diff options
| author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2018-04-11 15:59:10 +0100 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:49:37 +0000 |
| commit | e52a3000d2c13bc1b66ca66b3d12b6b836982394 (patch) | |
| tree | 70e8ef5ba216762604f84228805aac9bd65747b6 /src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp | |
| parent | dd03870b63784abe499761da2b26b209b33f2db2 (diff) | |
| download | ComputeLibrary-e52a3000d2c13bc1b66ca66b3d12b6b836982394.tar.gz | |
COMPMID-1026 - Add support for 4x4 output tile in CLWinogradConvolutionLayer
The performance achieved can be found at the following confluence page:
https://confluence.arm.com/display/MLENG/GEMM-based+convolution+vs+Winograd-based+convolution+on+OpenCL
Change-Id: I4b690cfdd4eb4ff0cd17b14fdd49ccaa1d1dc85c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127729
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp')
| -rw-r--r-- | src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp | 8 |
1 files changed, 5 insertions, 3 deletions
diff --git a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp index d3a33c01f9..41b3ac50b5 100644 --- a/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp +++ b/src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp @@ -55,9 +55,11 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); - ARM_COMPUTE_RETURN_ERROR_ON(kernel_size != Size2D(3U, 3U) && kernel_size != Size2D(5U, 5U)); - ARM_COMPUTE_RETURN_ERROR_ON(kernel_size == Size2D(3U, 3U) && output_tile_size != Size2D(2U, 2U) && output_tile_size != Size2D(4U, 4U)); - ARM_COMPUTE_RETURN_ERROR_ON(kernel_size == Size2D(5U, 5U) && output_tile_size != Size2D(4U, 4U)); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size != Size2D(3U, 3U) && kernel_size != Size2D(5U, 5U), "Winograd filter transform only supports 3x3 and 5x5 kernels"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(3U, 3U) && output_tile_size != Size2D(2U, 2U) + && output_tile_size != Size2D(4U, 4U), + "Winograd filter transform only supports 2x2 or 4x4 output tile for 3x3 kernels"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(kernel_size == Size2D(5U, 5U) && output_tile_size != Size2D(4U, 4U), "Winograd filter transform only supports 4x4 output tile for 5x5 kernels"); ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_w) != kernel_size.width || input->dimension(idx_h) != kernel_size.height); ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); |