From 8f43d745b170aefca269a087fc045d8af3813c33 Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Wed, 27 Mar 2019 09:28:32 +0000 Subject: COMPMID-2063: New Winograd implementation Refactoring of winograd code reducing the size of the binaries about 8X. Change-Id: If8845bda324573e1a5cf436f354ac8603e88a92e Signed-off-by: Pablo Tello Reviewed-on: https://review.mlplatform.org/c/959 Comments-Addressed: Arm Jenkins Tested-by: Anthony Barbier Reviewed-by: Georgios Pinitas --- .../convolution/winograd/transforms/input.hpp | 349 --------------------- .../convolution/winograd/transforms/kernel.hpp | 77 ----- .../convolution/winograd/transforms/output.hpp | 278 ---------------- 3 files changed, 704 deletions(-) delete mode 100644 arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp delete mode 100644 arm_compute/core/NEON/kernels/convolution/winograd/transforms/kernel.hpp delete mode 100644 arm_compute/core/NEON/kernels/convolution/winograd/transforms/output.hpp (limited to 'arm_compute/core/NEON/kernels/convolution/winograd/transforms') diff --git a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp deleted file mode 100644 index b813bbb25c..0000000000 --- a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp +++ /dev/null @@ -1,349 +0,0 @@ -/* - * Copyright (c) 2017 ARM Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ - -#pragma once -#include "../winograd_gemm.hpp" - -namespace winograd -{ - /***************************************************************************/ - /* Instance-less API */ - template - void InputTransformImpl::execute( - const T* const input, /** Input tensor data */ - const int n_batches, /** Number of batches in input tensor. */ - const int in_batch_stride, /** Stride between batches of the input. */ - const int n_rows, /** Number of rows in input tensor. */ - const int in_row_stride, /** Stride between rows of the input. */ - const int n_cols, /** Number of columns in input tensor. */ - const int in_col_stride, /** Stride between columns of the input. */ - const int n_channels, /** Number of channels in input tensor. */ - const PaddingType padding, /** Padding type. */ - const int tile_M, - const int tile_N, - T* const output, /** Base of output matrices. */ - const int matrix_stride, /** Stride between output matrices. */ - const int matrix_batch_stride, /** Stride between batches within the matrix. */ - const int matrix_row_stride /** Stride within matrices. */ - ) - { - // Compute the padding required on each edge of the image - const int pad_top = (padding == PADDING_SAME) ? (KernelRows - 1) / 2 : 0; - const int pad_left = (padding == PADDING_SAME) ? (KernelCols - 1) / 2 : 0; - - // Compute striding values (assuming NHWC ordered data) - const int output_col_stride = matrix_row_stride; - const int output_row_stride = tile_N * output_col_stride; - - // Loop over batches - for (int batch = 0; batch < n_batches; batch++) - { - // Pointer to the batch - const T* const input_base_batch = input + batch * in_batch_stride; - T* const outptr_base_batch = output + batch * matrix_batch_stride; - - // Loop over rows of tiles - for (int tile_i = 0; tile_i < tile_M; tile_i++) - { - // Padding (top + bottom) for the row - const int row_top = tile_i*(InnerTileRows - overlap_rows) - pad_top; - const int row_bottom = row_top + InnerTileRows; - const int row_pad_top = std::max(0, pad_top - tile_i*(InnerTileRows - overlap_rows)); - const int row_pad_bottom = (row_bottom <= n_rows) ? 0 : row_bottom - n_rows; - - // Pointer to the row - const int row_offset = std::min(0, row_pad_top - pad_top); - const T* const input_base_row = ( - input_base_batch + ((InnerTileRows - overlap_rows)*tile_i + row_offset)*in_row_stride - ); - T* const outptr_base_row = outptr_base_batch + tile_i*output_row_stride; - - // Process the row - process_tile_row( - tile_N, n_channels, - input_base_row, in_row_stride, in_col_stride, - outptr_base_row, matrix_stride, matrix_row_stride, - row_pad_top, pad_left, row_pad_bottom, n_cols - ); - } - } - } - - - template - void InputTransformImpl::execute( - const T* const input, /** Input tensor data */ - const int n_batches, /** Number of batches in input tensor. */ - const int in_batch_stride, /** Stride between batches of the input. */ - const int n_rows, /** Number of rows in input