Merge branch 'develop' into uif2-migraphx

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

@@ -15,6 +15,9 @@
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
 namespace ck {
 
 // GEMM:
@@ -26,7 +29,9 @@ namespace ck {
 //   E = cde_op(C, D0, D1, ...)
 // Assume:
 //   D0, D1, ... and E have the same layout
-template <typename ABDataType, // FIXME: don't assume A/B have same datatype
+template <typename ADataType,
+          typename BDataType,
+          typename AComputeDataType_,
           typename AccDataType,
           typename CShuffleDataType,
           typename DsDataType,
@@ -67,11 +72,14 @@ template <typename ABDataType, // FIXME: don't assume A/B have same datatype
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
           LoopScheduler LoopSched,
-          PipelineVersion PipelineVer = PipelineVersion::v1>
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename BComputeDataType   = AComputeDataType_>
 struct GridwiseGemmMultipleD_xdl_cshuffle
 {
     static constexpr index_t NumDTensor = DsDataType::Size();
 
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
@@ -92,15 +100,11 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
     using GridwiseGemmPipe = remove_cvref_t<
         decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
-    // denorm test fix, required to work around fp16 mfma issue
-    // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
-    // when mfma if fixed, remove this section and update
-    // ABDataTypeAdjusted -> ABDataType throughout this file
 #if CK_WORKAROUND_DENORM_FIX
-    using ABDataTypeAdjusted =
-        conditional_t<is_same_v<ABDataType, ck::half_t>, ck::bhalf_t, ABDataType>;
+    using AComputeDataType =
+        conditional_t<is_same_v<AComputeDataType_, ck::half_t>, ck::bhalf_t, AComputeDataType_>;
 #else
-    using ABDataTypeAdjusted = ABDataType;
+    using AComputeDataType = AComputeDataType_;
 #endif
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
@@ -169,8 +173,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
         constexpr auto c_block_size =
             c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
 
-        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
-                             sizeof(ABDataType),
+        return math::max(a_block_space_size_aligned * sizeof(AComputeDataType) +
+                             b_block_space_size_aligned * sizeof(BComputeDataType),
                          c_block_size * sizeof(CShuffleDataType));
     }
 
@@ -265,6 +269,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
         static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
                           (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
                       "Invalid tuning param!");
+        static_assert(KPerBlock % AK1Value == 0 && KPerBlock % BK1Value == 0,
+                      "KPerBlock must be divisible by AK1Value and BK1Value!");
 
         const auto M  = a_grid_desc_m_k.GetLength(I0);
         const auto N  = b_grid_desc_n_k.GetLength(I0);
@@ -313,8 +319,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
         // check tensor size: cannot be larger than 2GB each
         constexpr long_index_t TwoGB = (long_index_t{1} << 31);
 
-        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
-             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
              e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
         {
             return false;
@@ -332,14 +338,102 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
 
     using DsGridPointer = decltype(MakeDsGridPointer());
 
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
     template <bool HasMainKBlockLoop,
               typename AGridDesc_AK0_M_AK1,
               typename BGridDesc_BK0_N_BK1,
               typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
               typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
               typename Block2ETileMap>
-    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
-                               const ABDataType* __restrict__ p_b_grid,
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
                                DsGridPointer p_ds_grid,
                                EDataType* __restrict__ p_e_grid,
                                void* __restrict__ p_shared,
@@ -408,8 +502,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
                                                 Sequence<AK0PerBlock, MPerBlock, AK1>,
                                                 ABlockTransferThreadClusterLengths_AK0_M_AK1,
                                                 ABlockTransferThreadClusterArrangeOrder,
-                                                ABDataType,
-                                                ABDataTypeAdjusted,
+                                                ADataType,
+                                                AComputeDataType,
                                                 decltype(a_grid_desc_ak0_m_ak1),
                                                 decltype(a_block_desc_ak0_m_ak1),
                                                 ABlockTransferSrcAccessOrder,
@@ -439,8 +533,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
                                                 Sequence<BK0PerBlock, NPerBlock, BK1>,
                                                 BBlockTransferThreadClusterLengths_BK0_N_BK1,
                                                 BBlockTransferThreadClusterArrangeOrder,
-                                                ABDataType,
-                                                ABDataTypeAdjusted,
+                                                BDataType,
+                                                BComputeDataType,
                                                 decltype(b_grid_desc_bk0_n_bk1),
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
@@ -468,13 +562,15 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
         //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
         //       register
         // sanity check
-        constexpr index_t KPack =
-            math::max(math::lcm(AK1, BK1),
-                      MfmaSelector<ABDataTypeAdjusted, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1),
+            MfmaSelector<AComputeDataType, MPerXdl, NPerXdl, BComputeDataType>::selected_mfma
+                .k_per_blk);
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            ABDataTypeAdjusted,
+            AComputeDataType,
+            BComputeDataType,
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -492,11 +588,10 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
             a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ABDataTypeAdjusted*>(p_shared),
-            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+            static_cast<AComputeDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ABDataTypeAdjusted*>(p_shared) + a_block_space_size_aligned,
+            static_cast<BComputeDataType*>(p_shared) + a_block_space_size_aligned,
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
@@ -761,6 +856,85 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
             });
         }
     }
+
+    template <bool HasMainKBlockLoop,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ADataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const BDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        // tensor descriptors for problem definiton
+        const auto a_grid_desc_m_k = MakeAGridDescriptor_M_K<ALayout, GemmSpec>(M, K, StrideA);
+        const auto b_grid_desc_n_k = MakeBGridDescriptor_N_K<BLayout, GemmSpec>(K, N, StrideB);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_ak0_m_ak1 = MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+
+        const auto b_grid_desc_bk0_n_bk1 = MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        Run<HasMainKBlockLoop>(p_a_grid,
+                               p_b_grid,
+                               p_ds_grid,
+                               p_e_grid,
+                               p_shared,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               a_grid_desc_ak0_m_ak1,
+                               b_grid_desc_bk0_n_bk1,
+                               ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               block_2_etile_map);
+    }
 };
 
 } // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/amd_lds.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_direct_load.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_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename ADataType,
+          typename BDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_multiple_d_xdl_cshuffle_lds_direct_load(
+            const ADataType* __restrict__ p_a_grid,
+            const BDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx90a__) || defined(__gfx94__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore                 = p_a_grid;
+    ignore                 = p_b_grid;
+    ignore                 = p_ds_grid;
+    ignore                 = p_e_grid;
+    ignore                 = a_element_op;
+    ignore                 = b_element_op;
+    ignore                 = cde_element_op;
+    ignore                 = a_grid_desc_ak0_m_ak1;
+    ignore                 = b_grid_desc_bk0_n_bk1;
+    ignore                 = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore                 = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore                 = block_2_etile_map;
+#endif
+}
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AComputeDataType_,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferScalarPerVector,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferScalarPerVector,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v4,
+          typename BComputeDataType   = AComputeDataType_>
+struct GridwiseGemmMultipleD_Xdl_CShuffle_LdsDirectLoad
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+#if CK_WORKAROUND_DENORM_FIX
+    using AComputeDataType =
+        conditional_t<is_same_v<AComputeDataType_, ck::half_t>, ck::bhalf_t, AComputeDataType_>;
+#else
+    using AComputeDataType = AComputeDataType_;
+#endif
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, destination of blockwise copy.
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, destination of blockwise copy.
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment.
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle.
+        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(
+            NumGemmKPrefetchStage * a_block_space_size_aligned * sizeof(AComputeDataType) +
+                NumGemmKPrefetchStage * b_block_space_size_aligned * sizeof(BComputeDataType),
+            c_block_size * sizeof(CShuffleDataType));
+    }
+
+    __host__ __device__ static auto
+    MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    __host__ __device__ static auto
+    MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) { return MakeEGridDescriptor_M_N(MRaws[i], NRaws[i], DsStride[i]); },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N(1, 1, 1));
+
+    // A desc for source in blockwise copy.
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy.
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // E desc for destination in blockwise copy.
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_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 e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy.
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    using Block2ETileMap = remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    __host__ __device__ static constexpr bool CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                                                            const BGridDesc_N_K& b_grid_desc_n_k,
+                                                            const DsGridDesc_M_N& ds_grid_desc_m_n,
+                                                            const EGridDesc_M_N& e_grid_desc_m_n,
+                                                            const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+        static_assert(KPerBlock % AK1Value == 0 && KPerBlock % BK1Value == 0,
+                      "KPerBlock must be divisible by AK1Value and BK1Value!");
+
+        static_assert(
+            std::is_same_v<AElementwiseOperation,
+                           ck::tensor_operation::element_wise::PassThrough> &&
+                std::is_same_v<BElementwiseOperation,
+                               ck::tensor_operation::element_wise::PassThrough>,
+            "Direct load transfers do not support elementwise operations other than passthrough.");
+
+        const auto M  = a_grid_desc_m_k.GetLength(I0);
+        const auto N  = b_grid_desc_n_k.GetLength(I0);
+        const auto AK = a_grid_desc_m_k.GetLength(I1);
+        const auto BK = b_grid_desc_n_k.GetLength(I1);
+
+        // Check the consistency of descriptors.
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) && AK == BK))
+        {
+            return false;
+        }
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // Check the tile size.
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && AK % KPerBlock == 0))
+        {
+            return false;
+        }
+
+        // Check gridwise gemm pipeline.
+        const auto num_k_loop = AK / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // Check block-to-E-tile.
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // Check tensor size: cannot exceed 2GB.
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        // Elementwise operations are not supported for A and B, arguments left only for the API
+        // consistency.
+        (void)a_element_op;
+        (void)b_element_op;
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // Divide block work by [M, N].
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // This forces m/n_block_data_idx_on_grid into SGPR.
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, destination of blockwise copy.
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, destination of blockwise copy.
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_DirectLoad<ThisThreadBlock,
+                                                      Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                      ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                      ADataType,
+                                                      AComputeDataType,
+                                                      decltype(a_grid_desc_ak0_m_ak1),
+                                                      decltype(a_block_desc_ak0_m_ak1),
+                                                      ABlockTransferSrcVectorDim,
+                                                      2,
+                                                      ABlockTransferScalarPerVector>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0));
+
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_DirectLoad<ThisThreadBlock,
+                                                      Sequence<BK0PerBlock, NPerBlock, BK1>,
+                                                      BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                      BDataType,
+                                                      BComputeDataType,
+                                                      decltype(b_grid_desc_bk0_n_bk1),
+                                                      decltype(b_block_desc_bk0_n_bk1),
+                                                      BBlockTransferSrcVectorDim,
+                                                      2,
+                                                      BBlockTransferScalarPerVector>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1),
+            MfmaSelector<AComputeDataType, MPerXdl, NPerXdl, BComputeDataType>::selected_mfma
+                .k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            AComputeDataType,
+            BComputeDataType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment.
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        const auto a_buffers_offset = 0;
+        auto a_block_buffers =
+            ck::lds_utils::AllocateLdsBuffers<AComputeDataType, NumGemmKPrefetchStage>(
+                p_shared,
+                a_block_desc_ak0_m_ak1.GetElementSpaceSize(),
+                a_buffers_offset,
+                max_lds_align);
+        const auto b_buffers_offset = a_block_space_size_aligned * NumGemmKPrefetchStage;
+        auto b_block_buffers =
+            ck::lds_utils::AllocateLdsBuffers<BComputeDataType, NumGemmKPrefetchStage>(
+                p_shared,
+                b_block_desc_bk0_n_bk1.GetElementSpaceSize(),
+                b_buffers_offset,
+                max_lds_align);
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buffers,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buffers,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // Shuffle C and write out.
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            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 the origin of thread output tensor on global memory.
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_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{}};
+
+            // A tuple of reference to C/Ds tensor descriptors.
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+
+            // A tuple of reference to C/Ds grid buffers.
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // A tuple of starting index of C/Ds blockwise copy.
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            // Blockwise copy C/D/E between LDS and global.
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
+                 cde_element_op};
+
+            // Space filling curve for threadwise C in VGPR before shuffle.
+            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/D/E.
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // Make sure it's safe to write to LDS.
+                block_sync_lds();
+
+                // Each thread write its data from VGPR to LDS.
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // Make sure it's safe to read from LDS.
+                block_sync_lds();
+
+                // Each block copy its data from LDS to global.
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // Move on Ds.
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // Move on E.
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+        }
+    }
+
+    struct Argument : public tensor_operation::device::BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{MakeEGridDescriptor_M_N(MRaw, NRaw, StrideE)},
+              a_grid_desc_ak0_m_ak1_{MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              MRaw_{MRaw},
