Implement direct loads split-K GEMM kernel (#1137)

* WIP: Implement direct loads split-K GEMM kernel

* Clean the review

---------
Co-authored-by: Adam Osewski <19374865+aosewski@users.noreply.github.com>
Co-authored-by: Bartłomiej Kocot <barkocot@amd.com>

example/35_splitK_gemm/CMakeLists.txt

@@ -10,6 +10,9 @@ foreach(gpu IN LISTS GPU_TARGETS)
    add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
    add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
 
+   add_example_executable(example_splitK_gemm_xdl_lds_direct_load_fp16 splitK_gemm_xdl_lds_direct_load_fp16.cpp)
+   add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_lds_direct_load_fp16)
+
    add_example_executable(example_splitK_gemm_xdl_bf16 splitK_gemm_xdl_bf16.cpp)
    add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bf16)
 

example/35_splitK_gemm/splitK_gemm_xdl_lds_direct_load_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+#define DIRECT_LOAD 1
+
+#if DIRECT_LOAD
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp"
+#else
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#endif
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using CDataType   = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+#if DIRECT_LOAD
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle_LdsDirectLoad
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|         ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|         BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|          ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar| AddExtraM|          ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |     |     | Wave| Wave| Lengths_KBatch_K0_M_K1|               |               |      PerVector|          | Lengths_KBatch_K0_N_K1|               |              |      PerVector|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |         |      |      |      |      |    |     |     |     |     |                       |               |               |               |          |                       |               |              |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        2,   128,    32,    16,     4,  16,   16,   16,    1,    1,         S<1, 2, 8, 8>,  S<0, 2, 1, 3>,              3,              2,      true,         S<1, 2, 8, 8>,  S<0, 2, 1, 3>,             3,              2,      true,           1,           1,                   S<1, 32, 1, 4>,               4>;
+// clang-format on
+
+#else
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+using DeviceGemmInstance          = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#endif
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_splitk_lds_direct_load.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t NumGemmKPrefetchStage,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferScalarPerVector,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferScalarPerVector,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
+          typename ComputeType        = CDataType,
+          PipelineVersion PipelineVer = PipelineVersion::v4,
+          LoopScheduler LoopSched     = make_default_loop_scheduler()>
+
+struct DeviceGemmXdlSplitKCShuffle_LdsDirectLoad : public DeviceGemmSplitK<ALayout,
+                                                                           BLayout,
+                                                                           CLayout,
+                                                                           ADataType,
+                                                                           BDataType,
+                                                                           CDataType,
+                                                                           AElementwiseOperation,
+                                                                           BElementwiseOperation,
+                                                                           CElementwiseOperation,
+                                                                           ComputeType>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using GridwiseGemm = GridwiseGemm_xdlops_splitk_lds_direct_load<
+        BlockSize,
+        ADataType,
+        BDataType,
+        AccDataType,
+        CDataType,
+        ALayout,
+        BLayout,
+        CLayout,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        NumGemmKPrefetchStage,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferScalarPerVector,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferScalarPerVector,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXDL,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        LoopSched,
+        PipelineVer,
+        ComputeType>;
+
+    struct Argument : public GridwiseGemm::Argument
+    {
+        Argument(const ADataType* p_a_grid_,
+                 const BDataType* p_b_grid_,
+                 CDataType* p_c_grid_,
+                 index_t M_,
+                 index_t N_,
+                 index_t K_,
+                 index_t StrideA_,
+                 index_t StrideB_,
+                 index_t StrideC_,
+                 index_t MPadded_,
+                 index_t NPadded_,
+                 index_t KPadded_,
+                 index_t K0Padded_,
+                 index_t k_batch_,
+                 AElementwiseOperation a_element_op_,
+                 BElementwiseOperation b_element_op_,
+                 CElementwiseOperation c_element_op_)
+            : GridwiseGemm::Argument(p_a_grid_,
+                                     p_b_grid_,
+                                     p_c_grid_,
+                                     M_,
+                                     N_,
+                                     K_,
+                                     StrideA_,
+                                     StrideB_,
+                                     StrideC_,
+                                     MPadded_,
+                                     NPadded_,
+                                     KPadded_,
+                                     K0Padded_,
+                                     k_batch_),
+              a_element_op(a_element_op_),
+              b_element_op(b_element_op_),
+              c_element_op(c_element_op_)
+        {
+        }
+
+        AElementwiseOperation a_element_op;
+        BElementwiseOperation b_element_op;
+        CElementwiseOperation c_element_op;
+    };
+
+    using DefaultBlock2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+
+        void Print(const Argument& karg) { karg.Print(); }
+
+        float Run(const Argument& karg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                Print(karg);
+            }
+
+            const auto kbatch = karg.k_batch;
+
+            if(!GridwiseGemm::CheckValidity(karg))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 has invalid "
+                    "setting");
+            }
+
+            const auto b2c_map = DefaultBlock2CTileMap{};
+            index_t gdx, gdy, gdz;
+            std::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
+            const auto K0Padded     = karg.K0Padded;
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0Padded);
+
+            float ave_time = 0;
+
+            const auto Run = [&](const auto& kernel) {
+                if(kbatch > 1)
+                    hipGetErrorString(hipMemsetAsync(karg.p_c_grid,
+                                                     0,
+                                                     karg.M * karg.N * sizeof(CDataType),
+                                                     stream_config.stream_id_));
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(gdx, gdy, gdz),
+                                           dim3(BlockSize),
+                                           0,
+                                           static_cast<typename GridwiseGemm::Argument>(karg),
+                                           b2c_map,
+                                           karg.a_element_op,
+                                           karg.b_element_op,
+                                           karg.c_element_op);
+            };
+
+            if(has_main_k0_block_loop)
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_splitk_lds_direct_load<GridwiseGemm,
+                                                                  true,
+                                                                  InMemoryDataOperationEnum::Set,
+                                                                  DefaultBlock2CTileMap,
+                                                                  AElementwiseOperation,
+                                                                  BElementwiseOperation,
+                                                                  CElementwiseOperation>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_splitk_lds_direct_load<
+                        GridwiseGemm,
+                        true,
+                        InMemoryDataOperationEnum::AtomicAdd,
+                        DefaultBlock2CTileMap,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CElementwiseOperation>;
+
+                    Run(kernel);
+                }
+            }
+            else
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_splitk_lds_direct_load<GridwiseGemm,
+                                                                  false,
+                                                                  InMemoryDataOperationEnum::Set,
+                                                                  DefaultBlock2CTileMap,
+                                                                  AElementwiseOperation,
+                                                                  BElementwiseOperation,
+                                                                  CElementwiseOperation>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_splitk_lds_direct_load<
+                        GridwiseGemm,
+                        false,
+                        InMemoryDataOperationEnum::AtomicAdd,
+                        DefaultBlock2CTileMap,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CElementwiseOperation>;
+
+                    Run(kernel);
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& karg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(karg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op,
+                             index_t KBatch)
+    {
+        return Argument(p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        GridwiseGemm::CalculateMPadded(M),
+                        GridwiseGemm::CalculateNPadded(N),
+                        GridwiseGemm::CalculateKPadded(K, KBatch),
+                        GridwiseGemm::CalculateK0Padded(K, KBatch),
+                        KBatch,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op,
+                                                      ck::index_t KBatch = 1) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          GridwiseGemm::CalculateMPadded(M),
+                                          GridwiseGemm::CalculateNPadded(N),
+                                          GridwiseGemm::CalculateKPadded(K, KBatch),
+                                          GridwiseGemm::CalculateK0Padded(K, KBatch),
+                                          KBatch,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{
+            {PipelineVersion::v1, "v1"}, {PipelineVersion::v2, "v2"}, {PipelineVersion::v4, "v4"}};
+
+        // clang-format off
+        str << "DeviceGemmXdlSplitKCShuffle_LdsDirectLoad"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferScalarPerVector << ", "
+            << BBlockTransferScalarPerVector << ", "
+            << CShuffleMRepeatPerShuffle << ", "
+            << CShuffleNRepeatPerShuffle << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">"
+            << " LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer] << ", "
+            << "Prefetch: "
+            << NumGemmKPrefetchStage;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
+#include "ck/utility/amd_lds.hpp"
 #include "ck/utility/common_header.hpp"
 #include "ck/tensor_description/multi_index_transform_helper.hpp"
 #include "ck/tensor_description/tensor_descriptor.hpp"
@@ -491,22 +492,6 @@ struct GridwiseGemmMultipleD_Xdl_CShuffle_LdsDirectLoad
 
     __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
 
-    template <typename DataType>
-    __device__ static auto AllocateBlockBuffers(void* p_shared,
-                                                int32_t num_elems,
-                                                int32_t offset_elems,
-                                                int32_t max_lds_align)
-    {
-        const int32_t single_buffer_offset = math::integer_least_multiple(num_elems, max_lds_align);
-        return generate_tuple(
-            [&](auto i) {
-                const int32_t local_offset = i * single_buffer_offset;
-                return make_dynamic_buffer<AddressSpaceEnum::Lds>(
-                    static_cast<DataType*>(p_shared) + local_offset + offset_elems, num_elems);
-            },
-            Number<NumGemmKPrefetchStage>{});
-    }
-
     template <bool HasMainKBlockLoop,
               typename AGridDesc_AK0_M_AK1,
               typename BGridDesc_BK0_N_BK1,
@@ -640,14 +625,20 @@ struct GridwiseGemmMultipleD_Xdl_CShuffle_LdsDirectLoad
         constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
             a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
-        auto a_block_buffers = AllocateBlockBuffers<AComputeDataType>(
-            p_shared, a_block_desc_ak0_m_ak1.GetElementSpaceSize(), 0, max_lds_align);
+        const auto a_buffers_offset = 0;
+        auto a_block_buffers =
+            ck::lds_utils::AllocateLdsBuffers<AComputeDataType, NumGemmKPrefetchStage>(
+                p_shared,
+                a_block_desc_ak0_m_ak1.GetElementSpaceSize(),
+                a_buffers_offset,
+                max_lds_align);
         const auto b_buffers_offset = a_block_space_size_aligned * NumGemmKPrefetchStage;
         auto b_block_buffers =
-            AllocateBlockBuffers<BComputeDataType>(p_shared,
-                                                   b_block_desc_bk0_n_bk1.GetElementSpaceSize(),
-                                                   b_buffers_offset,
-                                                   max_lds_align);
+            ck::lds_utils::AllocateLdsBuffers<BComputeDataType, NumGemmKPrefetchStage>(
+                p_shared,
+                b_block_desc_bk0_n_bk1.GetElementSpaceSize(),
+                b_buffers_offset,
+                max_lds_align);
 
         constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
         constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);

include/ck/utility/amd_lds.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/amd_address_space.hpp"
+#include "ck/utility/dynamic_buffer.hpp"
+#include "ck/utility/math.hpp"
+
+namespace ck {
+
+namespace lds_utils {
+
+/** \brief Allocate a given number of buffers in LDS and return them as a tuple.
+ *
+ * \tparam DataType Data type of elements to be stored in LDS.
+ * \tparam NumBuffers Number of buffers to be allocated.
+ * \param lds_ptr Address of the beginning of LDS space.
+ * \param num_elems_per_buffer Number of elements to allocate per single buffer.
+ * \param start_offset_elems Number of elements to move from the start of LDS for the allocation of
+ * the first buffer. \param lds_alignment Alignment of every buffer allocation given as a number of
+ * elements. \return Tuple of dynamic buffers representing memory allocated in LDS.
+ */
+template <typename DataType, index_t NumBuffers>
+__device__ static auto AllocateLdsBuffers(void* lds_ptr,
+                                          int32_t num_elems_per_buffer,
+                                          int32_t start_offset_elems,
+                                          int32_t lds_alignment)
+{
+    const DataType* lds_start = static_cast<DataType*>(lds_ptr) + start_offset_elems;
+    const int32_t single_buffer_offset =
+        math::integer_least_multiple(num_elems_per_buffer, lds_alignment);
+    return generate_tuple(
+        [&](auto i) {
+            const int32_t local_offset = i * single_buffer_offset;
+            return make_dynamic_buffer<AddressSpaceEnum::Lds>(lds_start + local_offset,
+                                                              num_elems_per_buffer);
+        },
+        Number<NumBuffers>{});
+}
+
+} // namespace lds_utils
+} // namespace ck

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

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -36,6 +36,11 @@ void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
+
+void add_device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
 #endif
 #ifdef CK_ENABLE_FP32
 void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(
@@ -192,6 +197,7 @@ struct DeviceOperationInstanceFactory<
                               is_same_v<CLayout, Row>)
             {
                 add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
                               is_same_v<CLayout, Row>)

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

@@ -8,6 +8,7 @@ list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_in
                                   device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
                                   device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
                                   device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
+                                  device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp
                                   device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_instance.cpp
                                   device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp
                                   device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp

library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, 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_xdl_splitk_c_shuffle_lds_direct_load.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNPadding  = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#######################################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|         ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|         BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#######################################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization| Prefetch|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|          ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar| AddExtraM|          ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#######################################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |    Stage|      |      |      |      |    |     |     | Wave| Wave| Lengths_KBatch_K0_M_K1|               |               |      PerVector|          | Lengths_KBatch_K0_N_K1|               |              |      PerVector|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#######################################|     |      |      |        |        |        |        |            |            |            |              |         |      |      |      |      |    |     |     |     |     |                       |               |               |               |          |                       |               |              |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,       1,   256,    16,   128,     4,  16,   16,   16,    1,    2,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                  S<1, 16, 1, 16>,               4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,       1,    64,    16,    16,     8,   8,   16,   16,    1,    1,         S<1, 1, 16, 4>,  S<0, 2, 1, 3>,              3,              2,         0,         S<1, 1, 16, 4>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,       1,    64,    16,    16,     4,  16,   16,   16,    1,    1,          S<1, 1, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 1, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,       2,    64,    16,    16,     8,  16,   16,   16,    1,    1,          S<1, 1, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 1, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                   S<1, 16, 1, 4>,               4>,
+
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       1,   256,   128,   128,     4,  16,   32,   32,    2,    2,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                   S<1, 16, 1, 16>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       1,   256,    32,    32,     4,  16,   16,   16,    1,    1,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 8>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       1,   256,    16,    64,     8,  16,   16,   16,    1,    1,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 8>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       1,   128,    16,    64,     4,  32,   16,   16,    1,    2,         S<1, 2, 4, 16>,  S<0, 2, 1, 3>,              3,              2,         0,         S<1, 2, 4, 16>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 8>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       1,   128,    16,    32,     8,   8,   16,   16,    1,    1,         S<1, 2, 16, 4>,  S<0, 2, 1, 3>,              3,              2,         0,         S<1, 2, 16, 4>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 8>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       1,   128,    16,    32,     4,   8,   16,   16,    1,    1,         S<1, 2, 16, 4>,  S<0, 2, 1, 3>,              3,              2,         0,         S<1, 2, 16, 4>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                     S<1, 8, 1, 8>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       1,    64,    16,    16,     4,  32,   16,   16,    1,    1,         S<1, 1, 4, 16>,  S<0, 2, 1, 3>,              3,              2,         0,         S<1, 1, 4, 16>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 4>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       2,   256,    64,    16,     4,  16,   16,   16,    1,    1,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 32, 1, 4>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       2,   256,    16,    64,     4,  16,   16,   16,    1,    1,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 8>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,       2,    64,    16,    16,     8,  16,   16,   16,    1,    1,          S<1, 1, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 1, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 4>,              4>,
+
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNKPadding,       1,   256,   128,   128,     4,  16,   32,   32,    2,    2,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                   S<1, 16, 1, 16>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNKPadding,       1,   256,    16,   128,     4,  32,   16,   16,    1,    2,         S<1, 4, 4, 16>,  S<0, 2, 1, 3>,              3,              2,         0,         S<1, 4, 4, 16>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                   S<1, 16, 1, 16>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNKPadding,       1,   256,    32,    32,     8,  16,   16,   16,    1,    1,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 8>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNKPadding,       1,   256,    32,    32,     4,  16,   16,   16,    1,    1,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 8>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNKPadding,       1,   256,    16,    64,     4,  16,   16,   16,    1,    1,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 8>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNKPadding,       1,   128,    16,    32,     8,   8,   16,   16,    1,    1,         S<1, 2, 16, 4>,  S<0, 2, 1, 3>,              3,              2,         0,         S<1, 2, 16, 4>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNKPadding,       1,    64,    16,    16,     4,  32,   16,   16,    1,    1,         S<1, 1, 4, 16>,  S<0, 2, 1, 3>,              3,              2,         0,         S<1, 1, 4, 16>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 16, 1, 4>,              4>,
+        DeviceGemmXdlSplitKCShuffle_LdsDirectLoad<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNKPadding,       2,   256,    64,    16,     4,  16,   16,   16,    1,    1,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,              3,              2,         0,          S<1, 4, 8, 8>,  S<0, 2, 1, 3>,             3,              2,         0,           1,           1,                    S<1, 32, 1, 4>,              4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck