From 4074c995d2a88684fd4a9d1aa36d51de56bb8dab Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Tue, 30 Jan 2018 18:13:46 +0000 Subject: COMPMID-873: Integrate RSH NEON Depthwise Convolution routine Change-Id: Ida1e9a836bc518bfe5563e16bf7f92bde5fc13f7 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118472 Tested-by: Jenkins Reviewed-by: Pablo Tello --- .../convolution/winograd/transforms/input.hpp | 195 +++++++++++++++++++++ .../convolution/winograd/transforms/kernel.hpp | 77 ++++++++ .../convolution/winograd/transforms/output.hpp | 181 +++++++++++++++++++ 3 files changed, 453 insertions(+) create mode 100644 arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp create mode 100644 arm_compute/core/NEON/kernels/convolution/winograd/transforms/kernel.hpp create 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 new file mode 100644 index 0000000000..6dd8f5460a --- /dev/null +++ b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/input.hpp @@ -0,0 +1,195 @@ +/* + * 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 "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" + +namespace winograd +{ + /***************************************************************************/ + /* Instance-less API */ + template + template + void WinogradGEMM::InputTransform::execute( + const T *inptr, + const Tensor4DShape& input_shape, + const PaddingType padding_type, + const int tile_M, + const int tile_N, + T *outptr_base, + const int matrix_stride, + const int matrix_batch_stride, + const int matrix_row_stride + ) + { + // Compute the padding required on each edge of the image + const bool base_padding = (padding_type == PADDING_SAME) ? 1 : 0; + const int pad_top = base_padding; + const int pad_left = base_padding; + const int tile_overlap = kernel_rows - 1; + + // Compute striding values (assuming NHWC ordered data) + const int input_col_stride = input_shape.n_channels; + const int input_row_stride = input_shape.n_cols * input_col_stride; + const int input_batch_stride = input_shape.n_rows * input_row_stride; + 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 < input_shape.n_batches; batch++) + { + // Pointer to the batch + const T* const input_base_batch = inptr + batch * input_batch_stride; + T* const outptr_base_batch = outptr_base + batch * matrix_batch_stride; + + // Loop over rows of tiles + for (int tile_i = 0; tile_i < tile_M; tile_i++) + { + // Pointer to the row + const int row_offset = (tile_i == 0) ? + 0 : ((padding_type == PADDING_VALID) ? 0 : 1); + const T* const input_base_row = ( + input_base_batch + ((inner_tile_rows - (kernel_rows - 1))*tile_i - row_offset)*input_row_stride + ); + T* const outptr_base_row = outptr_base_batch + tile_i*output_row_stride; + + // Padding (top + bottom) for the row + const int row_top = tile_i*(inner_tile_rows - tile_overlap) - pad_top; + const int row_bottom = row_top + inner_tile_rows; + const int row_pad_top = (tile_i == 0) ? pad_top : 0; + const int row_pad_bottom = (row_bottom <= input_shape.n_rows) ? 0 : row_bottom - input_shape.n_rows; + + // Process the row + process_tile_row( + tile_N, input_shape.n_channels, + input_base_row, input_row_stride, input_col_stride, + outptr_base_row, matrix_stride, matrix_row_stride, + row_pad_top, pad_left, row_pad_bottom, input_shape.n_cols + ); + } + } + } + + template + template + void WinogradGEMM::InputTransform::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 + ) + { + constexpr int tile_overlap = kernel_cols - 1; + + // 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_pad_left = (tile_j == 0) ? row_pad_left : 0; + const int t_start = tile_j*(inner_tile_cols - tile_overlap) - row_pad_left; + const int t_end = t_start + inner_tile_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 = (tile_j == 0) ? 0 : row_pad_left; + const T* const input_base_col = ( + input_base + ((inner_tile_cols - tile_overlap)*tile_j - col_offset)*input_col_stride + ); + T* const outptr = matrix_base + tile_j*matrix_row_stride; + + // Apply the specific tile processing function + tile_fns[pad_top][t_pad_left][pad_bottom][t_pad_right]( + n_channels, + input_base_col, + input_row_stride, + input_col_stride, + outptr, + matrix_stride + ); + } + } + + /***************************************************************************/ + template + template + WinogradGEMM::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. */ + ) : _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 - 2, output_tile_rows)), + _tiles_N(iceildiv((padding == PADDING_SAME) ? n_cols : n_cols - 2, output_tile_cols)), + _padding_type(padding) + { + } + + template + template + unsigned int WinogradGEMM::InputTransform::get_window() const + { + // TODO When the input transform supports multithreading, return the total + // number of tile rows (allowing for multiple batches). For now we return 1 + // to indicate that the activations must be transformed as a single block. + return 1; // TODO _tiles_M * _n_batches; + } + + template + template + void WinogradGEMM::InputTransform::run( + const unsigned int start, const unsigned int stop + ) + { + // TODO When the input transform supports multithreading call execute for a + // portion of the tile rows. + (void) start; + (void) stop; + + // For now, just do all of the work. + const Tensor4DShape input_shape = { + _n_batches, _n_rows, _n_cols, _n_channels, NHWC + }; + execute( + _inptr, input_shape, _padding_type, _tiles_M, _tiles_N, _outptr, + _matrix_stride, _matrix_row_stride * _tiles_M * _tiles_N, _matrix_row_stride + ); + } +} diff --git a/arm_compute/core/NEON/kernels/convolution/winograd/transforms/kernel.hpp b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/kernel.hpp new file mode 100644 index 0000000000..bad3ef2249 --- /dev/null +++ b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/kernel.hpp @@ -0,0 +1,77 @@ +/* + * 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 new file mode 100644 index 0000000000..401b2816be --- /dev/null +++ b/arm_compute/core/NEON/kernels/convolution/winograd/transforms/output.hpp @@ -0,0 +1,181 @@ +/* + * 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 "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" + +namespace winograd +{ + template + template + void WinogradGEMM::OutputTransform::execute( + const Tensor4DShape &output_shape, + 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(output_shape.n_rows, output_tile_rows); + const int tile_N = iceildiv(output_shape.n_cols, output_tile_cols); + const int pad_bottom = output_tile_rows*tile_M - output_shape.n_rows; + const int pad_right = output_tile_cols*tile_N - output_shape.n_cols; + + const int matrix_tile_row_stride = tile_N * matrix_row_stride; + const int matrix_batch_stride = tile_M * matrix_tile_row_stride; + const int output_col_stride = output_shape.n_channels; + const int output_row_stride = output_shape.n_cols * output_col_stride; + const int output_batch_stride = output_shape.n_rows * output_row_stride; + + // Perform the output transformation for each batch + for (int batch = 0; batch < output_shape.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 + output_tile_rows*tile_i*output_row_stride; + + // Process the row + process_tile_row( + tile_N, output_shape.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 + template + void WinogradGEMM::OutputTransform::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 + output_tile_cols*tile_j*output_col_stride; + + // Perform the output transformation + tile_fns[row_pad_bottom][tile_pad_right]( + n_channels, matrix_row, matrix_stride, biases, + outptr, output_row_stride, output_col_stride + ); + } + } + + template + template + size_t WinogradGEMM::OutputTransform::bytes_read(const Tensor4DShape &shape) + { + const int M = iceildiv(shape.n_rows, output_tile_rows) * + iceildiv(shape.n_cols, output_tile_cols); + const int N = shape.n_channels; + return inner_tile_rows * inner_tile_cols * M * N * sizeof(T); + } + + template + template + size_t WinogradGEMM::OutputTransform::bytes_written(const Tensor4DShape &shape) + { + return shape.size() * sizeof(T); + } + + template + template + WinogradGEMM::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 + ) : _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, output_tile_rows)), + _tile_N(iceildiv(n_cols, output_tile_cols)) + { + } + + template + template + unsigned int WinogradGEMM::OutputTransform::get_window() const + { + // TODO When the output transform supports multithreading, return the total + // number of tile rows (allowing for multiple batches). For now we return 1 + // to indicate that the activations must be transformed as a single block. + return 1; // TODO _tile_M * _n_batches; + } + + template + template + void WinogradGEMM::OutputTransform::run( + const unsigned int start, const unsigned int stop + ) + { + // TODO When the output transform supports multithreading call execute for a + // portion of the tile rows. + (void) start; + (void) stop; + + // For now, just do all of the work. + const Tensor4DShape output_shape = { + _n_batches, _n_rows, _n_cols, _n_channels, NHWC + }; + execute( + output_shape, _matrix_base, _matrix_stride, _matrix_row_stride, _biases, + _outptr + ); + } +} // namespace winograd -- cgit v1.2.1