Merge branch 'develop' into aosewski/ggemm

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

@@ -533,6 +533,11 @@ struct DeviceElementwiseNormalizationImpl
                 return (false);
         }
 
+        if(p_arg_->x_lds_size_ >= 65536)
+        {
+            return (false);
+        }
+
         return true;
     };
 

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

@@ -669,6 +669,9 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
             {
                 throw std::runtime_error("wrong! GridwiseGemmWelford has invalid setting");
             }
+            if(arg.p_workspace_e_grid_ == nullptr || arg.p_workspace_mean_ == nullptr ||
+               arg.p_workspace_var_ == nullptr || arg.p_workspace_count_ == nullptr)
+                throw std::runtime_error("wrong! WorkSpace pointer has not been set");
 
             index_t grid_size = arg.block_2_etile_map_.CalculateGridSize(arg.gemm_e_grid_desc_m_n_);
 
@@ -939,7 +942,11 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
             }
         }
 
-        return true;
+        return GridwiseGemmWelford::CheckValidity(arg.a_grid_desc_m_k_,
+                                                  arg.b_grid_desc_n_k_,
+                                                  arg.ds_grid_desc_m_n_,
+                                                  arg.gemm_e_grid_desc_m_n_,
+                                                  arg.block_2_etile_map_);
     }
 
     // polymorphic
@@ -1055,7 +1062,12 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
             << GemmKPerBlock << ", "
             << AK1 << ", "
             << BK1 << ", "
-            << getGemmSpecializationString(GemmSpec)
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << PostShuffleThreadClusterSize_M_N::At(I0) << ", "
+            << PostShuffleThreadClusterSize_M_N::At(I1) << ", "
+            << LayernormThreadClusterSize_M_N::At(I0) << ", "
+            << LayernormThreadClusterSize_M_N::At(I1) << ", "
+            << LayernormThreadSliceSize_M
             << ">"
             << " LoopScheduler: "
             << LoopSchedToString[LoopSched] << ", "

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

@@ -49,6 +49,14 @@ struct Add
         y = x0 + x1;
     };
 
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x1);
+        y                  = x0 + x1_tmp;
+    }
+
     template <>
     __host__ __device__ constexpr void
     operator()<bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
@@ -67,6 +75,30 @@ struct Add
     };
 };
 
+struct ScaleAdd
+{
+    __host__ __device__ ScaleAdd(float scale) : scale_(scale) {}
+
+    template <typename Y, typename X0, typename X1>
+    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
+
+    template <>
+    __host__ __device__ void
+    operator()<float, float, half_t>(float& y, const float& x0, const half_t& x1) const
+    {
+        y = scale_ * x0 + ck::type_convert<float>(x1);
+    };
+
+    template <>
+    __host__ __device__ void
+    operator()<float, float, bhalf_t>(float& y, const float& x0, const bhalf_t& x1) const
+    {
+        y = scale_ * x0 + ck::type_convert<float>(x1);
+    };
+
+    float scale_;
+};
+
 struct Subtract
 {
     template <typename T>

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_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_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_softmax.hpp"
+
+namespace ck {
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename D0sDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename C0DEElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename C1DEElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename C1GridDesc_M_N,
+          typename D0sGridDesc_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,
+          bool PadN,
+          bool MaskOutUpperTriangle,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    // 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>{};
+
+    static constexpr auto Gemm0MWaves = MPerBlock / (MPerXdl * MXdlPerWave);
+    static constexpr auto Gemm0NWaves = NPerBlock / (NPerXdl * NXdlPerWave);
+
+    // Gemm1
+    static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
+    static constexpr auto B1K1 = Number<B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, 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()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
+                                         SharedMemTrait::b_block_space_size_aligned) *
+                                        sizeof(FloatAB);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(FloatAB);
+        const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
+                                           SharedMemTrait::reduction_space_size_aligned) *
+                                          sizeof(FloatGemmAcc);
+        const index_t c_block_bytes_end =
+            SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, softmax_bytes_end, c_block_bytes_end);
+    }
+
+    // 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 C1GridDesc_M_N& c1_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 == c1_grid_desc_m_n.GetLength(I0) && Gemm1N == c1_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;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c1_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
+    MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const C1GridDesc_M_N& c1_grid_desc_m_n)
+    {
+        const auto M = c1_grid_desc_m_n.GetLength(I0);
+        const auto N = c1_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(
+            c1_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 C1GridDesc_M_N& c1_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, C1GridDesc_M_N>(
+            c1_grid_desc_m_n);
+    }
+
+    __device__ static auto GetGemm0WaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+        constexpr auto WaveSize = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.wave_size;
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
+    {
+        constexpr auto WaveSize = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.wave_size;
+        constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(WaveSize / MPerXdl, MPerXdl))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    static constexpr auto MakeD0sGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+
+                return static_cast<const D0DataType*>(nullptr);
+            },
+            Number<NumD0Tensor>{});
+    }
+    // D0 desc for source in blockwise copy
+    template <typename D0GridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0GridDesc_M_N& d0_grid_desc_m_n)
+    {
+        const auto M = d0_grid_desc_m_n.GetLength(I0);
+        const auto N = d0_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto mfma = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma;
+        constexpr auto N3   = mfma.num_groups_per_blk;
+        constexpr auto N4   = mfma.num_input_blks;
+        constexpr auto N5   = mfma.group_size;
+        return transform_tensor_descriptor(
+            d0_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
+                       make_unmerge_transform(
+                           make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
+    }
+
+    // D0s desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0sGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ds_grid_desc_m_n[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    using D0sGridPointer                                  = decltype(MakeD0sGridPointer());
+    using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
+        MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(D0sGridDesc_M_N{}))>;
+
+    using C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(C1GridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(C1GridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A and B: be careful of alignment
+        static constexpr auto a_block_desc_ak0_m_ak1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b_block_desc_bk0_n_bk1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a_block_space_offset  = 0;
+        static constexpr auto b_block_space_offset  = a_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for reduction
+        static constexpr index_t reduction_space_size_aligned =
+            math::integer_least_multiple(BlockSize, max_lds_align);
+
+        static constexpr auto reduction_space_offset = 0;
+
+        // LDS allocation for C shuffle in LDS
+        static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c_block_space_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap, typename C0MatrixMask>
+    __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,
+                               D0sGridPointer p_d0s_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const C0DEElementwiseOperation& c0de_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const C1DEElementwiseOperation& c1de_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 C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
+                                   d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                               const Block2CTileMap& block_2_ctile_map,
+                               const C0MatrixMask& c0_matrix_mask)
+    {
+        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());
+        const auto d0s_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_d0s_grid[i],
+                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i].GetElementSpaceSize());
+            },
+            Number<NumD0Tensor>{});
+
+        // 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/gemm1_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 gemm1_n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // 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
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
+            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_Selector<PipelineVer,
+                                                                          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);
+
+        // d0 matrix threadwise copy
+        constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,   // MBlockId
+                                                           I1,   // NBlockID
+                                                           I1,   // MRepeat
+                                                           I1,   // NRepeat
+                                                           I1,   // MWaveId
+                                                           I1,   // NWaveId
+                                                           I1,   // MPerXdl
+                                                           I1,   // NGroupNum
+                                                           I1,   // NInputNum
+                                                           n4)); // registerNum
+
+        auto d0s_thread_buf = generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+                return StaticBuffer<
+                    AddressSpaceEnum::Vgpr,
+                    D0DataType,
+                    d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
+                    true>{};
+            },
+            Number<NumD0Tensor>{});
+
+        const auto wave_id     = GetGemm0WaveIdx();
+        const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
+
+        constexpr auto acc0_thread_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MXdlPerWave>{}, Number<NXdlPerWave>{}, n2, n4));
+
+        auto d0s_threadwise_copy = generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+                return ThreadwiseTensorSliceTransfer_v2<
+                    D0DataType,
+                    D0DataType,
+                    decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]),
+                    decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+                    Sequence<I1, I1, I1, I1, I1, I1, I1, I1, I1, n4>,
+                    Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
+                    9,
+                    n4,
+                    1,
+                    false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                           make_multi_index(block_work_idx[I0], // MBlockId
+                                            0,                  // NBlockId
+                                            0,                  // mrepeat
+                                            0,                  // nrepeat
+                                            wave_id[I0],        // MWaveId
+                                            wave_id[I1],        // NWaveId
+                                            wave_m_n_id[I1],    // MPerXdl
+                                            0,                  // group
+                                            wave_m_n_id[I0],    // NInputIndex
+                                            0));                // register number
+            },
+            Number<NumD0Tensor>{});
+        // 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),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{tensor_operation::element_wise::PassThrough{}};
+
+        // 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, gemm1_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) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        // selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
+        // selected_mfma.k_per_blk <= Gemm1KPack
+        //
+        // Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
+        // multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
+        // Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
+        // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
+        // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
+        // therefore we may just as well assign Gemm1KPack = group_size
+        constexpr index_t Gemm1KPack =
+            MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
+
+        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,
+            true,       // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            // BMmaKStride
+            make_tuple(0, 0, 0, 0)}; // A_origin
+
+        auto acc1_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
+
+        //
+        // Blockwise softmax
+        //
+        auto workspace_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatGemmAcc*>(p_shared) + SharedMemTrait::reduction_space_offset,
+            SharedMemTrait::reduction_space_size_aligned);
+
+        // get acc0 8D thread cluster
+        constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+        constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
+        constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
+        constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
+        constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
+        constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
+        constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
+        constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
+        constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
+
+        // get acc0 thread map
+        constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
+                       make_pass_through_transform(I1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(
+                make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+        const auto threadid_to_m_n_thread_cluster_adaptor =
+            chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
+
+        // get acc0 2D thread cluster & 2D thread slice
+        constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
+            make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
+        constexpr auto thread_slice_desc_m_n =
+            make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
+
+        auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
+                                                  FloatGemmAcc,
+                                                  decltype(threadid_to_m_n_thread_cluster_adaptor),
+                                                  decltype(thread_cluster_desc_m_n),
+                                                  decltype(thread_slice_desc_m_n)>{};
+
+        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
+        StaticBuffer<AddressSpaceEnum::Vgpr, FloatGemmAcc, acc1_thread_buf.Size(), true>
+            c_thread_buf;
+        c_thread_buf.Clear();
+
+        // Initialize running sum and max of exponentiating row vectors
+        using SoftmaxBuf = typename decltype(blockwise_softmax)::BufferType;
+        SoftmaxBuf running_sum, running_sum_new, running_max, running_max_new;
+        running_sum     = 0;
+        running_sum_new = 0;
+        running_max     = NumericLimits<FloatGemmAcc>::Lowest();
+        running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            auto n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
+            if(c0_matrix_mask.IsTileSkippable(
+                   m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
+            {
+                continue;
+            }
+            // 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);
+            // multiple d
+            if constexpr(NumD0Tensor)
+            {
+                static_for<0, MXdlPerWave, 1>{}([&](auto mr) {
+                    static_for<0, NXdlPerWave, 1>{}([&](auto nr) {
+                        static_for<0, n2, 1>{}([&](auto groupid) {
+                            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                                d0s_threadwise_copy(i).Run(
+                                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                    d0s_grid_buf[i],
+                                    d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                    make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                    d0s_thread_buf(i));
+                            });
+
+                            static_for<0, n4, 1>{}([&](auto i) {
+                                constexpr index_t c_offset = acc0_thread_desc.CalculateOffset(
+                                    make_tuple(mr, nr, groupid, i));
+
+                                // get reference to src data
+                                const auto src_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto iSrc) -> const auto& {
+                                        return d0s_thread_buf[iSrc][i];
+                                    },
+                                    Number<NumD0Tensor>{});
+
+                                // get reference to dst data
+                                auto dst_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto) -> auto& {
+                                        return acc_thread_buf(Number<c_offset>{});
+                                    },
+                                    Number<2>{});
+
+                                unpack2(c0de_element_op, dst_data_refs, src_data_refs);
+                            });
+                            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                                d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                    make_multi_index(0, 0, 0, 0, 0, 0, 0, 1, 0, 0));
+                            });
+                        });
+                        static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                            d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                                d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                make_multi_index(0, 0, 0, 1, 0, 0, 0, -n2.value, 0, 0));
+                        });
+                    });
+                    static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                        d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                            d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                            make_multi_index(0, 0, 1, -NXdlPerWave, 0, 0, 0, 0, 0, 0));
+                    });
+                });
+                static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                    d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                        d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                        make_multi_index(0, 1, -MXdlPerWave, 0, 0, 0, 0, 0, 0, 0));
+                });
+            }
+            else
+            {
+                static_for<0, acc_thread_buf.Size(), 1>{}(
+                    [&](auto i) { c0de_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
+            }
+
+            // do MNK padding or upper triangular masking
+            if constexpr(MaskOutUpperTriangle || PadN)
+            {
+                // 8d thread_desc in thread scope
+                constexpr auto c_thread_lengths =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                // 8d block_desc in block scope
+                constexpr auto c_block_lengths =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                constexpr auto M0 = c_block_lengths[I0];
+                constexpr auto N0 = c_block_lengths[I1];
+                constexpr auto M1 = c_block_lengths[I2];
+                constexpr auto N1 = c_block_lengths[I3];
+                constexpr auto M2 = c_block_lengths[I4];
+                constexpr auto N2 = c_block_lengths[I5];
+                constexpr auto N3 = c_block_lengths[I6];
+                constexpr auto N4 = c_block_lengths[I7];
+
+                // works like multi-dimension static_for (static_ford), but provides both the linear
+                // index as well as n-d index
+                using Acc0TileIterator = SpaceFillingCurve<
+                    decltype(c_thread_lengths),
+                    typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
+                    typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
+                    false>; // SnakeCurved
+
+                auto acc0_thread_origin = blockwise_gemm.CalculateCThreadOriginDataIndex8D(
+                    Number<0>{}, Number<0>{}, Number<0>{}, Number<0>{});
+
+                constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+                    make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2)),
+                               make_unmerge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+                static_for<0, Acc0TileIterator::GetNumOfAccess(), 1>{}([&](auto i) {
+                    auto acc0_thread_idx = Acc0TileIterator::GetIndex(i) + acc0_thread_origin;
+                    auto m_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
+                    auto n_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
+                    auto m_global = m_local + m_block_data_idx_on_grid;
+                    auto n_global = n_local + n_block_data_idx_on_grid;
+                    if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
+                    {
+                        acc_thread_buf(i) = -ck::NumericLimits<float>::Infinity();
+                    }
+                });
+            }
+
+            block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
+
+            // softmax
+            SoftmaxBuf& max = blockwise_softmax.max_value_buf;
+            SoftmaxBuf& sum = blockwise_softmax.sum_value_buf;
+
+            blockwise_softmax.Run(acc_thread_buf, workspace_buf);
+
+            // TODO: may convert to log domain
+            running_max_new = mathext::max(max, running_max);
+            running_sum_new = mathext::exp(running_max - running_max_new) * running_sum +
+                              mathext::exp(max - running_max_new) * sum;
+
+            // 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.
+
+                // Initialize acc1
+                acc1_thread_buf.Clear();
+
+                // 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 reduction 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, acc1_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, acc1_thread_buf);
+                }
+            } // end gemm1
+
+            // workaround compiler issue; see ck/ck.hpp
+            if constexpr(CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE == 1 &&
+                         is_same_v<FloatAB, bhalf_t> && MPerBlock == 256 && NPerBlock == 128 &&
+                         Gemm1NPerBlock == 128)
+            {
+                __builtin_amdgcn_sched_barrier(0);
+            }
+
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+            constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+            constexpr auto cn0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+            constexpr auto cm1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+            constexpr auto cn1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+            constexpr auto cm2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+            constexpr auto cn2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+            constexpr auto cn3 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+            constexpr auto cn4 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+            constexpr auto c_thread_slice_desc_m_n = make_naive_tensor_descriptor_packed(
+                make_tuple(cm0 * cm1 * cm2, cn0 * cn1 * cn2 * cn3 * cn4));
+            constexpr auto c_thread_buf_slice_m = c_thread_slice_desc_m_n.GetLength(I0);
+            constexpr auto c_thread_buf_slice_n = c_thread_slice_desc_m_n.GetLength(I1);
+
+            static_for<0, c_thread_buf_slice_m, 1>{}([&](auto iM) {
+                static_for<0, c_thread_buf_slice_n, 1>{}([&](auto iN) {
+                    auto I = Number<c_thread_slice_desc_m_n.CalculateOffset(make_tuple(iM, iN))>{};
+                    FloatGemmAcc acc1 = acc1_thread_buf[I]; // P*V
+                    FloatGemmAcc c    = c_thread_buf[I];    // O
+                    FloatGemmAcc c_new =
+                        (running_sum[iM] * math::exp(running_max[iM] - running_max_new[iM]) * c +
+                         math::exp(max[iM] - running_max_new[iM]) * acc1) /
+                        running_sum_new[iM]; // Formula by Dao et al.,
+                                             // https://arxiv.org/pdf/2205.14135v2.pdf section 3.1
+
+                    c_thread_buf(I) = c_new; // O_new
+                });
+            });
+
+            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
+
+            // update before next j iteration
+            running_max = running_max_new;
+            running_sum = running_sum_new;
+
+            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_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
+                gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
+            constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
+            constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_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_n2_n3_n4 = 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 = MPerXdl
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2,                                      // N2 * N3 * N4 = NPerXdl
+                        N3,
+                        N4))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 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_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_n3_n4_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_n2_n3_n4),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            I1,
+                                                            N2,
+                                                            I1,
+                                                            N4>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                    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],
+                                     n_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I3],
+                                     n_thread_data_on_block_idx[I4]),
+                    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
+                C1DEElementwiseOperation,   // 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),
+                 c1de_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, 1, N2, 1, N4>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           1,
+                                           N2,
+                                           1,
+                                           N4>>{};
+
+            // 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_n2_n3_n4,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              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/utility/math_v2.hpp

@@ -158,9 +158,9 @@ static inline __device__ bool isnan(half_t x)
     return (xx & 0x7FFF) > 0x7C00;
 };
 
-static inline __device__ float sqrt(float x) { return ::sqrtf(x); };
+static inline __device__ float sqrt(float x) { return __builtin_amdgcn_sqrtf(x); };
 
-static inline __device__ double sqrt(double x) { return ::sqrt(x); };
+static inline __device__ double sqrt(double x) { return __builtin_amdgcn_sqrt(x); };
 
 } // namespace math
 } // namespace ck

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -89,8 +89,10 @@ using Scale          = ck::tensor_operation::element_wise::Scale;
 using Bilinear       = ck::tensor_operation::element_wise::Bilinear;
 using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
 using AddFastGelu    = ck::tensor_operation::element_wise::AddFastGelu;
+using AddReluAdd     = ck::tensor_operation::element_wise::AddReluAdd;
 using FastGelu       = ck::tensor_operation::element_wise::FastGelu;
 using AddMultiply    = ck::tensor_operation::element_wise::AddMultiply;
+using ScaleAdd       = ck::tensor_operation::element_wise::ScaleAdd;
 
 template <typename Activation>
 using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;

library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.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_softmax_gemm_permute.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_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<F16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<F16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<BF16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<BF16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          MaskingSpecialization MaskingSpec>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      Acc0BiasDataType,
+                                                                      ck::Tuple<>,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      ScaleAdd,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      MaskingSpec>>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         ADataType,
+                                                         B0DataType,
+                                                         B1DataType,
+                                                         CDataType,
+                                                         Acc0BiasDataType,
+                                                         ck::Tuple<>,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         ScaleAdd,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         MaskingSpec>;
+
+    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> &&
+                     Acc0BiasDataType::Size() == 1 &&
+                     is_same_v<tuple_element_t<0, Acc0BiasDataType>, half_t>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
+                          is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16> &&
+                          Acc0BiasDataType::Size() == 1 &&
+                          is_same_v<tuple_element_t<0, Acc0BiasDataType>, BF16>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -10,7 +10,6 @@
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
-#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
 

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+// GEMM + Add + Relu + Add + Layernorm
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename HLayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    HLayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    GammaDataType,
+    BetaDataType,
+    HDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::AddReluAdd,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemmMultipleDLayernorm<ALayout,
+                                                  BLayout,
+                                                  ck::Tuple<D0Layout, D1Layout>,
+                                                  HLayout,
+                                                  ADataType,
+                                                  BDataType,
+                                                  ck::Tuple<D0DataType, D1DataType>,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  HDataType,
+                                                  ck::tensor_operation::element_wise::PassThrough,
+                                                  ck::tensor_operation::element_wise::PassThrough,
+                                                  ck::tensor_operation::element_wise::AddReluAdd,
+                                                  ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<GammaDataType, half_t> && is_same_v<BetaDataType, half_t> &&
+                     is_same_v<HDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

 add_instance_library(device_batched_gemm_softmax_gemm_permute_instance
     device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
     device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
+    device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+    device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
 )
 

library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_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/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using 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;
+using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          MaskingSpecialization MaskingSpec>
+using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        // #############################################|  NumDimG| NumDimM| NumDimN| NumDimK| NumDimO|  AData|  B0Data|  B1Data|  CData|     Acc0BiasData| Acc1BiasData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM|   ATensorSpec|  B0TensorSpec|  B1TensorSpec|   CTensorSpec| 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| MaskingSpec|
+        // #############################################|         |        |        |        |        |   Type|    Type|    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|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        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, MaskingSpec>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<BF16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskOutUpperTriangle>{});
+}
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<BF16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskDisabled>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_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/impl/device_batched_gemm_softmax_gemm_permute_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;
+using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          MaskingSpecialization MaskingSpec>
+using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        // #############################################|  NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| AData| B0Data| B1Data| CData|    Acc0BiasData| Acc1BiasData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM|   ATensorSpec|  B0TensorSpec|  B1TensorSpec|   CTensorSpec| 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| MaskingSpec|
+        // #############################################|         |        |        |        |        |  Type|   Type|   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|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        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, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        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, MaskingSpec>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<F16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskOutUpperTriangle>{});
+}
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<F16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskDisabled>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp

@@ -23,6 +23,11 @@ template <typename XElementwise, typename YElementwise, index_t Rank, index_t Re
 using device_elementwise_normalization_f16_instances =
     std::tuple <
         // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel for large N
         DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
         DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
         DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel

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

+add_instance_library(device_gemm_add_relu_add_layernorm_instance
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
+)

library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// h = layernorm(e, gamma, beta)
+// outout: h[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
+template <LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|            CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|     LoopScheduler|     Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise|    Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|      Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |               |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                  |             |
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|            CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|           PostShuffle|     PostShuffle|            Layernorm|       Layernorm|          LoopScheduler|              Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise|    Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|  ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|      Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                  _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |               |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                      |                |                     |                |                       |                      |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Interwave, PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances<
+            LoopScheduler::Interwave,
+            PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v2>{});
+#endif
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// h = layernorm(e, gamma, beta)
+// outout: h[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
+template <LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|     LoopScheduler|     Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                  |             |
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|          LoopScheduler|               Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                       |                       |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                       |                       |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                       |                       |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,              S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,   PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,              S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Interwave, PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,              S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,   PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances<
+            LoopScheduler::Interwave,
+            PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v2>{});
+#endif
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// h = layernorm(e, gamma, beta)
+// outout: h[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
+template <LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|     LoopScheduler|     Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                  |             |
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|          LoopScheduler|              Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                       |                      |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Interwave, PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances<
+            LoopScheduler::Interwave,
+            PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v2>{});
+#endif
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// h = layernorm(e, gamma, beta)
+// outout: h[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
+template <LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|     LoopScheduler|     Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                  |             |
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 4>,               8,             S<16, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 4>,               8,             S<16, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 4>,               8,             S<16, 4>,               1, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|          LoopScheduler|              Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                       |                      |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Interwave, PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances<
+            LoopScheduler::Interwave,
+            PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v2>{});
+#endif
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_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_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.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/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasesDataType,
+          typename Acc1BiasesDataType,
+          tensor_operation::device::MaskingSpecialization MaskingSpec>
+bool profile_batched_gemm_bias_softmax_gemm_permute_impl(bool do_verification,
+                                                         int init_method,
+                                                         bool do_log,
+                                                         bool time_kernel,
+                                                         int M,
+                                                         int N,
+                                                         int K,
+                                                         int O,
+                                                         int G0,
+                                                         int G1,
+                                                         float alpha = -1.f)
+
+{
+
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using ScaleAdd      = tensor_operation::element_wise::ScaleAdd;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using C0DEElementOp = ScaleAdd;
+    using Acc0ElementOp = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+    using D0DataType    = tuple_element_t<0, Acc0BiasesDataType>;
+    using tensor_operation::device::MaskingSpecialization;
+
+    // Ref Gemm0: various type in, fp32 out
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+    // Ref Softmax: fp32 in, various type out
+    using ReferenceSoftmaxInstance =
+        tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+    // Ref Gemm1: various type in, various type out
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    const int BatchCount = G0 * G1;
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<D0DataType> d0_gs_ms_ns(d0_gs_ms_ns_lengths, d0_gs_ms_ns_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    std::srand(1); // work around test flakiness
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        // Still unsure whether this kind of deterministic floating point accurary issue is expected
+        // or not. May want to try exact same approach as the GPU kernel in the host reference
+        // GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
+        // shrink the input value range as it is less likely to produce errors of around ~1e-3.
+        // a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        // b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        // b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+    d0_device_buf.ToDevice(d0_gs_ms_ns.mData.data());
+
+    if(alpha < 0)
+    {
+        alpha = 1.f / std::sqrt(K); // usually 1 / sqrt(head_dim)
+    }
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto c0de_element_op = C0DEElementOp{alpha};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp =
+        tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      Acc0BiasesDataType,
+                                                                      ck::Tuple<>,
+                                                                      AElementOp,
+                                                                      B0ElementOp,
+                                                                      C0DEElementOp,
+                                                                      B1ElementOp,
+                                                                      CElementOp,
+                                                                      MaskingSpec>;
+
+    // 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)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+        Tensor<D0DataType> d0_g_m_n({BatchCount, M, N});
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        d0_gs_ms_ns.ForEach([&](auto& self, auto idx) {
+            d0_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+
+        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, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        acc0_g_m_n.ForEach([&](auto&, auto idx) {
+            c0de_element_op(acc0_g_m_n(idx), acc0_g_m_n(idx), d0_g_m_n(idx));
+        });
+        // mask out upper triangle
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle && idx[1] < idx[2])
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_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);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+    }
+
+    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_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+            std::array<void*, 1>{
+                d0_device_buf.GetDeviceBuffer()}, // std::array<void*, 1> p_acc0_biases;
+            {},                                   // std::array<void*, 1> p_acc1_biases;
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_strides}, // std::array<std::vector<ck::index_t>,
+                                      // 1>{acc0_biases_gs_ms_ns_strides},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+            a_element_op,
+            b0_element_op,
+            c0de_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 +
+                                     sizeof(D0DataType) * M * N) *
+                                    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_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+                // default absolute error and relative error is 0.001
+                double rtol = 1e-3;
+                double atol = 1e-3;
+
+                // when BF16 is taken, set absolute error and relative error to 0.01
+                if(std::is_same_v<ADataType, ck::bhalf_t> &&
+                   std::is_same_v<B0DataType, ck::bhalf_t> &&
+                   std::is_same_v<B1DataType, ck::bhalf_t> &&
+                   std::is_same_v<CDataType, ck::bhalf_t> &&
+                   std::is_same_v<D0DataType, ck::bhalf_t>)
+                {
+                    rtol = 1e-2;
+                    atol = 1e-2;
+                }
+
+                pass = pass & ck::utils::check_err(c_gs_ms_os_device_result,
+                                                   c_gs_ms_os_host_result,
+                                                   "Error: Incorrect results!",
+                                                   rtol,
+                                                   atol);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_gs_ms_ks: ", a_gs_ms_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_gs_ns_ks : ", b0_gs_ns_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_gs_os_ns : ", b1_gs_os_ns.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_gs_ms_os_host_result : ", c_gs_ms_os_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_gs_ms_os_device_result : ",
+                                          c_gs_ms_os_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/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.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/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EMeanVarDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp,
+          typename HElementOp>
+void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<BDataType>& b_k_n,
+                         const Tensor<D0DataType>& d0_m_n,
+                         const Tensor<D1DataType>& d1_m_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<BetaDataType>& beta_n,
+                         AElementOp a_element_op,
+                         BElementOp b_element_op,
+                         CDEElementOp cde_element_op,
+                         HElementOp h_element_op,
+                         int M,
+                         int N,
+                         AccDataType epsilon = 1e-5)
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    using ReferenceGemm = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                    BDataType,
+                                                                    AccDataType,
+                                                                    AccDataType,
+                                                                    AElementOp,
+                                                                    BElementOp,
+                                                                    PassThrough>;
+
+    using ReferenceLayernorm = ck::tensor_operation::host::ReferenceLayernorm<EMeanVarDataType,
+                                                                              GammaDataType,
+                                                                              BetaDataType,
+                                                                              HDataType,
+                                                                              AccDataType,
+                                                                              HElementOp,
+                                                                              2,
+                                                                              1>;
+
+    Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
+    Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
+
+    auto ref_gemm         = ReferenceGemm{};
+    auto ref_gemm_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_gemm_argument =
+        ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+    ref_gemm_invoker.Run(ref_gemm_argument);
+
+    for(int n = 0; n < N; ++n)
+    {
+        for(int m = 0; m < M; ++m)
+        {
+            AccDataType e  = static_cast<AccDataType>(e_m_n(m, n));
+            AccDataType d0 = static_cast<AccDataType>(d0_m_n(m, n));
+            AccDataType d1 = static_cast<AccDataType>(d1_m_n(m, n));
+            cde_element_op(e, c_m_n(m, n), d0, d1);
+            e_m_n(m, n) = static_cast<EMeanVarDataType>(e);
+        }
+    }
+
+    ReferenceLayernorm ref_layernorm;
+    auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
+
+    auto ref_layernorm_argument = ref_layernorm.MakeArgument(
+        e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
+    ref_layernorm_invoker.Run(ref_layernorm_argument);
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EMeanVarDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename HLayout>
+bool profile_gemm_add_relu_add_layernorm_impl(int do_verification,
+                                              int init_method,
+                                              bool /*do_log*/,
+                                              bool time_kernel,
+                                              int M,
+                                              int N,
+                                              int K,
+                                              int StrideA,
+                                              int StrideB,
+                                              int StrideD0,
+                                              int StrideD1,
+                                              int StrideH,
+                                              AccDataType epsilon = 1e-5)
+{
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(std::is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<D1DataType> d0_m_n(f_host_tensor_descriptor2d(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, D1Layout{}));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<HDataType> h_m_n(f_host_tensor_descriptor2d(M, N, StrideH, HLayout{}));
+    Tensor<HDataType> h_m_n_host(f_host_tensor_descriptor2d(M, N, StrideH, HLayout{}));
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
+        gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+        beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+        break;
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddReluAdd;
+    using HElementOp   = PassThrough;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+    const auto h_element_op   = HElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        HLayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        GammaDataType,
+        BetaDataType,
+        HDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddReluAdd,
+        ck::tensor_operation::element_wise::PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        host_gemm_layernorm<ADataType,
+                            BDataType,
+                            AccDataType,
+                            D0DataType,
+                            D1DataType,
+                            EMeanVarDataType,
+                            GammaDataType,
+                            BetaDataType,
+                            HDataType>(h_m_n_host,
+                                       a_m_k,
+                                       b_k_n,
+                                       d0_m_n,
+                                       d1_m_n,
+                                       gamma_n,
+                                       beta_n,
+                                       a_element_op,
+                                       b_element_op,
+                                       cde_element_op,
+                                       h_element_op,
+                                       M,
+                                       N,
+                                       epsilon);
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
+    DeviceMem h_device_buf(sizeof(HDataType) * h_m_n.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    bool pass      = true;
+    int num_kernel = 0;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_m_n_device_buf.GetDeviceBuffer(), d1_m_n_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            DeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+            // re-init E to zero before profiling a kernel
+            h_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                (sizeof(D0DataType) + sizeof(D1DataType) + sizeof(HDataType)) * M * N +
+                (sizeof(GammaDataType) + sizeof(BetaDataType)) * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            if(time_kernel)
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec
+                          << " GB/s, " << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                h_device_buf.FromDevice(h_m_n.mData.data());
+
+                pass = pass && ck::utils::check_err(
+                                   h_m_n, h_m_n_host, "Error: Incorrect results h_m_n", 1e-2, 1e-2);
+            }
+        }
+        else
+        {
+            if(time_kernel)
+                std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        pass = false;
+    }
+    else
+    {
+        if(time_kernel)
+            std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                      << best_op_name << std::endl;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck