Fp16/fp8 mixed-precision Gemm with multiply+add fusion (#865)

* add compute_type

* add multiply_add ckProfiler

* add f8_fp16 support

* clean

* clean

* fixed lds size calc

* format

---------
Co-authored-by: Jing Zhang <jizha@amd.com>

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

@@ -543,9 +543,13 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
         EGridDesc_G_M_N e_grid_desc_g_m_n_;
     };
 
+    using ComputeDataType = ADataType;
+
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
-        ADataType, // TODO: distinguish A/B datatype
+        ADataType,
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,

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

@@ -331,8 +331,13 @@ struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<ALayout,
         EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
     };
 
+    using ComputeDataType = ADataType;
+
+    // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
-        ADataType, // TODO: distinguish A/B datatype
+        ADataType,
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         ck::Tuple<>, // DsDataType,

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

@@ -324,8 +324,12 @@ struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
         index_t BatchStrideE_;
     };
 
+    using ComputeDataType = ADataType;
+
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,

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

@@ -310,9 +310,13 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
     using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
     using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
 
+    using ComputeDataType = ADataType;
+
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,

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

@@ -20,7 +20,8 @@
 namespace ck {
 
 template <typename GridwiseGemm,
-          typename ABDataType,
+          typename ADataType,
+          typename BDataType,
           typename DsPointer,
           typename EDataType,
           typename AElementwiseOperation,
@@ -36,8 +37,8 @@ __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_gemm_multiple_d_xdl_cshuffle(const ABDataType* __restrict__ p_a_grid,
-                                            const ABDataType* __restrict__ p_b_grid,
+        kernel_gemm_multiple_d_xdl_cshuffle(const ADataType* __restrict__ p_a_grid,
+                                            const BDataType* __restrict__ p_b_grid,
                                             DsPointer p_ds_grid,
                                             EDataType* __restrict__ p_e_grid,
                                             const AElementwiseOperation a_element_op,
@@ -242,9 +243,13 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
     using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
     using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
 
+    using ComputeDataType = EDataType;
+
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,
@@ -442,6 +447,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
                 const auto kernel = kernel_gemm_multiple_d_xdl_cshuffle<
                     GridwiseGemm,
                     ADataType, // TODO: distiguish A/B datatype
+                    BDataType, // TODO: distiguish A/B datatype
                     typename GridwiseGemm::DsGridPointer,
                     EDataType,
                     AElementwiseOperation,

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

@@ -355,9 +355,13 @@ struct DeviceGroupedContractionMultipleD_Xdl_CShuffle
     using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
     using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
 
+    using ComputeDataType = ADataType;
+
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp

@@ -355,6 +355,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
 
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ABDataType, // TODO: distinguish A/B datatype
+        ABDataType, // TODO: distinguish A/B datatype
         ABDataType, // TODO: distinguish A/B datatype
         AccDataType,
         CShuffleDataType,

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

@@ -367,9 +367,13 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
     using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
     using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
 
+    using ComputeDataType = ADataType;
+
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,

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

@@ -228,9 +228,13 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
     using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
     using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
 
+    using ComputeDataType = ADataType;
+
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,

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

@@ -195,6 +195,51 @@ struct AddMultiply
     }
 };
 
+// C = A * B
+// E = C x D0 + D1
+struct MultiplyAdd
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ void operator()<half_t, half_t, half_t, half_t>(half_t& e,
+                                                                        const half_t& c,
+                                                                        const half_t& d0,
+                                                                        const half_t& d1) const
+    {
+        const half_t y = (c * d0) + d1;
+        e              = y;
+    }
+    template <>
+    __host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
+                                                                       const float& c,
+                                                                       const half_t& d0,
+                                                                       const half_t& d1) const
+    {
+        const half_t y = type_convert<half_t>(c) * d0 + d1;
+        e              = y;
+    }
+    template <>
+    __host__ __device__ void operator()<float, float, half_t, half_t>(float& e,
+                                                                      const float& c,
+                                                                      const half_t& d0,
+                                                                      const half_t& d1) const
+    {
+        const float y = c * d0 + d1;
+        e             = y;
+    }
+    template <>
+    __host__ __device__ void operator()<half_t, float, float, float>(half_t& e,
+                                                                     const float& c,
+                                                                     const float& d0,
+                                                                     const float& d1) const
+    {
+        const float y = c * d0 + d1;
+        e             = y;
+    }
+};
+
 // E = FastGelu(C + D0 + D1)
 struct AddAddFastGelu
 {

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

@@ -26,7 +26,9 @@ namespace ck {
 //   E = cde_op(C, D0, D1, ...)
 // Assume:
 //   D0, D1, ... and E have the same layout
-template <typename ABDataType, // FIXME: don't assume A/B have same datatype
+template <typename ADataType,
+          typename BDataType,
+          typename ComputeDataType_,
           typename AccDataType,
           typename CShuffleDataType,
           typename DsDataType,
@@ -92,15 +94,11 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
     using GridwiseGemmPipe = remove_cvref_t<
         decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
-    // denorm test fix, required to work around fp16 mfma issue
-    // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
-    // when mfma if fixed, remove this section and update
-    // ABDataTypeAdjusted -> ABDataType throughout this file
 #if CK_WORKAROUND_DENORM_FIX
-    using ABDataTypeAdjusted =
-        conditional_t<is_same_v<ABDataType, ck::half_t>, ck::bhalf_t, ABDataType>;
+    using ComputeDataType =
+        conditional_t<is_same_v<ComputeDataType_, ck::half_t>, ck::bhalf_t, ComputeDataType_>;
 #else
-    using ABDataTypeAdjusted = ABDataType;
+    using ComputeDataType = ComputeDataType_;
 #endif
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
@@ -170,7 +168,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
             c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
 
         return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
-                             sizeof(ABDataType),
+                             sizeof(ComputeDataType),
                          c_block_size * sizeof(CShuffleDataType));
     }
 
@@ -313,8 +311,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
         // check tensor size: cannot be larger than 2GB each
         constexpr long_index_t TwoGB = (long_index_t{1} << 31);
 
-        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
-             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
              e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
         {
             return false;
@@ -338,8 +336,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
               typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
               typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
               typename Block2ETileMap>
-    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
-                               const ABDataType* __restrict__ p_b_grid,
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
                                DsGridPointer p_ds_grid,
                                EDataType* __restrict__ p_e_grid,
                                void* __restrict__ p_shared,
@@ -408,8 +406,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
                                                 Sequence<AK0PerBlock, MPerBlock, AK1>,
                                                 ABlockTransferThreadClusterLengths_AK0_M_AK1,
                                                 ABlockTransferThreadClusterArrangeOrder,
-                                                ABDataType,
-                                                ABDataTypeAdjusted,
+                                                ADataType,
+                                                ComputeDataType,
                                                 decltype(a_grid_desc_ak0_m_ak1),
                                                 decltype(a_block_desc_ak0_m_ak1),
                                                 ABlockTransferSrcAccessOrder,
@@ -439,8 +437,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
                                                 Sequence<BK0PerBlock, NPerBlock, BK1>,
                                                 BBlockTransferThreadClusterLengths_BK0_N_BK1,
                                                 BBlockTransferThreadClusterArrangeOrder,
-                                                ABDataType,
-                                                ABDataTypeAdjusted,
+                                                BDataType,
+                                                ComputeDataType,
                                                 decltype(b_grid_desc_bk0_n_bk1),
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
@@ -470,11 +468,11 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
         // sanity check
         constexpr index_t KPack =
             math::max(math::lcm(AK1, BK1),
-                      MfmaSelector<ABDataTypeAdjusted, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+                      MfmaSelector<ComputeDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            ABDataTypeAdjusted,
+            ComputeDataType,
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -492,11 +490,10 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
             a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ABDataTypeAdjusted*>(p_shared),
-            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+            static_cast<ComputeDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<ABDataTypeAdjusted*>(p_shared) + a_block_space_size_aligned,
+            static_cast<ComputeDataType*>(p_shared) + a_block_space_size_aligned,
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);

include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.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/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ADataType, // FIXME: don't assume A/B have same datatype
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          index_t NumGemmKPrefetchStage,
+          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_KBatch_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_KBatch_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 CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       AK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       BK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __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;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_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(AK1, BK1);
+
+        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(ADataType) +
+                             b_block_space_size_aligned * sizeof(BDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch)
+    {
+        return math::integer_least_multiple(K, KPerBlock * K_Batch);
+    }
+
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_KBatch_AK0_M_AK1(index_t M, index_t K, index_t StrideA, index_t KBatch)
+    {
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto MPad = CalculateMPadded(M);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto AK0 = KPad / (KBatch * AK1);
+
+        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)
+        {
+            // const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_KBatch_BK0_N_BK1(index_t K, index_t N, index_t StrideB, index_t KBatch)
+    {
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto NPad = CalculateNPadded(N);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto BK0 = KPad / (KBatch * BK1);
+
+        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)
+        {
+            // const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    template <typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              GemmSpecialization GemmSpec>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const index_t M,
+                  const index_t N,
+                  const index_t K,
+                  const index_t StrideA,
+                  const index_t StrideB,
+                  const std::array<index_t, NumDTensor> StrideDs,
+                  const index_t StrideE,
+                  const index_t KBatch)
+    {
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        ignore = StrideDs;
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+#if 0
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+#endif
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              index_t NumDTensor_,
+              typename DsDataType_,
+              typename AGridDesc_KBatch_AK0_M_AK1,
+              typename BGridDesc_KBatch_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename CDEElementwiseOperation_,
+              typename Block2ETileMap>
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const index_t KBatch,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation_& cde_element_op,
+                               const AGridDesc_KBatch_AK0_M_AK1& a_grid_desc_kbatch_ak0_m_ak1,
+                               const BGridDesc_KBatch_BK0_N_BK1& b_grid_desc_kbatch_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor_>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_kbatch_ak0_m_ak1 =
+            GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_kbatch_bk0_n_bk1 =
+            GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ComputeType,
+                                                decltype(a_grid_desc_kbatch_ak0_m_ak1),
+                                                decltype(a_block_desc_kbatch_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_kbatch_ak0_m_ak1,
+                make_multi_index(kbatch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_kbatch_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BDataType,
+                                                ComputeType,
+                                                decltype(b_grid_desc_kbatch_bk0_n_bk1),
+                                                decltype(b_block_desc_kbatch_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_kbatch_bk0_n_bk1,
+                make_multi_index(kbatch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_kbatch_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // 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();
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ComputeType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+#if 1
+        if(block_work_idx[I0] == 0)
+        {
+            const index_t nThreadSize = CDEShuffleBlockTransferScalarPerVector_NPerBlock;
+            const index_t numNThreads = NPerBlock / nThreadSize;
+            const index_t numMThreads = BlockSize / numNThreads;
+            const index_t mThreadSize = MPerBlock / numMThreads;
+
+            const index_t m_tid = get_thread_local_1d_id() / numNThreads;
+            const index_t n_tid = get_thread_local_1d_id() % numNThreads;
+
+            auto c_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mThreadSize>{}, I1, Number<nThreadSize>{}));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr,
+                         EDataType,
+                         c_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize(),
+                         true>
+                e_thread_zero_buf;
+
+            auto c_thread_copy = ThreadwiseTensorSliceTransfer_v1r3<
+                EDataType,
+                EDataType,
+                decltype(c_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1, mThreadSize, 1, nThreadSize>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{e_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I1],
+                                       m_tid * mThreadSize,
+                                       block_work_idx[I2],
+                                       n_tid * nThreadSize),
+                      ck::tensor_operation::element_wise::PassThrough{}};
+
+            c_thread_copy.Run(c_thread_desc_mblock_mperblock_nblock_nperblock,
+                              make_tuple(I0, I0, I0, I0),
+                              e_thread_zero_buf,
+                              e_grid_desc_mblock_mperblock_nblock_nperblock,
+                              e_grid_buf);
+
+            __syncthreads();
+
+            if(threadIdx.x == 0)
+            {
+                atomicAdd(barrier_count_finished, 1);
+            }
+        }
+#endif
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop =
+            __builtin_amdgcn_readfirstlane((a_grid_desc_kbatch_ak0_m_ak1.GetLength(I1) *
+                                            a_grid_desc_kbatch_ak0_m_ak1.GetLength(I3)) /
+                                           KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_kbatch_ak0_m_ak1,
+                                                               a_block_desc_kbatch_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_kbatch_bk0_n_bk1,
+                                                               b_block_desc_kbatch_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            if(threadIdx.x == 0)
+            {
+                while(__atomic_load_n(barrier_count_finished, __ATOMIC_RELAXED) == 0) {}
+            }
+
+            __syncthreads();
+
+            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.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.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.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{}};
+
+            // 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[I1], 0, block_work_idx[I2], 0);
+                    },
+                    Number<NumDTensor_>{}));
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            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/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>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType_{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation_,
+                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 DimAccessOrder,
+                3,                    // index_t VectorDim,
+                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[I1], 0, block_work_idx[I2], 0)),
+                 cde_element_op};
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_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);
+                }
+            });
+
+            if(threadIdx.x == 0)
+            {
+                index_t k_id_finished_t = atomicAdd(barrier_count_finished, 1);
+
+                if(k_id_finished_t == KBatch)
+                {
+                    *barrier_count_finished = 0;
+                }
+            }
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const index_t KBatch,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ADataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const BDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        if(kbatch_id == KBatch - 1)
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, NumDTensor, DsDataType>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                cde_element_op,
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+        else
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, 0, Tuple<>>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                ck::tensor_operation::element_wise::PassThrough{},
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+    }
+};
+
+} // namespace ck

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

@@ -32,6 +32,8 @@ using F32_Tuple     = ck::Tuple<F32>;
 using I32_Tuple     = ck::Tuple<I32>;
 using I32_F32_Tuple = ck::Tuple<I32, F32>;
 
+using F32_F32_Tuple = ck::Tuple<F32, F32>;
+
 // GEMM layout
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -95,6 +97,7 @@ using AddFastGelu    = ck::tensor_operation::element_wise::AddFastGelu;
 using AddReluAdd     = ck::tensor_operation::element_wise::AddReluAdd;
 using FastGelu       = ck::tensor_operation::element_wise::FastGelu;
 using AddMultiply    = ck::tensor_operation::element_wise::AddMultiply;
+using 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;

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+// GEMM + Multiply + Add
+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::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::MultiplyAdd>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout, D1Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType, D1DataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::MultiplyAdd>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
+                     is_same_v<D0DataType, float> && is_same_v<D1DataType, float> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+add_instance_library(device_gemm_multiply_add_instance
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
+
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
+)

library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp

+// 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_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // M/N/K padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp

+// 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_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // M/N/K padding
+        // N % 8 == 0 && K % 1 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>
+
+        // clang-format on
+        >;
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp

+// 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_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8        = ck::f8_t;
+using F16       = ck::half_t;
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32, F32>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // M/N/K padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp

+// 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_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8        = ck::f8_t;
+using F16       = ck::half_t;
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32, F32>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // M/N/K padding
+        // N % 8 == 0 && K % 1 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>
+
+        // clang-format on
+        >;
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/include/profiler/profile_gemm_multiply_add_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout>
+bool profile_gemm_multiply_add_impl(int do_verification,
+                                    int init_method,
+                                    bool /*do_log*/,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int StrideA,
+                                    int StrideB,
+                                    int StrideD0,
+                                    int StrideD1,
+                                    int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = MultiplyAdd;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                          BLayout,
+                                                          ck::Tuple<D0Layout, D1Layout>,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          ck::Tuple<D0DataType, D1DataType>,
+                                                          EDataType,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck