From e52a3000d2c13bc1b66ca66b3d12b6b836982394 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Wed, 11 Apr 2018 15:59:10 +0100 Subject: 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 Reviewed-by: Georgios Pinitas --- src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) (limited to 'src/core/CL/kernels/CLWinogradFilterTransformKernel.cpp') 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); -- cgit v1.2.1