Merge branch 'develop' into aosewski/gemm_tile_loop

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, 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_v1.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_v6r1r2.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/grid/gridwise_gemm_pipeline_v3.hpp"
+#include "ck/utility/workgroup_barrier.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_streamk(const typename GridwiseGemm::FloatAB* p_a_grid,
+                                   const typename GridwiseGemm::FloatAB* p_b_grid,
+                                   typename GridwiseGemm::FloatC* p_c_grid,
+                                   void* p_workspace,
+                                   index_t M,
+                                   index_t N,
+                                   index_t K,
+                                   index_t StrideA,
+                                   index_t StrideB,
+                                   index_t StrideC,
+                                   typename GridwiseGemm::Block2CTileMap block_mapping)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
+
+    __shared__ uint8_t p_shared[shared_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_c_grid,
+                      p_workspace,
+                      M,
+                      N,
+                      K,
+                      StrideA,
+                      StrideB,
+                      StrideC,
+                      block_mapping,
+                      static_cast<void*>(p_shared));
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_workspace;
+    ignore = M;
+    ignore = N;
+    ignore = K;
+    ignore = StrideA;
+    ignore = StrideB;
+    ignore = StrideC;
+    ignore = block_mapping;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename Block2CTileMap_,
+          typename FloatAB_,
+          typename FloatAcc_,
+          typename FloatC_,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_streamk
+{
+    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 K1        = Number<K1Value>{};
+    static constexpr auto M01       = 1;
+    static constexpr auto N01       = 1;
+    static constexpr auto KPerBlock = K0PerBlock * K1;
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+    using FloatAcc        = FloatAcc_;
+    using FloatCShuffle   = FloatAcc;
+
+    using Block2CTileMap = Block2CTileMap_;
+    using FloatAB        = FloatAB_;
+    using FloatC         = FloatC_;
+
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        const FloatAB* p_a_grid;
+        const FloatAB* 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;
+        Block2CTileMap block_mapping;
+
+        Argument(const FloatAB* p_a_grid_,
+                 const FloatAB* 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_,
+                 uint32_t num_cu,
+                 uint32_t occupancy,
+                 uint32_t num_sk_blocks_)
+            : p_a_grid(p_a_grid_),
+              p_b_grid(p_b_grid_),
+              p_c_grid(p_c_grid_),
+              M(M_),
+              N(N_),
+              K(K_),
+              StrideA(StrideA_),
+              StrideB(StrideB_),
+              StrideC(StrideC_),
+              block_mapping(M, N, K, num_cu, occupancy, num_sk_blocks_)
+        {
+        }
+
+        void Print() const
+        {
+            std::cout << "arg {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << std::endl;
+        }
+    };
+
+    __host__ __device__ static auto CalculateGridSize(const Argument& karg)
+    {
+        return std::make_tuple(math::integer_divide_ceil(karg.N, NPerBlock),
+                               math::integer_divide_ceil(karg.M, MPerBlock),
+                               karg.k_batch);
+    }
+
+    __host__ __device__ static auto CalculateK0(index_t KPad) { return KPad / K1; }
+
+    __host__ __device__ static auto
+    MakeAGridDescriptor_K0_M_K1(index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA)
+    {
+        const index_t K0 = CalculateK0(KPad);
+
+        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 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>{}));
+
+        return transform_tensor_descriptor(a_grid_desc_m_kpad,
+                                           make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                      make_right_pad_transform(M, MPad - M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    __host__ __device__ static auto
+    MakeBGridDescriptor_K0_N_K1(index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB)
+    {
+        const index_t K0 = CalculateK0(KPad);
+
+        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 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>{}));
+
+        return transform_tensor_descriptor(b_grid_desc_kpad_n,
+                                           make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    __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_m_n = [&]() {
+            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));
+            }
+        }();
+
+        return transform_tensor_descriptor(c_grid_desc_m_n,
+                                           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>{}));
+    }
+
+    __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(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * K1, K1, 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(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * K1, K1, I1));
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto c_block_size =
+            GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle().GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    __host__ __device__ static constexpr bool CheckValidity(const Argument& karg)
+    {
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(karg.K % ABlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+        else
+        {
+            if(karg.M % ABlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(karg.N % BBlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+        else
+        {
+            if(karg.K % BBlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
+                return false;
+        }
+        else
+        {
+            if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
+                return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = K0 > K0PerBlock;
+
+        return has_main_k0_block_loop;
+    }
+
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto
+    MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        return transform_tensor_descriptor(
+            c_m_n_grid_desc,
+            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 block_id to C matrix tile idx (m0, n0) mapping
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc>(
+            c_m_n_grid_desc, 8, KBatch);
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1,
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       I1,
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MShuffleRepeat_MPerShuffle_NShuffleRepeat_NPerShuffle()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat / CShuffleMRepeatPerShuffle>{},
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       Number<NRepeat / CShuffleNRepeatPerShuffle>{},
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    __host__ __device__ static constexpr auto GetClusterLengthReduction()
+    {
+        // TODO: assume C is row major
+        // TODO: we always first loop over N, then M
+        constexpr auto NPerBlockPow2 = math::next_power_of_two<NPerBlock>();
+        constexpr auto NPerBlockReduction =
+            NPerBlockPow2 / CBlockTransferScalarPerVector_NWaveNPerXDL;
+        constexpr auto MPerBlockReduction =
+            (BlockSize + NPerBlockReduction - 1) / NPerBlockReduction;
+        return Sequence<MPerBlockReduction, NPerBlockReduction>{};
+    }
+
+    __host__ __device__ static constexpr auto GetPartialAccBlockDescriptor()
+    {
+        const auto c_partial_acc_block_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MPerBlock, NPerBlock),
+                                                    make_tuple(NPerBlock, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MPerBlock, NPerBlock),
+                                                    make_tuple(I1, MPerBlock));
+            }
+        }();
+        return c_partial_acc_block_m_n;
+    }
+
+    using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(1, 1, 1, 1, 1))>;
+
+    __device__ static void Run(const FloatAB* p_a_grid,
+                               const FloatAB* p_b_grid,
+                               FloatC* p_c_grid,
+                               void* p_workspace,
+                               index_t M,
+                               index_t N,
+                               index_t K,
+                               index_t StrideA,
+                               index_t StrideB,
+                               index_t StrideC,
+                               Block2CTileMap block_mapping,
+                               void* __restrict__ p_shared_block)
+    {
+        uint32_t m        = M;
+        uint32_t n        = N;
+        uint32_t k        = K;
+        uint32_t pad_m    = (m + MPerBlock - 1) / MPerBlock * MPerBlock;
+        uint32_t pad_n    = (n + NPerBlock - 1) / NPerBlock * NPerBlock;
+        uint32_t pad_k    = (k + KPerBlock - 1) / KPerBlock * KPerBlock;
+        uint32_t stride_a = StrideA;
+        uint32_t stride_b = StrideB;
+        uint32_t stride_c = StrideC;
+
+        const auto a_k0_m_k1_grid_desc = MakeAGridDescriptor_K0_M_K1(m, pad_m, k, pad_k, stride_a);
+        const auto b_k0_n_k1_grid_desc = MakeBGridDescriptor_K0_N_K1(k, pad_k, n, pad_n, stride_b);
+        const auto c_grid_desc_m_n     = MakeCGridDescriptor_M_N(m, pad_m, n, pad_n, stride_c);
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
+        const AElementwiseOperation a_element_op = AElementwiseOperation{};
+        const BElementwiseOperation b_element_op = BElementwiseOperation{};
+        const CElementwiseOperation c_element_op = CElementwiseOperation{};
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block = static_cast<FloatAB*>(p_shared_block);
+        FloatAB* p_b_block = static_cast<FloatAB*>(p_shared_block) + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block, a_block_desc_k0_m_k1.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block, b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_v3();
+
+        uint32_t block_idx = block_mapping.get_block_idx();
+        bool is_sk_block   = block_idx < block_mapping.sk_num_blocks;
+        bool is_dp_block   = block_idx >= block_mapping.dp_start_block_idx &&
+                           block_idx < block_mapping.reduction_start_block_idx;
+        bool is_reduction_block = block_idx >= block_mapping.reduction_start_block_idx;
+        bool is_padding_block   = block_idx >= block_mapping.sk_num_blocks &&
+                                block_idx < block_mapping.dp_start_block_idx;
+        uint32_t iter_start, iter_end;
+        block_mapping.get_block_itr(block_idx, iter_start, iter_end);
+        uint32_t total_iter_length = iter_end - iter_start;
+
+        if(is_padding_block)
+            return;
+
+        uint32_t* p_semaphore =
+            reinterpret_cast<uint32_t*>(reinterpret_cast<char*>(p_workspace) +
+                                        block_mapping.get_workspace_size_for_acc(sizeof(FloatAcc)));
+
+        if constexpr(Block2CTileMap::ReductionStrategy == StreamKReductionStrategy::Reduction)
+        {
+            if(is_reduction_block)
+            {
+                // descriptors
+                constexpr auto cluster_length_reduce = GetClusterLengthReduction();
+                constexpr auto reduce_desc = make_cluster_descriptor(cluster_length_reduce);
+                const auto reduce_thread_cluster_idx =
+                    reduce_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+                const auto thread_m_cluster_id = reduce_thread_cluster_idx[I0];
+                const auto thread_n_cluster_id = reduce_thread_cluster_idx[I1];
+
+                constexpr auto MReduceIters =
+                    math::integer_divide_ceil(Number<MPerBlock>{}, cluster_length_reduce.At(I0));
+                constexpr auto NReduceIters = math::integer_divide_ceil(
+                    Number<NPerBlock>{},
+                    cluster_length_reduce.At(I1) *
+                        Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{});
+
+                constexpr auto acc_thread_buf_load_desc = make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{}));
+                constexpr auto acc_thread_buf_store_desc = make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, I1, I1, Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{}));
+
+                constexpr auto c_partial_acc_block_m_n = GetPartialAccBlockDescriptor();
+
+                constexpr auto partial_acc_load_step_n = make_multi_index(
+                    0, cluster_length_reduce.At(I1) * CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_load_step_n_reverse =
+                    make_multi_index(0,
+                                     -1 * cluster_length_reduce.At(I1).value * (NReduceIters - 1) *
+                                         CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_load_step_m =
+                    make_multi_index(cluster_length_reduce.At(I0), 0);
+
+                constexpr auto partial_acc_store_step_n = make_multi_index(
+                    0,
+                    0,
+                    0,
+                    cluster_length_reduce.At(I1) * CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_store_step_n_reverse =
+                    make_multi_index(0,
+                                     0,
+                                     0,
+                                     -1 * cluster_length_reduce.At(I1).value * (NReduceIters - 1) *
+                                         CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_store_step_m =
+                    make_multi_index(0, cluster_length_reduce.At(I0), 0, 0);
+
+                StaticBuffer<AddressSpaceEnum::Vgpr,
+                             FloatAcc,
+                             CBlockTransferScalarPerVector_NWaveNPerXDL,
+                             true>
+                    parcial_acc_buf;
+                StaticBuffer<AddressSpaceEnum::Vgpr,
+                             FloatAcc,
+                             CBlockTransferScalarPerVector_NWaveNPerXDL,
+                             true>
+                    acc_buf;
+
+                // start to compute
+                auto reduction_idx = blockIdx.x - block_mapping.reduction_start_block_idx;
+                auto spatial_idx   = block_mapping.tile_to_spatial(reduction_idx, m, n);
+
+                workgroup_barrier wg_barrier(p_semaphore);
+
+                uint32_t tile_acc_offset_start =
+                    block_mapping.get_acc_buffer_offset_from_tile(reduction_idx);
+                uint32_t tile_acc_offset_end =
+                    block_mapping.get_acc_buffer_offset_from_tile(reduction_idx + 1);
+
+                auto acc_load = ThreadwiseTensorSliceTransfer_v2<
+                    FloatAcc,                                                // SrcData,
+                    FloatAcc,                                                // DstData,
+                    decltype(c_partial_acc_block_m_n),                       // SrcDesc,
+                    decltype(acc_thread_buf_load_desc),                      // DstDesc,
+                    Sequence<1, CBlockTransferScalarPerVector_NWaveNPerXDL>, // SliceLengths,
+                    Sequence<0, 1>,                                          // DimAccessOrder,
+                    1,                                                       // SrcVectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL,              // SrcScalarPerVector,
+                    1,    // SrcScalarStrideInVector,
+                    false // SrcResetCoordinateAfterRun,
+                    >{c_partial_acc_block_m_n,
+                      make_multi_index(thread_m_cluster_id,
+                                       thread_n_cluster_id *
+                                           CBlockTransferScalarPerVector_NWaveNPerXDL)};
+
+                auto acc_store = ThreadwiseTensorSliceTransfer_v1r3<
+                    FloatAcc,                                                // SrcData,
+                    FloatC,                                                  // DstData,
+                    decltype(acc_thread_buf_store_desc),                     // SrcDesc,
+                    decltype(c_grid_desc_mblock_mperblock_nblock_nperblock), // DstDesc,
+                    CElementwiseOperation, // ElementwiseOperation,
+                    Sequence<1, 1, 1, CBlockTransferScalarPerVector_NWaveNPerXDL>, // SliceLengths,
+                    Sequence<0, 1, 2, 3>,                       // DimAccessOrder,
+                    3,                                          // DstVectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // DstScalarPerVector,
+                    InMemoryDataOperationEnum::Set, // InMemoryDataOperationEnum DstInMemOp,
+                    1,                              // DstScalarStrideInVector,
+                    false                           // DstResetCoordinateAfterRun,
+                    >{c_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(__builtin_amdgcn_readfirstlane(spatial_idx[I0]),
+                                       thread_m_cluster_id,
+                                       __builtin_amdgcn_readfirstlane(spatial_idx[I1]),
+                                       thread_n_cluster_id *
+                                           CBlockTransferScalarPerVector_NWaveNPerXDL),
+                      CElementwiseOperation{}};
+
+                // block synchronization
+                wg_barrier.wait_eq(reduction_idx, tile_acc_offset_end - tile_acc_offset_start);
+
+#if 0
+                if(threadIdx.x == 0) {
+                    printf("bid:%d, rid:%d, os:%d,%d, spatial:%d,%d\n", static_cast<int>(blockIdx.x),
+                        reduction_idx, __builtin_amdgcn_readfirstlane(tile_acc_offset_start), __builtin_amdgcn_readfirstlane(tile_acc_offset_end),
+                        __builtin_amdgcn_readfirstlane(spatial_idx[I0]),
+                        __builtin_amdgcn_readfirstlane(spatial_idx[I1]));
+                }
+#endif
+
+                using Accumulation = ck::detail::
+                    AccumulateWithNanCheck<false /*PropagateNan*/, reduce::Add, FloatAcc>;
+
+                for(int i_m = 0; i_m < MReduceIters; i_m++)
+                {
+                    static_for<0, NReduceIters, 1>{}([&](auto i_n_reduce) {
+                        acc_buf.Clear();
+                        for(auto i = tile_acc_offset_start; i < tile_acc_offset_end; i++)
+                        {
+                            auto c_partial_acc_buf =
+                                make_dynamic_buffer<AddressSpaceEnum::Global,
+                                                    AmdBufferCoherenceEnum::GLC>(
+                                    reinterpret_cast<FloatAcc*>(p_workspace) +
+                                        i * c_partial_acc_block_m_n.GetElementSpaceSize(),
+                                    c_partial_acc_block_m_n.GetElementSpaceSize());
+
+                            acc_load.Run(c_partial_acc_block_m_n,
+                                         c_partial_acc_buf,
+                                         acc_thread_buf_load_desc,
+                                         make_tuple(I0, I0),
+                                         parcial_acc_buf);
+
+                            static_for<0, CBlockTransferScalarPerVector_NWaveNPerXDL, 1>{}(
+                                [&](auto i_vec) {
+                                    constexpr auto offset =
+                                        acc_thread_buf_load_desc.CalculateOffset(
+                                            make_tuple(0, i_vec));
+                                    Accumulation::Calculate(acc_buf(Number<offset>{}),
+                                                            parcial_acc_buf[Number<offset>{}]);
+                                });
+                        }
+
+                        if(thread_n_cluster_id * CBlockTransferScalarPerVector_NWaveNPerXDL <
+                           NPerBlock)
+                        {
+                            acc_store.Run(acc_thread_buf_store_desc,
+                                          make_tuple(I0, I0, I0, I0),
+                                          acc_buf,
+                                          c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                          c_grid_buf);
+                        }
+                        if constexpr(NReduceIters != 1)
+                        {
+                            if constexpr(i_n_reduce != (NReduceIters - 1))
+                            {
+                                acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
+                                                            partial_acc_load_step_n);
+                                acc_store.MoveDstSliceWindow(
+                                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    partial_acc_store_step_n);
