/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.3/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.3/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-6.4.3/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include <iostream>
 #include <sstream>
  
 #include "ck/utility/reduction_operator.hpp"
 #include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
 #include "ck/tensor_operation/gpu/device/welford_helper.hpp"
 #include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp"
 #include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp"
 #include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.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 XDataType,
           typename YDataType,
           typename AccDataType,
           typename ScaleDataType,
           typename BiasDataType,
           typename MeanVarDataType,
           typename YElementwiseOp,
           index_t Rank,
           index_t NumBatchNormReduceDim,
           bool UseMultiblockInK,
           index_t BlockSize,
           index_t MThreadClusterSize,
           index_t KThreadClusterSize,
           index_t MThreadSliceSize,
           index_t KThreadSliceSize,
           index_t XSrcYDstVectorDim,
           index_t XSrcVectorSize,
           index_t YDstVectorSize,
           index_t ScaleSrcVectorSize,
           index_t BiasSrcVectorSize,
           index_t MeanVarSrcDstVectorSize>
 struct DeviceBatchNormFwdImpl : public DeviceBatchNormFwd<XDataType,
                                                           YDataType,
                                                           AccDataType,
                                                           ScaleDataType,
                                                           BiasDataType,
                                                           MeanVarDataType,
                                                           YElementwiseOp,
                                                           Rank,
                                                           NumBatchNormReduceDim>
 {
     static_assert(Rank <= 6, "Bigger Rank size is not supported!");
     static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
                   "Invalid thread cluster size assignments!");
  
     static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
                       (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
                   "Invalid thread slice sizes and/or vector sizes configuration, please check!");
  
     static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
  
     static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
     static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
  
     static auto MakeXY2dDescriptor(const std::array<index_t, Rank>& xyLengths,
                                    const std::array<index_t, Rank>& xyStrides,
                                    int blkGroupSize,
                                    int numBlockTileIteration)
     {
         const auto tupleXYLengths =
             generate_tuple([&](auto I) { return xyLengths[I]; }, Number<Rank>{});
         const auto tupleXYStrides =
             generate_tuple([&](auto I) { return xyStrides[I]; }, Number<Rank>{});
  
         const auto raw_grid_desc = make_naive_tensor_descriptor(tupleXYLengths, tupleXYStrides);
  
         const auto grid_desc_m_k = [&]() {
             using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
             using ReduceDims    = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
  
             const auto reduceDimLengths =
                 generate_tuple([&](auto I) { return xyLengths[NumInvariantDim + I]; },
                                Number<NumBatchNormReduceDim>{});
             const auto invariantDimLengths =
                 generate_tuple([&](auto I) { return xyLengths[I]; }, Number<NumInvariantDim>{});
  
             return transform_tensor_descriptor(raw_grid_desc,
                                                make_tuple(make_merge_transform(invariantDimLengths),
                                                           make_merge_transform(reduceDimLengths)),
                                                make_tuple(InvariantDims{}, ReduceDims{}),
                                                make_tuple(Sequence<0>{}, Sequence<1>{}));
         }();
  
         const auto invariantLength = grid_desc_m_k.GetLength(Number<0>{});
         const auto reduceLength    = grid_desc_m_k.GetLength(Number<1>{});
  
         const int workSizePerBlock = K_BlockTileSize * numBlockTileIteration;
         const auto mPad =
             math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
         const auto kPad = workSizePerBlock * blkGroupSize - reduceLength;
  
         auto grid_desc_m_k_padded =
             transform_tensor_descriptor(grid_desc_m_k,
                                         make_tuple(make_right_pad_transform(invariantLength, mPad),
                                                    make_right_pad_transform(reduceLength, kPad)),
                                         make_tuple(Sequence<0>{}, Sequence<1>{}),
                                         make_tuple(Sequence<0>{}, Sequence<1>{}));
  
         return (grid_desc_m_k_padded);
     };
  
     static auto MakeMeanVarCountOutputMG2dDescriptor(int invariantLength, int blkGroupSize)
     {
         const auto grid_desc_m_g = make_naive_tensor_descriptor(
             make_tuple(invariantLength, blkGroupSize), make_tuple(1, invariantLength));
  
         const auto mPad =
             math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
  
         auto grid_desc_m_g_padded =
             transform_tensor_descriptor(grid_desc_m_g,
                                         make_tuple(make_right_pad_transform(invariantLength, mPad),
                                                    make_pass_through_transform(blkGroupSize)),
                                         make_tuple(Sequence<0>{}, Sequence<1>{}),
                                         make_tuple(Sequence<0>{}, Sequence<1>{}));
  
         return (grid_desc_m_g_padded);
     };
  
     static auto MakeMeanVarCountInputMK2dDescriptor(int invariantLength, int blkGroupSize)
     {
         const auto reduceLength  = blkGroupSize;
         const auto grid_desc_m_k = make_naive_tensor_descriptor(
             make_tuple(invariantLength, reduceLength), make_tuple(1, invariantLength));
  
         const auto mPad =
             math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
         const auto kPad =
             math::integer_least_multiple(reduceLength, KThreadClusterSize) - reduceLength;
  
         auto grid_desc_m_k_padded =
             transform_tensor_descriptor(grid_desc_m_k,
                                         make_tuple(make_right_pad_transform(invariantLength, mPad),
                                                    make_right_pad_transform(reduceLength, kPad)),
                                         make_tuple(Sequence<0>{}, Sequence<1>{}),
                                         make_tuple(Sequence<0>{}, Sequence<1>{}));
  
         return (grid_desc_m_k_padded);
     };
  
     static auto
     MakeScaleBiasMeanVar1dDescriptor(const std::array<index_t, NumInvariantDim>& lengths,
                                      const std::array<index_t, NumInvariantDim>& strides)
     {
         const auto tupleLengths =
             generate_tuple([&](auto I) { return lengths[I]; }, Number<NumInvariantDim>{});
         const auto tupleStrides =
             generate_tuple([&](auto I) { return strides[I]; }, Number<NumInvariantDim>{});
  
         auto raw_grid_desc = make_naive_tensor_descriptor(tupleLengths, tupleStrides);
  
         auto grid_desc_m = transform_tensor_descriptor(
             raw_grid_desc,
             make_tuple(make_merge_transform(tupleLengths)),
             make_tuple(typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type{}),
             make_tuple(Sequence<0>{}));
  
         const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
  
         const auto mPad =
             math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
  
         auto grid_desc_m_padded =
             transform_tensor_descriptor(grid_desc_m,
                                         make_tuple(make_right_pad_transform(invariantLength, mPad)),
                                         make_tuple(Sequence<0>{}),
                                         make_tuple(Sequence<0>{}));
         return (grid_desc_m_padded);
     };
  
     using XYGridDesc_M_K             = decltype(MakeXY2dDescriptor({1}, {1}, 1, 1));
     using ScaleBiasMeanVarGridDesc_M = decltype(MakeScaleBiasMeanVar1dDescriptor({1}, {1}));
  
     struct Argument : public BaseArgument
     {
         Argument(const std::array<index_t, Rank> xyLengths,
                  const std::array<index_t, Rank> xStrides,
                  const std::array<index_t, Rank> yStrides,
                  const std::array<int, NumBatchNormReduceDim> reduceDims,
                  const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
                  const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
                  const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
                  const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
                  const XDataType* p_x,
                  const ScaleDataType* p_scale,
                  const BiasDataType* p_bias,
                  const YElementwiseOp y_elementwise_op,
                  double epsilon,
                  YDataType* p_y,
                  MeanVarDataType* resultSaveMean,
                  MeanVarDataType* resultSaveInvVariance,
                  double averageFactor,
                  MeanVarDataType* resultRunningMean,
                  MeanVarDataType* resultRunningVariance)
             : bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
               bnScaleStrides_(bnScaleStrides),
               bnBiasStrides_(bnBiasStrides),
               bnMeanVarStrides_(bnMeanVarStrides),
               p_x_(p_x),
               p_scale_(p_scale),
               p_bias_(p_bias),
               y_elementwise_op_(y_elementwise_op),
               p_y_(p_y),
               resultSaveMean_(resultSaveMean),
               resultSaveInvVariance_(resultSaveInvVariance),
               resultRunningMean_(resultRunningMean),
               resultRunningVariance_(resultRunningVariance)
         {
             xyLengths_ =
                 shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xyLengths, reduceDims);
             xStrides_ =
                 shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xStrides, reduceDims);
             yStrides_ =
                 shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(yStrides, reduceDims);
  
             std::tie(invariant_length_, reduce_length_) =
                 get_2d_lengths<Rank, NumBatchNormReduceDim>(xyLengths_);
  
             epsilon_       = type_convert<AccDataType>(epsilon);
             averageFactor_ = type_convert<AccDataType>(averageFactor);
  
             updateMovingAverage_ =
                 (resultRunningMean != nullptr && resultRunningVariance != nullptr);
             saveMeanInvVariance_ = (resultSaveMean != nullptr && resultSaveInvVariance_ != nullptr);
  
             if(UseMultiblockInK)
             {
                 int iterations = 1;
                 while(true)
                 {
                     int testBlkGroupSize = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
                                            (K_BlockTileSize * iterations);
  
                     // we want the blkGroupSize be not more than 16
                     if(testBlkGroupSize <= 16)
                         break;
  
                     iterations++;
                 };
  
                 blkGroupSize_ = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
                                 (K_BlockTileSize * iterations);
  
                 numBlockTileIteration_ = iterations;
             }
             else
             {
                 blkGroupSize_          = 1;
                 numBlockTileIteration_ = (reduce_length_ + K_BlockTileSize - 1) / K_BlockTileSize;
             };
  
             gridSize_ = (invariant_length_ + M_BlockTileSize - 1) / M_BlockTileSize * blkGroupSize_;
  
             x_grid_desc_m_k_ =
                 MakeXY2dDescriptor(xyLengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
             y_grid_desc_m_k_ =
                 MakeXY2dDescriptor(xyLengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
             scale_grid_desc_m_ =
                 MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnScaleStrides_);
             bias_grid_desc_m_ =
                 MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnBiasStrides_);
             mean_var_grid_desc_m_ =
                 MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnMeanVarStrides_);
         }
  
         AccDataType epsilon_;
         AccDataType averageFactor_;
  
         bool updateMovingAverage_;
         bool saveMeanInvVariance_;
  
         std::array<index_t, Rank> xyLengths_;
         std::array<index_t, Rank> xStrides_;
         std::array<index_t, Rank> yStrides_;
  
         std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths_;
         std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides_;
         std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides_;
         std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides_;
  
         const XDataType* p_x_;
         const ScaleDataType* p_scale_;
         const BiasDataType* p_bias_;
         const YElementwiseOp y_elementwise_op_;
         YDataType* p_y_;
  
         MeanVarDataType* resultSaveMean_;
         MeanVarDataType* resultSaveInvVariance_;
  
         MeanVarDataType* resultRunningMean_;
         MeanVarDataType* resultRunningVariance_;
  
         long_index_t invariant_length_;
         long_index_t reduce_length_;
  
         int blkGroupSize_;
         int numBlockTileIteration_;
         size_t gridSize_;
  
         XYGridDesc_M_K x_grid_desc_m_k_;
         XYGridDesc_M_K y_grid_desc_m_k_;
         ScaleBiasMeanVarGridDesc_M scale_grid_desc_m_;
         ScaleBiasMeanVarGridDesc_M bias_grid_desc_m_;
         ScaleBiasMeanVarGridDesc_M mean_var_grid_desc_m_;
  
         void* workspace_mean_;
         void* workspace_variance_;
         void* workspace_count_;
  
         void* control_;
     };
  
     size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
     {
         const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
  
         size_t workspace_size = 0;
  
         if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
         {
             // workspace for welford intermediate mean
             workspace_size +=
                 pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
  
             // workspace for welford intermediate variance
             workspace_size +=
                 pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
  
             // workspace for welford intermediate count
             workspace_size +=
                 pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(int32_t) + 64;
  
             // workspace for barrier objects, each barrier object consists of two integers
             // TODO: allocate barrier object memory globally to reuse it by other operators
             workspace_size += (pArg_->invariant_length_ + M_BlockTileSize - 1) / M_BlockTileSize *
                               sizeof(int) * 2;
         }
  
         return (workspace_size);
     };
  
     void SetWorkSpacePointer(BaseArgument* pArg,
                              void* p_workspace,
                              const StreamConfig& = StreamConfig{}) const override
     {
         Argument* pArg_ = dynamic_cast<Argument*>(pArg);
  
         pArg_->p_workspace_ = p_workspace;
  
         if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
         {
             // setup buffer used for intermediate welford mean
             pArg_->workspace_mean_ = static_cast<char*>(pArg_->p_workspace_);
  
             index_t mean_space_sz =
                 pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
  
             mean_space_sz = math::integer_least_multiple(mean_space_sz, 64);
  
             // setup buffer used for intermediate welford varirance
             pArg_->workspace_variance_ =
                 reinterpret_cast<char*>(pArg_->workspace_mean_) + mean_space_sz;
  
             index_t variance_space_sz =
                 pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
  
             variance_space_sz = math::integer_least_multiple(variance_space_sz, 64);
  
             // setup buffer used for intermediate welfor count
             pArg_->workspace_count_ =
                 reinterpret_cast<char*>(pArg_->workspace_variance_) + variance_space_sz;
  
             index_t count_space_sz =
                 pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(int32_t);
  
             count_space_sz = math::integer_least_multiple(count_space_sz, 64);
  
             pArg_->control_ = reinterpret_cast<char*>(pArg_->workspace_count_) + count_space_sz;
  
             index_t control_space_sz = (pArg_->invariant_length_ + M_BlockTileSize - 1) /
                                        M_BlockTileSize * sizeof(int) * 2;
  
             hip_check_error(hipMemset(pArg_->control_, 0, control_space_sz));
         };
     };
  
     struct Invoker : public BaseInvoker
     {
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
             float avg_time = 0;
  
             if(UseMultiblockInK && arg.blkGroupSize_ > 1)
             {
                 using GetReduceCountPerThreadFunctor =
                     GetReduceCountPerThreadForMultiblockWelford<K_BlockTileSize, KThreadSliceSize>;
  
                 GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
                     arg.blkGroupSize_, arg.numBlockTileIteration_, arg.reduce_length_);
  
                 const auto mean_var_count_grid_desc_m_g =
                     DeviceBatchNormFwdImpl::MakeMeanVarCountOutputMG2dDescriptor(
                         arg.invariant_length_, arg.blkGroupSize_);
  
                 const auto mean_var_count_grid_desc_m_k =
                     DeviceBatchNormFwdImpl::MakeMeanVarCountInputMK2dDescriptor(
                         arg.invariant_length_, arg.blkGroupSize_);
  
                 using MeanVarCountGridDesc_M_G = decltype(mean_var_count_grid_desc_m_g);
                 using MeanVarCountGridDesc_M_K = decltype(mean_var_count_grid_desc_m_k);
  
                 using GridwiseMultiblockBatchNormForward_ =
                     GridwiseMultiblockBatchNormForward<XDataType,
                                                        YDataType,
                                                        AccDataType,
                                                        ScaleDataType,
                                                        BiasDataType,
                                                        MeanVarDataType,
                                                        YElementwiseOp,
                                                        XYGridDesc_M_K,
                                                        MeanVarCountGridDesc_M_G,
                                                        MeanVarCountGridDesc_M_K,
                                                        ScaleBiasMeanVarGridDesc_M,
                                                        ScaleBiasMeanVarGridDesc_M,
                                                        GetReduceCountPerThreadFunctor,
                                                        BlockSize,
                                                        MThreadClusterSize,
                                                        KThreadClusterSize,
                                                        MThreadSliceSize,
                                                        KThreadSliceSize,
                                                        XSrcYDstVectorDim,
                                                        XSrcVectorSize,
                                                        YDstVectorSize,
                                                        ScaleSrcVectorSize,
                                                        BiasSrcVectorSize,
                                                        MeanVarSrcDstVectorSize>;
  
                 using GridwiseMultiblockWelfordFirstHalf_ =
                     GridwiseMultiblockWelfordFirstHalf<XDataType,
                                                        AccDataType,
                                                        MeanVarDataType,
                                                        XYGridDesc_M_K,
                                                        MeanVarCountGridDesc_M_G,
                                                        GetReduceCountPerThreadFunctor,
                                                        BlockSize,
                                                        MThreadClusterSize,
                                                        KThreadClusterSize,
                                                        MThreadSliceSize,
                                                        KThreadSliceSize,
                                                        XSrcYDstVectorDim,
                                                        XSrcVectorSize>;
  
                 using GridwiseWelfordSecondHalfBatchNormForwardFinal_ =
                     GridwiseWelfordSecondHalfBatchNormForwardFinal<XDataType,
                                                                    YDataType,
                                                                    AccDataType,
                                                                    ScaleDataType,
                                                                    BiasDataType,
                                                                    MeanVarDataType,
                                                                    YElementwiseOp,
                                                                    XYGridDesc_M_K,
                                                                    MeanVarCountGridDesc_M_K,
                                                                    ScaleBiasMeanVarGridDesc_M,
                                                                    ScaleBiasMeanVarGridDesc_M,
                                                                    BlockSize,
                                                                    MThreadClusterSize,
                                                                    KThreadClusterSize,
                                                                    MThreadSliceSize,
                                                                    KThreadSliceSize,
                                                                    XSrcYDstVectorDim,
                                                                    XSrcVectorSize,
                                                                    YDstVectorSize,
                                                                    ScaleSrcVectorSize,
                                                                    BiasSrcVectorSize,
                                                                    MeanVarSrcDstVectorSize>;
  
                 // It is found that:
                 // 1) gfx1030 does not support the GLC enabled vector load/store, so using the
                 //    two-kernel method for gfx1030
                 // 2) Profiler on gfx908 could hang even though it works when running examples
                 // 3) Single-kernel method works on gfx1100, but the performance it not better
                 //    than two-kernel method (due to more warps participating the barrier)
                 if(ck::get_device_name() == "gfx90a")
                 {
                     const auto kern_multiblock_batchnorm_fwd_ =
                         kernel_multiblock_batchnorm_forward<GridwiseMultiblockBatchNormForward_,
                                                             XDataType,
                                                             YDataType,
                                                             AccDataType,
                                                             ScaleDataType,
                                                             BiasDataType,
                                                             MeanVarDataType,
                                                             YElementwiseOp,
                                                             XYGridDesc_M_K,
                                                             MeanVarCountGridDesc_M_G,
                                                             MeanVarCountGridDesc_M_K,
                                                             ScaleBiasMeanVarGridDesc_M,
                                                             ScaleBiasMeanVarGridDesc_M,
                                                             GetReduceCountPerThreadFunctor>;
  
                     avg_time += launch_and_time_kernel(
                         stream_config,
                         kern_multiblock_batchnorm_fwd_,
                         dim3(arg.gridSize_),
                         dim3(BlockSize),
                         0,
                         arg.x_grid_desc_m_k_,
                         arg.y_grid_desc_m_k_,
                         mean_var_count_grid_desc_m_g, // for writing to mean/variance/count
                                                       // workspace by multiple workgroups
                         mean_var_count_grid_desc_m_k, // for reading from mean/variance/count
                                                       // workspace by each workgroup
                         arg.scale_grid_desc_m_,
                         arg.bias_grid_desc_m_,
                         arg.mean_var_grid_desc_m_,
                         get_reduce_count_per_thread,
                         arg.numBlockTileIteration_,
                         arg.epsilon_,
                         arg.p_x_,
                         static_cast<MeanVarDataType*>(arg.workspace_mean_),
                         static_cast<MeanVarDataType*>(arg.workspace_variance_),
                         static_cast<int32_t*>(arg.workspace_count_),
                         static_cast<int*>(arg.control_),
                         arg.p_scale_,
                         arg.p_bias_,
                         arg.y_elementwise_op_,
                         arg.p_y_,
                         arg.updateMovingAverage_, // true or false
                         arg.averageFactor_,
                         arg.resultRunningMean_,
                         arg.resultRunningVariance_,
                         arg.saveMeanInvVariance_, // true or false
                         arg.resultSaveMean_,
                         arg.resultSaveInvVariance_);
                 }
                 else
                 {
                     const auto kern_multiblock_welford_first_half =
                         kernel_multiblock_welford_first_half<GridwiseMultiblockWelfordFirstHalf_,
                                                              XDataType,
                                                              MeanVarDataType,
                                                              XYGridDesc_M_K,
                                                              MeanVarCountGridDesc_M_G,
                                                              GetReduceCountPerThreadFunctor>;
  
                     const auto kern_welford_second_half_batchnorm_forward_final =
                         kernel_welford_second_half_batchnorm_forward_final<
                             GridwiseWelfordSecondHalfBatchNormForwardFinal_,
                             XDataType,
                             YDataType,
                             AccDataType,
                             ScaleDataType,
                             BiasDataType,
                             MeanVarDataType,
                             YElementwiseOp,
                             XYGridDesc_M_K,
                             MeanVarCountGridDesc_M_K,
                             ScaleBiasMeanVarGridDesc_M,
                             ScaleBiasMeanVarGridDesc_M>;
  
                     avg_time += launch_and_time_kernel(
                         stream_config,
                         kern_multiblock_welford_first_half,
                         dim3(arg.gridSize_),
                         dim3(BlockSize),
                         0,
                         arg.x_grid_desc_m_k_,
                         mean_var_count_grid_desc_m_g,
                         get_reduce_count_per_thread,
                         arg.numBlockTileIteration_,
                         arg.p_x_,
                         static_cast<MeanVarDataType*>(arg.workspace_mean_),
                         static_cast<MeanVarDataType*>(arg.workspace_variance_),
                         static_cast<int32_t*>(arg.workspace_count_));
  
                     avg_time += launch_and_time_kernel(
                         stream_config,
                         kern_welford_second_half_batchnorm_forward_final,
                         dim3(arg.gridSize_),
                         dim3(BlockSize),
                         0,
                         arg.x_grid_desc_m_k_,
                         arg.y_grid_desc_m_k_,
                         mean_var_count_grid_desc_m_k,
                         arg.scale_grid_desc_m_,
                         arg.bias_grid_desc_m_,
                         arg.mean_var_grid_desc_m_,
                         arg.blkGroupSize_,
                         arg.numBlockTileIteration_,
                         arg.epsilon_,
                         static_cast<MeanVarDataType*>(arg.workspace_mean_),
                         static_cast<MeanVarDataType*>(arg.workspace_variance_),
                         static_cast<int32_t*>(arg.workspace_count_),
                         arg.p_x_,
                         arg.p_scale_,
                         arg.p_bias_,
                         arg.y_elementwise_op_,
                         arg.p_y_,
                         arg.updateMovingAverage_,
                         arg.averageFactor_,
                         arg.resultRunningMean_,
                         arg.resultRunningVariance_,
                         arg.saveMeanInvVariance_,
                         arg.resultSaveMean_,
                         arg.resultSaveInvVariance_);
                 };
             }
             else
             {
                 using GetReduceCountPerThreadFunctor =
                     GetReduceCountPerThreadForBlockwiseWelford<K_BlockTileSize, KThreadSliceSize>;
  
