Merge branch 'develop' into sphinx_doc

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

@@ -1077,14 +1077,6 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle
                 }
             } // end gemm1
 
-            // workaround compiler issue; see ck/ck.hpp
-            if constexpr(CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE == 1 &&
-                         is_same_v<FloatAB, bhalf_t> && MPerBlock == 256 && NPerBlock == 128 &&
-                         Gemm1NPerBlock == 128)
-            {
-                __builtin_amdgcn_sched_barrier(0);
-            }
-
             constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
                 gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
             constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);

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

@@ -879,14 +879,6 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
                 }
             } // end gemm1
 
-            // workaround compiler issue; see ck/ck.hpp
-            if constexpr(CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE == 1 &&
-                         is_same_v<FloatAB, bhalf_t> && MPerBlock == 256 && NPerBlock == 128 &&
-                         Gemm1NPerBlock == 128)
-            {
-                __builtin_amdgcn_sched_barrier(0);
-            }
-
             constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
                 gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
             constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.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_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseOp,
+          typename ADataType,
+          typename BDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle(
+            const ADataType* __restrict__ p_a_grid,
+            const BDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                p_b_grid + b_batch_offset,
+                                                p_ds_grid_grp,
+                                                p_e_grid + e_batch_offset,
+                                                p_shared,
+                                                a_grid_desc_k0_m_k1,
+                                                b_grid_desc_k0_n_k1,
+                                                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                a_element_op,
+                                                b_element_op,
+                                                cde_element_op,
+                                                block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+template <typename GridwiseOp,
+          typename ADataType,
+          typename BDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_multiple_d_wmma_cshuffle(
+            const ADataType* __restrict__ p_a_grid,
+            const BDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const index_t batch_count,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+            const Block2CTileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
+    // printf("entry kernel launch");
+    __shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
+
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    // printf("before compute_ptr_offset call");
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    DsPointer p_ds_grid_grp;
+
+    // printf("before allocate pointer d");
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    // printf("before entry");
+
+    GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                p_b_grid + b_batch_offset,
+                                                p_ds_grid_grp,
+                                                p_e_grid + e_batch_offset,
+                                                p_shared,
+                                                a_grid_desc_k0_m_k1,
+                                                b_grid_desc_k0_n_k1,
+                                                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                a_element_op,
+                                                b_element_op,
+                                                cde_element_op,
+                                                block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+    ignore = compute_ptr_offset_of_batch;
+#endif
+}
+
+template <typename GridwiseOp,
+          typename ADataType,
+          typename BDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_mupltipe_d_wmma_cshuffle(
+            const ADataType* __restrict__ p_a_grid,
+            const BDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
+    __shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
+
+    GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                p_b_grid,
+                                                p_ds_grid,
+                                                p_e_grid,
+                                                p_shared,
+                                                a_grid_desc_k0_m_k1,
+                                                b_grid_desc_k0_n_k1,
+                                                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                a_element_op,
+                                                b_element_op,
+                                                cde_element_op,
+                                                block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx1100__))
+}
+
+template < // DataType Family
+    typename ADataType,
+    typename BDataType,
+    typename AccDataType,
+    typename CShuffleDataType,
+    typename DsDataType,
+    typename EDataType,
+    // InMemory Data Descriptor
+    typename AGridDesc_K0_M_K1,
+    typename BGridDesc_K0_N_K1,
+    typename DsGridDesc_M_N,
+    typename EGridDesc_M_N,
+    // ElementwiseOp Family
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CDEElementwiseOperation,
+    InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+    // Tiling Family
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t MPerWmma,
+    index_t NPerWmma,
+    index_t K1Value,
+    index_t MRepeat,
+    index_t NRepeat,
+    // ThreadCluster Family
+    index_t BlockSize,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_K1,
+    bool AThreadTransferSrcResetCoordinateAfterRun,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_K1,
+    bool BThreadTransferSrcResetCoordinateAfterRun,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMRepeatPerShuffle,
+    index_t CShuffleNRepeatPerShuffle,
+    typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+    index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+    index_t NumGemmKPrefetchStage = 1,
+    LoopScheduler LoopSched       = make_default_loop_scheduler(),
+    PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0perblock_mperblock_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0perblock_mperblock_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0perblock_nperblock_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0perblock_nperblock_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    // *Caution Here repeat is shuffle repeat
+    GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerWmma);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerWmma);
+
+        constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMRepeatPerShuffle * MWave * MPerWmma>{},
+                           I1,
+                           Number<CShuffleNRepeatPerShuffle * NWave * NPerWmma>{}));
+
+        return c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0perblock_mperblock_k1 =
+            GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0perblock_nperblock_k1 =
+            GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size_aligned * sizeof(ADataType) +
+                b_block_space_size_aligned * sizeof(BDataType));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const DsGridDesc_M_N& ds_grid_desc_m_n,
+                  const EGridDesc_M_N& e_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerWmma * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerWmma)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / (K0PerBlock * K1);
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N_& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N_& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const EGridDesc_M_N& e_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
+        MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(EGridDesc_M_N{}, 1, 1))>;
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                               const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                               const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        // printf("safe entry");
+        // clang-format off
+/*******************************************************************************/
+// Memory buffer zone.
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+/*******************************************************************************/
+// BlockIdx.x -> [BlockId.m, BlockId.n]
+        const auto block_work_idx = block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        { return; }
+
+        // Store BlockId into SGPR
+        const index_t m_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+        const index_t n_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+/*******************************************************************************/
+// BlockLevel, A/B Matrix ThreadMapping in LDS, As Destinaion of BlockWise_Copy
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+        constexpr auto max_lds_align = K1;
+        constexpr auto a_block_desc_k0perblock_mperblock_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+        constexpr auto b_block_desc_k0perblock_nperblock_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<        ThisThreadBlock,
+/* typename SrcElementwiseOperation,              */    AElementwiseOperation,
+/* typename DstElementwiseOperation,              */    ck::tensor_operation::element_wise::PassThrough,
+/* InMemoryDataOperationEnum DstInMemOp,          */    InMemoryDataOperationEnum::Set,
+/* typename BlockSliceLengths,                    */    Sequence<K0PerBlock, MPerBlock, K1>,
+/* typename ThreadClusterLengths,                 */    ABlockTransferThreadClusterLengths_K0_M_K1,
+/* typename ThreadClusterArrangeOrder,            */    ABlockTransferThreadClusterArrangeOrder,
+/* typename SrcData,                              */    ADataType,
+/* typename DstData,                              */    ADataType,
+/* typename SrcDesc,                              */    decltype(a_grid_desc_k0_m_k1),
+/* typename DstDesc,                              */    decltype(a_block_desc_k0perblock_mperblock_k1),
+/* typename SrcDimAccessOrder,                    */    ABlockTransferSrcAccessOrder,
+/* typename DstDimAccessOrder,                    */    Sequence<0, 1, 2>,
+/* index_t SrcVectorDim,                          */    ABlockTransferSrcVectorDim,
+/* index_t DstVectorDim,                          */    2,
+/* index_t SrcScalarPerVector,                    */    ABlockTransferSrcScalarPerVector,
+/* index_t DstScalarPerVector,                    */    ABlockTransferDstScalarPerVector_K1,
+/* index_t SrcScalarStrideInVector,               */    1,
+/* index_t DstScalarStrideInVector,               */    1,
+/* bool ThreadTransferSrcResetCoordinateAfterRun, */    AThreadTransferSrcResetCoordinateAfterRun,
+/* bool ThreadTransferDstResetCoordinateAfterRun, */    true>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0perblock_mperblock_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,
+                                                BDataType,
+                                                BDataType,
+                                                decltype(b_grid_desc_k0_n_k1),
+                                                decltype(b_block_desc_k0perblock_nperblock_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0perblock_nperblock_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+/*******************************************************************************/
+        // GEMM
+        constexpr auto WmmaK = 16;
+        constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
+
+        auto blockwise_gemm =
+            BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<BlockSize,
+                                                         ADataType,
+                                                         BDataType,
+                                                         AccDataType,
+                                                         decltype(a_block_desc_k0perblock_mperblock_k1),
+                                                         decltype(b_block_desc_k0perblock_nperblock_k1),
+                                                         MPerWmma,
+                                                         NPerWmma,
+                                                         MRepeat,
+                                                         NRepeat,
+                                                         KPack>{};
+
+        // Prepare Register for C matrix
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+/*******************************************************************************/
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<ADataType*>(p_shared), a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<BDataType*>(p_shared) + a_block_space_size_aligned, b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize());
+        
+        // Shift Per SUB_K
+        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);
+
+        // gridwise GEMM pipeline
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
+                                                          a_block_desc_k0perblock_mperblock_k1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_k0_n_k1,
+                                                          b_block_desc_k0perblock_nperblock_k1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          K0BlockMainLoop);
+/*******************************************************************************/
+        //printf("safe 1");
+        // write out to C, implement shuffle
+        {
+            constexpr auto c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs =  
+            blockwise_gemm.GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
+
+            // This API Provide All dimension (size) you need
+            constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp =
+                blockwise_gemm.GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
+
+            constexpr auto MWave              = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I1);
+            constexpr auto MSubGroup          = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I2);
+            constexpr auto NWave              = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I4);
+            constexpr auto NThreadPerSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I5);
+            constexpr auto MAccVgprs          = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I6);
+
+            // LDS descriptor, shuffle and write out in MRepeat x NRepeat times
+            constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
+                GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs = transform_tensor_descriptor(
+                c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMRepeatPerShuffle>{}, // MRepeat per shuffle repeat
+                        MWave,                               // MWave
+                        MSubGroup,                           // MSubGroup * MAccVgprs = MPerWmma
+                        MAccVgprs)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNRepeatPerShuffle>{}, // NRepeat per shuffle repeat
+                        NWave,                               // NWave
+                        NThreadPerSubGroup))),               // NThreadPerSubGroup = NPerWmma
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0, 1, 2, 6>{}, Sequence<>{}, Sequence<3, 4, 5>{}));
+
+            // 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);
+
+            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_mrepeat_mwave_msubgroup_maccvgprs_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(MRepeat, MWave, MSubGroup, MAccVgprs))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(NRepeat, NWave, NThreadPerSubGroup))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+            
+            const auto m_thread_data_on_block_idx = m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor.CalculateBottomIndex(
+                make_multi_index(m_thread_data_on_block));
+            
+            const auto n_thread_data_on_block_idx = n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor.CalculateBottomIndex(
+                make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
+                                                   decltype(c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            CShuffleNRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            MAccVgprs>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6>,
+                                                   6,
+                                                   1, // vector write pixel
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
+                    make_multi_index(0,
+                                     m_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     0,
+                                     n_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+            
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+            
+            // tuple of reference to C/Ds tensor buffers
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            // shuffle: blockwise copy C from LDS to global
+            auto cde_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,            // ThreadGroup
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation,    // ElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMRepeatPerShuffle * MWave * MPerWmma,
+                         1,
+                         CShuffleNRepeatPerShuffle * NWave * NPerWmma>, // BlockSliceLengths,
+                CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>,   // bool ThreadTransferSrcResetCoordinateAfterRun,
+                Sequence<false>> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
+                 cde_element_op};
+
+            // space filling curve for local reg & global memory
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MRepeat, 1, 1, NRepeat, 1, 1, MAccVgprs>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6>,
+                                  Sequence<CShuffleMRepeatPerShuffle,
+                                           1,
+                                           1,
+                                           CShuffleNRepeatPerShuffle,
+                                           1,
+                                           1,
+                                           MAccVgprs>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_cde_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMRepeatPerShuffle * MWave * MPerWmma,
+                                           1,
+                                           CShuffleNRepeatPerShuffle * NWave * NPerWmma>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_shuffle_block_copy_lds_to_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_global_step = sfc_cde_global.GetForwardStep(access_id);
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_shuffle_block_copy_lds_to_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_global_step);
+                    });
+
+                    // move on E
+                    cde_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_global_step);
+                }
+            });
+        }
+        // clang-format on
+    }
+};
+
+} // namespace ck

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

@@ -4,9 +4,8 @@
 #pragma once
 
 #include "ck/utility/data_type.hpp"
-#include "ck/utility/reduction_common.hpp"
+
 #include "ck/utility/reduction_operator.hpp"
-#include "ck/utility/reduction_functions_accumulate.hpp"
 #include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
 #include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
@@ -19,8 +18,8 @@ template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
           typename YDataType,
-          typename AccDataType,
-          typename AccElementwiseOperation,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
           typename GridDesc_M_K,
           index_t BlockSize,
           index_t MThreadClusterSize,
@@ -46,6 +45,10 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
                       (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
                   "Invalid thread slice sizes and/or vector sizes configuration, please check!");
 
+    static_assert(XSrcVectorSize == YDstVectorSize);
+    static_assert(XSrcVectorSize == GammaSrcVectorSize);
+    static_assert(XSrcVectorSize == BetaSrcVectorSize);
+
     static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
 
     using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
@@ -59,19 +62,23 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
     static constexpr auto thread_cluster_desc =
         make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
 
+    using ThreadBufferLengths_M_K                = Sequence<MThreadSliceSize, XSrcVectorSize>;
+    static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
     using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
-        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
     using ThreadReduceDstDesc_M =
         decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
 
-    using BlockwiseSumReduce = PartitionedBlockwiseReduction<AccDataType,
+    using BlockwiseSumReduce = PartitionedBlockwiseReduction<ComputeDataType,
                                                              BlockSize,
                                                              ThreadClusterLengths_M_K,
                                                              ThreadClusterArrangeOrder,
                                                              reduce::Add,
                                                              true>;
 
-    using ThreadwiseSumReduce = ThreadwiseReduction<AccDataType,
+    using ThreadwiseSumReduce = ThreadwiseReduction<ComputeDataType,
                                                     ThreadReduceSrcDesc_M_K,
                                                     ThreadReduceDstDesc_M,
                                                     reduce::Add,
@@ -81,64 +88,70 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
 
-    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
-    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t M_BlockTileSize     = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
+
+    static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
 
     __device__ static void Run(const GridDesc_M_K& x_grid_desc_m_k,
                                const GridDesc_M_K& gamma_grid_desc_m_k,
                                const GridDesc_M_K& beta_grid_desc_m_k,
                                const GridDesc_M_K& y_grid_desc_m_k,
                                index_t num_k_block_tile_iteration,
-                               AccDataType epsilon,
+                               ComputeDataType epsilon,
                                const XDataType* const __restrict__ p_x_global,
                                const GammaDataType* const __restrict__ p_gamma_global,
                                const BetaDataType* const __restrict__ p_beta_global,
                                YDataType* const __restrict__ p_y_global,
-                               const AccElementwiseOperation acc_elementwise_op)
+                               const YElementwiseOperation y_elementwise_op)
     {
-        if constexpr(SweepOnce)
-        {
-            num_k_block_tile_iteration = 1;
-        }
-
         // LDS
-        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
-
-        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
-            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
+        __shared__ ComputeDataType p_reduce_work_buffer[BlockSize];
 
         auto reduce_work_buf =
             make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
 
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
-            x_thread_buf;
-
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
-            gamma_thread_buf;
-
-        StaticBuffer<AddressSpaceEnum::Vgpr,
-                     AccDataType,
-                     MThreadSliceSize * KThreadSliceSize,
-                     true>& beta_thread_buf = gamma_thread_buf;
-
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
-            y_thread_buf;
-
-        StaticBuffer<AddressSpaceEnum::Vgpr,
-                     AccDataType,
-                     MThreadSliceSize * KThreadSliceSize,
-                     true>& x_square_thread_buf = y_thread_buf;
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
 
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
-            mean_square_thread_buf;
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>& var_thread_buf =
+        auto x_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * XSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto gamma_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * GammaSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto& beta_thread_buf = gamma_thread_buf;
+
+        auto y_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * YDstVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto& x_square_thread_buf = y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
             mean_square_thread_buf;
-
-        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
-            mean_thread_buf(I)        = reduce::Add::template GetIdentityValue<AccDataType>();
-            mean_square_thread_buf(I) = reduce::Add::template GetIdentityValue<AccDataType>();
-        });
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>&
+            var_thread_buf = mean_square_thread_buf;
 
         const index_t thread_local_id = get_thread_local_1d_id();
         const index_t block_global_id = get_block_1d_id();
@@ -149,12 +162,8 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
         const auto thread_m_cluster_id = thread_cluster_idx[I0];
         const auto thread_k_cluster_id = thread_cluster_idx[I1];
 
-        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
-        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
-            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
-
         auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
-                                                                  AccDataType,
+                                                                  ComputeDataType,
                                                                   GridDesc_M_K,
                                                                   decltype(thread_buffer_desc_m_k),
                                                                   ThreadBufferLengths_M_K,
@@ -166,11 +175,11 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
             x_grid_desc_m_k,
             make_multi_index(block_global_id * M_BlockTileSize +
                                  thread_m_cluster_id * MThreadSliceSize,
-                             thread_k_cluster_id * KThreadSliceSize));
+                             thread_k_cluster_id * XSrcVectorSize));
 
         auto threadwise_gamma_load =
             ThreadwiseTensorSliceTransfer_v2<GammaDataType,
-                                             AccDataType,
+                                             ComputeDataType,
                                              GridDesc_M_K,
                                              decltype(thread_buffer_desc_m_k),
                                              ThreadBufferLengths_M_K,
@@ -182,11 +191,11 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
                 gamma_grid_desc_m_k,
                 make_multi_index(block_global_id * M_BlockTileSize +
                                      thread_m_cluster_id * MThreadSliceSize,
-                                 thread_k_cluster_id * KThreadSliceSize));
+                                 thread_k_cluster_id * GammaSrcVectorSize));
 
         auto threadwise_beta_load =
             ThreadwiseTensorSliceTransfer_v2<BetaDataType,
-                                             AccDataType,
+                                             ComputeDataType,
                                              GridDesc_M_K,
                                              decltype(thread_buffer_desc_m_k),
                                              ThreadBufferLengths_M_K,
@@ -198,14 +207,14 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
                 beta_grid_desc_m_k,
                 make_multi_index(block_global_id * M_BlockTileSize +
                                      thread_m_cluster_id * MThreadSliceSize,
-                                 thread_k_cluster_id * KThreadSliceSize));
+                                 thread_k_cluster_id * BetaSrcVectorSize));
 
         auto threadwise_y_store =
-            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
                                                YDataType,
                                                decltype(thread_buffer_desc_m_k),
                                                GridDesc_M_K,
-                                               AccElementwiseOperation,
+                                               YElementwiseOperation,
                                                ThreadBufferLengths_M_K,
                                                ThreadBufferDimAccessOrder,
                                                YDstVectorDim,
@@ -216,13 +225,10 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
                 y_grid_desc_m_k,
                 make_multi_index(block_global_id * M_BlockTileSize +
                                      thread_m_cluster_id * MThreadSliceSize,
-                                 thread_k_cluster_id * KThreadSliceSize),
-                acc_elementwise_op);
+                                 thread_k_cluster_id * YDstVectorSize),
+                y_elementwise_op);
 
-        // Copy x from Cache
-        // one pass: fwd, second pass: bwd
-        constexpr auto thread_copy_fwd_step_m_k =
-            make_multi_index(0, SweepOnce ? 0 : K_BlockTileSize);
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
         constexpr auto thread_copy_bwd_step_m_k =
             make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
 
@@ -239,121 +245,260 @@ struct GridwiseNormalizationNaiveVariance_mk_to_mk
         // FIXME: Should not hack the transform from deviceOP
         int reduce_length = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
 
-        index_t reducedTiles = 0;
-        do
-        {
-            threadwise_x_load.Run(x_grid_desc_m_k,
-                                  x_global_val_buf,
-                                  thread_buffer_desc_m_k,
-                                  make_tuple(I0, I0),
-                                  x_thread_buf);
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I)        = reduce::Add::template GetIdentityValue<ComputeDataType>();
+            mean_square_thread_buf(I) = reduce::Add::template GetIdentityValue<ComputeDataType>();
+        });
 
-            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
-                    constexpr auto offset_m_k =
-                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
-                    x_square_thread_buf(Number<offset_m_k>{}) =
-                        x_thread_buf(Number<offset_m_k>{}) * x_thread_buf(Number<offset_m_k>{});
+        // Separate sweep once and sweep twice pipeline
+        if constexpr(SweepOnce)
+        {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_val_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf(i));
+
+                threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                          gamma_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          gamma_thread_buf(i));
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                        x_square_thread_buf(i)(Number<offset_m_k>{}) =
+                            x_thread_buf(i)(Number<offset_m_k>{}) *
+                            x_thread_buf(i)(Number<offset_m_k>{});
+                    });
                 });
-            });
 
-            ThreadwiseSumReduce::Reduce(x_thread_buf, mean_thread_buf);
-            ThreadwiseSumReduce::Reduce(x_square_thread_buf, mean_square_thread_buf);
+                ThreadwiseSumReduce::Reduce(x_thread_buf[i], mean_thread_buf);
+                ThreadwiseSumReduce::Reduce(x_square_thread_buf[i], mean_square_thread_buf);
 
-            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                if constexpr(i != ThreadBufferNumber - 1)
+                {
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                }
+            });
 
-            ++reducedTiles;
-        } while(reducedTiles < num_k_block_tile_iteration);
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
+                mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
 
-        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
-            if constexpr(I > 0)
                 block_sync_lds();
 
-            BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
-            mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
+                mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
+
+                // var(x) = E[x^2] - E[x]^2
+                var_thread_buf(I) =
+                    mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
+            });
 
-            block_sync_lds();
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // normalize
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                            divisor;
+
+                        // gamma & beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                            gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
 
-            BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
-            mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                         beta_global_val_buf,
+                                         thread_buffer_desc_m_k,
+                                         make_tuple(I0, I0),
+                                         beta_thread_buf(i));
 
-            // var(x) = E[x^2] - E[x]^2
-            var_thread_buf(I) =
-                mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
-        });
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+            });
 
-        // y = (x - E[x]) / sqrt(var[x] + epsilon)
-        auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k;
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                            beta_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
 
-        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
-        threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
-        threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
-        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                       make_tuple(I0, I0),
+                                       y_thread_buf(i),
+                                       y_grid_desc_m_k,
+                                       y_global_val_buf);
 
-        reducedTiles = 0;
-        do
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+            });
+        } // end of sweep once
+        else
         {
-            if constexpr(!SweepOnce)
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
             {
-                threadwise_x_load.Run(x_grid_desc_m_k,
-                                      x_global_val_buf,
-                                      thread_buffer_desc_m_k,
-                                      make_tuple(I0, I0),
-                                      x_thread_buf);
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+
+                    static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                            x_square_thread_buf(i)(Number<offset_m_k>{}) =
+                                x_thread_buf(i)(Number<offset_m_k>{}) *
+                                x_thread_buf(i)(Number<offset_m_k>{});
+                        });
+                    });
+
+                    ThreadwiseSumReduce::Reduce(x_thread_buf[i], mean_thread_buf);
+                    ThreadwiseSumReduce::Reduce(x_square_thread_buf[i], mean_square_thread_buf);
+                });
             }
 
-            threadwise_gamma_load.Run(gamma_grid_desc_m_k,
-                                      gamma_global_val_buf,
-                                      thread_buffer_desc_m_k,
-                                      make_tuple(I0, I0),
-                                      gamma_thread_buf);
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
 
-            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
-                    constexpr auto offset_m_k =
-                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
-
-                    // normalize
-                    y_thread_buf(Number<offset_m_k>{}) =
-                        (x_thread_buf(Number<offset_m_k>{}) - mean_thread_buf(iM)) /
-                        sqrt(var_thread_buf(iM) + epsilon);
-
-                    // gamma
-                    y_thread_buf(Number<offset_m_k>{}) =
-                        y_thread_buf(Number<offset_m_k>{}) * gamma_thread_buf(Number<offset_m_k>{});
-                });
-            });
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
+                mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
 
-            threadwise_beta_load.Run(beta_grid_desc_m_k,
-                                     beta_global_val_buf,
-                                     thread_buffer_desc_m_k,
-                                     make_tuple(I0, I0),
-                                     beta_thread_buf);
+                block_sync_lds();
 
-            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
-                    constexpr auto offset_m_k =
-                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
+                mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
 
-                    // beta
-                    y_thread_buf(Number<offset_m_k>{}) =
-                        y_thread_buf(Number<offset_m_k>{}) + beta_thread_buf(Number<offset_m_k>{});
-                });
+                // var(x) = E[x^2] - E[x]^2
+                var_thread_buf(I) =
+                    mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
             });
 
-            threadwise_y_store.Run(thread_buffer_desc_m_k,
-                                   make_tuple(I0, I0),
-                                   y_thread_buf,
-                                   y_grid_desc_m_k,
-                                   y_global_val_buf);
+            auto thread_copy_tail_m_k =
+                (num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k;
 
             threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
-            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_bwd_step_m_k);
-            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_bwd_step_m_k);
-            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                              gamma_global_val_buf,
+                                              thread_buffer_desc_m_k,
+                                              make_tuple(I0, I0),
+                                              gamma_thread_buf(i));
+
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                });
 
-            ++reducedTiles;
-        } while(reducedTiles < num_k_block_tile_iteration);
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // normalize
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                                divisor;
+
+                            // gamma
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                                gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                             beta_global_val_buf,
+                                             thread_buffer_desc_m_k,
+                                             make_tuple(I0, I0),
+                                             beta_thread_buf(i));
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // beta
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                                beta_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                           make_tuple(I0, I0),
+                                           y_thread_buf(i),
+                                           y_grid_desc_m_k,
+                                           y_global_val_buf);
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+                });
+
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         2 * thread_copy_bwd_step_m_k);
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        2 * thread_copy_bwd_step_m_k);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                      2 * thread_copy_bwd_step_m_k);
+            }
+        } // end of sweep twice
     }
 };
 

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/grid/gridwise_normalization_naive_variance.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp"
+
+namespace ck {
+template <typename GridwiseReduction,
+          typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename GridDesc_M_K>
+__global__ void kernel_normalization(const GridDesc_M_K x_grid_desc_m_k,
+                                     const GridDesc_M_K gamma_grid_desc_m_k,
+                                     const GridDesc_M_K beta_grid_desc_m_k,
+                                     const GridDesc_M_K y_grid_desc_m_k,
+                                     index_t num_k_block_tile_iteration,
+                                     ComputeDataType epsilon,
+                                     const XDataType* const __restrict__ p_x_global,
+                                     const GammaDataType* const __restrict__ p_gamma_global,
+                                     const BetaDataType* const __restrict__ p_beta_global,
+                                     YDataType* const __restrict__ p_y_global,
+                                     const YElementwiseOperation y_elementwise_op)
+{
+    GridwiseReduction::Run(x_grid_desc_m_k,
+                           gamma_grid_desc_m_k,
+                           beta_grid_desc_m_k,
+                           y_grid_desc_m_k,
+                           num_k_block_tile_iteration,
+                           epsilon,
+                           p_x_global,
+                           p_gamma_global,
+                           p_beta_global,
+                           p_y_global,
+                           y_elementwise_op);
+};
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool UseWelford>
+auto NormalizationKernelSelector(bool isSweepOnce)
+{
+    using GridwiseNormalizationGenericNaive =
+        GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    YDataType,
+                                                    ComputeDataType,
+                                                    YElementwiseOperation,
+                                                    GridDesc_M_K,
+                                                    BlockSize,
+                                                    MThreadClusterSize,
+                                                    KThreadClusterSize,
+                                                    MThreadSliceSize,
+                                                    KThreadSliceSize,
+                                                    XSrcVectorDim,
+                                                    XSrcVectorSize,
+                                                    GammaSrcVectorDim,
+                                                    GammaSrcVectorSize,
+                                                    BetaSrcVectorDim,
+                                                    BetaSrcVectorSize,
+                                                    YDstVectorDim,
+                                                    YDstVectorSize,
+                                                    false>;
+    using GridwiseNormalizationSweepOnceNaive =
+        GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    YDataType,
+                                                    ComputeDataType,
+                                                    YElementwiseOperation,
+                                                    GridDesc_M_K,
+                                                    BlockSize,
+                                                    MThreadClusterSize,
+                                                    KThreadClusterSize,
+                                                    MThreadSliceSize,
+                                                    KThreadSliceSize,
+                                                    XSrcVectorDim,
+                                                    XSrcVectorSize,
+                                                    GammaSrcVectorDim,
+                                                    GammaSrcVectorSize,
+                                                    BetaSrcVectorDim,
+                                                    BetaSrcVectorSize,
+                                                    YDstVectorDim,
+                                                    YDstVectorSize,
+                                                    true>;
+    using GridwiseNormalizationGenericWelford =
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      ComputeDataType,
+                                                      YElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      YDstVectorDim,
+                                                      YDstVectorSize,
+                                                      false>;
+    using GridwiseNormalizationSweepOnceWelford =
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      ComputeDataType,
+                                                      YElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      YDstVectorDim,
+                                                      YDstVectorSize,
+                                                      true>;
+
+    if constexpr(UseWelford)
+    {
+        return isSweepOnce ? kernel_normalization<GridwiseNormalizationSweepOnceWelford,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>
+                           : kernel_normalization<GridwiseNormalizationGenericWelford,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>;
+    }
+    else
+    {
+        return isSweepOnce ? kernel_normalization<GridwiseNormalizationSweepOnceNaive,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>
+                           : kernel_normalization<GridwiseNormalizationGenericNaive,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>;
+    }
+}
+
+} // namespace ck

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

@@ -16,8 +16,8 @@ template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
           typename YDataType,
-          typename AccDataType,
-          typename AccElementwiseOperation,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
           typename GridDesc_M_K,
           index_t BlockSize,
           index_t MThreadClusterSize,
@@ -43,6 +43,10 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                       (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
                   "Invalid thread slice sizes and/or vector sizes configuration, please check!");
 
+    static_assert(XSrcVectorSize == YDstVectorSize);
+    static_assert(XSrcVectorSize == GammaSrcVectorSize);
+    static_assert(XSrcVectorSize == BetaSrcVectorSize);
+
     static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
 
     using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
@@ -56,15 +60,19 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
     static constexpr auto thread_cluster_desc =
         make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
 
+    using ThreadBufferLengths_M_K                = Sequence<MThreadSliceSize, XSrcVectorSize>;
+    static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
     using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
         make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
     using ThreadReduceDstDesc_M =
         decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
 
     using ThreadwiseWelford =
-        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+        ThreadwiseWelford<ComputeDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
 
-    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+    using BlockwiseWelford = BlockwiseWelford<ComputeDataType,
                                               BlockSize,
                                               ThreadClusterLengths_M_K,
                                               ThreadClusterArrangeOrder>;
@@ -77,10 +85,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
     static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
     static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
 
-    static constexpr auto XThreadBufferNumber     = Number<KThreadSliceSize / XSrcVectorSize>{};
-    static constexpr auto GammaThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
-    static constexpr auto BetaThreadBufferNumber  = Number<KThreadSliceSize / XSrcVectorSize>{};
-    static constexpr auto YThreadBufferNumber     = Number<KThreadSliceSize / XSrcVectorSize>{};
+    static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
 
     __device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
                                         int thread_k_cluster_id)
@@ -93,7 +98,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
 
         if(kPerBlockTail > 0)
         {
-            static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
                 int thread_max_len =
                     (thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
                 int delta = thread_max_len - kPerBlockTail;
@@ -110,59 +115,41 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                                const GridDesc_M_K& beta_grid_desc_m_k,
                                const GridDesc_M_K& y_grid_desc_m_k,
                                index_t num_k_block_tile_iteration,
-                               AccDataType epsilon,
+                               ComputeDataType epsilon,
                                const XDataType* const __restrict__ p_x_global,
                                const GammaDataType* const __restrict__ p_gamma_global,
                                const BetaDataType* const __restrict__ p_beta_global,
                                YDataType* const __restrict__ p_y_global,
-                               const AccElementwiseOperation acc_elementwise_op)
+                               const YElementwiseOperation y_elementwise_op)
     {
-        if constexpr(SweepOnce)
-        {
-            num_k_block_tile_iteration = 1;
-        }
-
         auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
 
         auto x_thread_buf = generate_tuple(
             [&](auto) {
                 return StaticBuffer<AddressSpaceEnum::Vgpr,
-                                    AccDataType,
+                                    ComputeDataType,
                                     MThreadSliceSize * XSrcVectorSize,
                                     true>{};
             },
-            Number<XThreadBufferNumber>{});
+            Number<ThreadBufferNumber>{});
 
         auto gamma_thread_buf = generate_tuple(
             [&](auto) {
                 return StaticBuffer<AddressSpaceEnum::Vgpr,
-                                    AccDataType,
+                                    ComputeDataType,
                                     MThreadSliceSize * GammaSrcVectorSize,
                                     true>{};
             },
-            Number<GammaThreadBufferNumber>{});
-
-        auto beta_thread_buf = generate_tuple(
-            [&](auto) {
-                return StaticBuffer<AddressSpaceEnum::Vgpr,
-                                    AccDataType,
-                                    MThreadSliceSize * BetaSrcVectorSize,
-                                    true>{};
-            },
-            Number<BetaThreadBufferNumber>{});
+            Number<ThreadBufferNumber>{});
 
-        auto y_thread_buf = generate_tuple(
-            [&](auto) {
-                return StaticBuffer<AddressSpaceEnum::Vgpr,
-                                    AccDataType,
-                                    MThreadSliceSize * YDstVectorSize,
-                                    true>{};
-            },
-            Number<YThreadBufferNumber>{});
+        auto& beta_thread_buf = gamma_thread_buf;
+        auto& y_thread_buf    = x_thread_buf;
 
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
-        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            var_thread_buf;
 
         const index_t thread_local_id = get_thread_local_1d_id();
         const index_t block_global_id = get_block_1d_id();
@@ -173,12 +160,8 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
         const auto thread_m_cluster_id = thread_cluster_idx[I0];
         const auto thread_k_cluster_id = thread_cluster_idx[I1];
 
-        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, XSrcVectorSize>;
-        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
-            make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
-
         auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
-                                                                  AccDataType,
+                                                                  ComputeDataType,
                                                                   GridDesc_M_K,
                                                                   decltype(thread_buffer_desc_m_k),
                                                                   ThreadBufferLengths_M_K,
@@ -194,7 +177,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
 
         auto threadwise_gamma_load =
             ThreadwiseTensorSliceTransfer_v2<GammaDataType,
-                                             AccDataType,
+                                             ComputeDataType,
                                              GridDesc_M_K,
                                              decltype(thread_buffer_desc_m_k),
                                              ThreadBufferLengths_M_K,
@@ -210,7 +193,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
 
         auto threadwise_beta_load =
             ThreadwiseTensorSliceTransfer_v2<BetaDataType,
-                                             AccDataType,
+                                             ComputeDataType,
                                              GridDesc_M_K,
                                              decltype(thread_buffer_desc_m_k),
                                              ThreadBufferLengths_M_K,
@@ -225,11 +208,11 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                                  thread_k_cluster_id * BetaSrcVectorSize));
 
         auto threadwise_y_store =
-            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
                                                YDataType,
                                                decltype(thread_buffer_desc_m_k),
                                                GridDesc_M_K,
-                                               AccElementwiseOperation,
+                                               YElementwiseOperation,
                                                ThreadBufferLengths_M_K,
                                                ThreadBufferDimAccessOrder,
                                                YDstVectorDim,
@@ -241,7 +224,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                 make_multi_index(block_global_id * M_BlockTileSize +
                                      thread_m_cluster_id * MThreadSliceSize,
                                  thread_k_cluster_id * YDstVectorSize),
-                acc_elementwise_op);
+                y_elementwise_op);
 
         constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
         constexpr auto thread_copy_bwd_step_m_k =
@@ -260,67 +243,47 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
         threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k, thread_k_cluster_id);
 
         static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
-            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
-            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+            mean_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<ComputeDataType>(0.0f);
         });
 
-        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        // Separate sweep once and sweep twice pipeline
+        if constexpr(SweepOnce)
         {
-            static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
                 threadwise_x_load.Run(x_grid_desc_m_k,
                                       x_global_val_buf,
                                       thread_buffer_desc_m_k,
                                       make_tuple(I0, I0),
                                       x_thread_buf(i));
-                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
-                threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
-            });
-        }
-
-        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
-            if constexpr(I > 0)
-                block_sync_lds();
 
-            int count = threadwise_welford.cur_count_;
-            BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
-        });
-
-        auto thread_copy_tail_m_k =
-            (num_k_block_tile_iteration - 1) * XThreadBufferNumber * thread_copy_fwd_step_m_k;
-
-        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
-        threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
-        threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
-        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
-
-        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
-        {
-            if constexpr(!SweepOnce)
-            {
-                static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
-                    threadwise_x_load.Run(x_grid_desc_m_k,
-                                          x_global_val_buf,
-                                          thread_buffer_desc_m_k,
-                                          make_tuple(I0, I0),
-                                          x_thread_buf(i));
-                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
-                });
-            }
-
-            static_for<0, GammaThreadBufferNumber, 1>{}([&](auto i) {
                 threadwise_gamma_load.Run(gamma_grid_desc_m_k,
                                           gamma_global_val_buf,
                                           thread_buffer_desc_m_k,
                                           make_tuple(I0, I0),
                                           gamma_thread_buf(i));
 
-                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
-                                                         thread_copy_fwd_step_m_k);
+                threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                {
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                }
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                int count = threadwise_welford.cur_count_;
+                BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
             });
 
             static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
                 auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
-                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
                     static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
                         constexpr auto offset_m_k =
                             thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
@@ -330,7 +293,7 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                             (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
                             divisor;
 
-                        // gamma
+                        // gamma & beta
                         y_thread_buf(iK0)(Number<offset_m_k>{}) =
                             y_thread_buf(iK0)(Number<offset_m_k>{}) *
                             gamma_thread_buf(iK0)(Number<offset_m_k>{});
@@ -338,18 +301,20 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                 });
             });
 
-            static_for<0, BetaThreadBufferNumber, 1>{}([&](auto i) {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
                 threadwise_beta_load.Run(beta_grid_desc_m_k,
                                          beta_global_val_buf,
                                          thread_buffer_desc_m_k,
                                          make_tuple(I0, I0),
                                          beta_thread_buf(i));
-                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
-                                                        thread_copy_fwd_step_m_k);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
             });
 
             static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
-                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
                     static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
                         constexpr auto offset_m_k =
                             thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
@@ -362,22 +327,134 @@ struct GridwiseNormalizationWelfordVariance_mk_to_mk
                 });
             });
 
-            static_for<0, YThreadBufferNumber, 1>{}([&](auto i) {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
                 threadwise_y_store.Run(thread_buffer_desc_m_k,
                                        make_tuple(I0, I0),
                                        y_thread_buf(i),
                                        y_grid_desc_m_k,
                                        y_global_val_buf);
-                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_fwd_step_m_k);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+            });
+        } // end of sweep once
+        else
+        {
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                    threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
+                });
+            }
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                int count = threadwise_welford.cur_count_;
+                BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
             });
 
-            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
-            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
-                                                     2 * thread_copy_bwd_step_m_k);
-            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
-                                                    2 * thread_copy_bwd_step_m_k);
-            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
-        }
+            auto thread_copy_tail_m_k =
+                (num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k;
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                              gamma_global_val_buf,
+                                              thread_buffer_desc_m_k,
+                                              make_tuple(I0, I0),
+                                              gamma_thread_buf(i));
+
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // normalize
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                                divisor;
+
+                            // gamma
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                                gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                             beta_global_val_buf,
+                                             thread_buffer_desc_m_k,
+                                             make_tuple(I0, I0),
+                                             beta_thread_buf(i));
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // beta
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                                beta_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                           make_tuple(I0, I0),
+                                           y_thread_buf(i),
+                                           y_grid_desc_m_k,
+                                           y_global_val_buf);
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+                });
+
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         2 * thread_copy_bwd_step_m_k);
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        2 * thread_copy_bwd_step_m_k);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                      2 * thread_copy_bwd_step_m_k);
+            }
+        } // end of sweep twice
     }
 };
 

include/ck/utility/math_v2.hpp

@@ -83,6 +83,11 @@ static inline __host__ bool isnan(int4_t x)
 };
 #endif
 
+static inline __host__ half_t sqrt(half_t x)
+{
+    return static_cast<half_t>(std::sqrt(static_cast<float>(x)));
+};
+
 static inline __host__ float sqrt(float x) { return std::sqrt(x); };
 
 static inline __host__ double sqrt(double x) { return std::sqrt(x); };
@@ -158,6 +163,11 @@ static inline __device__ bool isnan(half_t x)
     return (xx & 0x7FFF) > 0x7C00;
 };
 
+static inline __device__ half_t sqrt(half_t x)
+{
+    return static_cast<half_t>(__builtin_amdgcn_sqrtf(static_cast<float>(x)));
+};
+
 static inline __device__ float sqrt(float x) { return __builtin_amdgcn_sqrtf(x); };
 
 static inline __device__ double sqrt(double x) { return __builtin_amdgcn_sqrt(x); };

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

@@ -244,6 +244,63 @@ void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(
                                                               PassThrough,
                                                               PassThrough>>>& instances);
 
+// grouped conv3d forward, NDHWGC/KZYXGC/NDHWGK
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
 template <ck::index_t NumDimSpatial,
           typename InLayout,
           typename WeiLayout,
@@ -385,6 +442,31 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(op_ptrs);
             }
         }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
+                          is_same_v<WeiLayout, KZYXGC> && 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_fwd_xdl_ndhwgc_kzyxgc_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_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances(op_ptrs);
+            }
+        }
 
         return op_ptrs;
     }

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt

@@ -3,4 +3,9 @@ add_instance_library(device_grouped_conv3d_fwd_instance
    device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
    device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
    device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
+
+   device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instance.cpp
+   device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instance.cpp
+   device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instance.cpp
+   device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instance.cpp
 )

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.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 BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
+using KZYXGC = ck::tensor_layout::convolution::KZYXGC;
+using NDHWGK = ck::tensor_layout::convolution::NDHWGK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k]
+using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.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 F16 = ck::half_t;
+using F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
+using KZYXGC = ck::tensor_layout::convolution::KZYXGC;
+using NDHWGK = ck::tensor_layout::convolution::NDHWGK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k]
+using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.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 F32 = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
+using KZYXGC = ck::tensor_layout::convolution::KZYXGC;
+using NDHWGK = ck::tensor_layout::convolution::NDHWGK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k]
+using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances =
+    std::tuple<
+        // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+
+        // Filter1x1Stride1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    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>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>
+        // clang-format on
+        >;
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.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 Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
+using KZYXGC = ck::tensor_layout::convolution::KZYXGC;
+using NDHWGK = ck::tensor_layout::convolution::NDHWGK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, wo, k]
+using device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances = std::tuple<
+    // clang-format off
+        // Default
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // Filter1x1Pad0
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+                                                                                                                                                                                                                
+        // Filter1x1Stride1Pad0                                                                                                                                                                                 
+        //########################################|  NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################| Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|        |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<       3, NDHWGC, KZYXGC, Empty_Tuple, NDHWGK, int8_t, int8_t, int32_t,   int8_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp

@@ -64,16 +64,15 @@ using device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances = st
         //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
-        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              2,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
-        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
-        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
-        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   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<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
-        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
         DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
         DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
         DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
-        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   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<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp

@@ -61,15 +61,17 @@ using device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances = st
         //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
-        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
-        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
-        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
-        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
         DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
         DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,    
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
         DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
         DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   256,    32,   8,   8,   32,   32,    1,    4,     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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
         DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
         DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        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>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
     // clang-format on

library/src/tensor_operation_instance/gpu/normalization/device_normalization_f16_instance.cpp

@@ -21,20 +21,25 @@ template <typename OutElementwise, index_t Rank, index_t Reduce>
 // clang-format off
 using device_normalization_f16_instances =
     std::tuple <
-        // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
         DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
         DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
         DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 2, 1, 2, 1, 2, 2>
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
     >;
 // clang-format on
 

library/src/tensor_operation_instance/gpu/normalization/device_normalization_f32_instance.cpp

@@ -19,17 +19,26 @@ using Pass = ck::tensor_operation::element_wise::PassThrough;
 template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_layernorm_f32_instances = std::tuple<
     // clang-format off
-        // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
         DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
         DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
     // clang-format on
     >;
 

profiler/include/profiler/profile_layernorm_impl.hpp

@@ -19,7 +19,7 @@ namespace profiler {
 template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
-          typename AccDataType,
+          typename ComputeDataType,
           typename YDataType,
           index_t Rank>
 bool profile_layernorm_impl(int do_verification,
@@ -86,7 +86,7 @@ bool profile_layernorm_impl(int do_verification,
     using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
                                                                        GammaDataType,
                                                                        BetaDataType,
-                                                                       AccDataType,
+                                                                       ComputeDataType,
                                                                        YDataType,
                                                                        PassThrough,
                                                                        Rank,
@@ -109,7 +109,7 @@ bool profile_layernorm_impl(int do_verification,
                                                                                  GammaDataType,
                                                                                  BetaDataType,
                                                                                  YDataType,
-                                                                                 AccDataType,
+                                                                                 ComputeDataType,
                                                                                  PassThrough,
                                                                                  Rank,
                                                                                  NumReduceDim>;
@@ -181,8 +181,8 @@ bool profile_layernorm_impl(int do_verification,
         {
             y_dev.FromDevice(y.mData.data());
 
-            bool pass = ck::utils::check_err(
-                y.mData, host_y.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+            bool pass =
+                ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
 
             if(do_log)
             {

script/cmake-ck-dev.sh

@@ -10,8 +10,8 @@ cmake
 -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
 -D CMAKE_CXX_FLAGS="-O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD"         \
 -D CMAKE_BUILD_TYPE=Release                                                                       \
--D BUILD_DEV=ON                                                                                   \
--D GPU_TARGETS="gfx908;gfx90a"                                                                    \
+-D BUILD_DEV=OFF                                                                                   \
+-D GPU_TARGETS="gfx90a"                                                                    \
 -D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                                 \
 -D USE_BITINT_EXTENSION_INT4=OFF                                                                  \
 ${MY_PROJECT_SOURCE}

test/normalization/test_groupnorm_fp16.cpp

@@ -12,11 +12,11 @@ template <typename Tuple>
 class TestGroupnorm : public ::testing::Test
 {
     protected:
-    using XDataType     = std::tuple_element_t<0, Tuple>;
-    using GammaDataType = std::tuple_element_t<1, Tuple>;
-    using BetaDataType  = std::tuple_element_t<2, Tuple>;
-    using AccDataType   = std::tuple_element_t<3, Tuple>;
-    using YDataType     = std::tuple_element_t<4, Tuple>;
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using GammaDataType   = std::tuple_element_t<1, Tuple>;
+    using BetaDataType    = std::tuple_element_t<2, Tuple>;
+    using ComputeDataType = std::tuple_element_t<3, Tuple>;
+    using YDataType       = std::tuple_element_t<4, Tuple>;
 
     void Run()
     {
@@ -36,7 +36,7 @@ class TestGroupnorm : public ::testing::Test
                 ck::profiler::profile_groupnorm_impl<XDataType,
                                                      GammaDataType,
                                                      BetaDataType,
-                                                     AccDataType,
+                                                     ComputeDataType,
                                                      YDataType>(true, 2, false, false, length);
             EXPECT_TRUE(success);
         }
@@ -44,7 +44,7 @@ class TestGroupnorm : public ::testing::Test
 };
 
 using KernelTypes = ::testing::Types<
-    // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
     std::tuple<F16, F16, F16, F32, F16>>;
 
 TYPED_TEST_SUITE(TestGroupnorm, KernelTypes);