+                            }
+                            else
+                            {
+                                acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
+                                                            partial_acc_load_step_n_reverse);
+                                acc_store.MoveDstSliceWindow(
+                                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    partial_acc_store_step_n_reverse);
+                            }
+                        }
+                    });
+                    {
+                        acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
+                                                    partial_acc_load_step_m);
+                        acc_store.MoveDstSliceWindow(c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                     partial_acc_store_step_m);
+                    }
+                }
+                return;
+            }
+        }
+
+        // offset for last acc buffer of this block
+        uint32_t block_acc_offset =
+            (block_mapping.get_acc_buffer_offset_from_block(block_idx + 1) - 1) * MPerBlock *
+            NPerBlock;
+
+        while(true)
+        {
+            uint32_t current_iter_length = __builtin_amdgcn_readfirstlane(
+                block_mapping.get_current_iter_length(iter_start, iter_end, total_iter_length));
+            uint32_t tile_idx, iter_offset;
+            block_mapping.get_tile_idx_with_offset(iter_end - 1, tile_idx, iter_offset);
+            iter_offset = __builtin_amdgcn_readfirstlane(iter_offset - current_iter_length + 1);
+            auto spatial_idx = block_mapping.tile_to_spatial(tile_idx, m, n);
+
+            const index_t m_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(spatial_idx[I0] * MPerBlock);
+
+            const index_t n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(spatial_idx[I1] * NPerBlock);
+
+            const index_t k0_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(iter_offset * K0PerBlock);
+
+            // A matrix blockwise copy
+            auto a_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                    AElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<K0PerBlock, MPerBlock, K1>,
+                                                    ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                    ABlockTransferThreadClusterArrangeOrder,
+                                                    FloatAB,
+                                                    FloatAB,
+                                                    decltype(a_k0_m_k1_grid_desc),
+                                                    decltype(a_block_desc_k0_m_k1),
+                                                    ABlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    ABlockTransferSrcVectorDim,
+                                                    2,
+                                                    ABlockTransferSrcScalarPerVector,
+                                                    ABlockTransferDstScalarPerVector_K1,
+                                                    1,
+                                                    1,
+                                                    AThreadTransferSrcResetCoordinateAfterRun,
+                                                    true>(
+                    a_k0_m_k1_grid_desc,
+                    make_multi_index(k0_block_data_idx_on_grid, m_block_data_idx_on_grid, 0),
+                    a_element_op,
+                    a_block_desc_k0_m_k1,
+                    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<K0PerBlock, NPerBlock, K1>,
+                                                    BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                    BBlockTransferThreadClusterArrangeOrder,
+                                                    FloatAB,
+                                                    FloatAB,
+                                                    decltype(b_k0_n_k1_grid_desc),
+                                                    decltype(b_block_desc_k0_n_k1),
+                                                    BBlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    BBlockTransferSrcVectorDim,
+                                                    2,
+                                                    BBlockTransferSrcScalarPerVector,
+                                                    BBlockTransferDstScalarPerVector_K1,
+                                                    1,
+                                                    1,
+                                                    BThreadTransferSrcResetCoordinateAfterRun,
+                                                    true>(
+                    b_k0_n_k1_grid_desc,
+                    make_multi_index(k0_block_data_idx_on_grid, n_block_data_idx_on_grid, 0),
+                    b_element_op,
+                    b_block_desc_k0_n_k1,
+                    make_multi_index(0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            const index_t num_k_block_main_loop = current_iter_length;
+
+            gridwise_gemm_pipeline.Run(a_k0_m_k1_grid_desc,
+                                       a_block_desc_k0_m_k1,
+                                       a_blockwise_copy,
+                                       a_grid_buf,
+                                       a_block_buf,
+                                       a_block_slice_copy_step,
+                                       b_k0_n_k1_grid_desc,
+                                       b_block_desc_k0_n_k1,
+                                       b_blockwise_copy,
+                                       b_grid_buf,
+                                       b_block_buf,
+                                       b_block_slice_copy_step,
+                                       blockwise_gemm,
+                                       c_thread_buf,
+                                       num_k_block_main_loop);
+
+            // output: register to global memory
+            {
+                constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+                constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+                constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+                constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+                constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+                constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+                constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+                constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+                constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+                constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+                constexpr auto c_block_desc_mblock_mpershuffle_nblock_npershuffle =
+                    GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle();
+
+                constexpr auto c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle =
+                    GetCBlockDescriptor_MShuffleRepeat_MPerShuffle_NShuffleRepeat_NPerShuffle();
+
+                auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                    reinterpret_cast<FloatCShuffle*>(p_shared_block),
+                    c_block_desc_mblock_mpershuffle_nblock_npershuffle.GetElementSpaceSize());
+
+                auto c_partial_acc_buf =
+                    make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
+                        reinterpret_cast<FloatAcc*>(p_workspace) + block_acc_offset,
+                        c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle
+                            .GetElementSpaceSize());
+
+                constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                    c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                    make_tuple(make_freeze_transform(I0), // freeze mblock
+                               make_unmerge_transform(
+                                   make_tuple(CShuffleMRepeatPerShuffle,
+                                              M1,
+                                              M2,
+                                              M3,
+                                              M4)),       // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                               make_freeze_transform(I0), // freeze nblock
+                               make_unmerge_transform(
+                                   make_tuple(CShuffleNRepeatPerShuffle,
+                                              N1,
+                                              N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                    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));
+
+                // VGPR to LDS
+                auto c_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3<
+                    FloatAcc,
+                    FloatCShuffle,
+                    decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                    decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                    ck::tensor_operation::element_wise::PassThrough,
+                    Sequence<CShuffleMRepeatPerShuffle,
+                             CShuffleNRepeatPerShuffle,
+                             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{}};
+
+                // LDS to global
+                auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1r2<
+                    ThisThreadBlock,       // index_t BlockSize,
+                    CElementwiseOperation, // ElementwiseOperation,
+                                           // InMemoryDataOperationEnum::Set, // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                             1,
+                             CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                    CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    FloatCShuffle,        // typename SrcData,
+                    FloatC,               // typename DstData,
+                    decltype(c_block_desc_mblock_mpershuffle_nblock_npershuffle),
+                    decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                    Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                    3,                                          // index_t VectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                    false, // bool ThreadTransferSrcResetCoordinateAfterRun,
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun
+                    {c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                     make_multi_index(0, 0, 0, 0),
+                     c_grid_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(__builtin_amdgcn_readfirstlane(spatial_idx[I0]),
+                                      0,
+                                      __builtin_amdgcn_readfirstlane(spatial_idx[I1]),
+                                      0),
+                     c_element_op};
+
+                // LDS to global partial acc
+                auto c_block_copy_lds_to_partial_acc = ThreadGroupTensorSliceTransfer_v6r1r2<
+                    ThisThreadBlock,       // index_t BlockSize,
+                    CElementwiseOperation, // ElementwiseOperation,
+                                           // InMemoryDataOperationEnum::Set, // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                             1,
+                             CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                    CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    FloatCShuffle,        // typename SrcData,
+                    FloatCShuffle,        // typename DstData,
+                    decltype(c_block_desc_mblock_mpershuffle_nblock_npershuffle),
+                    decltype(c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle),
+                    Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                    3,                                          // index_t VectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                    false, // bool ThreadTransferSrcResetCoordinateAfterRun, => need to be false,
+                           // othre wise has scratch
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun, => need to be false,
+                           // othre wise has scratch
+                    {c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                     make_multi_index(0, 0, 0, 0),
+                     c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
+                     make_multi_index(0, 0, 0, 0),
+                     c_element_op};
+
+                constexpr auto mxdlperwave_forward_step =
+                    make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
+                constexpr auto nxdlperwave_forward_step =
+                    make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+                constexpr auto nxdlperwave_backward_step =
+                    make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+
+                static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
+                    constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                    static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
+                        constexpr bool nxdlperwave_forward_sweep =
+                            (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
+
+                        constexpr index_t nxdlperwave_value =
+                            nxdlperwave_forward_sweep
+                                ? nxdlperwave_iter
+                                : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
+
+                        constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                        // make sure it's safe to do ds_write
+                        block_sync_lds();
+
+                        // VGPR to LDS
+                        c_thread_copy_vgpr_to_lds.Run(
+                            c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                            make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                            c_thread_buf,
+                            c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                            c_block_buf);
+
+                        // make sure it's safe to do ds_read
+                        block_sync_lds();
+
+                        c_block_copy_lds_to_global.SetSrcSliceOrigin(
+                            c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                            make_tuple(0, 0, 0, 0));
+
+                        // LDS to global
+                        if(is_dp_block)
+                            c_block_copy_lds_to_global.template Run<decltype(c_block_buf),
+                                                                    decltype(c_grid_buf),
+                                                                    InMemoryDataOperationEnum::Set>(
+                                c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                c_block_buf,
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                c_grid_buf);
+                        else if(is_sk_block)
+                        {
+                            if constexpr(Block2CTileMap::ReductionStrategy ==
+                                         StreamKReductionStrategy::Reduction)
+                            {
+                                // constexpr offset
+                                c_block_copy_lds_to_partial_acc.SetSrcSliceOrigin(
+                                    c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                    make_tuple(0, 0, 0, 0));
+
+                                c_block_copy_lds_to_partial_acc.SetDstSliceOrigin(
+                                    c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
+                                    make_tuple(mxdlperwave.value, 0, nxdlperwave.value, 0));
+
+                                c_block_copy_lds_to_partial_acc
+                                    .template Run<decltype(c_block_buf),
+                                                  decltype(c_partial_acc_buf),
+                                                  InMemoryDataOperationEnum::Set>(
+                                        c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                        c_block_buf,
+                                        c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
+                                        c_partial_acc_buf);
+                            }
+                            else if constexpr(Block2CTileMap::ReductionStrategy ==
+                                              StreamKReductionStrategy::Atomic)
+                            {
+                                c_block_copy_lds_to_global
+                                    .template Run<decltype(c_block_buf),
+                                                  decltype(c_grid_buf),
+                                                  InMemoryDataOperationEnum::AtomicAdd>(
+                                        c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                        c_block_buf,
+                                        c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                        c_grid_buf);
+                            }
+                        }
+
+                        // move on nxdlperwave dimension
+                        if constexpr(nxdlperwave_forward_sweep &&
+                                     (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
+                        {
+                            c_block_copy_lds_to_global.MoveDstSliceWindow(
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                nxdlperwave_forward_step);
+                        }
+                        else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                        {
+                            c_block_copy_lds_to_global.MoveDstSliceWindow(
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                nxdlperwave_backward_step);
+                        }
+                    });
+
+                    // move on mxdlperwave dimension
+                    if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            mxdlperwave_forward_step);
+                    }
+                });
+
+                if constexpr(Block2CTileMap::ReductionStrategy ==
+                             StreamKReductionStrategy::Reduction)
+                {
+                    if(is_sk_block)
+                    {
+                        // increase the counter for this tile
+                        workgroup_barrier wg_barrier(p_semaphore);
+                        wg_barrier.inc(tile_idx);
+                    }
+                }
+            }
+
+            // exit condition
+            iter_end -= current_iter_length;
+            if(iter_end <= iter_start)
+                break;
+
+            if constexpr(Block2CTileMap::ReductionStrategy == StreamKReductionStrategy::Reduction)
+            {
+                block_acc_offset -= MPerBlock * NPerBlock;
+            }
+            // make sure next loop LDS is ready for use
+            block_sync_lds();
+        }
+    }
+
+    template <typename Layout>
+    struct LStr
+    {
+        static std::string Get() { return ""; }
+    };
+
+    template <>
+    struct LStr<ck::tensor_layout::gemm::RowMajor>
+    {
+        static std::string Get() { return "R"; }
+    };
+
+    template <>
+    struct LStr<ck::tensor_layout::gemm::ColumnMajor>
+    {
+        static std::string Get() { return "C"; }
+    };
+
+    static std::string GetTypeString()
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "GemmXdlStreamK_"
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << "_"
+            << "B" << BlockSize << "_"
+            << "Vec" << ABlockTransferSrcScalarPerVector << "x"
+            << BBlockTransferSrcScalarPerVector << "x"
+            << CBlockTransferScalarPerVector_NWaveNPerXDL << "_"
+            << MPerBlock << "x"
+            << NPerBlock << "x"
+            << K0PerBlock << "x"
+            << K1 ;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool SrcResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r1r2
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r1r2(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const ElementwiseOperation& element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, InMemoryDataOperationEnum DstInMemOp>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        // loop over space-filling curve
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src_vector_type = vector_type_maker_t<SrcData, ScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                SrcData v;
+
+                // apply element-wise operation
+                element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
+
+                // apply type convert
+                dst_vector_container.template AsType<DstData>()(i) = type_convert<DstData>(v);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRun
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck

include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp

@@ -18,32 +18,53 @@ template <
     index_t NDimSpatial,
     typename ALayout,
     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization ConvBwdDataSpecialization>
-constexpr auto
-make_out_n_ho_wo_k_grid_desc(const index_t N,
+constexpr auto make_out_grid_desc(const index_t N,
+                                  const index_t Do,
                                   const index_t Ho,
                                   const index_t Wo,
                                   const index_t K,
                                   const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_strides)
 {
+    const auto KStride = Number<1>{};
 
     if constexpr(is_same_v<ALayout, tensor_layout::convolution::NHWGK>)
     {
         const index_t NStride  = out_g_n_k_wos_strides[1];
         const index_t HiStride = out_g_n_k_wos_strides[3];
         const index_t WiStride = out_g_n_k_wos_strides[4];
-        const auto CStride     = Number<1>{};
         if constexpr(ConvBwdDataSpecialization ==
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
                          Filter1x1Stride1Pad0)
         {
 
             return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, K),
-                                                make_tuple(WiStride, CStride));
+                                                make_tuple(WiStride, KStride));
         }
         else
         {
             return make_naive_tensor_descriptor(make_tuple(N, Ho, Wo, K),
-                                                make_tuple(NStride, HiStride, WiStride, CStride));
+                                                make_tuple(NStride, HiStride, WiStride, KStride));
+        }
+    }
+    else if constexpr(is_same_v<ALayout, tensor_layout::convolution::NDHWGK>)
+    {
+        const index_t NStride  = out_g_n_k_wos_strides[1];
+        const index_t DoStride = out_g_n_k_wos_strides[3];
+        const index_t HoStride = out_g_n_k_wos_strides[4];
+        const index_t WoStride = out_g_n_k_wos_strides[5];
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+
+            return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, K),
+                                                make_tuple(WoStride, KStride));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(N, Do, Ho, Wo, K),
+                make_tuple(NStride, DoStride, HoStride, WoStride, KStride));
         }
     }
     else if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK>)
@@ -60,12 +81,80 @@ make_out_n_ho_wo_k_grid_desc(const index_t N,
             return make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
         }
     }
+    else if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNDHWK>)
+    {
+        // assume packed
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(N, Do, Ho, Wo, K));
+        }
+    }
     else
     {
         throw std::runtime_error("wrong! unsupported layout: " + ALayout::name());
     }
 }
 
+template <typename BLayout>
+constexpr auto make_wei_grid_desc(
+    const index_t K, const index_t Z, const index_t Y, const index_t X, const index_t C)
+{
+
+    if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKYXC>)
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+    }
+    else if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKZYXC>)
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(K, Z, Y, X, C));
+    }
+    else
+    {
+        throw std::runtime_error("wrong! unsupported layout: " + BLayout::name());
+    }
+}
+
+template <index_t NDimSpatial, typename CLayout>
+constexpr auto make_in_grid_desc(const index_t N,
+                                 const index_t Di,
+                                 const index_t Hi,
+                                 const index_t Wi,
+                                 const index_t C,
+                                 const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_strides)
+{
+
+    if constexpr(is_same_v<CLayout, tensor_layout::convolution::GNHWC> ||
+                 is_same_v<CLayout, tensor_layout::convolution::NHWGC> ||
+                 is_same_v<CLayout, tensor_layout::convolution::G_NHW_C>)
+    {
+        return make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
+                                            make_tuple(in_g_n_c_wis_strides[1],
+                                                       in_g_n_c_wis_strides[3],
+                                                       in_g_n_c_wis_strides[4],
+                                                       in_g_n_c_wis_strides[2]));
+    }
+    else if constexpr(is_same_v<CLayout, tensor_layout::convolution::GNDHWC> ||
+                      is_same_v<CLayout, tensor_layout::convolution::NDHWGC>)
+    {
+        return make_naive_tensor_descriptor(make_tuple(N, Di, Hi, Wi, C),
+                                            make_tuple(in_g_n_c_wis_strides[1],
+                                                       in_g_n_c_wis_strides[3],
+                                                       in_g_n_c_wis_strides[4],
+                                                       in_g_n_c_wis_strides[5],
+                                                       in_g_n_c_wis_strides[2]));
+    }
+    else
+    {
+        throw std::runtime_error("wrong! unsupported layout: " + CLayout::name());
+    }
+}
+
 } // namespace
 
 template <
@@ -82,10 +171,26 @@ struct TransformConvBwdDataToGemm_v1
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
 
+    static constexpr auto NonSpatialDimsNum = Number<3>{};
+
+    static constexpr auto DIdx = Number<NonSpatialDimsNum>{};
+    static constexpr auto HIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+    static constexpr auto WIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{} : Number<NonSpatialDimsNum + 2>{};
+
+    static constexpr auto ZIdx = Number<NonSpatialDimsNum>{};
+    static constexpr auto YIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+    static constexpr auto XIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{} : Number<NonSpatialDimsNum + 2>{};
+
     template <typename ALayout,
-              typename std::enable_if<NDimSpatial == 2 &&
+              typename std::enable_if<(NDimSpatial == 2 || NDimSpatial == 3) &&
                                           (is_same_v<ALayout, tensor_layout::convolution::GNHWK> ||
-                                           is_same_v<ALayout, tensor_layout::convolution::NHWGK>),
+                                           is_same_v<ALayout, tensor_layout::convolution::GNDHWK> ||
+                                           is_same_v<ALayout, tensor_layout::convolution::NHWGK> ||
+                                           is_same_v<ALayout, tensor_layout::convolution::NDHWGK>),
                                       bool>::type = false>
     static auto MakeADescriptor_AK0_M_AK1(
         const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
@@ -100,35 +205,43 @@ struct TransformConvBwdDataToGemm_v1
         const std::array<index_t, NDimSpatial>& /* input_right_pads */,
         const std::array<index_t, NDimSpatial>& tildes)
     {
-        index_t i_ytilde = tildes[0];
-        index_t i_xtilde = tildes[1];
+        index_t i_ztilde = tildes[ZIdx - NonSpatialDimsNum];
+        index_t i_ytilde = tildes[YIdx - NonSpatialDimsNum];
+        index_t i_xtilde = tildes[XIdx - NonSpatialDimsNum];
 
         const index_t N = in_g_n_c_wis_lengths[1];
         const index_t K = wei_g_k_c_xs_lengths[1];
 
-        const index_t Hi = in_g_n_c_wis_lengths[3];
-        const index_t Wi = in_g_n_c_wis_lengths[4];
+        const index_t Di = NDimSpatial == 3 ? in_g_n_c_wis_lengths[DIdx] : 1;
+        const index_t Hi = in_g_n_c_wis_lengths[HIdx];
+        const index_t Wi = in_g_n_c_wis_lengths[WIdx];
 
-        const index_t Ho = out_g_n_k_wos_lengths[3];
-        const index_t Wo = out_g_n_k_wos_lengths[4];
+        const index_t Do = NDimSpatial == 3 ? out_g_n_k_wos_lengths[DIdx] : 1;
+        const index_t Ho = out_g_n_k_wos_lengths[HIdx];
+        const index_t Wo = out_g_n_k_wos_lengths[WIdx];
 
-        const index_t Y = wei_g_k_c_xs_lengths[3];
-        const index_t X = wei_g_k_c_xs_lengths[4];
+        const index_t Z = NDimSpatial == 3 ? wei_g_k_c_xs_lengths[ZIdx] : 1;
+        const index_t Y = wei_g_k_c_xs_lengths[YIdx];
+        const index_t X = wei_g_k_c_xs_lengths[XIdx];
 
-        const index_t InLeftPadH = input_left_pads[0];
-        const index_t InLeftPadW = input_left_pads[1];
+        const index_t InLeftPadD = input_left_pads[DIdx - NonSpatialDimsNum];
+        const index_t InLeftPadH = input_left_pads[HIdx - NonSpatialDimsNum];
+        const index_t InLeftPadW = input_left_pads[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvStrideH = conv_filter_strides[0];
-        const index_t ConvStrideW = conv_filter_strides[1];
+        const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+        const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+        const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvDilationH = conv_filter_dilations[0];
-        const index_t ConvDilationW = conv_filter_dilations[1];
+        const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+        const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+        const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
         const index_t AK0 = K / AK1;
 
-        const auto out_n_ho_wo_k_grid_desc =
-            make_out_n_ho_wo_k_grid_desc<NDimSpatial, ALayout, ConvBwdDataSpecialization>(
-                N, Ho, Wo, K, out_g_n_k_wos_strides);
+        // n_do_ho_wo_k for 3d or n_ho_wo_k for 2d
+        const auto out_grid_desc =
+            make_out_grid_desc<NDimSpatial, ALayout, ConvBwdDataSpecialization>(
+                N, Do, Ho, Wo, K, out_g_n_k_wos_strides);
 
         if constexpr(ConvBwdDataSpecialization ==
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
@@ -136,8 +249,8 @@ struct TransformConvBwdDataToGemm_v1
         {
             // A: output tensor
             const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
-                out_n_ho_wo_k_grid_desc,
-                make_tuple(make_pass_through_transform(N * Ho * Wo),
+                out_grid_desc,
+                make_tuple(make_pass_through_transform(N * Do * Ho * Wo),
                            make_unmerge_transform(make_tuple(AK0, AK1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
@@ -152,41 +265,54 @@ struct TransformConvBwdDataToGemm_v1
         }
         else
         {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
             const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
             const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
             const auto YTilde = ConvStrideH / GcdStrideDilationH;
             const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
             const auto YDot = math::integer_divide_ceil(Y, YTilde);
             const auto XDot = math::integer_divide_ceil(X, XTilde);
 
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
             const auto HTilde =
                 Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
             const auto WTilde =
                 Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
 
             // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
             const auto IHTildeSliceBegin = math::integer_divide_floor(
                 math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
             const auto IWTildeSliceBegin = math::integer_divide_floor(
                 math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
 
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
             const auto IHTildeSliceEnd = math::min(
                 HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
             const auto IWTildeSliceEnd = math::min(
                 WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
 
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
             const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
             const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
 
             // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
             const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
             const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
+            if constexpr(NDimSpatial == 2)
+            {
                 // A: output tensor
                 const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
-                out_n_ho_wo_k_grid_desc,
+                    out_grid_desc,
                     make_tuple(make_pass_through_transform(N),
                                make_pad_transform(Ho, I0, I0),
                                make_pad_transform(Wo, I0, I0),
@@ -244,11 +370,103 @@ struct TransformConvBwdDataToGemm_v1
 
                 return out_gemmak0_gemmm_gemmak1_grid_desc;
             }
+            else if constexpr(NDimSpatial == 3)
+            {
+                // A: output tensor
+                const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                    out_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_pad_transform(Do, I0, I0),
+                               make_pad_transform(Ho, I0, I0),
+                               make_pad_transform(Wo, I0, I0),
+                               make_pass_through_transform(K)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+                const auto out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc =
+                    transform_tensor_descriptor(
+                        out_n_hop_wop_k_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_embed_transform(
+                                       make_tuple(ZDot, DTilde),
+                                       make_tuple(-ConvDilationD / GcdStrideDilationD, I1)),
+                                   make_embed_transform(
+                                       make_tuple(YDot, HTilde),
+                                       make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                                   make_embed_transform(
+                                       make_tuple(XDot, WTilde),
+                                       make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                                   make_pass_through_transform(K)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1, 2>{},
+                                   Sequence<3, 4>{},
+                                   Sequence<5, 6>{},
+                                   Sequence<7>{}));
+
+                const auto
+                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc =
+                        transform_tensor_descriptor(
+                            out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc,
+                            make_tuple(make_pass_through_transform(N),
+                                       make_slice_transform(ZDot, I0, ZDotSlice),
+                                       make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                       make_slice_transform(YDot, I0, YDotSlice),
+                                       make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                       make_slice_transform(XDot, I0, XDotSlice),
+                                       make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                       make_unmerge_transform(make_tuple(AK0, AK1))),
+                            make_tuple(Sequence<0>{},
+                                       Sequence<1>{},
+                                       Sequence<2>{},
+                                       Sequence<3>{},
+                                       Sequence<4>{},
+                                       Sequence<5>{},
+                                       Sequence<6>{},
+                                       Sequence<7>{}),
+                            make_tuple(Sequence<0>{},
+                                       Sequence<1>{},
+                                       Sequence<2>{},
+                                       Sequence<3>{},
+                                       Sequence<4>{},
+                                       Sequence<5>{},
+                                       Sequence<6>{},
+                                       Sequence<7, 8>{}));
+
+                const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
+                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc,
+                    make_tuple(
+                        make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, AK0)),
+                        make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                        make_pass_through_transform(AK1)),
+                    make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}, Sequence<8>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                const auto out_gemmak0_gemmm_gemmak1_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        out_gemmak0_gemmmraw_gemmak1_grid_desc,
+                        make_tuple(AK0, GemmMPerBlock, AK1),
+                        Sequence<false, DoPadGemmM, false>{});
+
+                return out_gemmak0_gemmm_gemmak1_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
+        }
     }
 
     template <typename BLayout,
-              typename std::enable_if<NDimSpatial == 2 &&
-                                          is_same_v<BLayout, tensor_layout::convolution::GKYXC>,
+              typename std::enable_if<(NDimSpatial == 2 || NDimSpatial == 3) &&
+                                          (is_same_v<BLayout, tensor_layout::convolution::GKYXC> ||
+                                           is_same_v<BLayout, tensor_layout::convolution::GKZYXC>),
                                       bool>::type = false>
     static auto MakeBDescriptor_BK0_N_BK1(
         const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
@@ -263,30 +481,35 @@ struct TransformConvBwdDataToGemm_v1
         const std::array<index_t, NDimSpatial>& /* input_right_pads */,
         const std::array<index_t, NDimSpatial>& tildes)
     {
-        index_t i_ytilde = tildes[0];
-        index_t i_xtilde = tildes[1];
+        index_t i_ztilde = tildes[ZIdx - NonSpatialDimsNum];
+        index_t i_ytilde = tildes[YIdx - NonSpatialDimsNum];
+        index_t i_xtilde = tildes[XIdx - NonSpatialDimsNum];
 
         const index_t N = in_g_n_c_wis_lengths[1];
         const index_t K = wei_g_k_c_xs_lengths[1];
         const index_t C = wei_g_k_c_xs_lengths[2];
 
-        const index_t Ho = out_g_n_k_wos_lengths[3];
-        const index_t Wo = out_g_n_k_wos_lengths[4];
+        const index_t Do = NDimSpatial == 3 ? out_g_n_k_wos_lengths[DIdx] : 1;
+        const index_t Ho = out_g_n_k_wos_lengths[HIdx];
+        const index_t Wo = out_g_n_k_wos_lengths[WIdx];
 
-        const index_t Y = wei_g_k_c_xs_lengths[3];
-        const index_t X = wei_g_k_c_xs_lengths[4];
+        const index_t Z = NDimSpatial == 3 ? wei_g_k_c_xs_lengths[ZIdx] : 1;
+        const index_t Y = wei_g_k_c_xs_lengths[YIdx];
+        const index_t X = wei_g_k_c_xs_lengths[XIdx];
 
-        const index_t ConvStrideH = conv_filter_strides[0];
-        const index_t ConvStrideW = conv_filter_strides[1];
+        const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+        const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+        const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvDilationH = conv_filter_dilations[0];
-        const index_t ConvDilationW = conv_filter_dilations[1];
+        const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+        const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+        const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
         const index_t BK0 = K / BK1;
 
         // assume packed
-        const auto wei_k_y_x_c_grid_desc =
-            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+        // k_y_x_c for 2d or k_z_y_x_c for 3d
+        const auto wei_grid_desc = make_wei_grid_desc<BLayout>(K, Z, Y, X, C);
 
         if constexpr(ConvBwdDataSpecialization ==
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
@@ -299,7 +522,7 @@ struct TransformConvBwdDataToGemm_v1
                                                        make_pass_through_transform(C)),
                                             make_tuple(Sequence<0>{}, Sequence<1>{}),
                                             make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-            make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, C), make_tuple(I0, I1));
+            make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, C), make_tuple(I0, I1));
 
             const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
                 ck::tensor_operation::device::PadTensorDescriptor(
@@ -311,23 +534,30 @@ struct TransformConvBwdDataToGemm_v1
         }
         else
         {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
             const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
             const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
             const auto YTilde = ConvStrideH / GcdStrideDilationH;
             const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
             const auto YDot = math::integer_divide_ceil(Y, YTilde);
             const auto XDot = math::integer_divide_ceil(X, XTilde);
 
             // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
             const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
             const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
             // B weight tensor
+            if constexpr(NDimSpatial == 2)
+            {
                 const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
-                wei_k_y_x_c_grid_desc,
-                make_tuple(make_pass_through_transform(K),
+                    wei_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(K),
                         make_embed_transform(make_tuple(YDot, YTilde),
                                              make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
                         make_embed_transform(make_tuple(XDot, XTilde),
@@ -337,7 +567,8 @@ struct TransformConvBwdDataToGemm_v1
                     make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
 
                 const auto wei_bk0_bk1_ydotslice_xdotslice_c_grid_desc =
-                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                    transform_tensor_descriptor(
+                        wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
                         make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
                                    make_slice_transform(YDot, I0, YDotSlice),
                                    make_slice_transform(XDot, I0, XDotSlice),
@@ -368,18 +599,98 @@ struct TransformConvBwdDataToGemm_v1
                 const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
                     ck::tensor_operation::device::PadTensorDescriptor(
                         wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
+                        make_tuple(wei_gemmbk0_gemmnraw_gemmbk1_grid_desc.GetLength(I0),
+                                   GemmNPerBlock,
+                                   BK1),
+                        Sequence<false, DoPadGemmN, false>{});
+
+                return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                const auto wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc =
+                    transform_tensor_descriptor(
+                        wei_grid_desc,
+                        make_tuple(
+                            make_pass_through_transform(K),
+                            make_embed_transform(make_tuple(ZDot, ZTilde),
+                                                 make_tuple(ConvStrideD / GcdStrideDilationD, I1)),
+                            make_embed_transform(make_tuple(YDot, YTilde),
+                                                 make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                            make_embed_transform(make_tuple(XDot, XTilde),
+                                                 make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                            make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1, 2>{},
+                                   Sequence<3, 4>{},
+                                   Sequence<5, 6>{},
+                                   Sequence<7>{}));
+
+                const auto wei_bk0_bk1_zdotslice_ydotslice_xdotslice_c_grid_desc =
+                    transform_tensor_descriptor(
+                        wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                        make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                   make_slice_transform(ZDot, I0, ZDotSlice),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_freeze_transform(i_ztilde),
+                                   make_freeze_transform(i_ytilde),
+                                   make_freeze_transform(i_xtilde),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<3>{},
+                                   Sequence<5>{},
+                                   Sequence<2>{},
+                                   Sequence<4>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0, 1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<>{},
+                                   Sequence<>{},
+                                   Sequence<>{},
+                                   Sequence<5>{}));
+
+                const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc = transform_tensor_descriptor(
+                    wei_bk0_bk1_zdotslice_ydotslice_xdotslice_c_grid_desc,
                     make_tuple(
-                        wei_gemmbk0_gemmnraw_gemmbk1_grid_desc.GetLength(I0), GemmNPerBlock, BK1),
+                        make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, BK0)),
+                        make_pass_through_transform(C),
+                        make_pass_through_transform(BK1)),
+                    make_tuple(Sequence<2, 3, 4, 0>{}, Sequence<5>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
+                        make_tuple(wei_gemmbk0_gemmnraw_gemmbk1_grid_desc.GetLength(I0),
+                                   GemmNPerBlock,
+                                   BK1),
                         Sequence<false, DoPadGemmN, false>{});
 
                 return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
             }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
+        }
     }
 
     template <typename CLayout,
-              typename std::enable_if<NDimSpatial == 2 &&
+              typename std::enable_if<(NDimSpatial == 2 || NDimSpatial == 3) &&
                                           (is_same_v<CLayout, tensor_layout::convolution::GNHWC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::GNDHWC> ||
                                            is_same_v<CLayout, tensor_layout::convolution::NHWGC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::NDHWGC> ||
                                            is_same_v<CLayout, tensor_layout::convolution::G_NHW_C>),
                                       bool>::type = false>
     static auto
@@ -395,49 +706,57 @@ struct TransformConvBwdDataToGemm_v1
                         const std::array<index_t, NDimSpatial>& input_right_pads,
                         const std::array<index_t, NDimSpatial>& tildes)
     {
-        index_t i_ytilde = tildes[0];
-        index_t i_xtilde = tildes[1];
+        index_t i_ztilde = tildes[ZIdx - NonSpatialDimsNum];
+        index_t i_ytilde = tildes[YIdx - NonSpatialDimsNum];
+        index_t i_xtilde = tildes[XIdx - NonSpatialDimsNum];
 
         const index_t N = in_g_n_c_wis_lengths[1];
         const index_t C = wei_g_k_c_xs_lengths[2];
 
-        const index_t Hi = in_g_n_c_wis_lengths[3];
-        const index_t Wi = in_g_n_c_wis_lengths[4];
+        const index_t Di = NDimSpatial == 3 ? in_g_n_c_wis_lengths[DIdx] : 1;
+        const index_t Hi = in_g_n_c_wis_lengths[HIdx];
+        const index_t Wi = in_g_n_c_wis_lengths[WIdx];
 
-        const index_t Ho = out_g_n_k_wos_lengths[3];
-        const index_t Wo = out_g_n_k_wos_lengths[4];
+        const index_t Do = NDimSpatial == 3 ? out_g_n_k_wos_lengths[DIdx] : 1;
+        const index_t Ho = out_g_n_k_wos_lengths[HIdx];
+        const index_t Wo = out_g_n_k_wos_lengths[WIdx];
 
-        const index_t Y = wei_g_k_c_xs_lengths[3];
-        const index_t X = wei_g_k_c_xs_lengths[4];
+        const index_t Z = NDimSpatial == 3 ? wei_g_k_c_xs_lengths[ZIdx] : 1;
+        const index_t Y = wei_g_k_c_xs_lengths[YIdx];
+        const index_t X = wei_g_k_c_xs_lengths[XIdx];
 
-        const index_t InLeftPadH = input_left_pads[0];
-        const index_t InLeftPadW = input_left_pads[1];
+        const index_t InLeftPadD = input_left_pads[DIdx - NonSpatialDimsNum];
+        const index_t InLeftPadH = input_left_pads[HIdx - NonSpatialDimsNum];
+        const index_t InLeftPadW = input_left_pads[WIdx - NonSpatialDimsNum];
 
-        const index_t InRightPadH = input_right_pads[0];
-        const index_t InRightPadW = input_right_pads[1];
+        const index_t InRightPadD = input_right_pads[DIdx - NonSpatialDimsNum];
+        const index_t InRightPadH = input_right_pads[HIdx - NonSpatialDimsNum];
+        const index_t InRightPadW = input_right_pads[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvStrideH = conv_filter_strides[0];
-        const index_t ConvStrideW = conv_filter_strides[1];
+        const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+        const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+        const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvDilationH = conv_filter_dilations[0];
-        const index_t ConvDilationW = conv_filter_dilations[1];
+        const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+        const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+        const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
         // assume strided
-        const auto in_n_hi_wi_c_grid_desc =
-            make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
-                                         make_tuple(in_g_n_c_wis_strides[1],
-                                                    in_g_n_c_wis_strides[3],
-                                                    in_g_n_c_wis_strides[4],
-                                                    in_g_n_c_wis_strides[2]));
+        // n_hi_wi_c for 2d n_di_hi_wi_c for 3d
+        const auto in_grid_desc =
+            make_in_grid_desc<NDimSpatial, CLayout>(N, Di, Hi, Wi, C, in_g_n_c_wis_strides);
 
         if constexpr(ConvBwdDataSpecialization ==
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
                          Filter1x1Stride1Pad0)
         {
             // C: input tensor
+            if constexpr(NDimSpatial == 2)
+            {
                 const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
-                in_n_hi_wi_c_grid_desc,
-                make_tuple(make_pass_through_transform(N),
+                    in_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
                         make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
                         make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
                         make_pass_through_transform(C)),
@@ -453,7 +772,51 @@ struct TransformConvBwdDataToGemm_v1
                     make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
                     make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
 
-            const auto in_gemmm_gemmn_grid_desc = ck::tensor_operation::device::PadTensorDescriptor(
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+
+                // C: input tensor
+                const auto in_n_x_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                    in_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(I1, Do), make_tuple(I1, ConvStrideD)),
+                        make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                        make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                        make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+                const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_x_do_y_ho_x_wo_c_grid_desc,
+                    make_tuple(make_freeze_transform(I0),
+                               make_freeze_transform(I0),
+                               make_freeze_transform(I0),
+                               make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<1>{},
+                               Sequence<3>{},
+                               Sequence<5>{},
+                               Sequence<0, 2, 4, 6>{},
+                               Sequence<7>{}),
+                    make_tuple(
+                        Sequence<>{}, Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
                         in_gemmmraw_gemmnraw_grid_desc,
                         make_tuple(GemmMPerBlock, GemmNPerBlock),
                         Sequence<DoPadGemmM, DoPadGemmN>{});
@@ -462,34 +825,51 @@ struct TransformConvBwdDataToGemm_v1
             }
             else
             {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
+        }
+        else
+        {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
             const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
             const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
             const auto YTilde = ConvStrideH / GcdStrideDilationH;
             const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
             const auto HTilde =
                 Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
             const auto WTilde =
                 Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
 