+              NRaw_{NRaw},
+              KRaw_{KRaw}
+        {
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType      = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+                p_ds_grid_(i)        = static_cast<const DDataType*>(p_ds_grid[i]);
+                ds_grid_desc_m_n_(i) = MakeEGridDescriptor_M_N(MRaw, NRaw, StrideDs[i]);
+            });
+
+            if(CheckValidity(a_grid_desc_m_k_,
+                             b_grid_desc_n_k_,
+                             ds_grid_desc_m_n_,
+                             e_grid_desc_m_n_,
+                             block_2_etile_map_))
+            {
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n_);
+
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        // Pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // Tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // Tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise ops
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // For checking vector load/store
+        index_t MRaw_;
+        index_t NRaw_;
+        index_t KRaw_;
+    };
+};
+
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ADataType,
+          typename BDataType,
+          typename ComputeType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_KBatch_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 CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
+    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 AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       AK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       BK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(ComputeType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch)
+    {
+        return math::integer_least_multiple(K, KPerBlock * K_Batch);
+    }
+
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_KBatch_AK0_M_AK1(index_t M, index_t K, index_t StrideA, index_t KBatch)
+    {
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto MPad = CalculateMPadded(M);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            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 AK0 = KPad / (KBatch * AK1);
+
+        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)
+        {
+            // const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_KBatch_BK0_N_BK1(index_t K, index_t N, index_t StrideB, index_t KBatch)
+    {
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto NPad = CalculateNPadded(N);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto BK0 = KPad / (KBatch * BK1);
+
+        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)
+        {
+            // const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_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 e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    template <typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              GemmSpecialization GemmSpec>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const index_t M,
+                  const index_t N,
+                  const index_t K,
+                  const index_t StrideA,
+                  const index_t StrideB,
+                  const std::array<index_t, NumDTensor> StrideDs,
+                  const index_t StrideE,
+                  const index_t KBatch)
+    {
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        ignore = StrideDs;
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+#if 0
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+#endif
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              index_t NumDTensor_,
+              typename DsDataType_,
+              typename AGridDesc_KBatch_AK0_M_AK1,
+              typename BGridDesc_KBatch_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename CDEElementwiseOperation_,
+              typename Block2ETileMap>
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const index_t KBatch,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation_& cde_element_op,
+                               const AGridDesc_KBatch_AK0_M_AK1& a_grid_desc_kbatch_ak0_m_ak1,
+                               const BGridDesc_KBatch_BK0_N_BK1& b_grid_desc_kbatch_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor_>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_kbatch_ak0_m_ak1 =
+            GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_kbatch_bk0_n_bk1 =
+            GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ComputeType,
+                                                decltype(a_grid_desc_kbatch_ak0_m_ak1),
+                                                decltype(a_block_desc_kbatch_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_kbatch_ak0_m_ak1,
+                make_multi_index(kbatch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_kbatch_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BDataType,
+                                                ComputeType,
+                                                decltype(b_grid_desc_kbatch_bk0_n_bk1),
+                                                decltype(b_block_desc_kbatch_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_kbatch_bk0_n_bk1,
+                make_multi_index(kbatch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_kbatch_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // 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();
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ComputeType,
+            ComputeType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+#if 1
+        if(block_work_idx[I0] == 0)
+        {
+            const index_t nThreadSize = CDEShuffleBlockTransferScalarPerVector_NPerBlock;
+            const index_t numNThreads = NPerBlock / nThreadSize;
+            const index_t numMThreads = BlockSize / numNThreads;
+            const index_t mThreadSize = MPerBlock / numMThreads;
+
+            const index_t m_tid = get_thread_local_1d_id() / numNThreads;
+            const index_t n_tid = get_thread_local_1d_id() % numNThreads;
+
+            auto c_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mThreadSize>{}, I1, Number<nThreadSize>{}));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr,
+                         EDataType,
+                         c_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize(),
+                         true>
+                e_thread_zero_buf;
+
+            auto c_thread_copy = ThreadwiseTensorSliceTransfer_v1r3<
+                EDataType,
+                EDataType,
+                decltype(c_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1, mThreadSize, 1, nThreadSize>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{e_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I1],
+                                       m_tid * mThreadSize,
+                                       block_work_idx[I2],
+                                       n_tid * nThreadSize),
+                      ck::tensor_operation::element_wise::PassThrough{}};
+
+            c_thread_copy.Run(c_thread_desc_mblock_mperblock_nblock_nperblock,
+                              make_tuple(I0, I0, I0, I0),
+                              e_thread_zero_buf,
+                              e_grid_desc_mblock_mperblock_nblock_nperblock,
+                              e_grid_buf);
+
+            __syncthreads();
+
+            if(threadIdx.x == 0)
+            {
+                atomicAdd(barrier_count_finished, 1);
+            }
+        }
+#endif
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop =
+            __builtin_amdgcn_readfirstlane((a_grid_desc_kbatch_ak0_m_ak1.GetLength(I1) *
+                                            a_grid_desc_kbatch_ak0_m_ak1.GetLength(I3)) /
+                                           KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_kbatch_ak0_m_ak1,
+                                                               a_block_desc_kbatch_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_kbatch_bk0_n_bk1,
+                                                               b_block_desc_kbatch_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            if(threadIdx.x == 0)
+            {
+                while(__atomic_load_n(barrier_count_finished, __ATOMIC_RELAXED) == 0) {}
+            }
+
+            __syncthreads();
+
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            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.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{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0);
+                    },
+                    Number<NumDTensor_>{}));
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            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/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType_{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation_,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
+                                                                            // Sequence support
+                                                                            // arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor_,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0)),
+                 cde_element_op};
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor_, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+
+            if(threadIdx.x == 0)
+            {
+                index_t k_id_finished_t = atomicAdd(barrier_count_finished, 1);
+
+                if(k_id_finished_t == KBatch)
+                {
+                    *barrier_count_finished = 0;
+                }
+            }
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const index_t KBatch,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ADataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const BDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        if(kbatch_id == KBatch - 1)
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, NumDTensor, DsDataType>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                cde_element_op,
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+        else
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, 0, Tuple<>>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                ck::tensor_operation::element_wise::PassThrough{},
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+    }
+};
+
+} // namespace ck

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

@@ -3,18 +3,22 @@
 
 #pragma once
 
-#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
-#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp"
 #ifndef __HIPCC_RTC__
 #include <iostream>
 #endif
 
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v4_direct_load.hpp"
+
 namespace ck {
 
 enum struct PipelineVersion
 {
     v1,
     v2,
+    // v3 is only used in the Stream-K implementation.
+    v4,
 };
 
 template <PipelineVersion PipelineVer,
@@ -37,6 +41,10 @@ constexpr auto GridwiseGemmPipeline_Selector()
     {
         return GridwiseGemmPipeline_v2{};
     }
+    else if constexpr(PipelineVer == PipelineVersion::v4)
+    {
+        return GridwiseGemmPipeline_v4<NumPrefetch>{};
+    }
     else
     {
 #ifndef __HIPCC_RTC__

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp

@@ -4,7 +4,8 @@
 #pragma once
 
 #include "ck/utility/common_header.hpp"
-#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/utility/loop_scheduler.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 
 namespace ck {
 

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp

@@ -9,13 +9,13 @@ namespace ck {
 
 struct GridwiseGemmPipeline_v2
 {
-    __host__ __device__ static constexpr bool IsSupported(index_t num_loop)
+    __host__ __device__ static constexpr bool IsSupported(const index_t num_loop)
     {
         // TODO: improve applicability
         return num_loop % 2 == 0;
     }
 
-    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(const index_t num_loop)
     {
         return (num_loop / 2) > 1;
     }

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v4_direct_load.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/loop_scheduler.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+
+namespace lds_direct_load {
+
+__device__ void sched_barrier()
+{
+#if CK_USE_AMD_LDS_DIRECT_LOAD_INLINE_ASM
+    // When direct loads and `waitcnt` instructions are submitted using inline asm, the usage of
+    // `sched_barrier` is necessary to make sure no instructions that use the loaded memory
+    // are scheduled by the compiler before the `waitcnt` instruction.
+    __builtin_amdgcn_sched_barrier(0);
+#endif
+}
+
+} // namespace lds_direct_load
+
+namespace ck {
+
+template <index_t NumPrefetch>
+struct GridwiseGemmPipeline_v4;
+
+// 1-stage prefetch
+template <>
+struct GridwiseGemmPipeline_v4<1>
+{
+    static constexpr auto I0 = Number<0>{};
+
+    __host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return num_loop > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffers,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffers,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    __device__ static void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffers& a_block_bufs,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffers& b_block_bufs,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        static_assert(ABlockBuffers::Size() == 1 && BBlockBuffers::Size() == 1);
+        auto& a_block_buf = a_block_bufs.At(I0);
+        auto& b_block_buf = b_block_bufs.At(I0);
+
+        a_blockwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf);
+        b_blockwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                block_sync_lds_direct_load();
+                lds_direct_load::sched_barrier();
+
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds_direct_load();
+                lds_direct_load::sched_barrier();
+
+                a_blockwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf);
+                b_blockwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                ++i;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        {
+            block_sync_lds_direct_load();
+            lds_direct_load::sched_barrier();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+// 2-stages prefetch
+template <>
+struct GridwiseGemmPipeline_v4<2>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    __host__ __device__ static constexpr bool IsSupported(index_t num_loop)
+    {
+        return num_loop % 2 == 0;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return (num_loop / 2) > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffers,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffers,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    __device__ static void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffers& a_block_bufs,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffers& b_block_bufs,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        static_assert(ABlockBuffers::Size() == 2 && BBlockBuffers::Size() == 2);
+        auto& a_block_buf1 = a_block_bufs.At(I0);
+        auto& a_block_buf2 = a_block_bufs.At(I1);
+        auto& b_block_buf1 = b_block_bufs.At(I0);
+        auto& b_block_buf2 = b_block_bufs.At(I1);
+
+        a_blockwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf1);
+        b_blockwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf1);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                block_sync_lds_direct_load();
+                lds_direct_load::sched_barrier();
+
+                a_blockwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf2);
+                b_blockwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf2);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                blockwise_gemm.Run(a_block_buf1, b_block_buf1, c_thread_buf);
+
+                block_sync_lds_direct_load();
+                lds_direct_load::sched_barrier();
+
+                a_blockwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf1);
+                b_blockwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf1);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                blockwise_gemm.Run(a_block_buf2, b_block_buf2, c_thread_buf);
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        {
+            block_sync_lds_direct_load();
+            lds_direct_load::sched_barrier();
+
+            a_blockwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf2);
+            b_blockwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf2);
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+            blockwise_gemm.Run(a_block_buf1, b_block_buf1, c_thread_buf);
+
+            block_sync_lds_direct_load();
+            lds_direct_load::sched_barrier();
+
+            blockwise_gemm.Run(a_block_buf2, b_block_buf2, c_thread_buf);
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp

@@ -55,7 +55,7 @@ __global__ void
             const Block2CTileMap block_2_ctile_map)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -457,6 +457,7 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             FloatAB,
+            FloatAB,
             FloatGemmAcc,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),

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

@@ -588,6 +588,7 @@ struct GridwiseGemmSplitKMultipleD_xdl_cshuffle
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             ABDataType,
+            ABDataType,
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -1012,6 +1013,7 @@ struct GridwiseGemmSplitKMultipleD_xdl_cshuffle
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             ABDataType,
+            ABDataType,
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),

include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ADataType, // FIXME: don't assume A/B have same datatype
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_KBatch_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 CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
+    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 AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       AK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       BK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max(a_block_space_size_aligned * sizeof(ADataType) +
+                             b_block_space_size_aligned * sizeof(BDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch)
+    {
+        return math::integer_least_multiple(K, KPerBlock * K_Batch);
+    }
+
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_KBatch_AK0_M_AK1(index_t M, index_t K, index_t StrideA, index_t KBatch)
+    {
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto MPad = CalculateMPadded(M);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            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 AK0 = KPad / (KBatch * AK1);
+
+        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)
+        {
+            // const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_KBatch_BK0_N_BK1(index_t K, index_t N, index_t StrideB, index_t KBatch)
+    {
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto NPad = CalculateNPadded(N);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto BK0 = KPad / (KBatch * BK1);
+
+        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)
+        {
+            // const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_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 e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    template <typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              GemmSpecialization GemmSpec>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const index_t M,
+                  const index_t N,
+                  const index_t K,
+                  const index_t StrideA,
+                  const index_t StrideB,
+                  const std::array<index_t, NumDTensor> StrideDs,
+                  const index_t StrideE,
+                  const index_t KBatch)
+    {
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        ignore = StrideDs;
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+#if 0
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+#endif
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              index_t NumDTensor_,
+              typename DsDataType_,
+              typename AGridDesc_KBatch_AK0_M_AK1,
+              typename BGridDesc_KBatch_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename CDEElementwiseOperation_,
+              typename Block2ETileMap>
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const index_t KBatch,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation_& cde_element_op,
+                               const AGridDesc_KBatch_AK0_M_AK1& a_grid_desc_kbatch_ak0_m_ak1,
+                               const BGridDesc_KBatch_BK0_N_BK1& b_grid_desc_kbatch_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor_>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_kbatch_ak0_m_ak1 =
+            GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_kbatch_bk0_n_bk1 =
+            GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ComputeType,
+                                                decltype(a_grid_desc_kbatch_ak0_m_ak1),
+                                                decltype(a_block_desc_kbatch_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_kbatch_ak0_m_ak1,
+                make_multi_index(kbatch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_kbatch_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BDataType,
+                                                ComputeType,
+                                                decltype(b_grid_desc_kbatch_bk0_n_bk1),
+                                                decltype(b_block_desc_kbatch_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_kbatch_bk0_n_bk1,
+                make_multi_index(kbatch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_kbatch_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // 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();
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ComputeType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+#if 1
+        if(block_work_idx[I0] == 0)
+        {
+            const index_t nThreadSize = CDEShuffleBlockTransferScalarPerVector_NPerBlock;
+            const index_t numNThreads = NPerBlock / nThreadSize;
+            const index_t numMThreads = BlockSize / numNThreads;
+            const index_t mThreadSize = MPerBlock / numMThreads;
+
+            const index_t m_tid = get_thread_local_1d_id() / numNThreads;
+            const index_t n_tid = get_thread_local_1d_id() % numNThreads;
+
+            auto c_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mThreadSize>{}, I1, Number<nThreadSize>{}));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr,
+                         EDataType,
+                         c_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize(),
+                         true>
+                e_thread_zero_buf;
+
+            auto c_thread_copy = ThreadwiseTensorSliceTransfer_v1r3<
+                EDataType,
+                EDataType,
+                decltype(c_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1, mThreadSize, 1, nThreadSize>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{e_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I1],
+                                       m_tid * mThreadSize,
+                                       block_work_idx[I2],
+                                       n_tid * nThreadSize),
+                      ck::tensor_operation::element_wise::PassThrough{}};
+
+            c_thread_copy.Run(c_thread_desc_mblock_mperblock_nblock_nperblock,
+                              make_tuple(I0, I0, I0, I0),
+                              e_thread_zero_buf,
+                              e_grid_desc_mblock_mperblock_nblock_nperblock,
+                              e_grid_buf);
+
+            __syncthreads();
+
+            if(threadIdx.x == 0)
+            {
+                atomicAdd(barrier_count_finished, 1);
+            }
+        }
+#endif
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop =
+            __builtin_amdgcn_readfirstlane((a_grid_desc_kbatch_ak0_m_ak1.GetLength(I1) *
+                                            a_grid_desc_kbatch_ak0_m_ak1.GetLength(I3)) /
+                                           KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_kbatch_ak0_m_ak1,
+                                                               a_block_desc_kbatch_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_kbatch_bk0_n_bk1,
+                                                               b_block_desc_kbatch_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            if(threadIdx.x == 0)
+            {
+                while(__atomic_load_n(barrier_count_finished, __ATOMIC_RELAXED) == 0) {}
+            }
+
+            __syncthreads();
+
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            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.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{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0);
+                    },
+                    Number<NumDTensor_>{}));
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            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/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType_{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation_,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
+                                                                            // Sequence support
+                                                                            // arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor_,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0)),
+                 cde_element_op};
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor_, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+
+            if(threadIdx.x == 0)
+            {
+                index_t k_id_finished_t = atomicAdd(barrier_count_finished, 1);
+
+                if(k_id_finished_t == KBatch)
+                {
+                    *barrier_count_finished = 0;
+                }
+            }
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const index_t KBatch,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ADataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const BDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        if(kbatch_id == KBatch - 1)
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, NumDTensor, DsDataType>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                cde_element_op,
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+        else
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, 0, Tuple<>>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                ck::tensor_operation::element_wise::PassThrough{},
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+    }
+};
+
+} // namespace ck