                 GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
                     arg.numBlockTileIteration_, arg.reduce_length_);
  
                 using GridwiseBatchNormForwardWithBlockwiseWelford_ =
                     GridwiseBatchNormForwardWithBlockwiseWelford<XDataType,
                                                                  YDataType,
                                                                  AccDataType,
                                                                  ScaleDataType,
                                                                  BiasDataType,
                                                                  MeanVarDataType,
                                                                  YElementwiseOp,
                                                                  XYGridDesc_M_K,
                                                                  ScaleBiasMeanVarGridDesc_M,
                                                                  ScaleBiasMeanVarGridDesc_M,
                                                                  GetReduceCountPerThreadFunctor,
                                                                  BlockSize,
                                                                  MThreadClusterSize,
                                                                  KThreadClusterSize,
                                                                  MThreadSliceSize,
                                                                  KThreadSliceSize,
                                                                  XSrcYDstVectorDim,
                                                                  XSrcVectorSize,
                                                                  YDstVectorSize,
                                                                  ScaleSrcVectorSize,
                                                                  BiasSrcVectorSize,
                                                                  MeanVarSrcDstVectorSize>;
  
                 const auto kern_batchnorm_fwd = kernel_batchnorm_forward_with_blockwise_welford<
                     GridwiseBatchNormForwardWithBlockwiseWelford_,
                     XDataType,
                     YDataType,
                     AccDataType,
                     ScaleDataType,
                     BiasDataType,
                     MeanVarDataType,
                     YElementwiseOp,
                     XYGridDesc_M_K,
                     ScaleBiasMeanVarGridDesc_M,
                     ScaleBiasMeanVarGridDesc_M,
                     GetReduceCountPerThreadFunctor>;
  
                 avg_time += launch_and_time_kernel(stream_config,
                                                    kern_batchnorm_fwd,
                                                    dim3(arg.gridSize_),
                                                    dim3(BlockSize),
                                                    0,
                                                    arg.x_grid_desc_m_k_,
                                                    arg.y_grid_desc_m_k_,
                                                    arg.scale_grid_desc_m_,
                                                    arg.bias_grid_desc_m_,
                                                    arg.mean_var_grid_desc_m_,
                                                    get_reduce_count_per_thread,
                                                    arg.numBlockTileIteration_,
                                                    arg.epsilon_,
                                                    arg.p_x_,
                                                    arg.p_scale_,
                                                    arg.p_bias_,
                                                    arg.y_elementwise_op_,
                                                    arg.p_y_,
                                                    arg.updateMovingAverage_, // true or false
                                                    arg.averageFactor_,
                                                    arg.resultRunningMean_,
                                                    arg.resultRunningVariance_,
                                                    arg.saveMeanInvVariance_, // true or false
                                                    arg.resultSaveMean_,
                                                    arg.resultSaveInvVariance_);
             };
  
             return (avg_time);
         };
  
         float Run(const BaseArgument* pArg,
                   const StreamConfig& stream_config = StreamConfig{}) override
         {
             return Run(*dynamic_cast<const Argument*>(pArg), stream_config);
         };
     };
  
     bool IsSupportedArgument(const BaseArgument* pArg) override
     {
         const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
  
         if constexpr(XSrcYDstVectorDim == 0)
         {
             if(pArg_->xStrides_[NumInvariantDim - 1] != 1 ||
                pArg_->yStrides_[NumInvariantDim - 1] != 1)
                 return false;
  
