From 621965e3e9ef301d2668c60702f5fb79daea8d26 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Mon, 8 Jan 2018 17:11:26 +0000 Subject: COMPMID-769 Add asymmetric padding support in NEON kernels. - NEDirectConvolutionLayer - NEDepthwiseConvolutionLayer3x3 Change-Id: Id4d7d17ee334639c059015a290b8fc34712706ee Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/115430 Tested-by: Jenkins Reviewed-by: Anthony Barbier --- src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) (limited to 'src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp') diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp index 3cdb39ef94..40a8601aaa 100644 --- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp +++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -77,13 +77,13 @@ void NEDepthwiseConvolutionLayer3x3Kernel::configure(const ITensor *input, const _conv_info = conv_info; const unsigned int conv_stride_x = conv_info.stride().first; const unsigned int conv_stride_y = conv_info.stride().second; - const unsigned int conv_pad_x = conv_info.pad().first; - const unsigned int conv_pad_y = conv_info.pad().second; + const unsigned int conv_pad_left = conv_info.pad_left(); + const unsigned int conv_pad_top = conv_info.pad_top(); ARM_COMPUTE_ERROR_ON(conv_stride_x < 1 || conv_stride_x > 3); const unsigned int num_elems_written_per_iteration = 16 >> conv_stride_x; - _border_size = BorderSize(conv_pad_y, conv_pad_x); + _border_size = BorderSize(conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), conv_pad_left); // Configure kernel window Window win = calculate_max_window(*output->info(), Steps(num_elems_written_per_iteration)); @@ -91,7 +91,7 @@ void NEDepthwiseConvolutionLayer3x3Kernel::configure(const ITensor *input, const const unsigned int num_x_steps = (output_shape.x() + num_elems_written_per_iteration - 1) / num_elems_written_per_iteration; const int input_num_elems_processed = get_input_num_elems_processed(num_elems_written_per_iteration, conv_stride_x); - AccessWindowStatic input_access(input->info(), -conv_pad_x, -conv_pad_y, (num_x_steps - 1) * input_num_elems_processed + 12, conv_stride_y * (output_shape.y() - 1) + 2); + AccessWindowStatic input_access(input->info(), -conv_pad_left, -conv_pad_top, (num_x_steps - 1) * input_num_elems_processed + 12, conv_stride_y * (output_shape.y() - 1) + 2); AccessWindowStatic weights_access(weights->info(), 0, 0, weights->info()->dimension(0), weights->info()->dimension(1)); AccessWindowStatic output_access(output->info(), 0, 0, num_x_steps * num_elems_written_per_iteration, output_shape.y()); -- cgit v1.2.1