Merge branch 'develop' into aosewski/gemm_tile_loop

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

@@ -555,9 +555,7 @@ struct DeviceGemmMultipleDMultipleR_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
-             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-             ck::get_device_name() == "gfx942"))
+        if(!ck::is_xdl_supported())
         {
             return false;
         }

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

@@ -491,9 +491,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
-             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-             ck::get_device_name() == "gfx942"))
+        if(!ck::is_xdl_supported())
         {
             return false;
         }

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

@@ -645,6 +645,11 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<0, ReduceOperatio
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
         return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
                                            arg.b_grid_desc_bk0_n_bk1_,
                                            arg.c_grid_desc_m_n_,

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

@@ -188,9 +188,7 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
-             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-             ck::get_device_name() == "gfx942"))
+        if(!ck::is_xdl_supported())
         {
             return false;
         }

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

@@ -648,9 +648,7 @@ struct DeviceGemmLayerNorm_Xdl_CShuffle : public BaseOperator
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
-             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-             ck::get_device_name() == "gfx942"))
+        if(!ck::is_xdl_supported())
         {
             return false;
         }

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

@@ -416,6 +416,11 @@ struct DeviceGemmXdlSkipBLds : public DeviceGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
         return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
                                            arg.b_grid_desc_k0_n_k1_,
                                            arg.c_grid_desc_m_n_,

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

@@ -231,6 +231,11 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
 
     static bool IsSupportedArgument(const Argument& karg)
     {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
         return GridwiseGemm::CheckValidity(karg);
     }
 

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, 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_streamk.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/hip_check_error.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,
+          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 ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          ck::index_t ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          ck::index_t BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL>
+struct DeviceGemmXdlStreamK : public DeviceGemmStreamK<ALayout,
+                                                       BLayout,
+                                                       CLayout,
+                                                       ADataType,
+                                                       BDataType,
+                                                       CDataType,
+                                                       AElementwiseOperation,
+                                                       BElementwiseOperation,
+                                                       CElementwiseOperation>
+{
+    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_bk0mk1_bk0nk1_mn_xdlops_streamk<
+        BlockSize,
+        BlockToCTileMap_GemmStreamK<MPerBlock,
+                                    NPerBlock,
+                                    K0PerBlock * K1,
+                                    StreamKReductionStrategy::Atomic>,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        ALayout,
+        BLayout,
+        CLayout,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXDL,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    // 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);
+            }
+            if(!GridwiseGemm::CheckValidity(karg))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 has invalid "
+                    "setting");
+            }
+
+            dim3 grid_dims = karg.block_mapping.get_grid_dims();
+
+            float ave_time = 0;
+
+            const auto kernel = kernel_gemm_xdlops_streamk<GridwiseGemm>;
+
+            // TODO: remove clear buffer for streamk kernels
+            if constexpr(GridwiseGemm::Block2CTileMap::ReductionStrategy ==
+                         StreamKReductionStrategy::Atomic)
+            {
+                hipGetErrorString(hipMemset(karg.p_c_grid, 0, karg.M * karg.N * sizeof(CDataType)));
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  grid_dims,
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  karg.p_a_grid,
+                                                  karg.p_b_grid,
+                                                  karg.p_c_grid,
+                                                  karg.p_workspace_,
+                                                  karg.M,
+                                                  karg.N,
+                                                  karg.K,
+                                                  karg.StrideA,
+                                                  karg.StrideB,
+                                                  karg.StrideC,
+                                                  karg.block_mapping);
+            }
+            else if constexpr(GridwiseGemm::Block2CTileMap::ReductionStrategy ==
+                              StreamKReductionStrategy::Reduction)
+            {
+                char* workspace_semaphore = reinterpret_cast<char*>(karg.p_workspace_) +
+                                            karg.block_mapping.get_workspace_size_for_acc(
+                                                sizeof(typename GridwiseGemm::FloatAcc));
+                auto preprocess = [&]() {
+                    hipGetErrorString(
+                        hipMemsetAsync(workspace_semaphore,
+                                       0,
+                                       karg.block_mapping.get_workspace_size_for_semaphore(),
+                                       stream_config.stream_id_));
+                };
+
+                ave_time = launch_and_time_kernel_with_preprocess(stream_config,
+                                                                  preprocess,
+                                                                  kernel,
+                                                                  grid_dims,
+                                                                  dim3(BlockSize),
+                                                                  0,
+                                                                  karg.p_a_grid,
+                                                                  karg.p_b_grid,
+                                                                  karg.p_c_grid,
+                                                                  karg.p_workspace_,
+                                                                  karg.M,
+                                                                  karg.N,
+                                                                  karg.K,
+                                                                  karg.StrideA,
+                                                                  karg.StrideB,
+                                                                  karg.StrideC,
+                                                                  karg.block_mapping);
+            }
+
+            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);
+        }
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* p_arg = dynamic_cast<const Argument*>(pArg);
+        if constexpr(GridwiseGemm::Block2CTileMap::ReductionStrategy ==
+                     StreamKReductionStrategy::Reduction)
+        {
+            return p_arg->block_mapping.get_workspace_size(sizeof(typename GridwiseGemm::FloatAcc));
+        }
+        else
+        {
+            return 0;
+        }
+    }
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+    }
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& karg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+             ck::get_device_name() == "gfx942"))
+        {
+            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,
+                             BElementwiseOperation,
+                             CElementwiseOperation,
+                             uint32_t NumSKBlocks = 0xffffffff)
+    {
+        const auto kernel = kernel_gemm_xdlops_streamk<GridwiseGemm>;
+        int occupancy, num_cu;
+        hipError_t rtn;
+        rtn = hipOccupancyMaxActiveBlocksPerMultiprocessor(
+            &occupancy, kernel, BlockSize, GridwiseGemm::GetSharedMemoryNumberOfByte());
+        hip_check_error(rtn);
+
+        hipDeviceProp_t dev_prop;
+        hipDevice_t dev;
+        rtn = hipGetDevice(&dev);
+        hip_check_error(rtn);
+        rtn = hipGetDeviceProperties(&dev_prop, dev);
+        hip_check_error(rtn);
+        num_cu = dev_prop.multiProcessorCount;
+
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        static_cast<uint32_t>(num_cu),
+                        static_cast<uint32_t>(occupancy),
+                        NumSKBlocks};
+    }
+
+    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,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation,
+                                                      index_t NumSKBlocks = 0) override
+    {
+        const auto kernel = kernel_gemm_xdlops_streamk<GridwiseGemm>;
+        int occupancy, num_cu;
+        hipError_t rtn;
+        rtn = hipOccupancyMaxActiveBlocksPerMultiprocessor(
+            &occupancy, kernel, BlockSize, GridwiseGemm::GetSharedMemoryNumberOfByte());
+        hip_check_error(rtn);
+
+        hipDeviceProp_t dev_prop;
+        hipDevice_t dev;
+        rtn = hipGetDevice(&dev);
+        hip_check_error(rtn);
+        rtn = hipGetDeviceProperties(&dev_prop, dev);
+        hip_check_error(rtn);
+        num_cu = dev_prop.multiProcessorCount;
+
+        return std::make_unique<Argument>(reinterpret_cast<const ADataType*>(p_a),
+                                          reinterpret_cast<const BDataType*>(p_b),
+                                          reinterpret_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          static_cast<uint32_t>(num_cu),
+                                          static_cast<uint32_t>(occupancy),
+                                          static_cast<uint32_t>(NumSKBlocks));
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override { return GridwiseGemm::GetTypeString(); }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -417,9 +417,7 @@ struct DeviceGemm_Xdl_WaveletModel_CShuffle : public DeviceGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
-             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-             ck::get_device_name() == "gfx942"))
+        if(!ck::is_xdl_supported())
         {
             return false;
         }

