_scale_thread_copy_step);
+                HotLoopScheduler();
+                __builtin_amdgcn_sched_barrier(0);
+                i += 1;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        if constexpr(TailNum == TailNumber::Full)
+        {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    c_thread_buf_per_scale.Clear();
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        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))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                                 b_thread_vec.template AsType<mfma_input_type>(),
+                                                 c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                    });
+                    static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
+                        c_thread_buf(Number<c_offset>{}) +=
+                            c_thread_buf_per_scale[Number<t>{}] *
+                            // type_convert<AccDataType>(a_scale_thread_buf[I0]) *
+                            type_convert<AccDataType>(b_scale_thread_buf[I0]);
+                    });
+                });
+            });
+            __builtin_amdgcn_sched_barrier(0);
+        }
+    }
+
+    protected:
+    using Base::a_thread_copy_;
+    using Base::a_thread_desc_;
+    using Base::b_thread_copy_;
+    using Base::b_thread_desc_;
+    using Base::c_thread_desc_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1_b_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/utility/common_header.hpp"
+// #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1_b_scale.hpp"
+
+namespace ck {
+
+/**
+ * @brief Blockwise data transfer
+ *
+ * This version does following things to avoid scratch memory issue
+ * 1. Use StaticallyIndexedArray instead of C array for thread buffer
+ * 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+ * 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+ *
+ */
+template <typename ThreadGroup,
+          typename SrcElementwiseOperation,
+          typename DstElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename BlockSliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool ThreadTransferSrcResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun,
+          index_t NumThreadScratch = 1>
+struct ThreadGroupTensorSliceTransfer_v4r1_b_scale
+{
+    static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = BlockSliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v4r1_b_scale(
+        const SrcDesc& src_desc,
+        const Index& src_block_slice_origin,
+        const SrcElementwiseOperation& src_element_op,
+        const DstDesc& dst_desc,
+        const Index& dst_block_slice_origin,
+        const DstElementwiseOperation& dst_element_op)
+        : threadwise_transfer_(src_desc,
+                               make_zero_multi_index<nDim>(),
+                               src_element_op,
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<BlockSliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_block_slice_origin)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename SrcBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDesc& src_desc,
+                            const SrcBuffer& src_buf,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunRead(src_desc, src_buf, thread_scratch_id);
+        }
+    }
+
+    template <typename DstBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDesc& dst_desc,
+                             DstBuffer& dst_buf,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunWrite(dst_desc, dst_buf, thread_scratch_id);
+        }
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, index_t ThreadScratchId>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf,
+                        Number<ThreadScratchId> thread_scratch_id)
+    {
+        RunRead(src_desc, src_buf, thread_scratch_id);
+        RunWrite(dst_desc, dst_buf, thread_scratch_id);
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v3r1_b_scale<decltype(thread_slice_lengths),
+                                                   SrcElementwiseOperation,
+                                                   DstElementwiseOperation,
+                                                   DstInMemOp,
+                                                   SrcData,
+                                                   DstData,
+                                                   SrcDesc,
+                                                   DstDesc,
+                                                   SrcDimAccessOrder,
+                                                   DstDimAccessOrder,
+                                                   SrcVectorDim,
+                                                   DstVectorDim,
+                                                   SrcScalarPerVector,
+                                                   DstScalarPerVector,
+                                                   SrcScalarStrideInVector,
+                                                   DstScalarStrideInVector,
+                                                   ThreadTransferSrcResetCoordinateAfterRun,
+                                                   ThreadTransferDstResetCoordinateAfterRun,
+                                                   NumThreadScratch>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : 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 AScaleType,
+          typename BDataType,
+          typename BScaleType,
+          typename DsDataType,
+          typename EDataType,
+        //   index_t ScaleBlockM,
+          index_t ScaleBlockN,
+          index_t ScaleBlockK,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGemmMultipleD_ABScale : 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,
+                        const ck::index_t M,
+                        const ck::index_t N,
+                        const ck::index_t K,
+                        const ck::index_t StrideA,
+                        const ck::index_t StrideB,
+                        const std::array<ck::index_t, NumDTensor> StrideDs,
+                        const ck::index_t StrideE,
+                        // const void* p_a_scale,
+                        const void* p_b_scale,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_b_scale.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_scale.hpp"
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename CLayout,
+          typename ADataType,
+          //   typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename DsDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          //   index_t ScaleBlockM,
+          index_t ScaleBlockN,
+          index_t ScaleBlockK,
+          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_BScale_Xdl_CShuffle_V3
+    : public DeviceGemmMultipleD_ABScale<ALayout,
+                                         BLayout,
+                                         DsLayout,
+                                         CLayout,
+                                         ADataType,
+                                         //  AScaleDataType,
+                                         BDataType,
+                                         BScaleDataType,
+                                         DsDataType,
+                                         CDataType,
+                                         //  ScaleBlockM,
+                                         ScaleBlockN,
+                                         ScaleBlockK,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CElementwiseOperation>
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultiD_BScale_xdl_cshuffle_v3<
+        ALayout,
+        BLayout,
+        DsLayout,
+        CLayout,
+        ADataType,
+        BDataType,
+        GemmAccDataType,
+        CShuffleDataType,
+        DsDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        BlockSize,
+        // ScaleBlockM,
+        ScaleBlockN,
+        ScaleBlockK,
+        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;
+
+    // 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(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 &&
+                 MPerBlock * NPerBlock / BlockSize > 64)
+                    ? 1
+                    : 2;
+
+            if(has_main_k_block_loop)
+            {
+                // Tail number always 1
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 ||
+                             BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                        true,
+                                                        InMemoryDataOperationEnum::Set,
+                                                        minimum_occupancy>;
+                        Run(kernel);
+                    }
+                }
+                // Tail number could be One to Seven
+                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
+                {
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
+                        {
+                            const auto kernel =
+                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                            true,
+                                                            InMemoryDataOperationEnum::Set,
+                                                            minimum_occupancy,
+                                                            TailNumber::One>;
+                            Run(kernel);
+                        }
+                        else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
+                                TailNumber::Full)
+                        {
+                            const auto kernel =
+                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                            true,
+                                                            InMemoryDataOperationEnum::Set,
+                                                            minimum_occupancy,
+                                                            TailNumber::Full>;
+                            Run(kernel);
+                        }
+
+                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
+                        {
+                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
+                            {
+                                const auto kernel =
+                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                                true,
+                                                                InMemoryDataOperationEnum::Set,
+                                                                minimum_occupancy,
+                                                                TailNumber::Two>;
+                                Run(kernel);
+                            }
+                        }
+
+                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
+                        {
+                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
+                               TailNumber::Three)
+                            {
+                                const auto kernel =
+                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                                true,
+                                                                InMemoryDataOperationEnum::Set,
+                                                                minimum_occupancy,
+                                                                TailNumber::Three>;
+                                Run(kernel);
+                            }
+                        }
+
+                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
+                        {
+                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
+                               TailNumber::Four)
+                            {
+                                const auto kernel =
+                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                                true,
+                                                                InMemoryDataOperationEnum::Set,
+                                                                minimum_occupancy,
+                                                                TailNumber::Four>;
+                                Run(kernel);
+                            }
+                        }
+
+                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
+                        {
+                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
+                               TailNumber::Five)
+                            {
+                                const auto kernel =
+                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                                true,
+                                                                InMemoryDataOperationEnum::Set,
+                                                                minimum_occupancy,
+                                                                TailNumber::Five>;
+                                Run(kernel);
+                            }
+                        }
+
+                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
+                        {
+                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
+                            {
+                                const auto kernel =
+                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                                true,
+                                                                InMemoryDataOperationEnum::Set,
+                                                                minimum_occupancy,
+                                                                TailNumber::Six>;
+                                Run(kernel);
+                            }
+                        }
+
+                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
+                        {
+                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
+                               TailNumber::Seven)
+                            {
+                                const auto kernel =
+                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                                true,
+                                                                InMemoryDataOperationEnum::Set,
+                                                                minimum_occupancy,
+                                                                TailNumber::Seven>;
+                                Run(kernel);
+                            }
+                        }
+                    }
+                }
+            }
+            else
+            {
+                // Tail number always 1
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3<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(ScaleBlockN % NPerBlock != 0 || ScaleBlockK != KPerBlock)
+        {
+            printf("Return 1\n");
+            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;
+        }
+
+        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,
+                             const index_t M,
+                             const index_t N,
+                             const index_t K,
+                             const index_t StrideA,
+                             const index_t StrideB,
+                             const std::array<index_t, NumDTensor> StrideDs,
+                             const index_t StrideC,
+                             //  const void* p_a_scale,
+                             const void* p_b_scale,
+                             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,
+                        // static_cast<const AScaleDataType*>(p_a_scale),
+                        static_cast<const BScaleDataType*>(p_b_scale),
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_c,
+                        const index_t M,
+                        const index_t N,
+                        const index_t K,
+                        const index_t StrideA,
+                        const index_t StrideB,
+                        const std::array<ck::index_t, NumDTensor> StrideDs,
+                        const index_t StrideC,
+                        // const void* p_a_scale,
+                        const void* p_b_scale,
+                        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,
+                                          //   static_cast<const AScaleDataType*>(p_a_scale),
+                                          static_cast<const BScaleDataType*>(p_b_scale),
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        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"}};
+
+        // 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/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#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/block/blockwise_gemm_pipeline_xdlops_b_scale_selector.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/utility/common_header.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_v4r1_b_scale.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.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::Full>
+__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(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_a_grid,
+        karg.p_b_grid,
+        karg.p_ds_grid,
+        karg.p_c_grid,
+        // karg.p_a_scale_grid,
+        karg.p_b_scale_grid,
+        p_shared,
+        karg,
+        karg.a_element_op,
+        karg.b_element_op,
+        karg.c_element_op);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+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 ScaleBlockM,
+          index_t ScaleBlockN,
+          index_t ScaleBlockK,
+          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::v1,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA,
+          typename LDSTypeA                           = ADataType,
+          typename LDSTypeB                           = BDataType>
+struct GridwiseGemmMultiD_BScale_xdl_cshuffle_v3
+{
+    // using AScaleType = float;
+    // using BScaleType = float;
+    using BScaleType = ck::half_t;
+
+    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 index_t NumDTensor = DsDataType::Size();
+
+    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());
+
+    static constexpr index_t KPack = math::max(
+        math::lcm(AK1Number, BK1Number),
+        MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>::selected_mfma.k_per_blk);
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, KBatch);
+    }
+
+    __host__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ 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__ 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__ 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__ 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__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock);
+    }
+
+    __host__ 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>{}));
+    }
+
+    __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;
+        }
+    }
+
+    __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&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor<NXdlPerWave, NWaves, 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));
+            }
+        }();
+
+        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;
+        }
+    }
+
+    __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__ 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_)}
+        {
+        }
+
+        __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;
+    };
+
+    // 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_,
+                          //   const AScaleType* p_a_scale_grid_,
+                          const BScaleType* p_b_scale_grid_,
+                          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_},
+              //   p_a_scale_grid{p_a_scale_grid_},
+              p_b_scale_grid{p_b_scale_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 AScaleType* p_a_scale_grid;
+        const BScaleType* p_b_scale_grid;
+
+        const AElementwiseOperation a_element_op;
+        const BElementwiseOperation b_element_op;
+        const CElementwiseOperation c_element_op;
+    };
+
+    struct SplitKBatchOffset
+    {
+        __device__ SplitKBatchOffset(Argument& karg)
+        {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                a_k_split_offset = blockIdx.z * karg.KRead;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                a_k_split_offset = blockIdx.z * karg.KRead * karg.M;
+            }
+
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
+            {
+                b_k_split_offset = blockIdx.z * karg.KRead * karg.N;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
+            {
+                b_k_split_offset = blockIdx.z * karg.KRead;
+            }
+
+            if(blockIdx.z < static_cast<uint32_t>(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 index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        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(BlockGemmBScalePipeline_Selector<
+                                BlkGemmPipelineVer,
+                                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();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // 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);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.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) +
+                          b_block_space_size_aligned * sizeof(LDSTypeB)),
+                         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))
+        {
+            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))
+        {
+            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__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockHasHotloop(num_loop);
+    }
+
+    __host__ 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 Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
+    // using Block2CTileMap = BlockToCTileMap_3DGrid_KSplit<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,
+                               //    const AScaleType* p_a_scale_grid,
+                               const BScaleType* p_b_scale_grid,
+                               void* p_shared,
+                               const Problem& problem,
+                               AElementwiseOperation a_element_op,
+                               BElementwiseOperation b_element_op,
+                               CElementwiseOperation c_element_op)
+    {
+        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_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
+            problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N<CLayout>(
+            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
+
+        // const auto a_scale_grid_desc_am_ak = make_naive_tensor_descriptor(
+        //     make_tuple(math::integer_divide_ceil(problem.M, ScaleBlockM),
+        //                math::integer_divide_ceil(problem.K, ScaleBlockK)),
+        //     make_tuple(math::integer_divide_ceil(problem.K, ScaleBlockK), 1));
+        const auto b_scale_grid_desc_bn_ak = make_naive_tensor_descriptor(
+            make_tuple(math::integer_divide_ceil(problem.N, ScaleBlockN),
+                       math::integer_divide_ceil(problem.K, ScaleBlockK)),
+            make_tuple(math::integer_divide_ceil(problem.K, ScaleBlockK), 1));
+
+        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_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // const auto a_scale_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+        //     p_a_scale_grid, a_scale_grid_desc_am_ak.GetElementSpaceSize());
+
+        const auto b_scale_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_scale_grid, b_scale_grid_desc_bn_ak.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4};
+
+        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);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        // 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_b_scale<
+            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{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy = ThreadGroupTensorSliceTransfer_v4r1_b_scale<
+            ThisThreadBlock,
+            BElementwiseOperation,
+            ck::tensor_operation::element_wise::PassThrough,
+            InMemoryDataOperationEnum::Set,
+            Sequence<BK0Number, NPerBlock, BK1Number>,
+            BBlockTransferThreadClusterLengths_BK0_N_BK1,
+            BBlockTransferThreadClusterArrangeOrder,
+            BDataType,
+            LDSTypeB,
+            decltype(b_grid_desc_bk0_n_bk1),
+            decltype(b_block_desc_bk0_n_bk1),
+            BBlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2>,
+            BBlockTransferSrcVectorDim,
+            2,
+            BBlockTransferSrcScalarPerVector,
+            BBlockTransferDstScalarPerVector_BK1,
+            1,
+            1,
+            BThreadTransferSrcResetCoordinateAfterRun,
+            true,
+            BlockwiseGemmPipe::GlobalBufferNum>(b_grid_desc_bk0_n_bk1,
+                                                make_multi_index(0, n_block_data_idx_on_grid, 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
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        // 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 b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeB*>(p_shared) +
+                a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB),
+            b_block_desc_bk0_n_bk1.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(KPerBlock / BK1Number, 0, 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);
+
+        const index_t ScaleSliceSizeM = 1;
+        const index_t ScaleSliceSizeN = 1;
+        const index_t ScaleSliceSizeK = 1;
+
+        // constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed(
+        //     make_tuple(Number<ScaleSliceSizeM>{}, Number<ScaleSliceSizeK>{}));
+
+        constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<ScaleSliceSizeM>{}, Number<ScaleSliceSizeK>{}));
+
+        // auto a_scale_thread_copy =
+        //     ThreadwiseTensorSliceTransfer_v2<AScaleType,
+        //                                      AScaleType,
+        //                                      decltype(a_scale_grid_desc_am_ak),
+        //                                      decltype(a_scale_thread_desc),
+        //                                      Sequence<ScaleSliceSizeM, ScaleSliceSizeK>,
+        //                                      Sequence<0, 1>,
+        //                                      1,
+        //                                      1,
+        //                                      1,
+        //                                      false>(
+        //         a_scale_grid_desc_am_ak, make_multi_index(block_m_id * MPerBlock / ScaleBlockM,
+        //         0));
+
+        auto b_scale_thread_copy =
+            ThreadwiseTensorSliceTransfer_v2<BScaleType,
+                                             BScaleType,
+                                             decltype(b_scale_grid_desc_bn_ak),
+                                             decltype(b_scale_thread_desc),
+                                             Sequence<ScaleSliceSizeN, ScaleSliceSizeK>,
+                                             Sequence<0, 1>,
+                                             1,
+                                             1,
+                                             1,
+                                             false>(
+                b_scale_grid_desc_bn_ak, make_multi_index(block_n_id * NPerBlock / ScaleBlockN, 0));
+
+        // constexpr auto a_scale_thread_slice_copy_step = make_multi_index(0, 1);
+        constexpr auto b_scale_thread_slice_copy_step = make_multi_index(0, 1);
+
+        const index_t num_k_block_per_scale = ScaleBlockK / 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_slice_copy_step,
+            b_grid_desc_bk0_n_bk1,
+            b_block_desc_bk0_n_bk1,
+            b_blockwise_copy,
+            b_grid_buf,
+            b_block_buf,
+            b_block_slice_copy_step,
+            c_thread_buf,
+
+            // a_scale_grid_desc_am_ak,
+            // a_scale_thread_desc,
+            // a_scale_thread_copy,
+            // a_scale_grid_buf,
+            // a_scale_thread_slice_copy_step,
+
+            b_scale_grid_desc_bn_ak,
+            b_scale_thread_desc,
+            b_scale_thread_copy,
+            b_scale_grid_buf,
+            b_scale_thread_slice_copy_step,
+
+            num_k_block_main_loop,
+            num_k_block_per_scale);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // 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_b_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor/static_tensor.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/is_detected.hpp"
+
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
+
+namespace ck {
+
+namespace detail {
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor
+template <index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector>
+struct lambda_scalar_per_access_for_src_and_dst_idle
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        if(i == SrcVectorDim && i == DstVectorDim)
+        {
+            return math::lcm(SrcScalarPerVector, DstScalarPerVector);
+        }
+        else if(i == SrcVectorDim)
+        {
+            return SrcScalarPerVector;
+        }
+        else if(i == DstVectorDim)
+        {
+            return DstScalarPerVector;
+        }
+        else
+        {
+            return 1;
+        }
+    }
+};
+
+} // namespace detail
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          typename SrcElementwiseOperation,
+          typename DstElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun, // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v3r1_b_scale
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v3r1_b_scale(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const SrcElementwiseOperation& src_element_op,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const DstElementwiseOperation& dst_element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          src_element_op_(src_element_op),
+          dst_element_op_(dst_element_op)
+    {
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDesc& src_desc,
+                            const SrcBuffer& src_buf,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(src_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_scalar_per_access;
+            }();
+
+            constexpr auto src_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            using src_vector_type = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            // copy data from src_vector_container into src_thread_scratch_
+            src_thread_scratch_tuple_(thread_scratch_id)
+                .template SetAsType<src_vector_t>(
+                    src_data_idx_seq, src_vector_container.template AsType<src_vector_t>()[I0]);
+
+            constexpr auto move_on_dim = [&]() constexpr {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }();
+
+            // move src coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <index_t ThreadScratchId>
+    __device__ void
+    TransferDataFromSrcThreadScratchToDstThreadScratch(Number<ThreadScratchId> thread_scratch_id)
+    {
+#if !CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE
+        static_ford<SliceLengths>{}([&](auto idx) {
+            // convert from SrcData to DstData here
+            dst_thread_scratch_(idx) =
+                type_convert<DstData>(src_thread_scratch_tuple_[thread_scratch_id][idx]);
+        });
+#else
+        // sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
+        // TODO make this logic more generic for more sub-dword datatype
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     ((is_same<half_t, remove_cvref_t<SrcData>>::value &&
+                       is_same<half_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<int8_t, remove_cvref_t<SrcData>>::value &&
+                       is_same<int8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
+        {
+            // each transpose does
+            // DstScalarPerVector # of src vectors in src_thread_scratch_
+            // SrcScalarPerVector # of dst vectors in dst_thread_scratch_
+            constexpr index_t num_src_vector = Number<DstScalarPerVector>{};
+            constexpr index_t num_dst_vector = Number<SrcScalarPerVector>{};
+
+            // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose
+            // TODO: make this logic generic for all scenario
+            static_assert(SrcVectorDim != DstVectorDim, "wrong");
+
+            constexpr auto src_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+            constexpr auto dst_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+            constexpr auto scalar_per_access = generate_sequence(
+                detail::lambda_scalar_per_access_for_src_and_dst_idle<SrcVectorDim,
+                                                                      SrcScalarPerVector,
+                                                                      DstVectorDim,
+                                                                      DstScalarPerVector>{},
+                Number<nDim>{});
+
+            constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+            static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+                constexpr auto data_idx = access_idx * scalar_per_access;
+
+                constexpr auto data_idx_seq = generate_sequence_v2(
+                    [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
+
+                using src_vector_t = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+                using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
+
+                // get DstScalarPerVector # of read-only references to src vectors from
+                // src_thread_scratch_
+                const auto src_vector_refs = generate_tie(
+                    [&](auto i) -> const src_vector_t& {
+                        // i increment corresponds to movement in DstVectorDim
+                        return src_thread_scratch_tuple_[thread_scratch_id].GetVectorTypeReference(
+                            data_idx_seq + i * dst_scalar_step_in_vector);
+                    },
+                    Number<num_src_vector>{});
+
+                // get SrcScalarPerVector # of references to dst vectors from dst_thread_scratch_
+                auto dst_vector_refs = generate_tie(
+                    [&](auto i) -> dst_vector_t& {
+                        // i increment corresponds to movement in SrcVectorDim
+                        return dst_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * src_scalar_step_in_vector);
+                    },
+                    Number<num_dst_vector>{});
+
+                // do data transpose
+                transpose_vectors<SrcData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+
+        // Do fast numeric convert
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access_for_src_and_dst_idle<SrcVectorDim,
+                                                                  SrcScalarPerVector,
+                                                                  DstVectorDim,
+                                                                  DstScalarPerVector>{},
+            Number<nDim>{});
+
+        constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+        using src_vector_type = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+        using src_vector_t    = typename src_vector_type::type;
+
+        using src_converted_vector_type = vector_type_maker_t<DstData, SrcScalarPerVector>;
+        using src_converted_vector_t    = typename src_converted_vector_type::type;
+        // Vector-wise type convert
+        static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+            auto src_vector_container = src_vector_type{
+                src_thread_scratch_tuple_[thread_scratch_id].template GetAsType<src_vector_t>(
+                    access_idx)};
+
+            auto src_converted_vector_container =
+                src_converted_vector_type{fast_numeric_converter(src_vector_container)};
+
+            src_converted_thread_scratch_.template SetAsType<src_converted_vector_t>(
+                access_idx,
+                src_converted_vector_container.template AsType<src_converted_vector_t>()[I0]);
+        });
+
+        // Element-scale operation, expect packed multiplication
+        static_ford<SliceLengths>{}([&](auto idx) {
+            DstData dst_v;
+            // constexpr auto scale_idx = Sequence<I0, idx.At(1), I0>{};
+            // printf("Tid: %03d, scale: %04x\n", get_thread_local_1d_id(),
+            // *(reinterpret_cast<const uint16_t*>(&scale_thread_scratch_[scale_idx])));
+            src_element_op_(dst_v, src_converted_thread_scratch_[idx]);
+            dst_thread_scratch_(idx) = dst_v;
+        });
+#endif
+    }
+
+    template <typename DstBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDesc& dst_desc,
+                             DstBuffer& dst_buf,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // if there is transpose, it's done here
+        // if there is oob check, it's done here
+        // TODO move this elsewhere
+        TransferDataFromSrcThreadScratchToDstThreadScratch(thread_scratch_id);
+
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        // src scalar per access on each dim
+        // TODO: don't use this
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, forward_step_idx);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(dst_desc, backward_step_idx);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_dst_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_scalar_per_access;
+            }();
+
+            constexpr auto dst_data_idx_seq = generate_sequence_v2(
+                [&](auto i) { return Number<dst_data_idx[i]>{}; }, Number<dst_data_idx.Size()>{});
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            using dst_vector_type = vector_type_maker_t<DstData, DstScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            // copy data from dst_thread_scratch_ into dst_vector_container
+            auto dst_vector_container = dst_vector_type{
+                dst_thread_scratch_.template GetAsType<dst_vector_t>(dst_data_idx_seq)};
+
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                DstData dst_v;
+
+                // apply DstElementwiseOperation
+                dst_element_op_(dst_v, dst_vector_container.template AsType<DstData>()[i]);
+
+                dst_vector_container.template AsType<DstData>()(i) = dst_v;
+            });
+
+            // copy data from dst_vector_container to dst_buf
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            constexpr auto move_on_dim = [&]() constexpr {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }();
+
+            // move dst coord
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(src_access_lengths), Number<SrcScalarPerVector>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto GetSrcOOBThreadScratchDescriptor()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        return make_naive_tensor_descriptor_packed(src_access_lengths);
+    }
+
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(dst_access_lengths), Number<DstScalarPerVector>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    private:
+    static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
+    static constexpr auto src_oob_thread_scratch_desc_ =
+        decltype(GetSrcThreadScratchDescriptor()){};
+    static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
+
+    using SrcThreadScratch =
+        StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                        DstData, // apply data_convert with SrcThreadScratch
+                                        SrcScalarPerVector,
+                                        decltype(src_thread_scratch_desc_),
+                                        true>;
+
+    using SrcOOBThreadScratch =
+        StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                        bool, // apply data_convert with SrcThreadScratch
+                                        1,
+                                        decltype(src_oob_thread_scratch_desc_),
+                                        true>;
+
+    // Registers, contain fast converted data
+    using SrcThreadConvertedScratch =
+        StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                        DstData,
+                                        SrcScalarPerVector,
+                                        decltype(src_thread_scratch_desc_),
+                                        true>;
+
+    using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                                             DstData,
+                                                             DstScalarPerVector,
+                                                             decltype(dst_thread_scratch_desc_),
+                                                             true>;
+
+    using FastTypeConverter = tensor_operation::element_wise::
+        FastNumericArrayConverter<SrcData, DstData, SrcScalarPerVector>;
+
+    StaticallyIndexedArray<SrcThreadScratch, NumThreadScratch> src_thread_scratch_tuple_;
+    StaticallyIndexedArray<SrcOOBThreadScratch, NumThreadScratch> src_oob_thread_scratch_tuple_;
+    SrcThreadConvertedScratch src_converted_thread_scratch_;
+
+    DstThreadScratch dst_thread_scratch_;
+    FastTypeConverter fast_numeric_converter;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const SrcElementwiseOperation src_element_op_;
+    const DstElementwiseOperation dst_element_op_;
+};
+
+} // namespace ck