Added gemm_fp8xint4_Bpreshuffle files, function not checked yet

example/01_gemm/gemm_xdl_fp8_pk_i4_bpreshuffle_v3.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
-#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_preshuffle.hpp"
 
 using F8 = ck::f8_t;
 using I4  = ck::pk_i4_t;
@@ -63,7 +63,7 @@ static constexpr ck::index_t KPerBlock = 128;
 
 // clang-format off
 using DeviceGemmV2Instance = 
-    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3_BPreshuffle<
         ALayout,   BLayout,  CLayout,   
         ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, 
         AElementOp, BElementOp, CElementOp, GemmDefault, 
@@ -77,7 +77,7 @@ using DeviceGemmV2Instance =
         S<4, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
         2, 32, 32, 0,
         1, 1, S<1, 16, 1, 8>, 4,
-        ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ADataType, ADataType, PermuteA, PermuteB>;
+        ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, F8, F8, PermuteA, PermuteB>;
 
 // clang-format on
 
@@ -174,8 +174,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
     DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n_permute.mDesc.GetElementSpaceSize());
     DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
 
-    // int NperXdl= 16;
-    // preShuffleBuffer(b_k_n.mData.data(), b_k_n_preshuffled.mData.data(), N, K, NperXdl);
+    int NperXdl = GetPreShuffleParameters;
+    preShuffleBuffer(b_k_n.mData.data(), b_k_n_preshuffled.mData.data(), N, K, NperXdl);
 
     // weight permute
     if constexpr(PermuteB)
@@ -190,7 +190,7 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
             {
                 for(int jj = 0; jj < K1; jj++)
                 {
-                    b_k_n_permute(j * N * K1 + i * K1 + jj) = b_k_n(i * K + (j * K1 + jj));
+                    b_k_n_permute(j * N * K1 + i * K1 + jj) = b_k_n_preshuffled(i * K + (j * K1 + jj));
                 }
             }
         }
@@ -201,7 +201,7 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
         {
             for(int j = 0; j < K; j++)
             {
-                b_k_n_permute(i * K + j) = b_k_n(i * K + j);
+                b_k_n_permute(i * K + j) = b_k_n_preshuffled(i * K + j);
             }
         }
     }

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp"
+
+namespace ck {
+
+template <BlockGemmPipelineVersion BlkGemmPipelineVer,
+          BlockGemmPipelineScheduler BlkGemmPipeSche,
+          index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+constexpr auto BlockGemmBPreshufflePipeline_Selector()
+{
+    if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+    {
+        return BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlkGemmPipeSche,
+                                                           BlockSize,
+                                                           ADataType,
+                                                           BDataType,
+                                                           ComputeDataType,
+                                                           AccDataType,
+                                                           ATileDesc,
+                                                           BTileDesc,
+                                                           AMmaTileDesc,
+                                                           BMmaTileDesc,
+                                                           ABlockTransferSrcScalarPerVector,
+                                                           BBlockTransferSrcScalarPerVector,
+                                                           MPerBlock,
+                                                           NPerBlock,
+                                                           KPerBlock,
+                                                           MPerXDL,
+                                                           NPerXDL,
+                                                           MRepeat,
+                                                           NRepeat,
+                                                           KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
+    {
+        return BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlkGemmPipeSche,
+                                                           BlockSize,
+                                                           ADataType,
+                                                           BDataType,
+                                                           ComputeDataType,
+                                                           AccDataType,
+                                                           ATileDesc,
+                                                           BTileDesc,
+                                                           AMmaTileDesc,
+                                                           BMmaTileDesc,
+                                                           ABlockTransferSrcScalarPerVector,
+                                                           BBlockTransferSrcScalarPerVector,
+                                                           MPerBlock,
+                                                           NPerBlock,
+                                                           KPerBlock,
+                                                           MPerXDL,
+                                                           NPerXDL,
+                                                           MRepeat,
+                                                           NRepeat,
+                                                           KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+    {
+        return BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlkGemmPipeSche,
+                                                           BlockSize,
+                                                           ADataType,
+                                                           BDataType,
+                                                           ComputeDataType,
+                                                           AccDataType,
+                                                           ATileDesc,
+                                                           BTileDesc,
+                                                           AMmaTileDesc,
+                                                           BMmaTileDesc,
+                                                           ABlockTransferSrcScalarPerVector,
+                                                           BBlockTransferSrcScalarPerVector,
+                                                           MPerBlock,
+                                                           NPerBlock,
+                                                           KPerBlock,
+                                                           MPerXDL,
+                                                           NPerXDL,
+                                                           MRepeat,
+                                                           NRepeat,
+                                                           KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
+    {
+        return BlockwiseGemmXdlops_pipeline_v4<BlkGemmPipeSche,
+                                               BlockSize,
+                                               ADataType,
+                                               BDataType,
+                                               ComputeDataType,
+                                               AccDataType,
+                                               ATileDesc,
+                                               BTileDesc,
+                                               AMmaTileDesc,
+                                               BMmaTileDesc,
+                                               ABlockTransferSrcScalarPerVector,
+                                               BBlockTransferSrcScalarPerVector,
+                                               MPerBlock,
+                                               NPerBlock,
+                                               KPerBlock,
+                                               MPerXDL,
+                                               NPerXDL,
+                                               MRepeat,
+                                               NRepeat,
+                                               KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v5)
+    {
+        return BlockwiseGemmXdlops_pipeline_v5<BlkGemmPipeSche,
+                                               BlockSize,
+                                               ADataType,
+                                               BDataType,
+                                               ComputeDataType,
+                                               AccDataType,
+                                               ATileDesc,
+                                               BTileDesc,
+                                               AMmaTileDesc,
+                                               BMmaTileDesc,
+                                               ABlockTransferSrcScalarPerVector,
+                                               BBlockTransferSrcScalarPerVector,
+                                               MPerBlock,
+                                               NPerBlock,
+                                               KPerBlock,
+                                               MPerXDL,
+                                               NPerXDL,
+                                               MRepeat,
+                                               NRepeat,
+                                               KPack>{};
+    }
+    else
+    {
+        std::cerr << "BlockGemmPipeline configuration is not available" << std::endl;
+    }
+}
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Compute optimized pipeline
+// GlobalPrefetchStages: 2
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 1
+// LocalSharedMemoryBuffer: 1
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPacks>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1
+{
+};
+
+template <index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack
+          // ,bool TransposeC //disable transposec right now...
+          >
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::Intrawave,
+                                                   BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>
+    : BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                        ADataType,
+                                        BDataType,
+                                        ComputeDataType,
+                                        AccDataType,
+                                        ATileDesc,
+                                        BTileDesc,
+                                        AMmaTileDesc,
+                                        BMmaTileDesc,
+                                        ABlockTransferSrcScalarPerVector,
+                                        BBlockTransferSrcScalarPerVector,
+                                        MPerBlock,
+                                        NPerBlock,
+                                        KPerBlock,
+                                        MPerXDL,
+                                        NPerXDL,
+                                        MRepeat,
+                                        NRepeat,
+                                        KPack>
+
+{
+    using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>;
+    using Base::A_K1;
+    using Base::B_K1;
+    using Base::I0;
+    using Base::I1;
+    using Base::KRepeat;
+    using Base::xdlops_gemm;
+    using typename Base::HotLoopInstList;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::CalculateCThreadOriginDataIndex8D;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+
+    static constexpr index_t PrefetchStages  = 2;
+    static constexpr index_t PrefillStages   = 1;
+    static constexpr index_t GlobalBufferNum = 2;
+
+    template <typename TileDesc_M0_M1_M2_K>
+    __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
+    {
+        constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
+        constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
+        constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
+        constexpr index_t K2 = KPack;
+        constexpr index_t K1 = 64 / NPerXDL;
+        constexpr index_t K0 = KRepeat;
+
+        return transform_tensor_descriptor(
+            TileDesc_M0_M1_M2_K{},
+            make_tuple(
+                make_pass_through_transform(Number<M0>{}),
+                make_pass_through_transform(Number<M1>{}),
+                make_pass_through_transform(Number<M2>{}),
+                make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
+        MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
+
+    __host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
+    }
+
+    __device__ static constexpr auto HotLoopScheduler()
+    {
+        constexpr auto num_ds_read_inst_a     = HotLoopInstList::A_LDS_Read_Inst_Num;
+        constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
+        constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
+
+        // B global
+        static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+        });
+
+        // A global
+        static_for<0, num_buffer_load_inst_a, 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(0x020, 1, 0); // VMEM read
+        });
+
+        // A local
+        static_for<0, num_ds_read_inst_a / 2, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read
+        });
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              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,
+                        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
+    {
+        ignore = b_block_buf;
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+        constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
+
+        // Global prefetch A1 B1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        b_blockwise_copy.Run(b_grid_desc,
+                             b_grid_buf,
+                             b_block_desc_n0_n1_k0_k1,
+                             b_block_origin_idx,
+                             b_thread_bufs(I0));
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+        __builtin_amdgcn_sched_barrier(0);
+
+        // // Local prefill A1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
+
+        // // Global prefetch A2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Local prefetch A1
+        block_sync_lds();
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                                   make_tuple(m0, I0, I0, k0, I0, I0),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, I0, k0, I0, I0),
+                                   a_thread_buf);
+            });
+        });
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
+                    b_blockwise_copy.Run(b_grid_desc,
+                                         b_grid_buf,
+                                         b_block_desc_n0_n1_k0_k1,
+                                         b_block_origin_idx,
+                                         b_thread_bufs(local_read_buf));
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    block_sync_lds();
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, mfma_reg_buf);
+
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, local_read_buf);
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                vector_type<ComputeDataType, KPack> a_thread_vec;
+                                vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                                static_for<0, KPack, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                            make_tuple(m0, I0, I0, k0, I0, ik))>{}];
+                                    b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        b_thread_bufs[mfma_reg_buf]
+                                                     [Number<b_thread_desc_.CalculateOffset(
+                                                         make_tuple(n0, I0, k0, ik))>{}];
+                                });
+
+                                using mfma_input_type =
+                                    typename vector_type<ComputeDataType,
+                                                         xdlops_gemm.K1PerXdlops>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                xdlops_gemm.Run(
+                                    a_thread_vec.template AsType<mfma_input_type>(),
+                                    b_thread_vec.template AsType<mfma_input_type>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                            });
+                        });
+                    });
+
+                    block_sync_lds();
+
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                                               make_tuple(m0, I0, I0, k0, I0, I0),
+                                               a_block_buf,
+                                               a_thread_desc_,
+                                               make_tuple(m0, I0, I0, k0, I0, I0),
+                                               a_thread_buf);
+                        });
+                    });
+
+                    HotLoopScheduler();
+                    __builtin_amdgcn_sched_barrier(0);
+                };
+
+                LoopFunc(I0, I1);
+                LoopFunc(I1, I0);
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+        // tail
+        if constexpr(TailNum == TailNumber::Even)
+        {
+            b_blockwise_copy.Run(b_grid_desc,
+                                 b_grid_buf,
+                                 b_block_desc_n0_n1_k0_k1,
+                                 b_block_origin_idx,
+                                 b_thread_bufs(I1));
+
+            block_sync_lds();
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            block_sync_lds();
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                                       make_tuple(m0, I0, I0, k0, I0, I0),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, I0, k0, I0, I0),
+                                       a_thread_buf);
+                });
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+            // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle
+            // latency
+            // __builtin_amdgcn_sched_barrier(0);
+        }
+        else
+        {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+        }
+    }
+
+    protected:
+    // MRepeat MWave MLane KRepeat KLane KPack
+    // KRepeat -> MRepeat-> Mwave->KLane->MLane->KPack
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, I1, I1, Number<KRepeat>{}, I1, Number<KPack>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
+                                                         ComputeDataType,
+                                                         decltype(a_block_desc_m0_m1_m2_k0_k1_k2),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3, 4, 5>,
+                                                         5,
+                                                         A_K1,
+                                                         A_K1>;
+
+    AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex6D()};
+
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
+
+    static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
+
+    using Base::c_thread_desc_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Compute optimized pipeline
+// GlobalPrefetchStages: 3
+// LocalPreFillStages: 2
+// LocalPreFetchStages: 2
+// LocalSharedMemoryBuffer: 2
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPacks>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2
+{
+};
+
+template <index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack
+          // ,bool TransposeC //disable transposec right now...
+          >
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::Intrawave,
+                                                   BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>
+    : BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                        ADataType,
+                                        BDataType,
+                                        ComputeDataType,
+                                        AccDataType,
+                                        ATileDesc,
+                                        BTileDesc,
+                                        AMmaTileDesc,
+                                        BMmaTileDesc,
+                                        ABlockTransferSrcScalarPerVector,
+                                        BBlockTransferSrcScalarPerVector,
+                                        MPerBlock,
+                                        NPerBlock,
+                                        KPerBlock,
+                                        MPerXDL,
+                                        NPerXDL,
+                                        MRepeat,
+                                        NRepeat,
+                                        KPack>
+
+{
+    using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>;
+    using Base::A_K1;
+    using Base::B_K1;
+    using Base::I0;
+    using Base::I1;
+    using Base::KRepeat;
+    using Base::xdlops_gemm;
+    using typename Base::HotLoopInstList;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::CalculateCThreadOriginDataIndex8D;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+
+    static constexpr index_t PrefetchStages  = 3;
+    static constexpr index_t PrefillStages   = 2;
+    static constexpr index_t GlobalBufferNum = 2;
+
+    template <typename TileDesc_M0_M1_M2_K>
+    __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
+    {
+        constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
+        constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
+        constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
+        constexpr index_t K2 = KPack;
+        constexpr index_t K1 = 64 / NPerXDL;
+        constexpr index_t K0 = KRepeat;
+
+        return transform_tensor_descriptor(
+            TileDesc_M0_M1_M2_K{},
+            make_tuple(
+                make_pass_through_transform(Number<M0>{}),
+                make_pass_through_transform(Number<M1>{}),
+                make_pass_through_transform(Number<M2>{}),
+                make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
+        MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
+
+    __host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+
+        return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
+    }
+
+    __device__ static constexpr auto HotLoopScheduler()
+    {
+        // constexpr auto num_ds_read_inst_a     = HotLoopInstList::A_LDS_Read_Inst_Num;
+        constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
+        constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
+
+        // B global + A local
+        static_for<0, num_buffer_load_inst_b / 2, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read B
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read   A
+        });
+
+        static_for<0, num_buffer_load_inst_b / 2, 1>{}([&](auto i) {
+            ignore = i;
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read B
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read   A
+        });
+
+        // A global
+        static_for<0, num_buffer_load_inst_a, 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(0x020, 1, 0); // VMEM read
+        });
+
+        // A local
+        // static_for<0, num_ds_read_inst_a / 2, 1>{}([&](auto i) {
+        //     ignore = i;
+        //     __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+        //     __builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read
+        // });
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              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,
+                        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
+    {
+        ignore = b_block_buf;
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+        constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
+
+        // Global prefetch A1, B1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        b_blockwise_copy.Run(b_grid_desc,
+                             b_grid_buf,
+                             b_block_desc_n0_n1_k0_k1,
+                             b_block_origin_idx,
+                             b_thread_bufs(I0));
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Local prefill A1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0), I0);
+
+        // Global prefetch A2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Local prefetch A1
+        block_sync_lds();
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                                   make_tuple(m0, I0, I0, k0, I0, I0),
+                                   a_block_buf.At(I0),
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, I0, k0, I0, I0),
+                                   a_thread_bufs(I0));
+            });
+        });
+
+        // Local prefill A2
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1), I1);
+
+        // // Global prefetch A3
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
+                    block_sync_lds();
+
+                    b_blockwise_copy.Run(b_grid_desc,
+                                         b_grid_buf,
+                                         b_block_desc_n0_n1_k0_k1,
+                                         b_block_origin_idx,
+                                         b_thread_bufs(local_read_buf));
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                                               make_tuple(m0, I0, I0, k0, I0, I0),
+                                               a_block_buf.At(local_read_buf),
+                                               a_thread_desc_,
+                                               make_tuple(m0, I0, I0, k0, I0, I0),
+                                               a_thread_bufs(local_read_buf));
+                        });
+                    });
+
+                    a_blockwise_copy.RunWrite(
+                        a_block_desc, a_block_buf.At(mfma_reg_buf), mfma_reg_buf);
+
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, local_read_buf);
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                vector_type<ComputeDataType, KPack> a_thread_vec;
+                                vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                                static_for<0, KPack, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        a_thread_bufs[mfma_reg_buf]
+                                                     [Number<a_thread_desc_.CalculateOffset(
+                                                         make_tuple(m0, I0, I0, k0, I0, ik))>{}];
+                                    b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        b_thread_bufs[mfma_reg_buf]
+                                                     [Number<b_thread_desc_.CalculateOffset(
+                                                         make_tuple(n0, I0, k0, ik))>{}];
+                                });
+
+                                using mfma_input_type =
+                                    typename vector_type<ComputeDataType,
+                                                         xdlops_gemm.K1PerXdlops>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                xdlops_gemm.Run(
+                                    a_thread_vec.template AsType<mfma_input_type>(),
+                                    b_thread_vec.template AsType<mfma_input_type>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                            });
+                        });
+                    });
+
+                    HotLoopScheduler();
+                    __builtin_amdgcn_sched_barrier(0);
+                };
+
+                LoopFunc(I0, I1);
+                LoopFunc(I1, I0);
+
+                i += 2;
+            } while(i < (num_loop - 3));
+        }
+        // tail
+
+        auto ReadWriteCompFunc = [&](auto mfma_reg, auto local_read_reg) {
+            block_sync_lds();
+
+            b_blockwise_copy.Run(b_grid_desc,
+                                 b_grid_buf,
+                                 b_block_desc_n0_n1_k0_k1,
+                                 b_block_origin_idx,
+                                 b_thread_bufs(local_read_reg));
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                                       make_tuple(m0, I0, I0, k0, I0, I0),
+                                       a_block_buf.At(local_read_reg),
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, I0, k0, I0, I0),
+                                       a_thread_bufs(local_read_reg));
+                });
+            });
+
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(mfma_reg), mfma_reg);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_bufs[mfma_reg][Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[mfma_reg][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            HotLoopScheduler();
+            __builtin_amdgcn_sched_barrier(0);
+        };
+
+        auto ReadCompFunc = [&](auto mfma_reg, auto local_read_reg) {
+            block_sync_lds();
+
+            b_blockwise_copy.Run(b_grid_desc,
+                                 b_grid_buf,
+                                 b_block_desc_n0_n1_k0_k1,
+                                 b_block_origin_idx,
+                                 b_thread_bufs(local_read_reg));
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                                       make_tuple(m0, I0, I0, k0, I0, I0),
+                                       a_block_buf.At(local_read_reg),
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, I0, k0, I0, I0),
+                                       a_thread_bufs(local_read_reg));
+                });
+            });
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_bufs[mfma_reg][Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[mfma_reg][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            HotLoopScheduler();
+            __builtin_amdgcn_sched_barrier(0);
+        };
+
+        auto CompFunc = [&](auto mfma_reg) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_bufs[mfma_reg][Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[mfma_reg][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+        };
+
+        if constexpr(TailNum == TailNumber::Even)
+        {
+            ReadCompFunc(I0, I1);
+            CompFunc(I1);
+        }
+        else if constexpr(TailNum == TailNumber::Odd)
+        {
+            ReadWriteCompFunc(I0, I1);
+            ReadCompFunc(I1, I0);
+            CompFunc(I0);
+        }
+    }
+
+    protected:
+    // MRepeat MWave MLane KRepeat KLane KPack
+    // KRepeat -> MRepeat-> Mwave->KLane->MLane->KPack
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, I1, I1, Number<KRepeat>{}, I1, Number<KPack>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
+                                                         ComputeDataType,
+                                                         decltype(a_block_desc_m0_m1_m2_k0_k1_k2),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3, 4, 5>,
+                                                         5,
+                                                         A_K1,
+                                                         A_K1>;
+
+    AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex6D()};
+
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
+
+    static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
+
+    using Base::c_thread_desc_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Compute optimized pipeline
+// GlobalPrefetchStages: 2
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 1
+// LocalSharedMemoryBuffer: 1
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPacks>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3
+{
+};
+
+template <index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack
+          // ,bool TransposeC //disable transposec right now...
+          >
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::Intrawave,
+                                                   BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>
+    : BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                        ADataType,
+                                        BDataType,
+                                        ComputeDataType,
+                                        AccDataType,
+                                        ATileDesc,
+                                        BTileDesc,
+                                        AMmaTileDesc,
+                                        BMmaTileDesc,
+                                        ABlockTransferSrcScalarPerVector,
+                                        BBlockTransferSrcScalarPerVector,
+                                        MPerBlock,
+                                        NPerBlock,
+                                        KPerBlock,
+                                        MPerXDL,
+                                        NPerXDL,
+                                        MRepeat,
+                                        NRepeat,
+                                        KPack>
+
+{
+    using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>;
+    using Base::A_K1;
+    using Base::B_K1;
+    using Base::I0;
+    using Base::I1;
+    using Base::I2;
+    using Base::KRepeat;
+    using Base::xdlops_gemm;
+    using typename Base::HotLoopInstList;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::CalculateCThreadOriginDataIndex8D;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+
+    using Base::MWaves;
+
+    static constexpr index_t PrefetchStages        = 2;
+    static constexpr index_t PrefillStages         = 1;
+    static constexpr index_t GlobalBufferNum       = 1;
+    static constexpr index_t HotloopLocalBufSwitch = MRepeat % 2 == 0 ? 0 : 1;
+
+    template <typename TileDesc_M0_M1_M2_K>
+    __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
+    {
+        constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
+        constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
+        constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
+        constexpr index_t K2 = KPack;
+        constexpr index_t K1 = 64 / NPerXDL;
+        constexpr index_t K0 = KRepeat;
+
+        return transform_tensor_descriptor(
+            TileDesc_M0_M1_M2_K{},
+            make_tuple(
+                make_pass_through_transform(Number<M0>{}),
+                make_pass_through_transform(Number<M1>{}),
+                make_pass_through_transform(Number<M2>{}),
+                make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
+        MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
+
+    __host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
+    }
+
+    template <typename Stage>
+    __device__ static constexpr auto HotLoopScheduler(Stage stage)
+    {
+        constexpr auto num_ds_read_inst_a     = HotLoopInstList::A_LDS_Read_Inst_Num;
+        constexpr auto num_ds_write_inst_a    = HotLoopInstList::A_LDS_Write_Inst_Num;
+        constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
+        constexpr auto num_buffer_load_inst_b = MWaves * HotLoopInstList::B_Buffer_Load_Inst_Num;
+
+        constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num;
+
+        constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat;
+        constexpr auto staged_num_mfma           = num_mfma / MRepeat;
+
+        constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a;
+
+        if constexpr(stage.value == 0)
+        {
+            constexpr auto staged_num_buffer_load_b_per_ds_read_a =
+                num_buffer_load_inst_b / staged_num_ds_read_inst_a;
+            constexpr auto staged_num_mfma_per_buffer_load_b =
+                staged_num_mfma / num_buffer_load_inst_b;
+            // B global
+            static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+                ignore = i_inst;
+
+                static_for<0, staged_num_buffer_load_b_per_ds_read_a - 1, 1>{}([&](auto ibuf_inst) {
+                    ignore = ibuf_inst;
+                    __builtin_amdgcn_sched_group_barrier(
+                        0x008, staged_num_mfma_per_buffer_load_b, 0);  // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM 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, staged_num_mfma_per_buffer_load_b - 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0);    // VMEM read
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+        }
+        else if constexpr(stage.value == 1)
+        {
+            constexpr auto staged_num_mfma_per_ds_write_a =
+                math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a);
+
+            constexpr auto stage_more_mfma =
+                staged_num_mfma - (staged_num_mfma_per_ds_write_a - 1) * num_ds_write_inst_a;
+
+            // A local write
+            static_for<0, num_ds_write_inst_a, 1>{}([&](auto i_inst) {
+                if constexpr(i_inst.value < stage_more_mfma)
+                {
+                    if(i_inst.value < staged_num_ds_read_inst_a)
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a - 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
+                    }
+                    else
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a, 0);     // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                    }
+                }
+                else
+                {
+                    if(i_inst.value < staged_num_ds_read_inst_a)
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a - 2, 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
+                    }
+                    else
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                    }
+                }
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+        }
+        else if constexpr(stage.value == 2)
+        {
+            constexpr auto staged_num_mfma_per_buffer_load_a =
+                math::integer_divide_ceil(staged_num_mfma, num_buffer_load_inst_a);
+
+            constexpr auto stage_more_mfma =
+                staged_num_mfma - (staged_num_mfma_per_buffer_load_a - 1) * num_buffer_load_inst_a;
+
+            // A global
+            static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i_inst) {
+                if constexpr(i_inst.value < stage_more_mfma)
+                {
+                    if(i_inst.value < staged_num_ds_read_inst_a)
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_buffer_load_a - 1, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x020, 1, 0);    // VMEM read
+                        __builtin_amdgcn_sched_group_barrier(0x008, 1, 0);    // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x100, 1, 0);    // DS read
+                    }
+                    else
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_buffer_load_a, 0);  // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                    }
+                }
+                else
+                {
+                    if(i_inst.value < staged_num_ds_read_inst_a)
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_buffer_load_a - 2, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x020, 1, 0);    // VMEM read
+                        __builtin_amdgcn_sched_group_barrier(0x008, 1, 0);    // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x100, 1, 0);    // DS read
+                    }
+                    else
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_buffer_load_a - 1, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x020, 1, 0);    // VMEM read
+                    }
+                }
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+        }
+        else
+        {
+            // A local Read
+            static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+                ignore = i_inst;
+                __builtin_amdgcn_sched_group_barrier(
+                    0x008, staged_num_mfma_per_ds_read_a, 0);      // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+        }
+    }
+
+    template <typename Stage>
+    __device__ static constexpr auto EpilogueScheduler_1(Stage stage)
+    {
+        constexpr auto num_ds_read_inst_a     = HotLoopInstList::A_LDS_Read_Inst_Num;
+        constexpr auto num_ds_write_inst_a    = HotLoopInstList::A_LDS_Write_Inst_Num;
+        constexpr auto num_buffer_load_inst_b = MWaves * HotLoopInstList::B_Buffer_Load_Inst_Num;
+
+        constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num;
+
+        constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat;
+        constexpr auto staged_num_mfma           = num_mfma / MRepeat;
+
+        constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a;
+
+        if constexpr(stage.value == 0)
+        {
+            constexpr auto staged_num_buffer_load_b_per_ds_read_a =
+                num_buffer_load_inst_b / staged_num_ds_read_inst_a;
+            constexpr auto staged_num_mfma_per_buffer_load_b =
+                staged_num_mfma / num_buffer_load_inst_b;
+            // B global
+            static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+                ignore = i_inst;
+
+                static_for<0, staged_num_buffer_load_b_per_ds_read_a, 1>{}([&](auto ibuf_inst) {
+                    ignore = ibuf_inst;
+                    __builtin_amdgcn_sched_group_barrier(
+                        0x008, staged_num_mfma_per_buffer_load_b, 0);  // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM 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, staged_num_mfma_per_buffer_load_b - 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0);    // VMEM read
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+        }
+        else if constexpr(stage.value == 1)
+        {
+#if 0
+            constexpr auto staged_num_ds_write_a_per_ds_read_a =
+                num_ds_write_inst_a / staged_num_ds_read_inst_a;
+            constexpr auto staged_num_mfma_per_ds_write_a = staged_num_mfma / num_ds_write_inst_a;
+            // A local write
+            static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+                ignore = i_inst;
+
+                static_for<0, staged_num_ds_write_a_per_ds_read_a, 1>{}([&](auto idswrite_inst) {
+                    ignore = idswrite_inst;
+                    __builtin_amdgcn_sched_group_barrier(
+                        0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                });
+
+                __builtin_amdgcn_sched_group_barrier(
+                    0x008, staged_num_ds_write_a_per_ds_read_a, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0);  // DS read
+            });
+#elif 1
+            constexpr auto staged_num_mfma_per_ds_write_a =
+                math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a);
+
+            constexpr auto stage_more_mfma =
+                staged_num_mfma - (staged_num_mfma_per_ds_write_a - 1) * num_ds_write_inst_a;
+
+            // A local write
+            static_for<0, num_ds_write_inst_a, 1>{}([&](auto i_inst) {
+                if constexpr(i_inst.value < stage_more_mfma)
+                {
+                    if(i_inst.value < staged_num_ds_read_inst_a)
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a - 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
+                    }
+                    else
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a, 0);     // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                    }
+                }
+                else
+                {
+                    if(i_inst.value < staged_num_ds_read_inst_a)
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a - 2, 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
+                    }
+                    else
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                    }
+                }
+            });
+#endif
+            __builtin_amdgcn_sched_barrier(0);
+        }
+        else
+        {
+            // A local Read
+            static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+                ignore = i_inst;
+                __builtin_amdgcn_sched_group_barrier(
+                    0x008, staged_num_mfma_per_ds_read_a, 0);      // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+        }
+    }
+
+    __device__ static constexpr auto EpilogueScheduler_2()
+    {
+        constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num;
+
+        constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num;
+
+        constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat;
+        constexpr auto staged_num_mfma           = num_mfma / MRepeat;
+
+        constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a;
+
+        // A local Read
+        static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+            ignore = i_inst;
+            __builtin_amdgcn_sched_group_barrier(0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+        });
+
+        __builtin_amdgcn_sched_barrier(0);
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              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,
+                        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
+    {
+        ignore = b_block_buf;
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+        constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
+
+        // Global prefetch A1 B1
+        b_blockwise_copy.Run(b_grid_desc,
+                             b_grid_buf,
+                             b_block_desc_n0_n1_k0_k1,
+                             b_block_origin_idx,
+                             b_thread_bufs(I0));
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        __builtin_amdgcn_sched_barrier(0);
+
+        // // Local prefill A1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
+
+        // // Global prefetch A2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Local prefetch A1
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k0) {
+            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                               make_tuple(I0, I0, I0, k0, I0, I0),
+                               a_block_buf.At(I0),
+                               a_thread_desc_,
+                               make_tuple(I0, I0, I0, k0, I0, I0),
+                               a_thread_buf);
+        });
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        if constexpr(m0.value == 0)
+                        {
+                            b_blockwise_copy.Run(b_grid_desc,
+                                                 b_grid_buf,
+                                                 b_block_desc_n0_n1_k0_k1,
+                                                 b_block_origin_idx,
+                                                 b_thread_bufs(local_read_buf));
+                            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+                        }
+                        else if constexpr(m0.value == 1)
+                        {
+                            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf));
+                        }
+                        else if constexpr(m0.value == 2)
+                        {
+                            a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                        }
+
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                                vector_type<ComputeDataType, KPack> a_thread_vec;
+                                vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                                static_for<0, KPack, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                            make_tuple((m0 + HotloopLocalBufSwitch * mfma_reg_buf) %
+                                                           2,
+                                                       I0,
+                                                       I0,
+                                                       k0,
+                                                       I0,
+                                                       ik))>{}];
+                                    b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        b_thread_bufs[mfma_reg_buf]
+                                                     [Number<b_thread_desc_.CalculateOffset(
+                                                         make_tuple(n0, I0, k0, ik))>{}];
+                                });
+
+                                using mfma_input_type =
+                                    typename vector_type<ComputeDataType,
+                                                         xdlops_gemm.K1PerXdlops>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                xdlops_gemm.Run(
+                                    a_thread_vec.template AsType<mfma_input_type>(),
+                                    b_thread_vec.template AsType<mfma_input_type>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                            });
+                        });
+
+                        if constexpr(m0.value == MRepeat - 1)
+                        {
+                            block_sync_lds();
+
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                a_thread_copy_.Run(
+                                    a_block_desc_m0_m1_m2_k0_k1_k2,
+                                    make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0),
+                                    a_block_buf.At(local_read_buf),
+                                    a_thread_desc_,
+                                    make_tuple(
+                                        Number<(m0 + 1 + HotloopLocalBufSwitch * mfma_reg_buf) %
+                                               2>{},
+                                        I0,
+                                        I0,
+                                        k0,
+                                        I0,
+                                        I0),
+                                    a_thread_buf);
+                            });
+                        }
+                        else
+                        {
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                a_thread_copy_.Run(
+                                    a_block_desc_m0_m1_m2_k0_k1_k2,
+                                    make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0),
+                                    a_block_buf.At(mfma_reg_buf),
+                                    a_thread_desc_,
+                                    make_tuple(
+                                        Number<(m0 + 1 + HotloopLocalBufSwitch * mfma_reg_buf) %
+                                               2>{},
+                                        I0,
+                                        I0,
+                                        k0,
+                                        I0,
+                                        I0),
+                                    a_thread_buf);
+                            });
+                        }
+
+                        HotLoopScheduler(m0);
+                    });
+                };
+
+                LoopFunc(I0, I1);
+                LoopFunc(I1, I0);
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+        // tail
+        if constexpr(TailNum == TailNumber::Even)
+        {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                if constexpr(m0.value == 0)
+                {
+                    b_blockwise_copy.Run(b_grid_desc,
+                                         b_grid_buf,
+                                         b_block_desc_n0_n1_k0_k1,
+                                         b_block_origin_idx,
+                                         b_thread_bufs(I1));
+                }
+                else if constexpr(m0.value == MRepeat - 1)
+                {
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
+                }
+
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0 % 2, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+
+                if constexpr(m0.value == MRepeat - 1)
+                {
+                    block_sync_lds();
+
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        a_thread_copy_.Run(
+                            a_block_desc_m0_m1_m2_k0_k1_k2,
+                            make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0),
+                            a_block_buf.At(I1),
+                            a_thread_desc_,
+                            make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0),
+                            a_thread_buf);
+                    });
+                }
+                else
+                {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        a_thread_copy_.Run(
+                            a_block_desc_m0_m1_m2_k0_k1_k2,
+                            make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0),
+                            a_block_buf.At(I0),
+                            a_thread_desc_,
+                            make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0),
+                            a_thread_buf);
+                    });
+                }
+
+                EpilogueScheduler_1(m0);
+            });
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(make_tuple(
+                                    (m0 + HotloopLocalBufSwitch) % 2, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+
+                if constexpr(m0.value != (MRepeat - 1))
+                {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        a_thread_copy_.Run(
+                            a_block_desc_m0_m1_m2_k0_k1_k2,
+                            make_tuple(Number<m0 + 1>{}, I0, I0, k0, I0, I0),
+                            a_block_buf.At(I1),
+                            a_thread_desc_,
+                            make_tuple(
+                                Number<(m0 + 1 + HotloopLocalBufSwitch) % 2>{}, I0, I0, k0, I0, I0),
+                            a_thread_buf);
+                    });
+
+                    EpilogueScheduler_2();
+                }
+            });
+            // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle
+            // latency
+            // __builtin_amdgcn_sched_barrier(0);
+        }
+        else
+        {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0 % 2, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+
+                if constexpr(m0.value != (MRepeat - 1))
+                {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                                           make_tuple(Number<m0 + 1>{}, I0, I0, k0, I0, I0),
+                                           a_block_buf.At(I0),
+                                           a_thread_desc_,
+                                           make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0),
+                                           a_thread_buf);
+                    });
+
+                    EpilogueScheduler_2();
+                }
+            });
+        }
+    }
+
+    protected:
+    // MRepeat MWave MLane KRepeat KLane KPack
+    // KRepeat -> MRepeat-> Mwave->KLane->MLane->KPack
+    // Reduce the vgpr usage here.
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(I2, I1, I1, Number<KRepeat>{}, I1, Number<KPack>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
+                                                         ComputeDataType,
+                                                         decltype(a_block_desc_m0_m1_m2_k0_k1_k2),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3, 4, 5>,
+                                                         5,
+                                                         A_K1,
+                                                         A_K1>;
+
+    AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex6D()};
+
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
+
+    static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
+
+    using Base::c_thread_desc_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/device/device_gemm_v2.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -114,6 +114,38 @@ struct DeviceGemmV2BScale : public BaseOperator
     virtual ck::index_t GetKPerBlock() = 0;
 };
 
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmV2BPreshuffle : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        ck::index_t KSplit,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+
+    virtual bool GetPermuteB()         = 0;
+    virtual ck::index_t GetKPerBlock() = 0;
+};
+
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_preshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_v2.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_b_preshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/flush_cache.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA,
+          bool PermuteA                               = false,
+          bool PermuteB                               = false>
+struct DeviceGemm_Xdl_CShuffleV3_BPreshuffle : public DeviceGemmV2BPreshuffle<ALayout,
+                                                                              BLayout,
+                                                                              CLayout,
+                                                                              ADataType,
+                                                                              BDataType,
+                                                                              CDataType,
+                                                                              AElementwiseOperation,
+                                                                              BElementwiseOperation,
+                                                                              CElementwiseOperation>
+{
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_xdl_cshuffle_v3_b_preshuffle<
+        ALayout,
+        BLayout,
+        CLayout,
+        ADataType,
+        BDataType,
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        BlkGemmPipeSched,
+        BlkGemmPipelineVer,
+        ComputeTypeA,
+        ComputeTypeB,
+        PermuteA,
+        PermuteB>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    int GetPreShuffleParameters() override { return NPerXDL; }
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+                GridwiseGemm::BlockwiseGemmPipe::HotLoopInstList::Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            index_t gdx, gdy, gdz;
+            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch);
+
+            float ave_time = 0;
+
+            index_t k_grain = arg.KBatch * KPerBlock;
+            index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
+
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
+
+            const auto Run = [&](const auto& kernel) {
+                if(stream_config.flush_cache)
+                {
+                    Argument arg_ = arg;
+
+                    const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
+                        arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
+                    const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
+                        arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
+
+                    auto size_a_buffer =
+                        a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
+                    auto size_b_buffer =
+                        b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
+
+                    ck::utility::RotatingMemWrapper<Argument> rotating_mem(
+                        arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer);
+                    rotating_mem.Print();
+
+                    auto run_flush_cache = [&]() {
+                        // flush icache
+                        ck::utility::flush_icache();
+                        // rotating mem
+                        rotating_mem.Next();
+                        // clear c mem
+                        if(arg_.KBatch > 1)
+                            hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
+                                                             0,
+                                                             arg_.M * arg_.N * sizeof(CDataType),
+                                                             stream_config.stream_id_));
+                    };
+
+                    ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
+                        stream_config,
+                        run_flush_cache,
+                        kernel,
+                        dim3(gdx, gdy, gdz),
+                        dim3(BlockSize),
+                        0,
+                        arg_);
+                }
+                else
+                {
+                    if(arg.KBatch > 1)
+                        hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
+                                                         0,
+                                                         arg.M * arg.N * sizeof(CDataType),
+                                                         stream_config.stream_id_));
+
+                    ave_time = launch_and_time_kernel(
+                        stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+                }
+            };
+
+            constexpr auto estimated_reg_a = MPerBlock * KPerBlock * sizeof(ADataType) / BlockSize /
+                                             4 * (1 + GridwiseGemm::NWave);
+            constexpr auto estimated_reg_b =
+                NPerBlock * KPerBlock * sizeof(BDataType) / BlockSize / 4 * (2);
+            constexpr auto estimated_reg_c =
+                MPerBlock * NPerBlock * sizeof(GemmAccDataType) / BlockSize / 4;
+            constexpr auto estimated_reg_total =
+                estimated_reg_a + estimated_reg_b + estimated_reg_c;
+
+            constexpr index_t minimum_occupancy = (estimated_reg_total >= 256) ? 1 : 2;
+
+            if(has_main_k_block_loop)
+            {
+                // Tail number always full
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+                    if(arg.KBatch > 1)
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::AtomicAdd,
+                                minimum_occupancy,
+                                TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::AtomicAdd,
+                                minimum_occupancy,
+                                TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                    else
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::Set,
+                                minimum_occupancy,
+                                TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::Set,
+                                minimum_occupancy,
+                                TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                }
+                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2 ||
+                                 BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+                    if(arg.KBatch > 1)
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::AtomicAdd,
+                                minimum_occupancy,
+                                TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::AtomicAdd,
+                                minimum_occupancy,
+                                TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                    else
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::Set,
+                                minimum_occupancy,
+                                TailNumber::Odd>;
+                            Run(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
+                                GridwiseGemm,
+                                true,
+                                InMemoryDataOperationEnum::Set,
+                                minimum_occupancy,
+                                TailNumber::Even>;
+                            Run(kernel);
+                        }
+                    }
+                }
+                else
+                {
+                    throw std::runtime_error("Only support pipeline ver v1, v2, v3 now!");
+            }
+#if 0
+            else
+            {
+                // Tail number always 1
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+                    if(arg.KBatch > 1)
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                        false,
+                                                        InMemoryDataOperationEnum::AtomicAdd,
+                                                        minimum_occupancy>;
+                        Run(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
+                                                        false,
+                                                        InMemoryDataOperationEnum::Set,
+                                                        minimum_occupancy>;
+                        Run(kernel);
+                    }
+                }
+            }
+#endif
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
+        {
+            return false;
+        }
+
+        if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
+                                                       GemmSpec == GemmSpecialization::NKPadding ||
+                                                       GemmSpec == GemmSpecialization::MNKPadding ||
+                                                       GemmSpec == GemmSpecialization::KPadding))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    index_t GetKPerBlock() override { return KPerBlock; }
+
+    bool GetPermuteA() override { return PermuteA; }
+    bool GetPermuteB() override { return PermuteB; }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             index_t KBatch,
+                             AElementwiseOperation,
+                             BElementwiseOperation,
+                             CElementwiseOperation)
+    {
+        return Argument{p_a, p_b, p_c, M, N, K, StrideA, StrideB, StrideC, KBatch};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      index_t KBatch,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          KBatch);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
+            {BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
+            {BlockGemmPipelineScheduler::Interwave, "Interwave"}};
+
+        std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
+            {BlockGemmPipelineVersion::v1, "v1"},
+            {BlockGemmPipelineVersion::v2, "v2"},
+            {BlockGemmPipelineVersion::v3, "v3"},
+            {BlockGemmPipelineVersion::v4, "v4"},
+            {BlockGemmPipelineVersion::v5, "v5"}};
+
+        // clang-format off
+        str << "DeviceGemmXdlUniversal"
+            << "<"
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << ">"
+            << " BlkSize: "
+            << BlockSize << ", "
+            << "BlkTile: "
+            << MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
+            << "WaveTile: "
+            << MPerXDL<<"x"<<NPerXDL << ", "
+            << "WaveMap: "
+            << MXdlPerWave<<"x" << NXdlPerWave<<", "
+            << "VmemReadVec: "
+            << ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
+            << "BlkGemmPipelineScheduler: "
+            << BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
+            << "BlkGemmPipelineVersion: "
+            << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
+            << "BlkGemmPipelinePrefetchStages: "
+            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages << ", "
+            << "Kpack: "
+            << GridwiseGemm::BlockwiseGemmPipe::AMmaKStride;
+        // clang-format on
+
+        return str.str();
+    }
+    REGISTER_EXTRA_PRINTING_METHODS
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_b_preshuffle.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same
+// kernel function Blockers:
+// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on
+// two lds chunks.
+// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds
+// buffer when we declare __shared__ inside blkgemmpipe
+template <typename GridwiseGemm,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t MinimumOccupancy = 1,
+          TailNumber TailNum       = TailNumber::Even>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
+#endif
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_gemm_xdl_cshuffle_v3_b_preshuffle(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
+        karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
+        karg.p_c_grid + splitk_batch_offset.c_reduce_offset,
+        p_shared,
+        karg);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx9__))
+}
+
+template <typename GridwiseGemm,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t MinimumOccupancy = 1,
+          TailNumber TailNum       = TailNumber::Even>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
+#endif
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
+    // Pass two lds pointer is the key to tell compiler that ds_read/write
+    // operate on different lds chunk at same time without order dependecy
+    __shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    __shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
+
+    GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
+        karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
+        karg.p_c_grid + splitk_batch_offset.c_reduce_offset,
+        p_shared_0,
+        p_shared_1,
+        karg);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx9__))
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA,
+          bool PermuteA                               = false,
+          bool PermuteB                               = false>
+struct GridwiseGemm_xdl_cshuffle_v3_b_preshuffle
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // 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 mfma_selector = MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>;
+
+    static constexpr index_t KPack =
+        math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk);
+
+    static constexpr index_t KLane =
+        mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops();
+    static constexpr index_t KRepeat = KPerBlock / KLane / KPack;
+    static constexpr index_t NLane   = NPerXdl;
+    static constexpr index_t NWave   = NPerBlock / NPerXdl / NXdlPerWave;
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t APackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<ADataType>, pk_i4_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    static constexpr index_t BPackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, KBatch);
+    }
+
+    __host__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+    
+    __host__ __device__ static auto CalculateBN0Shuffled(index_t N)
+    {
+        return math::integer_divide_ceil(N, NLane);
+    }
+
+    __host__ __device__ static auto CalculateBK0Shuffled(index_t K)
+    {
+        return math::integer_divide_ceil(K, KLane * KPack);
+    }
+
+    __host__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ static auto CalculateAK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / AK1Value);
+    }
+
+    __host__ static auto CalculateBK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / BK1Value);
+    }
+
+    __host__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * KPerBlock;
+    }
+
+    __host__ static auto CalculateKRead(index_t K, index_t K_Batch = 1)
+    {
+        constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+        auto K_t                = K_Batch * KReadVec;
+        return (K + K_t - 1) / K_t * KReadVec;
+    }
+
+    __host__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock);
+    }
+
+    __host__ static auto CalculateNBlock(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock);
+    }
+
+    template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+    __host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&)
+    {
+        constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
+        constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
+
+        return transform_tensor_descriptor(
+            TileDesc_K0_MN_K1{},
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
+                       make_unmerge_transform(make_tuple(
+                           Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
+        index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(M, MPad - M),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(MPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    __host__ __device__ static auto MakeBGridDescriptor_Preshuffled(index_t N0, index_t K0)
+    {
+        constexpr index_t NkSwizzleNumber = Number<warpSize * KPack>{};
+        return make_naive_tensor_descriptor(
+            make_tuple(N0 / NWave, NWave, K0, NkSwizzleNumber),
+            make_tuple(NWave * K0 * NkSwizzleNumber, K0 * NkSwizzleNumber, NkSwizzleNumber, I1));
+    }
+
+    __host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
+        index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        static_assert(!(is_same_v<remove_cvref_t<ADataType>, pk_i4_t> &&
+                        GemmSpec != GemmSpecialization::Default),
+                      "pk_i4_t does not support padding");
+
+        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
+        {
+            if constexpr(!PermuteB)
+            {
+                // 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;
+            }
+            else
+            {
+                // Pre-shuffled Weight
+                // BGlobal[K / KPerBlock, N, KPerBlock / K1, K1] -> BTile[K / K1, N, K1]
+                constexpr index_t BK01 = KPerBlock / BK1Value;
+                const index_t BK0_     = StrideB / BK1Value;
+                const index_t BK00     = BK0_ / BK01;
+
+                const auto b_grid_desc_bk00_n_bk01_bk1_permute =
+                    make_naive_tensor_descriptor_packed(make_tuple(BK00, N, BK01, BK1Value));
+
+                const auto b_grid_desc_bk0_n_bk1_permute = transform_tensor_descriptor(
+                    b_grid_desc_bk00_n_bk01_bk1_permute,
+                    make_tuple(make_merge_transform(make_tuple(BK00, BK01)),
+                               make_pass_through_transform(make_tuple(N)),
+                               make_pass_through_transform(BK1Value)),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                return b_grid_desc_bk0_n_bk1_permute;
+            }
+        }
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor<MXdlPerWave, MWaves, MPerXdl>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        // constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        // return MakeGemmMmaTileDescriptor<NXdlPerWave, NWaves, NPerXdl>(BBlockDesc_BK0_N_BK1{});
+
+        return MakeGemmMmaTileDescriptor<NXdlPerWave, NWave, 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));
+            }
+        }();
+
+        // pad M and N
+        return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                           make_tuple(make_right_pad_transform(M, MPad - M),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}));
+#if 0
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(M, MPad - M),
+                                                          make_right_pad_transform(N, NPad - N)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+#endif
+    }
+
+    struct Problem
+    {
+        __host__ Problem(index_t M_,
+                         index_t N_,
+                         index_t K_,
+                         index_t StrideA_,
+                         index_t StrideB_,
+                         index_t StrideC_,
+                         index_t KBatch_)
+            : M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideC{StrideC_},
+              KBatch{KBatch_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              KRead{CalculateKRead(K_, KBatch_)},
+              KPadded{CalculateKPadded(K_, KBatch_)},
+              AK0{CalculateAK0Padded(K_, KBatch_)},
+              BK0{CalculateBK0Padded(K_, KBatch_)},
+              MBlock{CalculateMBlock(M_)},
+              NBlock{CalculateNBlock(N_)},
+              BN0Shuffled{CalculateBN0Shuffled(N_)},
+              BK0Shuffled{CalculateBK0Shuffled(K_)}
+        {
+        }
+
+        __host__ void Print() const
+        {
+            std::cout << "problem {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "KRead:" << KRead << ", "
+                      << "KP:" << KPadded << ", "
+                      << "AK0:" << AK0 << ", "
+                      << "BK0:" << BK0 << ", "
+                      << "MBlock: " << MBlock << ", "
+                      << "NBlock: " << NBlock << "}" << std::endl;
+        }
+
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        index_t StrideC;
+        index_t KBatch;
+        index_t MPadded;
+        index_t NPadded;
+        index_t KRead;
+        index_t KPadded;
+        index_t AK0;
+        index_t BK0;
+        index_t MBlock;
+        index_t NBlock;
+        // For B pre-shuffle only
+        index_t BN0Shuffled;
+        index_t BK0Shuffled;
+    };
+
+    // Argument
+    struct Argument : public tensor_operation::device::BaseArgument, public Problem
+    {
+        __host__ Argument(const ADataType* p_a_grid_,
+                          const BDataType* p_b_grid_,
+                          CDataType* p_c_grid_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          index_t StrideC_,
+                          index_t k_batch_,
+                          bool is_reduce_ = false)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_, k_batch_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_c_grid{p_c_grid_},
+              is_reduce(is_reduce_)
+        {
+        }
+
+        __host__ __device__ inline bool IsReduceAdd() const
+        {
+            return (Problem::KBatch > 1) && is_reduce;
+        }
+
+        __host__ __device__ inline bool IsAtomicAdd() const
+        {
+            return (Problem::KBatch > 1) && (!is_reduce);
+        }
+
+        const ADataType* p_a_grid;
+        const BDataType* p_b_grid;
+        CDataType* p_c_grid;
+        bool is_reduce;
+    };
+
+    struct SplitKBatchOffset
+    {
+
+        __device__ SplitKBatchOffset(Argument& karg)
+        {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                a_k_split_offset = blockIdx.z * karg.KRead / APackedSize;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                a_k_split_offset = blockIdx.z * karg.KRead * karg.StrideA;
+            }
+
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
+            {
+                b_k_split_offset = blockIdx.z * karg.KRead * karg.StrideB;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
+            {
+                if constexpr(!PermuteB)
+                {
+                    // b_k_split_offset = blockIdx.z * karg.KRead / BPackedSize;
+
+                    b_k_split_offset = blockIdx.z * karg.KRead * NLane / BPackedSize;
+                }
+                else
+                {
+                    const int k0_offset = karg.KRead * karg.N;
+                    b_k_split_offset    = blockIdx.z * k0_offset / BPackedSize;
+                }
+            }
+
+            if(blockIdx.z < static_cast<uint32_t>(karg.KBatch - 1))
+            {
+                karg.K = karg.KRead;
+            }
+            else
+            {
+                karg.K = karg.K - karg.KRead * (karg.KBatch - 1);
+            }
+
+            if(karg.IsReduceAdd())
+            {
+                c_reduce_offset = blockIdx.z * karg.M * karg.N;
+            }
+            else
+            {
+                c_reduce_offset = 0;
+            }
+        }
+
+        index_t a_k_split_offset;
+        index_t b_k_split_offset;
+        index_t c_reduce_offset;
+    };
+
+    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        if constexpr(ABlockLdsExtraM)
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock + ABlockLdsExtraM>{}, I1));
+        }
+        // xor tensor transformation request more unnecessary vgpr usage, would cause register spill
+        // in some cases.
+        else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            constexpr auto a_lds_block_desc =
+                make_naive_tensor_descriptor(make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+                                             make_tuple(AK1Number, Number<KPerBlock>{}, I1));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(make_xor_with_modulo_transform(
+                               make_tuple(Number<MPerBlock>{}, Number<AK0Number>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_permuted;
+        }
+        else // ColumnMajor A
+        {
+            // kfold and mpair dimension is not always required.
+            // more dimension in merge_transform increase the difficulty of generating immarg offset
+            // for compiler.
+            constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
+            constexpr auto M1 = MPerBlock / M0;
+
+            constexpr auto KThreadWrite     = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0);
+            constexpr auto K0PerThreadWrite = AK0Number / KThreadWrite;
+            constexpr auto KThreadRead      = 64 / MPerXdl;
+            constexpr auto K0PerThreadRead  = AK0Number / KThreadRead;
+
+            constexpr auto kfold = (AK1Number * M0 * sizeof(ADataType) > 128)
+                                       ? 1
+                                       : 128 / (AK1Number * M0 * sizeof(ADataType));
+            constexpr auto KThreadReadPerm =
+                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
+                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
+                    : KThreadRead;
+
+            // 1<=mpair<=n0
+            constexpr auto mpair = (AK1Number * MPerXdl * sizeof(ADataType) > 128)
+                                       ? 1
+                                       : ((128 / (AK1Number * MPerXdl * sizeof(ADataType))) > M0
+                                              ? M0
+                                              : 128 / (AK1Number * MPerXdl * sizeof(ADataType)));
+
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                           Number<K0PerThreadWrite>{},
+                           Number<KThreadReadPerm * M1>{},
+                           Number<kfold * M0 / mpair>{},
+                           Number<mpair>{},
+                           AK1Number));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_xor_with_modulo_transform(
+                        make_tuple(Number<KThreadReadPerm * M1>{}, Number<kfold * M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
+
+            constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<M1>{})),
+                    make_unmerge_transform(make_tuple(Number<kfold>{}, Number<M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<0, 3>{},
+                           Sequence<4, 5>{},
+                           Sequence<6>{},
+                           Sequence<7>{}));
+
+            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_unmerged,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(Number<KThreadReadPerm>{},
+                                          Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                                          Number<kfold>{},
+                                          Number<K0PerThreadWrite>{})),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<M0 / mpair>{}, Number<mpair>{}, Number<M1>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_ak0_m_ak1;
+        }
+    }
+
+    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}));  //??? BK1Value same as KPack?
+    }
+
+    __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    using BlockwiseGemmPipe =
+        remove_cvref_t<decltype(BlockGemmBPreshufflePipeline_Selector<
+                                BlkGemmPipelineVer,
+                                BlkGemmPipeSched,
+                                BlockSize,
+                                ADataType,
+                                BDataType,
+                                ComputeTypeA,
+                                AccDataType,
+                                decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()),
+                                decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()),
+                                decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(
+                                    GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())),
+                                decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(
+                                    GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())),
+                                ABlockTransferSrcScalarPerVector,
+                                BBlockTransferSrcScalarPerVector,
+                                MPerBlock,
+                                NPerBlock,
+                                KPerBlock,
+                                MPerXdl,
+                                NPerXdl,
+                                MXdlPerWave,
+                                NXdlPerWave,
+                                KPack>())>;
+
+    __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max(a_block_space_size_aligned * sizeof(ADataType) / APackedSize,
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ static constexpr bool CheckValidity(const Argument& karg)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
+                     !(is_same<tensor_layout::gemm::RowMajor, ALayout>::value))
+        {
+            if(!(karg.M % MPerBlock == 0))
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " "
+                              << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                              << std::endl;
+                }
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
+                     (is_same<tensor_layout::gemm::RowMajor, BLayout>::value))
+        {
+            if(!(karg.N % NPerBlock == 0))
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " "
+                              << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                              << std::endl;
+                }
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+
+            auto K_t = karg.KBatch * KPerBlock;
+            if(!(karg.K % K_t == 0))
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
+                              << karg.K << " " << __FILE__ << ":" << __LINE__
+                              << ", in function: " << __func__ << std::endl;
+                }
+                return false;
+            }
+        }
+        else
+        {
+            constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+            auto K_t                = karg.KBatch * KReadVec;
+            auto KReadPadSplited    = math::integer_divide_ceil(karg.K, K_t) * KReadVec;
+            if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(karg.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << "Arg K (" << karg.K
+                              << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                              << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                              << __LINE__ << ", in function: " << __func__ << std::endl;
+                }
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << "Arg M (" << karg.M
+                              << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                              << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                              << __LINE__ << ", in function: " << __func__ << std::endl;
+                }
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(karg.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << "Arg N (" << karg.N
+                              << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                              << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                              << __LINE__ << ", in function: " << __func__ << std::endl;
+                }
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << "Arg K (" << karg.K
+                              << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                              << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                              << __LINE__ << ", in function: " << __func__ << std::endl;
+                }
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << "Arg N (" << karg.N
+                              << ") value is not a multiple of "
+                                 "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                              << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! "
+                              << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                              << std::endl;
+                }
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << "Arg M (" << karg.M
+                              << ") value is not a multiple of "
+                                 "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                              << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! "
+                              << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                              << std::endl;
+                }
+                return false;
+            }
+        }
+
+        if constexpr(!(is_same<remove_cvref_t<CDataType>, half_t>::value ||
+                       is_same<remove_cvref_t<CDataType>, float>::value ||
+                       is_same<remove_cvref_t<CDataType>, bhalf_t>::value ||
+                       is_same<remove_cvref_t<CDataType>, int32_t>::value))
+        {
+            if(!karg.IsReduceAdd())
+            {
+                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+                {
+                    std::cout << " KBatch: " << karg.KBatch << " > 1 is not support yet" << __FILE__
+                              << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+                }
+                if(karg.KBatch > 1)
+                {
+                    return false;
+                }
+            }
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value);
+
+        if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages)
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockHasHotloop(num_loop);
+    }
+
+    __host__ static constexpr TailNumber CalculateKBlockLoopTailNum(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockLoopTailNum(num_loop);
+    }
+
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    // if arch = gfx942
+    using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
+    // using Block2CTileMap = BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>;
+
+    template <typename AGridDesc_AK0_M_K1,
+              typename BGridDesc_BPreshuffled,
+              typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run(const ADataType* p_a_grid,
+                               const BDataType* p_b_grid,
+                               CDataType* p_c_grid,
+                               void* p_shared,
+                               const Problem& problem,
+                               const AGridDesc_AK0_M_K1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BPreshuffled& b_grid_desc_bpreshuffled,
+                               const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bpreshuffled.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const 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, 4};
+
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+        const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
+        
+        // N0, K0, Blocksize*KPack
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0Number, MPerBlock, AK1Number>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ADataType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                BlockwiseGemmPipe::GlobalBufferNum>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix threadwise copy, using threadwiseTensorSliceTransfer_v2
+        auto b_block_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BDataType>(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto b_blockwise_copy = ThreadwiseTensorSliceTransfer_v2<
+            BDataType,
+            BDataType,
+            decltype(b_grid_desc_bpreshuffled),
+            decltype(b_block_desc_bk0_n_bk1),
+            Sequence<Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}>,
+            Sequence<1, 2, 0, 3>,
+            3,
+            BBlockTransferSrcScalarPerVector,
+            BThreadTransferSrcResetCoordinateAfterRun,
+            true>(b_grid_desc_bpreshuffled,
+                  make_multi_index(n_block_data_idx_on_grid,
+                                   get_warp_local_1d_id() % NWave,
+                                   0,
+                                   KPack * (get_thread_local_1d_id() % warpSize)));
+
+        // LDS allocation for A and B: be careful of alignment
+
+
+        // Cast after lds
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ADataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        // auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+        //     reinterpret_cast<BDataType*>(static_cast<char*>(p_shared) + a_block_space_size_aligned *
+        //                                                                     sizeof(ADataType) /
+        //                                                                     APackedSize),
+        //     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(0, 0, KRepeat, 0);
+
+        // Blockwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
+        auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
+        auto c_thread_buf            = blockwise_gemm_pipeline.GetCThreadBuffer();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
+                                                                         a_block_desc_ak0_m_ak1,
+                                                                         a_blockwise_copy,
+                                                                         a_grid_buf,
+                                                                         a_block_buf,
+                                                                         a_block_slice_copy_step,
+                                                                         b_grid_desc_bpreshuffled,
+                                                                         b_blockwise_copy,
+                                                                         b_grid_buf,
+                                                                         b_block_buf,
+                                                                         b_block_slice_copy_step,
+                                                                         c_thread_buf,
+                                                                         num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // 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,
+                CShuffleDataType,     // typename SrcData,
+                CDataType,            // 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_m_id, 0, block_n_id, 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // 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);
+                }
+            });
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run(const ADataType* p_a_grid,
+                               const BDataType* p_b_grid,
+                               CDataType* p_c_grid,
+                               void* 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_bpreshuffled =
+            MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled);
+        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);
+
+        Run<decltype(a_grid_desc_ak0_m_ak1),
+            decltype(b_grid_desc_bpreshuffled),
+            decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+            HasMainKBlockLoop,
+            CGlobalMemoryDataOperation,
+            TailNum>(p_a_grid,
+                     p_b_grid,
+                     p_c_grid,
+                     p_shared,
+                     problem,
+                     a_grid_desc_ak0_m_ak1,
+                     b_grid_desc_bpreshuffled,
+                     c_grid_desc_mblock_mperblock_nblock_nperblock);
+    }
+
+    template <typename AGridDesc_AK0_M_K1,
+              typename BGridDesc_BPreshuffled,
+              typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run_2Lds(const ADataType* p_a_grid,
+                                    const BDataType* p_b_grid,
+                                    CDataType* p_c_grid,
+                                    void* p_shared_0,
+                                    void* p_shared_1,
+                                    const Problem& problem,
+                                    const AGridDesc_AK0_M_K1& a_grid_desc_ak0_m_ak1,
+                                    const BGridDesc_BPreshuffled& b_grid_desc_bpreshuffled,
+                                    const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                        c_grid_desc_mblock_mperblock_nblock_nperblock)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bpreshuffled.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const 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, 4};
+
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+        const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
+
+        // N0, K0, Blocksize*KPack
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0Number, MPerBlock, AK1Number>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ADataType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                2>(
+                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
+        // Thread-wise copy
+        // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack
+        auto b_block_buf_ping = make_static_buffer<AddressSpaceEnum::Vgpr, BDataType>(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto b_block_buf_pong = make_static_buffer<AddressSpaceEnum::Vgpr, BDataType>(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong);
+
+        auto b_blockwise_copy = ThreadwiseTensorSliceTransfer_v2<
+            BDataType,
+            BDataType,
+            decltype(b_grid_desc_bpreshuffled),
+            decltype(b_block_desc_bk0_n_bk1),
+            Sequence<Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}>,
+            Sequence<1, 2, 0, 3>,
+            3,
+            BBlockTransferSrcScalarPerVector,
+            BThreadTransferSrcResetCoordinateAfterRun,
+            true>(b_grid_desc_bpreshuffled,
+                  make_multi_index(n_block_data_idx_on_grid,
+                                   get_warp_local_1d_id() % NWave,
+                                   0,
+                                   KPack * (get_thread_local_1d_id() % warpSize)));
+
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ADataType*>(p_shared_0), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ADataType*>(p_shared_1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto a_block_bufs = make_tuple(a_block_buf_ping, a_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(, 0, KRepeat, 0);
+
+        // Blockwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
+        auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
+        auto c_thread_buf            = blockwise_gemm_pipeline.GetCThreadBuffer();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
+                                                                         a_block_desc_ak0_m_ak1,
+                                                                         a_blockwise_copy,
+                                                                         a_grid_buf,
+                                                                         a_block_bufs,
+                                                                         a_block_slice_copy_step,
+                                                                         b_grid_desc_bpreshuffled,
+                                                                         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);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared_0),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // 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,
+                CShuffleDataType,     // typename SrcData,
+                CDataType,            // 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_m_id, 0, block_n_id, 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // 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);
+                }
+            });
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run_2Lds(const ADataType* p_a_grid,
+                                    const BDataType* p_b_grid,
+                                    CDataType* p_c_grid,
+                                    void* p_shared_0,
+                                    void* p_shared_1,
+                                    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_bpreshuffled =
+            MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled);
+        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);
+
+        Run_2Lds<decltype(a_grid_desc_ak0_m_ak1),
+                 decltype(b_grid_desc_bpreshuffled),
+                 decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                 HasMainKBlockLoop,
+                 CGlobalMemoryDataOperation,
+                 TailNum>(p_a_grid,
+                          p_b_grid,
+                          p_c_grid,
+                          p_shared_0,
+                          p_shared_1,
+                          problem,
+                          a_grid_desc_ak0_m_ak1,
+                          b_grid_desc_bpreshuffled,
+                          c_grid_desc_mblock_mperblock_nblock_nperblock);
+    }
+};
+
+} // namespace ck