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

@@ -705,9 +705,7 @@ struct DeviceGroupedContractionMultipleD_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
-             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-             ck::get_device_name() == "gfx942"))
+        if(!ck::is_xdl_supported())
         {
             return false;
         }

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

@@ -258,7 +258,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                                                CDEElementwiseOp>
 {
     // FIXME
-    static_assert(NDimSpatial == 2, "wrong! only implemented for 2D now");
+    static_assert(NDimSpatial == 2 || NDimSpatial == 3,
+                  "wrong! only implemented for 2D and 3D now");
 
     using DeviceOp = DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1;
 
@@ -491,35 +492,73 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                 compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_c_wis_strides[i][0];
             });
 
+            static constexpr auto NonSpatialDimsNum = Number<3>{};
+
+            static constexpr auto DIdx = Number<NonSpatialDimsNum>{};
+            static constexpr auto HIdx =
+                NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+            static constexpr auto WIdx = NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{}
+                                                          : Number<NonSpatialDimsNum + 2>{};
+
+            static constexpr auto ZIdx = Number<NonSpatialDimsNum>{};
+            static constexpr auto YIdx =
+                NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+            static constexpr auto XIdx = NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{}
+                                                          : Number<NonSpatialDimsNum + 2>{};
+
             // problem definition
-            const index_t Y = b_g_k_c_xs_lengths[3];
-            const index_t X = b_g_k_c_xs_lengths[4];
+            const index_t Z = b_g_k_c_xs_lengths[ZIdx];
+            const index_t Y = b_g_k_c_xs_lengths[YIdx];
+            const index_t X = b_g_k_c_xs_lengths[XIdx];
 
-            const index_t ConvStrideH = conv_filter_strides_[0];
-            const index_t ConvStrideW = conv_filter_strides_[1];
+            const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+            const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+            const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
 
-            const index_t ConvDilationH = conv_filter_dilations_[0];
-            const index_t ConvDilationW = conv_filter_dilations_[1];
+            const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+            const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+            const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
             const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
             const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
+            const auto ZTilde = NDimSpatial == 3 ? ConvStrideD / GcdStrideDilationD : 1;
             const auto YTilde = ConvStrideH / GcdStrideDilationH;
             const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
+            for(index_t i_ztilde = 0; i_ztilde < ZTilde; ++i_ztilde)
+            {
+
                 for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
                 {
                     for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
                     {
                         // check slice is valid
+                        const auto ZDotSlice =
+                            NDimSpatial == 3 ? math::integer_divide_ceil(Z - i_ztilde, ZTilde) : 1;
                         const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
                         const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
-                    if(YDotSlice * XDotSlice <= 0)
+                        if(YDotSlice * XDotSlice * ZDotSlice <= 0)
                         {
                             continue;
                         }
 
+                        std::array<index_t, NDimSpatial> tildes;
+                        if constexpr(NDimSpatial == 2)
+                        {
+                            tildes = {i_ytilde, i_xtilde};
+                        }
+                        else if constexpr(NDimSpatial == 3)
+                        {
+                            tildes = {i_ztilde, i_ytilde, i_xtilde};
+                        }
+                        else
+                        {
+                            throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+                        }
+
                         const auto a_grid_desc_ak0_m_ak1 =
                             transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
                                 a_g_n_k_wos_lengths,
@@ -532,7 +571,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                                 conv_filter_dilations,
                                 input_left_pads,
                                 input_right_pads,
-                            {i_ytilde, i_xtilde});
+                                tildes);
 
                         const auto b_grid_desc_bk0_n_bk1 =
                             transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
@@ -546,7 +585,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                                 conv_filter_dilations,
                                 input_left_pads,
                                 input_right_pads,
-                            {i_ytilde, i_xtilde});
+                                tildes);
 
                         DsGridDesc_M_N ds_grid_desc_m_n;
 
@@ -566,7 +605,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                                     conv_filter_dilations,
                                     input_left_pads,
                                     input_right_pads,
-                                {i_ytilde, i_xtilde});
+                                    tildes);
                         });
 
                         const auto e_grid_desc_m_n =
@@ -581,11 +620,13 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                                 conv_filter_dilations,
                                 input_left_pads,
                                 input_right_pads,
-                            {i_ytilde, i_xtilde});
+                                tildes);
 
                         // desc for problem definition
-                    const auto a_grid_desc_m_k = transform_k0_m_k1_to_m_k(a_grid_desc_ak0_m_ak1);
-                    const auto b_grid_desc_n_k = transform_k0_m_k1_to_m_k(b_grid_desc_bk0_n_bk1);
+                        const auto a_grid_desc_m_k =
+                            transform_k0_m_k1_to_m_k(a_grid_desc_ak0_m_ak1);
+                        const auto b_grid_desc_n_k =
+                            transform_k0_m_k1_to_m_k(b_grid_desc_bk0_n_bk1);
 
                         a_grid_desc_m_k_container_.push_back(a_grid_desc_m_k);
                         b_grid_desc_n_k_container_.push_back(b_grid_desc_n_k);
@@ -609,7 +650,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                                                        block_2_etile_map))
                         {
                             ds_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
-                            GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                GridwiseGemm::
+                                    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
                                         ds_grid_desc_m_n));
 
                             e_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
@@ -619,6 +661,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                     }
                 }
             }
+        }
 
         void Print() const
         {
@@ -783,6 +826,11 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
         const index_t ConvK = arg.b_g_k_c_xs_lengths_[1];
         const index_t ConvC = arg.b_g_k_c_xs_lengths_[2];
 
@@ -803,7 +851,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
 
         // vector load for A matrix from global memory to LDS
         if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK> ||
-                     is_same_v<ALayout, tensor_layout::convolution::NHWGK>)
+                     is_same_v<ALayout, tensor_layout::convolution::GNDHWK> ||
+                     is_same_v<ALayout, tensor_layout::convolution::NHWGK> ||
+                     is_same_v<ALayout, tensor_layout::convolution::NDHWGK>)
         {
             if(!(ABlockTransferSrcVectorDim == 2 && ConvK % ABlockTransferSrcScalarPerVector == 0))
             {
@@ -816,7 +866,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
         }
 
         // vector load for B matrix from global memory to LDS
-        if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKYXC>)
+        if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKYXC> ||
+                     is_same_v<BLayout, tensor_layout::convolution::GKZYXC>)
         {
             if(!(BBlockTransferSrcVectorDim == 1 && ConvC % BBlockTransferSrcScalarPerVector == 0))
             {
@@ -835,7 +886,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
             using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
 
             if constexpr(is_same_v<DLayout, tensor_layout::convolution::GNHWC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::GNDHWC> ||
                          is_same_v<DLayout, tensor_layout::convolution::NHWGC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::NDHWGC> ||
                          is_same_v<DLayout, tensor_layout::convolution::G_NHW_C> ||
                          is_same_v<DLayout, tensor_layout::convolution::GC> ||
                          is_same_v<DLayout, tensor_layout::convolution::G_C>)
@@ -859,7 +912,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
 
         // vector store for E
         if constexpr(is_same_v<ELayout, tensor_layout::convolution::GNHWC> ||
-                     is_same_v<ELayout, tensor_layout::convolution::NHWGC>)
+                     is_same_v<ELayout, tensor_layout::convolution::GNDHWC> ||
+                     is_same_v<ELayout, tensor_layout::convolution::NHWGC> ||
+                     is_same_v<ELayout, tensor_layout::convolution::NDHWGC>)
         {
             // vector store C matrix into global memory
             if(!(ConvC % CDEBlockTransferScalarPerVector_NPerBlock == 0))

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

@@ -195,17 +195,17 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
 
     template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
     static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
-        ck::index_t N,
-        ck::index_t K,
-        ck::index_t C,
-        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
-        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
-        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
-        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
-        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
-        std::array<ck::index_t, NDimSpatial> input_left_pads,
-        std::array<ck::index_t, NDimSpatial> input_right_pads,
-        ck::index_t batch_k)
+        const ck::index_t N,
+        const ck::index_t K,
+        const ck::index_t C,
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+        const ck::index_t batch_k)
     {
         using namespace ck;
 
@@ -347,17 +347,17 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     } // function end
     template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
     static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
-        ck::index_t N,
-        ck::index_t K,
-        ck::index_t C,
-        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
-        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
-        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
-        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
-        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
-        std::array<ck::index_t, NDimSpatial> input_left_pads,
-        std::array<ck::index_t, NDimSpatial> input_right_pads,
-        ck::index_t batch_k)
+        const ck::index_t N,
+        const ck::index_t K,
+        const ck::index_t C,
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+        const ck::index_t batch_k)
     {
         using namespace ck;
 
@@ -515,17 +515,17 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
 
     template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
     static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
-        ck::index_t N,
-        ck::index_t K,
-        ck::index_t C,
-        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
-        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
-        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
-        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
-        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
-        std::array<ck::index_t, NDimSpatial> input_left_pads,
-        std::array<ck::index_t, NDimSpatial> input_right_pads,
-        ck::index_t batch_k)
+        const ck::index_t N,
+        const ck::index_t K,
+        const ck::index_t C,
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+        const ck::index_t batch_k)
     {
         using namespace ck;
 
@@ -784,17 +784,19 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
         Argument(const InDataType* p_in_grid,
                  WeiDataType* p_wei_grid,
                  const OutDataType* p_out_grid,
-                 ck::index_t G,
-                 ck::index_t N,
-                 ck::index_t K,
-                 ck::index_t C,
-                 std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
-                 std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
-                 std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
-                 std::array<ck::index_t, NDimSpatial> conv_filter_strides,
-                 std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
-                 std::array<ck::index_t, NDimSpatial> input_left_pads,
-                 std::array<ck::index_t, NDimSpatial> input_right_pads,
+                 const ck::index_t G,
+                 const ck::index_t N,
+                 const ck::index_t K,
+                 const ck::index_t C,
+                 const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+                 const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+                 const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+                 const std::array<ck::index_t, NDimSpatial + 3>& /*input_strides*/,
+                 const std::array<ck::index_t, NDimSpatial + 3>& /*output_strides*/,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                 const std::array<ck::index_t, NDimSpatial>& input_right_pads,
                  InElementwiseOperation in_element_op,
                  WeiElementwiseOperation wei_element_op,
                  OutElementwiseOperation out_element_op,
@@ -897,18 +899,18 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
         InElementwiseOperation c_element_op_;
 
         // for checking IsSupportedArgument()
-        index_t Conv_G_;
-        index_t Conv_N_;
-        index_t Conv_K_;
-        index_t Conv_C_;
-
-        std::array<ck::index_t, NDimSpatial> input_spatial_lengths_;
-        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths_;
-        std::array<ck::index_t, NDimSpatial> output_spatial_lengths_;
-        std::array<ck::index_t, NDimSpatial> conv_filter_strides_;
-        std::array<ck::index_t, NDimSpatial> conv_filter_dilations_;
-        std::array<ck::index_t, NDimSpatial> input_left_pads_;
-        std::array<ck::index_t, NDimSpatial> input_right_pads_;
+        const index_t Conv_G_;
+        const index_t Conv_N_;
+        const index_t Conv_K_;
+        const index_t Conv_C_;
+
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths_;
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths_;
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths_;
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides_;
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations_;
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads_;
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads_;
         index_t k_batch_;
     };
 
@@ -1111,17 +1113,19 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     static auto MakeArgument(const InDataType* p_in_grid,
                              WeiDataType* p_wei_grid,
                              const OutDataType* p_out_grid,
-                             ck::index_t G,
-                             ck::index_t N,
-                             ck::index_t K,
-                             ck::index_t C,
-                             std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
-                             std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
-                             std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
-                             std::array<ck::index_t, NDimSpatial> conv_filter_strides,
-                             std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
-                             std::array<ck::index_t, NDimSpatial> input_left_pads,
-                             std::array<ck::index_t, NDimSpatial> input_right_pads,
+                             const ck::index_t G,
+                             const ck::index_t N,
+                             const ck::index_t K,
+                             const ck::index_t C,
+                             const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+                             const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+                             const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+                             const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
+                             const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
+                             const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                             const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                             const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                             const std::array<ck::index_t, NDimSpatial>& input_right_pads,
                              InElementwiseOperation in_element_op,
                              WeiElementwiseOperation wei_element_op,
                              OutElementwiseOperation out_element_op,
@@ -1137,6 +1141,8 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
                         input_spatial_lengths,
                         filter_spatial_lengths,
                         output_spatial_lengths,
+                        input_strides,
+                        output_strides,
                         conv_filter_strides,
                         conv_filter_dilations,
                         input_left_pads,
@@ -1153,17 +1159,19 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     MakeArgumentPointer(const void* p_in_grid,
                         void* p_wei_grid,
                         const void* p_out_grid,
-                        ck::index_t G,
-                        ck::index_t N,
-                        ck::index_t K,
-                        ck::index_t C,
-                        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
-                        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
-                        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
-                        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
-                        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
-                        std::array<ck::index_t, NDimSpatial> input_left_pads,
-                        std::array<ck::index_t, NDimSpatial> input_right_pads,
+                        const ck::index_t G,
+                        const ck::index_t N,
+                        const ck::index_t K,
+                        const ck::index_t C,
+                        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+                        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+                        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+                        const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
+                        const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
+                        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
                         InElementwiseOperation in_element_op,
                         WeiElementwiseOperation wei_element_op,
                         OutElementwiseOperation out_element_op,
@@ -1179,6 +1187,8 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
                                           input_spatial_lengths,
                                           filter_spatial_lengths,
                                           output_spatial_lengths,
+                                          input_strides,
+                                          output_strides,
                                           conv_filter_strides,
                                           conv_filter_dilations,
                                           input_left_pads,

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

@@ -681,9 +681,7 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
-             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-             ck::get_device_name() == "gfx942"))
+        if(!ck::is_xdl_supported())
         {
             return false;
         }

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

@@ -600,6 +600,11 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
         if((ck::type_convert<ck::index_t>(arg.gemm_desc_kernel_arg_.size()) +
             arg.skipped_group_count_) != arg.group_count_)
         {

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

@@ -496,6 +496,11 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
         if((ck::type_convert<ck::index_t>(arg.gemm_kernel_args_.size()) +
             arg.skipped_group_count_) != arg.group_count_)
         {

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

@@ -939,9 +939,7 @@ struct DeviceSplitKContractionMultipleD_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
-             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-             ck::get_device_name() == "gfx942"))
+        if(!ck::is_xdl_supported())
         {
             return false;
         }

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

@@ -7,6 +7,8 @@
 #include "ck/utility/number.hpp"
 #include "ck/tensor_description/tensor_adaptor.hpp"
 #include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include <limits>
+#include <stdlib.h>
 
 namespace ck {
 
@@ -728,4 +730,406 @@ struct LocalBlockToCTileMap
     index_t local_block_id_;
 };
 
+enum StreamKReductionStrategy
+{
+    Atomic = 0, // sk block use atomic to do reduction
+    Reduction,  // let some workgroup responsible for doing the reduction operation
+};
+
+template <uint32_t MPerBlock_,
+          uint32_t NPerBlock_,
+          uint32_t KPerBlock_,
+          StreamKReductionStrategy ReductionStrategy_ = StreamKReductionStrategy::Atomic,
+          uint32_t TileSwizzleSubM_                   = 8>
+struct BlockToCTileMap_GemmStreamK
+{
+    static constexpr uint32_t min_k_iters_per_sk_block          = 2;
+    static constexpr uint32_t MPerBlock                         = MPerBlock_;
+    static constexpr uint32_t NPerBlock                         = NPerBlock_;
+    static constexpr uint32_t KPerBlock                         = KPerBlock_;
+    static constexpr StreamKReductionStrategy ReductionStrategy = ReductionStrategy_;
+    static constexpr uint32_t tile_swizzle_sub_m                = TileSwizzleSubM_;
+
+    //--------------------------------------
+    // pass to device
+    uint32_t sk_num_blocks;
+    uint32_t sk_num_big_blocks;
+    uint32_t dp_start_block_idx;
+    uint32_t reduction_start_block_idx;
+    uint32_t k_iters_per_big_block;
+    MDiv2 n_tiles;
+    MDiv k_iters_per_tile;
+    MDiv eqav_tiles_big;    // for reduction
+    MDiv eqav_tiles_little; // for reduction
+
+    // MDiv tile_swizzle_sub_m_rem;
+    //--------------------------------------
+
+    // prefer construct on host
+    BlockToCTileMap_GemmStreamK(uint32_t m,
+                                uint32_t n,
+                                uint32_t k,
+                                uint32_t num_cu,
+                                uint32_t occupancy,
+                                uint32_t sk_blocks = 0xffffffff)
+    {
+        uint32_t num_tiles =
+            math::integer_divide_ceil(m, MPerBlock) * math::integer_divide_ceil(n, NPerBlock);
+        k_iters_per_tile = MDiv(math::integer_divide_ceil(k, KPerBlock));
+
+        // one cu can hold one wg at one time, from the whole chip's point of view
+        // if number of wg is same as num_cu, we call it 1 dispatch
+        // if number of wg is 2x num_cu, we call it 2 dispatches.
+        // one dispatch can deliver wg same as num_cu (full dispatch), or less than num_cu (partial
+        // dispatch)
+        //
+        uint32_t full_dispatches         = num_tiles / num_cu;
+        uint32_t full_dispatch_tiles     = full_dispatches * num_cu;
+        uint32_t partial_dispatche_tiles = num_tiles - full_dispatch_tiles;
+
+        uint32_t sk_occupancy = occupancy;
+        uint32_t dp_tiles     = full_dispatch_tiles;
+        uint32_t sk_tiles     = partial_dispatche_tiles;
+
+        if(full_dispatches < occupancy)
+        {
+            // in this case, we allocate all blocks as sk blocks
+            // sk_occupancy = occupancy - full_dispatches;
+            sk_occupancy = 1; // TODO: single occ seems better
+            dp_tiles     = full_dispatch_tiles;
+            sk_tiles     = partial_dispatche_tiles;
+        }
+        else if((occupancy > 1) && (full_dispatches % occupancy == occupancy - 1))
+        {
+            // e.g. occupancy = 2, full_dispatches = 3, 5, 7 ...
+            //      occupancy = 3, full_dispatches = 5, 8, 11 ...
+            //      occupancy = 4, full_dispatches = 7, 11 ...
+            sk_occupancy = 1; // left 1 slot for sk occupancy
+            dp_tiles     = full_dispatch_tiles;
+            sk_tiles     = partial_dispatche_tiles;
+        }
+        else
+        {
+            // others, we reduce 1 dispatch from dp, together with partial dispatch,
+            // to construct sk dispatch
+            sk_occupancy = occupancy - ((full_dispatches - 1) % occupancy);
+            dp_tiles     = full_dispatch_tiles - num_cu;
+            sk_tiles     = partial_dispatche_tiles + num_cu;
+        }
+
+        // uint32_t dp_iters_per_block = k_iters_per_tile.get();
+        uint32_t sk_total_iters = k_iters_per_tile.get() * sk_tiles;
+        uint32_t dp_num_blocks  = 0;
+
+        {
+            uint32_t min_sk_tiles = (sk_tiles >= num_cu) ? num_cu : (sk_tiles + 1);
+            uint32_t max_sk_tiles =
+                (sk_tiles >= num_cu) ? num_cu * sk_occupancy
+                                     : math::min(num_cu, sk_total_iters / min_k_iters_per_sk_block);
+
+            // if use dp for sk-block, how many iters do we need
+            uint32_t dp_for_sk_iters = k_iters_per_tile.get();
+
+            uint32_t best_sk_score =
+                std::numeric_limits<int>::max(); // we need to find the smallest sk iters
+            for(uint32_t tentative_sk_blocks = min_sk_tiles; tentative_sk_blocks < max_sk_tiles;
+                tentative_sk_blocks++)
+            {
+                uint32_t tentative_sk_iters_per_block =
+                    (sk_total_iters + tentative_sk_blocks - 1) / tentative_sk_blocks;
+                uint32_t tentative_sk_iters = tentative_sk_iters_per_block;
+                uint32_t sk_blocks_per_tile = (tentative_sk_blocks + sk_tiles - 1) / sk_tiles;
+
+                // TODO: carefully adjust this parameter
+                //       the more sk_blocks_per_tile, the worse the overhead
+                uint32_t cross_sk_blocks_overhead = sk_blocks_per_tile;
+                if(tentative_sk_blocks % sk_tiles != 0)
+                {
+                    // penalty for uneven divide
+                    cross_sk_blocks_overhead +=
+                        sk_blocks_per_tile * tentative_sk_iters_per_block / 50;
+                }
+
+                uint32_t tentative_sk_score = tentative_sk_iters + cross_sk_blocks_overhead;
+
+                if(tentative_sk_score < best_sk_score)
+                {
+                    best_sk_score = tentative_sk_score;
+                    sk_num_blocks = tentative_sk_blocks;
+                }
+            }
+
+            if(best_sk_score >= dp_for_sk_iters)
+            {
+                sk_num_blocks = 0;
+            }
+
+            // give a chance to control num of sk blocks
+            sk_num_blocks = sk_blocks != 0xffffffff ? sk_blocks : sk_num_blocks;
+
+            if(sk_num_blocks == 0)
+            {
+                sk_num_big_blocks     = 0;
+                k_iters_per_big_block = 0;
+
+                dp_num_blocks      = num_tiles; // all tile to be dp block
+                dp_start_block_idx = 0;
+                sk_total_iters     = 0; // clear this tiles
+            }
+            else
+            {
+                // k_iters_per_sk_block is the floor of avg each ck block loop over tiles.
+                // we need to decide how many iters for each sk block
+                // let m = k_iters_per_sk_block
+                // some of the sk block (little) will cover m iters, some (big) will cover m+1
+                // we have
+                // 1) l + b = sk_blocks
+                // 2) l * m + b * (m + 1) = sk_total_iters
+                //      => (l + b) * m + b = sk_total_iters
+                //      => sk_blocks * m + b = sk_total_iters
+                //      => b = sk_total_iters - m * sk_blocks
+                //      NOTE: big could be zero
+                uint32_t k_iters_per_sk_block = sk_total_iters / sk_num_blocks;
+                sk_num_big_blocks     = sk_total_iters - k_iters_per_sk_block * sk_num_blocks;
+                k_iters_per_big_block = k_iters_per_sk_block + 1;
+
+                dp_num_blocks      = dp_tiles;
+                dp_start_block_idx = (sk_num_blocks + num_cu - 1) / num_cu * num_cu;
+            }
+        }
+        n_tiles                   = MDiv2(math::integer_divide_ceil(n, NPerBlock));
+        reduction_start_block_idx = dp_start_block_idx + dp_num_blocks;
+
+        if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
+        {
+            uint32_t upper_big    = math::lcm(k_iters_per_big_block, k_iters_per_tile.get());
+            uint32_t upper_little = math::lcm(k_iters_per_big_block - 1, k_iters_per_tile.get());
+            eqav_tiles_big        = MDiv(upper_big / k_iters_per_tile.get());
+            eqav_tiles_little     = MDiv(upper_little / k_iters_per_tile.get());
+        }
+
+#if 0
+        printf("cu:%d, occupancy:%d, grids:%d, num_tiles:%d, dp_tiles:%d, sk_num_big_blocks:%d, "
+               "sk_num_blocks:%d, "
+               "sk_total_iters:%d, dp_start_block_idx:%d, dp_iters_per_block:%d, dp_num_blocks:%d, "
+               "k_iters_per_tile:%d, k_iters_per_big_block:%d, reduction_start_block_idx:%u, "
+               "sk_tiles:%u, workspace(acc float):%u\n",
+               num_cu,
+               occupancy,
+               get_grid_dims().x,
+               num_tiles,
+               dp_tiles,
+               sk_num_big_blocks,
+               sk_num_blocks,
+               sk_total_iters,
+               dp_start_block_idx,
+               dp_iters_per_block,
+               dp_num_blocks,
+               k_iters_per_tile.get(),
+               k_iters_per_big_block,
+               reduction_start_block_idx,
+               get_sk_tiles(),
+               get_workspace_size(sizeof(float)));
+#endif
+    }
+
+    __host__ __device__ uint32_t get_sk_total_iters() const
+    {
+        uint32_t sk_total_iters = sk_num_big_blocks * k_iters_per_big_block +
+                                  (sk_num_blocks - sk_num_big_blocks) * (k_iters_per_big_block - 1);
+        return sk_total_iters;
+    }
+
+    __host__ __device__ uint32_t get_sk_tiles() const
+    {
+        // tiles for sk
+        uint32_t sk_total_iters = get_sk_total_iters();
+        return k_iters_per_tile.div(sk_total_iters);
+    }
+
+    __host__ __device__ dim3 get_grid_dims() const
+    {
+        if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
+        {
+            return dim3(reduction_start_block_idx + get_sk_tiles(), 1, 1);
+        }
+        else
+            return dim3(reduction_start_block_idx, 1, 1);
+    }
+
+    __device__ uint32_t get_block_idx() const
+    {
+        // TODO: swizzle block index for better locality
+        return __builtin_amdgcn_readfirstlane(blockIdx.x);
+    }
+
+    __device__ void
+    get_block_itr(uint32_t block_idx, uint32_t& iter_start, uint32_t& iter_end) const
+    {
+        if(block_idx < sk_num_big_blocks)
+        {
+            iter_start = block_idx * k_iters_per_big_block;
+            iter_end   = iter_start + k_iters_per_big_block;
+        }
+        else if(block_idx < sk_num_blocks)
+        {
+            iter_start = (sk_num_big_blocks * k_iters_per_big_block) +
+                         (block_idx - sk_num_big_blocks) * (k_iters_per_big_block - 1);
+            iter_end = iter_start + (k_iters_per_big_block - 1);
+        }
+        else if(block_idx >= dp_start_block_idx)
+        {
+            uint32_t sk_total_iters     = get_sk_total_iters();
+            uint32_t dp_iters_per_block = k_iters_per_tile.get();
+            iter_start = sk_total_iters + (block_idx - dp_start_block_idx) * dp_iters_per_block;
+            iter_end   = iter_start + dp_iters_per_block;
+        }
+    }
+
+    __device__ uint32_t get_current_iter_length(uint32_t iter_start,
+                                                uint32_t iter_end,
+                                                uint32_t total_iter_length) const
+    {
+        uint32_t iter_length_mod, iter_length_quo /*unused*/;
+        k_iters_per_tile.divmod(iter_end, iter_length_quo, iter_length_mod);
+        uint32_t current_iter_length = math::min(
+            iter_length_mod == 0 ? (iter_end - iter_start) : iter_length_mod, total_iter_length);
+        return current_iter_length;
+    }
+
+    __device__ uint32_t get_tile_idx(uint32_t iter) const { return k_iters_per_tile.div(iter); }
+
+    __device__ void
+    get_tile_idx_with_offset(uint32_t iter, uint32_t& tile_idx, uint32_t& iter_offset) const
+    {
+        k_iters_per_tile.divmod(iter, tile_idx, iter_offset);
+    }
+
+    __device__ auto tile_to_spatial(uint32_t tile_idx, uint32_t m, uint32_t n) const
+    {
+        uint32_t m_tile_idx, n_tile_idx;
+        uint32_t n_tiles_value = math::integer_divide_ceil(n, NPerBlock);
+        n_tiles.divmod(tile_idx, n_tiles_value, m_tile_idx, n_tile_idx);
+
+        // swizzle tile
+        uint32_t m_tiles = math::integer_divide_ceil(m, MPerBlock);
+
+        uint32_t tile_swizzle_sub_m_rem = m_tiles % tile_swizzle_sub_m;
+
+        const auto sub_m_adapt = (m_tile_idx < (m_tiles - tile_swizzle_sub_m_rem))
+                                     ? tile_swizzle_sub_m
+                                     : tile_swizzle_sub_m_rem;
+
+        uint32_t m_tile_idx_sub0, m_tile_idx_sub1;
+        m_tile_idx_sub0 = m_tile_idx / tile_swizzle_sub_m;
+        m_tile_idx_sub1 = m_tile_idx % tile_swizzle_sub_m;
+
+        uint32_t tile_idx_local = n_tile_idx + m_tile_idx_sub1 * n_tiles_value;
+
+        uint32_t m_tile_idx_with_adapt, n_tile_idx_with_adapt;
+
+        n_tile_idx_with_adapt = tile_idx_local / sub_m_adapt;
+        m_tile_idx_with_adapt = tile_idx_local % sub_m_adapt;
+        return make_tuple(m_tile_idx_with_adapt + m_tile_idx_sub0 * tile_swizzle_sub_m,
+                          n_tile_idx_with_adapt);
+    }
+
+    __host__ __device__ uint32_t get_workspace_size_for_acc(uint32_t acc_element_bytes) const
+    {
+        static constexpr uint32_t alignment = 128;
+        uint32_t acc_buffer_bytes =
+            MPerBlock * NPerBlock * get_total_acc_buffers() * acc_element_bytes;
+        return (acc_buffer_bytes + alignment - 1) / alignment * alignment;
+    }
+
+    __host__ __device__ uint32_t get_workspace_size_for_semaphore() const
+    {
+        return get_sk_tiles() * sizeof(uint32_t);
+    }
+
+    __host__ __device__ uint32_t get_workspace_size(uint32_t acc_element_bytes) const
+    {
+        return get_workspace_size_for_acc(acc_element_bytes) + get_workspace_size_for_semaphore();
+    }
+
+    __host__ __device__ uint32_t get_tile_intersections(uint32_t tiles_,
+                                                        const MDiv& eqav_tiles_) const
+    {
+        uint32_t tile_idx_       = tiles_ == 0 ? 0 : (tiles_ - 1);
+        uint32_t max_eqav_tiles_ = eqav_tiles_.get() - 1;
+        uint32_t quo_, rem_;
+        eqav_tiles_.divmod(tile_idx_, quo_, rem_);
+        return quo_ * max_eqav_tiles_ + rem_;
+    }
+
+    __host__ __device__ uint32_t get_tiles_cover_sk_block(uint32_t num_sk_blocks_,
+                                                          uint32_t iters_per_sk_block_) const
+    {
+        return k_iters_per_tile.div(num_sk_blocks_ * iters_per_sk_block_ + k_iters_per_tile.get() -
+                                    1);
+    }
+
+    __host__ __device__ uint32_t get_total_acc_buffers() const
+    {
+        uint32_t tiles_cover_big_blocks =
+            get_tiles_cover_sk_block(sk_num_big_blocks, k_iters_per_big_block);
+        uint32_t tiles_cover_little_blocks =
+            get_tiles_cover_sk_block(sk_num_blocks - sk_num_big_blocks, k_iters_per_big_block - 1);
+
+        uint32_t total_intersec_big =
+            get_tile_intersections(tiles_cover_big_blocks, eqav_tiles_big);
+        uint32_t total_intersec_little =
+            get_tile_intersections(tiles_cover_little_blocks, eqav_tiles_little);
+
+        return sk_num_blocks + total_intersec_big + total_intersec_little;
+    }
+
+    __device__ uint32_t get_acc_buffer_offset_from_tile(uint32_t tile_idx_) const
+    {
+        // TODO: from big to little
+        uint32_t tiles_cover_big_blocks =
+            get_tiles_cover_sk_block(sk_num_big_blocks, k_iters_per_big_block);
+        if(tile_idx_ < tiles_cover_big_blocks)
+        {
+            uint32_t touched_sk_blocks =
+                (tile_idx_ * k_iters_per_tile.get() + k_iters_per_big_block - 1) /
+                k_iters_per_big_block;
+            uint32_t current_intersec = get_tile_intersections(tile_idx_, eqav_tiles_big);
+            return touched_sk_blocks + current_intersec;
+        }
+        else
+        {
+            uint32_t iters_per_little_sk_block = k_iters_per_big_block - 1;
+            uint32_t tile_idx_little_reverse   = get_sk_tiles() - tile_idx_;
+            uint32_t touched_sk_blocks =
+                (tile_idx_little_reverse * k_iters_per_tile.get() + iters_per_little_sk_block - 1) /
+                iters_per_little_sk_block;
+            uint32_t current_intersec =
+                get_tile_intersections(tile_idx_little_reverse, eqav_tiles_little);
+            return get_total_acc_buffers() - (touched_sk_blocks + current_intersec);
+        }
+    }
+
+    __device__ uint32_t get_acc_buffer_offset_from_block(uint32_t block_idx_) const
+    {
+        uint32_t iters_per_big_sk_block    = k_iters_per_big_block;
+        uint32_t iters_per_little_sk_block = k_iters_per_big_block - 1;
+        if(block_idx_ < sk_num_big_blocks)
+        {
+            uint32_t touched_tiles    = k_iters_per_tile.div(block_idx_ * iters_per_big_sk_block +
+                                                          k_iters_per_tile.get() - 1);
+            uint32_t current_intersec = get_tile_intersections(touched_tiles, eqav_tiles_big);
+            return block_idx_ + current_intersec;
+        }
+        else
+        {
+            uint32_t block_idx_little_reverse = sk_num_blocks - block_idx_;
+            uint32_t touched_tiles            = k_iters_per_tile.div(
+                block_idx_little_reverse * iters_per_little_sk_block + k_iters_per_tile.get() - 1);
+            uint32_t current_intersec = get_tile_intersections(touched_tiles, eqav_tiles_little);
+            return get_total_acc_buffers() - (block_idx_little_reverse + current_intersec);
+        }
+    }
+};
+
 } // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp

@@ -67,6 +67,8 @@ template <typename A0B0B1DataType, // FIXME: don't assume A0/B0/B1 have same dat
           index_t B0BlockTransferDstScalarPerVector_BK1,
           bool B0ThreadTransferSrcResetCoordinateAfterRun, // ignored
           index_t B0BlockLdsExtraN,
+          index_t CDE0BlockTransferSrcVectorDim,
+          index_t CDE0BlockTransferSrcScalarPerVector,
           typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
           typename B1BlockTransferThreadClusterArrangeOrder,
           typename B1BlockTransferSrcAccessOrder,
@@ -710,13 +712,13 @@ struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
         constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
             make_naive_tensor_descriptor_packed(make_tuple(I1,   // MBlockId
                                                            I1,   // NBlockID
-                                                           I1,   // MRepeat
-                                                           I1,   // NRepeat
-                                                           I1,   // MWaveId
-                                                           I1,   // NWaveId
-                                                           I1,   // MPerXdl
-                                                           I1,   // NGroupNum
-                                                           I1,   // NInputNum
+                                                           m0,   // MRepeat
+                                                           n0,   // NRepeat
+                                                           m1,   // MWaveId
+                                                           n1,   // NWaveId
+                                                           m2,   // MPerXdl
+                                                           n2,   // NGroupNum
+                                                           n3,   // NInputNum
                                                            n4)); // registerNum
 
         auto d0s_thread_buf = generate_tuple(
@@ -732,8 +734,9 @@ struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
         const auto wave_id     = GetGemm0WaveIdx();
         const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
 
-        constexpr auto acc0_thread_desc = make_naive_tensor_descriptor_packed(
-            make_tuple(Number<Gemm0MXdlPerWave>{}, Number<Gemm0NXdlPerWave>{}, n2, n4));
+        static_assert(CDE0BlockTransferSrcScalarPerVector <= n4,
+                      "vector load must be not greater than n4");
+        static_assert(n4 % CDE0BlockTransferSrcScalarPerVector == 0);
 
         auto d0s_threadwise_copy = generate_tuple(
             [&](auto i) {
@@ -742,10 +745,19 @@ struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
                     A0B0B1DataType,
                     decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]),
                     decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
-                    Sequence<I1, I1, I1, I1, I1, I1, I1, I1, I1, n4>,
+                    Sequence<I1, // MBlockId
+                             I1, // NBlockID
+                             m0, // MRepeat
+                             n0, // NRepeat
+                             m1, // MWaveId
+                             n1, // NWaveId
+                             m2, // MPerXdl
+                             n2, // NGroupNum
+                             n3, // NInputNum
+                             n4>,
                     Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
-                    9,
-                    n4,
+                    9, // CDE0BlockTransferSrcVectorDim
+                    CDE0BlockTransferSrcScalarPerVector,
                     1,
                     false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
                            make_multi_index(block_work_idx[I0], // MBlockId
@@ -898,38 +910,27 @@ struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
                                                                     blockwise_gemm0,
                                                                     acc0_thread_buf,
                                                                     num_k_block_main_loop);
-            // bias+gelu
+            // multiple d
+            if constexpr(NumD0Tensor)
             {
-                static_for<0, Gemm0MXdlPerWave, 1>{}([&](auto mr) {
-                    static_for<0, Gemm0NXdlPerWave, 1>{}([&](auto nr) {
-                        static_for<0, n2, 1>{}([&](auto groupid) {
                 static_for<0, NumD0Tensor, 1>{}([&](auto i) {
-                                d0s_threadwise_copy(i).Run(
-                                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                    d0s_threadwise_copy(i).Run(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
                                                d0s_grid_buf[i],
                                                d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
                                                make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
                                                d0s_thread_buf(i));
                 });
-
-                            static_for<0, n4, 1>{}([&](auto i) {
-                                constexpr index_t c_offset = acc0_thread_desc.CalculateOffset(
-                                    make_tuple(mr, nr, groupid, i));
-
+                static_for<0, m0 * n0 * n2 * n4, 1>{}([&](auto i) {
                     // get reference to src data
                     const auto src_data_refs = generate_tie(
                         // return type should be lvalue
-                                    [&](auto iSrc) -> const auto& {
-                                        return d0s_thread_buf[iSrc][i];
-                                    },
+                        [&](auto iSrc) -> const auto& { return d0s_thread_buf[iSrc][i]; },
                         Number<NumD0Tensor>{});
 
                     // get reference to dst data
                     auto dst_data_refs = generate_tie(
                         // return type should be lvalue
-                                    [&](auto) -> auto& {
-                                        return acc0_thread_buf(Number<c_offset>{});
-                                    },
+                        [&](auto) -> auto& { return acc0_thread_buf(i); },
                         Number<2>{});
 
                     unpack2(cde0_element_op, dst_data_refs, src_data_refs);
@@ -937,27 +938,14 @@ struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
                 static_for<0, NumD0Tensor, 1>{}([&](auto i) {
                     d0s_threadwise_copy(i).MoveSrcSliceWindow(
                         d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
-                                    make_multi_index(0, 0, 0, 0, 0, 0, 0, 1, 0, 0));
-                            });
-                        });
-                        static_for<0, NumD0Tensor, 1>{}([&](auto i) {
-                            d0s_threadwise_copy(i).MoveSrcSliceWindow(
-                                d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
-                                make_multi_index(0, 0, 0, 1, 0, 0, 0, -n2.value, 0, 0));
-                        });
-                    });
-                    static_for<0, NumD0Tensor, 1>{}([&](auto i) {
-                        d0s_threadwise_copy(i).MoveSrcSliceWindow(
-                            d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
-                            make_multi_index(0, 0, 1, -Gemm0NXdlPerWave, 0, 0, 0, 0, 0, 0));
-                    });
-                });
-                static_for<0, NumD0Tensor, 1>{}([&](auto i) {
-                    d0s_threadwise_copy(i).MoveSrcSliceWindow(
-                        d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
-                        make_multi_index(0, 1, -Gemm0MXdlPerWave, 0, 0, 0, 0, 0, 0, 0));
+                        make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0));
                 });
             }
+            else
+            {
+                static_for<0, acc0_thread_buf.Size(), 1>{}(
+                    [&](auto i) { cde0_element_op(acc_thread_buf(i), acc0_thread_buf[i]); });
+            }
             // gemm1
             {
                 // TODO: explore using dynamic buffer for a1 thread buffer

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+
+struct GridwiseGemmPipeline_v3
+{
+    __host__ __device__ static constexpr bool IsSupported(index_t)
+    {
+        // TODO: improve applicability
+        return true;
+    }
+
+    template <typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    __device__ static void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // global read 0
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // LDS write 0
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        num_loop--;
+
+        while(num_loop > 0)
+        {
+            a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+            block_sync_lds();
+            b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+            block_sync_lds();
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+            num_loop--;
+        }
+        // tail
+        {
+            block_sync_lds();
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+} // namespace ck