tensor. */ - const int in_row_stride, /** Stride between rows of the input. */ - const int n_cols, /** Number of columns in input tensor. */ - const int in_col_stride, /** Stride between columns of the input. */ - const int n_channels, /** Number of channels in input tensor. */ - const PaddingType padding, /** Padding type. */ - const int tile_M, - const int tile_N, - T* const output, /** Base of output matrices. */ - const int matrix_stride, /** Stride between output matrices. */ - const int matrix_batch_stride, /** Stride between batches within the matrix. */ - const int matrix_row_stride /** Stride within matrices. */ - ) - { - // If an Nx1 kernel then transpose and redirect to the 1xN implementation - InputTransformImpl<1, KernelRows, 1, InnerTileRows, T>::execute( - input, - n_batches, in_batch_stride, - n_cols, in_col_stride, - n_rows, in_row_stride, - n_channels, padding, - tile_N, tile_M, - output, matrix_stride, matrix_batch_stride, matrix_row_stride - ); - } - - template - void InputTransformImpl::process_tile_row( - const int tile_N, - int n_channels, - const T* const input_base, - const int input_row_stride, - const int input_col_stride, - T* const matrix_base, - const int matrix_stride, - const int matrix_row_stride, - const int pad_top, - const int row_pad_left, - const int pad_bottom, - const int n_cols - ) - { - // Loop over columns of tiles - for (int tile_j = 0; tile_j < tile_N; tile_j++) - { - // Padding (left + right) for the tile - const int t_start = tile_j*(InnerTileCols - overlap_cols) - row_pad_left; - const int t_end = t_start + InnerTileCols; - const int t_pad_left = std::max(0, row_pad_left - tile_j*(InnerTileCols - overlap_cols)); - const int t_pad_right = (t_end <= n_cols) ? 0 : t_end - n_cols; - - // Get pointers into the inputs and outputs - const int col_offset = std::min(0, t_pad_left - row_pad_left); - const T* const input_base_col = ( - input_base + ((InnerTileCols - overlap_cols)*tile_j + col_offset)*input_col_stride - ); - T* const outptr = matrix_base + tile_j*matrix_row_stride; - - // Apply the specific tile processing function - const typename Tiles::TileFn tilefn = Tiles::get_tile_specialization( - pad_top, t_pad_left, pad_bottom, t_pad_right - ); - - tilefn( - n_channels, - input_base_col, input_row_stride, input_col_stride, - outptr, matrix_stride, - pad_top, t_pad_left, pad_bottom, t_pad_right - ); - } - } - - /***************************************************************************/ - template - InputTransform::InputTransform( - const T* const input, /** Input tensor data */ - const int n_batches, /** Number of batches in input tensor. */ - const int n_rows, /** Number of rows in input tensor. */ - const int n_cols, /** Number of columns in input tensor. */ - const int n_channels, /** Number of channels in input tensor. */ - const PaddingType padding, /** Padding type. */ - T* const output, /** Base of output matrices. */ - const int matrix_stride, /** Stride between output matrices. */ - const int matrix_row_stride, /** Stride within matrices. */ - const int in_batch_stride, /** Stride between input batches. */ - const int in_row_stride, /** Stride between input rows. */ - const int in_col_stride /** Stride between input columns. */ - ) : _inptr(input), _outptr(output), - _n_batches(n_batches), _n_rows(n_rows), _n_cols(n_cols), _n_channels(n_channels), - _matrix_stride(matrix_stride), _matrix_row_stride(matrix_row_stride), - _tiles_M(iceildiv((padding == PADDING_SAME) ? n_rows : n_rows - KernelRows + 1, - InnerTileRows - KernelRows + 1)), - _tiles_N(iceildiv((padding == PADDING_SAME) ? n_cols : n_cols - KernelCols + 1, - InnerTileCols - KernelCols + 1)), - _in_col_stride(in_col_stride ? in_col_stride : n_channels), - _in_row_stride(in_row_stride ? in_row_stride : n_cols * _in_col_stride), - _in_batch_stride(in_batch_stride ? in_batch_stride : n_rows * _in_row_stride), - _padding_type(padding) - { - } - - template - unsigned int InputTransform::get_window() const - { - // The final window includes the tail, all other windows will be a multiple - // of the window block in size. - return iceildiv(_n_channels, WINDOW_BLOCK); - } - - template - void InputTransform::run( - const unsigned int start, const unsigned int stop - ) - { - if (start >= get_window()) - { - return; - } - - // Determine the window of work to perform - const unsigned int start_channel = start * WINDOW_BLOCK; - const unsigned int stop_channel = std::min( - stop * WINDOW_BLOCK, _n_channels - ); - const unsigned int n_channels = stop_channel - start_channel; - - // Perform the work - execute( - _inptr + start_channel, - _n_batches, _in_batch_stride, - _n_rows, _in_row_stride, - _n_cols, _in_col_stride, - n_channels, - _padding_type, - _tiles_M, - _tiles_N, - _outptr + start_channel, - _matrix_stride, - _matrix_row_stride * _tiles_M * _tiles_N, - _matrix_row_stride - ); - } - - template - void InputTransform::execute( - const T* const input, /** Input tensor data */ - const int n_batches, /** Number of batches in input tensor. */ - const int in_batch_stride, /** Stride between batches of the input. */ - const int n_rows, /** Number of rows in input tensor. */ - const int in_row_stride, /** Stride between rows of the input. */ - const int n_cols, /** Number of columns in input tensor. */ - const int in_col_stride, /** Stride between columns of the input. */ - const int n_channels, /** Number of channels in input tensor. */ - const PaddingType padding, /** Padding type. */ - const int tile_M, - const int tile_N, - T* const output, /** Base of output matrices. */ - const int matrix_stride, /** Stride between output matrices. */ - const int matrix_batch_stride, /** Stride between batches within the matrix. */ - const int matrix_row_stride /** Stride within matrices. */ - ) - { - Transform::execute( - input, n_batches, in_batch_stride, n_rows, in_row_stride, n_cols, - in_col_stride, n_channels, padding, tile_M, tile_N, output, - matrix_stride, matrix_batch_stride, matrix_row_stride - ); - } - - template - typename InputTransformImplTiles::TileFn - InputTransformImplTiles:: - get_tile_specialization( - const int pad_top, - const int pad_left, - const int pad_bottom, - const int pad_right - ) - { - if (!(pad_top || pad_left || pad_bottom || pad_right)) - { - // No padding, return unpadded specialisation - return tilefn_unpadded; - } - else if (pad_top && !(pad_left || pad_bottom || pad_right)) - { - // Top padding only - const int index = (pad_top - min_pad_top) / (InnerTileRows - overlap_rows); - return tilefn_top_padded[index]; - } - else if (!(pad_top) && pad_left && !(pad_bottom || pad_right)) - { - // Left padding only - const int index = (pad_left - min_pad_left) / (InnerTileCols - overlap_cols); - return tilefn_left_padded[index]; - } - else if (!(pad_top || pad_left) && pad_bottom && !(pad_right)) - { - // Bottom padding only - return tilefn_bottom_padded[pad_bottom - 1]; - } - else if (!(pad_top || pad_left || pad_bottom) && pad_right) - { - // Right padding only - return tilefn_right_padded[pad_right - 1]; - } - else - { - // Combination of paddings, return an unspecialised method - return tilefn_generic; - } - } - - template - typename InputTransformImplTiles<1, KernelCols, 1, InnerTileCols, T>::TileFn - InputTransformImplTiles<1, KernelCols, 1, InnerTileCols, T>:: - get_tile_specialization( - const int pad_top, - const int pad_left, - const int pad_bottom, - const int pad_right - ) - { - (void) pad_top; - (void) pad_bottom; - - if (!(pad_left || pad_right)) - { - // No padding, return unpadded specialisation - return tilefn_unpadded; - } - else if (pad_left && !pad_right) - { - // Left padding only - const int index = (pad_left - min_pad_left) / (InnerTileCols - overlap_cols); - return tilefn_left_padded[index]; - } - else if (!pad_left && pad_right) - { - // Right padding only - return tilefn_right_padded[pad_right - 1]; - } - else - { - // Combination of paddings, return an unspecialised method - return tilefn_generic; - } - } -} - - diff --git a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/kernel.hpp b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/kernel.hpp deleted file mode 100644 index bad3ef2249..0000000000 --- a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/kernel.hpp +++ /dev/null @@ -1,77 +0,0 @@ -/* - * Copyright (c) 2017 ARM Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ - -#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" -using namespace winograd; - - -template -template -WinogradGEMM::WeightsTransform::WeightsTransform( - const T* const input, - T* const output, - const int matrix_stride, /** Stride across matrices in the output. */ - const int matrix_row_stride, /** Stride across rows of the matrix. */ - const int n_output_channels, - const int n_input_channels -) : inptr(input), outptr(output), - matrix_stride(matrix_stride), matrix_row_stride(matrix_row_stride), - n_output_channels(n_output_channels), n_input_channels(n_input_channels) -{ -} - - -template -template -unsigned int WinogradGEMM::WeightsTransform::get_window() const -{ - // TODO When the weights transform supports multithreading, return the number - // of output channels. For now we return 1 to indicate that the weights must - // be transformed as a single block. - // return n_output_channels; - return 1; -} - - -template -template -void WinogradGEMM::WeightsTransform::run( - const unsigned int start, const unsigned int stop -) -{ - // TODO When the weights transform supports multithreading call execute for a - // portion of the output channels. - (void) start; - (void) stop; - - // For now, just do all of the work. - execute( - n_output_channels, - n_input_channels, - inptr, - outptr, - matrix_stride, - matrix_row_stride - ); -} diff --git a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/output.hpp b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/output.hpp deleted file mode 100644 index 77cd9de513..0000000000 --- a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/output.hpp +++ /dev/null @@ -1,278 +0,0 @@ -/* - * Copyright (c) 2017 ARM Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ - -#pragma once -#include "../winograd_gemm.hpp" - -namespace winograd -{ -/***************************************************************************/ - /* Instance-less API */ - template - void OutputTransformImpl::execute( - const int n_batches, - const int output_batch_stride, - const int n_rows, - const int output_row_stride, - const int n_cols, - const int output_col_stride, - const int n_channels, - const T* const matrix_base, - const int matrix_stride, - const int matrix_row_stride, - const T* const biases, - T* const output - ) - { - // Compute the number of tiles and hence the padding required on the bottom - // and right of the image. - const int tile_M = iceildiv(n_rows, OutputTileRows); - const int tile_N = iceildiv(n_cols, OutputTileCols); - const int pad_bottom = OutputTileRows*tile_M - n_rows; - const int pad_right = OutputTileCols*tile_N - n_cols; - - const int matrix_tile_row_stride = tile_N * matrix_row_stride; - const int matrix_batch_stride = tile_M * matrix_tile_row_stride; - - // Perform the output transformation for each batch - for (int batch = 0; batch < n_batches; batch++) - { - // Get batch offset for input and outputs. - const T* const matrix_batch = matrix_base + batch*matrix_batch_stride; - T* const outptr_batch = output + batch*output_batch_stride; - - // Perform the output transformation for each row of the output tensor. - for (int tile_i = 0; tile_i < tile_M; tile_i++) - { - // Compute properties of this row of output tiles - const int row_pad_bottom = (tile_i < tile_M - 1) ? 0: pad_bottom; - const T* const matrix_tile_row = matrix_batch + tile_i * matrix_tile_row_stride; - T* const outptr_row = outptr_batch + OutputTileRows*tile_i*output_row_stride; - - // Process the row - process_tile_row( - tile_N, n_channels, matrix_tile_row, matrix_stride, - matrix_row_stride, biases, - outptr_row, output_row_stride, output_col_stride, row_pad_bottom, - pad_right - ); - } - } - } - -template - void OutputTransformImpl::execute( - const int n_batches, - const int output_batch_stride, - const int n_rows, - const int output_row_stride, - const int n_cols, - const int output_col_stride, - const int n_channels, - const T* const matrix_base, - const int matrix_stride, - const int matrix_row_stride, - const T* const biases, - T* const output - ) - { - // If an Nx1 kernel then transpose and redirect to the 1xN implementation. - OutputTransformImpl<1, KernelRows, 1, InnerTileRows, T>::execute( - n_batches, - output_batch_stride, - n_cols, output_col_stride, - n_rows, output_row_stride, - n_channels, - matrix_base, matrix_stride, matrix_row_stride, - biases, output - ); - } - - template - void OutputTransformImpl::process_tile_row( - const int tile_N, - const int n_channels, - const T* const matrix_base, - const int matrix_stride, - const int matrix_row_stride, - const T* const biases, - T* const output, - const int output_row_stride, - const int output_col_stride, - const int row_pad_bottom, - const int row_pad_right - ) - { - // Loop over columns of tiles - for (int tile_j = 0; tile_j < tile_N; tile_j++) - { - // Properties of this tile - const int tile_pad_right = (tile_j < tile_N - 1) ? 0 : row_pad_right; - const T* const matrix_row = matrix_base + tile_j * matrix_row_stride; - T* const outptr = output + OutputTileCols *tile_j*output_col_stride; - - // Perform the output transformation - const typename Tiles::TileFn tilefn = Tiles::get_tile_specialization(row_pad_bottom, tile_pad_right); - tilefn( - n_channels, matrix_row, matrix_stride, biases, - outptr, output_row_stride, output_col_stride, - row_pad_bottom, tile_pad_right - ); - } - } - -/***************************************************************************/ - template - OutputTransform::OutputTransform( - const T* const matrix_base, - const int matrix_stride, - const int matrix_row_stride, - const T* const biases, - T* const output, - const int n_batches, - const int n_rows, - const int n_cols, - const int n_channels, - const int out_batch_stride, - const int out_row_stride, - const int out_col_stride - ) : _matrix_base(matrix_base), _biases(biases), - _matrix_stride(matrix_stride), _matrix_row_stride(matrix_row_stride), - _outptr(output), _n_batches(n_batches), _n_rows(n_rows), _n_cols(n_cols), - _n_channels(n_channels), _tile_M(iceildiv(n_rows, OutputTileRows)), - _tile_N(iceildiv(n_cols, OutputTileCols)), - _out_col_stride(out_col_stride ? out_col_stride : n_channels), - _out_row_stride(out_row_stride ? out_row_stride : n_cols * _out_col_stride), - _out_batch_stride(out_batch_stride ? out_batch_stride : n_rows * _out_row_stride) - { - } - - template - unsigned int OutputTransform::get_window() const - { - // The final window includes the tail, all other windows will be a multiple - // of the window block in size. - return iceildiv(_n_channels, WINDOW_BLOCK); - } - -template - void OutputTransform::run( - const unsigned int start, const unsigned int stop - ) - { - if (start >= get_window()) - { - return; - } - - // Determine the window of work to perform - const unsigned int start_channel = start * WINDOW_BLOCK; - const unsigned int stop_channel = std::min( - stop * WINDOW_BLOCK, _n_channels - ); - const unsigned int n_channels = stop_channel - start_channel; - - execute( - _n_batches, - _out_batch_stride, - _n_rows, - _out_row_stride, - _n_cols, - _out_col_stride, - n_channels, - _matrix_base + start_channel, - _matrix_stride, - _matrix_row_stride, - (_biases != nullptr) ? _biases + start_channel : nullptr, - _outptr + start_channel - ); - } - - template - void OutputTransform::execute( - const int n_batches, - const int out_batch_stride, - const int n_rows, - const int out_row_stride, - const int n_cols, - const int out_col_stride, - const int n_channels, - const T* const matrix_base, - const int matrix_stride, - const int matrix_row_stride, - const T* const biases, - T* const output - ) - { - Transform::execute( - n_batches, out_batch_stride, - n_rows, out_row_stride, - n_cols, out_col_stride, n_channels, - matrix_base, matrix_stride, matrix_row_stride, - biases, output - ); - } - - template - typename OutputTransformImplTiles<1, KernelCols, 1, InnerTileCols, T>::TileFn - OutputTransformImplTiles<1, KernelCols, 1, InnerTileCols, T>:: - get_tile_specialization(const int pad_bottom, const int pad_right) - { - (void) pad_bottom; - - if (!pad_right) - { - // No padding, return unpadded specialisation - return tilefn_unpadded; - } - else - { - return tilefn_right_padded[pad_right - 1]; - } - } - - template - typename OutputTransformImplTiles::TileFn - OutputTransformImplTiles:: - get_tile_specialization(const int pad_bottom, const int pad_right) - { - if (!(pad_bottom || pad_right)) - { - // No padding, return unpadded specialisation - return tilefn_unpadded; - } - else if (pad_bottom && !pad_right) - { - return tilefn_bottom_padded[pad_bottom - 1]; - } - else if (!pad_bottom && pad_right) - { - return tilefn_right_padded[pad_right - 1]; - } - else - { - return tilefn_generic; - } - } -} // namespace winograd -- cgit v1.2.1