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

@@ -49,8 +49,7 @@ __global__ void
             const CElementwiseOperation c_element_op,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
-    defined(__gfx1102__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -75,7 +74,7 @@ __global__ void
     ignore = b_element_op;
     ignore = c_element_op;
     ignore = block_2_ctile_map;
-#endif // end of if (defined(__gfx1100__))
+#endif // end of if (defined(__gfx11__))
 }
 
 template <index_t BlockSize,

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

@@ -25,7 +25,7 @@ __global__ void
         kernel_gemm_xdl_cshuffle_v1(typename GridwiseGemm::Argument karg)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(
@@ -50,7 +50,7 @@ __global__ void
                                     typename GridwiseGemm::Problem problem)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid, p_b_grid, p_c_grid, p_shared, problem);
@@ -108,7 +108,8 @@ template <typename ALayout,
           index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
           LoopScheduler LoopSched,
           PipelineVersion PipelineVer = PipelineVersion::v1,
-          typename ComputeType        = FloatC>
+          typename ComputeTypeA       = FloatC,
+          typename ComputeTypeB       = ComputeTypeA>
 struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
 {
     static constexpr auto I0 = Number<0>{};
@@ -547,8 +548,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         constexpr auto c_block_size =
             c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
 
-        return math::max((a_block_space_size_aligned * sizeof(ComputeType) +
-                          b_block_space_size_aligned * sizeof(ComputeType)),
+        return math::max((a_block_space_size_aligned * sizeof(ComputeTypeA) +
+                          b_block_space_size_aligned * sizeof(ComputeTypeB)),
                          c_block_size * sizeof(FloatCShuffle));
     }
 
@@ -750,7 +751,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
                                                 ABlockTransferThreadClusterLengths_AK0_M_AK1,
                                                 ABlockTransferThreadClusterArrangeOrder,
                                                 FloatA,
-                                                ComputeType,
+                                                ComputeTypeA,
                                                 decltype(a_grid_desc_ak0_m_ak1),
                                                 decltype(a_block_desc_ak0_m_ak1),
                                                 ABlockTransferSrcAccessOrder,
@@ -781,7 +782,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
                                                 BBlockTransferThreadClusterLengths_BK0_N_BK1,
                                                 BBlockTransferThreadClusterArrangeOrder,
                                                 FloatB,
-                                                ComputeType,
+                                                ComputeTypeB,
                                                 decltype(b_grid_desc_bk0_n_bk1),
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
@@ -809,13 +810,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
         //       register
         // sanity check
-        constexpr index_t KPack =
-            math::max(math::lcm(AK1Number, BK1Number),
-                      MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1Number, BK1Number),
+            MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>::selected_mfma.k_per_blk);
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            ComputeType,
+            ComputeTypeA,
+            ComputeTypeB,
             FloatGemmAcc,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -833,10 +835,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
             a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+            static_cast<ComputeTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
+            static_cast<ComputeTypeB*>(p_shared) + a_block_space_size_aligned,
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm, bool HasMainKBlockLoop, index_t TailNum = 3>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, 1)
+#endif
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_gemm_xdl_cshuffle_v2(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx94__))
+    // 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()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, TailNum>(
+        karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, p_shared_0, p_shared_1, karg);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename GridwiseGemm,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, 1)
+#endif
+        kernel_gemm_xdl_cshuffle_v2(const FloatA* p_a_grid,
+                                    const FloatB* p_b_grid,
+                                    FloatC* p_c_grid,
+                                    typename GridwiseGemm::Problem problem)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx94__))
+    __shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    __shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        p_a_grid, p_b_grid, p_c_grid, p_shared_0, p_shared_1, problem);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = problem;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename FloatA,
+          typename FloatB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeTypeA       = FloatC,
+          typename ComputeTypeB       = ComputeTypeA>
+struct GridwiseGemm_xdl_cshuffle_v2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0Number = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0Number = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1Number = Number<AK1Value>{};
+    static constexpr auto BK1Number = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, 1);
+    }
+
+    __host__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock) * MPerBlock;
+    }
+
+    __host__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock) * NPerBlock;
+    }
+
+    __host__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ static auto CalculateAK0(index_t K)
+    {
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding ||
+                     GemmSpec == GemmSpecialization::KPadding ||
+                     GemmSpec == GemmSpecialization::NKPadding)
+        {
+            return CalculateKPadded(K) / AK1Value;
+        }
+        else
+        {
+            return K / AK1Value;
+        }
+    }
+
+    __host__ static auto CalculateBK0(index_t K)
+    {
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding ||
+                     GemmSpec == GemmSpecialization::KPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding)
+        {
+            return CalculateKPadded(K) / BK1Value;
+        }
+        else
+        {
+            return K / BK1Value;
+        }
+    }
+
+    __host__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_floor(M, MPerBlock);
+    }
+
+    __host__ static auto CalculateNBlock(index_t N)
+    {
+        return math::integer_divide_floor(N, NPerBlock);
+    }
+
+    template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+    __host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&)
+    {
+        constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
+        constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
+
+        return transform_tensor_descriptor(
+            TileDesc_K0_MN_K1{},
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
+                       make_unmerge_transform(make_tuple(
+                           Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
+        index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(M, MPad - M),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(MPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
+        index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(N, NPad - N),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(NPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor<MXdlPerWave, MWaves, MPerXdl>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor<NXdlPerWave, NWaves, NPerXdl>(BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static auto
+    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(M, MPad - M),
+                                                          make_right_pad_transform(N, NPad - N)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    struct Problem
+    {
+        __host__ Problem(index_t M_,
+                         index_t N_,
+                         index_t K_,
+                         index_t StrideA_,
+                         index_t StrideB_,
+                         index_t StrideC_)
+            : M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideC{StrideC_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              KPadded{CalculateKPadded(K_)},
+              AK0{CalculateAK0(K_)},
+              BK0{CalculateBK0(K_)},
+              MBlock{CalculateMBlock(M_)},
+              NBlock{CalculateNBlock(N_)}
+        {
+        }
+
+        __host__ void Print() const
+        {
+            std::cout << "problem {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "KP:" << KPadded << ", "
+                      << "AK0:" << AK0 << ", "
+                      << "BK0:" << BK0 << ", "
+                      << "MBlock: " << MBlock << ", "
+                      << "NBlock: " << NBlock << "}" << std::endl;
+        }
+
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        index_t StrideC;
+        index_t MPadded;
+        index_t NPadded;
+        index_t KPadded;
+        index_t AK0;
+        index_t BK0;
+        index_t MBlock;
+        index_t NBlock;
+    };
+
+    // Argument
+    struct Argument : public tensor_operation::device::BaseArgument, public Problem
+    {
+        __host__ Argument(const FloatA* p_a_grid_,
+                          const FloatB* p_b_grid_,
+                          FloatC* p_c_grid_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          index_t StrideC_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_c_grid{p_c_grid_}
+        {
+        }
+
+        const FloatA* p_a_grid;
+        const FloatB* p_b_grid;
+        FloatC* p_c_grid;
+    };
+
+    // FIXME: pass GridwiseGemmPipe as a template arguement into GridwiseGemm
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1Number, AK1Number, I1));
+    }
+
+    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1Number, BK1Number, I1));
+    }
+
+    __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned * sizeof(ComputeTypeA) +
+                          b_block_space_size_aligned * sizeof(ComputeTypeB)),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ static constexpr bool CheckValidity(const Problem& problem)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(problem.M % MPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(problem.N % NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding)
+        {
+            if(!(CalculateKPadded(problem.K) % AK1Value == 0) ||
+               !(CalculateKPadded(problem.K) % BK1Value == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(problem.K % AK1Value == 0) || !(problem.K % BK1Value == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(problem.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(problem.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(problem.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                return false;
+            }
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = (CalculateAK0(problem.K) * AK1Value) / KPerBlock;
+
+        if(num_k_loop < 4)
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return num_loop > 3;
+    }
+
+    __host__ static constexpr index_t CalculateKBlockLoopTailNum(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        if(num_loop % 2 == 1)
+            return 3;
+        else
+            return 2;
+    }
+
+    template <typename CGridDesc>
+    __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    // if arch = gfx942
+    using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
+
+    template <bool HasMainKBlockLoop, index_t TailNum = 3>
+    __device__ static void Run(const FloatA* p_a_grid,
+                               const FloatB* p_b_grid,
+                               FloatC* 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_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
+            problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
+            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                c_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const AElementwiseOperation a_element_op{};
+        const BElementwiseOperation b_element_op{};
+        const CElementwiseOperation c_element_op{};
+
+        // divide block work by [M, N]
+        const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 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;
+        }
+#if 0
+        if(threadIdx.x == 0){
+            printf("Hardware assigned No. %03d workgroup of logical C tile (%02d, %02d) on %d th XCC Die, %d th SE, %d th CU\n",
+               get_block_1d_id(),
+               block_work_idx[I0],
+               block_work_idx[I1],
+               __smid()>>6 & 0xf,
+               __smid()>>4 & 0x3,
+               __smid() & 0xf);
+        }
+#endif
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0Number, MPerBlock, AK1Number>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatA,
+                                                ComputeTypeA,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0Number, NPerBlock, BK1Number>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatB,
+                                                ComputeTypeB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1Number, BK1Number),
+                      MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        // auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+        //     BlockSize,
+        //     ComputeType,
+        //     FloatGemmAcc,
+        //     decltype(a_block_desc_ak0_m_ak1),
+        //     decltype(b_block_desc_bk0_n_bk1),
+        //     MPerXdl,
+        //     NPerXdl,
+        //     MXdlPerWave,
+        //     NXdlPerWave,
+        //     KPack,
+        //     LoopSched>();
+        auto blockwise_gemm_pipeline = BlockwiseGemmXdlops_pipeline_v4<
+            BlockSize,
+            ComputeTypeA,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack>{}; // TransposeC
+
+        auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeA*>(p_shared_0), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeB*>(p_shared_0) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeA*>(p_shared_1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeTypeB*>(p_shared_1) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong);
+        auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong);
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 0);
+
+        // gridwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<GridwiseGemmPipe>);
+        // const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
+                                                                         a_block_desc_ak0_m_ak1,
+                                                                         a_blockwise_copy,
+                                                                         a_grid_buf,
+                                                                         a_block_bufs,
+                                                                         a_block_slice_copy_step,
+                                                                         b_grid_desc_bk0_n_bk1,
+                                                                         b_block_desc_bk0_n_bk1,
+                                                                         b_blockwise_copy,
+                                                                         b_grid_buf,
+                                                                         b_block_bufs,
+                                                                         b_block_slice_copy_step,
+                                                                         c_thread_buf,
+                                                                         num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(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<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp

@@ -58,7 +58,7 @@ __global__ void
             const Block2CTileMap block_2_ctile_map)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     // TODO ANT: separate into MMA + Epilogue
@@ -495,6 +495,7 @@ struct GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             FloatAB,
+            FloatAB,
             FloatGemmAcc,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),

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

@@ -494,6 +494,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle
             auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<
                 TileMathThreadGroupSize,
                 ABDataType,
+                ABDataType,
                 FloatGemmAcc,
                 decltype(a_block_desc_ak0_m_ak1),
                 decltype(b_block_desc_bk0_n_bk1),

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

@@ -139,7 +139,8 @@ __host__ __device__ constexpr auto make_merge_transform_v4_no_carry(const LowLen
 }
 
 template <typename GridwiseGemm,
-          typename FloatAB,
+          typename FloatA,
+          typename FloatB,
           typename FloatC,
           typename AGridDesc_B_K0_M_K1,
           typename BGridDesc_B_K0_N_K1,
@@ -153,8 +154,8 @@ __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_gemm_xdlops_bwd_weight(const FloatAB* __restrict__ p_a_grid,
-                                      const FloatAB* __restrict__ p_b_grid,
+        kernel_gemm_xdlops_bwd_weight(const FloatA* __restrict__ p_a_grid,
+                                      const FloatB* __restrict__ p_b_grid,
                                       FloatC* __restrict__ p_c_grid,
                                       const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
                                       const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
@@ -166,7 +167,7 @@ __global__ void
                                       const CBlockClusterAdaptor c_block_cluster_adaptor)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -181,21 +182,22 @@ __global__ void
                                                   c_element_op,
                                                   c_block_cluster_adaptor);
 #else
-    ignore                = p_a_grid;
-    ignore                = p_b_grid;
-    ignore                = p_c_grid;
-    ignore                = a_b_k0_m_k1_grid_desc;
-    ignore                = b_b_k0_n_k1_grid_desc;
-    ignore                = c_grid_desc_mblock_mperblock_nblock_nperblock;
-    ignore                = a_element_op;
-    ignore                = b_element_op;
-    ignore                = c_element_op;
-    ignore                = c_block_cluster_adaptor;
+    ignore               = p_a_grid;
+    ignore               = p_b_grid;
+    ignore               = p_c_grid;
+    ignore               = a_b_k0_m_k1_grid_desc;
+    ignore               = b_b_k0_n_k1_grid_desc;
+    ignore               = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore               = a_element_op;
+    ignore               = b_element_op;
+    ignore               = c_element_op;
+    ignore               = c_block_cluster_adaptor;
 #endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
 template <index_t BlockSize,
-          typename FloatAB,
+          typename FloatA,
+          typename FloatB,
           typename FloatAcc,
           typename FloatC,
           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
@@ -242,7 +244,9 @@ template <index_t BlockSize,
           bool ABlockLdsExtraM1Wrw      = false,
           bool BBlockLdsExtraN1Wrw      = false,
           index_t NumGemmKPrefetchStage = 1,
-          PipelineVersion PipelineVer   = PipelineVersion::v1>
+          PipelineVersion PipelineVer   = PipelineVersion::v1,
+          typename ComputeTypeA         = FloatA,
+          typename ComputeTypeB         = ComputeTypeA>
 struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
 {
     static constexpr auto I0 = Number<0>{};
@@ -265,11 +269,16 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
     // denorm test fix, required to work around fp16 mfma issue
     // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
     // when mfma if fixed, remove this section and update
-    // FloatABAdjusted -> FloatAB throughout this file
+    // FloatAAdjusted -> ComputeTypeA, FloatBAdjusted -> ComputeTypeB,
+    // throughout this file
 #if CK_WORKAROUND_DENORM_FIX
-    using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
+    using FloatAAdjusted =
+        conditional_t<is_same_v<ComputeTypeA, ck::half_t>, ck::bhalf_t, ComputeTypeA>;
+    using FloatBAdjusted =
+        conditional_t<is_same_v<ComputeTypeB, ck::half_t>, ck::bhalf_t, ComputeTypeB>;
 #else
-    using FloatABAdjusted = FloatAB;
+    using FloatAAdjusted = ComputeTypeA;
+    using FloatBAdjusted = ComputeTypeB;
 #endif
 
     // M0/M1/M1Padding
@@ -506,7 +515,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
         constexpr auto c_block_size =
             GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock().GetElementSpaceSize();
 
-        return math::max((a_block_space_size + b_block_space_size) * sizeof(FloatAB),
+        return math::max((a_block_space_size * sizeof(FloatAAdjusted) +
+                          b_block_space_size * sizeof(FloatBAdjusted)),
                          c_block_size * sizeof(FloatC));
     }
 
@@ -610,8 +620,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
     using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
 
     template <bool HasMainKBlockLoop>
-    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
-                               const FloatAB* __restrict__ p_b_grid,
+    __device__ static void Run(const FloatA* __restrict__ p_a_grid,
+                               const FloatB* __restrict__ p_b_grid,
                                FloatC* __restrict__ p_c_grid,
                                void* __restrict__ p_shared,
                                const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
@@ -673,8 +683,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
                                                 Sequence<1, K0PerBlock, MPerBlock, K1>,
                                                 ABlockTransferThreadClusterLengths_K0_M_K1,
                                                 ABlockTransferThreadClusterArrangeOrder,
-                                                FloatAB,
-                                                FloatABAdjusted,
+                                                FloatA,
+                                                FloatAAdjusted,
                                                 decltype(a_b_k0_m_k1_grid_desc),
                                                 decltype(a_b_k0_m_k1_block_desc),
                                                 ABlockTransferSrcAccessOrder,
@@ -703,8 +713,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
                                                 Sequence<1, K0PerBlock, NPerBlock, K1>,
                                                 BBlockTransferThreadClusterLengths_K0_N_K1,
                                                 BBlockTransferThreadClusterArrangeOrder,
-                                                FloatAB,
-                                                FloatABAdjusted,
+                                                FloatB,
+                                                FloatBAdjusted,
                                                 decltype(b_b_k0_n_k1_grid_desc),
                                                 decltype(b_b_k0_n_k1_block_desc),
                                                 BBlockTransferSrcAccessOrder,
@@ -733,11 +743,14 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
         // sanity check
 
         constexpr index_t KPack =
-            math::max(K1, MfmaSelector<FloatABAdjusted, MPerXDL, NPerXDL>::selected_mfma.k_per_blk);
+            math::max(K1,
+                      MfmaSelector<FloatAAdjusted, MPerXDL, NPerXDL, FloatBAdjusted>::selected_mfma
+                          .k_per_blk);
 
         auto blockwise_gemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
-                                                                FloatABAdjusted,
+                                                                FloatAAdjusted,
+                                                                FloatBAdjusted,
                                                                 FloatAcc,
                                                                 decltype(a_k0_m_k1_block_desc),
                                                                 decltype(b_k0_n_k1_block_desc),
@@ -757,10 +770,10 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
         constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<FloatABAdjusted*>(p_shared), a_k0_m_k1_block_desc.GetElementSpaceSize());
+            static_cast<FloatAAdjusted*>(p_shared), a_k0_m_k1_block_desc.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<FloatABAdjusted*>(p_shared) + a_block_space_size,
+            static_cast<FloatBAdjusted*>(p_shared) + a_block_space_size,
             b_k0_n_k1_block_desc.GetElementSpaceSize());
 
         // gridwise GEMM pipeline

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp

@@ -45,7 +45,7 @@ __global__ void
             const Block2CTileMap block_2_ctile_map)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    defined(__gfx94__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,