diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-07-30 12:21:07 +0100 |
---|---|---|
committer | Sheri Zhang <sheri.zhang@arm.com> | 2021-08-04 11:04:22 +0000 |
commit | 663419457b02238687cb329afcddc73719bdb8fa (patch) | |
tree | 040abeb9878085df78841cc907036b9293e7a644 /src/core | |
parent | 4164814a099773c0a512889473c980bc148e590f (diff) | |
download | ComputeLibrary-663419457b02238687cb329afcddc73719bdb8fa.tar.gz |
Avoid over-allocation of temporary buffers within CpuWinogradConv2d
Resolves: COMPMID-4716
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: Ie036d2bb7a243301a62f089b3920ebee0f409190
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6028
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Diffstat (limited to 'src/core')
-rw-r--r-- | src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp | 13 |
1 files changed, 5 insertions, 8 deletions
diff --git a/src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp b/src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp index 74b031b226..5620d36e2c 100644 --- a/src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp +++ b/src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp @@ -195,8 +195,7 @@ unsigned int CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTi { const KernelShape shape(num_output_channels, KernelRows, KernelCols, num_input_channels); return static_cast<unsigned int>( - // WinogradConv returns the size in bytes, we divide by `sizeof(T)` to express that in units of T - WinogradConv::get_kernel_storage_size(num_input_channels, num_output_channels) / sizeof(T)); + WinogradConv::get_kernel_storage_size(num_input_channels, num_output_channels)); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> @@ -298,14 +297,13 @@ unsigned int CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTile // Construct shapes for the input and kernel tensors. const Tensor4DShape input_shape(num_batches, num_rows, num_cols, num_channels); const KernelShape kern_shape(1, KernelRows, KernelCols, num_channels); - // Return the size, converted into units of TIn - return static_cast<unsigned int>(WinogradConv::get_input_storage_size(num_batches, num_rows, num_cols, num_channels, same_padding) / sizeof(T)); + return static_cast<unsigned int>(WinogradConv::get_input_storage_size(num_batches, num_rows, num_cols, num_channels, same_padding)); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> unsigned int CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_working_space_size(unsigned int num_threads) const { - return _transform->get_working_space_size(num_threads) / sizeof(T); + return _transform->get_working_space_size(num_threads); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> @@ -434,9 +432,8 @@ unsigned int CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTil // Construct shapes for the input and kernel tensors. const Tensor4DShape input_shape(num_batches, num_rows, num_cols, 1); const KernelShape kern_shape(num_output_channels, KernelRows, KernelCols, 1); - // Return the size, converted into units of TOut return static_cast<unsigned int>( - WinogradConv::get_output_storage_size(num_batches, num_rows, num_cols, num_output_channels) / sizeof(T)); + WinogradConv::get_output_storage_size(num_batches, num_rows, num_cols, num_output_channels)); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> @@ -448,7 +445,7 @@ CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, Kernel template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> unsigned int CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_working_space_size(unsigned int num_threads) const { - return _transform->get_working_space_size(num_threads) / sizeof(T); + return _transform->get_working_space_size(num_threads); } template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> |