tempsave, failed to generate ideal code based on iglp

example/01_gemm/gemm_xdl_fp16.cpp

@@ -7,7 +7,6 @@
 #include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp"
 
-
 using ADataType        = ck::half_t;
 using BDataType        = ck::half_t;
 using AccDataType      = float;
@@ -18,7 +17,7 @@ using F16 = ck::half_t;
 using F32 = float;
 
 using ALayout = Row;
-using BLayout = Row;
+using BLayout = Col;
 using CLayout = Row;
 
 using AElementOp = PassThrough;
@@ -213,12 +212,11 @@ using DeviceGemmFactory =
         S<8, 32, 1>,  S<0, 2, 1>,  S<0, 2, 1>, 
         1, 4, 2, 0,
         1, 1, S<1, 32, 1, 8>, 8>,
-#endif
     ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle<
         Row,   Row,  Row,   
         F16,   F16,  F16,  F32,  F16, 
         PassThrough, PassThrough, PassThrough, GemmDefault, 
-        2,   256,
+        1,   256,
         256, 256, 
         32, 8, 4,
         32,   32,
@@ -228,24 +226,25 @@ using DeviceGemmFactory =
         S<4, 64, 1>,  S<0, 2, 1>,  S<0, 2, 1>, 
         1, 4, 4, 1,
         1, 1, S<1, 32, 1, 8>, 8,
-        ck::LoopScheduler::Default, ck::PipelineVersion::v1>,
+        ck::LoopScheduler::Default, ck::PipelineVersion::v1>
+#endif
     ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle<
-        Row,   Row,  Row,   
+        Row,   Col,  Row,   
         F16,   F16,  F16,  F32,  F16, 
         PassThrough, PassThrough, PassThrough, GemmDefault, 
-        1,   256,
+        2,   256,
         256, 256, 
-        32, 8, 4,
+        32, 8, 8,
         32,   32,
         4,    4, 
         S<4, 64, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
-        2, 8, 8, 1,
-        S<4, 64, 1>,  S<0, 2, 1>,  S<0, 2, 1>, 
-        1, 4, 4, 1,
+        2, 8, 8, 0,
+        S<4, 64, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 8, 8, 0,
         1, 1, S<1, 32, 1, 8>, 8,
         ck::LoopScheduler::Default, ck::PipelineVersion::v1>
-    >;
 
+    >;
 
 // clang-format on
 

example/01_gemm/gemm_xdl_fp8.cpp

@@ -7,12 +7,14 @@
 
 using ADataType        = ck::f8_t;
 using BDataType        = ck::f8_t;
-using CDataType        = ck::half_t;
+using CDataType        = ck::f8_t;
 using AccDataType      = float;
 using CShuffleDataType = float;
 
+using F8      = ck::f8_t;
+using F32     = float;
 using ALayout = Row;
-using BLayout = Col;
+using BLayout = Row;
 using CLayout = Row;
 
 using AElementOp = PassThrough;
@@ -21,15 +23,138 @@ using CElementOp = PassThrough;
 
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 
-// clang-format off
-using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
-// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| 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|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  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|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
-// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   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,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 64, 1, 4>,               8>;
-// clang-format on
-
+using DeviceGemmFactory = std::tuple<
+#if 1
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle<Row,
+                                                          Row,
+                                                          Row,
+                                                          F8,
+                                                          F8,
+                                                          F8,
+                                                          F32,
+                                                          F8,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          GemmDefault,
+                                                          1,
+                                                          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,
+                                                          1,
+                                                          S<8, 32, 1>,
+                                                          S<0, 2, 1>,
+                                                          S<0, 2, 1>,
+                                                          1,
+                                                          4,
+                                                          4,
+                                                          0,
+                                                          1,
+                                                          1,
+                                                          S<1, 64, 1, 4>,
+                                                          16,
+                                                          ck::LoopScheduler::Interwave,
+                                                          ck::PipelineVersion::v1>,
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle<Row,
+                                                          Row,
+                                                          Row,
+                                                          F8,
+                                                          F8,
+                                                          F8,
+                                                          F32,
+                                                          F8,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          GemmDefault,
+                                                          1,
+                                                          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,
+                                                          1,
+                                                          S<4, 64, 1>,
+                                                          S<0, 2, 1>,
+                                                          S<0, 2, 1>,
+                                                          1,
+                                                          2,
+                                                          16,
+                                                          1,
+                                                          1,
+                                                          1,
+                                                          S<1, 64, 1, 4>,
+                                                          16,
+                                                          ck::LoopScheduler::Interwave,
+                                                          ck::PipelineVersion::v1>,
+#endif
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle<Row,
+                                                          Row,
+                                                          Row,
+                                                          F8,
+                                                          F8,
+                                                          F8,
+                                                          F32,
+                                                          F8,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          GemmDefault,
+                                                          1,
+                                                          256,
+                                                          256,
+                                                          128,
+                                                          64,
+                                                          16,
+                                                          8,
+                                                          32,
+                                                          32,
+                                                          4,
+                                                          2,
+                                                          S<4, 64, 1>,
+                                                          S<1, 0, 2>,
+                                                          S<1, 0, 2>,
+                                                          2,
+                                                          16,
+                                                          16,
+                                                          1,
+                                                          S<8, 32, 1>,
+                                                          S<0, 2, 1>,
+                                                          S<0, 2, 1>,
+                                                          1,
+                                                          4,
+                                                          8,
+                                                          1,
+                                                          1,
+                                                          1,
+                                                          S<1, 64, 1, 4>,
+                                                          16,
+                                                          ck::LoopScheduler::Interwave,
+                                                          ck::PipelineVersion::v1>>;
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 

example/01_gemm/run_gemm_example.inc

@@ -3,8 +3,6 @@
 
 #pragma once
 
-#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
-
 template <typename ProblemType>
 bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 {
@@ -68,6 +66,26 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
         ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
         ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
         break;
+    case 2:
+        ck::utils::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    case 3:
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+        break;
+    case 4:
+        ck::utils::FillUniformDistribution<ADataType>{0.0f, 0.1f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-0.01f, 0.01f}(b_k_n);
+        break;
+    case 5:
+        ck::utils::FillConstant<ADataType>{static_cast<ADataType>(1.f)}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    case 6:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillConstant<BDataType>{static_cast<BDataType>(1.f)}(b_k_n);
+        break;
     default:
         ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
         ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
@@ -99,62 +117,24 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
     a_m_k_device_buf.ToDevice(a_m_k.mData.data());
     b_k_n_device_buf.ToDevice(b_k_n.mData.data());
 #endif
-    DeviceMem workspace;
 
     auto a_element_op = AElementOp{};
     auto b_element_op = BElementOp{};
     auto c_element_op = CElementOp{};
 
-    using BaseStreamK = ck::tensor_operation::device::DeviceGemmStreamK<ALayout,
-                                                                        BLayout,
-                                                                        CLayout,
-                                                                        ADataType,
-                                                                        BDataType,
-                                                                        CDataType,
-                                                                        AElementOp,
-                                                                        BElementOp,
-                                                                        CElementOp>;
-
-    // do GEMM
-    auto gemm      = DeviceGemmInstance{};
-    auto invoker   = gemm.MakeInvoker();
-    float ave_time = 0;
-
-    if constexpr(std::is_same<ProblemType, ProblemSize>::value &&
-                 !std::is_base_of<BaseStreamK, DeviceGemmInstance>::value)
-    {
-        auto argument = gemm.MakeArgument(
-#ifdef BUILD_INT4_EXAMPLE
-            static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
-            static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
-            static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
-#else
-            static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
-            static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
-            static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
-#endif
-            M,
-            N,
-            K,
-            StrideA,
-            StrideB,
-            StrideC,
-            a_element_op,
-            b_element_op,
-            c_element_op);
+    float best_perf         = .0;
+    float best_time         = .0;
+    std::string best_kernel = "";
 
-        if(!gemm.IsSupportedArgument(argument))
-        {
-            std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+    ck::static_for<0, std::tuple_size_v<DeviceGemmFactory>, 1>{}([&](auto i) -> void {
+        const auto device_gemm_instance = std::get<i>(DeviceGemmFactory{});
 
-            return true;
-        }
+        using DeviceGemmInstance = ck::remove_cvref_t<decltype(device_gemm_instance)>;
+        // do GEMM
+        auto gemm      = DeviceGemmInstance{};
+        auto invoker   = gemm.MakeInvoker();
+        float ave_time = 0;
 
-        ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
-    }
-    else if constexpr(std::is_same<ProblemType, ProblemSizeStreamK>::value &&
-                      std::is_base_of<BaseStreamK, DeviceGemmInstance>::value)
-    {
         auto argument = gemm.MakeArgument(
 #ifdef BUILD_INT4_EXAMPLE
             static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
@@ -173,51 +153,47 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
             StrideC,
             a_element_op,
             b_element_op,
-            c_element_op,
-            problem_size.NumSKBlocks);
-
+            c_element_op);
+#if 0
         if(!gemm.IsSupportedArgument(argument))
         {
             std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
 
             return true;
         }
-
-        std::size_t workspace_size = gemm.GetWorkSpaceSize(&argument);
-        if(workspace_size != 0)
-        {
-            workspace.Realloc(workspace_size);
-            gemm.SetWorkSpacePointer(&argument, workspace.GetDeviceBuffer());
-        }
-
-        ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
-
-#if 0
-        // TODO!!!!!
-        if(workspace_size != 0){
-            float * ws_ptr = reinterpret_cast<float*>(malloc(workspace_size));
-            size_t ws_dwords = workspace_size / sizeof(float);
-            workspace.FromDevice(ws_ptr);
-
-            for(size_t i = 0; i < ws_dwords; i++) {
-                uint32_t rere = reinterpret_cast<uint32_t*>(ws_ptr)[i];
-                printf("%4lu : %f(0x%08x)\n", i, ws_ptr[i], rere);
-            }
-            free(ws_ptr);
-        }
 #endif
-    }
-
-    std::size_t flop = 2_uz * M * N * K;
-    std::size_t num_btype =
-        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
 
-    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+        std::size_t flop = 2_uz * M * N * K;
+        std::size_t num_btype =
+            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
 
-    float gb_per_sec = num_btype / 1.E6 / ave_time;
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+        if(tflops > best_perf)
+        {
+            best_perf   = tflops;
+            best_time   = ave_time;
+            best_kernel = gemm.GetTypeString();
+        }
 
-    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
-              << gemm.GetTypeString() << std::endl;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    });
+
+    std::cout << "---------------------------------------------------------------------------------"
+                 "-----------"
+              << std::endl;
+    std::cout << "Problem Size: M: " << M << ", N: " << N << ", K: " << K << std::endl;
+    std::cout << "---------------------------------------------------------------------------------"
+                 "-----------"
+              << std::endl;
+    std::cout << "Best kernel: " << best_kernel << " , " << best_perf << " TFlops , " << best_time
+              << " ms" << std::endl;
+    std::cout << "---------------------------------------------------------------------------------"
+                 "-----------"
+              << std::endl;
 
     if(config.do_verification)
     {

example/27_layernorm2d_fwd/run_layernorm_example.inc

@@ -44,9 +44,9 @@ int run_layernorm2d_fwd_example()
         {0, 1},
         std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
         std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
-                                    save_mean.mDesc.GetStrides().end()},
+                                 save_mean.mDesc.GetStrides().end()},
         std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
-                                    save_mean.mDesc.GetStrides().end()},
+                                 save_mean.mDesc.GetStrides().end()},
         {1},
         1e-4,
         x_dev.GetDeviceBuffer(),

example/42_groupnorm_fwd/run_groupnorm_fwd_example.inc

@@ -65,9 +65,9 @@ int run_groupnorm_fwd_example(int argc, char* argv[])
         {0, 0, 0, C, 1},
         std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
         std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
-                                    save_mean.mDesc.GetStrides().end()},
+                                 save_mean.mDesc.GetStrides().end()},
         std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
-                                    save_mean.mDesc.GetStrides().end()},
+                                 save_mean.mDesc.GetStrides().end()},
         {1, 2, 4}, // reduction dimension: [H, W, C]
         1e-6,
         x_dev.GetDeviceBuffer(),

example/63_layernorm4d_fwd/run_layernorm4d_fwd_example.inc

@@ -46,9 +46,9 @@ int run_layernorm4d_fwd_example()
         {0, W * C, C, 1},
         std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
         std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
-                                    save_mean.mDesc.GetStrides().end()},
+                                 save_mean.mDesc.GetStrides().end()},
         std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
-                                    save_mean.mDesc.GetStrides().end()},
+                                 save_mean.mDesc.GetStrides().end()},
         {1, 2, 3},
         1e-4,
         x_dev.GetDeviceBuffer(),

include/ck/ck.hpp

@@ -22,7 +22,7 @@
 #ifdef CK_USE_LAUNCH_BOUNDS
 // for most kernels
 #define CK_MAX_THREAD_PER_BLOCK 256
-#define CK_MIN_BLOCK_PER_CU 2
+#define CK_MIN_BLOCK_PER_CU 1
 
 // for wavelet GEMM kernel
 #define CK_WAVELET_MAX_THREAD_PER_BLOCK 512

include/ck/host_utility/kernel_launch.hpp

@@ -51,7 +51,7 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
         hip_check_error(hipDeviceSynchronize());
         hip_check_error(hipEventRecord(start, stream_config.stream_id_));
 
-        for(int i = 0; i < profile_repeat; ++i)
+        for(int i = 0; i < nrepeat; ++i)
         {
             kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
             hip_check_error(hipGetLastError());
@@ -64,7 +64,7 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
 
         hip_check_error(hipEventElapsedTime(&total_time, start, stop));
 
-        return total_time / profile_repeat;
+        return total_time / nrepeat;
     }
     else
     {

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1.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/utility/loop_scheduler.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+#if 0
+template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+__host__ __device__ static constexpr auto
+MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(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>{}));
+}
+#endif
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename ATileDesc,
+    typename BTileDesc,
+    typename AMmaTileDesc,
+    typename BMmaTileDesc,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t KPerBlock,
+    index_t MPerXDL,
+    index_t NPerXDL,
+    index_t MRepeat,
+    index_t NRepeat,
+    index_t KPack,
+    bool TransposeC = false,
+    index_t AMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops,
+    index_t BMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops>
+struct BlockwiseGemmXdlops_pipeline_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm =
+        XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{};
+    static constexpr auto xdlops_gemm_sp =
+        XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack / 2, FloatAB, TransposeC>{};
+
+    static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
+    static constexpr index_t KRepeat    = KPerThread / KPack;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    static_assert(KPerThread % KPack == 0,
+                  "Wrong KPack setting; try increasing KPerThread or decreasing KPack");
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPack * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPack * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(
+            m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
+    }
+
+    using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
+
+    __host__ __device__
+    BlockwiseGemmXdlops_pipeline_v1(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
+                                    Tuple4 b_origin = CalculateBThreadOriginDataIndex())
+        : a_thread_copy_(a_origin), b_thread_copy_(b_origin)
+    {
+        static_assert(AMmaTileDesc::IsKnownAtCompileTime() && BMmaTileDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k;
+    static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k;
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        const BBlockDesc& b_block_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        const BGridBuffer& b_grid_buf,
+                        BBlockBuffer& b_block_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        index_t num_loop) const
+    {
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+        // Global prefetch 1th
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // Write to LDS immediately to save VGPR
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // Global prefetch 2th, Hold in VGPR
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Wait all wave produce this K-loop data
+        block_sync_lds();
+        // preload data into VGPR, 1 kpack of K-loop
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            // read A
+            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                               make_tuple(m0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_,
+                               make_tuple(m0, I0, I0, I0),
+                               a_thread_buf);
+        });
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            // read B
+            b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                               make_tuple(n0, I0, I0, I0),
+                               b_block_buf,
+                               b_thread_desc_,
+                               make_tuple(n0, I0, I0, I0),
+                               b_thread_buf);
+        });
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+#if 0
+            if(get_thread_local_1d_id() == 0)
+            {
+                printf("HasMainLoop=True\n");
+            }
+#endif
+            index_t i = 0;
+
+            do
+            {
+                // Here only KRepeat-1 times read (1~KRepeat) & compute (0~KRepat-1) of this k-loop
+                static_for<1, KRepeat, 1>{}([&](auto k) { // k=1,2 instead of kpack*1, ...
+                    /* Read N+1 */
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        // read A
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(m0, Number<k % 2>{}, I0, I0),
+                                           a_thread_buf);
+                    });
+
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        // read B
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf,
+                                           b_thread_desc_,
+                                           make_tuple(n0, Number<k % 2>{}, I0, I0),
+                                           b_thread_buf);
+                    });
+
+                    /* Compute N */
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, (k - 1) % 2, 0, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, (k - 1) % 2, 0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+                });
+                /*
+                __builtin_amdgcn_sched_group_barrier(0x020, 2,  0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x020, 4,  0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x100, 3,  0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 3,  0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                */
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_barrier(0);
+
+                // Wait all wave consume this k-loop data
+                block_sync_lds();
+
+                // Write 2th prefetch k-loop
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                /* We have entire loop to cover buffer_load latency */
+
+                // Next Global prefetch
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                ++i;
+                // compute(idx=KRepeat) this K-loop can hide ds_write latency
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        /* Compute N */
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / 2, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, (KRepeat - 1) % 2, 0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, (KRepeat - 1) % 2, 0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm_sp.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                __builtin_amdgcn_sched_barrier(0);
+                // Wait all wave produce next k-loop data
+                block_sync_lds();
+                // Here 1 time prefetch read(idx=0) of next K-loop
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    // read A
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, I0),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, I0, I0),
+                                       a_thread_buf);
+                });
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    // read B
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, I0),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(n0, I0, I0, I0),
+                                       b_thread_buf);
+                });
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        /* Compute N */
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / 2, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, (KRepeat - 1) % 2, 0, ik + KPack / 2))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, (KRepeat - 1) % 2, 0, ik + KPack / 2))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm_sp.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+                // Current best = 112T at float initialization 3840x4096x4096
+                // __builtin_amdgcn_sched_group_barrier(0x020, 4, 0); // VMEM read
+                // __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+                // __builtin_amdgcn_sched_group_barrier(0x020, 2, 0); // VMEM read
+                // __builtin_amdgcn_sched_group_barrier(0x100, 6, 0); // DS read
+                // __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+                // __builtin_amdgcn_sched_group_barrier(0x200, 6, 0); // DS write
+                // __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        {
+            // N-2 th k-loop
+            // Here only KRepeat-1 times read & compute
+            static_for<1, KRepeat, 1>{}([&](auto k) { // k=1,2 instead of kpack*1, ...
+                /* Read N+1 */
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    // read A
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, Number<k % 2>{}, I0, I0),
+                                       a_thread_buf);
+                });
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    // read B
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(n0, Number<k % 2>{}, I0, I0),
+                                       b_thread_buf);
+                });
+                /* Compute N */
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, (k - 1) % 2, 0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, (k - 1) % 2, 0, ik))>{}];
+                        });
+#if 0
+                        if(get_thread_local_1d_id() == 0)
+                        {
+                            printf("rep of m.n.k (%01d, %01d, %01d)\n",
+                                   m0.value,
+                                   n0.value,
+                                   k.value - 1);
+                        }
+                        printf(
+                            "Tid: %03d, A_compute_buf: %04x %04x %04x %04x %04x %04x %04x %04x\n",
+                            get_thread_local_1d_id(),
+                            *(reinterpret_cast<uint16_t*>(
+                                &(a_thread_vec.template AsType<FloatAB>()(Number<0>{})))),
+                            *(reinterpret_cast<uint16_t*>(
+                                &(a_thread_vec.template AsType<FloatAB>()(Number<1>{})))),
+                            *(reinterpret_cast<uint16_t*>(
+                                &(a_thread_vec.template AsType<FloatAB>()(Number<2>{})))),
+                            *(reinterpret_cast<uint16_t*>(
+                                &(a_thread_vec.template AsType<FloatAB>()(Number<3>{})))),
+                            *(reinterpret_cast<uint16_t*>(
+                                &(a_thread_vec.template AsType<FloatAB>()(Number<4>{})))),
+                            *(reinterpret_cast<uint16_t*>(
+                                &(a_thread_vec.template AsType<FloatAB>()(Number<5>{})))),
+                            *(reinterpret_cast<uint16_t*>(
+                                &(a_thread_vec.template AsType<FloatAB>()(Number<6>{})))),
+                            *(reinterpret_cast<uint16_t*>(
+                                &(a_thread_vec.template AsType<FloatAB>()(Number<7>{})))));
+#endif
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            __builtin_amdgcn_sched_group_barrier(0x100, 8, 0);  // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 32, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+
+            block_sync_lds();
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+            // compute(idx=KRepeat) this K-loop can hide ds_write latency
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    /* Compute N */
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack / 2, 1>{}([&](auto ik) {
+                        a_thread_vec.template AsType<FloatAB>()(ik) =
+                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                make_tuple(m0, (KRepeat - 1) % 2, 0, ik))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(ik) =
+                            b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                make_tuple(n0, (KRepeat - 1) % 2, 0, ik))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm_sp.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+
+            __builtin_amdgcn_sched_group_barrier(0x200, 8, 0);  // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 32, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+            // wait final data producer finished
+            block_sync_lds();
+
+            // N-1 th k-loop
+            // Here 1 time prefetch read(idx=0) of next K-loop
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                // read A
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, I0),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, I0, I0),
+                                   a_thread_buf);
+            });
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                   make_tuple(n0, I0, I0, I0),
+                                   b_block_buf,
+                                   b_thread_desc_,
+                                   make_tuple(n0, I0, I0, I0),
+                                   b_thread_buf);
+            });
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    /* Compute N */
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack / 2, 1>{}([&](auto ik) {
+                        a_thread_vec.template AsType<FloatAB>()(ik) =
+                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                make_tuple(m0, (KRepeat - 1) % 2, 0, ik + KPack / 2))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(ik) =
+                            b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                make_tuple(n0, (KRepeat - 1) % 2, 0, ik + KPack / 2))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm_sp.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+            __builtin_amdgcn_sched_group_barrier(0x100, 8, 0);  // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 32, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+            // Here only KRepeat-1 times read & compute
+            static_for<1, KRepeat, 1>{}([&](auto k) { // k=1,2 instead of kpack*1, ...
+                /* Read N+1 */
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    // read A
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, Number<k % 2>{}, I0, I0),
+                                       a_thread_buf);
+                });
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    // read B
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(n0, Number<k % 2>{}, I0, I0),
+                                       b_thread_buf);
+                });
+                /* Compute N */
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, (k - 1) % 2, 0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, (k - 1) % 2, 0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+            __builtin_amdgcn_sched_group_barrier(0x100, 8, 0);  // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 32, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+            /* Final Compute issue */
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<FloatAB>()(i) =
+                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                make_tuple(m0, (KRepeat - 1) % 2, 0, i))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(i) =
+                            b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                make_tuple(n0, (KRepeat - 1) % 2, 0, i))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        }
+    }
+
+    protected:
+    // M1, N1 as double buffer index
+    // Read buffer + Compute buffer
+    // A[M0, M1, M2, KPack]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, I2, I1, Number<KPack>{}));
+
+    // B[N0, N1, N2, KPack]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<NRepeat>{}, I2, I1, Number<KPack>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2.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/utility/loop_scheduler.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+#if 0
+template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+__host__ __device__ static constexpr auto
+MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(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>{}));
+}
+#endif
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename ATileDesc,
+    typename BTileDesc,
+    typename AMmaTileDesc,
+    typename BMmaTileDesc,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t KPerBlock,
+    index_t MPerXDL,
+    index_t NPerXDL,
+    index_t MRepeat,
+    index_t NRepeat,
+    index_t KPack,
+    bool TransposeC = false,
+    index_t AMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops,
+    index_t BMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops>
+struct BlockwiseGemmXdlops_pipeline_v2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm =
+        XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{};
+
+    static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
+    static constexpr index_t KRepeat    = KPerThread / KPack;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    static_assert(KPerThread % KPack == 0,
+                  "Wrong KPack setting; try increasing KPerThread or decreasing KPack");
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPack * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPack * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(
+            m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
+    }
+
+    using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
+
+    __host__ __device__
+    BlockwiseGemmXdlops_pipeline_v2(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
+                                    Tuple4 b_origin = CalculateBThreadOriginDataIndex())
+        : a_thread_copy_(a_origin), b_thread_copy_(b_origin)
+    {
+        static_assert(AMmaTileDesc::IsKnownAtCompileTime() && BMmaTileDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k;
+    static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k;
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        const BBlockDesc& b_block_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        const BGridBuffer& b_grid_buf,
+                        BBlockBuffer& b_block_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        index_t num_loop) const
+    {
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+        // Global prefetch 1th
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // Write to LDS immediately to save VGPR
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // Global prefetch 2th, Hold in VGPR
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+#if 0
+            if(get_thread_local_1d_id() == 0)
+            {
+                printf("HasMainLoop=True\n");
+            }
+#endif
+            index_t i = 0;
+
+            do
+            {
+                // Wait all wave produce this K-loop data
+                block_sync_lds();
+                static_for<0, KRepeat, 1>{}([&](auto k) { // k=1,2 instead of kpack*1, ...
+                    /* Read N+1 */
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        // read A
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(m0, Number<k % 2>{}, I0, I0),
+                                           a_thread_buf);
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            // read B
+                            b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                               make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                               b_block_buf,
+                                               b_thread_desc_,
+                                               make_tuple(n0, Number<k % 2>{}, I0, I0),
+                                               b_thread_buf);
+                        });
+                    });
+                });
+
+                /* Compute N */
+                static_for<0, MRepeat / 2, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, 0, 0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, 0, 0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+
+                // schedule 16 ds_read_b128 and 16 v_mfma
+                __builtin_amdgcn_sched_group_barrier(0x100, 8, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                __builtin_amdgcn_sched_barrier(0);
+                block_sync_lds();
+
+                // Write 2th prefetch k-loop
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                /* We have entire loop to cover buffer_load latency */
+
+                // Next Global prefetch
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                static_for<MRepeat / 2, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, 0, 0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, 0, 0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I1, 0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I1, 0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+
+                // schedule 8 ds_write_b128, 8 buffer_load_dwordx4, 48 v_mfma
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(0x008, 6, 0); // MFMA
+                __builtin_amdgcn_sched_barrier(0);
+
+                ++i;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        {
+            // Wait all wave produce this K-loop data
+            block_sync_lds();
+            static_for<0, KRepeat, 1>{}([&](auto k) { // k=1,2 instead of kpack*1, ...
+                /* Read N+1 */
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    // read A
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, Number<k % 2>{}, I0, I0),
+                                       a_thread_buf);
+
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        // read B
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf,
+                                           b_thread_desc_,
+                                           make_tuple(n0, Number<k % 2>{}, I0, I0),
+                                           b_thread_buf);
+                    });
+                });
+            });
+
+            /* Compute N */
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto ik) {
+                        a_thread_vec.template AsType<FloatAB>()(ik) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(m0, 0, 0, ik))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(ik) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(n0, 0, 0, ik))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+
+            // schedule 16 ds_read_b128 and 32 v_mfma
+            __builtin_amdgcn_sched_group_barrier(0x100, 8, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+            block_sync_lds();
+
+            // Write 2th prefetch k-loop
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto ik) {
+                        a_thread_vec.template AsType<FloatAB>()(ik) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(m0, I1, 0, ik))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(ik) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(n0, I1, 0, ik))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+
+            // schedule 8 ds_write_b128, 8 buffer_load_dwordx4, 32 v_mfma
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+
+            // Wait all wave produce this K-loop data
+            block_sync_lds();
+            static_for<0, KRepeat, 1>{}([&](auto k) { // k=1,2 instead of kpack*1, ...
+                /* Read N+1 */
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    // read A
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, Number<k % 2>{}, I0, I0),
+                                       a_thread_buf);
+                });
+
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    // read B
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(n0, Number<k % 2>{}, I0, I0),
+                                       b_thread_buf);
+                });
+            });
+
+            /* Compute N */
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, k % 2, 0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, k % 2, 0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            // schedule 16 ds_read_b128 and 64 v_mfma
+            __builtin_amdgcn_sched_group_barrier(0x100, 8, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+
+            __builtin_amdgcn_sched_group_barrier(0x100, 5, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+        }
+    }
+
+    protected:
+    // M1, N1 as double buffer index
+    // Read buffer + Compute buffer
+    // A[M0, M1, M2, KPack]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, I2, I1, Number<KPack>{}));
+
+    // B[N0, N1, N2, KPack]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<NRepeat>{}, I2, I1, Number<KPack>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.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/utility/loop_scheduler.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+#if 0
+template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+__host__ __device__ static constexpr auto
+MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(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>{}));
+}
+#endif
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename ATileDesc,
+    typename BTileDesc,
+    typename AMmaTileDesc,
+    typename BMmaTileDesc,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t KPerBlock,
+    index_t MPerXDL,
+    index_t NPerXDL,
+    index_t MRepeat,
+    index_t NRepeat,
+    index_t KPack,
+    bool TransposeC = false,
+    index_t AMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops,
+    index_t BMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops>
+struct BlockwiseGemmXdlops_pipeline_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>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm =
+        XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{};
+
+    static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
+    static constexpr index_t KRepeat    = KPerThread / KPack;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    static_assert(KPerThread % KPack == 0,
+                  "Wrong KPack setting; try increasing KPerThread or decreasing KPack");
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPack * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPack * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(
+            m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
+    }
+
+    using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
+
+    __host__ __device__
+    BlockwiseGemmXdlops_pipeline_v3(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
+                                    Tuple4 b_origin = CalculateBThreadOriginDataIndex())
+        : a_thread_copy_(a_origin), b_thread_copy_(b_origin)
+    {
+        static_assert(AMmaTileDesc::IsKnownAtCompileTime() && BMmaTileDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k;
+    static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k;
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        const BBlockDesc& b_block_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        const BGridBuffer& b_grid_buf,
+                        BBlockBuffer& b_block_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        index_t num_loop) const
+    {
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        // Inst List:
+        // ds_read_b128: 16
+        // ds_write_b128: 8
+        // buffer_load_dwordx4: 16
+        // v_mfma: 0
+        // -------------------------------------------------------------------------------------------
+
+        // Global prefetch 1th
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        // Global prefetch 2th
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, k, I0),
+                                   a_thread_buf);
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(n0, I0, k, I0),
+                                       b_thread_buf);
+                });
+            });
+        });
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            // This hot loop has two legacy loopover, to implement the double local buffer strategy
+            do
+            {
+                // Inst List:
+                // ds_read_b128: 32
+                // ds_write_b128: 16
+                // buffer_load_dwordx4: 16
+                // v_mfma: 128
+                // -------------------------------------------------------------------------------------------
+                // I0 As Compute, I1 As Read
+                {
+                    using ComputeLocalBuffer = Number<0>;
+                    using MemoryLocalBuffer  = Number<1>;
+
+                    block_sync_lds();
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                    b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                    b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, ComputeLocalBuffer{}, I0, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, ComputeLocalBuffer{}, I0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+
+                    static_for<0, MRepeat / 2, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, ComputeLocalBuffer{}, I1, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, ComputeLocalBuffer{}, I1, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+
+                    // schedule 8 ds_write_b128, 8 buffer_load_dwordx4, 48 v_mfma
+                    // two single MFMA for hide ds_write & VMEM read issue latency
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_barrier(0);
+
+                    block_sync_lds();
+
+                    static_for<0, KRepeat, 1>{}([&](auto k) {
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                               make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                               a_block_buf,
+                                               a_thread_desc_,
+                                               make_tuple(m0, MemoryLocalBuffer{}, k, I0),
+                                               a_thread_buf);
+                            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                                b_thread_copy_.Run(
+                                    b_block_desc_n0_n1_n2_k,
+                                    make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                    b_block_buf,
+                                    b_thread_desc_,
+                                    make_tuple(n0, MemoryLocalBuffer{}, k, I0),
+                                    b_thread_buf);
+                            });
+                        });
+                    });
+
+                    static_for<MRepeat / 2, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, ComputeLocalBuffer{}, I1, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, ComputeLocalBuffer{}, I1, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+
+                    // schedule 16 ds_read_b128 and 16 v_mfma
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_barrier(0);
+                }
+
+                // -------------------------------------------------------------------------------------------
+                // I1 As Compute, I0 As Read
+                {
+                    using ComputeLocalBuffer = Number<1>;
+                    using MemoryLocalBuffer  = Number<0>;
+
+                    block_sync_lds();
+
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                    b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                    b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, ComputeLocalBuffer{}, I0, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, ComputeLocalBuffer{}, I0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+
+                    static_for<0, MRepeat / 2, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, ComputeLocalBuffer{}, I1, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, ComputeLocalBuffer{}, I1, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+
+                    // schedule 8 ds_write_b128, 8 buffer_load_dwordx4, 48 v_mfma
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+                    __builtin_amdgcn_sched_barrier(0);
+
+                    block_sync_lds();
+
+                    static_for<0, KRepeat, 1>{}([&](auto k) {
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                               make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                               a_block_buf,
+                                               a_thread_desc_,
+                                               make_tuple(m0, MemoryLocalBuffer{}, k, I0),
+                                               a_thread_buf);
+                            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                                b_thread_copy_.Run(
+                                    b_block_desc_n0_n1_n2_k,
+                                    make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                    b_block_buf,
+                                    b_thread_desc_,
+                                    make_tuple(n0, MemoryLocalBuffer{}, k, I0),
+                                    b_thread_buf);
+                            });
+                        });
+                    });
+
+                    static_for<MRepeat / 2, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, ComputeLocalBuffer{}, I1, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, ComputeLocalBuffer{}, I1, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+
+                    // schedule 16 ds_read_b128 and 16 v_mfma
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    __builtin_amdgcn_sched_barrier(0);
+                }
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        {
+            // Inst List:
+            // ds_read_b128: 16
+            // ds_write_b128: 8
+            // buffer_load_dwordx4: 0
+            // v_mfma: 128
+
+            block_sync_lds();
+
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto ik) {
+                        a_thread_vec.template AsType<FloatAB>()(ik) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(m0, I0, I0, ik))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(ik) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(n0, I0, I0, ik))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+            // schedule 8 ds_write_b128 and 32 v_mfma
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 4, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_barrier(0);
+
+            block_sync_lds();
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, I1, k, I0),
+                                       a_thread_buf);
+
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf,
+                                           b_thread_desc_,
+                                           make_tuple(n0, I1, k, I0),
+                                           b_thread_buf);
+                    });
+                });
+            });
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto ik) {
+                        a_thread_vec.template AsType<FloatAB>()(ik) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(m0, I0, I1, ik))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(ik) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(n0, I0, I1, ik))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+
+            // schedule 16 ds_read_b128 and 32 v_mfma
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_barrier(0);
+
+            block_sync_lds();
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I1, k, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I1, k, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            // 64 v_mfma
+            __builtin_amdgcn_sched_group_barrier(0x008, 64, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+        }
+    }
+
+    protected:
+    // M1, N1 as double buffer index
+    // Read buffer + Compute buffer
+    // A[M0, M1, M2, KPack]
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, I2, Number<KRepeat>{}, Number<KPack>{}));
+
+    // B[N0, N1, N2, KPack]
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I2, Number<KRepeat>{}, Number<KPack>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.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/utility/loop_scheduler.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+// Double LDS buffer
+// Prefetech 2 stage
+// Local prefetch 1 stage
+
+namespace ck {
+
+#if 0
+template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+__host__ __device__ static constexpr auto
+MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(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>{}));
+}
+#endif
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename ATileDesc,
+    typename BTileDesc,
+    typename AMmaTileDesc,
+    typename BMmaTileDesc,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t KPerBlock,
+    index_t MPerXDL,
+    index_t NPerXDL,
+    index_t MRepeat,
+    index_t NRepeat,
+    index_t KPack,
+    bool TransposeC = false,
+    index_t AMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops,
+    index_t BMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops>
+struct BlockwiseGemmXdlops_pipeline_v4
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm =
+        XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{};
+
+    static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
+    static constexpr index_t KRepeat    = KPerThread / KPack;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    static_assert(KPerThread % KPack == 0,
+                  "Wrong KPack setting; try increasing KPerThread or decreasing KPack");
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPack * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPack * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(
+            m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
+    }
+
+    using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
+
+    __host__ __device__
+    BlockwiseGemmXdlops_pipeline_v4(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
+                                    Tuple4 b_origin = CalculateBThreadOriginDataIndex())
+        : a_thread_copy_(a_origin), b_thread_copy_(b_origin)
+    {
+        static_assert(AMmaTileDesc::IsKnownAtCompileTime() && BMmaTileDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    __device__ static constexpr auto HotLoopScheduler()
+    {
+        // schedule
+        constexpr auto num_ds_read_inst = 16;
+        constexpr auto num_ds_write_inst = 8;
+        constexpr auto num_buffer_load_inst = 8;
+        constexpr auto num_mfma_inst = 64;
+
+        constexpr auto num_issue = num_buffer_load_inst;
+
+        static_for<0, num_issue, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, num_ds_read_inst/num_buffer_load_inst, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, num_ds_write_inst/num_buffer_load_inst, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+            __builtin_amdgcn_sched_group_barrier(0x008, num_mfma_inst/num_buffer_load_inst-3, 0); // MFMA
+        });
+    }
+
+    template <index_t stage>
+    __device__ static constexpr auto TailScheduler()
+    {
+    }
+
+    template <>
+    __device__ static constexpr auto TailScheduler<1>()
+    {
+        // schedule
+        constexpr auto num_ds_read_inst = 16;
+        constexpr auto num_ds_write_inst = 8;
+        constexpr auto num_mfma_inst = 64;
+
+        constexpr auto num_issue = num_ds_write_inst;
+
+        static_for<0, num_issue, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, num_ds_read_inst/num_ds_write_inst-1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, num_mfma_inst/num_ds_write_inst-3, 0); // MFMA
+        });
+    }
+
+    template <>
+    __device__ static constexpr auto TailScheduler<2>()
+    {
+        // schedule
+        constexpr auto num_ds_read_inst = 16;
+        constexpr auto num_mfma_inst = 64;
+
+        constexpr auto num_issue = num_ds_read_inst;
+
+        static_for<0, num_issue, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, num_mfma_inst/num_ds_read_inst, 0); // MFMA
+        });
+    }
+
+    static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k;
+    static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k;
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        const BBlockDesc& b_block_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        const BGridBuffer& b_grid_buf,
+                        BBlockBuffer& b_block_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        index_t num_loop) const
+    {
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+        // Inst List:
+        // ds_read_b128: 16
+        // ds_write_b128: 8
+        // buffer_load_dwordx4: 16
+        // v_mfma: 0
+        // -------------------------------------------------------------------------------------------
+
+        // Global prefetch 1th, Fill Ping LDS
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0));
+
+        // Local prefetch 1th, Fill Ping Reg
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                   a_block_buf.At(I0),
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, k, I0),
+                                   a_thread_bufs(I0));
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf.At(I0),
+                                       b_thread_desc_,
+                                       make_tuple(n0, I0, k, I0),
+                                       b_thread_bufs(I0));
+                });
+            });
+        });
+
+        // Global prefetch 2th, Fill Pong LDS
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1));
+
+        // Global prefetch 3rd
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            // This hot loop has two legacy loopover, to implement the double local buffer strategy
+            do
+            {
+                // -------------------------------------------------------------------------------------------
+                using PingP1 = Number<0>;
+                using PongP1 = Number<1>;
+                // MFMA: Ping Reg
+                // DS_WRITE: To Ping LDS
+                // DS_READ: Pong LDS to Pong Reg
+                block_sync_lds();
+
+                static_for<0, KRepeat, 1>{}([&](auto k) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                           a_block_buf.At(PongP1{}),
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k, I0),
+                                           a_thread_bufs(PongP1{}));
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                               make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                               b_block_buf.At(PongP1{}),
+                                               b_thread_desc_,
+                                               make_tuple(n0, I0, k, I0),
+                                               b_thread_bufs(PongP1{}));
+                        });
+                    });
+                });
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(PingP1{}));
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(PingP1{}));
+
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_bufs[PingP1{}][Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, I0, k0, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_bufs[PingP1{}][Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, I0, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+                });
+
+                HotLoopScheduler();
+                __builtin_amdgcn_sched_barrier(0);
+
+                // -------------------------------------------------------------------------------------------
+                using PingP2 = Number<1>;
+                using PongP2 = Number<0>;
+                // MFMA: Pong Reg
+                // DS_WRITE: To Pong LDS
+                // DS_READ: Ping LDS to Ping Reg
+                block_sync_lds();
+
+                static_for<0, KRepeat, 1>{}([&](auto k) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                           a_block_buf.At(PongP2{}),
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k, I0),
+                                           a_thread_bufs(PongP2{}));
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                               make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                               b_block_buf.At(PongP2{}),
+                                               b_thread_desc_,
+                                               make_tuple(n0, I0, k, I0),
+                                               b_thread_bufs(PongP2{}));
+                        });
+                    });
+                });
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(PingP2{}));
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(PingP2{}));
+
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_bufs[PingP2{}][Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, I0, k0, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_bufs[PingP2{}][Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, I0, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+                });
+
+                HotLoopScheduler();
+                __builtin_amdgcn_sched_barrier(0);
+
+                i += 2;
+            } while(i < (num_loop - 3));
+        }
+
+        // tail
+        {
+            using PingP1 = Number<0>;
+            using PongP1 = Number<1>;
+            // MFMA: Ping Reg
+            // DS_WRITE: To Ping LDS
+            // DS_READ: Pong LDS to Pong Reg
+            block_sync_lds();
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf.At(PongP1{}),
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, I0),
+                                       a_thread_bufs(PongP1{}));
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf.At(PongP1{}),
+                                           b_thread_desc_,
+                                           make_tuple(n0, I0, k, I0),
+                                           b_thread_bufs(PongP1{}));
+                    });
+                });
+            });
+
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(PingP1{}));
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(PingP1{}));
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_bufs[PingP1{}][Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_bufs[PingP1{}][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            TailScheduler<1>();
+            __builtin_amdgcn_sched_barrier(0);
+
+            // -------------------------------------------------------------------------------------------
+            using PingP2 = Number<1>;
+            using PongP2 = Number<0>;
+            // MFMA: Pong Reg
+            // DS_WRITE: To Pong LDS
+            // DS_READ: Ping LDS to Ping Reg
+            block_sync_lds();
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf.At(PongP2{}),
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, I0),
+                                       a_thread_bufs(PongP2{}));
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf.At(PongP2{}),
+                                           b_thread_desc_,
+                                           make_tuple(n0, I0, k, I0),
+                                           b_thread_bufs(PongP2{}));
+                    });
+                });
+            });
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_bufs[PingP2{}][Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_bufs[PingP2{}][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            TailScheduler<2>();
+            __builtin_amdgcn_sched_barrier(0);
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_bufs[PongP2{}][Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_bufs[PongP2{}][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            // 64 v_mfma
+            __builtin_amdgcn_sched_group_barrier(0x008, 64, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+        }
+    }
+
+    protected:
+    // M1, N1 as double buffer index
+    // Read buffer + Compute buffer
+    // A[M0, M1, M2, KPack]
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
+        make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}),
+        make_tuple(
+            Number<KPack>{}, Number<KPack * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
+
+    // B[N0, N1, N2, KPack]
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor(
+        make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}),
+        make_tuple(
+            Number<KPack>{}, Number<KPack * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4_r2.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/utility/loop_scheduler.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+// Double LDS buffer
+// Prefetech 2 stage
+// Local prefetch 1 stage
+
+namespace ck {
+
+#if 0
+template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+__host__ __device__ static constexpr auto
+MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(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>{}));
+}
+#endif
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename ATileDesc,
+    typename BTileDesc,
+    typename AMmaTileDesc,
+    typename BMmaTileDesc,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t KPerBlock,
+    index_t MPerXDL,
+    index_t NPerXDL,
+    index_t MRepeat,
+    index_t NRepeat,
+    index_t KPack,
+    bool TransposeC = false,
+    index_t AMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops,
+    index_t BMmaKStride =
+        KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops>
+struct BlockwiseGemmXdlops_pipeline_v4
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm =
+        XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{};
+
+    static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
+    static constexpr index_t KRepeat    = KPerThread / KPack;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    static_assert(KPerThread % KPack == 0,
+                  "Wrong KPack setting; try increasing KPerThread or decreasing KPack");
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPack * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPack * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(
+            m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
+    }
+
+    using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
+
+    __host__ __device__
+    BlockwiseGemmXdlops_pipeline_v4(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
+                                    Tuple4 b_origin = CalculateBThreadOriginDataIndex())
+        : a_thread_copy_(a_origin), b_thread_copy_(b_origin)
+    {
+        static_assert(AMmaTileDesc::IsKnownAtCompileTime() && BMmaTileDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    __device__ static constexpr auto HotLoopScheduler()
+    {
+        // schedule
+        constexpr auto num_ds_read_inst = 16;
+        constexpr auto num_ds_write_inst = 8;
+        constexpr auto num_buffer_load_inst = 8;
+        constexpr auto num_mfma_inst = 64;
+
+        constexpr auto num_issue = num_buffer_load_inst;
+
+        static_for<0, num_issue, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, num_ds_read_inst/num_buffer_load_inst, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, num_ds_write_inst/num_buffer_load_inst, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+            __builtin_amdgcn_sched_group_barrier(0x008, num_mfma_inst/num_buffer_load_inst-3, 0); // MFMA
+        });
+    }
+
+    template <index_t stage>
+    __device__ static constexpr auto TailScheduler()
+    {
+    }
+
+    template <>
+    __device__ static constexpr auto TailScheduler<1>()
+    {
+        // schedule
+        constexpr auto num_ds_read_inst = 16;
+        constexpr auto num_ds_write_inst = 8;
+        constexpr auto num_mfma_inst = 64;
+
+        constexpr auto num_issue = num_ds_write_inst;
+
+        static_for<0, num_issue, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, num_ds_read_inst/num_ds_write_inst-1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, num_mfma_inst/num_ds_write_inst-3, 0); // MFMA
+        });
+    }
+
+    template <>
+    __device__ static constexpr auto TailScheduler<2>()
+    {
+        // schedule
+        constexpr auto num_ds_read_inst = 16;
+        constexpr auto num_mfma_inst = 64;
+
+        constexpr auto num_issue = num_ds_read_inst;
+
+        static_for<0, num_issue, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            __builtin_amdgcn_sched_group_barrier(0x008, num_mfma_inst/num_ds_read_inst, 0); // MFMA
+        });
+    }
+
+    static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k;
+    static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k;
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        const BBlockDesc& b_block_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        const BGridBuffer& b_grid_buf,
+                        BBlockBuffer& b_block_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        index_t num_loop) const
+    {
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        // -------------------------------------------------------------------------------------------
+
+        // Global prefetch 1th, Fill Ping LDS
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0));
+
+        // Local prefetch 1th, Fill Ping Reg
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                   a_block_buf.At(I0),
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, k, I0),
+                                   a_thread_buf);
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf.At(I0),
+                                       b_thread_desc_,
+                                       make_tuple(n0, I0, k, I0),
+                                       b_thread_buf);
+                });
+            });
+        });
+
+        // Global prefetch 2th, Fill Pong LDS
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1));
+
+        // Global prefetch 3rd
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            // This hot loop has two legacy loopover, to implement the double local buffer strategy
+            do
+            {
+                // -------------------------------------------------------------------------------------------
+                using PingP1 = Number<0>;
+                using PongP1 = Number<1>;
+                // MFMA: Ping Reg
+                // DS_WRITE: To Ping LDS
+                // DS_READ: Pong LDS to Pong Reg
+                block_sync_lds();
+
+                static_for<0, KRepeat, 1>{}([&](auto k) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                           a_block_buf.At(PongP1{}),
+                                           a_thread_desc_,
+                                           make_tuple(m0, PongP1{}, k, I0),
+                                           a_thread_buf);
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                               make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                               b_block_buf.At(PongP1{}),
+                                               b_thread_desc_,
+                                               make_tuple(n0, PongP1{}, k, I0),
+                                               b_thread_buf);
+                        });
+                    });
+                });
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(PingP1{}));
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(PingP1{}));
+
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, PingP1{}, k0, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, PingP1{}, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+                });
+
+                HotLoopScheduler();
+                __builtin_amdgcn_sched_barrier(0);
+
+                // -------------------------------------------------------------------------------------------
+                using PingP2 = Number<1>;
+                using PongP2 = Number<0>;
+                // MFMA: Pong Reg
+                // DS_WRITE: To Pong LDS
+                // DS_READ: Ping LDS to Ping Reg
+                block_sync_lds();
+
+                static_for<0, KRepeat, 1>{}([&](auto k) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                           a_block_buf.At(PongP2{}),
+                                           a_thread_desc_,
+                                           make_tuple(m0, PongP2{}, k, I0),
+                                           a_thread_buf);
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                               make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                               b_block_buf.At(PongP2{}),
+                                               b_thread_desc_,
+                                               make_tuple(n0, PongP2{}, k, I0),
+                                               b_thread_buf);
+                        });
+                    });
+                });
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(PingP2{}));
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(PingP2{}));
+
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            vector_type<FloatAB, KPack> a_thread_vec;
+                            vector_type<FloatAB, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<FloatAB>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, PingP2{}, k0, ik))>{}];
+                                b_thread_vec.template AsType<FloatAB>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, PingP2{}, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            xdlops_gemm.template Run(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+                });
+
+                HotLoopScheduler();
+                __builtin_amdgcn_sched_barrier(0);
+
+                i += 2;
+            } while(i < (num_loop - 3));
+        }
+
+        // tail
+        {
+            using PingP1 = Number<0>;
+            using PongP1 = Number<1>;
+            // MFMA: Ping Reg
+            // DS_WRITE: To Ping LDS
+            // DS_READ: Pong LDS to Pong Reg
+            block_sync_lds();
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf.At(PongP1{}),
+                                       a_thread_desc_,
+                                       make_tuple(m0, PongP1{}, k, I0),
+                                       a_thread_buf);
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf.At(PongP1{}),
+                                           b_thread_desc_,
+                                           make_tuple(n0, PongP1{}, k, I0),
+                                           b_thread_buf);
+                    });
+                });
+            });
+
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(PingP1{}));
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(PingP1{}));
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, PingP1{}, k0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, PingP1{}, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            TailScheduler<1>();
+            __builtin_amdgcn_sched_barrier(0);
+
+            // -------------------------------------------------------------------------------------------
+            using PingP2 = Number<1>;
+            using PongP2 = Number<0>;
+            // MFMA: Pong Reg
+            // DS_WRITE: To Pong LDS
+            // DS_READ: Ping LDS to Ping Reg
+            block_sync_lds();
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf.At(PongP2{}),
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, I0),
+                                       a_thread_buf);
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf.At(PongP2{}),
+                                           b_thread_desc_,
+                                           make_tuple(n0, I0, k, I0),
+                                           b_thread_buf);
+                    });
+                });
+            });
+
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, PingP2{}, k0, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, PingP2{}, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            TailScheduler<2>();
+            __builtin_amdgcn_sched_barrier(0);
+
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<FloatAB>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, PongP2{}, k, ik))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, PongP2{}, k, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            // 64 v_mfma
+            __builtin_amdgcn_sched_group_barrier(0x008, 64, 0); // MFMA
+            __builtin_amdgcn_sched_barrier(0);
+        }
+    }
+
+    protected:
+    // M1, N1 as double buffer index
+    // Read buffer + Compute buffer
+    // A[M0, M1, M2, KPack]
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
+        make_tuple(Number<MRepeat>{}, I2, Number<KRepeat>{}, Number<KPack>{}),
+        make_tuple(
+            Number<KPack>{}, Number<KPack * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
+
+    // B[N0, N1, N2, KPack]
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor(
+        make_tuple(Number<NRepeat>{}, I2, Number<KRepeat>{}, Number<KPack>{}),
+        make_tuple(
+            Number<KPack>{}, Number<KPack * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck

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

@@ -37,7 +37,8 @@ template <index_t BlockSize,
           index_t NPerXDL,
           index_t MRepeat,
           index_t NRepeat,
-          index_t KPack>
+          index_t KPack,
+          bool TransposeC = false>
 struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
 {
     static constexpr auto I0 = Number<0>{};
@@ -59,7 +60,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
     static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
 
-    static constexpr auto xdlops_gemm = XdlopsGemm<FloatA, MPerXDL, NPerXDL, KPack, FloatB>{};
+    static constexpr auto xdlops_gemm =
+        XdlopsGemm<FloatA, MPerXDL, NPerXDL, KPack, FloatB, TransposeC>{};
 
     static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
 
@@ -185,6 +187,20 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
             make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
     }
 
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
+    }
+
     __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
     {
         constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
@@ -211,6 +227,20 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
         return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
     }
 
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
     __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
     {
         constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
@@ -300,33 +330,35 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
         auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
             b_thread_desc_.GetElementSpaceSize());
 
-        static_for<0, MRepeat, 1>{}([&](auto m0) {
-            // read A
-            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
-                               make_tuple(m0, I0, I0, I0),
-                               a_block_buf,
-                               a_thread_desc_,
-                               make_tuple(I0, I0, I0, I0),
-                               a_thread_buf);
+        static_for<0, KPerThread / KPack, 1>{}([&](auto k) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                // read A
+                a_thread_copy_.Run(
+                    a_block_desc_m0_m1_m2_k,
+                    make_tuple(m0, I0, I0, Number<(k.value * KPack)>{}),
+                    a_block_buf,
+                    a_thread_desc_,
+                    make_tuple(Number<(k.value % 2)>{}, Number<(m0.value % 2)>{}, I0, I0),
+                    a_thread_buf);
 
-            static_for<0, NRepeat, 1>{}([&](auto n0) {
-                // read B
-                b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
-                                   make_tuple(n0, I0, I0, I0),
-                                   b_block_buf,
-                                   b_thread_desc_,
-                                   make_tuple(I0, I0, I0, I0),
-                                   b_thread_buf);
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    // read B
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<(k.value * KPack)>{}),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(Number<(k.value % 2)>{}, n0, I0, I0),
+                                       b_thread_buf);
 
-                static_for<0, KPerThread, KPack>{}([&](auto k) {
                     vector_type<FloatA, KPack> a_thread_vec;
                     vector_type<FloatB, KPack> b_thread_vec;
 
                     static_for<0, KPack, 1>{}([&](auto i) {
-                        a_thread_vec.template AsType<FloatA>()(i) = a_thread_buf
-                            [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
+                        a_thread_vec.template AsType<FloatA>()(i) =
+                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                make_tuple(k % 2, m0 % 2, 0, i))>{}];
                         b_thread_vec.template AsType<FloatB>()(i) = b_thread_buf
-                            [Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(k % 2, n0, 0, i))>{}];
                     });
 
                     using mfma_input_type_a =
@@ -349,11 +381,11 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     protected:
     // A[M0, M1, M2, KPerThread]
     static constexpr auto a_thread_desc_ =
-        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+        make_naive_tensor_descriptor_packed(make_tuple(I2, I2, I1, Number<KPack>{}));
 
     // B[N0, N1, N2, KPerThread]
     static constexpr auto b_thread_desc_ =
-        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+        make_naive_tensor_descriptor_packed(make_tuple(I2, Number<NRepeat>{}, I1, Number<KPack>{}));
 
     // C[M, N, NumRegXdlops]
     static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
@@ -363,7 +395,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                                                          FloatA,
                                                          decltype(a_block_desc_m0_m1_m2_k),
                                                          decltype(a_thread_desc_),
-                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<1, 1, 1, KPack>,
                                                          Sequence<0, 1, 2, 3>,
                                                          3,
                                                          A_K1,
@@ -373,7 +405,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                                                          FloatB,
                                                          decltype(b_block_desc_n0_n1_n2_k),
                                                          decltype(b_thread_desc_),
-                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<1, 1, 1, KPack>,
                                                          Sequence<0, 1, 2, 3>,
                                                          3,
                                                          B_K1,
@@ -398,6 +430,7 @@ template <index_t BlockSize,
           index_t MRepeat,
           index_t NRepeat,
           index_t KPack,
+          bool TransposeC        = false,
           index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS>
 struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     : public BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
@@ -410,7 +443,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                                                                  NPerXDL,
                                                                  MRepeat,
                                                                  NRepeat,
-                                                                 KPack>
+                                                                 KPack,
+                                                                 TransposeC>
 {
     using Base = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
                                                                      FloatA,
@@ -422,7 +456,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
                                                                      NPerXDL,
                                                                      MRepeat,
                                                                      NRepeat,
-                                                                     KPack>;
+                                                                     KPack,
+                                                                     TransposeC>;
 
 #if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
     using Base::a_block_desc_m0_m1_m2_k;
@@ -586,7 +621,8 @@ template <index_t BlockSize,
           index_t MRepeat,
           index_t NRepeat,
           index_t KPack,
-          LoopScheduler LoopSched>
+          LoopScheduler LoopSched,
+          bool TransposeC = false>
 constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
 {
     if constexpr(LoopSched == LoopScheduler::Default)
@@ -601,7 +637,8 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
                                                                    NPerXDL,
                                                                    MRepeat,
                                                                    NRepeat,
-                                                                   KPack>{};
+                                                                   KPack,
+                                                                   TransposeC>{};
     }
     else if constexpr(LoopSched == LoopScheduler::Interwave)
     {
@@ -615,7 +652,8 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
                                                                             NPerXDL,
                                                                             MRepeat,
                                                                             NRepeat,
-                                                                            KPack>{};
+                                                                            KPack,
+                                                                            TransposeC>{};
     }
 };
 

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

@@ -12,7 +12,7 @@
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
-#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1_doublelds.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 
@@ -155,10 +155,15 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
             index_t gdx, gdy, gdz;
             std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N);
 
-            const auto K = GridwiseGemm::CalculateAK0(arg.K) * AK1;
-
             float ave_time = 0;
+#if 1
+            const auto kernel = kernel_gemm_xdl_cshuffle_v1<GridwiseGemm, true>;
 
+            ave_time = launch_and_time_kernel(
+                stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+#endif
+#if 0
+            const auto K = GridwiseGemm::CalculateAK0(arg.K) * AK1;
             if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
             {
                 const auto kernel = kernel_gemm_xdl_cshuffle_v1<GridwiseGemm, true>;
@@ -173,7 +178,7 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
                 ave_time = launch_and_time_kernel(
                     stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
             }
-
+#endif
             return ave_time;
         }
 

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp

@@ -9,7 +9,7 @@
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
 #include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
-#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.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"
@@ -22,7 +22,8 @@ __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_gemm_xdl_cshuffle_v1(typename GridwiseGemm::Argument karg)
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_gemm_xdl_cshuffle_v1(typename GridwiseGemm::Argument karg)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
     defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
@@ -357,6 +358,26 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         }
     }
 
+    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_MN0_MN1_MN2_K<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_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
     __host__ __device__ static auto
     MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
     {
@@ -810,25 +831,40 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
         //       register
         // sanity check
-        constexpr index_t KPack = math::max(
-            math::lcm(AK1Number, BK1Number),
-            MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>::selected_mfma.k_per_blk);
-
-        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1Number, BK1Number),
+                      MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        // auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+        //     BlockSize,
+        //     ComputeType,
+        //     FloatGemmAcc,
+        //     decltype(a_block_desc_ak0_m_ak1),
+        //     decltype(b_block_desc_bk0_n_bk1),
+        //     MPerXdl,
+        //     NPerXdl,
+        //     MXdlPerWave,
+        //     NXdlPerWave,
+        //     KPack,
+        //     LoopSched>();
+        auto blockwise_gemm_pipeline = BlockwiseGemmXdlops_pipeline_v3<
             BlockSize,
             ComputeTypeA,
-            ComputeTypeB,
             FloatGemmAcc,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
             MPerXdl,
             NPerXdl,
             MXdlPerWave,
             NXdlPerWave,
-            KPack,
-            LoopSched>();
+            KPack>{}; // TransposeC
 
-        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+        auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
 
         // LDS allocation for A and B: be careful of alignment
         constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
@@ -846,27 +882,26 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
 
         // gridwise GEMM pipeline
         static_assert(std::is_default_constructible_v<GridwiseGemmPipe>);
-        const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
+        // const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
 
         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);
 
-        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(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,
-                                                               blockwise_gemm,
-                                                               c_thread_buf,
-                                                               num_k_block_main_loop);
+        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop>(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
         {
@@ -879,12 +914,12 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
 
             // TODO: hacky, fix it!
             constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
-                blockwise_gemm.GetCThreadDescriptor_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.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+                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);
@@ -924,7 +959,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
             // calculate origin of thread output tensor on global memory
             //     blockwise GEMM c matrix starting index
             const auto c_thread_mtx_on_block =
-                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+                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];

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1_doublelds.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/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.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"
+
+namespace ck {
+
+template <typename GridwiseGemm, bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_gemm_xdl_cshuffle_v1(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, p_shared, karg);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename GridwiseGemm,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdl_cshuffle_v1(const FloatA* __restrict__ p_a_grid,
+                                    const FloatB* __restrict__ p_b_grid,
+                                    FloatC* __restrict__ p_c_grid,
+                                    typename GridwiseGemm::Problem problem)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid, p_b_grid, p_c_grid, p_shared, problem);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = problem;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename FloatA,
+          typename FloatB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t NumGemmKPrefetchStage,
+          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,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeTypeA       = FloatC,
+          typename ComputeTypeB       = ComputeTypeA>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0Number = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0Number = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1Number = Number<AK1Value>{};
+    static constexpr auto BK1Number = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, 1);
+    }
+
+    __host__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock) * MPerBlock;
+    }
+
+    __host__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock) * NPerBlock;
+    }
+
+    __host__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ static auto CalculateAK0(index_t K)
+    {
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding ||
+                     GemmSpec == GemmSpecialization::KPadding ||
+                     GemmSpec == GemmSpecialization::NKPadding)
+        {
+            return CalculateKPadded(K) / AK1Value;
+        }
+        else
+        {
+            return K / AK1Value;
+        }
+    }
+
+    __host__ static auto CalculateBK0(index_t K)
+    {
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding ||
+                     GemmSpec == GemmSpecialization::KPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding)
+        {
+            return CalculateKPadded(K) / BK1Value;
+        }
+        else
+        {
+            return K / BK1Value;
+        }
+    }
+
+    __host__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_floor(M, MPerBlock);
+    }
+
+    __host__ static auto CalculateNBlock(index_t N)
+    {
+        return math::integer_divide_floor(N, NPerBlock);
+    }
+
+    __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_MN0_MN1_MN2_K<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_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __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, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        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;
+        }
+    }
+
+    struct Problem
+    {
+        __host__ Problem(index_t M_,
+                         index_t N_,
+                         index_t K_,
+                         index_t StrideA_,
+                         index_t StrideB_,
+                         index_t StrideC_)
+            : M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideC{StrideC_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              KPadded{CalculateKPadded(K_)},
+              AK0{CalculateAK0(K_)},
+              BK0{CalculateBK0(K_)},
+              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 << ", "
+                      << "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;
+        index_t StrideC;
+        index_t MPadded;
+        index_t NPadded;
+        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 FloatA* p_a_grid_,
+                          const FloatB* p_b_grid_,
+                          FloatC* p_c_grid_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          index_t StrideC_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_c_grid{p_c_grid_}
+        {
+        }
+
+        const FloatA* p_a_grid;
+        const FloatB* p_b_grid;
+        FloatC* p_c_grid;
+    };
+
+    // FIXME: pass GridwiseGemmPipe as a template arguement into GridwiseGemm
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1Number, AK1Number, I1));
+    }
+
+    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1Number, BK1Number, I1));
+    }
+
+    __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;
+    }
+
+    __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(2 * (a_block_space_size_aligned * sizeof(ComputeTypeA) +
+                              b_block_space_size_aligned * sizeof(ComputeTypeB)),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ static constexpr bool CheckValidity(const Problem& problem)
+    {
+        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(!(problem.M % MPerBlock == 0))
+            {
+                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(!(problem.N % NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding)
+        {
+            if(!(CalculateKPadded(problem.K) % AK1Value == 0) ||
+               !(CalculateKPadded(problem.K) % BK1Value == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(problem.K % AK1Value == 0) || !(problem.K % BK1Value == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(problem.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(problem.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(problem.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                return false;
+            }
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = (CalculateAK0(problem.K) * AK1Value) / KPerBlock;
+
+        if(num_k_loop < 3 || (num_k_loop - 3) % 2 != 0)
+        {
+            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 num_loop > 3;
+    }
+
+    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
+    using Block2CTileMap = BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock>;
+
+    template <bool HasMainKBlockLoop>
+    __device__ static void Run(const FloatA* __restrict__ p_a_grid,
+                               const FloatB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const Problem& problem)
+    {
+        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(
+            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());
+
+        const AElementwiseOperation a_element_op{};
+        const BElementwiseOperation b_element_op{};
+        const CElementwiseOperation c_element_op{};
+
+        // divide block work by [M, N]
+        const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N};
+
+        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;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // 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,
+                                                FloatA,
+                                                ComputeTypeA,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                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,
+                                                FloatB,
+                                                ComputeTypeB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                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{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1Number, BK1Number),
+                      MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        // auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+        //     BlockSize,
+        //     ComputeType,
+        //     FloatGemmAcc,
+        //     decltype(a_block_desc_ak0_m_ak1),
+        //     decltype(b_block_desc_bk0_n_bk1),
+        //     MPerXdl,
+        //     NPerXdl,
+        //     MXdlPerWave,
+        //     NXdlPerWave,
+        //     KPack,
+        //     LoopSched>();
+        auto blockwise_gemm_pipeline = BlockwiseGemmXdlops_pipeline_v4<
+            BlockSize,
+            ComputeTypeA,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack>{}; // TransposeC
+
+        auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
+
+        // 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);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeA*>(p_shared) + a_block_space_size_aligned +
+                b_block_space_size_aligned,
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeB*>(p_shared) + a_block_space_size_aligned +
+                a_block_space_size_aligned + b_block_space_size_aligned,
+            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);
+
+        // gridwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<GridwiseGemmPipe>);
+        // const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
+
+        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>(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<FloatCShuffle*>(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<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   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{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 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>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.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
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1_transpose.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/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.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"
+
+namespace ck {
+
+template <typename GridwiseGemm, bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdl_cshuffle_v1(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, p_shared, karg);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename GridwiseGemm,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdl_cshuffle_v1(const FloatA* __restrict__ p_a_grid,
+                                    const FloatB* __restrict__ p_b_grid,
+                                    FloatC* __restrict__ p_c_grid,
+                                    typename GridwiseGemm::Problem problem)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid, p_b_grid, p_c_grid, p_shared, problem);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = problem;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename FloatA,
+          typename FloatB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t NumGemmKPrefetchStage,
+          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,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeTypeA       = FloatC,
+          typename ComputeTypeB       = ComputeTypeA>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0Number = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0Number = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1Number = Number<AK1Value>{};
+    static constexpr auto BK1Number = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, 1);
+    }
+
+    __host__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock) * MPerBlock;
+    }
+
+    __host__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock) * NPerBlock;
+    }
+
+    __host__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ static auto CalculateAK0(index_t K)
+    {
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding ||
+                     GemmSpec == GemmSpecialization::KPadding ||
+                     GemmSpec == GemmSpecialization::NKPadding)
+        {
+            return CalculateKPadded(K) / AK1Value;
+        }
+        else
+        {
+            return K / AK1Value;
+        }
+    }
+
+    __host__ static auto CalculateBK0(index_t K)
+    {
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding ||
+                     GemmSpec == GemmSpecialization::KPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding)
+        {
+            return CalculateKPadded(K) / BK1Value;
+        }
+        else
+        {
+            return K / BK1Value;
+        }
+    }
+
+    __host__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_floor(M, MPerBlock);
+    }
+
+    __host__ static auto CalculateNBlock(index_t N)
+    {
+        return math::integer_divide_floor(N, NPerBlock);
+    }
+
+    __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;
+        }
+    }
+
+    __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, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        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;
+        }
+    }
+
+    struct Problem
+    {
+        __host__ Problem(index_t M_,
+                         index_t N_,
+                         index_t K_,
+                         index_t StrideA_,
+                         index_t StrideB_,
+                         index_t StrideC_)
+            : M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideC{StrideC_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              KPadded{CalculateKPadded(K_)},
+              AK0{CalculateAK0(K_)},
+              BK0{CalculateBK0(K_)},
+              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 << ", "
+                      << "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;
+        index_t StrideC;
+        index_t MPadded;
+        index_t NPadded;
+        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 FloatA* p_a_grid_,
+                          const FloatB* p_b_grid_,
+                          FloatC* p_c_grid_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          index_t StrideC_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_c_grid{p_c_grid_}
+        {
+        }
+
+        const FloatA* p_a_grid;
+        const FloatB* p_b_grid;
+        FloatC* p_c_grid;
+    };
+
+    // FIXME: pass GridwiseGemmPipe as a template arguement into GridwiseGemm
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1Number, AK1Number, I1));
+    }
+
+    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1Number, BK1Number, I1));
+    }
+
+    __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;
+    }
+
+    __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(ComputeTypeA) +
+                          b_block_space_size_aligned * sizeof(ComputeTypeB)),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ static constexpr bool CheckValidity(const Problem& problem)
+    {
+        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(!(problem.M % MPerBlock == 0))
+            {
+                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(!(problem.N % NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding)
+        {
+            if(!(CalculateKPadded(problem.K) % AK1Value == 0) ||
+               !(CalculateKPadded(problem.K) % BK1Value == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(problem.K % AK1Value == 0) || !(problem.K % BK1Value == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(problem.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(problem.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(problem.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                return false;
+            }
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = (CalculateAK0(problem.K) * AK1Value) / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            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 GridwiseGemmPipe::CalculateHasMainLoop(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
+    using Block2CTileMap = BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock>;
+
+    template <bool HasMainKBlockLoop>
+    __device__ static void Run(const FloatA* __restrict__ p_a_grid,
+                               const FloatB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const Problem& problem)
+    {
+        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(
+            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());
+
+        const AElementwiseOperation a_element_op{};
+        const BElementwiseOperation b_element_op{};
+        const CElementwiseOperation c_element_op{};
+
+        // divide block work by [M, N]
+        const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N};
+
+        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;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // 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,
+                                                FloatA,
+                                                ComputeTypeA,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                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,
+                                                FloatB,
+                                                ComputeTypeB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                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{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1Number, BK1Number),
+            MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ComputeTypeA,
+            ComputeTypeB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched,
+            true>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // 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 = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeB*>(p_shared) + a_block_space_size_aligned,
+            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);
+
+        // gridwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<GridwiseGemmPipe>);
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
+
+        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);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(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,
+                                                               blockwise_gemm,
+                                                               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_n2_n3_n4 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
+            constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
+            constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_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<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = 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 = MPerXdl
+                        )),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2,
+                        N3,
+                        N4))), // N2 * N3 * N4 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 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.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_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_n3_n4_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<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            I1,
+                                                            N2,
+                                                            I1,
+                                                            N4>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   N4,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                    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],
+                                     n_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I3],
+                                     n_thread_data_on_block_idx[I4]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, 1, N2, 1, N4>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           1,
+                                           N2,
+                                           1,
+                                           N4>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.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_n2_n3_n4,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              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
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck

script/clang-format-overwrite.sh

-#find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' -o -iname '*.inc' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
-git status --porcelain | awk '$1 != "D" && (match($2, "\\.cpp|hpp|inc")) {print $2}' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
+find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' -o -iname '*.inc' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
+# git status --porcelain | awk '$1 != "D" && (match($2, "\\.cpp|hpp|inc")) {print $2}' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'

test/transpose/test_transpose_ut_cases.inc

@@ -14,7 +14,6 @@ TYPED_TEST(TestTranspose, Test1)
     this->Run();
 }
 
-
 TYPED_TEST(TestTranpose, Test2)
 {
     std::vector<int> Ms{127, 255, 312, 799, 1573};
@@ -27,4 +26,3 @@ TYPED_TEST(TestTranpose, Test2)
 
     this->Run();
 }
-