Merge branch 'develop' into amd-develop

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp

@@ -1080,6 +1080,19 @@ struct GridwiseGemm_xdl_cshuffle_v3
             }
         }
 
+        if constexpr(is_same<remove_cvref_t<CDataType>, bhalf_t>::value)
+        {
+#if DEBUG_LOG
+            std::cout << " KBatch: " << karg.KBatch << " > 1 is not support yet" << __FILE__ << ":"
+                      << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+            if(karg.KBatch > 1)
+            {
+                return false;
+            }
+        }
+
         // check gridwise gemm pipeline
         const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value);
 

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_abd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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/block/blockwise_gemm_pipeline_xdlops_selector.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same
+// kernel function Blockers:
+// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on
+// two lds chunks.
+// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds
+// buffer when we declare __shared__ inside blkgemmpipe
+template <typename GridwiseGemm,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t MinimumOccupancy = 1,
+          TailNumber TailNum       = TailNumber::Full>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
+#endif
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_gemm_xdl_cshuffle_v3(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_as_grid,
+        karg.p_bs_grid,
+        karg.p_ds_grid,
+        karg.p_c_grid,
+        p_shared,
+        karg,
+        karg.a_element_op,
+        karg.b_element_op,
+        karg.c_element_op);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename GridwiseGemm,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t MinimumOccupancy = 1,
+          TailNumber TailNum       = TailNumber::Full>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
+#endif
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_gemm_xdl_cshuffle_v3_2lds(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    // Pass two lds pointer is the key to tell compiler that ds_read/write
+    // operate on different lds chunk at same time without order dependecy
+    __shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    __shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_as_grid,
+        karg.p_bs_grid,
+        karg.p_ds_grid,
+        karg.p_c_grid,
+        p_shared_0,
+        p_shared_1,
+        karg,
+        karg.a_element_op,
+        karg.b_element_op,
+        karg.c_element_op);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AsDataType,
+          typename BsDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v4,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA>
+struct GridwiseGemm_xdl_cshuffle_v3
+{
+    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>{};
+
+    using LDSTypeA = ComputeTypeA;
+    using LDSTypeB = ComputeTypeB;
+
+    // K1 should be Number<...>
+    static constexpr auto AK0Number = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0Number = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1Number = Number<AK1Value>{};
+    static constexpr auto BK1Number = Number<BK1Value>{};
+
+    static constexpr index_t NumATensor = AsDataType::Size();
+    static constexpr index_t NumBTensor = BsDataType::Size();
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto MakeAsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using ADataType_ = remove_cvref_t<tuple_element_t<i.value, AsDataType>>;
+
+                return static_cast<const ADataType_*>(nullptr);
+            },
+            Number<NumATensor>{});
+    }
+
+    static constexpr auto MakeBsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using BDataType_ = remove_cvref_t<tuple_element_t<i.value, BsDataType>>;
+
+                return static_cast<const BDataType_*>(nullptr);
+            },
+            Number<NumBTensor>{});
+    }
+
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AsGridPointer = decltype(MakeAsGridPointer());
+    using BsGridPointer = decltype(MakeBsGridPointer());
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    static constexpr index_t KPack = math::max(
+        math::lcm(AK1Number, BK1Number),
+        MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>::selected_mfma.k_per_blk);
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, KBatch);
+    }
+
+    __host__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ static auto CalculateAK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / AK1Value);
+    }
+
+    __host__ static auto CalculateBK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / BK1Value);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * KPerBlock;
+    }
+
+    __host__ static auto CalculateKRead(index_t K, index_t K_Batch = 1)
+    {
+        constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+        auto K_t                = K_Batch * KReadVec;
+        return (K + K_t - 1) / K_t * KReadVec;
+    }
+
+    __host__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock);
+    }
+
+    __host__ static auto CalculateNBlock(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock);
+    }
+
+    template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+    __host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&)
+    {
+        constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
+        constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
+
+        return transform_tensor_descriptor(
+            TileDesc_K0_MN_K1{},
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
+                       make_unmerge_transform(make_tuple(
+                           Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
+        index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(M, MPad - M),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(MPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    __host__ __device__ static auto
+    MakeAsGridDescriptor_AK0_M_AK1(const index_t M,
+                                   const index_t MPad,
+                                   const index_t K,
+                                   const index_t KPad,
+                                   const std::array<index_t, NumATensor>& StrideAs,
+                                   const index_t AK0)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeAGridDescriptor_AK0_M_AK1(M, MPad, K, KPad, StrideAs[i], AK0);
+            },
+            Number<NumATensor>{});
+    }
+
+    __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
+        index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(N, NPad - N),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(NPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    __host__ __device__ static auto
+    MakeBsGridDescriptor_BK0_N_BK1(const index_t K,
+                                   const index_t KPad,
+                                   const index_t N,
+                                   const index_t NPad,
+                                   const std::array<index_t, NumBTensor>& StrideBs,
+                                   const index_t BK0)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeBGridDescriptor_BK0_N_BK1(K, KPad, N, NPad, StrideBs[i], BK0);
+            },
+            Number<NumBTensor>{});
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor<MXdlPerWave, MWaves, MPerXdl>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor<NXdlPerWave, NWaves, NPerXdl>(BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static auto
+    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(M, MPad - M),
+                                                          make_right_pad_transform(N, NPad - N)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    __host__ __device__ static auto MakeDsGridDescriptor_M_N(
+        index_t M, index_t MPad, index_t N, index_t NPad, std::array<index_t, NumDTensor> StrideDs)
+    {
+        return generate_tuple(
+            [&](auto i) { return MakeCGridDescriptor_M_N(M, MPad, N, NPad, StrideDs[i]); },
+            Number<NumDTensor>{});
+    }
+
+    struct Problem
+    {
+        __host__ Problem(index_t M_,
+                         index_t N_,
+                         index_t K_,
+                         std::array<index_t, NumATensor> StrideAs_,
+                         std::array<index_t, NumBTensor> StrideBs_,
+                         std::array<index_t, NumDTensor> StrideDs_,
+                         index_t StrideC_,
+                         index_t KBatch_)
+            : M{M_},
+              N{N_},
+              K{K_},
+              StrideAs{StrideAs_},
+              StrideBs{StrideBs_},
+              StrideDs{StrideDs_},
+              StrideC{StrideC_},
+              KBatch{KBatch_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              KRead{CalculateKRead(K_, KBatch_)},
+              KPadded{CalculateKPadded(K_, KBatch_)},
+              AK0{CalculateAK0Padded(K_, KBatch_)},
+              BK0{CalculateBK0Padded(K_, KBatch_)},
+              MBlock{CalculateMBlock(M_)},
+              NBlock{CalculateNBlock(N_)}
+        {
+        }
+
+        __host__ void Print() const
+        {
+            std::cout << "problem {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "KRead:" << KRead << ", "
+                      << "KP:" << KPadded << ", "
+                      << "AK0:" << AK0 << ", "
+                      << "BK0:" << BK0 << ", "
+                      << "MBlock: " << MBlock << ", "
+                      << "NBlock: " << NBlock << "}" << std::endl;
+        }
+
+        index_t M;
+        index_t N;
+        index_t K;
+
+        std::array<index_t, NumATensor> StrideAs;
+        std::array<index_t, NumBTensor> StrideBs;
+        std::array<index_t, NumDTensor> StrideDs;
+        index_t StrideC;
+
+        index_t KBatch;
+        index_t MPadded;
+        index_t NPadded;
+        index_t KRead;
+        index_t KPadded;
+        index_t AK0;
+        index_t BK0;
+        index_t MBlock;
+        index_t NBlock;
+    };
+
+    // Argument
+    struct Argument : public tensor_operation::device::BaseArgument, public Problem
+    {
+        __host__ Argument(std::array<const void*, NumATensor> p_as_grid_,
+                          std::array<const void*, NumBTensor> p_bs_grid_,
+                          std::array<const void*, NumDTensor> p_ds_grid_,
+                          void* p_c_grid_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          std::array<index_t, NumATensor> StrideAs_,
+                          std::array<index_t, NumBTensor> StrideBs_,
+                          std::array<index_t, NumDTensor> StrideDs_,
+                          index_t StrideC_,
+                          index_t k_batch_,
+                          AElementwiseOperation a_element_op_,
+                          BElementwiseOperation b_element_op_,
+                          CElementwiseOperation c_element_op_)
+            : Problem{M_, N_, K_, StrideAs_, StrideBs_, StrideDs_, StrideC_, k_batch_},
+              p_as_grid{},
+              p_bs_grid{},
+              p_ds_grid{},
+              p_c_grid{static_cast<CDataType*>(p_c_grid_)},
+              a_element_op{a_element_op_},
+              b_element_op{b_element_op_},
+              c_element_op{c_element_op_}
+
+        {
+            // populate pointer, desc for As
+            static_for<0, NumATensor, 1>{}([&](auto i) {
+                using ADataType_ = remove_cvref_t<tuple_element_t<i.value, AsDataType>>;
+
+                // A pointer
+                p_as_grid(i) = static_cast<const ADataType_*>(p_as_grid_[i]);
+            });
+
+            // populate pointer, desc for Bs
+            static_for<0, NumBTensor, 1>{}([&](auto i) {
+                using BDataType_ = remove_cvref_t<tuple_element_t<i.value, BsDataType>>;
+
+                // B pointer
+                p_bs_grid(i) = static_cast<const BDataType_*>(p_bs_grid_[i]);
+            });
+
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType_ = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid(i) = static_cast<const DDataType_*>(p_ds_grid_[i]);
+            });
+        }
+
+        AsGridPointer p_as_grid;
+        BsGridPointer p_bs_grid;
+        DsGridPointer p_ds_grid;
+        CDataType* p_c_grid;
+
+        const AElementwiseOperation a_element_op;
+        const BElementwiseOperation b_element_op;
+        const CElementwiseOperation c_element_op;
+    };
+
+    struct SplitKBatchOffset
+    {
+        __device__ SplitKBatchOffset(Argument& karg)
+        {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                a_k_split_offset = blockIdx.z * karg.KRead;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                a_k_split_offset = blockIdx.z * karg.KRead * karg.M;
+            }
+
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
+            {
+                b_k_split_offset = blockIdx.z * karg.KRead * karg.N;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
+            {
+                b_k_split_offset = blockIdx.z * karg.KRead;
+            }
+
+            if(blockIdx.z < static_cast<uint32_t>(karg.KBatch - 1))
+            {
+                karg.K = karg.KRead;
+            }
+            else
+            {
+                karg.K = karg.K - karg.KRead * (karg.KBatch - 1);
+            }
+        }
+
+        index_t a_k_split_offset;
+        index_t b_k_split_offset;
+    };
+
+#if 0
+    struct SplitKBatchOffsetMultiABD
+    {
+        __device__ SplitKBatchOffsetMultiABD(AsGridPointer& p_as_grid,
+                                             BsGridPointer& p_bs_grid,
+                                             Argument& karg)
+        {
+            static_for<0, NumATensor, 1>{}([&](auto i) {
+                using ALayout_ = remove_cvref_t<tuple_element_t<i.value, AsLayout>>;
+                if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout_>)
+                {
+                    as_k_split_offset[i] = blockIdx.z * karg.KRead;
+                }
+                else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout_>)
+                {
+                    as_k_split_offset[i] = blockIdx.z * karg.KRead * karg.StrideAs[i];
+                }
+
+                p_as_grid_(i) = p_as_grid[i] + as_k_split_offset[i];
+            });
+
+            static_for<0, NumBTensor, 1>{}([&](auto i) {
+                using BLayout_ = remove_cvref_t<tuple_element_t<i.value, BsLayout>>;
+                if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout_>)
+                {
+                    bs_k_split_offset[i] = blockIdx.z * karg.KRead * karg.StrideBs[i];
+                }
+                else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout_>)
+                {
+                    bs_k_split_offset[i] = blockIdx.z * karg.KRead;
+                }
+
+                p_bs_grid_(i) = p_bs_grid[i] + bs_k_split_offset[i];
+            });
+
+            if(blockIdx.z < static_cast<uint32_t>(karg.KBatch - 1))
+            {
+                karg.K = karg.KRead;
+            }
+            else
+            {
+                karg.K = karg.K - karg.KRead * (karg.KBatch - 1);
+            }
+        }
+
+        AsGridPointer p_as_grid_;
+        BsGridPointer p_bs_grid_;
+        std::array<index_t, NumATensor> as_k_split_offset;
+        std::array<index_t, NumBTensor> bs_k_split_offset;
+    };
+#endif
+
+    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        if constexpr(ABlockLdsExtraM)
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock + ABlockLdsExtraM>{}, I1));
+        }
+        // xor tensor transformation request more unnecessary vgpr usage, would cause register spill
+        // in some cases.
+        else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            constexpr auto MLdsLayer        = 32 * 4 / KPerBlock / sizeof(LDSTypeA) < 1
+                                                  ? 1
+                                                  : 32 * 4 / KPerBlock / sizeof(LDSTypeA);
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
+                make_tuple(
+                    AK0Number * Number<MLdsLayer>{}, Number<MPerBlock / MLdsLayer>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock * MLdsLayer>{}, I1));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(make_xor_transform(make_tuple(Number<MPerBlock / MLdsLayer>{},
+                                                         Number<AK0Number * MLdsLayer>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
+
+            constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number<MLdsLayer>{})),
+                           make_pass_through_transform(Number<MPerBlock / MLdsLayer>{}),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
+
+            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_ak0_mldslayer_m_ak1,
+                make_tuple(make_pass_through_transform(AK0Number),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<MPerBlock / MLdsLayer>{}, Number<MLdsLayer>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_ak0_m_ak1;
+        }
+        else // ColumnMajor A
+        {
+            // kfold and mpair dimension is not always required.
+            // more dimension in merge_transform increase the difficulty of generating immarg offset
+            // for compiler.
+            constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
+            constexpr auto M1 = MPerBlock / M0;
+
+            constexpr auto KThreadWrite     = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0);
+            constexpr auto K0PerThreadWrite = AK0Number / KThreadWrite;
+            constexpr auto KThreadRead      = 64 / MPerXdl;
+            constexpr auto K0PerThreadRead  = AK0Number / KThreadRead;
+
+            constexpr auto kfold = (AK1Number * M0 * sizeof(LDSTypeA) > 128)
+                                       ? 1
+                                       : 128 / (AK1Number * M0 * sizeof(LDSTypeA));
+            constexpr auto KThreadReadPerm =
+                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
+                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
+                    : KThreadRead;
+
+            // 1<=mpair<=n0
+            constexpr auto mpair = (AK1Number * MPerXdl * sizeof(LDSTypeA) > 128)
+                                       ? 1
+                                       : ((128 / (AK1Number * MPerXdl * sizeof(LDSTypeA))) > M0
+                                              ? M0
+                                              : 128 / (AK1Number * MPerXdl * sizeof(LDSTypeA)));
+
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                           Number<K0PerThreadWrite>{},
+                           Number<KThreadReadPerm * M1>{},
+                           Number<kfold * M0 / mpair>{},
+                           Number<mpair>{},
+                           AK1Number));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_xor_transform(
+                        make_tuple(Number<KThreadReadPerm * M1>{}, Number<kfold * M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
+
+            constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<M1>{})),
+                    make_unmerge_transform(make_tuple(Number<kfold>{}, Number<M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<0, 3>{},
+                           Sequence<4, 5>{},
+                           Sequence<6>{},
+                           Sequence<7>{}));
+
+            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_unmerged,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(Number<KThreadReadPerm>{},
+                                          Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                                          Number<kfold>{},
+                                          Number<K0PerThreadWrite>{})),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<M0 / mpair>{}, Number<mpair>{}, Number<M1>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_ak0_m_ak1;
+        }
+    }
+
+    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        if constexpr(BBlockLdsExtraN)
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
+                make_tuple(BK1Number, Number<KPerBlock + BBlockLdsExtraN>{}, I1));
+        }
+        else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+        {
+            // NLdsLayer * K0 as logical Bank
+            constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeB) < 1
+                                           ? 1
+                                           : 32 * 4 / KPerBlock / sizeof(LDSTypeB);
+            ;
+            constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
+                make_tuple(
+                    BK0Number * Number<NLdsLayer>{}, Number<NPerBlock / NLdsLayer>{}, BK1Number),
+                make_tuple(BK1Number, Number<KPerBlock * NLdsLayer>{}, I1));
+
+            constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
+                b_lds_block_desc,
+                make_tuple(make_xor_transform(make_tuple(Number<NPerBlock / NLdsLayer>{},
+                                                         Number<BK0Number * NLdsLayer>{})),
+                           make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
+
+            constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor(
+                b_lds_block_desc_permuted,
+                make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number<NLdsLayer>{})),
+                           make_pass_through_transform(Number<NPerBlock / NLdsLayer>{}),
+                           make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
+
+            constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_lds_block_desc_bk0_nldslayer_n_bk1,
+                make_tuple(make_pass_through_transform(BK0Number),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<NPerBlock / NLdsLayer>{}, Number<NLdsLayer>{})),
+                           make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return b_lds_block_desc_bk0_n_bk1;
+        }
+        else // RowMajor B
+        {
+            constexpr auto N0 = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I1);
+            constexpr auto N1 = NPerBlock / N0;
+
+            constexpr auto KThreadWrite     = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I0);
+            constexpr auto K0PerThreadWrite = BK0Number / KThreadWrite;
+            constexpr auto KThreadRead      = 64 / NPerXdl;
+            constexpr auto K0PerThreadRead  = BK0Number / KThreadRead;
+
+            constexpr auto kfold = (BK1Number * N0 * sizeof(LDSTypeB) > 128)
+                                       ? 1
+                                       : 128 / (BK1Number * N0 * sizeof(LDSTypeB));
+            constexpr auto KThreadReadPerm =
+                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
+                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
+                    : KThreadRead;
+
+            // 1<=npair<=n0
+            constexpr auto npair = (BK1Number * NPerXdl * sizeof(LDSTypeB) > 128)
+                                       ? 1
+                                       : ((128 / (BK1Number * NPerXdl * sizeof(LDSTypeB))) > N0
+                                              ? N0
+                                              : 128 / (BK1Number * NPerXdl * sizeof(LDSTypeB)));
+
+            constexpr auto b_lds_block_desc = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                           Number<K0PerThreadWrite>{},
+                           Number<KThreadReadPerm * N1>{},
+                           Number<kfold * N0 / npair>{},
+                           Number<npair>{},
+                           BK1Number));
+
+            constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
+                b_lds_block_desc,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_xor_transform(
+                        make_tuple(Number<KThreadReadPerm * N1>{}, Number<kfold * N0 / npair>{})),
+                    make_pass_through_transform(Number<npair>{}),
+                    make_pass_through_transform(BK1Number)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
+
+            constexpr auto b_lds_block_desc_unmerged = transform_tensor_descriptor(
+                b_lds_block_desc_permuted,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<N1>{})),
+                    make_unmerge_transform(make_tuple(Number<kfold>{}, Number<N0 / npair>{})),
+                    make_pass_through_transform(Number<npair>{}),
+                    make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<0, 3>{},
+                           Sequence<4, 5>{},
+                           Sequence<6>{},
+                           Sequence<7>{}));
+
+            constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_lds_block_desc_unmerged,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(Number<KThreadReadPerm>{},
+                                          Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                                          Number<kfold>{},
+                                          Number<K0PerThreadWrite>{})),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<N0 / npair>{}, Number<npair>{}, Number<N1>{})),
+                           make_pass_through_transform(BK1Number)),
+                make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return b_lds_block_desc_bk0_n_bk1;
+        }
+    }
+
+    __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * 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;
+    }
+
+    using BlockwiseGemmPipe =
+        remove_cvref_t<decltype(BlockGemmPipeline_Selector<
+                                BlkGemmPipelineVer,
+                                BlkGemmPipeSched,
+                                BlockSize,
+                                LDSTypeA,
+                                LDSTypeB,
+                                ComputeTypeA,
+                                AccDataType,
+                                decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()),
+                                decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()),
+                                decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(
+                                    GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())),
+                                decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(
+                                    GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())),
+                                ABlockTransferSrcScalarPerVector,
+                                BBlockTransferSrcScalarPerVector,
+                                MPerBlock,
+                                NPerBlock,
+                                KPerBlock,
+                                MPerXdl,
+                                NPerXdl,
+                                MXdlPerWave,
+                                NXdlPerWave,
+                                KPack>())>;
+
+    __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned * sizeof(LDSTypeA) +
+                          b_block_space_size_aligned * sizeof(LDSTypeB)),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ static constexpr bool CheckValidity(const Argument& karg)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(karg.M % MPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(karg.N % NPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+
+            auto K_t = karg.KBatch * KPerBlock;
+            if(!(karg.K % K_t == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
+                          << karg.K << " " << __FILE__ << ":" << __LINE__
+                          << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+            auto K_t                = karg.KBatch * KReadVec;
+            auto KReadPadSplited    = math::integer_divide_ceil(karg.K, K_t) * KReadVec;
+            if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(karg.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M (" << karg.M
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(karg.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N (" << karg.N
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N (" << karg.N
+                          << ") value is not a multiple of "
+                             "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                          << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
+                          << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M (" << karg.M
+                          << ") value is not a multiple of "
+                             "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                          << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
+                          << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value);
+
+        if constexpr(BlkGemmPipelineVer != BlockGemmPipelineVersion::v1)
+        {
+            if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages)
+            {
+                return false;
+            }
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockHasHotloop(num_loop);
+    }
+
+    __host__ static constexpr TailNumber CalculateKBlockLoopTailNum(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockLoopTailNum(num_loop);
+    }
+
+    template <typename CGridDesc>
+    __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    template <typename DsGridDesc>
+    __device__ static constexpr auto MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDesc& ds_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n[i], MBlock, NBlock);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N(0, 0, 0, 0, {}))>;
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    // if arch = gfx942
+    using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
+    // using Block2CTileMap = BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>;
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run(AsGridPointer& p_as_grid,
+                               BsGridPointer& p_bs_grid,
+                               DsGridPointer& p_ds_grid,
+                               CDataType* p_c_grid,
+                               void* p_shared,
+                               const Problem& problem,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CElementwiseOperation& c_element_op)
+    {
+        // std::array<index_t, NumATensor> StrideAs = {problem.StrideA};
+        // std::array<index_t, NumBTensor> StrideBs = {problem.StrideB};
+
+        // AsGridPointer p_as_grid;
+        // BsGridPointer p_bs_grid;
+        // DsGridPointer p_ds_grid;
+
+        // const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+        //    problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
+        // const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
+        //    problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
+        const auto as_grid_desc_ak0_m_ak1 = MakeAsGridDescriptor_AK0_M_AK1(
+            problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideAs, problem.AK0);
+        const auto bs_grid_desc_bk0_n_bk1 = MakeBsGridDescriptor_BK0_N_BK1(
+            problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideBs, problem.BK0);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
+            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                c_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+        const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N(
+            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs);
+
+#if 0
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_m_n(j) = MakeCGridDescriptor_M_N(
+                problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs[j]);
+        });
+#endif
+
+        const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                ds_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+        // 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_bs_grid[I0], b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto as_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_as_grid[i], as_grid_desc_ak0_m_ak1[i].GetElementSpaceSize());
+            },
+            Number<NumATensor>{});
+
+        const auto bs_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_bs_grid[i], bs_grid_desc_bk0_n_bk1[i].GetElementSpaceSize());
+            },
+            Number<NumBTensor>{});
+
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i], ds_grid_desc_m_n[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        // divide block work by [M, N]
+        const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4};
+
+        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;
+        }
+
+        const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+        const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        // 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();
+
+#if 0
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0Number, MPerBlock, AK1Number>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ADataType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                BlockwiseGemmPipe::GlobalBufferNum>(
+                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),
+                ck::tensor_operation::element_wise::PassThrough{});
+#else
+        const auto idx_as_block_begin =
+            generate_tuple([&](auto) { return make_multi_index(0, m_block_data_idx_on_grid, 0); },
+                           Number<NumATensor>{});
+
+        auto a_blockwise_copy = ThreadGroupTensorSliceTransfer_v7r2<
+            ThisThreadBlock,
+            AsDataType,
+            Tuple<LDSTypeA>,
+            decltype(as_grid_desc_ak0_m_ak1),
+            decltype(tie(a_block_desc_ak0_m_ak1)),
+            AElementwiseOperation,
+            Sequence<static_cast<index_t>(InMemoryDataOperationEnum::Set)>,
+            Sequence<AK0Number, MPerBlock, AK1Number>,
+            ABlockTransferThreadClusterLengths_AK0_M_AK1,
+            ABlockTransferThreadClusterArrangeOrder,
+            ABlockTransferSrcAccessOrder,
+            Sequence<1, 0, 2>,
+            ABlockTransferSrcVectorDim,
+            2,
+            ABlockTransferSrcScalarPerVector,
+            ABlockTransferDstScalarPerVector_AK1,
+            uniform_sequence_gen_t<NumATensor, AThreadTransferSrcResetCoordinateAfterRun>,
+            Sequence<true>,
+            BlockwiseGemmPipe::GlobalBufferNum>{as_grid_desc_ak0_m_ak1,
+                                                idx_as_block_begin,
+                                                tie(a_block_desc_ak0_m_ak1),
+                                                make_tuple(make_multi_index(0, 0, 0)),
+                                                a_element_op};
+#endif
+
+#if 0
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0Number, NPerBlock, BK1Number>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BDataType,
+                                                BDataType,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                BlockwiseGemmPipe::GlobalBufferNum>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+#else
+        const auto idx_bs_block_begin =
+            generate_tuple([&](auto) { return make_multi_index(0, n_block_data_idx_on_grid, 0); },
+                           Number<NumBTensor>{});
+
+        auto b_blockwise_copy = ThreadGroupTensorSliceTransfer_v7r2<
+            ThisThreadBlock,
+            BsDataType,
+            Tuple<LDSTypeB>,
+            decltype(bs_grid_desc_bk0_n_bk1),
+            decltype(tie(b_block_desc_bk0_n_bk1)),
+            BElementwiseOperation,
+            Sequence<static_cast<index_t>(InMemoryDataOperationEnum::Set)>,
+            Sequence<BK0Number, NPerBlock, BK1Number>,
+            BBlockTransferThreadClusterLengths_BK0_N_BK1,
+            BBlockTransferThreadClusterArrangeOrder,
+            BBlockTransferSrcAccessOrder,
+            Sequence<1, 0, 2>,
+            BBlockTransferSrcVectorDim,
+            2,
+            BBlockTransferSrcScalarPerVector,
+            BBlockTransferDstScalarPerVector_BK1,
+            uniform_sequence_gen_t<NumBTensor, BThreadTransferSrcResetCoordinateAfterRun>,
+            Sequence<true>,
+            BlockwiseGemmPipe::GlobalBufferNum>{bs_grid_desc_bk0_n_bk1,
+                                                idx_bs_block_begin,
+                                                tie(b_block_desc_bk0_n_bk1),
+                                                make_tuple(make_multi_index(0, 0, 0)),
+                                                b_element_op};
+
+#endif
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        // Cast after lds
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeB*>(p_shared) +
+                a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB),
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 0);
+
+        // Blockwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
+        auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
+        auto c_thread_buf            = blockwise_gemm_pipeline.GetCThreadBuffer();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (as_grid_desc_ak0_m_ak1[I0].GetLength(I0) * as_grid_desc_ak0_m_ak1[I0].GetLength(I2)) /
+            KPerBlock);
+
+        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(as_grid_desc_ak0_m_ak1,
+                                                                         a_block_desc_ak0_m_ak1,
+                                                                         a_blockwise_copy,
+                                                                         as_grid_buf,
+                                                                         a_block_buf,
+                                                                         a_block_slice_copy_step,
+                                                                         bs_grid_desc_bk0_n_bk1,
+                                                                         b_block_desc_bk0_n_bk1,
+                                                                         b_blockwise_copy,
+                                                                         bs_grid_buf,
+                                                                         b_block_buf,
+                                                                         b_block_slice_copy_step,
+                                                                         c_thread_buf,
+                                                                         num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+#if 0
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                CShuffleDataType,     // typename SrcData,
+                CDataType,            // 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_m_id, 0, block_n_id, 0),
+                 c_element_op};
+#else
+            using EDataType = CDataType;
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                c_grid_desc_mblock_mperblock_nblock_nperblock;
+
+            using CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock =
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock;
+            const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
+            const auto CDEShuffleBlockTransferScalarPerVector_NPerBlock =
+                CShuffleBlockTransferScalarPerVector_NPerBlock;
+
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r2<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                            // support arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
+                Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
+                3,                    // index_t SrcVectorDim,
+                3,                    // index_t DstVectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_m_id, 0, block_n_id, 0)),
+                 c_element_op};
+
+#endif
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+#if 0
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+#else
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+#endif
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+#if 0
+                // 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);
+                }
+#else
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(c_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+#endif
+            });
+        }
+    }
+
+#if 1
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run_2Lds(AsGridPointer& p_as_grid,
+                                    BsGridPointer& p_bs_grid,
+                                    DsGridPointer& p_ds_grid,
+                                    CDataType* p_c_grid,
+                                    void* p_shared_0,
+                                    void* p_shared_1,
+                                    const Problem& problem,
+                                    const AElementwiseOperation& a_element_op,
+                                    const BElementwiseOperation& b_element_op,
+                                    const CElementwiseOperation& c_element_op)
+    {
+        // const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+        //    problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
+        // const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
+        //    problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
+        const auto as_grid_desc_ak0_m_ak1 = MakeAsGridDescriptor_AK0_M_AK1(
+            problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideAs, problem.AK0);
+        const auto bs_grid_desc_bk0_n_bk1 = MakeBsGridDescriptor_BK0_N_BK1(
+            problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideBs, problem.BK0);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
+            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                c_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+        const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N(
+            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs);
+
+        const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                ds_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+        // 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 as_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_as_grid[i], as_grid_desc_ak0_m_ak1[i].GetElementSpaceSize());
+            },
+            Number<NumATensor>{});
+
+        const auto bs_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_bs_grid[i], bs_grid_desc_bk0_n_bk1[i].GetElementSpaceSize());
+            },
+            Number<NumBTensor>{});
+
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i], ds_grid_desc_m_n[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        // divide block work by [M, N]
+        const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4};
+
+        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;
+        }
+
+        const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+        const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        // 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();
+
+#if 0
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0Number, MPerBlock, AK1Number>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ADataType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                BlockwiseGemmPipe::GlobalBufferNum>(
+                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),
+                ck::tensor_operation::element_wise::PassThrough{});
+#else
+        const auto idx_as_block_begin =
+            generate_tuple([&](auto) { return make_multi_index(0, m_block_data_idx_on_grid, 0); },
+                           Number<NumATensor>{});
+
+        auto a_blockwise_copy = ThreadGroupTensorSliceTransfer_v7r2<
+            ThisThreadBlock,
+            AsDataType,
+            Tuple<LDSTypeA>,
+            decltype(as_grid_desc_ak0_m_ak1),
+            decltype(tie(a_block_desc_ak0_m_ak1)),
+            AElementwiseOperation,
+            Sequence<static_cast<index_t>(InMemoryDataOperationEnum::Set)>,
+            Sequence<AK0Number, MPerBlock, AK1Number>,
+            ABlockTransferThreadClusterLengths_AK0_M_AK1,
+            ABlockTransferThreadClusterArrangeOrder,
+            ABlockTransferSrcAccessOrder,
+            Sequence<1, 0, 2>,
+            ABlockTransferSrcVectorDim,
+            2,
+            ABlockTransferSrcScalarPerVector,
+            ABlockTransferDstScalarPerVector_AK1,
+            uniform_sequence_gen_t<NumATensor, false>,
+            Sequence<true>,
+            BlockwiseGemmPipe::GlobalBufferNum>{as_grid_desc_ak0_m_ak1,
+                                                idx_as_block_begin,
+                                                tie(a_block_desc_ak0_m_ak1),
+                                                make_tuple(make_multi_index(0, 0, 0)),
+                                                a_element_op};
+
+#endif
+
+#if 0
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0Number, NPerBlock, BK1Number>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BDataType,
+                                                BDataType,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                BlockwiseGemmPipe::GlobalBufferNum>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+#else
+        const auto idx_bs_block_begin =
+            generate_tuple([&](auto) { return make_multi_index(0, n_block_data_idx_on_grid, 0); },
+                           Number<NumBTensor>{});
+
+        auto b_blockwise_copy = ThreadGroupTensorSliceTransfer_v7r2<
+            ThisThreadBlock,
+            BsDataType,
+            Tuple<LDSTypeB>,
+            decltype(bs_grid_desc_bk0_n_bk1),
+            decltype(tie(b_block_desc_bk0_n_bk1)),
+            BElementwiseOperation,
+            Sequence<static_cast<index_t>(InMemoryDataOperationEnum::Set)>,
+            Sequence<BK0Number, NPerBlock, BK1Number>,
+            BBlockTransferThreadClusterLengths_BK0_N_BK1,
+            BBlockTransferThreadClusterArrangeOrder,
+            BBlockTransferSrcAccessOrder,
+            Sequence<1, 0, 2>,
+            BBlockTransferSrcVectorDim,
+            2,
+            BBlockTransferSrcScalarPerVector,
+            BBlockTransferDstScalarPerVector_BK1,
+            uniform_sequence_gen_t<NumBTensor, false>,
+            Sequence<true>,
+            BlockwiseGemmPipe::GlobalBufferNum>{bs_grid_desc_bk0_n_bk1,
+                                                idx_bs_block_begin,
+                                                tie(b_block_desc_bk0_n_bk1),
+                                                make_tuple(make_multi_index(0, 0, 0)),
+                                                b_element_op};
+#endif
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeA*>(p_shared_0), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeB*>(p_shared_0) +
+                a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB),
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeA*>(p_shared_1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeB*>(p_shared_1) +
+                a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB),
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong);
+        auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong);
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 0);
+
+        // Blockwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
+        auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
+        auto c_thread_buf            = blockwise_gemm_pipeline.GetCThreadBuffer();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (as_grid_desc_ak0_m_ak1[I0].GetLength(I0) * as_grid_desc_ak0_m_ak1[I0].GetLength(I2)) /
+            KPerBlock);
+
+        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(as_grid_desc_ak0_m_ak1,
+                                                                         a_block_desc_ak0_m_ak1,
+                                                                         a_blockwise_copy,
+                                                                         as_grid_buf,
+                                                                         a_block_bufs,
+                                                                         a_block_slice_copy_step,
+                                                                         bs_grid_desc_bk0_n_bk1,
+                                                                         b_block_desc_bk0_n_bk1,
+                                                                         b_blockwise_copy,
+                                                                         bs_grid_buf,
+                                                                         b_block_bufs,
+                                                                         b_block_slice_copy_step,
+                                                                         c_thread_buf,
+                                                                         num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared_0),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+#if 0
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                CShuffleDataType,     // typename SrcData,
+                CDataType,            // 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_m_id, 0, block_n_id, 0),
+                 c_element_op};
+#else
+            using EDataType = CDataType;
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                c_grid_desc_mblock_mperblock_nblock_nperblock;
+
+            using CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock =
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock;
+            const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
+            const auto CDEShuffleBlockTransferScalarPerVector_NPerBlock =
+                CShuffleBlockTransferScalarPerVector_NPerBlock;
+
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r2<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                            // support arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
+                Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
+                3,                    // index_t SrcVectorDim,
+                3,                    // index_t DstVectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
+                 c_element_op};
+
+#endif
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  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!");
+
+#if 1
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+#endif
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+#if 0
+                // 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);
+                }
+#else
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(c_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+#endif
+            });
+        }
+    }
+#endif
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp

@@ -42,7 +42,8 @@ template <typename SrcDatas,
           index_t SrcScalarPerVector,
           index_t DstScalarPerVector,
           typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
-          typename DstResetCoordinateAfterRunFlags> // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          index_t NumThreadScratch = 1>
 struct ThreadwiseTensorSliceTransfer_v7r2
 {
     static constexpr auto I0 = Number<0>{};
@@ -139,14 +140,19 @@ struct ThreadwiseTensorSliceTransfer_v7r2
     // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
     // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
     template <typename SrcBuffers,
+              index_t ThreadScratchId                                   = 0,
               enable_if_t<SrcDescs::Size() == SrcBuffers::Size(), bool> = false>
-    __device__ void RunRead(const SrcDescs& src_descs, const SrcBuffers& src_bufs)
+    __device__ void RunRead(const SrcDescs& src_descs,
+                            const SrcBuffers& src_bufs,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
     {
         // loop over space-filling curve
         static_for<0, src_num_access, 1>{}([&](auto iAccess) {
             auto src_vectors = generate_vectors<SrcDatas, SrcScalarPerVector>();
             auto elm_vectors = generate_vectors<DstDatas, SrcScalarPerVector>();
 
+            bool oob_val = true;
+
             // copy data from src_bufs into src_vectors
             static_for<0, nSrc, 1>{}([&](auto i) {
                 using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
@@ -155,9 +161,10 @@ struct ThreadwiseTensorSliceTransfer_v7r2
                     coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
                                                                                 src_coords_[i]);
 
+                oob_val = oob_val & is_src_valid;
+
                 src_vectors(i).template AsType<src_vector_t>()(I0) =
-                    src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(),
-                                                           is_src_valid);
+                    src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(), true);
             });
 
             constexpr auto get_elem_op_vec_len = []() {
@@ -218,7 +225,8 @@ struct ThreadwiseTensorSliceTransfer_v7r2
                 unpack2(element_op_, dst_data_refs, src_data_refs);
             });
 
-            elm_vectors_tuple_(iAccess) = elm_vectors;
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+            oob_vectors_tuple_(thread_scratch_id)(iAccess) = oob_val;
 
             // move coordinate
             if constexpr(iAccess.value != src_num_access - 1)
@@ -245,17 +253,38 @@ struct ThreadwiseTensorSliceTransfer_v7r2
         });
     }
 
-    __device__ void TransposeFromElmToDst()
+#if 1
+    template <index_t ThreadScratchId = 0>
+    __device__ void OOBCheck(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // loop over space-filling curve
+        static_for<0, src_num_access, 1>{}([&](auto iAccess) {
+            auto elm_vectors = elm_vectors_tuple_[thread_scratch_id][iAccess];
+            auto oob_val     = oob_vectors_tuple_[thread_scratch_id][iAccess];
+
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using elm_vector_t = typename remove_cvref_t<decltype(elm_vectors[i])>::type;
+                elm_vectors(i).template AsType<elm_vector_t>()(I0) =
+                    oob_val ? elm_vectors(i).template AsType<elm_vector_t>()[I0] : elm_vector_t{0};
+            });
+
+            elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors;
+        });
+    }
+#endif
+
+    template <index_t ThreadScratchId = 0>
+    __device__ void
+    TransposeFromElmToDst(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
     {
         using DstData = remove_cvref_t<decltype(DstDatas{}[I0])>;
 
-        using SrcThreadScratch =
+        using ElmThreadScratch =
             StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
                                             DstData,
                                             SrcScalarPerVector,
                                             decltype(GetSrcThreadScratchDescriptor()),
                                             true>;
-
         using DstThreadScratch =
             StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
                                             DstData,
@@ -263,15 +292,17 @@ struct ThreadwiseTensorSliceTransfer_v7r2
                                             decltype(GetDstThreadScratchDescriptor()),
                                             true>;
 
-        SrcThreadScratch elm_thread_scratch_;
+        ElmThreadScratch elm_thread_scratch_;
         DstThreadScratch dst_thread_scratch_;
 
         elm_thread_scratch_.data_ =
-            bit_cast<decltype(elm_thread_scratch_.data_)>(elm_vectors_tuple_);
+            bit_cast<decltype(elm_thread_scratch_.data_)>(elm_vectors_tuple_[thread_scratch_id]);
 
         if constexpr(SrcVectorDim != DstVectorDim &&
                      ((is_same<half_t, remove_cvref_t<DstData>>::value &&
                        SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<f8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0) ||
                       (is_same<int8_t, remove_cvref_t<DstData>>::value &&
                        SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
         {
@@ -338,20 +369,24 @@ struct ThreadwiseTensorSliceTransfer_v7r2
                 [&](auto idx) { dst_thread_scratch_(idx) = elm_thread_scratch_[idx]; });
         }
 
-        dst_vectors_tuple_ = bit_cast<decltype(dst_vectors_tuple_)>(dst_thread_scratch_.data_);
+        dst_vectors_tuple_(thread_scratch_id) = bit_cast<DstVectorTuple>(dst_thread_scratch_.data_);
     }
 
     // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
     // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
     template <typename DstBuffers,
+              index_t ThreadScratchId                                             = 0,
               enable_if_t<DstDescs::Size() == 1 && DstBuffers::Size() == 1, bool> = false>
-    __device__ void RunWrite(const DstDescs& dst_descs, DstBuffers dst_bufs)
+    __device__ void RunWrite(const DstDescs& dst_descs,
+                             DstBuffers dst_bufs,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
     {
-        TransposeFromElmToDst();
+        OOBCheck(thread_scratch_id);
+        TransposeFromElmToDst(thread_scratch_id);
 
         // loop over space-filling curve
         static_for<0, dst_num_access, 1>{}([&](auto iAccess) {
-            auto dst_vectors = dst_vectors_tuple_[Number<iAccess>{}];
+            auto dst_vectors = dst_vectors_tuple_[thread_scratch_id][iAccess];
 
             // copy data from buf_vectors into dst_bufs
             static_for<0, nDst, 1>{}([&](auto i) {
@@ -578,8 +613,14 @@ struct ThreadwiseTensorSliceTransfer_v7r2
     static constexpr auto src_num_access = SrcSpaceFillingCurve::GetNumOfAccess();
     static constexpr auto dst_num_access = DstSpaceFillingCurve::GetNumOfAccess();
 
-    StaticallyIndexedArray<ElmVectorsType, src_num_access> elm_vectors_tuple_;
-    StaticallyIndexedArray<DstVectorsType, dst_num_access> dst_vectors_tuple_;
+    using ElmVectorTuple = StaticallyIndexedArray<ElmVectorsType, src_num_access>;
+    using DstVectorTuple = StaticallyIndexedArray<DstVectorsType, dst_num_access>;
+
+    StaticallyIndexedArray<ElmVectorTuple, NumThreadScratch> elm_vectors_tuple_;
+    StaticallyIndexedArray<DstVectorTuple, NumThreadScratch> dst_vectors_tuple_;
+
+    using OOBVectorTuple = StaticallyIndexedArray<bool, src_num_access>;
+    StaticallyIndexedArray<OOBVectorTuple, NumThreadScratch> oob_vectors_tuple_;
 
     SrcCoords src_coords_;
     DstCoords dst_coords_;

include/ck/utility/debug.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef UTILITY_DEBUG_HPP
 #define UTILITY_DEBUG_HPP
@@ -79,6 +79,13 @@ __device__ void print_shared(T const* p_shared, index_t num_elements)
     __syncthreads();
 }
 
+template <index_t... Ids>
+__device__ static bool is_thread_local_1d_id_idx()
+{
+    const auto tid = get_thread_local_1d_id();
+    return ((tid == Ids) || ...);
+}
+
 } // namespace debug
 } // namespace ck
 

include/ck/utility/flush_icache.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+namespace ck {
+static __global__ void flush_icache()
+{
+    asm __volatile__("s_icache_inv \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t"
+                     "s_nop 0 \n\t" ::
+                         :);
+}
+} // namespace ck

include/ck/utility/loop_scheduler.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+#include <ostream>
 
 #pragma once
 
@@ -24,3 +25,14 @@ constexpr LoopScheduler make_default_loop_scheduler()
 }
 
 } // namespace ck
+
+inline std::ostream& operator<<(std::ostream& os, const ck::LoopScheduler& s)
+{
+    switch(s)
+    {
+    case ck::LoopScheduler::Default: os << "Default"; break;
+    case ck::LoopScheduler::Interwave: os << "Interwave"; break;
+    default: os << "";
+    }
+    return os;
+}

include/ck/utility/sequence.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
+#include <ostream>
+
 #include "ck/utility/integral_constant.hpp"
 #include "ck/utility/type.hpp"
 #include "ck/utility/functional.hpp"
@@ -897,3 +899,14 @@ template <index_t NSize, index_t I>
 using uniform_sequence_gen_t = typename uniform_sequence_gen<NSize, I>::type;
 
 } // namespace ck
+
+template <ck::index_t... Is>
+std::ostream& operator<<(std::ostream& os, const ck::Sequence<Is...>)
+{
+    using S = ck::Sequence<Is...>;
+    os << "{";
+    ck::static_for<0, S::Size() - ck::Number<1>{}, 1>{}(
+        [&](auto i) { os << S::At(i).value << ", "; });
+    os << S::At(S::Size() - ck::Number<1>{}).value << "}";
+    return os;
+}

include/ck/utility/type.hpp

@@ -40,23 +40,10 @@ inline constexpr bool is_pointer_v = std::is_pointer<T>::value;
 template <typename Y, typename X, typename enable_if<sizeof(X) == sizeof(Y), bool>::type = false>
 __host__ __device__ constexpr Y bit_cast(const X& x)
 {
-#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST
-    Y y;
-
-    // auto t = reinterpret_cast<const Y*>(&x);
-    // y      = *t;
-    __builtin_memcpy(&y, &x, sizeof(X));
-
-    return y;
-#else
-    union AsType
-    {
-        X x;
-        Y y;
-    };
-
-    return AsType{x}.y;
-#endif
+    static_assert(__has_builtin(__builtin_bit_cast), "");
+    static_assert(sizeof(X) == sizeof(Y), "Do not support cast between different size of type");
+
+    return __builtin_bit_cast(Y, x);
 }
 
 } // namespace ck

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

@@ -27,14 +27,16 @@ using Empty_Tuple = ck::Tuple<>;
 
 using BF16_Tuple = ck::Tuple<BF16>;
 
-using F16_Tuple     = ck::Tuple<F16>;
-using F16_F16_Tuple = ck::Tuple<F16, F16>;
+using F16_Tuple       = ck::Tuple<F16>;
+using F16_F16_Tuple   = ck::Tuple<F16, F16>;
+using BF16_BF16_Tuple = ck::Tuple<BF16, BF16>;
 
 using F64_Tuple     = ck::Tuple<F64>;
 using F32_Tuple     = ck::Tuple<F32>;
 using I32_Tuple     = ck::Tuple<I32>;
 using I32_F32_Tuple = ck::Tuple<I32, F32>;
 using I8_Tuple      = ck::Tuple<I8>;
+using BF16_Tuple    = ck::Tuple<BF16>;
 
 using F32_F32_Tuple = ck::Tuple<F32, F32>;
 
@@ -91,23 +93,26 @@ using GK_Tuple    = ck::Tuple<G_K>;
 using GK_GK_Tuple = ck::Tuple<G_K, G_K>;
 
 // pointwise functor
-using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
-using Relu           = ck::tensor_operation::element_wise::Relu;
-using TanH           = ck::tensor_operation::element_wise::TanH;
-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 AddRelu        = ck::tensor_operation::element_wise::AddRelu;
-using AddSilu        = ck::tensor_operation::element_wise::AddSilu;
-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 MultiplyAdd    = ck::tensor_operation::element_wise::MultiplyAdd;
-using ScaleAdd       = ck::tensor_operation::element_wise::ScaleAdd;
-using Gelu           = ck::tensor_operation::element_wise::Gelu;
-using Swish          = ck::tensor_operation::element_wise::Swish;
-using Add            = ck::tensor_operation::element_wise::Add;
+using PassThrough         = ck::tensor_operation::element_wise::PassThrough;
+using Relu                = ck::tensor_operation::element_wise::Relu;
+using TanH                = ck::tensor_operation::element_wise::TanH;
+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 MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
+using AddRelu             = ck::tensor_operation::element_wise::AddRelu;
+using AddSilu             = ck::tensor_operation::element_wise::AddSilu;
+using AddReluAdd          = ck::tensor_operation::element_wise::AddReluAdd;
+using FastGelu            = ck::tensor_operation::element_wise::FastGelu;
+using MultiplyFastGelu    = ck::tensor_operation::element_wise::MultiplyFastGelu;
+using AddMultiply         = ck::tensor_operation::element_wise::AddMultiply;
+using MultiplyAdd         = ck::tensor_operation::element_wise::MultiplyAdd;
+using ScaleAdd            = ck::tensor_operation::element_wise::ScaleAdd;
+using Gelu                = ck::tensor_operation::element_wise::Gelu;
+using Swish               = ck::tensor_operation::element_wise::Swish;
+using Add                 = ck::tensor_operation::element_wise::Add;
+using Multiply            = ck::tensor_operation::element_wise::Multiply;
 
 template <typename Activation>
 using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;

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

@@ -17,7 +17,7 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
 
@@ -33,7 +33,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_bias_gelu_v1_instances(
                                                       ck::Tuple<BF16>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       AddFastGelu>>>& instances);
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_bias_v1_instances(
@@ -46,7 +46,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_bias_v1_instances(
                                                       ck::Tuple<BF16>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       Add>>>& instances);
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_gelu_v1_instances(
@@ -59,7 +59,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_gelu_v1_instances(
                                                       ck::Tuple<>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       FastGelu>>>& instances);
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_v1_instances(
@@ -72,7 +72,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_v1_instances(
                                                       ck::Tuple<>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       PassThrough>>>& instances);
 
 // RCR
@@ -86,7 +86,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_bias_gelu_v1_instances(
                                                       ck::Tuple<BF16>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       AddFastGelu>>>& instances);
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_bias_v1_instances(
@@ -99,7 +99,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_bias_v1_instances(
                                                       ck::Tuple<BF16>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       Add>>>& instances);
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_gelu_v1_instances(
@@ -112,7 +112,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_gelu_v1_instances(
                                                       ck::Tuple<>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       FastGelu>>>& instances);
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_v1_instances(
@@ -125,7 +125,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_v1_instances(
                                                       ck::Tuple<>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       PassThrough>>>& instances);
 
 // CRR
@@ -139,7 +139,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_km_kn_mn_bias_gelu_v1_instances(
                                                       ck::Tuple<BF16>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       AddFastGelu>>>& instances);
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_km_kn_mn_bias_v1_instances(
@@ -152,7 +152,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_km_kn_mn_bias_v1_instances(
                                                       ck::Tuple<BF16>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       Add>>>& instances);
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_km_kn_mn_gelu_v1_instances(
@@ -165,7 +165,7 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_km_kn_mn_gelu_v1_instances(
                                                       ck::Tuple<>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       FastGelu>>>& instances);
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_km_kn_mn_v1_instances(
@@ -178,8 +178,62 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_km_kn_mn_v1_instances(
                                                       ck::Tuple<>,
                                                       BF16,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       PassThrough>>>& instances);
+
+// Multiply
+void add_device_gemm_xdl_multi_abd_multiply_bf16_i8_bf16_mk_kn_mn_bias_gelu_v1_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleABD<ck::Tuple<Row>,
+                                                      ck::Tuple<Row>,
+                                                      ck::Tuple<Row, Row>,
+                                                      Row,
+                                                      ck::Tuple<BF16>,
+                                                      ck::Tuple<I8>,
+                                                      ck::Tuple<BF16, BF16>,
+                                                      BF16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      MultiplyAddFastGelu>>>& instances);
+
+void add_device_gemm_xdl_multi_abd_multiply_bf16_i8_bf16_mk_kn_mn_bias_v1_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleABD<ck::Tuple<Row>,
+                                                      ck::Tuple<Row>,
+                                                      ck::Tuple<Row>,
+                                                      Row,
+                                                      ck::Tuple<BF16>,
+                                                      ck::Tuple<I8>,
+                                                      ck::Tuple<BF16, BF16>,
+                                                      BF16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      MultiplyAdd>>>& instances);
+
+void add_device_gemm_xdl_multi_abd_multiply_bf16_i8_bf16_mk_kn_mn_gelu_v1_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleABD<ck::Tuple<Row>,
+                                                      ck::Tuple<Row>,
+                                                      ck::Tuple<Row>,
+                                                      Row,
+                                                      ck::Tuple<BF16>,
+                                                      ck::Tuple<I8>,
+                                                      ck::Tuple<BF16>,
+                                                      BF16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      MultiplyFastGelu>>>& instances);
+
+void add_device_gemm_xdl_multi_abd_multiply_bf16_i8_bf16_mk_kn_mn_v1_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleABD<ck::Tuple<Row>,
+                                                      ck::Tuple<Row>,
+                                                      ck::Tuple<Row>,
+                                                      Row,
+                                                      ck::Tuple<BF16>,
+                                                      ck::Tuple<I8>,
+                                                      ck::Tuple<BF16>,
+                                                      BF16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Multiply>>>& instances);
+
 #endif
 
 // GEMM + Add + Gelu
@@ -201,7 +255,7 @@ struct DeviceOperationInstanceFactory<
                                                         DsDataType,
                                                         EDataType,
                                                         PassThrough,
-                                                        Scales,
+                                                        Multiply,
                                                         AddFastGelu>>
 {
     using DeviceOp = DeviceGemmMultipleABD<AsLayout,
@@ -213,7 +267,7 @@ struct DeviceOperationInstanceFactory<
                                            DsDataType,
                                            EDataType,
                                            PassThrough,
-                                           Scales,
+                                           Multiply,
                                            AddFastGelu>;
 
     static auto GetInstances()
@@ -271,7 +325,7 @@ struct DeviceOperationInstanceFactory<
                                                         DsDataType,
                                                         EDataType,
                                                         PassThrough,
-                                                        Scales,
+                                                        Multiply,
                                                         Add>>
 {
     using DeviceOp = DeviceGemmMultipleABD<AsLayout,
@@ -283,7 +337,7 @@ struct DeviceOperationInstanceFactory<
                                            DsDataType,
                                            EDataType,
                                            PassThrough,
-                                           Scales,
+                                           Multiply,
                                            Add>;
 
     static auto GetInstances()
@@ -341,7 +395,7 @@ struct DeviceOperationInstanceFactory<
                                                         DsDataType,
                                                         EDataType,
                                                         PassThrough,
-                                                        Scales,
+                                                        Multiply,
                                                         FastGelu>>
 {
     using DeviceOp = DeviceGemmMultipleABD<AsLayout,
@@ -353,7 +407,7 @@ struct DeviceOperationInstanceFactory<
                                            DsDataType,
                                            EDataType,
                                            PassThrough,
-                                           Scales,
+                                           Multiply,
                                            FastGelu>;
 
     static auto GetInstances()
@@ -411,7 +465,7 @@ struct DeviceOperationInstanceFactory<
                                                         DsDataType,
                                                         EDataType,
                                                         PassThrough,
-                                                        Scales,
+                                                        Multiply,
                                                         PassThrough>>
 {
     using DeviceOp = DeviceGemmMultipleABD<AsLayout,
@@ -423,7 +477,7 @@ struct DeviceOperationInstanceFactory<
                                            DsDataType,
                                            EDataType,
                                            PassThrough,
-                                           Scales,
+                                           Multiply,
                                            PassThrough>;
 
     static auto GetInstances()
@@ -462,6 +516,234 @@ struct DeviceOperationInstanceFactory<
     }
 };
 
+// Multiply
+// GEMM + Add + Gelu
+template <typename AsLayout,
+          typename BsLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename AsDataType,
+          typename BsDataType,
+          typename DsDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                        BsLayout,
+                                                        DsLayout,
+                                                        ELayout,
+                                                        AsDataType,
+                                                        BsDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        MultiplyAddFastGelu>>
+{
+    using DeviceOp = DeviceGemmMultipleABD<AsLayout,
+                                           BsLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           AsDataType,
+                                           BsDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           PassThrough,
+                                           PassThrough,
+                                           MultiplyAddFastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+#ifdef CK_ENABLE_INT8
+        if constexpr(is_same_v<AsDataType, ck::Tuple<BF16>> &&
+                     is_same_v<BsDataType, ck::Tuple<I8>> &&
+                     is_same_v<DsDataType, ck::Tuple<BF16, BF16>> && is_same_v<EDataType, BF16>)
+        {
+            if constexpr(is_same_v<AsLayout, ck::Tuple<Row>> &&
+                         is_same_v<BsLayout, ck::Tuple<Row>> &&
+                         is_same_v<DsLayout, ck::Tuple<Row, Row>> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_xdl_multi_abd_multiply_bf16_i8_bf16_mk_kn_mn_bias_gelu_v1_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+
+        return op_ptrs;
+    }
+};
+
+// GEMM + Add
+template <typename AsLayout,
+          typename BsLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename AsDataType,
+          typename BsDataType,
+          typename DsDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                        BsLayout,
+                                                        DsLayout,
+                                                        ELayout,
+                                                        AsDataType,
+                                                        BsDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        MultiplyAdd>>
+{
+    using DeviceOp = DeviceGemmMultipleABD<AsLayout,
+                                           BsLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           AsDataType,
+                                           BsDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           PassThrough,
+                                           PassThrough,
+                                           MultiplyAdd>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+#ifdef CK_ENABLE_INT8
+        if constexpr(is_same_v<AsDataType, ck::Tuple<BF16>> &&
+                     is_same_v<BsDataType, ck::Tuple<I8>> &&
+                     is_same_v<DsDataType, ck::Tuple<BF16, BF16>> && is_same_v<EDataType, BF16>)
+        {
+            if constexpr(is_same_v<AsLayout, ck::Tuple<Row>> &&
+                         is_same_v<BsLayout, ck::Tuple<Row>> &&
+                         is_same_v<DsLayout, ck::Tuple<Row, Row>> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_xdl_multi_abd_multiply_bf16_i8_bf16_mk_kn_mn_bias_v1_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+
+        return op_ptrs;
+    }
+};
+
+// GEMM + Gelu
+template <typename AsLayout,
+          typename BsLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename AsDataType,
+          typename BsDataType,
+          typename DsDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                        BsLayout,
+                                                        DsLayout,
+                                                        ELayout,
+                                                        AsDataType,
+                                                        BsDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        MultiplyFastGelu>>
+{
+    using DeviceOp = DeviceGemmMultipleABD<AsLayout,
+                                           BsLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           AsDataType,
+                                           BsDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           PassThrough,
+                                           PassThrough,
+                                           MultiplyFastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+#ifdef CK_ENABLE_INT8
+        if constexpr(is_same_v<AsDataType, ck::Tuple<BF16>> &&
+                     is_same_v<BsDataType, ck::Tuple<I8>> &&
+                     is_same_v<DsDataType, ck::Tuple<BF16>> && is_same_v<EDataType, BF16>)
+        {
+            if constexpr(is_same_v<AsLayout, ck::Tuple<Row>> &&
+                         is_same_v<BsLayout, ck::Tuple<Row>> &&
+                         is_same_v<DsLayout, ck::Tuple<Row>> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_xdl_multi_abd_multiply_bf16_i8_bf16_mk_kn_mn_gelu_v1_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+
+        return op_ptrs;
+    }
+};
+
+// GEMM
+template <typename AsLayout,
+          typename BsLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename AsDataType,
+          typename BsDataType,
+          typename DsDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                        BsLayout,
+                                                        DsLayout,
+                                                        ELayout,
+                                                        AsDataType,
+                                                        BsDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        Multiply>>
+{
+    using DeviceOp = DeviceGemmMultipleABD<AsLayout,
+                                           BsLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           AsDataType,
+                                           BsDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           PassThrough,
+                                           PassThrough,
+                                           Multiply>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+#ifdef CK_ENABLE_INT8
+        if constexpr(is_same_v<AsDataType, ck::Tuple<BF16>> &&
+                     is_same_v<BsDataType, ck::Tuple<I8>> &&
+                     is_same_v<DsDataType, ck::Tuple<BF16>> && is_same_v<EDataType, BF16>)
+        {
+            if constexpr(is_same_v<AsLayout, ck::Tuple<Row>> &&
+                         is_same_v<BsLayout, ck::Tuple<Row>> &&
+                         is_same_v<DsLayout, ck::Tuple<Row>> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_xdl_multi_abd_multiply_bf16_i8_bf16_mk_kn_mn_v1_instances(op_ptrs);
+            }
+        }
+#endif
+
+        return op_ptrs;
+    }
+};
+
 } // namespace instance
 } // namespace device
 } // namespace tensor_operation

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

@@ -315,6 +315,107 @@ void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instanc
         DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
+#ifdef CK_ENABLE_FP16
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_default_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_mnpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_mnkpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v1_mnkpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_kpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_mnkpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_default_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_mnpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_mnkpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
 
 template <typename ADataType,
           typename BDataType,
@@ -494,6 +595,64 @@ struct DeviceOperationInstanceFactory<
                     op_ptrs);
             }
         }
+#endif
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
+                     is_same_v<CDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_mnpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v1_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v1_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v1_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_v2_mnkpadding_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_mnpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
+                    op_ptrs);
+            }
+        }
 #endif
         return op_ptrs;
     }

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

@@ -17,7 +17,7 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
 
@@ -32,7 +32,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_kn_mn_bias_g
                                                                  ck::Tuple<BF16>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  AddFastGelu>>>& instances);
 
 void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_kn_mn_bias_instances(
@@ -45,7 +45,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_kn_mn_bias_i
                                                                  ck::Tuple<BF16>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  Add>>>& instances);
 
 void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_kn_mn_gelu_instances(
@@ -58,7 +58,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_kn_mn_gelu_i
                                                                  ck::Tuple<>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  FastGelu>>>& instances);
 
 void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_kn_mn_instances(
@@ -71,7 +71,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_kn_mn_instan
                                                                  ck::Tuple<>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  PassThrough>>>& instances);
 
 // RCR
@@ -85,7 +85,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_nk_mn_bias_g
                                                                  ck::Tuple<BF16>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  AddFastGelu>>>& instances);
 
 void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_nk_mn_bias_instances(
@@ -98,7 +98,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_nk_mn_bias_i
                                                                  ck::Tuple<BF16>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  Add>>>& instances);
 
 void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_nk_mn_gelu_instances(
@@ -111,7 +111,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_nk_mn_gelu_i
                                                                  ck::Tuple<>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  FastGelu>>>& instances);
 
 void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_nk_mn_instances(
@@ -124,7 +124,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_mk_nk_mn_instan
                                                                  ck::Tuple<>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  PassThrough>>>& instances);
 
 // CRR
@@ -138,7 +138,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_km_kn_mn_bias_g
                                                                  ck::Tuple<BF16>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  AddFastGelu>>>& instances);
 
 void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_km_kn_mn_bias_instances(
@@ -151,7 +151,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_km_kn_mn_bias_i
                                                                  ck::Tuple<BF16>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  Add>>>& instances);
 
 void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_km_kn_mn_gelu_instances(
@@ -164,7 +164,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_km_kn_mn_gelu_i
                                                                  ck::Tuple<>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  FastGelu>>>& instances);
 
 void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_km_kn_mn_instances(
@@ -177,7 +177,7 @@ void add_device_grouped_gemm_xdl_fixed_nk_multi_abd_bf16_i8_bf16_km_kn_mn_instan
                                                                  ck::Tuple<>,
                                                                  BF16,
                                                                  PassThrough,
-                                                                 Scales,
+                                                                 Multiply,
                                                                  PassThrough>>>& instances);
 
 // GEMM + Add + Gelu
@@ -199,7 +199,7 @@ struct DeviceOperationInstanceFactory<
                                                                    DsDataType,
                                                                    EDataType,
                                                                    PassThrough,
-                                                                   Scales,
+                                                                   Multiply,
                                                                    AddFastGelu>>
 {
     using DeviceOp = DeviceGroupedGemmMultiABDFixedNK<AsLayout,
@@ -211,7 +211,7 @@ struct DeviceOperationInstanceFactory<
                                                       DsDataType,
                                                       EDataType,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       AddFastGelu>;
 
     static auto GetInstances()
@@ -270,7 +270,7 @@ struct DeviceOperationInstanceFactory<
                                                                    DsDataType,
                                                                    EDataType,
                                                                    PassThrough,
-                                                                   Scales,
+                                                                   Multiply,
                                                                    Add>>
 {
     using DeviceOp = DeviceGroupedGemmMultiABDFixedNK<AsLayout,
@@ -282,7 +282,7 @@ struct DeviceOperationInstanceFactory<
                                                       DsDataType,
                                                       EDataType,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       Add>;
 
     static auto GetInstances()
@@ -341,7 +341,7 @@ struct DeviceOperationInstanceFactory<
                                                                    DsDataType,
                                                                    EDataType,
                                                                    PassThrough,
-                                                                   Scales,
+                                                                   Multiply,
                                                                    FastGelu>>
 {
     using DeviceOp = DeviceGroupedGemmMultiABDFixedNK<AsLayout,
@@ -353,7 +353,7 @@ struct DeviceOperationInstanceFactory<
                                                       DsDataType,
                                                       EDataType,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       FastGelu>;
 
     static auto GetInstances()
@@ -412,7 +412,7 @@ struct DeviceOperationInstanceFactory<
                                                                    DsDataType,
                                                                    EDataType,
                                                                    PassThrough,
-                                                                   Scales,
+                                                                   Multiply,
                                                                    PassThrough>>
 {
     using DeviceOp = DeviceGroupedGemmMultiABDFixedNK<AsLayout,
@@ -424,7 +424,7 @@ struct DeviceOperationInstanceFactory<
                                                       DsDataType,
                                                       EDataType,
                                                       PassThrough,
-                                                      Scales,
+                                                      Multiply,
                                                       PassThrough>;
 
     static auto GetInstances()

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_tile_loop.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#ifdef CK_ENABLE_FP16
+// fp16_output
+void add_device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Empty_Tuple,
+                                                          Row,
+                                                          F16,
+                                                          F16,
+                                                          Empty_Tuple,
+                                                          F16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Col,
+                                                          Empty_Tuple,
+                                                          Row,
+                                                          F16,
+                                                          F16,
+                                                          Empty_Tuple,
+                                                          F16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          PassThrough>>>& instances);
+#endif
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
+                                                            BLayout,
+                                                            Empty_Tuple,
+                                                            ELayout,
+                                                            ADataType,
+                                                            BDataType,
+                                                            Empty_Tuple,
+                                                            EDataType,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            PassThrough>>
+{
+    using DeviceOp = DeviceGroupedGemmTileLoop<ALayout,
+                                               BLayout,
+                                               Empty_Tuple,
+                                               ELayout,
+                                               ADataType,
+                                               BDataType,
+                                               Empty_Tuple,
+                                               EDataType,
+                                               PassThrough,
+                                               PassThrough,
+                                               PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+#ifdef CK_ENABLE_FP16
+        // fp16_output
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_tile_loop_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_tile_loop.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          Multiply>>>& instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
+                                                            BLayout,
+                                                            ck::Tuple<D0Layout>,
+                                                            ELayout,
+                                                            ADataType,
+                                                            BDataType,
+                                                            ck::Tuple<D0DataType>,
+                                                            EDataType,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Multiply>>
+{
+    using DeviceOp = DeviceGroupedGemmTileLoop<ALayout,
+                                               BLayout,
+                                               ck::Tuple<D0Layout>,
+                                               ELayout,
+                                               ADataType,
+                                               BDataType,
+                                               ck::Tuple<D0DataType>,
+                                               EDataType,
+                                               PassThrough,
+                                               PassThrough,
+                                               Multiply>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        // fp16_output
+        if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
+                     is_same_v<EDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn_instances(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyFastGelu>>>& instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
+                                                            BLayout,
+                                                            ck::Tuple<D0Layout>,
+                                                            ELayout,
+                                                            ADataType,
+                                                            BDataType,
+                                                            ck::Tuple<D0DataType>,
+                                                            EDataType,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MultiplyFastGelu>>
+{
+    using DeviceOp = DeviceGroupedGemmTileLoop<ALayout,
+                                               BLayout,
+                                               ck::Tuple<D0Layout>,
+                                               ELayout,
+                                               ADataType,
+                                               BDataType,
+                                               ck::Tuple<D0DataType>,
+                                               EDataType,
+                                               PassThrough,
+                                               PassThrough,
+                                               MultiplyFastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        // fp16_output
+        if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
+                     is_same_v<EDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_tile_loop_multiply_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bias_bf16_i8_bf16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyAdd>>>& instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
+                                                            BLayout,
+                                                            ck::Tuple<D0Layout, D1Layout>,
+                                                            ELayout,
+                                                            ADataType,
+                                                            BDataType,
+                                                            ck::Tuple<D0DataType, D1DataType>,
+                                                            EDataType,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MultiplyAdd>>
+{
+    using DeviceOp = DeviceGroupedGemmTileLoop<ALayout,
+                                               BLayout,
+                                               ck::Tuple<D0Layout, D1Layout>,
+                                               ELayout,
+                                               ADataType,
+                                               BDataType,
+                                               ck::Tuple<D0DataType, D1DataType>,
+                                               EDataType,
+                                               PassThrough,
+                                               PassThrough,
+                                               MultiplyAdd>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        // fp16_output
+        if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
+                     is_same_v<EDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bias_bf16_i8_bf16_mk_kn_mn_instances(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+void add_device_grouped_gemm_xdl_tile_loop_multiply_bias_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmTileLoop<Row,
+                                                          Row,
+                                                          Row_Row_Tuple,
+                                                          Row,
+                                                          BF16,
+                                                          I8,
+                                                          BF16_BF16_Tuple,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyAddFastGelu>>>& instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGroupedGemmTileLoop<ALayout,
+                                                            BLayout,
+                                                            ck::Tuple<D0Layout, D1Layout>,
+                                                            ELayout,
+                                                            ADataType,
+                                                            BDataType,
+                                                            ck::Tuple<D0DataType, D1DataType>,
+                                                            EDataType,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MultiplyAddFastGelu>>
+{
+    using DeviceOp = DeviceGroupedGemmTileLoop<ALayout,
+                                               BLayout,
+                                               ck::Tuple<D0Layout, D1Layout>,
+                                               ELayout,
+                                               ADataType,
+                                               BDataType,
+                                               ck::Tuple<D0DataType, D1DataType>,
+                                               EDataType,
+                                               PassThrough,
+                                               PassThrough,
+                                               MultiplyAddFastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        // fp16_output
+        if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
+                     is_same_v<EDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_tile_loop_multiply_bias_fastgelu_bf16_i8_bf16_mk_kn_mn_instances(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -4,7 +4,8 @@ set(GEMM_MULTI_ABD_INSTANCES)
 list(APPEND GEMM_MULTI_ABD_INSTANCES 
 	device_gemm_xdl_multi_abd_bias_gelu_bf16_i8_bf16_mk_kn_mn_v1_instance.cpp
 	device_gemm_xdl_multi_abd_bias_gelu_bf16_i8_bf16_mk_nk_mn_v1_instance.cpp
-	device_gemm_xdl_multi_abd_bias_gelu_bf16_i8_bf16_km_kn_mn_v1_instance.cpp
+
+	device_gemm_xdl_multi_abd_multiply_bias_gelu_bf16_i8_bf16_mk_kn_mn_v1_instance.cpp
 	)
 
 add_instance_library(device_gemm_multi_abd_instance ${GEMM_MULTI_ABD_INSTANCES})

library/src/tensor_operation_instance/gpu/gemm_multi_abd/device_gemm_xdl_multi_abd_bf16_i8_bf16_km_kn_mn_common.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, 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_gemm_multiple_abd_xdl_cshuffle.hpp"
-
-#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
-
-#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-template <ck::index_t... Is>
-using S = ck::Sequence<Is...>;
-
-using BF16 = ck::bhalf_t;
-using I8   = int8_t;
-using F32  = float;
-
-using Row = ck::tensor_layout::gemm::RowMajor;
-using Col = ck::tensor_layout::gemm::ColumnMajor;
-
-using A0DataType       = BF16;
-using AsDataType       = ck::Tuple<A0DataType>;
-using B0DataType       = I8;
-using B1DataType       = BF16;
-using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
-using AccDataType      = F32;
-using CShuffleDataType = BF16;
-using D0DataType       = BF16;
-// using DsDataType       = ck::Tuple<D0DataType>;
-using EDataType = BF16;
-
-using A0Layout = Col;
-using AsLayout = ck::Tuple<A0Layout>;
-using B0Layout = Row;
-using B1Layout = B0Layout;
-using BsLayout = ck::Tuple<B0Layout, B1Layout>;
-using D0Layout = Row;
-// using DsLayout = ck::Tuple<D0Layout>;
-using ELayout = Row;
-
-using Scales      = ck::tensor_operation::element_wise::Scales;
-using PassThrough = ck::tensor_operation::element_wise::PassThrough;
-using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
-using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
-using Add         = ck::tensor_operation::element_wise::Add;
-
-using AElementOp = PassThrough;
-using BElementOp = Scales;
-// using CDEElementOp = AddFastGelu;
-
-static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
-static constexpr auto GemmMNPadding  = ck::tensor_operation::device::GemmSpecialization::MNPadding;
-static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
-
-// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
-template <typename DsLayout,
-          typename DsDataType,
-          typename CDEElementOp,
-          ck::tensor_operation::device::GemmSpecialization GemmSpec,
-          ck::PipelineVersion PipVer,
-          ck::LoopScheduler LoopSche>
-using device_gemm_xdl_multi_abd_bf16_i8_bf16_km_kn_mn_instances = std::tuple<
-    // clang-format off
-        //###############################|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer| K0Per| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
-        //###############################|         |         |         |        |       Type|       Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
-        //###############################|         |         |         |        |           |           |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
-        //###############################|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
-        //PipelineVersion::v1
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 48, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 24, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    32,    32,    32,   8,   8,   32,   32,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    16,    32,    32,   8,   8,   16,   16,    1,    2,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               4, LoopSche, PipVer>
-    // clang-format on
-    >;
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multi_abd/device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_common.hpp

@@ -27,13 +27,13 @@ using F32  = float;
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
 
-using A0DataType       = BF16;
-using AsDataType       = ck::Tuple<A0DataType>;
-using B0DataType       = I8;
-using B1DataType       = BF16;
-using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using A0DataType = BF16;
+using AsDataType = ck::Tuple<A0DataType>;
+using B0DataType = I8;
+using B1DataType = BF16;
+// using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
 using AccDataType      = F32;
-using CShuffleDataType = BF16;
+using CShuffleDataType = F32;
 using D0DataType       = BF16;
 // using DsDataType       = ck::Tuple<D0DataType>;
 using EDataType = BF16;
@@ -42,57 +42,84 @@ using A0Layout = Row;
 using AsLayout = ck::Tuple<A0Layout>;
 using B0Layout = Row;
 using B1Layout = B0Layout;
-using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+// using BsLayout = ck::Tuple<B0Layout, B1Layout>;
 using D0Layout = Row;
 // using DsLayout = ck::Tuple<D0Layout>;
 using ELayout = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply            = ck::tensor_operation::element_wise::Multiply;
+using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
+using MultiplyFastGelu    = ck::tensor_operation::element_wise::MultiplyFastGelu;
+using MultiplyAdd         = ck::tensor_operation::element_wise::MultiplyAdd;
+
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
 using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
 using Add         = ck::tensor_operation::element_wise::Add;
 
 using AElementOp = PassThrough;
-using BElementOp = Scales;
+// using BElementOp = Multiply;
 // using CDEElementOp = AddFastGelu;
 
+static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
+static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
+
 static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
 static constexpr auto GemmMNPadding  = ck::tensor_operation::device::GemmSpecialization::MNPadding;
 static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
 
 // Compilation parameters for a[m, k] * b[k, n] = c[m, n]
-template <typename DsLayout,
+template <typename BsLayout,
+          typename DsLayout,
+          typename BsDataType,
+          typename DsDataType,
+          typename BElementOp,
+          typename CDEElementOp,
+          ck::tensor_operation::device::GemmSpecialization GemmSpec,
+          BlockGemmPipelineScheduler BlkGemmPipeSched>
+using device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_comp_instances = std::tuple<
+    // clang-format off
+        //###############################|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer| K0Per| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //###############################|         |         |         |        |       Type|       Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //###############################|         |         |         |        |           |           |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###############################|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   224,   256,    64,   8,   4,  16,   16,    7,    8,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           2,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   256,    32,   8,   4,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+    // clang-format on
+    >;
+
+template <typename BsLayout,
+          typename DsLayout,
+          typename BsDataType,
           typename DsDataType,
+          typename BElementOp,
           typename CDEElementOp,
           ck::tensor_operation::device::GemmSpecialization GemmSpec,
-          ck::PipelineVersion PipVer,
-          ck::LoopScheduler LoopSche>
-using device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_instances = std::tuple<
+          BlockGemmPipelineScheduler BlkGemmPipeSched>
+using device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_mem_instances = std::tuple<
     // clang-format off
         //###############################|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer| K0Per| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
         //###############################|         |         |         |        |       Type|       Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
         //###############################|         |         |         |        |           |           |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
         //###############################|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
-        //PipelineVersion::v1
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 48, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 24, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    32,    32,    32,   8,   8,   32,   32,    1,    1,     S<2, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    16,    32,    32,   8,   8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               4, LoopSche, PipVer>
+        // Latency friendly
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    16,   16,   256,   8,   4,  16,   16,    1,    1,     S<32, 2, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 1, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,   32,   256,   8,   4,  16,   16,    1,    1,     S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 2, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        // Memory friendly
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    16,   16,   256,   8,   4,  16,   16,    1,    1,     S<32, 2, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 1, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,   32,   256,   8,   4,  16,   16,    1,    1,     S<32, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<64, 2, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,   64,   128,   8,   4,  16,   16,    1,    2,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,   64,   128,   8,   4,  32,   32,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<32, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,  128,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,  128,    64,   8,   4,  32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,    16,  256,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,    32,  256,    64,   8,   4,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              16,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
     // clang-format on
     >;
 } // namespace instance

library/src/tensor_operation_instance/gpu/gemm_multi_abd/device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_common.hpp

@@ -47,14 +47,18 @@ using D0Layout = Row;
 // using DsLayout = ck::Tuple<D0Layout>;
 using ELayout = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
 using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
 using Add         = ck::tensor_operation::element_wise::Add;
 
 using AElementOp = PassThrough;
-using BElementOp = Scales;
+using BElementOp = Multiply;
+
+static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
+static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
+
 // using CDEElementOp = AddFastGelu;
 
 static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
@@ -66,33 +70,52 @@ template <typename DsLayout,
           typename DsDataType,
           typename CDEElementOp,
           ck::tensor_operation::device::GemmSpecialization GemmSpec,
-          ck::PipelineVersion PipVer,
-          ck::LoopScheduler LoopSche>
-using device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_instances = std::tuple<
+          BlockGemmPipelineScheduler BlkGemmPipeSched>
+using device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_comp_instances = std::tuple<
+    // clang-format off
+        //###############################|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer| K0Per| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //###############################|         |         |         |        |       Type|       Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //###############################|         |         |         |        |           |           |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###############################|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+	//Compute-bound
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,  16,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,            16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,                8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   224,   256,    64,   8,  16,  16,   16,    7,    8,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,            16,             16,          0,          1,           2,                   S<1, 32, 1, 8>,                8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,  16,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,            16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,                8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,  16,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,            16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,                8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,  16,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,            16,             16,          0,          1,           1,                   S<1, 32, 1, 8>,                8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+    // clang-format on
+    >;
+template <typename DsLayout,
+          typename DsDataType,
+          typename CDEElementOp,
+          ck::tensor_operation::device::GemmSpecialization GemmSpec,
+          BlockGemmPipelineScheduler BlkGemmPipeSched>
+using device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_mem_instances = std::tuple<
     // clang-format off
         //###############################|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer| K0Per| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
         //###############################|         |         |         |        |       Type|       Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
         //###############################|         |         |         |        |           |           |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
         //###############################|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
-        //PipelineVersion::v1
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 48, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 24, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        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,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    32,    32,    32,   8,   8,   32,   32,    1,    1,     S<2, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8, LoopSche, PipVer>,
-       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    16,    32,    32,   8,   8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               4, LoopSche, PipVer>
+        // Latency friendly 
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,   16,   128,   8,   16,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    16,   16,   128,   8,   16,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 8, 1>,      S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    16,   16,   128,   8,   16,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 8, 1>,      S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,   32,   128,   8,   16,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
+        // Memory friendly
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,   128,   32,   128,   8,   16,  32,   32,    2,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,   128,   16,   128,   8,   16,  16,   16,    4,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    64,   32,   128,   8,   16,  32,   32,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    64,   16,   128,   8,   16,  16,   16,    2,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,   16,   128,   8,   16,  16,   16,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    16,   16,   128,   8,   16,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 8, 1>,      S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,    64,    16,   16,   128,   8,   16,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 8, 1>,      S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,   32,   128,   8,   16,  16,   16,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,   64,   128,   8,   16,  16,   16,    1,    2,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,   64,   128,   8,   16,  32,   32,    1,    1,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,  128,   128,   8,   16,  16,   16,    1,    4,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    32,  128,   128,   8,   16,  32,   32,    1,    2,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,    16,  256,   128,   8,   16,  16,   16,    1,    4,     S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+       DeviceGemmMultipleABD_Xdl_CShuffle< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,    32,  256,   128,   8,   16,  32,   32,    1,    2,     S<16, 16, 1>,    S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,          1,           1,                   S<1, 16, 1, 16>,              8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
     // clang-format on
     >;
 } // namespace instance

library/src/tensor_operation_instance/gpu/gemm_multi_abd/device_gemm_xdl_multi_abd_bias_gelu_bf16_i8_bf16_mk_kn_mn_v1_instance.cpp

@@ -19,94 +19,143 @@ namespace instance {
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_bias_gelu_v1_instances(
     std::vector<std::unique_ptr<DeviceGemmMultipleABD<AsLayout,
-                                                      BsLayout,
+                                                      ck::Tuple<B0Layout, B1Layout>,
                                                       ck::Tuple<D0Layout>,
                                                       ELayout,
                                                       AsDataType,
-                                                      BsDataType,
+                                                      ck::Tuple<B0DataType, B1DataType>,
                                                       ck::Tuple<D0DataType>,
                                                       EDataType,
                                                       AElementOp,
-                                                      BElementOp,
+                                                      Multiply,
                                                       AddFastGelu>>>& instances)
 {
-    add_device_operation_instances(
-        instances,
-        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_instances<ck::Tuple<D0Layout>,
-                                                                  ck::Tuple<D0DataType>,
-                                                                  AddFastGelu,
-                                                                  GemmMNKPadding,
-                                                                  PipelineVersion::v1,
-                                                                  LoopScheduler::Default>{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_comp_instances<
+                                       ck::Tuple<B0Layout, B1Layout>,
+                                       ck::Tuple<D0Layout>,
+                                       ck::Tuple<B0DataType, B1DataType>,
+                                       ck::Tuple<D0DataType>,
+                                       Multiply,
+                                       AddFastGelu,
+                                       GemmMNKPadding,
+                                       Interwave>{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_mem_instances<
+                                       ck::Tuple<B0Layout, B1Layout>,
+                                       ck::Tuple<D0Layout>,
+                                       ck::Tuple<B0DataType, B1DataType>,
+                                       ck::Tuple<D0DataType>,
+                                       Multiply,
+                                       AddFastGelu,
+                                       GemmMNKPadding,
+                                       Interwave>{});
 }
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_bias_v1_instances(
     std::vector<std::unique_ptr<DeviceGemmMultipleABD<AsLayout,
-                                                      BsLayout,
+                                                      ck::Tuple<B0Layout, B1Layout>,
                                                       ck::Tuple<D0Layout>,
                                                       ELayout,
                                                       AsDataType,
-                                                      BsDataType,
+                                                      ck::Tuple<B0DataType, B1DataType>,
                                                       ck::Tuple<D0DataType>,
                                                       EDataType,
                                                       AElementOp,
-                                                      BElementOp,
+                                                      Multiply,
                                                       Add>>>& instances)
 {
-    add_device_operation_instances(
-        instances,
-        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_instances<ck::Tuple<D0Layout>,
-                                                                  ck::Tuple<D0DataType>,
-                                                                  Add,
-                                                                  GemmMNKPadding,
-                                                                  PipelineVersion::v1,
-                                                                  LoopScheduler::Default>{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_comp_instances<
+                                       ck::Tuple<B0Layout, B1Layout>,
+                                       ck::Tuple<D0Layout>,
+                                       ck::Tuple<B0DataType, B1DataType>,
+                                       ck::Tuple<D0DataType>,
+                                       Multiply,
+                                       Add,
+                                       GemmMNKPadding,
+                                       Interwave>{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_mem_instances<
+                                       ck::Tuple<B0Layout, B1Layout>,
+                                       ck::Tuple<D0Layout>,
+                                       ck::Tuple<B0DataType, B1DataType>,
+                                       ck::Tuple<D0DataType>,
+                                       Multiply,
+                                       Add,
+                                       GemmMNKPadding,
+                                       Interwave>{});
 }
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_v1_instances(
     std::vector<std::unique_ptr<DeviceGemmMultipleABD<AsLayout,
-                                                      BsLayout,
+                                                      ck::Tuple<B0Layout, B1Layout>,
                                                       ck::Tuple<>,
                                                       ELayout,
                                                       AsDataType,
-                                                      BsDataType,
+                                                      ck::Tuple<B0DataType, B1DataType>,
                                                       ck::Tuple<>,
                                                       EDataType,
                                                       AElementOp,
-                                                      BElementOp,
+                                                      Multiply,
                                                       PassThrough>>>& instances)
 {
-    add_device_operation_instances(
-        instances,
-        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_instances<ck::Tuple<>,
-                                                                  ck::Tuple<>,
-                                                                  PassThrough,
-                                                                  GemmMNKPadding,
-                                                                  PipelineVersion::v1,
-                                                                  LoopScheduler::Default>{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_comp_instances<
+                                       ck::Tuple<B0Layout, B1Layout>,
+                                       ck::Tuple<>,
+                                       ck::Tuple<B0DataType, B1DataType>,
+                                       ck::Tuple<>,
+                                       Multiply,
+                                       PassThrough,
+                                       GemmMNKPadding,
+                                       Interwave>{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_mem_instances<
+                                       ck::Tuple<B0Layout, B1Layout>,
+                                       ck::Tuple<>,
+                                       ck::Tuple<B0DataType, B1DataType>,
+                                       ck::Tuple<>,
+                                       Multiply,
+                                       PassThrough,
+                                       GemmMNKPadding,
+                                       Interwave>{});
 }
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_gelu_v1_instances(
     std::vector<std::unique_ptr<DeviceGemmMultipleABD<AsLayout,
-                                                      BsLayout,
+                                                      ck::Tuple<B0Layout, B1Layout>,
                                                       ck::Tuple<>,
                                                       ELayout,
                                                       AsDataType,
-                                                      BsDataType,
+                                                      ck::Tuple<B0DataType, B1DataType>,
                                                       ck::Tuple<>,
                                                       EDataType,
                                                       AElementOp,
-                                                      BElementOp,
+                                                      Multiply,
                                                       FastGelu>>>& instances)
 {
-    add_device_operation_instances(
-        instances,
-        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_instances<ck::Tuple<>,
-                                                                  ck::Tuple<>,
-                                                                  FastGelu,
-                                                                  GemmMNKPadding,
-                                                                  PipelineVersion::v1,
-                                                                  LoopScheduler::Default>{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_comp_instances<
+                                       ck::Tuple<B0Layout, B1Layout>,
+                                       ck::Tuple<>,
+                                       ck::Tuple<B0DataType, B1DataType>,
+                                       ck::Tuple<>,
+                                       Multiply,
+                                       FastGelu,
+                                       GemmMNKPadding,
+                                       Interwave>{});
+
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_kn_mn_mem_instances<
+                                       ck::Tuple<B0Layout, B1Layout>,
+                                       ck::Tuple<>,
+                                       ck::Tuple<B0DataType, B1DataType>,
+                                       ck::Tuple<>,
+                                       Multiply,
+                                       FastGelu,
+                                       GemmMNKPadding,
+                                       Interwave>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/gemm_multi_abd/device_gemm_xdl_multi_abd_bias_gelu_bf16_i8_bf16_mk_nk_mn_v1_instance.cpp

@@ -32,12 +32,11 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_bias_gelu_v1_instances(
 {
     add_device_operation_instances(
         instances,
-        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_instances<ck::Tuple<D0Layout>,
-                                                                  ck::Tuple<D0DataType>,
-                                                                  AddFastGelu,
-                                                                  GemmMNKPadding,
-                                                                  PipelineVersion::v1,
-                                                                  LoopScheduler::Default>{});
+        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_comp_instances<ck::Tuple<D0Layout>,
+                                                                       ck::Tuple<D0DataType>,
+                                                                       AddFastGelu,
+                                                                       GemmMNKPadding,
+                                                                       Interwave>{});
 }
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_bias_v1_instances(
@@ -55,12 +54,11 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_bias_v1_instances(
 {
     add_device_operation_instances(
         instances,
-        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_instances<ck::Tuple<D0Layout>,
-                                                                  ck::Tuple<D0DataType>,
-                                                                  Add,
-                                                                  GemmMNKPadding,
-                                                                  PipelineVersion::v1,
-                                                                  LoopScheduler::Default>{});
+        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_comp_instances<ck::Tuple<D0Layout>,
+                                                                       ck::Tuple<D0DataType>,
+                                                                       Add,
+                                                                       GemmMNKPadding,
+                                                                       Interwave>{});
 }
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_v1_instances(
@@ -78,12 +76,11 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_v1_instances(
 {
     add_device_operation_instances(
         instances,
-        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_instances<ck::Tuple<>,
-                                                                  ck::Tuple<>,
-                                                                  PassThrough,
-                                                                  GemmMNKPadding,
-                                                                  PipelineVersion::v1,
-                                                                  LoopScheduler::Default>{});
+        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_comp_instances<ck::Tuple<>,
+                                                                       ck::Tuple<>,
+                                                                       PassThrough,
+                                                                       GemmMNKPadding,
+                                                                       Interwave>{});
 }
 
 void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_gelu_v1_instances(
@@ -101,12 +98,11 @@ void add_device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_gelu_v1_instances(
 {
     add_device_operation_instances(
         instances,
-        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_instances<ck::Tuple<>,
-                                                                  ck::Tuple<>,
-                                                                  FastGelu,
-                                                                  GemmMNKPadding,
-                                                                  PipelineVersion::v1,
-                                                                  LoopScheduler::Default>{});
+        device_gemm_xdl_multi_abd_bf16_i8_bf16_mk_nk_mn_comp_instances<ck::Tuple<>,
+                                                                       ck::Tuple<>,
+                                                                       FastGelu,
+                                                                       GemmMNKPadding,
+                                                                       Interwave>{});
 }
 
 } // namespace instance