Merge remote-tracking branch 'origin/develop' into fix_0813

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

@@ -46,13 +46,14 @@ template <typename XDataType,
           typename AccElementwiseOperation,
           index_t Rank,
           index_t NumReduceDim>
-struct DeviceNormalization2 : public BaseOperator
+struct DeviceLayernorm : public BaseOperator
 {
     virtual std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const std::vector<index_t> lengths,
                         const std::vector<index_t> xStrides,
                         const std::vector<index_t> gammaStrides,
                         const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
                         const std::vector<index_t> reduceDims,
                         AccDataType epsilon,
                         const void* p_x,
@@ -72,14 +73,14 @@ template <typename XDataType,
           typename AccElementwiseOperation,
           index_t Rank,
           index_t NumReduceDim>
-using DeviceNormalization2Ptr = std::unique_ptr<DeviceNormalization2<XDataType,
-                                                                     GammaDataType,
-                                                                     BetaDataType,
-                                                                     AccDataType,
-                                                                     YDataType,
-                                                                     AccElementwiseOperation,
-                                                                     Rank,
-                                                                     NumReduceDim>>;
+using DeviceLayernormPtr = std::unique_ptr<DeviceLayernorm<XDataType,
+                                                           GammaDataType,
+                                                           BetaDataType,
+                                                           AccDataType,
+                                                           YDataType,
+                                                           AccElementwiseOperation,
+                                                           Rank,
+                                                           NumReduceDim>>;
 
 } // namespace device
 } // namespace tensor_operation

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t B1K1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched>
+struct GridwiseBatchedGemmGemm_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    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<...>
+    // Gemm0
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+    // Gemm1
+    static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
+    static constexpr auto B1K1 = Number<B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
+            ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1  = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1  = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b0_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::max(b0_block_space_size_aligned.value, b1_block_space_size_aligned.value);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // 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_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+        const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && Gemm1N == c_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        assert(num_gemm1_k_outer_loop * num_gemm1_k_inner_loop == N / Gemm1KPerBlock);
+
+        if(!block_2_ctile_map.CheckValidity(c_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 / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               const FloatAB* __restrict__ p_b1_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const AccElementwiseOperation& acc_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, 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(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        //
+        // set up Gemm0
+        //
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+        const auto a_block_reset_copy_step =
+            make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b_block_reset_copy_step =
+            make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto AccN3 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 =
+            make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            FloatGemmAcc,
+            FloatAB,
+            decltype(acc_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(acc_element_op),
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{acc_element_op};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b1_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b1_element_op,
+                b1_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr index_t Gemm1KPack = math::max(
+            math::lcm(MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size, B1K1),
+            MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            false,
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            make_tuple(0, 0, 0, 0)}; // TransposeC
+
+        auto c_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            // gemm0
+            gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                                   a_block_desc_ak0_m_ak1,
+                                                                   a_blockwise_copy,
+                                                                   a_grid_buf,
+                                                                   a_block_buf,
+                                                                   a_block_slice_copy_step,
+                                                                   b_grid_desc_bk0_n_bk1,
+                                                                   b_block_desc_bk0_n_bk1,
+                                                                   b_blockwise_copy,
+                                                                   b_grid_buf,
+                                                                   b_block_buf,
+                                                                   b_block_slice_copy_step,
+                                                                   blockwise_gemm,
+                                                                   acc_thread_buf,
+                                                                   num_k_block_main_loop);
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for gemm0 LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, c_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+                    block_sync_lds();
+
+                    gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, c_thread_buf);
+                }
+            } // end gemm1
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_ak0_m_ak1,
+                                                a_block_reset_copy_step); // rewind K
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_bk0_n_bk1,
+                                                b_block_reset_copy_step); // rewind K and step N
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                gemm1_blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_layernorm.hpp → include/ck/tensor_operation/gpu/grid/gridwise_layernorm_naive_variance.hpp

@@ -14,40 +14,6 @@
 
 namespace ck {
 
-template <typename GridwiseReduction,
-          typename XDataType,
-          typename GammaDataType,
-          typename BetaDataType,
-          typename YDataType,
-          typename AccDataType,
-          typename AccElementwiseOperation,
-          typename GridDesc_M_K,
-          typename GridDesc_K>
-__global__ void kernel_layernorm(const GridDesc_M_K x_grid_desc_m_k,
-                                 const GridDesc_K gamma_grid_desc_k,
-                                 const GridDesc_K beta_grid_desc_k,
-                                 const GridDesc_M_K y_grid_desc_m_k,
-                                 index_t num_k_block_tile_iteration,
-                                 AccDataType 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)
-{
-    GridwiseReduction::Run(x_grid_desc_m_k,
-                           gamma_grid_desc_k,
-                           beta_grid_desc_k,
-                           y_grid_desc_m_k,
-                           num_k_block_tile_iteration,
-                           epsilon,
-                           p_x_global,
-                           p_gamma_global,
-                           p_beta_global,
-                           p_y_global,
-                           acc_elementwise_op);
-};
-
 // Y = LayerNorm(X, Beta, Gamma)
 template <typename XDataType,
           typename GammaDataType,
@@ -69,7 +35,7 @@ template <typename XDataType,
           index_t YDstVectorDim,
           index_t YDstVectorSize,
           bool SweepOnce>
-struct GridwiseLayernorm_mk_to_mk
+struct GridwiseLayernormNaiveVariance_mk_to_mk
 {
     static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
                       (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// Y = LayerNorm(X, Beta, Gamma)
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename AccElementwiseOperation,
+          typename GridDesc_M_K,
+          typename GridDesc_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool SweepOnce>
+struct GridwiseLayernormWelfordVariance_mk_to_mk
+{
+    static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((YDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    __device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
+                                        int thread_k_cluster_id)
+    {
+        int kPerBlock = x_grid_desc_m_k.GetTransforms()[I0].GetUpperLengths()[I1];
+        int kPerThread =
+            kPerBlock < K_BlockTileSize ? 0 : KThreadSliceSize * (kPerBlock / K_BlockTileSize);
+        int kPerBlockTail = kPerBlock - kPerThread * KThreadClusterSize;
+
+        if(kPerBlockTail > 0)
+        {
+            int thread_max_len = (thread_k_cluster_id + 1) * KThreadSliceSize;
+            int delta          = thread_max_len - kPerBlockTail;
+            delta              = math::clamp(thread_max_len - kPerBlockTail, 0, KThreadSliceSize);
+            kPerThread += KThreadSliceSize - delta;
+        }
+
+        return kPerThread;
+    }
+
+    __device__ static void Run(const GridDesc_M_K& x_grid_desc_m_k,
+                               const GridDesc_K& gamma_grid_desc_k,
+                               const GridDesc_K& beta_grid_desc_k,
+                               const GridDesc_M_K& y_grid_desc_m_k,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType 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)
+    {
+        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());
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, KThreadSliceSize, true> gamma_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, KThreadSliceSize, true>& beta_thread_buf =
+            gamma_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, 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();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        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>;
+        using ThreadBufferLengths_K           = Sequence<KThreadSliceSize>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_k =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<KThreadSliceSize>{}));
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  GridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_gamma_load =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             AccDataType,
+                                             GridDesc_K,
+                                             decltype(thread_buffer_desc_k),
+                                             ThreadBufferLengths_K,
+                                             Sequence<0>,
+                                             0,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_k, make_multi_index(thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_beta_load = ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                                                     AccDataType,
+                                                                     GridDesc_K,
+                                                                     decltype(thread_buffer_desc_k),
+                                                                     ThreadBufferLengths_K,
+                                                                     Sequence<0>,
+                                                                     0,
+                                                                     BetaSrcVectorSize,
+                                                                     1,
+                                                                     true>(
+            beta_grid_desc_k, make_multi_index(thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               GridDesc_M_K,
+                                               AccElementwiseOperation,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               YDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                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);
+
+        // Copy x from Cache
+        // one pass: fwd, second pass: bwd
+        constexpr auto thread_copy_fwd_step_k = make_multi_index(SweepOnce ? 0 : K_BlockTileSize);
+        constexpr auto thread_copy_bwd_step_k = make_multi_index(SweepOnce ? 0 : -K_BlockTileSize);
+
+        constexpr auto thread_copy_fwd_step_m_k =
+            make_multi_index(0, SweepOnce ? 0 : K_BlockTileSize);
+        constexpr auto thread_copy_bwd_step_m_k =
+            make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x_global, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_global, gamma_grid_desc_k.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_global, beta_grid_desc_k.GetElementSpaceSize());
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        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);
+        });
+
+        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);
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            threadwise_welford.Run(x_thread_buf, 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) * thread_copy_fwd_step_m_k;
+        auto thread_copy_tail_k   = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_k;
+
+        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_k, thread_copy_tail_k);
+        threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_k, thread_copy_tail_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)
+            {
+                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);
+            }
+
+            threadwise_gamma_load.Run(gamma_grid_desc_k,
+                                      gamma_global_val_buf,
+                                      thread_buffer_desc_k,
+                                      make_tuple(I0),
+                                      gamma_thread_buf);
+
+            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));
+
+                    constexpr auto offset_k = thread_buffer_desc_k.CalculateOffset(make_tuple(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_k>{});
+                });
+            });
+
+            threadwise_beta_load.Run(beta_grid_desc_k,
+                                     beta_global_val_buf,
+                                     thread_buffer_desc_k,
+                                     make_tuple(I0),
+                                     beta_thread_buf);
+
+            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));
+
+                    constexpr auto offset_k = thread_buffer_desc_k.CalculateOffset(make_tuple(iK));
+
+                    // beta
+                    y_thread_buf(Number<offset_m_k>{}) =
+                        y_thread_buf(Number<offset_m_k>{}) + beta_thread_buf(Number<offset_k>{});
+                });
+            });
+
+            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);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_k, thread_copy_bwd_step_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_k, thread_copy_bwd_step_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        }
+    }
+};
+
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/math_v2.hpp"
+
+namespace ck {
+
+// Assume
+//  1) XDesc is known at compile-time
+//  2) MeanVarDesc is known at compile-time
+//  3) XBuffer is static buffer
+//  4) MeanBuffer is static buffer
+//  5) VarBuffer is static buffer
+template <typename T, typename XThreadDesc_M_K, typename MeanVarThreadDesc_M>
+struct ThreadwiseWelford
+{
+    static constexpr auto x_thread_desc_m_k      = XThreadDesc_M_K{};
+    static constexpr auto mean_var_thread_desc_m = MeanVarThreadDesc_M{};
+
+    static constexpr auto thread_x_length_m        = x_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto thread_x_length_k        = x_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto thread_mean_var_length_m = mean_var_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(thread_x_length_m == thread_mean_var_length_m,
+                  "lengths of source and mean/var buffer must match!");
+
+    __device__ constexpr ThreadwiseWelford() : cur_count_(0), max_count_(0) {}
+
+    __device__ inline void Update(T& mean, T& var, T x)
+    {
+        using ck::math::isnan;
+
+        if(isnan(x))
+        {
+            mean = x;
+            var  = x;
+        }
+        else
+        {
+            T delta = x - mean;
+            mean += delta / cur_count_;
+            T delta2 = x - mean;
+            var += delta * delta2;
+        }
+    }
+
+    template <typename XBufferType, typename MeanBufferType, typename VarBufferType>
+    __device__ void
+    Run(const XBufferType& x_buf_m_k, MeanBufferType& mean_buf_m, VarBufferType& var_buf_m)
+    {
+        // FIXME - Better naming for var_buf_m
+
+        static_for<0, thread_x_length_k, 1>{}([&](auto iK) {
+            if(cur_count_ < max_count_)
+            {
+                ++cur_count_;
+
+                static_for<0, thread_x_length_m, 1>{}([&](auto iM) {
+                    constexpr index_t out_offset =
+                        mean_var_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+                    constexpr auto in_offset =
+                        x_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                    Update(mean_buf_m(Number<out_offset>{}),
+                           var_buf_m(Number<out_offset>{}),
+                           x_buf_m_k[Number<in_offset>{}]);
+                });
+            }
+        });
+    };
+
+    int cur_count_;
+    int max_count_;
+};
+
+} // namespace ck

include/ck/utility/math.hpp

@@ -144,6 +144,12 @@ __host__ __device__ constexpr auto min(X x, Ys... ys)
     return min(x, min(ys...));
 }
 
+template <typename T>
+__host__ __device__ constexpr T clamp(const T& x, const T& lowerbound, const T& upperbound)
+{
+    return min(max(x, lowerbound), upperbound);
+}
+
 // disallow implicit type casting
 template <typename T>
 __device__ T exp(T x);

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmGemm<ALayout,
+                                                        B0Layout,
+                                                        B1Layout,
+                                                        CLayout,
+                                                        ADataType,
+                                                        B0DataType,
+                                                        B1DataType,
+                                                        CDataType,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>
+{
+    using DeviceOp = DeviceBatchedGemmGemm<ALayout,
+                                           B0Layout,
+                                           B1Layout,
+                                           CLayout,
+                                           ADataType,
+                                           B0DataType,
+                                           B1DataType,
+                                           CDataType,
+                                           PassThrough,
+                                           PassThrough,
+                                           PassThrough,
+                                           PassThrough,
+                                           PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
+                         is_same_v<B1Layout, Row> && is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/CMakeLists.txt

@@ -13,6 +13,7 @@ add_subdirectory(gemm_reduce)
 add_subdirectory(gemm_bias_add_reduce)
 add_subdirectory(batched_gemm)
 add_subdirectory(batched_gemm_reduce)
+add_subdirectory(batched_gemm_gemm)
 add_subdirectory(batched_gemm_softmax_gemm)
 add_subdirectory(grouped_gemm)
 add_subdirectory(contraction_scale)

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

+set(DEVICE_BATCHED_GEMM_GEMM_INSTANCE_SOURCE
+    device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+)
+
+add_instance_library(device_batched_gemm_gemm_instance OBJECT ${DEVICE_BATCHED_GEMM_GEMM_INSTANCE_SOURCE})
+target_compile_features(device_batched_gemm_gemm_instance PUBLIC)
+set_target_properties(device_batched_gemm_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
+clang_tidy_check(device_batched_gemm_gemm_instance)
\ No newline at end of file

library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_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/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm_xdl_cshuffle.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 Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances = std::tuple<
+    // clang-format off
+        //################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################|        |         |         |        |  Type|   Type|   Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f16_instance.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/device_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp"
 #include "ck/utility/data_type.hpp"
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
@@ -21,28 +21,28 @@ template <index_t Rank, index_t Reduce>
 using device_layernorm_f16_instances = std::tuple<
     // clang-format off
         // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>, // fallback kernel
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8, 8, 8>,
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8, 8, 8>,
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8, 8, 8>,
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8, 8, 8>,
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8, 8, 8>,
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8, 8, 8>,
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8, 8, 8>,
-        DeviceLayernorm<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8, 8, 8>
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>, // fallback kernel
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8, 8, 8>,
+        DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8, 8, 8>
     // clang-format on
     >;
 
 void add_device_layernorm_f16_rank2_instances(
-    std::vector<DeviceNormalization2Ptr<F16, F16, F16, F32, F16, Pass, 2, 1>>& instances)
+    std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Pass, 2, 1>>& instances)
 {
     add_device_operation_instances(instances, device_layernorm_f16_instances<2, 1>{});
 }
 
 void add_device_layernorm_f16_rank4_instances(
-    std::vector<DeviceNormalization2Ptr<F16, F16, F16, F32, F16, Pass, 4, 3>>& instances)
+    std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Pass, 4, 3>>& instances)
 {
     add_device_operation_instances(instances, device_layernorm_f16_instances<4, 3>{});
 }

library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f32_instance.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/device_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp"
 #include "ck/utility/data_type.hpp"
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
@@ -20,27 +20,27 @@ template <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>
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>,
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4, 4, 4>,
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4, 4, 4>,
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4, 4, 4>,
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4, 4, 4>,
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4, 4, 4>,
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4, 4, 4>,
-        DeviceLayernorm<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4, 4, 4>
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4, 4, 4>,
+        DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4, 4, 4>
     // clang-format on
     >;
 
 void add_device_layernorm_f32_rank2_instances(
-    std::vector<DeviceNormalization2Ptr<F32, F32, F32, F32, F32, Pass, 2, 1>>& instances)
+    std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Pass, 2, 1>>& instances)
 {
     add_device_operation_instances(instances, device_layernorm_f32_instances<2, 1>{});
 }
 
 void add_device_layernorm_f32_rank4_instances(
-    std::vector<DeviceNormalization2Ptr<F32, F32, F32, F32, F32, Pass, 4, 3>>& instances)
+    std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Pass, 4, 3>>& instances)
 {
     add_device_operation_instances(instances, device_layernorm_f32_instances<4, 3>{});
 }

profiler/include/profile_batched_gemm_gemm_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout>
+bool profile_batched_gemm_gemm_impl(bool do_verification,
+                                    int init_method,
+                                    bool do_log,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int O,
+                                    int BatchCount    = 1,
+                                    int StrideA       = -1,
+                                    int StrideB0      = -1,
+                                    int StrideB1      = -1,
+                                    int StrideC       = -1,
+                                    int BatchStrideA  = -1,
+                                    int BatchStrideB0 = -1,
+                                    int BatchStrideB1 = -1,
+                                    int BatchStrideC  = -1)
+
+{
+
+    using Row           = tensor_layout::gemm::RowMajor;
+    using Col           = tensor_layout::gemm::ColumnMajor;
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using Acc0ElementOp = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+
+    // Ref Gemm0
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+    // Ref Gemm
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    // Host verification: Output of Gemm0 is input A of Gemm1
+    Tensor<ADataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 3});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 3});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 3});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) * c_g_m_o_device_result.mDesc.GetElementSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp = tensor_operation::device::DeviceBatchedGemmGemm<ALayout,
+                                                                     B0Layout,
+                                                                     B1Layout,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     B0DataType,
+                                                                     B1DataType,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     B0ElementOp,
+                                                                     Acc0ElementOp,
+                                                                     B1ElementOp,
+                                                                     CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, PassThrough{});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            acc0_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+            M,
+            N,
+            K,
+            O,
+            BatchCount,
+            StrideA,
+            StrideB0,
+            StrideB1,
+            StrideC,
+            BatchStrideA,
+            BatchStrideB0,
+            BatchStrideB1,
+            BatchStrideC,
+            a_element_op,
+            b0_element_op,
+            acc0_element_op,
+            b1_element_op,
+            c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+                pass = pass &
+                       ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_g_m_k: ", a_g_m_k.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_g_k_n : ", b0_g_k_n.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_g_n_o : ", b1_g_n_o.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_host_result : ", c_g_m_o_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_device_result : ", c_g_m_o_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profile_layernorm_impl.hpp

@@ -7,7 +7,7 @@
 
 #include "ck/ck.hpp"
 #include "profiler/include/data_type_enum.hpp"
-#include "ck/tensor_operation/gpu/device/device_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
@@ -25,10 +25,10 @@ using F32         = float;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
 void add_device_layernorm_f16_rank2_instances(
-    std::vector<DeviceNormalization2Ptr<F16, F16, F16, F32, F16, PassThrough, 2, 1>>&);
+    std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, PassThrough, 2, 1>>&);
 
 void add_device_layernorm_f32_rank2_instances(
-    std::vector<DeviceNormalization2Ptr<F32, F32, F32, F32, F32, PassThrough, 2, 1>>&);
+    std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, PassThrough, 2, 1>>&);
 
 } // namespace instance
 } // namespace device
@@ -105,14 +105,14 @@ void profile_layernorm_impl(int do_verification,
 
     // add device normalization instances
     constexpr int NumReduceDim = Rank - 1;
-    std::vector<tensor_operation::device::DeviceNormalization2Ptr<XDataType,
-                                                                  GammaDataType,
-                                                                  BetaDataType,
-                                                                  AccDataType,
-                                                                  YDataType,
-                                                                  PassThrough,
-                                                                  Rank,
-                                                                  NumReduceDim>>
+    std::vector<tensor_operation::device::DeviceLayernormPtr<XDataType,
+                                                             GammaDataType,
+                                                             BetaDataType,
+                                                             AccDataType,
+                                                             YDataType,
+                                                             PassThrough,
+                                                             Rank,
+                                                             NumReduceDim>>
         instances;
 
     if constexpr(is_same<XDataType, F16>::value && is_same<GammaDataType, F16>::value &&
@@ -163,6 +163,7 @@ void profile_layernorm_impl(int do_verification,
                                                           strideXY,
                                                           strideGamma,
                                                           strideBeta,
+                                                          strideXY,
                                                           reduce_dim,
                                                           1e-4,
                                                           x_dev.GetDeviceBuffer(),

test/CMakeLists.txt

@@ -40,6 +40,7 @@ add_subdirectory(gemm_split_k)
 add_subdirectory(gemm_reduce)
 add_subdirectory(batched_gemm)
 add_subdirectory(batched_gemm_reduce)
+add_subdirectory(batched_gemm_gemm)
 add_subdirectory(batched_gemm_softmax_gemm)
 add_subdirectory(grouped_gemm)
 add_subdirectory(reduce)

test/batched_gemm_gemm/CMakeLists.txt

+add_custom_target(test_batched_gemm_gemm)
+
+add_gtest_executable(test_batched_gemm_gemm_fp16 test_batched_gemm_gemm_fp16.cpp)
+target_link_libraries(test_batched_gemm_gemm_fp16 PRIVATE utility device_batched_gemm_gemm_instance)
+add_dependencies(test_batched_gemm_gemm test_batched_gemm_gemm_fp16)
\ No newline at end of file

test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_gemm_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmGemmFP16 : public TestBatchedGemmGemm<Tuple>
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<F16, F16, F16, F16, Row, Col, Row, Row>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmGemmFP16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmGemmFP16, DISABLED_Bench_FP16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 768},
+        {256, 256, 128, 128, 768},
+        {512, 512, 64, 64, 768},
+        {512, 512, 128, 128, 768},
+        {1024, 1024, 64, 64, 768},
+        {1024, 1024, 128, 128, 768},
+        {2048, 2048, 64, 64, 768},
+        {2048, 2048, 128, 128, 768},
+        {4096, 4096, 64, 64, 768},
+        {4096, 4096, 128, 128, 768},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}

test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include <vector>
+#include "profiler/include/profile_batched_gemm_gemm_impl.hpp"
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename Tuple>
+struct TestBatchedGemmGemm : public ::testing::Test
+{
+    using ADataType  = std::tuple_element_t<0, Tuple>;
+    using B0DataType = std::tuple_element_t<1, Tuple>;
+    using B1DataType = std::tuple_element_t<2, Tuple>;
+    using CDataType  = std::tuple_element_t<3, Tuple>;
+    using ALayout    = std::tuple_element_t<4, Tuple>;
+    using B0Layout   = std::tuple_element_t<5, Tuple>;
+    using B1Layout   = std::tuple_element_t<6, Tuple>;
+    using CLayout    = std::tuple_element_t<7, Tuple>;
+
+    std::vector<std::vector<int>> lengths_ = {
+        {256, 256, 64, 64, 4},
+        {256, 256, 128, 128, 4},
+        {512, 512, 64, 64, 2},
+        {512, 512, 128, 128, 2},
+        {1024, 1024, 64, 64, 1},
+        {1024, 1024, 128, 128, 1},
+    };
+    bool bench_  = false;
+    bool verify_ = true;
+
+    void RunSingle(int M, int N, int K, int O, int BatchCount)
+    {
+        bool pass = ck::profiler::profile_batched_gemm_gemm_impl<ADataType,
+                                                                 B0DataType,
+                                                                 B1DataType,
+                                                                 CDataType,
+                                                                 ALayout,
+                                                                 B0Layout,
+                                                                 B1Layout,
+                                                                 CLayout>(
+            verify_, 1, false, bench_, M, N, K, O, BatchCount);
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run()
+    {
+        for(auto lengths : this->lengths_)
+        {
+            int M          = lengths[0];
+            int N          = lengths[1];
+            int K          = lengths[2];
+            int O          = lengths[3];
+            int BatchCount = lengths[4];
+
+            this->RunSingle(M, N, K, O, BatchCount);
+        }
+    }
+};

test/layernorm/test_layernorm_util.hpp

@@ -9,7 +9,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/utility/number.hpp"
-#include "ck/tensor_operation/gpu/device/device_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/host_tensor.hpp"
@@ -63,24 +63,24 @@ class TestLayernorm : public ::testing::Test
                                                                          Rank,
                                                                          NumReduceDim>;
 
-    using DeviceInstance = tensor_operation::device::DeviceLayernorm<XDataType,
-                                                                     GammaDataType,
-                                                                     BetaDataType,
-                                                                     AccDataType,
-                                                                     YDataType,
-                                                                     PassThrough,
-                                                                     Rank,
-                                                                     NumReduceDim,
-                                                                     BlockSize,
-                                                                     MThreadClusterSize,
-                                                                     KThreadClusterSize,
-                                                                     MThreadSliceSize,
-                                                                     KThreadSliceSize,
-                                                                     XYSrcVectorDim,
-                                                                     XSrcVectorSize,
-                                                                     GammaSrcVectorSize,
-                                                                     BetaSrcVectorSize,
-                                                                     YDstVectorSize>;
+    using DeviceInstance = tensor_operation::device::DeviceLayernormImpl<XDataType,
+                                                                         GammaDataType,
+                                                                         BetaDataType,
+                                                                         AccDataType,
+                                                                         YDataType,
+                                                                         PassThrough,
+                                                                         Rank,
+                                                                         NumReduceDim,
+                                                                         BlockSize,
+                                                                         MThreadClusterSize,
+                                                                         KThreadClusterSize,
+                                                                         MThreadSliceSize,
+                                                                         KThreadSliceSize,
+                                                                         XYSrcVectorDim,
+                                                                         XSrcVectorSize,
+                                                                         GammaSrcVectorSize,
+                                                                         BetaSrcVectorSize,
+                                                                         YDstVectorSize>;
 
     TestLayernorm() : ref_instance_invoker_(ReferenceInstance{}.MakeInvoker()) {}
 
@@ -119,6 +119,7 @@ class TestLayernorm : public ::testing::Test
                                      gamma.mDesc.GetStrides().end()},
             std::vector<ck::index_t>{beta.mDesc.GetStrides().begin(),
                                      beta.mDesc.GetStrides().end()},
+            std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
             reduceDims,
             1e-4,
             x_dev.GetDeviceBuffer(),