-            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            // only work on DTilde, HTilde and WTilde that contribute to
+            // non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
             const auto IHTildeSliceBegin = math::integer_divide_floor(
                 math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
             const auto IWTildeSliceBegin = math::integer_divide_floor(
                 math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
 
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
             const auto IHTildeSliceEnd = math::min(
                 HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
             const auto IWTildeSliceEnd = math::min(
                 WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
 
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
             const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
             const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
 
             // C: input tensor
+            if constexpr(NDimSpatial == 2)
+            {
                 const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
-                in_n_hi_wi_c_grid_desc,
+                    in_grid_desc,
                     make_tuple(make_pass_through_transform(N),
                                make_pad_transform(Hi, InLeftPadH, InRightPadH),
                                make_pad_transform(Wi, InLeftPadW, InRightPadW),
@@ -497,7 +877,8 @@ struct TransformConvBwdDataToGemm_v1
                     make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
                     make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
 
-            const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+                    transform_tensor_descriptor(
                         in_n_hip_wip_c_grid_desc,
                         make_tuple(make_pass_through_transform(N),
                                    make_embed_transform(make_tuple(YTilde, HTilde),
@@ -506,7 +887,8 @@ struct TransformConvBwdDataToGemm_v1
                                                         make_tuple(ConvDilationW, ConvStrideW)),
                                    make_pass_through_transform(C)),
                         make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
-                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+                        make_tuple(
+                            Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
 
                 const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
                     in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
@@ -536,13 +918,98 @@ struct TransformConvBwdDataToGemm_v1
                     make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
                     make_tuple(Sequence<0>{}, Sequence<1>{}));
 
-            const auto in_gemmm_gemmn_grid_desc = ck::tensor_operation::device::PadTensorDescriptor(
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
                         in_gemmmraw_gemmnraw_grid_desc,
                         make_tuple(GemmMPerBlock, GemmNPerBlock),
                         Sequence<DoPadGemmM, DoPadGemmN>{});
 
                 return in_gemmm_gemmn_grid_desc;
             }
+            else if(NDimSpatial == 3)
+            {
+                const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                    in_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_pad_transform(Di, InLeftPadD, InRightPadD),
+                               make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                               make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                               make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+                const auto in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+                    transform_tensor_descriptor(
+                        in_n_dip_hip_wip_c_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_embed_transform(make_tuple(ZTilde, DTilde),
+                                                        make_tuple(ConvDilationD, ConvStrideD)),
+                                   make_embed_transform(make_tuple(YTilde, HTilde),
+                                                        make_tuple(ConvDilationH, ConvStrideH)),
+                                   make_embed_transform(make_tuple(XTilde, WTilde),
+                                                        make_tuple(ConvDilationW, ConvStrideW)),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1, 2>{},
+                                   Sequence<3, 4>{},
+                                   Sequence<5, 6>{},
+                                   Sequence<7>{}));
+
+                const auto in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc =
+                    transform_tensor_descriptor(
+                        in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_freeze_transform(i_ztilde),
+                                   make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                   make_freeze_transform(i_ytilde),
+                                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                   make_freeze_transform(i_xtilde),
+                                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<>{},
+                                   Sequence<1>{},
+                                   Sequence<>{},
+                                   Sequence<2>{},
+                                   Sequence<>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{}));
+
+                const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc,
+                    make_tuple(
+                        make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                        make_pass_through_transform(C)),
+                    make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
+        }
     }
 
     // for input bias

include/ck/utility/amd_buffer_addressing.hpp

@@ -629,7 +629,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
 {
     static_assert(
         (is_same<T, double>::value && (N == 1 || N == 2)) ||
-            (is_same<T, float>::value && (N == 1 || N == 2 || N == 4)) ||
+            (is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
             (is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
             (is_same<T, bhalf_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
             (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4)) ||
@@ -682,6 +682,20 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
                                                 dst_wave_addr_offset,
                                                 static_cast<index_t>(coherence));
         }
+        else if constexpr(N == 8)
+        {
+            vector_type<float, 8> tmp{src_thread_data};
+            llvm_amdgcn_raw_buffer_store_fp32x4(tmp.AsType<float4_t>()[Number<0>{}],
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+            llvm_amdgcn_raw_buffer_store_fp32x4(tmp.AsType<float4_t>()[Number<1>{}],
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset + 4 * sizeof(float),
+                                                static_cast<index_t>(coherence));
+        }
     }
     else if constexpr(is_same<T, half_t>::value)
     {

include/ck/utility/inner_product.hpp

@@ -13,13 +13,13 @@ __device__ void inner_product(const TA& a, const TB& b, TC& c);
 template <>
 __device__ void inner_product<float, float, float>(const float& a, const float& b, float& c)
 {
-#if CK_USE_AMD_INNER_PRODUCT_INLINE_ASM && defined(CK_USE_AMD_V_MAC_F32)
+#if CK_USE_AMD_V_MAC_INLINE_ASM && defined(CK_USE_AMD_V_MAC_F32)
     asm volatile("\n \
             v_mac_f32 %0, %1, %2 \n \
             "
                  : "=v"(c)
                  : "v"(a), "v"(b), "0"(c));
-#elif CK_USE_AMD_INNER_PRODUCT_INLINE_ASM && defined(CK_USE_AMD_V_FMAC_F32)
+#elif CK_USE_AMD_V_MAC_INLINE_ASM && defined(CK_USE_AMD_V_FMAC_F32)
     asm volatile("\n \
             v_fmac_f32 %0, %1, %2 \n \
             "
@@ -76,14 +76,18 @@ template <>
 __device__ void inner_product<half2_t, half2_t, float>(const half2_t& a, const half2_t& b, float& c)
 {
 #if defined(CK_USE_AMD_V_DOT2_F32_F16)
-#if CK_USE_AMD_INNER_PRODUCT_INLINE_ASM
+#if CK_USE_AMD_V_DOT_INLINE_ASM
+    // Use 3 x s_nop to avoid hazard (mi200 cdna2 isa page 47
+    // https://www.amd.com/system/files/TechDocs/instinct-mi200-cdna2-instruction-set-architecture.pdf
+    // ) s_nop with parameter 2 is equal to 3 x s_nop
     asm volatile("\n \
             v_dot2_f32_f16 %0, %1, %2, %0\n \
+            s_nop 2 \n \
             "
                  : "=v"(c)
                  : "v"(a), "v"(b), "0"(c));
 #else
-    c = __builtin_amdgcn_sdot2(a, b, c, false);
+    c = __builtin_amdgcn_fdot2(a, b, c, false);
 #endif
 #else
     const vector_type<half_t, 2> a_vector{a};
@@ -163,9 +167,13 @@ __device__ void
 inner_product<int8x4_t, int8x4_t, int32_t>(const int8x4_t& a, const int8x4_t& b, int32_t& c)
 {
 #if defined(CK_USE_AMD_V_DOT4_I32_I8)
-#if CK_USE_AMD_INNER_PRODUCT_INLINE_ASM
+#if CK_USE_AMD_V_DOT_INLINE_ASM
+    // Use 3 x s_nop to avoid hazard (mi200 cdna2 isa page 47
+    // https://www.amd.com/system/files/TechDocs/instinct-mi200-cdna2-instruction-set-architecture.pdf
+    // ) s_nop with parameter 2 is equal to 3 x s_nop
     asm volatile("\n \
             v_dot4_i32_i8 %0, %1, %2, %0\n \
+            s_nop 2 \n \
             "
                  : "=v"(c)
                  : "v"(bit_cast<int32_t>(a)), "v"(bit_cast<int32_t>(b)), "0"(c));

include/ck/utility/magic_division.hpp

@@ -157,4 +157,76 @@ struct MagicDivision
     }
 };
 
+struct MDiv
+{
+    // 1 dword -> 3 dword storage
+    uint32_t divisor;
+    uint32_t multiplier;
+    uint32_t shift; // TODO: 8 bit is enough
+
+    // prefer construct on host
+    __host__ __device__ MDiv(uint32_t divisor_) : divisor(divisor_)
+    {
+        auto tmp = MagicDivision::CalculateMagicNumbers(divisor_);
+
+        multiplier = tmp[Number<0>{}];
+        shift      = tmp[Number<1>{}];
+    }
+
+    __host__ __device__ MDiv() : divisor(0), multiplier(0), shift(0) {}
+
+    __host__ __device__ void update(uint32_t divisor_)
+    {
+        divisor  = divisor_;
+        auto tmp = MagicDivision::CalculateMagicNumbers(divisor_);
+
+        multiplier = tmp[Number<0>{}];
+        shift      = tmp[Number<1>{}];
+    }
+
+    __host__ __device__ uint32_t div(uint32_t dividend_) const
+    {
+        return MagicDivision::DoMagicDivision(dividend_, multiplier, shift);
+    }
+
+    __host__ __device__ void
+    divmod(uint32_t dividend_, uint32_t& quotient_, uint32_t& remainder_) const
+    {
+        quotient_  = div(dividend_);
+        remainder_ = dividend_ - (quotient_ * divisor);
+    }
+
+    __host__ __device__ uint32_t get() const { return divisor; }
+};
+
+struct MDiv2
+{
+    // 1 dword -> 2 dword storage, divisor need compute from runtime
+    uint32_t multiplier;
+    uint32_t shift; // TODO: 8 bit is enough
+
+    // prefer construct on host
+    __host__ __device__ MDiv2(uint32_t divisor_)
+    {
+        auto tmp = MagicDivision::CalculateMagicNumbers(divisor_);
+
+        multiplier = tmp[Number<0>{}];
+        shift      = tmp[Number<1>{}];
+    }
+
+    __host__ __device__ MDiv2() : multiplier(0), shift(0) {}
+
+    __host__ __device__ uint32_t div(uint32_t dividend_) const
+    {
+        return MagicDivision::DoMagicDivision(dividend_, multiplier, shift);
+    }
+
+    __host__ __device__ void
+    divmod(uint32_t dividend_, uint32_t divisor_, uint32_t& quotient_, uint32_t& remainder_) const
+    {
+        quotient_  = div(dividend_);
+        remainder_ = dividend_ - (quotient_ * divisor_);
+    }
+};
+
 } // namespace ck

include/ck/utility/math.hpp

@@ -240,5 +240,21 @@ struct less
     __host__ __device__ constexpr bool operator()(T x, T y) const { return x < y; }
 };
 
+template <index_t X>
+__host__ __device__ constexpr auto next_power_of_two()
+{
+    // TODO: X need to be 2 ~ 0x7fffffff. 0, 1, or larger than 0x7fffffff will compile fail
+    constexpr index_t Y = 1 << (32 - __builtin_clz(X - 1));
+    return Y;
+}
+
+template <index_t X>
+__host__ __device__ constexpr auto next_power_of_two(Number<X> x)
+{
+    // TODO: X need to be 2 ~ 0x7fffffff. 0, 1, or larger than 0x7fffffff will compile fail
+    constexpr index_t Y = 1 << (32 - __builtin_clz(x.value - 1));
+    return Number<Y>{};
+}
+
 } // namespace math
 } // namespace ck

include/ck/utility/type_convert.hpp

@@ -62,24 +62,6 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, half_t>(half_
     return type_convert<bhalf_t>(x_fp32);
 }
 
-// convert bfp16 to int32 via fp32
-template <>
-inline __host__ __device__ constexpr int32_t type_convert<int32_t, bhalf_t>(bhalf_t x)
-{
-    float x_fp32 = type_convert<float>(x);
-
-    return static_cast<int32_t>(x_fp32);
-}
-
-// convert int32 to bfp16 via fp32
-template <>
-inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int32_t>(int32_t x)
-{
-    float x_fp32 = static_cast<float>(x);
-
-    return type_convert<bhalf_t>(x_fp32);
-}
-
 // convert bfp16 to int8 via fp32
 template <>
 inline __host__ __device__ constexpr int8_t type_convert<int8_t, bhalf_t>(bhalf_t x)

include/ck/utility/workgroup_barrier.hpp

+#pragma once
+#include <hip/hip_runtime.h>
+#include <stdint.h>
+
+namespace ck {
+struct workgroup_barrier
+{
+    __device__ workgroup_barrier(uint32_t* ptr) : base_ptr(ptr) {}
+
+    __device__ uint32_t ld(uint32_t offset)
+    {
+#if 0
+        float d = llvm_amdgcn_raw_buffer_load_fp32(
+                        amdgcn_make_buffer_resource(base_ptr),
+                        0,
+                        offset,
+                        AMDGCN_BUFFER_GLC);
+        union cvt {
+            float f32;
+            uint32_t u32;
+        };
+        cvt x;
+        x.f32 = d;
+        return x.u32;
+#endif
+        return __atomic_load_n(base_ptr + offset, __ATOMIC_RELAXED);
+    }
+
+    __device__ void wait_eq(uint32_t offset, uint32_t value)
+    {
+        if(threadIdx.x == 0)
+        {
+            while(ld(offset) != value) {}
+        }
+        __syncthreads();
+    }
+
+    __device__ void wait_lt(uint32_t offset, uint32_t value)
+    {
+        if(threadIdx.x == 0)
+        {
+            while(ld(offset) < value) {}
+        }
+        __syncthreads();
+    }
+
+    __device__ void wait_set(uint32_t offset, uint32_t compare, uint32_t value)
+    {
+        if(threadIdx.x == 0)
+        {
+            while(atomicCAS(base_ptr + offset, compare, value) != compare) {}
+        }
+        __syncthreads();
+    }
+
+    // enter critical zoon, assume buffer is zero when launch kernel
+    __device__ void aquire(uint32_t offset) { wait_set(offset, 0, 1); }
+
+    // exit critical zoon, assume buffer is zero when launch kernel
+    __device__ void release(uint32_t offset) { wait_set(offset, 1, 0); }
+
+    __device__ void inc(uint32_t offset)
+    {
+        __syncthreads();
+        if(threadIdx.x == 0)
+        {
+            atomicAdd(base_ptr + offset, 1);
+        }
+    }
+
+    uint32_t* base_ptr;
+};
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp

@@ -14,6 +14,7 @@
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
 
+#ifdef DL_KERNELS
 namespace ck {
 namespace tensor_operation {
 namespace device {
@@ -335,3 +336,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck
+#endif

library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp

@@ -39,7 +39,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
     std::vector<std::unique_ptr<
         DeviceConvBwdData<1, NWC, KXC, NWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#ifdef __int8__
 void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<1,
                                                   NWC,
@@ -51,7 +51,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#endif
 // conv2d backward data
 void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<2,
@@ -88,7 +88,7 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#ifdef __int8__
 void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<2,
                                                   NHWC,
@@ -100,7 +100,7 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#endif
 // conv2d dl
 void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<2,
@@ -125,7 +125,7 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#ifdef __int8__
 void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<2,
                                                   NHWC,
@@ -137,6 +137,7 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
+#endif
 // conv3d backward data
 void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<3,
@@ -173,7 +174,7 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#ifdef __int8__
 void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
     std::vector<std::unique_ptr<DeviceConvBwdData<3,
                                                   NDHWC,
@@ -185,7 +186,7 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
                                                   PassThrough,
                                                   PassThrough,
                                                   PassThrough>>>& instances);
-
+#endif
 template <ck::index_t NumDimSpatial,
           typename InLayout,
           typename WeiLayout,
@@ -239,11 +240,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
             {
                 add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(op_ptrs);
             }
+#ifdef __int8__
             else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
                               is_same_v<OutDataType, int8_t>)
             {
                 add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(op_ptrs);
             }
+#endif
         }
         else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
                           is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
@@ -266,12 +269,14 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
             {
                 add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
             }
+#ifdef __int8__
             else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
                               is_same_v<OutDataType, int8_t>)
             {
                 add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
                 add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
             }
+#endif
         }
         else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
                           is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
@@ -292,11 +297,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
             {
                 add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(op_ptrs);
             }
+#ifdef __int8__
             else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
                               is_same_v<OutDataType, int8_t>)
             {
                 add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(op_ptrs);
             }
+#endif
         }
 
         return op_ptrs;

library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp

@@ -17,6 +17,7 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
+#if defined(__fp16__) && defined(DL_KERNELS)
 void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
@@ -56,11 +57,11 @@ void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#endif
+#if defined(__fp32__) && defined(DL_KERNELS)
 void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
-
         instances);
 
 void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(
@@ -77,7 +78,8 @@ void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#endif
+#if defined(__int8__) && defined(DL_KERNELS)
 void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
@@ -117,30 +119,32 @@ void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
-void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
+#endif
+#ifdef __int8__
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
-void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
+#endif
+#ifdef __fp16__
+void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
 void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
@@ -163,64 +167,67 @@ void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
         DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+#endif
+#ifdef __bf16__
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
-void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
+#endif
+#ifdef __fp32__
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
 void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(
@@ -242,7 +249,8 @@ void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#endif
+#ifdef __fp64__
 void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
     std::vector<std::unique_ptr<
 
@@ -263,7 +271,7 @@ void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
         instances);
-
+#endif
 template <typename ALayout,
           typename BLayout,
           typename CLayout,
@@ -302,28 +310,36 @@ struct DeviceOperationInstanceFactory<
                          is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+#endif
                 add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+#endif
                 add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+#endif
                 add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+#endif
                 add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs);
             }
         }
@@ -334,16 +350,20 @@ struct DeviceOperationInstanceFactory<
                          is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs);
+#endif
                 add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs);
+#endif
                 add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
                 add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
             }
@@ -351,16 +371,20 @@ struct DeviceOperationInstanceFactory<
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs);
+#endif
                 add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs);
+#endif
                 add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
             }
         }
@@ -388,6 +412,7 @@ struct DeviceOperationInstanceFactory<
                 add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(op_ptrs);
             }
         }
+#ifdef __int8__
         else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
                           is_same_v<CDataType, int8_t>)
         {
@@ -395,32 +420,40 @@ struct DeviceOperationInstanceFactory<
                          is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
                 add_device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instances(op_ptrs);
+#endif
             }
             else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
                 add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(op_ptrs);
+#endif
             }
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(op_ptrs);
                 add_device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances(op_ptrs);
+#endif
             }
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
                 add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(op_ptrs);
                 add_device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances(op_ptrs);
+#endif
             }
         }
-
+#endif
         return op_ptrs;
     }
 };

library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmStreamK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmStreamK<
+    ALayout,
+    BLayout,
+    CLayout,
+    ADataType,
+    BDataType,
+    CDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemmStreamK<ALayout,
+                                       BLayout,
+                                       CLayout,
+                                       ADataType,
+                                       BDataType,
+                                       CDataType,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#if 0
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<CDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+#endif
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_instance.hpp → library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp

@@ -24,13 +24,7 @@ using Empty_Tuple = ck::Tuple<>;
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
 
-using NHWGC = ck::tensor_layout::convolution::NHWGC;
-using GNHWC = ck::tensor_layout::convolution::GNHWC;
-
-using GKYXC = ck::tensor_layout::convolution::GKYXC;
-
-using NHWGK = ck::tensor_layout::convolution::NHWGK;
-using GNHWK = ck::tensor_layout::convolution::GNHWK;
+using namespace ck::tensor_layout::convolution;
 
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
@@ -40,97 +34,103 @@ static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
 
 // f16_f16_f32_f16
-template <typename ALayout,
+template <index_t NDimSpatial,
+          typename ALayout,
           typename BLayout,
           typename DsLayout,
           typename ELayout,
           ConvolutionBackwardDataSpecialization ConvSpec>
-using device_grouped_conv2d_bwd_data_xdl_f16_instances = std::tuple<
+using device_grouped_conv_bwd_data_xdl_f16_instances =
+    std::tuple<
         // clang-format off
         // ##############################################|       NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
         // ##############################################|    Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
         // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
         // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
-        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>
-        
-#ifdef CK_WORKAROUND_SWDEV_3318619
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>
+        
+#ifndef CK_WORKAROUND_GITHUB_ISSUE_824
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
 #endif
         // clang-format on
         >;
 
 // bf16_bf16_f32_bf16
-template <typename ALayout,
+template <index_t NDimSpatial,
+          typename ALayout,
           typename BLayout,
           typename DsLayout,
           typename ELayout,
           ConvolutionBackwardDataSpecialization ConvSpec>
-using device_grouped_conv2d_bwd_data_xdl_bf16_instances = std::tuple<
+using device_grouped_conv_bwd_data_xdl_bf16_instances =
+    std::tuple<
         // clang-format off
         // ##############################################|       NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
         // ##############################################|    Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
         // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
         // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
-        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   BF16,   BF16,     F32,      BF16, Empty_Tuple,   BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>
-        
-#ifdef CK_WORKAROUND_SWDEV_3318619
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   BF16,   BF16,     F32,      BF16, Empty_Tuple,   BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>
+        
+#ifndef CK_WORKAROUND_GITHUB_ISSUE_824
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+            DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>
 #endif
         // clang-format on
         >;
 
 // f32_f32_f32_f32
-template <typename ALayout,
+template <index_t NDimSpatial,
+          typename ALayout,
           typename BLayout,
           typename DsLayout,
           typename ELayout,
           ConvolutionBackwardDataSpecialization ConvSpec>
-using device_grouped_conv2d_bwd_data_xdl_f32_instances = std::tuple<
+using device_grouped_conv_bwd_data_xdl_f32_instances =
+    std::tuple<
         // clang-format off
          // ##############################################|       NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
          // ##############################################|    Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
          // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
          // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
-         DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  F32,  F32,      F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                4>
-
-#ifdef CK_WORKAROUND_SWDEV_3318619
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+         DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  F32,  F32,      F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                4>
+
+#ifndef CK_WORKAROUND_GITHUB_ISSUE_824
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
 #endif
         // clang-format on
         >;

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using namespace ck::tensor_layout::convolution;
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec>
+using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |         |          |          |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec>
+using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |         |          |          |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec>
+using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |         |          |          |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 32, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp

@@ -16,7 +16,7 @@ namespace device {
 namespace instance {
 
 // conv2d backward data
-void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   GNHWK,
                                                                   GKYXC,
@@ -30,7 +30,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   GNHWK,
                                                                   GKYXC,
@@ -44,7 +44,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   GNHWK,
                                                                   GKYXC,
@@ -58,7 +58,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   NHWGK,
                                                                   GKYXC,
@@ -72,7 +72,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   NHWGK,
                                                                   GKYXC,
@@ -86,7 +86,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   NHWGK,
                                                                   GKYXC,
@@ -100,6 +100,91 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
+// conv3d backward data
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
 template <ck::index_t NumDimSpatial,
           typename OutLayout,
           typename WeiLayout,
@@ -139,43 +224,96 @@ struct DeviceOperationInstanceFactory<
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(NumDimSpatial == 2)
+        {
 
-        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
-                     is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+            if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
+                         is_same_v<OutLayout, GNHWK>)
             {
                 if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
                              is_same_v<OutDataType, F16>)
                 {
-                add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
+                    add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(op_ptrs);
                 }
                 else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
                                   is_same_v<OutDataType, F32>)
                 {
-                add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
+                    add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(op_ptrs);
                 }
                 else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
                                   is_same_v<OutDataType, BF16>)
                 {
-                add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs);
+                    add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
+                        op_ptrs);
                 }
             }
-        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
-                          is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
+            else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
+                              is_same_v<OutLayout, NHWGK>)
             {
                 if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
                              is_same_v<OutDataType, F16>)
                 {
-                add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
+                    add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(op_ptrs);
                 }
                 else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
                                   is_same_v<OutDataType, F32>)
                 {
-                add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
+                    add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(op_ptrs);
                 }
                 else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
                                   is_same_v<OutDataType, BF16>)
                 {
-                add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
+                    add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
+                        op_ptrs);
+                }
+            }
+        }
+        else if constexpr(NumDimSpatial == 3)
+        {
+
+            if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
+                         is_same_v<OutLayout, GNDHWK>)
+            {
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
+                        op_ptrs);
+                }
+            }
+            else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
+                              is_same_v<OutLayout, NDHWGK>)
+            {
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
+                        op_ptrs);
+                }
             }
         }
 

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp

@@ -91,6 +91,42 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
                                                            PassThrough,
                                                            PassThrough>>>& instances);
 
+void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
 // conv3d backward weight
 void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
@@ -128,6 +164,42 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances
                                                            PassThrough,
                                                            PassThrough>>>& instances);
 
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
 template <ck::index_t NumDimSpatial,
           typename InLayout,
           typename WeiLayout,
@@ -162,8 +234,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
 
-        if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC> &&
-                     is_same_v<WeiLayout, GKXC> && is_same_v<OutLayout, GNWK>)
+        if constexpr(NumDimSpatial == 1)
+        {
+            if constexpr(is_same_v<InLayout, GNWC> && is_same_v<WeiLayout, GKXC> &&
+                         is_same_v<OutLayout, GNWK>)
             {
                 if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
                              is_same_v<OutDataType, float>)
@@ -175,35 +249,68 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
                 }
-            else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
                                   is_same_v<OutDataType, ck::bhalf_t>)
                 {
                     add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
                         op_ptrs);
                 }
             }
-        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
-                          is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+        }
+        else if constexpr(NumDimSpatial == 2)
+        {
+            if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
+                         is_same_v<OutLayout, GNHWK>)
             {
                 if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
                              is_same_v<OutDataType, float>)
                 {
-                add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+                        op_ptrs);
                 }
                 else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
                                   is_same_v<OutDataType, half_t>)
                 {
-                add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+                        op_ptrs);
                 }
-            else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
                                   is_same_v<OutDataType, ck::bhalf_t>)
                 {
                     add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
                         op_ptrs);
                 }
             }
-        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> &&
-                          is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, GNDHWK>)
+            else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
+                              is_same_v<OutLayout, NHWGK>)
+            {
+                if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                             is_same_v<OutDataType, float>)
+                {
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                                  is_same_v<OutDataType, half_t>)
+                {
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
+                                  is_same_v<OutDataType, ck::bhalf_t>)
+                {
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
+                        op_ptrs);
+                }
+            }
+        }
+        else if constexpr(NumDimSpatial == 3)
+        {
+            if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
+                         is_same_v<OutLayout, GNDHWK>)
             {
                 if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
                              is_same_v<OutDataType, float>)
@@ -217,13 +324,38 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                     add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
                         op_ptrs);
                 }
-            else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> &&
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
                                   is_same_v<OutDataType, ck::bhalf_t>)
                 {
                     add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
                         op_ptrs);
                 }
             }
+            else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
+                              is_same_v<OutLayout, NDHWGK>)
+            {
+                if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                             is_same_v<OutDataType, float>)
+                {
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                                  is_same_v<OutDataType, half_t>)
+                {
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
+                                  is_same_v<OutDataType, ck::bhalf_t>)
+                {
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
+                        op_ptrs);
+                }
+            }
+        }
 
         return op_ptrs;
     }

library/include/ck/library/utility/device_memory.hpp

@@ -20,8 +20,9 @@ __global__ void set_buffer_value(T* p, T x, uint64_t buffer_element_size)
  */
 struct DeviceMem
 {
-    DeviceMem() = delete;
+    DeviceMem() : mpDeviceBuf(nullptr), mMemSize(0) {}
     DeviceMem(std::size_t mem_size);
+    void Realloc(std::size_t mem_size);
     void* GetDeviceBuffer() const;
     std::size_t GetBufferSize() const;
     void ToDevice(const void* p) const;

library/src/tensor_operation_instance/gpu/CMakeLists.txt

@@ -12,9 +12,46 @@ set(CK_DEVICE_INSTANCES)
 FOREACH(subdir_path ${dir_list})
 set(target_dir)
 IF(IS_DIRECTORY "${subdir_path}")
+    set(cmake_instance)
+    file(READ "${subdir_path}/CMakeLists.txt" cmake_instance)
+    set(add_inst 0)
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp8\" " AND DTYPES MATCHES "fp8")
+            #message("fp8 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp16\"" AND DTYPES MATCHES "fp16")
+            #message("fp16 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp32\"" AND DTYPES MATCHES "fp32")
+            #message("fp32 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp64\"" AND DTYPES MATCHES "fp64")
+            #message("fp64 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf16\"" AND DTYPES MATCHES "bf16")
+            #message("bf16 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"int8\"" AND DTYPES MATCHES "int8")
+            #message("int8 instance found!")
+            set(add_inst 1)
+    endif()
+    if(NOT "${cmake_instance}" MATCHES "DTYPES" OR NOT DEFINED DTYPES)
+            #message("instance should be built for all types!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "ONLY DL_KERNELS" AND NOT DEFINED DL_KERNELS)
+	    message("Found only dl instances, but DL_KERNELS is not set. Skipping.")
+            set(add_inst 0)
+    endif()
+    if(add_inst EQUAL 1)
       get_filename_component(target_dir ${subdir_path} NAME)
       add_subdirectory(${target_dir})
       list(APPEND CK_DEVICE_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+    endif()
 ENDIF()
 ENDFOREACH()
 

library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp

@@ -31,22 +31,24 @@ using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
 using device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
     std::tuple<
         // clang-format off
-        //##################################################| A0Layout| B0Layout|       D0Layout| B1Layout|      D1sLayout| E1Layout| A0Data| B0Data| Acc0DataType|     D0DataType| B1Data| Acc1CData| CShuffle|        D1sData| E1Data|          A0|          B0|          CDE0|          B1|          CDE1| PadGemm0M| PadGemm0N| PadGemm0K| PadGemm1N| PadGemm1K|NumGemm0K| Block|  Gemm0| Gemm0| Gemm0| Gemm1| Gemm1|A0K1|B0K1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|   C1Shuffle|   C1Shuffle| CDE1BlockTransferClusterLengths| CDE1BlockTransfer|
-        //##################################################|         |         |               |         |               |         |   Type|   Type|         Type|           Type|   Type|      Type| DataType|           Type|   Type| Elementwise| Elementwise|   Elementwise| Elementwise|   Elementwise|          |          |          |          |          | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|  ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MWaveMPerXdl|   ScalarPerVector|
-        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |   Operation|   Operation|     Operation|   Operation|     Operation|          |          |          |          |          |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per|Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|            _NBlock_NWaveNPerXdl|     _NWaveNPerXdl|
-        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |            |            |              |            |              |          |          |          |          |          |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|               |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                                |                  |
+        //##################################################| A0Layout| B0Layout|       D0Layout| B1Layout|      D1sLayout| E1Layout| A0Data| B0Data| Acc0DataType|     D0DataType| B1Data| Acc1CData| CShuffle|        D1sData| E1Data|          A0|          B0|          CDE0|          B1|          CDE1| PadGemm0M| PadGemm0N| PadGemm0K| PadGemm1N| PadGemm1K|NumGemm0K| Block|  Gemm0| Gemm0| Gemm0| Gemm1| Gemm1|A0K1|B0K1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CDE0BlockTransfer| CDE0BlockTransfer|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|   C1Shuffle|   C1Shuffle| CDE1BlockTransferClusterLengths| CDE1BlockTransfer|
+        //##################################################|         |         |               |         |               |         |   Type|   Type|         Type|           Type|   Type|      Type| DataType|           Type|   Type| Elementwise| Elementwise|   Elementwise| Elementwise|   Elementwise|          |          |          |          |          | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|  ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|      SrcVectorDim|         SrcScalar|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MWaveMPerXdl|   ScalarPerVector|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |   Operation|   Operation|     Operation|   Operation|     Operation|          |          |          |          |          |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per|Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |                  |         PerVector|  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|            _NBlock_NWaveNPerXdl|     _NWaveNPerXdl|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |            |            |              |            |              |          |          |          |          |          |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|               |               |               |               |               |               |          |                 |                |                |                |                |                |           |                  |                  |                 |                |                |                |                |                |           |            |            |                                |                  |
+        //generic
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 1,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
         // no padding
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,                  S<1, 16, 1,16>,               8>,
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,                  S<1, 32, 1, 8>,               8>,
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,                  S<1, 16, 1,16>,               8>,
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,                 9,                 4,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,                  S<1, 32, 1, 8>,               8>,
         // Padded fallback kernel
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
-        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>
         // clang-format on
         >;