             if(pArg_->xyLengths_[NumInvariantDim - 1] % XSrcVectorSize != 0 ||
                pArg_->xyLengths_[NumInvariantDim - 1] % YDstVectorSize != 0)
                 return false;
         }
         else
         {
             if(pArg_->xStrides_[Rank - 1] != 1 || pArg_->yStrides_[Rank - 1] != 1)
                 return false;
  
             if(pArg_->xyLengths_[Rank - 1] % XSrcVectorSize != 0 ||
                pArg_->xyLengths_[Rank - 1] % YDstVectorSize != 0)
                 return false;
         };
  
         if(pArg_->bnScaleStrides_[NumInvariantDim - 1] != 1 && ScaleSrcVectorSize != 1)
             return false;
         if(pArg_->bnBiasStrides_[NumInvariantDim - 1] != 1 && BiasSrcVectorSize != 1)
             return false;
  
         if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % ScaleSrcVectorSize != 0)
             return false;
         if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % BiasSrcVectorSize != 0)
             return false;
  
         if(pArg_->bnMeanVarStrides_[NumInvariantDim - 1] != 1 && MeanVarSrcDstVectorSize != 1)
             return false;
  
         if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % MeanVarSrcDstVectorSize != 0)
             return false;
  
         bool is_valid = true;
  
         static_for<0, NumInvariantDim, 1>{}([&](auto I) {
             if(pArg_->xyLengths_[I] != pArg_->bnScaleBiasMeanVarLengths_[I])
                 is_valid = false;
         });
  
         if(!is_valid)
             return false;
  
         return true;
     };
  
     std::unique_ptr<BaseArgument> MakeArgumentPointer(
         const std::array<index_t, Rank> xyLengths,
         const std::array<index_t, Rank> xStrides,
         const std::array<index_t, Rank> yStrides,
         const std::array<int, NumBatchNormReduceDim> reduceDims,
         const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
         const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
         const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
         const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
         const void* p_x,
         const void* p_scale,
         const void* p_bias,
         double epsilon,
         const YElementwiseOp y_elementwise_op,
         void* p_y,
         void* resultSaveMean,
         void* resultSaveInvVariance,
         double averageFactor,
         void* resultRunningMean,
         void* resultRunningVariance) override
     {
         return std::make_unique<Argument>(xyLengths,
                                           xStrides,
                                           yStrides,
                                           reduceDims,
                                           bnScaleBiasMeanVarLengths,
                                           bnScaleStrides,
                                           bnBiasStrides,
                                           bnMeanVarStrides,
                                           static_cast<const XDataType*>(p_x),
                                           static_cast<const ScaleDataType*>(p_scale),
                                           static_cast<const BiasDataType*>(p_bias),
                                           y_elementwise_op,
                                           epsilon,
                                           static_cast<YDataType*>(p_y),
                                           static_cast<MeanVarDataType*>(resultSaveMean),
                                           static_cast<MeanVarDataType*>(resultSaveInvVariance),
                                           averageFactor,
                                           static_cast<MeanVarDataType*>(resultRunningMean),
                                           static_cast<MeanVarDataType*>(resultRunningVariance));
     };
  
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
         return std::make_unique<Invoker>();
     };
  
     std::string GetTypeString() const override
     {
         auto str = std::stringstream();
  
         // clang-format off
         str << "DeviceBatchNormFwdImpl<" << BlockSize << ",";
         str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
         str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
         str << "XSrcYDstVectorDim_" << XSrcYDstVectorDim  << ",";
         str << "VectorSize_X" << XSrcVectorSize << "_scale_" << ScaleSrcVectorSize << "_bias_" << BiasSrcVectorSize << "_mean_var_" << MeanVarSrcDstVectorSize << "_Y" << YDstVectorSize << ">";
         // clang-format on
  
         return str.str();
     }
 };
  
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck