Merge branch 'develop' into amd-develop

.azuredevops/rocm-ci.yml

+resources:
+  repositories:
+  - repository: pipelines_repo
+    type: github
+    endpoint: ROCm
+    name: ROCm/ROCm
+
+variables:
+- group: common
+- template: /.azuredevops/variables-global.yml@pipelines_repo
+
+trigger:
+  batch: true
+  branches:
+    include:
+    - develop
+  paths:
+    exclude:
+    - .github
+    - docs
+    - '.*.y*ml'
+    - '*.md'
+    - Jenkinsfile
+    - LICENSE
+
+pr:
+  autoCancel: true
+  branches:
+    include:
+    - develop
+  paths:
+    exclude:
+    - .github
+    - docs
+    - '.*.y*ml'
+    - '*.md'
+    - Jenkinsfile
+    - LICENSE
+  drafts: false
+
+jobs:
+  - template: ${{ variables.CI_COMPONENT_PATH }}/composable_kernel.yml@pipelines_repo

Jenkinsfile

@@ -911,9 +911,8 @@ pipeline {
                         execute_args = """ cmake -D CMAKE_PREFIX_PATH=/opt/rocm \
                                            -D CMAKE_CXX_COMPILER="${build_compiler()}" \
                                            -D CMAKE_BUILD_TYPE=Release \
-                                           -D GPU_TARGETS="gfx90a;gfx1030;gfx1101" \
                                            -D INSTANCES_ONLY=ON \
-                                           -DCMAKE_CXX_FLAGS=" -O3 " .. && make -j32 """
+                                           -DCMAKE_CXX_FLAGS=" -O3 " .. && make -j64 """
                    }
                     steps{
                         buildHipClangJobAndReboot(setup_cmd: "",  build_cmd: "", no_reboot:true, build_type: 'Release', execute_cmd: execute_args)

docs/sphinx/requirements.in

-rocm-docs-core==1.1.2
+rocm-docs-core==1.2.0
 sphinxcontrib-bibtex==2.6.2

docs/sphinx/requirements.txt

@@ -103,7 +103,7 @@ requests==2.31.0
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core==1.1.2
+rocm-docs-core==1.2.0
     # via -r requirements.in
 six==1.16.0
     # via

example/65_gemm_multiply_multiply/CMakeLists.txt

+add_example_executable(example_gemm_multiply_multiply_xdl_fp16 gemm_multiply_multiply_xdl_fp16.cpp)

example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using FP8 = ck::f8_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = FP8;
+using B0DataType       = FP8;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = F16;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using D0Layout = Row;
+using D1Layout = Col;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+struct MultiplyMultiply
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, float, float>(
+        ck::half_t& e, const float& c, const float& d0, const float& d1) const
+    {
+        const float x0_f = c * d0 * d1;
+
+        e = ck::type_convert<ck::half_t>(x0_f);
+    }
+};
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MultiplyMultiply;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3
+    // clang-format off
+///######|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     DsData|     EData|     AccData|         CShuffle|           A|           B|          CDE|           GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+///######|         |         |         |        |       Type|       Type|       Type|      Type|        Type|         DataType| Elementwise| Elementwise|  Elementwise| Spacialization|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+///######|         |         |         |        |           |           |           |          |            |                 |   Operation|   Operation|    Operation|               |      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+///######|         |         |         |        |           |           |           |          |            |                 |            |            |             |               |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |    S<C, D0, D1>|
+///###### RRR
+      ///<      Row,      Row, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   256,   128,    64,  16,   4,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,               S<1, 32, 1, 8>,      S<8, 8, 1>,  ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>;
+///###### RCR
+         <      Row,      Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   256,   128,    64,  16,  16,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,     S<4, 64, 1>,    S<1, 0, 2>,     S<1, 0, 2>,             2,              16,             16,          0,          1,           1,               S<1, 32, 1, 8>,      S<8, 8, 1>,  ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideD = 0;
+    ck::index_t StrideE = N;
+
+    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 == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{0, 2});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{0, 2});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{0, 2});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+
+    constexpr auto I0 = ck::Number<0>{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(),
+                               b0_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer(),
+                                                                   d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, NumDTensor>{I0, I0},
+                               StrideE,
+                               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<CShuffleDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B0DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/ck_tile/01_fmha/README.md

@@ -34,6 +34,7 @@ args:
                  if not equal to h, then this is GQA/MQA case
           -s    seqlen_q. if group-mode, means the average value of seqlen_q (default:3328)
                  total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary
+                 also with "-s=s0,s1,s2..." comma seperated int to set per batch seqlen(group-mode)
         -s_k    seqlen_k, -1 means equal to s (default:-1)
           -d    head dim for q, k (default:128)
         -d_v    head dim for v, -1 means equal to d (default:-1)

example/ck_tile/01_fmha/fmha_fwd.cpp

@@ -44,11 +44,18 @@ auto create_args(int argc, char* argv[])
                 "-1",
                 "num of head, for k/v, -1 means equal to h\n"
                 "if not equal to h, then this is GQA/MQA case")
-        .insert("s",
-                "3328",
-                "seqlen_q. if group-mode, means the average value of seqlen_q\n"
-                "total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary")
+        .insert(
+            "s",
+            "3328",
+            "seqlen_q. if group-mode, means the average value of seqlen_q\n"
+            "total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary\n"
+            "also with \"-s=s0,s1,s2...\" comma seperated int to set per batch seqlen(group-mode)")
         .insert("s_k", "-1", "seqlen_k, -1 means equal to s")
+        .insert("s_kpad",
+                "-1",
+                "seqlen_k stride between 2 tokens, currently used in group-mode only\n"
+                "for kv-cache case, each batch [1,s,h,d]/[1,h,s,d] can have a stride\n"
+                "along seqlen, instead of packed. same as xformer kv_padding")
         .insert("d", "128", "head dim for q, k")
         .insert("d_v", "-1", "head dim for v, -1 means equal to d")
         .insert("scale_s",
@@ -103,6 +110,7 @@ auto create_args(int argc, char* argv[])
                 "11939",
                 "random seed used for initializing input tensors. 0 for "
                 "non-deterministic seed")
+        .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
         .insert("warmup", "5", "number of iterations before benchmark the kernel")
         .insert("repeat", "20", "number of iterations to benchmark the kernel");
 
@@ -177,10 +185,20 @@ bool run(const ck_tile::ArgParser& arg_parser)
         return false;
     }
 
-    ck_tile::index_t seqlen_q = arg_parser.get_int("s");
-    ck_tile::index_t seqlen_k = arg_parser.get_int("s_k");
-    if(seqlen_k < 0)
-        seqlen_k = seqlen_q;
+    auto [seqlen_qs, seqlen_ks, seqlen_kpads] = decode_seqlen(mode,
+                                                              batch,
+                                                              arg_parser.get_str("s"),
+                                                              arg_parser.get_str("s_k"),
+                                                              arg_parser.get_str("s_kpad"));
+
+#if 0
+    // clang-format off
+    std::cout << "seqlen_qs:"; for(auto xx : seqlen_qs) { std::cout << xx << ","; } std::cout << std::endl;
+    std::cout << "seqlen_ks:"; for(auto xx : seqlen_ks) { std::cout << xx << ","; } std::cout << std::endl;
+    std::cout << "seqlen_kpads:"; for(auto xx : seqlen_kpads) { std::cout << xx << ","; } std::cout << std::endl;
+    // clang-format on
+#endif
+
     ck_tile::index_t hdim_q = arg_parser.get_int("d");
     ck_tile::index_t hdim_v = arg_parser.get_int("d_v");
     if(hdim_v < 0)
@@ -229,7 +247,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     bool lse            = arg_parser.get_bool("lse");
 
     bias_info bias = bias_info::decode(arg_parser.get_str("bias"));
-    mask_info mask = mask_info::decode(arg_parser.get_str("mask"), seqlen_q, seqlen_k);
+    mask_info mask = mask_info::decode(
+        arg_parser.get_str("mask"), seqlen_qs[0], seqlen_ks[0]); // TODO: we don't need x/y anymore
 
     std::string init_method      = arg_parser.get_str("init");
     std::optional<uint32_t> seed = arg_parser.get_uint32("seed");
@@ -242,11 +261,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
     int stream_repeat = arg_parser.get_int("repeat");
     bool kname        = arg_parser.get_bool("kname");
 
-    ck_tile::stream_config stream_config{
-        nullptr, true, /* log_level = */ (kname ? 1 : 0), stream_warmup, stream_repeat};
+    ck_tile::stream_config stream_config{nullptr,
+                                         true,
+                                         /* log_level = */ (kname ? 1 : 0),
+                                         stream_warmup,
+                                         stream_repeat,
+                                         arg_parser.get_str("timer") == std::string("gpu")};
 
-    const auto seqstart_q_host = generate_seqstarts(mode, batch, seqlen_q);
-    const auto seqstart_k_host = generate_seqstarts(mode, batch, seqlen_k);
+    const auto seqstart_q_host              = to_seqstarts(seqlen_qs);
+    const auto seqstart_k_host              = to_seqstarts(seqlen_ks);
+    const auto seqstart_k_with_padding_host = to_seqstarts(seqlen_kpads);
 
     using TypeConfig = FmhaFwdTypeConfig<DataType>;
 
@@ -302,9 +326,11 @@ bool run(const ck_tile::ArgParser& arg_parser)
     // host memory for storing all the tensor elements
     const ck_tile::index_t shape_batch = (mode == mode_enum::batch ? batch : 1);
     const ck_tile::index_t shape_seqlen_q =
-        (mode == mode_enum::batch ? seqlen_q : seqstart_q_host.back());
+        (mode == mode_enum::batch ? seqlen_qs[0] : seqstart_q_host.back());
     const ck_tile::index_t shape_seqlen_k =
-        (mode == mode_enum::batch ? seqlen_k : seqstart_k_host.back());
+        (mode == mode_enum::batch ? seqlen_ks[0]
+                                  : (seqlen_kpads[0] < 0 ? seqstart_k_host.back()
+                                                         : seqstart_k_with_padding_host.back()));
 
     ck_tile::HostTensor<QDataType> q_host(
         get_lengths(i_perm, shape_batch, nhead, shape_seqlen_q, hdim_q));
@@ -407,6 +433,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
     ck_tile::DeviceMem o_buf(o_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t));
     ck_tile::DeviceMem seqstart_k(seqstart_k_host.size() * sizeof(int32_t));
+    ck_tile::DeviceMem seqlen_k_buf(seqlen_kpads[0] < 0 ? 0 : seqlen_ks.size() * sizeof(int32_t));
     ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes());
 
     q_buf.ToDevice(q_host.data());
@@ -414,7 +441,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
     v_buf.ToDevice(v_host.data());
     bias_buf.ToDevice(bias_host.data());
     seqstart_q.ToDevice(seqstart_q_host.data());
-    seqstart_k.ToDevice(seqstart_k_host.data());
+    seqstart_k.ToDevice(seqlen_kpads[0] < 0 ? seqstart_k_host.data()
+                                            : seqstart_k_with_padding_host.data());
+    seqlen_k_buf.ToDevice(seqlen_kpads[0] < 0 ? nullptr : seqlen_ks.data());
     alibi_slope_buf.ToDevice(alibi_slope_host.data());
 
     // clang-format off
@@ -430,7 +459,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
     const std::string prec = arg_parser.get_str("prec");
 
     std::cout << "[" << prec << "|" << mode << "|" << io_layout(i_perm, o_perm) << "] b:" << batch
-              << ", h:" << nhead << "/" << nhead_k << ", s:" << seqlen_q << "/" << seqlen_k
+              << ", h:" << nhead << "/" << nhead_k << ", s:" << seqlen_qs[0] << "/" << seqlen_ks[0]
+              << (seqlen_kpads[0] < 0 ? ""
+                                      : (std::string("(") + std::to_string(seqlen_kpads[0]) + ")"))
               << ", d:" << hdim_q << "/" << hdim_v << ", scale_s:" << scale_s << ", bias:" << bias
               << ", lse:" << lse << ", squant:" << squant << ", mask:" << mask << ", v:" << vlayout
               << std::flush;
@@ -460,7 +491,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
             return ck_tile::identity{};
     }();
 
-    auto fmha_args = [&]() {
+    auto fmha_args = [&, k_paddings_ = seqlen_kpads]() {
         assert(nhead % nhead_k == 0);
         /// NOTE: we broadcast bias from [1, 1, seqlen_q, seqlen_k] to [batch, nhead, seqlen_q,
         ///       seqlen_k] in this example, hence both the 'batch_stride_bias' &
@@ -506,7 +537,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
                              o_buf.GetDeviceBuffer(),
                              seqstart_q.GetDeviceBuffer(),
                              seqstart_k.GetDeviceBuffer(),
-                             nullptr,
+                             k_paddings_[0] < 0 ? nullptr : seqlen_k_buf.GetDeviceBuffer(),
                              shape_seqlen_q,
                              shape_seqlen_k,
                              batch,
@@ -576,7 +607,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
         // adjust matrix index according to the mode
         const ck_tile::index_t b            = (mode == mode_enum::batch ? wb : 0);
         const ck_tile::index_t query_offset = (mode == mode_enum::batch ? 0 : seqstart_q_host[wb]);
-        const ck_tile::index_t key_offset   = (mode == mode_enum::batch ? 0 : seqstart_k_host[wb]);
+        const ck_tile::index_t key_offset =
+            (mode == mode_enum::batch
+                 ? 0
+                 : (seqlen_kpads[0] < 0 ? seqstart_k_host[wb] : seqstart_k_with_padding_host[wb]));
 
         const auto v_host_ref_lengths =
             std::array<ck_tile::index_t, 3>{nhead, hdim_v, real_seqlen_k};
@@ -661,7 +695,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
                 else
                 {
                     return ck_tile::Alibi<SaccDataType, true>{
-                        0, real_seqlen_q, real_seqlen_k, ck_tile::AlibiMode::VERTICAL};
+                        0, real_seqlen_q, real_seqlen_k, ck_tile::AlibiMode::FROM_BOTTOM_RIGHT};
                 }
             }();
 
@@ -671,7 +705,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
             for(auto i_h = 0; i_h < nhead; i_h++)
             {
                 SaccDataType current_slope = alibi_slope_host(i_b_slope, i_h);
-                alibi_host.slope           = current_slope;
+                alibi_host.slope = alibi_host.mode == ck_tile::AlibiMode::VERTICAL ? current_slope
+                                                                                   : -current_slope;
                 for(auto i_r = 0; i_r < real_seqlen_q; i_r++)
                 {
                     for(auto i_c = 0; i_c < real_seqlen_k; i_c++)

example/ck_tile/01_fmha/generate.py

@@ -78,6 +78,11 @@ BOOL_MAP = {
     "f" : "false"
 }
 
+TILE_PARTITIONER_MAP = {
+    "shb" : "ck_tile::FmhaFwdTilePartitioner_SHB",
+    "hbs" : "ck_tile::FmhaFwdTilePartitioner_HBS",
+}
+
 DIRECTIONS = ["fwd"]
 GEN_DIR = ""    # in Cmake, have to generate files in same folder
 
@@ -107,7 +112,7 @@ using fmha_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad},
                                                     {F_dvpad},
                                                     {F_bias},
                                                     {F_lse},
-                                                    {F_squant},
+                                                    {F_squant}, 
                                                     {F_occupancy}>;
 using fmha_mask_{F_idx} = {F_mask};
 
@@ -136,7 +141,7 @@ using fmha_epilogue_{F_idx} =
                                            {F_spad}, {F_dvpad}>>;
 
 using fmha_kernel_{F_idx} =
-    ck_tile::FmhaFwdKernel<ck_tile::FmhaFwdTilePartitioner<fmha_shape_{F_idx}>,
+    ck_tile::FmhaFwdKernel<{F_tile_partitioner}<fmha_shape_{F_idx}>,
                   fmha_pipeline_{F_idx},
                   fmha_epilogue_{F_idx}>;
 
@@ -154,7 +159,7 @@ float fmha_fwd_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_args a)
     auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
     constexpr dim3 blocks             = k_::BlockSize();
     constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
-    return ck_tile::launch_kernel<blocks.x, kBlockPerCu>(s, k_{{}}, grids, blocks, 0, kargs);
+    return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
 }}
 """
 
@@ -389,6 +394,12 @@ class FmhaFwdKernel:
     F_pipeline      : FmhaFwdPipeline
     mask_impl       : str
 
+    def get_tp(self) -> str:
+        if self.F_mode == 'group':
+            return 'hbs'
+        else:
+            return 'shb'
+
     @property
     def template(self) -> str:
         kernel_body = str()
@@ -413,7 +424,7 @@ class FmhaFwdKernel:
                 F_spad          = BOOL_MAP[self.F_pipeline.F_spad],
                 F_skpad         = BOOL_MAP[self.F_pipeline.F_skpad],
                 F_dpad          = BOOL_MAP[self.F_pipeline.F_dpad],
-                F_dvpad         = BOOL_MAP[self.F_pipeline.F_dvpad],
+                F_dvpad         = BOOL_MAP[self.F_pipeline.F_dvpad],    
                 F_bias          = BIAS_MAP[self.F_pipeline.F_bias],
                 F_lse           = BOOL_MAP[self.F_pipeline.F_lse],
                 F_squant        = BOOL_MAP[self.F_pipeline.F_squant],
@@ -421,12 +432,13 @@ class FmhaFwdKernel:
                 F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag],
                 F_mask          = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask],
                 F_mode          = MODE_MAP[self.F_mode],
-                F_pipeline      = PIPELINE_MAP[self.F_pipeline.tag])
+                F_pipeline      = PIPELINE_MAP[self.F_pipeline.tag],
+                F_tile_partitioner = TILE_PARTITIONER_MAP[self.get_tp()])
 
     @property
     def name(self) -> str:
         # TODO: we don't encode idx here
-        return f"fmha_{self.direction}_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" +\
+        return f"fmha_{self.direction}_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_{self.get_tp()}_" + \
                 self.F_tile.name + '_' + self.F_pipeline.name
 
     @property

example/ck_tile/01_fmha/script/smoke_test.sh

@@ -28,6 +28,7 @@ $EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -d_v=24 -s=3 -s_k=99 -bias=$bias
 $EXE -prec=$prec -mode=$mode -b=3 -h=2 -h_k=1 -d=$hdim -s=200 -s_k=520 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=t:128,30 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
 $EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -s=99 -s_k=32 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=b:4,35 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
 $EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=33 -s_k=0 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
+$EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1 -s_k=10 -s_kpad=32 -bias=$bias -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -kname=$KNAME $COMMON_ARGS
 
 done
 done

example/ck_tile/01_fmha/utils.hpp

@@ -4,12 +4,14 @@
 #pragma once
 
 #include <cstdint>
+#include <cstdlib>
 #include <optional>
 #include <ostream>
 #include <tuple>
 #include <utility>
 #include <vector>
 #include <functional>
+#include <string>
 
 #include "ck_tile/core/container/span.hpp"
 
@@ -37,12 +39,14 @@ std::vector<int32_t> to_seqstarts(ck_tile::span<const int32_t> seqlens)
 
 std::vector<int32_t> generate_seqlens(mode_enum mode,
                                       unsigned count,
-                                      int32_t seqlens_sum,
+                                      int32_t seqlen_avg,
+                                      int32_t seqlen_max = -1, // if not negative, clamp max
                                       std::optional<unsigned> seed = std::nullopt)
 {
     assert(0 < count);
 
-    std::vector<int32_t> seqlens(count, seqlens_sum);
+    std::vector<int32_t> seqlens(
+        count, seqlen_max > 0 ? (seqlen_avg < seqlen_max ? seqlen_avg : seqlen_max) : seqlen_avg);
 
     if(mode == mode_enum::group && 1 < count)
     {
@@ -55,7 +59,7 @@ std::vector<int32_t> generate_seqlens(mode_enum mode,
         std::uniform_int_distribution<size_type> step_dist(1, count - 1);
         auto next_step = std::bind(step_dist, std::ref(random_engine));
 
-        for(unsigned repeat = seqlens_sum * (count / 2); 0 < repeat; --repeat)
+        for(unsigned repeat = seqlen_avg * (count / 2); 0 < repeat; --repeat)
         {
             const size_type to_decrease = next_idx();
             // make sure each elements of seqlens is always greater than 0
@@ -66,6 +70,11 @@ std::vector<int32_t> generate_seqlens(mode_enum mode,
 
             const size_type to_increase = (to_decrease + next_step()) % count;
 
+            if(seqlen_max > 0 && seqlens[to_increase] >= seqlen_max)
+            {
+                continue;
+            }
+
             --seqlens[to_decrease];
             ++seqlens[to_increase];
         }
@@ -76,10 +85,91 @@ std::vector<int32_t> generate_seqlens(mode_enum mode,
 
 std::vector<int32_t> generate_seqstarts(mode_enum mode,
                                         unsigned count,
-                                        int32_t seqlens_sum,
+                                        int32_t seqlen_avg,
+                                        int32_t seqlen_max           = -1,
                                         std::optional<unsigned> seed = std::nullopt)
 {
-    return to_seqstarts(generate_seqlens(mode, count, seqlens_sum, seed));
+    return to_seqstarts(generate_seqlens(mode, count, seqlen_avg, seqlen_max, seed));
+}
+
+/*
+ * decode the seqlen string from cmdline
+ * example (assume batch=3)
+ *   q_val=1,2,3 k_val=4,5,6 -> OK
+ *   q_val=1,2,3             -> OK, k same as q
+ *   q_val=1,2               -> OK, q will rand remaining 1 element, k same as q
+ *   q_val=1,2   k_val=4,5   -> OK, q/k will rand remaining 1 element
+ *   q_val=1,2,3,4           -> OK, but ignore exceed one
+ *
+ *   q_val=1,2   k_val=4,5,6 -> not OK, k must have same splits with q
+ *   q_val=1,2   k_val=4     -> not OK, k must have same splits with q
+ */
+std::tuple<std::vector<ck_tile::index_t>,
+           std::vector<ck_tile::index_t>,
+           std::vector<ck_tile::index_t>>
+decode_seqlen(mode_enum mode,
+              ck_tile::index_t batch,
+              std::string q_val,
+              std::string k_val,
+              std::string k_pad_val,
+              std::optional<unsigned> seed = std::nullopt)
+{
+#define _S2I_(str_) static_cast<ck_tile::index_t>(std::atoi((str_).c_str()))
+    if(mode == mode_enum::batch)
+    {
+        ck_tile::index_t q = _S2I_(q_val);
+        ck_tile::index_t k = _S2I_(k_val);
+        auto s_q           = std::vector<ck_tile::index_t>(batch, q);
+        auto s_k           = std::vector<ck_tile::index_t>(batch, k < 0 ? q : k);
+        auto s_kpad = std::vector<ck_tile::index_t>(batch, -1); // TODO: batch not support k_padding
+        return std::make_tuple(s_q, s_k, s_kpad);
+    }
+    else
+    {
+        ck_tile::index_t idx          = 0;
+        std::string::size_type pos_q  = 0;
+        std::string::size_type pos_k  = 0;
+        std::string::size_type pos_kp = 0;
+        std::vector<ck_tile::index_t> s_q;
+        std::vector<ck_tile::index_t> s_k;
+        std::vector<ck_tile::index_t> s_kpad;
+        while(true)
+        {
+            auto found_q  = q_val.find(',', pos_q);
+            auto found_k  = k_val.find(',', pos_k);
+            auto found_kp = k_pad_val.find(',', pos_kp);
+
+            ck_tile::index_t q = _S2I_(
+                q_val.substr(pos_q, found_q == std::string::npos ? found_q : found_q - pos_q));
+            ck_tile::index_t k = _S2I_(
+                k_val.substr(pos_k, found_k == std::string::npos ? found_k : found_k - pos_k));
+            ck_tile::index_t kp = _S2I_(k_pad_val.substr(
+                pos_kp, found_kp == std::string::npos ? found_kp : found_kp - pos_kp));
+
+            s_q.push_back(q);
+            s_k.push_back(k < 0 ? q : k);
+            s_kpad.push_back(kp);
+            idx++;
+            if(found_q == std::string::npos || idx >= batch)
+            {
+                break;
+            }
+            pos_q  = found_q + 1;
+            pos_k  = found_k == std::string::npos ? pos_k : found_k + 1;
+            pos_kp = found_kp == std::string::npos ? pos_kp : found_kp + 1;
+        }
+        if(idx < batch)
+        {
+            auto rem_q = generate_seqlens(mode, batch - idx, s_q.back(), s_kpad.back(), seed);
+            auto rem_k = generate_seqlens(mode, batch - idx, s_k.back(), s_kpad.back(), seed);
+
+            s_q.insert(s_q.end(), rem_q.begin(), rem_q.end());
+            s_k.insert(s_k.end(), rem_k.begin(), rem_k.end());
+            s_kpad.insert(s_kpad.end(), batch - idx, s_kpad.back());
+        }
+        return std::make_tuple(s_q, s_k, s_kpad);
+    }
+#undef _S2I_
 }
 
 int env_get_int(const char* var_name, int default_int)
@@ -87,6 +177,6 @@ int env_get_int(const char* var_name, int default_int)
     char* v = getenv(var_name);
     int r   = default_int;
     if(v)
-        r = atoi(v);
+        r = std::atoi(v);
     return r;
 }

include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r3.hpp"
+#include "ck/utility/is_detected.hpp"
+
+namespace ck {
+
+// Thread-group level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//
+// Does following things to avoid scratch memory issue
+//   1. Pass tensor descritpors by reference (or tuple of references)
+//   2. Does not keep reference to tensor descriptor
+//   3. Does not construct new tensor coordinate when call Run()
+template <typename ThreadGroup,
+          typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          typename SrcScalarPerVectors,
+          index_t DstScalarPerVector,
+          typename ThreadTransferSrcResetCoordinateAfterRunFlags,
+          typename ThreadTransferDstResetCoordinateAfterRunFlags,
+          index_t NumThreadScratch = 1>
+struct ThreadGroupTensorSliceTransfer_v7r3
+{
+    static constexpr index_t nDim =
+        remove_cvref_t<tuple_element_t<0, SrcDescs>>::GetNumOfDimension();
+
+    static constexpr index_t nSrc = remove_cvref_t<SrcDescs>::Size();
+    static constexpr index_t nDst = remove_cvref_t<DstDescs>::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v7r3(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_block_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_block_slice_origins,
+        const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src_descs,
+                               StaticallyIndexedArray<Index, nSrc>{},
+                               dst_descs,
+                               StaticallyIndexedArray<Index, nDst>{},
+                               element_op)
+    {
+        static_assert(nSrc == SrcDatas::Size() && nSrc == SrcDescs::Size() &&
+                          nSrc == ThreadTransferSrcResetCoordinateAfterRunFlags::Size() &&
+                          nDst == DstDatas::Size() && nDst == DstDescs::Size() &&
+                          nDst == ThreadTransferDstResetCoordinateAfterRunFlags::Size(),
+                      "wrong!");
+
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            static_assert(
+                nDim == remove_cvref_t<tuple_element_t<i.value, SrcDescs>>::GetNumOfDimension(),
+                "wrong!");
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            static_assert(
+                nDim == remove_cvref_t<tuple_element_t<i.value, DstDescs>>::GetNumOfDimension(),
+                "wrong!");
+        });
+
+        static_assert(nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            const auto src_thread_slice_origins = generate_tuple(
+                [&](auto i) { return src_block_slice_origins[i] + thread_data_idx_begin; },
+                Number<nSrc>{});
+
+            const auto dst_thread_slice_origins = generate_tuple(
+                [&](auto i) { return dst_block_slice_origins[i] + thread_data_idx_begin; },
+                Number<nDst>{});
+
+            threadwise_transfer_.SetSrcSliceOrigins(src_descs, src_thread_slice_origins);
+            threadwise_transfer_.SetDstSliceOrigins(dst_descs, dst_thread_slice_origins);
+        }
+    }
+
+    template <typename SrcBuffers, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDescs& src_descs,
+                            const SrcBuffers& src_bufs,
+                            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_descs, src_bufs, thread_scratch_id);
+        }
+    }
+
+    template <typename T>
+    using is_tuple = decltype(std::declval<T&>().IsTuple());
+
+    template <typename DstBuffers, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDescs& dst_descs,
+                             DstBuffers dst_bufs,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            if constexpr(is_detected<is_tuple, decltype(dst_bufs)>::value)
+                threadwise_transfer_.RunWrite(dst_descs, dst_bufs, thread_scratch_id);
+            else
+                threadwise_transfer_.RunWrite(dst_descs, tie(dst_bufs), thread_scratch_id);
+        }
+    }
+
+    template <typename SrcBuffers, typename DstBuffers>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        RunRead(src_descs, src_bufs);
+        RunWrite(dst_descs, dst_bufs);
+    }
+
+    template <index_t ISrc>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDescs& src_descs, Number<ISrc> iSrc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_descs, iSrc, step);
+        }
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs, const Index& step)
+    {
+        static_for<0, SrcDescs::Size(), 1>{}(
+            [&](auto i) { MoveSrcSliceWindow(src_descs, i, step); });
+    }
+
+    template <index_t IDst>
+    __device__ void
+    MoveDstSliceWindow(const DstDescs& dst_descs, Number<IDst> iDst, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_descs, iDst, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs, const Index& step)
+    {
+        static_for<0, DstDescs::Size(), 1>{}(
+            [&](auto i) { MoveDstSliceWindow(dst_descs, i, step); });
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v7r3<SrcDatas,
+                                           DstDatas,
+                                           SrcDescs,
+                                           DstDescs,
+                                           ElementwiseOperation,
+                                           DstInMemOps,
+                                           decltype(thread_slice_lengths),
+                                           SrcDimAccessOrder,
+                                           DstDimAccessOrder,
+                                           SrcVectorDim,
+                                           DstVectorDim,
+                                           SrcScalarPerVectors,
+                                           DstScalarPerVector,
+                                           ThreadTransferSrcResetCoordinateAfterRunFlags,
+                                           ThreadTransferDstResetCoordinateAfterRunFlags,
+                                           NumThreadScratch>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename CDEShuffleBlockTransferScalarPerVectors,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA,
+          typename LDSTypeA                           = ComputeTypeA,
+          typename LDSTypeB                           = ComputeTypeB>
+struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD<ALayout,
+                                                                     BLayout,
+                                                                     DsLayout,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     DsDataType,
+                                                                     CDataType,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     CElementwiseOperation>
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultiD_xdl_cshuffle_v3<
+        ALayout,
+        BLayout,
+        DsLayout,
+        CLayout,
+        ADataType,
+        BDataType,
+        GemmAccDataType,
+        CShuffleDataType,
+        DsDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEShuffleBlockTransferScalarPerVectors,
+        BlkGemmPipeSched,
+        BlkGemmPipelineVer,
+        ComputeTypeA,
+        ComputeTypeB,
+        LDSTypeA,
+        LDSTypeB>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    // 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 ? 1 : 2;
+
+            if(has_main_k_block_loop)
+            {
+                // Tail number always full
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 ||
+                             BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+#if 0
+                    if(arg.KBatch > 1)
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                        true,
+                                                        InMemoryDataOperationEnum::AtomicAdd,
+                                                        minimum_occupancy>;
+                        Run(kernel);
+                    }
+                    else
+#endif
+                    {
+                        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 0
+                    if(arg.KBatch > 1)
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
+                        {
+                            const auto kernel =
+                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                            true,
+                                                            InMemoryDataOperationEnum::AtomicAdd,
+                                                            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::AtomicAdd,
+                                                            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::AtomicAdd,
+                                    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::AtomicAdd,
+                                    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::AtomicAdd,
+                                    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::AtomicAdd,
+                                    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::AtomicAdd,
+                                    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::AtomicAdd,
+                                    minimum_occupancy,
+                                    TailNumber::Seven>;
+                                Run(kernel);
+                            }
+                        }
+                    }
+                    else
+#endif
+                    {
+                        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);
+                            }
+                        }
+                    }
+                }
+                // Tail number could be Odd or Even
+                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
+                {
+#if 0
+                    if(arg.KBatch > 1)
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::AtomicAdd,
+                                minimum_occupancy,
+                                TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::AtomicAdd,
+                                minimum_occupancy,
+                                TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                    else
+#endif
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel =
+                                kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
+                                                                 true,
+                                                                 InMemoryDataOperationEnum::Set,
+                                                                 minimum_occupancy,
+                                                                 TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel =
+                                kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
+                                                                 true,
+                                                                 InMemoryDataOperationEnum::Set,
+                                                                 minimum_occupancy,
+                                                                 TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                }
+                else
+                {
+#if 0
+                    if(arg.KBatch > 1)
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel =
+                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                            true,
+                                                            InMemoryDataOperationEnum::AtomicAdd,
+                                                            minimum_occupancy,
+                                                            TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel =
+                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                            true,
+                                                            InMemoryDataOperationEnum::AtomicAdd,
+                                                            minimum_occupancy,
+                                                            TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                    else
+#endif
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel =
+                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                            true,
+                                                            InMemoryDataOperationEnum::Set,
+                                                            minimum_occupancy,
+                                                            TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel =
+                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                            true,
+                                                            InMemoryDataOperationEnum::Set,
+                                                            minimum_occupancy,
+                                                            TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                }
+            }
+            else
+            {
+                // Tail number always 1
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+#if 0
+                    if(arg.KBatch > 1)
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                        false,
+                                                        InMemoryDataOperationEnum::AtomicAdd,
+                                                        minimum_occupancy>;
+                        Run(kernel);
+                    }
+                    else
+#endif
+                    {
+                        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((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,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<index_t, NumDTensor> StrideDs,
+                             index_t StrideC,
+                             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,
+                        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,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      std::array<ck::index_t, NumDTensor> StrideDs,
+                                                      index_t StrideC,
+                                                      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,
+                                          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"},
+            {BlockGemmPipelineVersion::v4, "v4"},
+            {BlockGemmPipelineVersion::v5, "v5"}};
+
+        // clang-format off
+        str << "DeviceGemmXdlUniversal"
+            << "<"
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << ">"
+            << " BlkSize: "
+            << BlockSize << ", "
+            << "BlkTile: "
+            << MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
+            << "WaveTile: "
+            << MPerXDL<<"x"<<NPerXDL << ", "
+            << "WaveMap: "
+            << MXdlPerWave<<"x" << NXdlPerWave<<", "
+            << "VmemReadVec: "
+            << ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
+            << "BlkGemmPipelineScheduler: "
+            << BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
+            << "BlkGemmPipelineVersion: "
+            << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
+            << "BlkGemmPipelinePrefetchStages: "
+            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_selector.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r3.hpp"
+
+#define DEBUG_LOG 0
+
+namespace ck {
+
+// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same
+// kernel function Blockers:
+// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on
+// two lds chunks.
+// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds
+// buffer when we declare __shared__ inside blkgemmpipe
+template <typename GridwiseGemm,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t MinimumOccupancy = 1,
+          TailNumber TailNum       = TailNumber::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()];
+
+    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
+        karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
+        karg.p_ds_grid,
+        karg.p_c_grid,
+        p_shared,
+        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 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_2lds(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    // Pass two lds pointer is the key to tell compiler that ds_read/write
+    // operate on different lds chunk at same time without order dependecy
+    __shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    __shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
+
+    GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
+        karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
+        karg.p_ds_grid,
+        karg.p_c_grid,
+        p_shared_0,
+        p_shared_1,
+        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 MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename CDEShuffleBlockTransferScalarPerVectors,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v4,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA,
+          typename LDSTypeA                           = ADataType,
+          typename LDSTypeB                           = BDataType>
+struct GridwiseGemmMultiD_xdl_cshuffle_v3
+{
+    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_,
+                          index_t k_batch_,
+                          AElementwiseOperation a_element_op_,
+                          BElementwiseOperation b_element_op_,
+                          CElementwiseOperation c_element_op_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideDs_, StrideC_, k_batch_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_ds_grid{},
+              p_c_grid{p_c_grid_},
+              a_element_op{a_element_op_},
+              b_element_op{b_element_op_},
+              c_element_op{c_element_op_}
+        {
+
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType_ = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid(i) = static_cast<const DDataType_*>(p_ds_grid_[i]);
+            });
+        }
+
+        const ADataType* p_a_grid;
+        const BDataType* p_b_grid;
+        DsGridPointer p_ds_grid;
+        CDataType* p_c_grid;
+
+        const AElementwiseOperation a_element_op;
+        const BElementwiseOperation b_element_op;
+        const CElementwiseOperation c_element_op;
+    };
+
+    struct SplitKBatchOffset
+    {
+        __device__ SplitKBatchOffset(Argument& karg)
+        {
+            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(BlockGemmPipeline_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,
+                               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 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());
+
+        // 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<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<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);
+
+        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,
+                                                                         num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            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);
+                }
+            });
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run_2Lds(const ADataType* p_a_grid,
+                                    const BDataType* p_b_grid,
+                                    DsGridPointer& p_ds_grid,
+                                    CDataType* p_c_grid,
+                                    void* p_shared_0,
+                                    void* p_shared_1,
+                                    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 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());
+
+        // 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<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<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);
+
+        auto a_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeA*>(p_shared_0), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeB*>(p_shared_0) +
+                a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB),
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeA*>(p_shared_1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeB*>(p_shared_1) +
+                a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB),
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong);
+        auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong);
+
+        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);
+
+        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_bufs,
+                                                                         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_bufs,
+                                                                         b_block_slice_copy_step,
+                                                                         c_thread_buf,
+                                                                         num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            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_0),
+                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_v7r3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+#include "ck/utility/is_detected.hpp"
+#include "ck/tensor/static_tensor.hpp"
+
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp"
+
+namespace ck {
+// Thread-level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//   6. Does not need to know src_descs and dst_descs at compile-time
+//   7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
+//
+// Does following things to avoid scratch memory issue
+//   1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
+//   2. Pass tensor descritpors by reference (or tuple of references)
+//   3. Does not keep reference to tensor descriptor
+//   4. Does not construct new tensor coordinate when call Run()
+template <typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          typename SrcScalarPerVectors,
+          index_t DstScalarPerVector,
+          typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v7r3
+{
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr auto SrcScalarPerVector = SrcScalarPerVectors{}[I0];
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    // scalar per access on each dim
+    // FIXME: don't use lambda_scalar_per_access
+    static constexpr auto src_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+    static constexpr auto dst_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+    using SrcSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   SrcDimAccessOrder,
+                                                   remove_cv_t<decltype(src_scalar_per_access)>,
+                                                   false>;
+
+    using DstSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   DstDimAccessOrder,
+                                                   remove_cv_t<decltype(dst_scalar_per_access)>,
+                                                   false>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v7r3(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto i) {
+            dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename DataTypes, index_t ScalarPerVector>
+    __device__ static auto generate_vectors()
+    {
+        auto data_types = DataTypes{};
+
+        constexpr index_t num = data_types.Size();
+
+        return generate_tuple(
+            [&](auto i) {
+                using DataType = remove_cvref_t<decltype(data_types[i])>;
+
+                return vector_type_maker_t<DataType, ScalarPerVector>{};
+            },
+            Number<num>{});
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    template <typename SrcBuffers,
+              index_t ThreadScratchId                                   = 0,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size(), bool> = false>
+    __device__ void RunRead(const SrcDescs& src_descs,
+                            const SrcBuffers& src_bufs,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // loop over space-filling curve
+        static_for<0, src_num_access, 1>{}([&](auto iAccess) {
+            auto src_vectors = generate_vectors<SrcDatas, SrcScalarPerVector>();
+            auto elm_vectors = generate_vectors<DstDatas, SrcScalarPerVector>();
+
+            bool oob_val = true;
+
+            // copy data from src_bufs into src_vectors
+            static_for<0, nSrc, 1>{}([&](auto i) {
+                using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
+
+                const bool is_src_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
+                                                                                src_coords_[i]);
+
+                oob_val = oob_val & is_src_valid;
+
+                if constexpr(SrcScalarPerVectors{}[i] == 1)
+                {
+                    auto data_types = SrcDatas{};
+                    using DataType  = remove_cvref_t<decltype(data_types[i])>;
+                    const auto tmp =
+                        src_bufs[i].template Get<DataType>(src_coords_[i].GetOffset(), true);
+
+                    static_for<0, SrcScalarPerVector, 1>{}(
+                        [&](auto j) { src_vectors(i).template AsType<DataType>()(j) = tmp; });
+                }
+                else
+                {
+                    src_vectors(i).template AsType<src_vector_t>()(I0) =
+                        src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(), true);
+                }
+            });
+
+            constexpr auto get_elem_op_vec_len = []() {
+                if constexpr(is_detected<is_pack8_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack8_invocable)
+                        return math::min(8, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack4_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack4_invocable)
+                        return math::min(4, SrcScalarPerVector);
+                }
+                if constexpr(is_detected<is_pack2_invocable_t, decltype(element_op_)>::value)
+                {
+                    if constexpr(decltype(element_op_)::is_pack2_invocable)
+                        return math::min(2, SrcScalarPerVector);
+                }
+                return 1;
+            };
+
+            constexpr index_t elem_op_vec_len = get_elem_op_vec_len();
+
+            // apply pointwise function
+            static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) {
+                // get reference to src data
+                const auto src_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iSrc) -> const auto& {
+                        using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
+
+                        return src_vectors[iSrc].template AsType<elem_op_vec_t>()[i];
+                    },
+                    Number<nSrc>{});
+
+                // get reference to dst data
+                auto dst_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iDst) -> auto& {
+                        using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                        using elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
+
+                        return elm_vectors(iDst).template AsType<elem_op_vec_t>()(i);
+                    },
+                    Number<nDst>{});
+
+                // apply pointwise function
+                // pointwise function signature:
+                // element_op_(dst_data_refs[I0],
+                //             dst_data_refs[I1],
+                //             ...,
+                //             src_data_refs[I0],
+                //             src_data_refs[I1],
+                //             ...)
+                unpack2(element_op_, dst_data_refs, src_data_refs);
+            });
+
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+            oob_vectors_tuple_(thread_scratch_id)(iAccess) = oob_val;
+
+            // move coordinate
+            if constexpr(iAccess.value != src_num_access - 1)
+            {
+                constexpr auto forward_step = SrcSpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nSrc, 1>{}([&](auto i) {
+                    move_tensor_coordinate(src_descs[i],
+                                           src_coords_(i),
+                                           make_tensor_coordinate_step(src_descs[i], forward_step));
+                });
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto src_reset_step =
+                    make_tensor_coordinate_step(src_descs[i], GetSrcCoordinateResetStep());
+
+                move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
+            }
+        });
+    }
+
+#if 1
+    template <index_t ThreadScratchId = 0>
+    __device__ void OOBCheck(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // loop over space-filling curve
+        static_for<0, src_num_access, 1>{}([&](auto iAccess) {
+            auto elm_vectors = elm_vectors_tuple_[thread_scratch_id][iAccess];
+            auto oob_val     = oob_vectors_tuple_[thread_scratch_id][iAccess];
+
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using elm_vector_t = typename remove_cvref_t<decltype(elm_vectors[i])>::type;
+                elm_vectors(i).template AsType<elm_vector_t>()(I0) =
+                    oob_val ? elm_vectors(i).template AsType<elm_vector_t>()[I0] : elm_vector_t{0};
+            });
+
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+        });
+    }
+#endif
+
+    template <index_t ThreadScratchId = 0>
+    __device__ void
+    TransposeFromElmToDst(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        using DstData = remove_cvref_t<decltype(DstDatas{}[I0])>;
+
+        using ElmThreadScratch =
+            StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                            DstData,
+                                            SrcScalarPerVector,
+                                            decltype(GetSrcThreadScratchDescriptor()),
+                                            true>;
+        using DstThreadScratch =
+            StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                            DstData,
+                                            DstScalarPerVector,
+                                            decltype(GetDstThreadScratchDescriptor()),
+                                            true>;
+
+        ElmThreadScratch elm_thread_scratch_;
+        DstThreadScratch dst_thread_scratch_;
+
+        elm_thread_scratch_.data_ =
+            bit_cast<decltype(elm_thread_scratch_.data_)>(elm_vectors_tuple_[thread_scratch_id]);
+
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     ((is_same<half_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<f8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0) ||
+                      (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
+
+            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<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<DstData, 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 elm_thread_scratch_.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<DstData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+        else
+        {
+            static_ford<SliceLengths>{}(
+                [&](auto idx) { dst_thread_scratch_(idx) = elm_thread_scratch_[idx]; });
+        }
+
+        dst_vectors_tuple_(thread_scratch_id) = bit_cast<DstVectorTuple>(dst_thread_scratch_.data_);
+    }
+
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename DstBuffers,
+              index_t ThreadScratchId                                             = 0,
+              enable_if_t<DstDescs::Size() == 1 && DstBuffers::Size() == 1, bool> = false>
+    __device__ void RunWrite(const DstDescs& dst_descs,
+                             DstBuffers dst_bufs,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        OOBCheck(thread_scratch_id);
+        TransposeFromElmToDst(thread_scratch_id);
+
+        // loop over space-filling curve
+        static_for<0, dst_num_access, 1>{}([&](auto iAccess) {
+            auto dst_vectors = dst_vectors_tuple_[thread_scratch_id][iAccess];
+
+            // copy data from buf_vectors into dst_bufs
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
+
+                const bool is_dst_valid =
+                    coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
+                                                                                dst_coords_[i]);
+
+                constexpr InMemoryDataOperationEnum DstInMemOp =
+                    static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
+
+                dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
+                    dst_coords_[i].GetOffset(),
+                    is_dst_valid,
+                    dst_vectors[i].template AsType<dst_vector_t>()[I0]);
+            });
+
+            // move coordinate
+            if constexpr(iAccess.value != dst_num_access - 1)
+            {
+                constexpr auto forward_step = DstSpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nDst, 1>{}([&](auto i) {
+                    move_tensor_coordinate(dst_descs[i],
+                                           dst_coords_(i),
+                                           make_tensor_coordinate_step(dst_descs[i], forward_step));
+                });
+            }
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            if constexpr(DstResetCoordinateAfterRunFlags::At(i))
+            {
+                const auto dst_reset_step =
+                    make_tensor_coordinate_step(dst_descs[i], GetDstCoordinateResetStep());
+
+                move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
+            }
+        });
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename SrcBuffers,
+              typename DstBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
+                              DstDescs::Size() == DstBuffers::Size(),
+                          bool> = false>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        RunRead(src_descs, src_bufs);
+        RunWrite(dst_descs, dst_bufs);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        if constexpr(src_num_access == 0)
+        {
+            return typename SrcSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            return SrcSpaceFillingCurve::GetStepBetween(Number<src_num_access - 1>{}, Number<0>{});
+        }
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        if constexpr(dst_num_access == 0)
+        {
+            return typename DstSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            return DstSpaceFillingCurve::GetStepBetween(Number<dst_num_access - 1>{}, Number<0>{});
+        }
+    }
+
+    __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 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);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t ISrc>
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       Number<ISrc> iSrc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRunFlags::At(iSrc)
+                ? src_slice_origin_step_idx
+                : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
+
+        move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t IDst>
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       Number<IDst> iDst,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRunFlags::At(iDst)
+                ? dst_slice_origin_step_idx
+                : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx);
+
+        move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
+    }
+
+    private:
+    using SrcVectorsType = decltype(generate_vectors<SrcDatas, SrcScalarPerVector>());
+    using ElmVectorsType = decltype(generate_vectors<DstDatas, SrcScalarPerVector>());
+    using DstVectorsType = decltype(generate_vectors<DstDatas, DstScalarPerVector>());
+
+    static constexpr auto src_num_access = SrcSpaceFillingCurve::GetNumOfAccess();
+    static constexpr auto dst_num_access = DstSpaceFillingCurve::GetNumOfAccess();
+
+    using ElmVectorTuple = StaticallyIndexedArray<ElmVectorsType, src_num_access>;
+    using DstVectorTuple = StaticallyIndexedArray<DstVectorsType, dst_num_access>;
+
+    StaticallyIndexedArray<ElmVectorTuple, NumThreadScratch> elm_vectors_tuple_;
+    StaticallyIndexedArray<DstVectorTuple, NumThreadScratch> dst_vectors_tuple_;
+
+    using OOBVectorTuple = StaticallyIndexedArray<bool, src_num_access>;
+    StaticallyIndexedArray<OOBVectorTuple, NumThreadScratch> oob_vectors_tuple_;
+
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck_tile/core/arch/amd_buffer_addressing.hpp

@@ -29,6 +29,25 @@ CK_TILE_DEVICE int32x4_t make_wave_buffer_resource(const void* ptr, uint32_t siz
     return __builtin_bit_cast(int32x4_t, res);
 }
 
+namespace impl {
+// below type indicate the data type used for buffer load inline asm
+// clang-format off
+template<index_t N, typename T> struct buffer_load_trait;
+
+template<typename T> struct buffer_load_trait<16, T> { using payload_t = fp32x4_t; };
+template<typename T> struct buffer_load_trait<8 , T> { using payload_t = fp32x2_t; };
+template<typename T> struct buffer_load_trait<4 , T> { using payload_t = float; };
+template<typename T> struct buffer_load_trait<2 , T> { using payload_t = float; };
+template<typename T> struct buffer_load_trait<1 , T> { using payload_t = float; };
+
+#if CK_TILE_BUFFER_LOAD_RAW_BF16_WA
+template<> struct buffer_load_trait<16, thread_buffer<bf16_t, 8>> { using payload_t = bf16x8_t; };
+template<> struct buffer_load_trait<8 , thread_buffer<bf16_t, 4>> { using payload_t = bf16x4_t; };
+template<> struct buffer_load_trait<4 , thread_buffer<bf16_t, 2>> { using payload_t = bf16x2_t; };
+#endif
+// clang-format on
+} // namespace impl
+
 // TODO: glc/slc/...
 template <index_t bytes>
 struct buffer_load;
@@ -48,7 +67,7 @@ struct buffer_load<16>
                                    index_t /*flag*/ = 0)
     {
         static_assert(sizeof(T) == 16);
-        using mbuf_t = fp32x4_t;
+        using mbuf_t = typename impl::buffer_load_trait<16, T>::payload_t;
         asm volatile("buffer_load_dwordx4 %0, %1, %2, %3 offen offset:%4"
                      : "+v"(reinterpret_cast<mbuf_t&>(value))
                      : "v"(v_offset), "s"(res), "s"(s_offset), "n"(i_offset)
@@ -68,7 +87,7 @@ struct buffer_load<8>
                                    index_t /*flag*/ = 0)
     {
         static_assert(sizeof(T) == 8);
-        using mbuf_t = fp32x2_t;
+        using mbuf_t = typename impl::buffer_load_trait<8, T>::payload_t;
         asm volatile("buffer_load_dwordx2 %0, %1, %2, %3 offen offset:%4"
                      : "+v"(reinterpret_cast<mbuf_t&>(value))
                      : "v"(v_offset), "s"(res), "s"(s_offset), "n"(i_offset)
@@ -88,7 +107,7 @@ struct buffer_load<4>
                                    index_t /*flag*/ = 0)
     {
         static_assert(sizeof(T) == 4);
-        using mbuf_t = float;
+        using mbuf_t = typename impl::buffer_load_trait<4, T>::payload_t;
         asm volatile("buffer_load_dword %0, %1, %2, %3 offen offset:%4"
                      : "+v"(reinterpret_cast<mbuf_t&>(value))
                      : "v"(v_offset), "s"(res), "s"(s_offset), "n"(i_offset)
@@ -108,7 +127,7 @@ struct buffer_load<2>
                                    index_t /*flag*/ = 0)
     {
         static_assert(sizeof(T) == 4); // subdword is buggy, use dword buf and convert manually
-        using mbuf_t = float;
+        using mbuf_t = typename impl::buffer_load_trait<2, T>::payload_t;
         asm volatile("buffer_load_ushort %0, %1, %2, %3 offen offset:%4"
                      : "+v"(reinterpret_cast<mbuf_t&>(value))
                      : "v"(v_offset), "s"(res), "s"(s_offset), "n"(i_offset)
@@ -128,7 +147,7 @@ struct buffer_load<1>
                                    index_t /*flag*/ = 0)
     {
         static_assert(sizeof(T) == 4);
-        using mbuf_t = float;
+        using mbuf_t = typename impl::buffer_load_trait<1, T>::payload_t;
         asm volatile("buffer_load_ubyte %0, %1, %2, %3 offen offset:%4"
                      : "+v"(reinterpret_cast<mbuf_t&>(value))
                      : "v"(v_offset), "s"(res), "s"(s_offset), "n"(i_offset)
@@ -152,7 +171,7 @@ struct buffer_load_if<16>
     {
         static_assert(sizeof(T) == 16);
         auto saved_exec = __builtin_amdgcn_read_exec();
-        using mbuf_t    = fp32x4_t;
+        using mbuf_t    = typename impl::buffer_load_trait<16, T>::payload_t;
         static_assert(sizeof(mbuf_t) == sizeof(T));
         asm volatile(
             "v_cmpx_le_u32 exec, 1, %5\n"
@@ -177,7 +196,7 @@ struct buffer_load_if<8>
     {
         static_assert(sizeof(T) == 8);
         auto saved_exec = __builtin_amdgcn_read_exec();
-        using mbuf_t    = fp32x2_t;
+        using mbuf_t    = typename impl::buffer_load_trait<8, T>::payload_t;
         asm volatile(
             "v_cmpx_le_u32 exec, 1, %5\n"
             "buffer_load_dwordx2 %0, %1, %2, %3 offen offset:%4\n"
@@ -201,7 +220,7 @@ struct buffer_load_if<4>
     {
         static_assert(sizeof(T) == 4);
         auto saved_exec = __builtin_amdgcn_read_exec();
-        using mbuf_t    = float;
+        using mbuf_t    = typename impl::buffer_load_trait<4, T>::payload_t;
         asm volatile(
             "v_cmpx_le_u32 exec, 1, %5\n"
             "buffer_load_dword %0, %1, %2, %3 offen offset:%4\n"
@@ -225,7 +244,7 @@ struct buffer_load_if<2>
     {
         static_assert(sizeof(T) == 4);
         auto saved_exec = __builtin_amdgcn_read_exec();
-        using mbuf_t    = float;
+        using mbuf_t    = typename impl::buffer_load_trait<2, T>::payload_t;
         asm volatile(
             "v_cmpx_le_u32 exec, 1, %5\n"
             "buffer_load_ushort %0, %1, %2, %3 offen offset:%4\n"
@@ -249,7 +268,7 @@ struct buffer_load_if<1>
     {
         static_assert(sizeof(T) == 4);
         auto saved_exec = __builtin_amdgcn_read_exec();
-        using mbuf_t    = float;
+        using mbuf_t    = typename impl::buffer_load_trait<1, T>::payload_t;
         asm volatile(
             "v_cmpx_le_u32 exec, 1, %5\n"
             "buffer_load_ubyte %0, %1, %2, %3 offen offset:%4\n"

include/ck_tile/core/config.hpp

@@ -171,3 +171,7 @@
 #ifndef CK_TILE_FMHA_FWD_FAST_EXP2
 #define CK_TILE_FMHA_FWD_FAST_EXP2 0
 #endif
+
+#ifndef CK_TILE_BUFFER_LOAD_RAW_BF16_WA
+#define CK_TILE_BUFFER_LOAD_RAW_BF16_WA 1
+#endif

include/ck_tile/host.hpp

@@ -20,3 +20,4 @@
 #include "ck_tile/host/reference/reference_reduce.hpp"
 #include "ck_tile/host/reference/reference_softmax.hpp"
 #include "ck_tile/host/stream_config.hpp"
+#include "ck_tile/host/timer.hpp"

include/ck_tile/host/device_memory.hpp

@@ -27,7 +27,14 @@ struct DeviceMem
     DeviceMem() : mpDeviceBuf(nullptr), mMemSize(0) {}
     DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
     {
-        HIP_CHECK_ERROR(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
+        if(mMemSize != 0)
+        {
+            HIP_CHECK_ERROR(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
+        }
+        else
+        {
+            mpDeviceBuf = nullptr;
+        }
     }
     void Realloc(std::size_t mem_size)
     {
@@ -36,7 +43,14 @@ struct DeviceMem
             HIP_CHECK_ERROR(hipFree(mpDeviceBuf));
         }
         mMemSize = mem_size;
-        HIP_CHECK_ERROR(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
+        if(mMemSize != 0)
+        {
+            HIP_CHECK_ERROR(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
+        }
+        else
+        {
+            mpDeviceBuf = nullptr;
+        }
     }
     void* GetDeviceBuffer() const { return mpDeviceBuf; }
     std::size_t GetBufferSize() const { return mMemSize; }
@@ -47,15 +61,18 @@ struct DeviceMem
             HIP_CHECK_ERROR(
                 hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
         }
-        else
-        {
-            throw std::runtime_error("ToDevice with an empty pointer");
-        }
+        // else
+        // {
+        //     throw std::runtime_error("ToDevice with an empty pointer");
+        // }
     }
     void ToDevice(const void* p, const std::size_t cpySize) const
     {
-        HIP_CHECK_ERROR(
-            hipMemcpy(mpDeviceBuf, const_cast<void*>(p), cpySize, hipMemcpyHostToDevice));
+        if(mpDeviceBuf)
+        {
+            HIP_CHECK_ERROR(
+                hipMemcpy(mpDeviceBuf, const_cast<void*>(p), cpySize, hipMemcpyHostToDevice));
+        }
     }
     void FromDevice(void* p) const
     {
@@ -63,14 +80,17 @@ struct DeviceMem
         {
             HIP_CHECK_ERROR(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
         }
-        else
-        {
-            throw std::runtime_error("FromDevice with an empty pointer");
-        }
+        // else
+        // {
+        //     throw std::runtime_error("FromDevice with an empty pointer");
+        // }
     }
     void FromDevice(void* p, const std::size_t cpySize) const
     {
-        HIP_CHECK_ERROR(hipMemcpy(p, mpDeviceBuf, cpySize, hipMemcpyDeviceToHost));
+        if(mpDeviceBuf)
+        {
+            HIP_CHECK_ERROR(hipMemcpy(p, mpDeviceBuf, cpySize, hipMemcpyDeviceToHost));
+        }
     }
     void SetZero() const
     {
@@ -82,13 +102,16 @@ struct DeviceMem
     template <typename T>
     void SetValue(T x) const
     {
-        if(mMemSize % sizeof(T) != 0)
+        if(mpDeviceBuf)
         {
-            throw std::runtime_error("wrong! not entire DeviceMem will be set");
-        }
+            if(mMemSize % sizeof(T) != 0)
+            {
+                throw std::runtime_error("wrong! not entire DeviceMem will be set");
+            }
 
-        // TODO: call a gpu kernel to set the value (?)
-        set_buffer_value<T><<<1, 1024>>>(static_cast<T*>(mpDeviceBuf), x, mMemSize / sizeof(T));
+            // TODO: call a gpu kernel to set the value (?)
+            set_buffer_value<T><<<1, 1024>>>(static_cast<T*>(mpDeviceBuf), x, mMemSize / sizeof(T));
+        }
     }
     ~DeviceMem()
     {

include/ck_tile/host/kernel_launch.hpp

@@ -6,6 +6,7 @@
 #include "ck_tile/core/config.hpp"
 #include "ck_tile/host/stream_config.hpp"
 #include "ck_tile/host/hip_check_error.hpp"
+#include "ck_tile/host/timer.hpp"
 #include <hip/hip_runtime.h>
 #include <cstddef>
 
@@ -14,153 +15,92 @@ template <int MaxThreadPerBlock, int MinBlockPerCu, typename Kernel, typename...
 #if CK_TILE_USE_LAUNCH_BOUNDS
 __launch_bounds__(MaxThreadPerBlock, MinBlockPerCu)
 #endif
-    __global__ void kentry(Kernel f, Args... args)
+    __global__ void kentry(Args... args)
 {
-    f(args...);
+    Kernel{}(args...);
 }
 
-template <typename... Args, typename F>
-CK_TILE_HOST float launch_and_time_kernel(const stream_config& s,
-                                          F kernel,
-                                          dim3 grid_dim,
-                                          dim3 block_dim,
-                                          std::size_t lds_byte,
-                                          Args... args)
+//
+// return a anonymous functor(lambda) to be called later
+// the KernelImpl should be a class without non-static data member, or let's say
+// can be instantiate with "KernelImpl{}"
+//
+// the "static __device__ operator()(some_arg)" is the entry point of KernelImpl
+//
+template <int MaxThreadPerBlock = CK_TILE_MAX_THREAD_PER_BLOCK,
+          int MinBlockPerCu     = CK_TILE_MIN_BLOCK_PER_CU,
+          typename KernelImpl,
+          typename... Args>
+CK_TILE_HOST auto
+make_kernel(KernelImpl /*f*/, dim3 grid_dim, dim3 block_dim, std::size_t lds_byte, Args... args)
 {
-#if CK_TILE_TIME_KERNEL
-    if(s.time_kernel_)
-    {
-        // warm up
-        for(int i = 0; i < s.cold_niters_; ++i)
-        {
-            kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
-            hip_check_error(hipGetLastError());
-        }
-
-        const int nrepeat = s.nrepeat_;
-        hipEvent_t start, stop;
-
-        HIP_CHECK_ERROR(hipEventCreate(&start));
-        HIP_CHECK_ERROR(hipEventCreate(&stop));
-
-        HIP_CHECK_ERROR(hipDeviceSynchronize());
-        HIP_CHECK_ERROR(hipEventRecord(start, s.stream_id_));
-
-        for(int i = 0; i < nrepeat; ++i)
-        {
-            kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
-            hip_check_error(hipGetLastError());
-        }
-
-        HIP_CHECK_ERROR(hipEventRecord(stop, s.stream_id_));
-        HIP_CHECK_ERROR(hipEventSynchronize(stop));
-
-        float total_time = 0;
-
-        HIP_CHECK_ERROR(hipEventElapsedTime(&total_time, start, stop));
+    const auto kernel = kentry<MaxThreadPerBlock, MinBlockPerCu, KernelImpl, Args...>;
 
-        return total_time / nrepeat;
-    }
-    else
-    {
+    return [=](const stream_config& s) {
         kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
-        hip_check_error(hipGetLastError());
-        return 0;
-    }
-#else
-    kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
-    hip_check_error(hipGetLastError());
-    return 0;
-#endif
+    };
 }
 
-template <typename... Args, typename F, typename PreProcessFunc>
-CK_TILE_HOST float launch_and_time_kernel_with_preprocess(const stream_config& s,
-                                                          PreProcessFunc preprocess,
-                                                          F kernel,
-                                                          dim3 grid_dim,
-                                                          dim3 block_dim,
-                                                          std::size_t lds_byte,
-                                                          Args... args)
+// clang-format off
+/*
+ * launch_kernel()
+ *
+ * this is the function to launch arbitrary number of kernels with optional timer(selected by stream_config)
+ * the callables should have signature as "operator()(const stream_config& s){ ... }" to call
+ * 
+ * the simplest way is pass in a lambda function, with "[=](const stream_config& s){ call_your_kernel_here() }"
+ * as signature, for the callable (pay attention to the capture list)
+ * 
+ * e.g.
+ *  ck_tile::launch_kernel(s,
+ *                      [=](const stream_config& s){ hipMemset(ptr, 0, size) },
+ *                      [=](const stream_config& s){ some_kernel<<<grids, blocks>>>(arg); }
+ *                      );
+ * 
+ * if you use ck_tile kernel, or similiar to this style (structure with "static __device__ operator()(...){}")
+ * you can pass your kernel to ck_tile::make_kernel(), which will create a anonymous functor for you,
+ * then pass it to ck_tile::launch_kernel()
+ * 
+ * e.g.
+ *  ck_tile::launch_kernel(s,
+ *                      ck_tile::make_kernel<T0, B0>(kernel_0{}, grids0, blocks0, 0, kargs0),
+ *                      ck_tile::make_kernel<T0, B1>(kernel_1{}, grids1, blocks1, 0, kargs1),
+ *                       ...);
+ **/
+// clang-format on
+template <typename... Callables>
+CK_TILE_HOST float launch_kernel(const stream_config& s, Callables... callables)
 {
-#if CK_TILE_TIME_KERNEL
-    if(s.time_kernel_)
-    {
-#if CK_TILE_DEBUG_LOG
-        printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
-               __func__,
-               grid_dim.x,
-               grid_dim.y,
-               grid_dim.z,
-               block_dim.x,
-               block_dim.y,
-               block_dim.z);
-
-        printf("Warm up 1 time\n");
-#endif
-        // warm up
-        preprocess();
-        kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
-        hip_check_error(hipGetLastError());
-
-        const int nrepeat = 10;
-#if CK_TILE_DEBUG_LOG
-        printf("Start running %d times...\n", nrepeat);
-#endif
-        hipEvent_t start, stop;
-
-        HIP_CHECK_ERROR(hipEventCreate(&start));
-        HIP_CHECK_ERROR(hipEventCreate(&stop));
-
-        HIP_CHECK_ERROR(hipDeviceSynchronize());
-        HIP_CHECK_ERROR(hipEventRecord(start, s.stream_id_));
+    // clang-format off
+    if(!s.time_kernel_) {
+        (callables(s),...); hip_check_error(hipGetLastError());
+        return 0;
+    }
+    if(s.is_gpu_timer_) {
+        gpu_timer timer {};
 
-        for(int i = 0; i < nrepeat; ++i)
-        {
-            preprocess();
-            kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
-            hip_check_error(hipGetLastError());
-        }
+        // warmup
+        for(int i = 0; i < s.cold_niters_; i++) { (callables(s),...); } hip_check_error(hipGetLastError());
 
-        HIP_CHECK_ERROR(hipEventRecord(stop, s.stream_id_));
-        HIP_CHECK_ERROR(hipEventSynchronize(stop));
+        timer.start(s.stream_id_);
+        for(int i = 0; i < s.nrepeat_; i++) { (callables(s),...); } hip_check_error(hipGetLastError());
+        timer.stop(s.stream_id_);
 
-        float total_time = 0;
+        return timer.duration() / s.nrepeat_;
+    }
+    else {
+        cpu_timer timer {};
 
-        HIP_CHECK_ERROR(hipEventElapsedTime(&total_time, start, stop));
+        // warmup
+        for(int i = 0; i < s.cold_niters_; i++) { (callables(s),...); } hip_check_error(hipGetLastError());
 
-        return total_time / nrepeat;
-    }
-    else
-    {
-        preprocess();
-        kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
-        hip_check_error(hipGetLastError());
+        timer.start(s.stream_id_);
+        for(int i = 0; i < s.nrepeat_; i++) { (callables(s),...); } hip_check_error(hipGetLastError());
+        timer.stop(s.stream_id_);
 
-        return 0;
+        return timer.duration() / s.nrepeat_;
     }
-#else
-    kernel<<<grid_dim, block_dim, lds_byte, s.stream_id_>>>(args...);
-    hip_check_error(hipGetLastError());
-
-    return 0;
-#endif
+    // clang-format on
 }
 
-template <int MaxThreadPerBlock = CK_TILE_MAX_THREAD_PER_BLOCK,
-          int MinBlockPerCu     = CK_TILE_MIN_BLOCK_PER_CU,
-          typename KernelImpl,
-          typename... Args>
-CK_TILE_HOST float launch_kernel(const stream_config& s,
-                                 KernelImpl kernel_impl,
-                                 dim3 grid_dim,
-                                 dim3 block_dim,
-                                 std::size_t dynamic_smem_byte,
-                                 Args... args)
-{
-    const auto kernel = kentry<MaxThreadPerBlock, MinBlockPerCu, KernelImpl, Args...>;
-
-    return launch_and_time_kernel(
-        s, kernel, grid_dim, block_dim, dynamic_smem_byte, kernel_impl, args...);
-}
 } // namespace ck_tile