Merge remote-tracking branch 'origin/develop' into aosewski/gemm_tile_loop

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

@@ -361,15 +361,19 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
     }
 
     // desc for problem definition
-    using AGridDesc_M_K  = remove_cvref_t<decltype(
-        MakeAGridDescriptor_M_K<ALayout>({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using AGridDesc_M_K  = remove_cvref_t<decltype(MakeAGridDescriptor_M_K<ALayout>(
+        {}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
     using BGridDesc_N_K  = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
     using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
     using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
 
+    using ComputeDataType = ADataType;
+
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,
@@ -412,14 +416,18 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
         LoopSched>;
 
     // desc for blockwise copy
-    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
-        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
-    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
-        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
-    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
-    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
-        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
 
     // block-to-e-tile map
     using Block2ETileMap =

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

@@ -228,9 +228,13 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
     using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
     using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
 
+    using ComputeDataType = ADataType;
+
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,
@@ -272,14 +276,18 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
         CDEBlockTransferScalarPerVector_NPerBlock,
         LoopSched>;
 
-    using AGridDesc_AK0_M_AK1                          = remove_cvref_t<decltype(
-        GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
-    using BGridDesc_BK0_N_BK1                          = remove_cvref_t<decltype(
-        GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
-    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
-    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
-        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
 
     struct GroupedGemmBlock2ETileMap
     {

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GemmDesc,
+          GemmSpecialization GemmSpec,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename DsDataType,
+          typename Block2ETileMap,
+          typename GroupedGemmBlock2ETileMap,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_xdl_fixed_nk(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                         uint32_t* barrier_count,
+                                         const index_t barrier_size_grp,
+                                         const index_t group_count,
+                                         const index_t grid_size_grp,
+                                         const index_t KBatch,
+                                         const AElementwiseOperation a_element_op,
+                                         const BElementwiseOperation b_element_op,
+                                         const CDEElementwiseOperation c_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto gemm_desc_ptr =
+        reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
+
+    const index_t group_id = block_id / grid_size_grp;
+
+    if(group_id >= group_count)
+        return;
+
+    const index_t M = gemm_desc_ptr[group_id].M;
+    const index_t N = gemm_desc_ptr[group_id].N;
+    const index_t K = gemm_desc_ptr[group_id].K;
+
+    if(M * N * K == 0)
+        return;
+
+    const auto StrideA  = gemm_desc_ptr[group_id].StrideA;
+    const auto StrideB  = gemm_desc_ptr[group_id].StrideB;
+    const auto StrideDs = gemm_desc_ptr[group_id].StrideDs;
+    const auto StrideE  = gemm_desc_ptr[group_id].StrideE;
+
+    const auto e_grid_desc_m_n =
+        GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+    const index_t BlockStart = group_id * grid_size_grp;
+
+    const auto local_b2e_tile_map = Block2ETileMap{e_grid_desc_m_n, KBatch};
+
+    const auto local_grid_size = local_b2e_tile_map.CalculateGridSize(e_grid_desc_m_n);
+
+    constexpr auto NumDTensor = DsDataType::Size();
+
+    using DsGridPointer = decltype(GridwiseGemm::MakeDsGridPointer());
+
+    DsGridPointer p_ds_grid_;
+
+    static_for<0, NumDTensor, 1>{}([&](auto i) {
+        using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+        // D pointer
+        p_ds_grid_(i) = static_cast<const DDataType*>(gemm_desc_ptr[group_id].p_ds_grid[i]);
+    });
+
+    index_t id_off   = 0;
+    index_t id_local = get_block_1d_id() - BlockStart;
+
+    const index_t mn_blocks = local_grid_size / KBatch;
+
+    while(id_local < local_grid_size)
+    {
+        const auto block_2_etile_map =
+            GroupedGemmBlock2ETileMap(local_b2e_tile_map, BlockStart, id_off);
+
+        auto barrier_count_finished =
+            barrier_count + group_id * barrier_size_grp + id_local % mn_blocks;
+
+        GridwiseGemm::template Run<HasMainKBlockLoop,
+                                   EGlobalMemoryDataOperation,
+                                   GemmSpec,
+                                   ALayout,
+                                   BLayout,
+                                   DsLayout,
+                                   ELayout>(gemm_desc_ptr[group_id].p_a_grid,
+                                            gemm_desc_ptr[group_id].p_b_grid,
+                                            p_ds_grid_,
+                                            gemm_desc_ptr[group_id].p_e_grid,
+                                            p_shared,
+                                            barrier_count_finished,
+                                            a_element_op,
+                                            b_element_op,
+                                            c_element_op,
+                                            M,
+                                            N,
+                                            K,
+                                            StrideA,
+                                            StrideB,
+                                            StrideDs,
+                                            StrideE,
+                                            KBatch,
+                                            block_2_etile_map);
+
+        id_off += grid_size_grp;
+        id_local += grid_size_grp;
+    }
+#else
+    ignore = gemm_descs_const;
+    ignore = barrier_count;
+    ignore = barrier_size_grp;
+    ignore = group_count;
+    ignore = grid_size_grp;
+    ignore = KBatch;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t NumPrefetch,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t KPerBlock,
+          ck::index_t AK1,
+          ck::index_t BK1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
+                                                                        BLayout,
+                                                                        DsLayout,
+                                                                        ELayout,
+                                                                        ADataType,
+                                                                        BDataType,
+                                                                        DsDataType,
+                                                                        EDataType,
+                                                                        AElementwiseOperation,
+                                                                        BElementwiseOperation,
+                                                                        CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedGemm_Xdl_Fixed_NK;
+
+    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>{};
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_splitk_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        NumPrefetch, // NumGemmKPrefetchStage
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    template <typename UnderlyingBlockToCTileMap>
+    struct OffsettedBlockToCTileMapMLoops
+    {
+        using underlying_type = UnderlyingBlockToCTileMap;
+
+        __host__ __device__ OffsettedBlockToCTileMapMLoops(
+            UnderlyingBlockToCTileMap block_to_ctile_map, index_t block_start, index_t id_off = 0)
+        {
+            block_to_ctile_map_ = block_to_ctile_map;
+            block_start_        = block_start;
+            id_off_             = id_off;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            auto idx_bot = block_to_ctile_map_.CalculateBottomIndex(
+                make_multi_index(idx_top[Number<0>{}] - block_start_ + id_off_));
+
+            return make_tuple(idx_bot[Number<0>{}], idx_bot[Number<1>{}], idx_bot[Number<2>{}]);
+        }
+
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                 const CTileDim& c_tile_dim) const
+        {
+            return block_to_ctile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+        {
+            return block_to_ctile_map_.CheckValidity(c_grid_desc_m_n);
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+        {
+            return block_to_ctile_map_.CalculateGridSize(c_grid_desc_m_n);
+        }
+
+        UnderlyingBlockToCTileMap block_to_ctile_map_;
+        index_t block_start_;
+        index_t id_off_;
+    };
+
+    template <index_t MPerBlock_, index_t NPerBlock_>
+    struct BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops
+    {
+        static constexpr auto I0 = Number<0>{};
+        static constexpr auto I1 = Number<1>{};
+
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops() = default;
+
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
+            const BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&) = default;
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
+            BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&&) = default;
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&
+        operator=(const BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&) = default;
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&
+        operator=(BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&&) = default;
+
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(index_t M,
+                                                                          index_t N,
+                                                                          index_t KBatch,
+                                                                          index_t M01 = 8)
+            : M_(M), N_(N), KBatch_(KBatch), M01_(M01)
+        {
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
+            const CGridDesc_M_N& c_grid_desc_m_n, index_t KBatch, index_t M01 = 8)
+            : BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
+                  c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1), KBatch, M01)
+        {
+        }
+
+        __host__ __device__ constexpr index_t CalculateGridSize(index_t M, index_t N) const
+        {
+            const auto M0 = math::integer_divide_ceil(M, MPerBlock);
+            const auto N0 = math::integer_divide_ceil(N, NPerBlock);
+
+            return M0 * N0 * KBatch_;
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ __device__ constexpr index_t
+        CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+        {
+            return CalculateGridSize(c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1));
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
+        {
+            return true;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            auto block_1d_id = idx_top[I0];
+
+            const auto M0 = math::integer_divide_ceil(M_, MPerBlock_);
+            const auto N0 = math::integer_divide_ceil(N_, NPerBlock_);
+
+            block_1d_id = block_1d_id % (M0 * N0 * KBatch_); // hide groups
+
+            const index_t idx_ksplit = block_1d_id / (M0 * N0);
+            block_1d_id              = block_1d_id % (M0 * N0);
+
+            index_t idx_N0 = block_1d_id % N0;
+            index_t idx_M0 = block_1d_id / N0;
+
+            const auto M01_adapt = (idx_M0 < M0 - M0 % M01_) ? M01_ : M0 % M01_;
+
+            index_t idx_M00          = idx_M0 / M01_;
+            index_t idx_M01          = idx_M0 % M01_;
+            index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
+
+            return make_tuple(idx_ksplit,
+                              idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
+                              idx_N0_M01_local / M01_adapt);
+        }
+
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
+                                                 const CTileDim& /* c_tile_dim */) const
+        {
+            return true; // always valid provided that user gets grid size from CalculateGridSize()
+        }
+
+        private:
+        index_t M_;
+        index_t N_;
+        index_t KBatch_;
+        index_t M01_;
+    };
+
+    using Block2ETileMap = BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops<MPerBlock, NPerBlock>;
+    using GroupedGemmBlock2ETileMap = OffsettedBlockToCTileMapMLoops<Block2ETileMap>;
+
+    struct GemmBiasTransKernelArg
+    {
+        // pointers
+        const void* a_ptr_;
+        const void* b_ptr_;
+        std::array<const void*, NumDTensor> ds_ptr_;
+        void* e_ptr_;
+
+        index_t M_, N_, K_;
+        index_t StrideA_, StrideB_;
+        std::array<index_t, NumDTensor> StrideDs_;
+        index_t StrideE_;
+    };
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+
+        void UpdateKBatch(index_t k_batch)
+        {
+            k_batch_ = k_batch;
+
+            if(k_batch_ < 1)
+            {
+
+                throw std::runtime_error("wrong! k_batch must be > 0");
+            }
+
+            const index_t AverM = math::integer_divide_ceil(sum_of_m, group_count_);
+
+            const index_t StrideE = gemm_desc_kernel_arg_[0].StrideE_;
+            const index_t N       = gemm_desc_kernel_arg_[0].N_;
+
+            const auto e_grid_desc_m_n =
+                GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
+                    AverM, N, StrideE);
+
+            const auto local_b2c_tile_map = Block2ETileMap{e_grid_desc_m_n, k_batch_};
+
+            grid_size_grp_ = local_b2c_tile_map.CalculateGridSize(e_grid_desc_m_n);
+
+            grid_size_ = grid_size_grp_ * group_count_;
+        }
+
+        Argument(std::vector<const void*>&,
+                 std::vector<const void*>&,
+                 std::vector<std::array<const void*, NumDTensor>>&,
+                 std::vector<void*>&,
+                 std::vector<GemmDesc>& gemm_descs,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation c_element_op)
+            : a_element_op_{a_element_op}, b_element_op_{b_element_op}, c_element_op_{c_element_op}
+        {
+            grid_size_ = 0;
+
+            k_batch_ = 1;
+
+            grouped_gemm_kernel_args_dev = nullptr;
+
+            group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
+
+            gemm_desc_kernel_arg_.reserve(group_count_);
+
+            index_t group_id = 0;
+
+            sum_of_m            = gemm_descs[0].M_;
+            const index_t AverM = math::integer_divide_ceil(sum_of_m, group_count_);
+            const index_t N     = gemm_descs[0].N_;
+            const index_t K     = gemm_descs[0].K_;
+
+            for(std::size_t i = 0; i < gemm_descs.size(); i++)
+            {
+                if(sum_of_m != gemm_descs[i].M_ || N != gemm_descs[i].N_ || K != gemm_descs[i].K_)
+                {
+                    throw std::runtime_error("wrong! M/N/K is not identical");
+                }
+
+                a_mtx_mraw_kraw_.emplace_back(sum_of_m, K);
+                b_mtx_nraw_kraw_.emplace_back(N, K);
+
+                const index_t StrideA = gemm_descs[i].stride_A_;
+                const index_t StrideB = gemm_descs[i].stride_B_;
+                const index_t StrideE = gemm_descs[i].stride_C_;
+
+                // pointer
+                std::array<const void*, NumDTensor> p_ds_grid;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) { p_ds_grid[j] = nullptr; });
+
+                std::array<index_t, NumDTensor> StrideDs;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    // using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+                    if(gemm_descs[i].stride_Ds_.size() != NumDTensor)
+                    {
+                        throw std::runtime_error(
+                            "wrong! gemm_descs[i].stride_Ds_.size() does not match NumDTensor");
+                    }
+
+                    StrideDs[j] = gemm_descs[i].stride_Ds_[j];
+                });
+
+                const auto e_grid_desc_m_n =
+                    GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
+                        AverM, N, StrideE);
+
+                // block-to-e-tile map
+                const auto local_b2c_tile_map = Block2ETileMap{e_grid_desc_m_n, k_batch_};
+
+                grid_size_grp_ = local_b2c_tile_map.CalculateGridSize(e_grid_desc_m_n);
+
+                if(group_id * grid_size_grp_ != grid_size_)
+                {
+                    throw std::runtime_error("wrong! grid_size_grp_ is not identical!");
+                }
+
+                grid_size_ += grid_size_grp_;
+
+                // check block-to-E-tile
+                if(!local_b2c_tile_map.CheckValidity(e_grid_desc_m_n))
+                {
+                    throw std::runtime_error("wrong! block_2_etile_map validation failed");
+                }
+
+                if(!GridwiseGemm::
+                       template CheckValidity<ALayout, BLayout, DsLayout, ELayout, GemmSpec>(
+                           AverM, N, K, StrideA, StrideB, StrideDs, StrideE, 1))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+                }
+
+                gemm_desc_kernel_arg_.push_back(GemmBiasTransKernelArg{
+                    nullptr,
+                    nullptr,
+                    p_ds_grid,
+                    nullptr,
+                    AverM,
+                    N,
+                    K,
+                    StrideA,
+                    StrideB,
+                    StrideDs,
+                    StrideE,
+                });
+
+                group_id++;
+            }
+
+            const auto e_grid_desc_sum_m_n =
+                GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
+                    sum_of_m, gemm_desc_kernel_arg_[0].N_, gemm_desc_kernel_arg_[0].StrideE_);
+
+            const auto local_b2c_tile_map = Block2ETileMap{e_grid_desc_sum_m_n, 1};
+
+            barrier_size_grp_ = local_b2c_tile_map.CalculateGridSize(e_grid_desc_sum_m_n);
+        }
+
+        //  private:
+        index_t group_count_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation c_element_op_;
+
+        std::vector<GemmBiasTransKernelArg> gemm_desc_kernel_arg_;
+        std::vector<Tuple<index_t, index_t>> a_mtx_mraw_kraw_;
+        std::vector<Tuple<index_t, index_t>> b_mtx_nraw_kraw_;
+
+        const void* grouped_gemm_kernel_args_dev;
+
+        index_t grid_size_;
+        index_t grid_size_grp_;
+        index_t barrier_size_grp_;
+        index_t sum_of_m;
+
+        index_t k_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            bool has_main_k_block_loop = true;
+
+            for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
+            {
+                const auto KPad =
+                    GridwiseGemm::CalculateKPadded(arg.gemm_desc_kernel_arg_[i].K_, arg.k_batch_);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(KPad) != has_main_k_block_loop)
+                {
+                    throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
+                }
+            }
+
+            if(arg.grouped_gemm_kernel_args_dev == nullptr)
+            {
+                throw std::runtime_error("wrong! grouped_gemm_kernel_args_dev is nullpr");
+            }
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_, auto e_global_memory_operation_) {
+                const auto kernel =
+                    kernel_grouped_gemm_xdl_fixed_nk<GridwiseGemm,
+                                                     GroupedGemmKernelArgument<NumDTensor>,
+                                                     GemmSpec,
+                                                     ALayout,
+                                                     BLayout,
+                                                     DsLayout,
+                                                     ELayout,
+                                                     DsDataType,
+                                                     Block2ETileMap,
+                                                     GroupedGemmBlock2ETileMap,
+                                                     AElementwiseOperation,
+                                                     BElementwiseOperation,
+                                                     CDEElementwiseOperation,
+                                                     e_global_memory_operation_,
+                                                     has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.grouped_gemm_kernel_args_dev),
+                    reinterpret_cast<uint32_t*>(arg.p_workspace_),
+                    arg.barrier_size_grp_,
+                    arg.gemm_desc_kernel_arg_.size(),
+                    arg.grid_size_grp_,
+                    arg.k_batch_,
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.c_element_op_);
+            };
+
+            constexpr auto AtomicAdd = InMemoryDataOperationEnum::AtomicAdd;
+            constexpr auto Set       = InMemoryDataOperationEnum::Set;
+
+            if(arg.k_batch_ > 1)
+            {
+                if(has_main_k_block_loop)
+                {
+                    ave_time =
+                        launch_kernel(integral_constant<bool, true>{},
+                                      integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
+                }
+                else
+                {
+                    ave_time =
+                        launch_kernel(integral_constant<bool, false>{},
+                                      integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
+                }
+            }
+            else
+            {
+                if(has_main_k_block_loop)
+                {
+                    ave_time = launch_kernel(integral_constant<bool, true>{},
+                                             integral_constant<InMemoryDataOperationEnum, Set>{});
+                }
+                else
+                {
+                    ave_time = launch_kernel(integral_constant<bool, false>{},
+                                             integral_constant<InMemoryDataOperationEnum, Set>{});
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::type_convert<ck::index_t>(arg.gemm_desc_kernel_arg_.size()) != arg.group_count_)
+        {
+            return false;
+        }
+
+        bool supported = true;
+
+        // If we use padding we do not support vector loads for dimensions not divisible by vector
+        // load size.
+        if constexpr(GemmSpec != GemmSpecialization::Default)
+        {
+            // [A|B]BlockTransferSrcVectorDim value define dimension in the block {K0,M,K1} layout,
+            // thus we have to adapt it to the {M,K} or {N,K} layout.
+            const auto a_raw_vector_dim = ABlockTransferSrcVectorDim != 1 ? 1 : 0;
+            const auto b_raw_vector_dim = BBlockTransferSrcVectorDim != 1 ? 1 : 0;
+
+            for(index_t i = 0; i < arg.group_count_; ++i)
+            {
+                const auto a_vector_dim = arg.a_mtx_mraw_kraw_[i].At(Number<a_raw_vector_dim>{});
+                const auto b_vector_dim = arg.b_mtx_nraw_kraw_[i].At(Number<b_raw_vector_dim>{});
+
+                supported = supported & (a_vector_dim % ABlockTransferSrcScalarPerVector == 0);
+                supported = supported & (b_vector_dim % BBlockTransferSrcScalarPerVector == 0);
+            }
+        }
+
+        return supported;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*>& p_As,
+                             std::vector<const void*>& p_Bs,
+                             std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                             std::vector<void*>& p_Es,
+                             std::vector<GemmDesc> gemm_descs,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation c_element_op)
+    {
+        return Argument{
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_As,
+                        std::vector<const void*>& p_Bs,
+                        std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                        std::vector<void*>& p_Es,
+                        std::vector<GemmDesc>& gemm_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemm_Xdl_Fixed_NK"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    static void SetDeviceKernelArgs(Argument& arg, const void* kernel_args)
+    {
+        arg.grouped_gemm_kernel_args_dev = kernel_args;
+    }
+
+    // polymorphic
+    void SetDeviceKernelArgs(BaseArgument* p_arg, const void* kernel_args) const override
+    {
+        return SetDeviceKernelArgs(*dynamic_cast<Argument*>(p_arg), kernel_args);
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        auto arg = *dynamic_cast<const Argument*>(p_arg);
+
+        return arg.group_count_ * arg.barrier_size_grp_ * sizeof(uint32_t);
+    }
+
+    size_t GetDeviceKernelArgSize(const BaseArgument* p_arg) const override
+    {
+        auto arg = *dynamic_cast<const Argument*>(p_arg);
+
+        return arg.group_count_ * sizeof(GroupedGemmKernelArgument<NumDTensor>);
+    }
+
+    void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const override
+    {
+        auto p_arg_          = dynamic_cast<Argument*>(p_arg);
+        p_arg_->p_workspace_ = p_workspace;
+
+        hip_check_error(hipMemset(p_workspace, 0, GetWorkSpaceSize(p_arg)));
+    }
+
+    static void SetKBatch(Argument& arg, index_t k_batch) { arg.UpdateKBatch(k_batch); }
+
+    // polymorphic
+    void SetKBatch(BaseArgument* p_arg, index_t k_batch) const override
+    {
+        return SetKBatch(*dynamic_cast<Argument*>(p_arg), k_batch);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_image_to_column.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InputGridDesc,
+          typename InputDataType,
+          typename OutputGridDesc,
+          typename OutputDataType,
+          typename Block2ETileMap,
+          typename GridwiseImageToColumnKernel>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_image_to_column(const InputGridDesc in_grid_desc,
+                               const InputDataType* __restrict__ p_in_global,
+                               const OutputGridDesc out_grid_desc,
+                               OutputDataType* __restrict__ p_out_global,
+                               const Block2ETileMap block_2_tile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||             \
+    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
+    defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx941__) || defined(__gfx942__))
+    GridwiseImageToColumnKernel::Run(
+        in_grid_desc, p_in_global, out_grid_desc, p_out_global, block_2_tile_map);
+#else
+    ignore = in_grid_desc;
+    ignore = p_in_global;
+    ignore = out_grid_desc;
+    ignore = p_out_global;
+    ignore = block_2_tile_map;
+#endif
+}
+
+// Image to column for input layout NDHWC:
+//   input : input image [N, Di, Hi, Wi, C],
+//   output : output image [N * Do * Ho * Wo, Z * Y * X * C]
+template <index_t NDimSpatial,
+          typename InputLayout,
+          typename InputDataType,
+          typename OutputDataType,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t KPerBlock,
+          typename ThreadClusterLengths,
+          index_t ScalarPerVector>
+struct DeviceImageToColumnImpl
+    : public DeviceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>
+{
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvolutionForwardSpecialization::Default>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpecialization::MKPadding, index_t, index_t, index_t>{
+            MPerBlock, 0 /* NPerBlock*/, KPerBlock};
+
+    // Use MakeADescriptor_M_K from grouped convolution forward
+    static auto
+    MakeInputDescriptor_M_K(const ck::index_t N,
+                            const ck::index_t C,
+                            const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                            const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                            const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{1};
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{1};
+        std::array<index_t, NDimSpatial + 3> c_g_n_k_wos_lengths{1};
+
+        auto copy = [](const auto& x, auto& y, index_t dst_offset) {
+            std::copy(x.begin(), x.end(), y.begin() + dst_offset);
+        };
+
+        constexpr index_t spatial_offset = 3;
+
+        copy(input_spatial_lengths, a_g_n_c_wis_lengths, spatial_offset);
+        copy(filter_spatial_lengths, b_g_k_c_xs_lengths, spatial_offset);
+        copy(output_spatial_lengths, c_g_n_k_wos_lengths, spatial_offset);
+
+        // fill only significant values (C and N)
+        a_g_n_c_wis_lengths[I1] = N;
+        a_g_n_c_wis_lengths[I2] = C;
+        b_g_k_c_xs_lengths[I2]  = C;
+        c_g_n_k_wos_lengths[I1] = N;
+
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<InputLayout>(
+                a_g_n_c_wis_lengths,
+                input_g_n_c_wis_strides,
+                b_g_k_c_xs_lengths,
+                {}, // not needed for A Descriptor
+                c_g_n_k_wos_lengths,
+                {}, // not needed for A Descriptor
+                conv_filter_strides,
+                conv_filter_dilations,
+                input_left_pads,
+                input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+        return in_gemmm_gemmk_desc;
+    }
+
+    static auto
+    MakeOutDescriptor_M_K(const ck::index_t N,
+                          const ck::index_t C,
+                          const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                          const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                          const std::array<index_t, 2>& output_m_k_strides)
+    {
+        const index_t NDoHoWo =
+            N * ck::accumulate_n<index_t>(
+                    output_spatial_lengths.begin(), NDimSpatial, 1, std::multiplies<>());
+        const index_t CZYX =
+            C * ck::accumulate_n<index_t>(
+                    filter_spatial_lengths.begin(), NDimSpatial, 1, std::multiplies<>());
+        const auto desc_mraw_kraw = make_naive_tensor_descriptor(
+            make_tuple(NDoHoWo, CZYX), make_tuple(output_m_k_strides[I0], output_m_k_strides[I1]));
+
+        const auto desc_m_k = matrix_padder.PadADescriptor_M_K(desc_mraw_kraw);
+        return desc_m_k;
+    }
+
+    using InputGridDesc =
+        remove_cvref_t<decltype(MakeInputDescriptor_M_K(1, 1, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using OutputGridDesc = remove_cvref_t<decltype(MakeOutDescriptor_M_K(1, 1, {}, {}, {}))>;
+
+    using Block2ETileMap = remove_cvref_t<
+        decltype(BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, OutputGridDesc>(
+            OutputGridDesc{}))>;
+
+    using GridwiseImageToColumnKernel = GridwiseImageToColumn<InputGridDesc,
+                                                              InputDataType,
+                                                              OutputGridDesc,
+                                                              OutputDataType,
+                                                              BlockSize,
+                                                              MPerBlock,
+                                                              KPerBlock,
+                                                              ThreadClusterLengths,
+                                                              ScalarPerVector,
+                                                              Block2ETileMap>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_in, // input image
+                 void* p_out,      // output image
+                 const ck::index_t N,
+                 const ck::index_t C,
+                 const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                 const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                 const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                 const std::array<index_t, 2>& output_m_k_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads)
+            : C_(C),
+              X_(filter_spatial_lengths[NDimSpatial - I1]),
+              p_in_{static_cast<const InputDataType*>(p_in)},
+              p_out_{static_cast<OutputDataType*>(p_out)},
+              input_g_n_c_wis_strides_{input_g_n_c_wis_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+
+            in_grid_desc_m_k_ = MakeInputDescriptor_M_K(N,
+                                                        C,
+                                                        input_spatial_lengths,
+                                                        filter_spatial_lengths,
+                                                        output_spatial_lengths,
+                                                        input_g_n_c_wis_strides,
+
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+
+            out_grid_desc_m_k_ = MakeOutDescriptor_M_K(
+                N, C, filter_spatial_lengths, output_spatial_lengths, output_m_k_strides);
+        }
+
+        void Print() const
+        {
+            std::cout << in_grid_desc_m_k_ << std::endl;
+            std::cout << out_grid_desc_m_k_ << std::endl;
+        }
+
+        const ck::index_t C_;
+        const ck::index_t X_;
+
+        const InputDataType* p_in_;
+        OutputDataType* p_out_;
+
+        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides_;
+        const std::array<index_t, NDimSpatial>& conv_filter_strides_;
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations_;
+        const std::array<index_t, NDimSpatial>& input_left_pads_;
+        const std::array<index_t, NDimSpatial>& input_right_pads_;
+
+        InputGridDesc in_grid_desc_m_k_;
+        OutputGridDesc out_grid_desc_m_k_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            const auto block_2_tile_map =
+                BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, OutputGridDesc>(
+                    arg.out_grid_desc_m_k_);
+            const index_t grid_size = block_2_tile_map.CalculateGridSize(arg.out_grid_desc_m_k_);
+            const auto kernel       = kernel_image_to_column<InputGridDesc,
+                                                       InputDataType,
+                                                       OutputGridDesc,
+                                                       OutputDataType,
+                                                       Block2ETileMap,
+                                                       GridwiseImageToColumnKernel>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(grid_size),
+                                                        dim3(BlockSize),
+                                                        0,
+                                                        arg.in_grid_desc_m_k_,
+                                                        arg.p_in_,
+                                                        arg.out_grid_desc_m_k_,
+                                                        arg.p_out_,
+                                                        block_2_tile_map);
+            return elapsed_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const Argument& arg)
+    {
+        using namespace tensor_layout::convolution;
+        if(!(std::is_same_v<InputLayout, GNWC> || std::is_same_v<InputLayout, GNHWC> ||
+             std::is_same_v<InputLayout, GNDHWC>))
+        {
+            return false;
+        }
+        if(!(NDimSpatial >= 1 && NDimSpatial <= 3))
+        {
+            return false;
+        }
+
+        const auto w_pad_left  = arg.input_left_pads_[NDimSpatial - I1];
+        const auto w_pad_right = arg.input_right_pads_[NDimSpatial - I1];
+        const auto dilation_x  = arg.conv_filter_dilations_[NDimSpatial - I1];
+        const auto stride_x    = arg.conv_filter_strides_[NDimSpatial - I1];
+        bool is_w_packed       = arg.input_g_n_c_wis_strides_[NDimSpatial + I2] == arg.C_;
+        bool is_c_packed       = arg.input_g_n_c_wis_strides_[I2] == 1;
+
+        // check vector acces with c not packed
+        if(!is_c_packed && ScalarPerVector != 1)
+            return false;
+        // check vector access of filter window row (only C if C is not packed)
+        if(!is_w_packed && arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of filter window row (X * C)
+        if(arg.X_ * arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of pads (w_pad_left/w_pad_right * C)
+        if(w_pad_left * arg.C_ % ScalarPerVector != 0 ||
+           w_pad_right * arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of with stride and pad
+        if((w_pad_left != 0 || w_pad_right != 0) && stride_x > 1 && arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of with dilation
+        if(dilation_x > 1 && arg.C_ % ScalarPerVector != 0)
+            return false;
+
+        return GridwiseImageToColumnKernel::CheckValidity(arg.in_grid_desc_m_k_,
+                                                          arg.out_grid_desc_m_k_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_in, // input image
+                             void* p_out,      // output image
+                             const ck::index_t N,
+                             const ck::index_t C,
+                             const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                             const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                             const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                             const std::array<index_t, 2>& output_m_k_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                             const std::array<index_t, NDimSpatial>& input_left_pads,
+                             const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        return Argument{static_cast<const InputDataType*>(p_in),
+                        static_cast<OutputDataType*>(p_out),
+                        N,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        input_g_n_c_wis_strides,
+                        output_m_k_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in, // input image
+                        void* p_out,      // output image
+                        const ck::index_t N,
+                        const ck::index_t C,
+                        const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                        const std::array<index_t, 2>& output_m_k_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads) override
+    {
+        return std::make_unique<Argument>(static_cast<const InputDataType*>(p_in),
+                                          static_cast<OutputDataType*>(p_out),
+                                          N,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          input_g_n_c_wis_strides,
+                                          output_m_k_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceImageToColumn"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << KPerBlock << ", "
+            << ScalarPerVector
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_index_pool_bwd_impl.hpp → include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp

@@ -8,7 +8,7 @@
 
 #include "ck/tensor_description/tensor_descriptor.hpp"
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
-#include "ck/tensor_operation/gpu/device/device_index_pool_bwd.hpp"
+#include "ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
@@ -25,7 +25,7 @@ template <typename DOutDataType,
           typename IndexDataType,
           typename DInDataType,
           ck::index_t InOutVectorSize>
-struct DeviceIndexPoolBwdImpl : public DeviceIndexPoolBwd<DOutDataType, IndexDataType, DInDataType>
+struct DeviceMaxPoolBwdImpl : public DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>
 {
     using DInDataType_AutomicAddPreCast =
         conditional_t<is_same_v<DInDataType, float> || is_same_v<DInDataType, double>,
@@ -91,7 +91,8 @@ struct DeviceIndexPoolBwdImpl : public DeviceIndexPoolBwd<DOutDataType, IndexDat
                  index_t dout_length,
                  index_t din_length,
                  const std::vector<ck::index_t>& window_lengths,
-                 const std::vector<ck::index_t>& window_strides)
+                 const std::vector<ck::index_t>& window_strides,
+                 const std::vector<ck::index_t>& window_dilations)
             : p_dout_{p_dout},
               p_indices_{p_indices},
               p_din_{p_din},
@@ -102,7 +103,8 @@ struct DeviceIndexPoolBwdImpl : public DeviceIndexPoolBwd<DOutDataType, IndexDat
         {
             for(size_t i = 0; i < window_lengths.size(); ++i)
             {
-                windowOverlap_ |= window_lengths.at(i) > window_strides.at(i);
+                auto eff = (window_lengths.at(i) - 1) * window_dilations.at(i) + 1;
+                windowOverlap_ |= eff > window_strides.at(i);
             }
         }
 
@@ -228,6 +230,11 @@ struct DeviceIndexPoolBwdImpl : public DeviceIndexPoolBwd<DOutDataType, IndexDat
                 }
                 else
                 {
+                    hip_check_error(hipMemsetAsync(arg.p_din_,
+                                                   0,
+                                                   arg.din_length_raw_ * sizeof(DInDataType),
+                                                   stream_config.stream_id_));
+
                     const auto put_kernel = kernel_put_element_1d<GridwisePutElementSet,
                                                                   InOutGrid1dDesc,
                                                                   DOutDataType,
@@ -292,7 +299,8 @@ struct DeviceIndexPoolBwdImpl : public DeviceIndexPoolBwd<DOutDataType, IndexDat
                         index_t dout_length,
                         index_t din_length,
                         std::vector<ck::index_t> window_lengths,
-                        std::vector<ck::index_t> window_strides) override
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> window_dilations) override
     {
         // Assume p_dout, p_indices, p_din are packed memory space, dout_length and din_length are
         // physical size of the packed tensor
@@ -302,7 +310,8 @@ struct DeviceIndexPoolBwdImpl : public DeviceIndexPoolBwd<DOutDataType, IndexDat
                                           dout_length,
                                           din_length,
                                           window_lengths,
-                                          window_strides);
+                                          window_strides,
+                                          window_dilations);
     }
 
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override

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

@@ -617,10 +617,12 @@ struct DeviceSplitKContractionMultipleD_Xdl_CShuffle
         CDEBlockTransferScalarPerVector_NPerBlock,
         LoopSched>;
 
-    using AGridDesc_AKB_AK0_M_AK1 = remove_cvref_t<decltype(
-        GridwiseGemm::MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(AGridDesc_M_K{}, 1))>;
-    using BGridDesc_BKB_BK0_N_BK1 = remove_cvref_t<decltype(
-        GridwiseGemm::MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(BGridDesc_N_K{}, 1))>;
+    using AGridDesc_AKB_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(
+            AGridDesc_M_K{}, 1))>;
+    using BGridDesc_BKB_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(
+            BGridDesc_N_K{}, 1))>;
 
     using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
 

include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp

@@ -36,6 +36,13 @@ struct Add
         y = x0 + type_convert<half_t>(x1);
     };
 
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const float& x0, const float& x1) const
+    {
+        y = type_convert<half_t>(x0 + x1);
+    };
+
     template <>
     __host__ __device__ constexpr void
     operator()<half_t>(half_t& y, const float& x0, const half_t& x1) const

include/ck/tensor_operation/gpu/element/element_wise_operation.hpp

@@ -195,6 +195,51 @@ struct AddMultiply
     }
 };
 
+// C = A * B
+// E = C x D0 + D1
+struct MultiplyAdd
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ void operator()<half_t, half_t, half_t, half_t>(half_t& e,
+                                                                        const half_t& c,
+                                                                        const half_t& d0,
+                                                                        const half_t& d1) const
+    {
+        const half_t y = (c * d0) + d1;
+        e              = y;
+    }
+    template <>
+    __host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
+                                                                       const float& c,
+                                                                       const half_t& d0,
+                                                                       const half_t& d1) const
+    {
+        const half_t y = type_convert<half_t>(c) * d0 + d1;
+        e              = y;
+    }
+    template <>
+    __host__ __device__ void operator()<float, float, half_t, half_t>(float& e,
+                                                                      const float& c,
+                                                                      const half_t& d0,
+                                                                      const half_t& d1) const
+    {
+        const float y = c * d0 + d1;
+        e             = y;
+    }
+    template <>
+    __host__ __device__ void operator()<half_t, float, float, float>(half_t& e,
+                                                                     const float& c,
+                                                                     const float& d0,
+                                                                     const float& d1) const
+    {
+        const float y = c * d0 + d1;
+        e             = y;
+    }
+};
+
 // E = FastGelu(C + D0 + D1)
 struct AddAddFastGelu
 {

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -39,6 +39,12 @@ struct PassThrough
         y = x;
     }
 
+    template <>
+    __host__ __device__ void operator()<half_t, float>(half_t& y, const float& x) const
+    {
+        y = type_convert<half_t>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x) const
     {

include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp

@@ -136,8 +136,8 @@ struct GridwiseMultiblockBatchNormForward
     using ThreadReduceDstDesc_M =
         decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
 
-    using ThreadReduceSrcDesc_M_1 = decltype(
-        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceSrcDesc_M_1 = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
 
     using ThreadwiseWelford1 =
         ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;

include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp

@@ -118,8 +118,8 @@ struct GridwiseReduceSecondHalfBatchNormBackwardFinal
     static constexpr auto thread_cluster_desc =
         make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
 
-    using ThreadReduceSrcDesc_M_1 = decltype(
-        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceSrcDesc_M_1 = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
     using ThreadReduceDstDesc_M =
         decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
 

include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp

@@ -121,8 +121,8 @@ struct GridwiseWelfordSecondHalfBatchNormForwardFinal
     static constexpr auto thread_cluster_desc =
         make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
 
-    using ThreadReduceSrcDesc_M_1 = decltype(
-        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceSrcDesc_M_1 = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
     using ThreadReduceDstDesc_M =
         decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
 

include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp

@@ -115,8 +115,8 @@ struct GridwiseWelfordSecondHalfReduceFirstHalf
 
     using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
         make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
-    using ThreadReduceSrcDesc_M_1 = decltype(
-        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceSrcDesc_M_1 = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
     using ThreadReduceDstDesc_M =
         decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
 

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

@@ -588,7 +588,8 @@ struct OffsettedBlockToCTileMap
 {
     using underlying_type = UnderlyingBlockToCTileMap;
 
-    OffsettedBlockToCTileMap(UnderlyingBlockToCTileMap block_to_ctile_map, index_t block_start)
+    __host__ __device__ OffsettedBlockToCTileMap(UnderlyingBlockToCTileMap block_to_ctile_map,
+                                                 index_t block_start)
     {
         block_to_ctile_map_ = block_to_ctile_map;
         block_start_        = block_start;

include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp

@@ -101,8 +101,8 @@ struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
     {
@@ -346,14 +346,18 @@ struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
         remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
     using DefaultBGridDesc_BK0_N_BK1 =
         remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
-    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
-        MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
-    using MeanVarGridDescriptor_MBlock_MPerBlock_NBlock      = remove_cvref_t<decltype(
-        MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(MeanVarGridDesc_M_NBlock{}))>;
-    using CountGridDescriptor_MBlock_MPerBlock_NBlock        = remove_cvref_t<decltype(
-        MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(CountGridDesc_M_NBlock{}))>;
-    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+    using MeanVarGridDescriptor_MBlock_MPerBlock_NBlock =
+        remove_cvref_t<decltype(MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(
+            MeanVarGridDesc_M_NBlock{}))>;
+    using CountGridDescriptor_MBlock_MPerBlock_NBlock =
+        remove_cvref_t<decltype(MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(
+            CountGridDesc_M_NBlock{}))>;
+    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
 
     using DefaultBlock2ETileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp

@@ -102,8 +102,8 @@ struct GridwiseBatchedGemmGemm_Xdl_CShuffle
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     template <typename ABlockDesc_AK0_M_AK1>
     __host__ __device__ static constexpr auto
@@ -286,8 +286,9 @@ struct GridwiseBatchedGemmGemm_Xdl_CShuffle
             c_grid_desc_m_n);
     }
 
-    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
 
     using DefaultBlock2CTileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp

@@ -446,14 +446,17 @@ struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
             e1_grid_desc_m_n);
     }
 
-    using E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(E1GridDesc_M_N{}))>;
+    using E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            E1GridDesc_M_N{}))>;
 
-    using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
-        MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(D0sGridDesc_M_N{}))>;
+    using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 =
+        remove_cvref_t<decltype(MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+            D0sGridDesc_M_N{}))>;
 
-    using D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeD1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(D1sGridDesc_M_N{}))>;
+    using D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeD1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            D1sGridDesc_M_N{}))>;
 
     using DefaultBlock2E1TileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2E1TileMap(E1GridDesc_M_N{}))>;

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp

@@ -114,8 +114,8 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     template <typename ABlockDesc_AK0_M_AK1>
     __host__ __device__ static constexpr auto
@@ -369,11 +369,13 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle
     }
 
     using D0sGridPointer = decltype(MakeD0sGridPointer());
-    using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
-        MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(D0sGridDesc_M_N{}))>;
+    using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 =
+        remove_cvref_t<decltype(MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+            D0sGridDesc_M_N{}))>;
 
-    using C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(C1GridDesc_M_N{}))>;
+    using C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            C1GridDesc_M_N{}))>;
 
     using DefaultBlock2CTileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(C1GridDesc_M_N{}))>;

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp

@@ -113,8 +113,8 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     template <typename ABlockDesc_AK0_M_AK1>
     __host__ __device__ static constexpr auto
@@ -300,8 +300,9 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
             c_grid_desc_m_n);
     }
 
-    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
-        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
 
     using DefaultBlock2CTileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;