From 9fb1159e2501f276a27d32264bece54b3d42d258 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Thu, 26 Apr 2018 20:34:58 +0100 Subject: COMPMID-1074: Rename WinograLayer.cpp to WinogradConvolutionLayer.cpp Change-Id: Iccac7cd6cb458469568d0cd6fb36b262353f4188 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/129261 Tested-by: Jenkins Reviewed-by: Pablo Tello --- .../kernels/NEWinogradConvolutionLayerKernel.cpp | 561 +++++++++++++++++++++ src/core/NEON/kernels/NEWinogradLayerKernel.cpp | 561 --------------------- src/graph/backends/NEON/NEFunctionFactory.cpp | 4 +- src/graph/backends/NEON/NENodeValidator.cpp | 2 +- src/runtime/NEON/functions/NEConvolutionLayer.cpp | 4 +- .../NEON/functions/NEWinogradConvolutionLayer.cpp | 390 ++++++++++++++ src/runtime/NEON/functions/NEWinogradLayer.cpp | 390 -------------- 7 files changed, 956 insertions(+), 956 deletions(-) create mode 100644 src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp delete mode 100644 src/core/NEON/kernels/NEWinogradLayerKernel.cpp create mode 100644 src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp delete mode 100644 src/runtime/NEON/functions/NEWinogradLayer.cpp (limited to 'src') diff --git a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp new file mode 100644 index 0000000000..fa76194529 --- /dev/null +++ b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp @@ -0,0 +1,561 @@ +/* + * Copyright (c) 2017-2018 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/NEWinogradConvolutionLayerKernel.h" + +#include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/IAccessWindow.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "support/ToolchainSupport.h" + +namespace arm_compute +{ +//Batched Gemms + +namespace +{ +Status validate_arguments_winograd_gemm(const ITensorInfo *a, const ITensorInfo *b, const ITensor *c, const ITensorInfo *output, const float alpha, const float beta, + const GEMMInfo &gemm_info = GEMMInfo()) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(a); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(b); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported"); + + if(c != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, c->info()); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != c->info()->dimension(1), "The matrix C must have the same number of rows as the matrix A"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != c->info()->dimension(0), "The matrix C must have the same number of columns as the matrix B"); + } + + if(output->total_size() != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, output); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != output->dimension(0), "The output matrix must have the same number of columns as the matrix B"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != output->dimension(1), "The output matrix must have the same number of rows as the matrix A"); + ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() != a->num_dimensions()); + } + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(0) != b->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); + ARM_COMPUTE_UNUSED(alpha, beta); + return Status{}; +} + +Status validate_arguments_winograd_weight_trans(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + + const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); + const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != 3 && input->dimension(idx_width) != 5); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != input->dimension(idx_height)); + ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); + const Size2D &output_tile = winograd_info.output_tile_size; + ARM_COMPUTE_RETURN_ERROR_ON(output_tile != Size2D(2U, 2U) && output_tile != Size2D(4U, 4U)); + + // Checks performed when output is configured + if(output->total_size() != 0) + { + const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_filter_transform_shape(*input, winograd_info)); + + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + } + + return Status{}; +} + +std::pair validate_and_configure_window_winograd_weight_trans(ITensorInfo *input, ITensorInfo *output, const WinogradInfo &winograd_info) +{ + const Size2D kernel_dims = winograd_info.kernel_size; + // Output tensor auto inizialitation if not yet initialized + auto_init_if_empty(*output, input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_filter_transform_shape(*input, winograd_info))); + + unsigned int num_elems_processed_per_iteration_x = kernel_dims.width; + unsigned int num_elems_processed_per_iteration_y = kernel_dims.height; + + Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); + bool window_changed = false; + + AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); + AccessWindowStatic output_access(output, 0, 0, output->dimension(0), output->dimension(1)); + window_changed = update_window_and_padding(win, input_access, output_access); + output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape())); + + Window win_collapsed = win.collapse(win, Window::DimZ); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + + return std::make_pair(err, win_collapsed); +} + +Status validate_arguments_winograd_input_trans(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) +{ + const Size2D &kernel_dims = winograd_info.kernel_size; + const PadStrideInfo &conv_info = winograd_info.convolution_info; + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd input transform only supports unit strides"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd input transform only supports 3x3 and 5x5 kernels"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd input transform only supports 3x3 and 5x5 kernels"); + + // Validate configured output + if(output->total_size() != 0) + { + const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); + + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + } + + return Status{}; +} + +std::pair validate_and_configure_window_winograd_input_trans(ITensorInfo *input, ITensorInfo *output, const WinogradInfo &winograd_info) +{ + const PadStrideInfo conv_info = winograd_info.convolution_info; + const Size2D output_tile_size = winograd_info.output_tile_size; + const Size2D kernel_dims = winograd_info.kernel_size; + const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); + // Output auto inizialitation if not yet initialized + auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape)); + + unsigned int num_elems_read_per_iteration_x = (output_tile_size.width + kernel_dims.width - 1); + unsigned int num_elems_read_per_iteration_y = (output_tile_size.height + kernel_dims.height - 1); + + Window win = calculate_max_window(*input, Steps(1, 1)); + + AccessWindowRectangle input_access(input, -conv_info.pad_left(), -conv_info.pad_top(), num_elems_read_per_iteration_x, num_elems_read_per_iteration_y); + + bool window_changed = update_window_and_padding(win, input_access); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} + +Status validate_arguments_winograd_output_trans(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info) +{ + const PadStrideInfo &conv_info = winograd_info.convolution_info; + const Size2D kernel_dims = winograd_info.kernel_size; + + // Number of tiles along the X and Y direction + const unsigned int num_tiles_x = std::ceil((winograd_info.input_dimensions.x() - (kernel_dims.width - 1) + conv_info.pad_left() + conv_info.pad_right()) / 2.f); + const unsigned int num_tiles_y = std::ceil((winograd_info.input_dimensions.y() - (kernel_dims.height - 1) + conv_info.pad_top() + conv_info.pad_bottom()) / 2.f); + const Size2D num_tiles = Size2D(num_tiles_x, num_tiles_y); + + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(winograd_info.output_data_layout != DataLayout::NCHW); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area()); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd output transform only supports 3x3 and 5x5 kernels"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd output transform only supports 3x3 and 5x5 kernels"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((input->dimension(2) != size_t(16U)) && (input->dimension(2) != size_t(36U))), "Only 2x2 and 4x4 output tile is supported"); + ARM_COMPUTE_UNUSED(kernel_dims); + if(bias != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0)); + ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() != size_t(1)); + } + + // Checks performed when output is configured + if(output->total_size() != 0) + { + const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_output_transform_shape(*input, winograd_info)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + } + return Status{}; +} + +std::pair validate_and_configure_window_winograd_output_trans(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output, const WinogradInfo &winograd_info) +{ + // Output tensor auto initialization if not yet initialized + auto_init_if_empty(*output, input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_output_transform_shape(*input, winograd_info))); + + constexpr unsigned int num_elems_processed_per_iteration = 1; + + Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); + bool window_changed = false; + + AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration); + AccessWindowStatic output_access(output, 0, 0, ceil_to_multiple(output->dimension(0), 2), ceil_to_multiple(output->dimension(1), 2)); + + if(bias != nullptr) + { + AccessWindowStatic bias_access(bias, 0, 0, bias->dimension(0), bias->dimension(1)); + window_changed = update_window_and_padding(win, input_access, bias_access, output_access); + } + else + { + window_changed = update_window_and_padding(win, input_access, output_access); + } + output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape())); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} +} // namespace +template +NEWinogradLayerBatchedGEMMKernel::NEWinogradLayerBatchedGEMMKernel() + : _gemms() +{ +} + +template +void NEWinogradLayerBatchedGEMMKernel::configure( + const unsigned int n_gemms, + const int M, const int K, const int N, + const int a_matrix_stride, + const int a_row_stride, + const int b_matrix_stride, + const int b_row_stride, + const int c_matrix_stride, + const int c_row_stride, + const TIn *const a_ptr, + const TIn *const b_ptr, + TOut *const c_ptr) +{ + _gemms = support::cpp14::make_unique(n_gemms, M, K, N, a_matrix_stride, a_row_stride, b_matrix_stride, b_row_stride, c_matrix_stride, c_row_stride, a_ptr, b_ptr, c_ptr); + Window win; + auto win_last = _gemms->get_window(); + win.set(Window::DimX, Window::Dimension(0, win_last, 1)); + INEKernel::configure(win); +} + +template +void NEWinogradLayerBatchedGEMMKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + const size_t first_gemm = window.x().start(); + const size_t last_gemm = window.x().end(); + _gemms->run(first_gemm, last_gemm); +} + +template +unsigned int NEWinogradLayerBatchedGEMMKernel::get_number_gemms() const +{ + return WinogradBase::N_GEMMS; +} + +template +int NEWinogradLayerBatchedGEMMKernel::get_output_tile_rows() const +{ + return _output_tile_rows; +} + +template +int NEWinogradLayerBatchedGEMMKernel::get_output_tile_cols() const +{ + return _output_tile_cols; +} + +template +int NEWinogradLayerBatchedGEMMKernel::get_number_blocks() const +{ + return WinogradConv::N_BLOCK; +} + +template +Status NEWinogradLayerBatchedGEMMKernel::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensor *c, + const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_gemm(a, b, c, output, alpha, beta, gemm_info)); + return Status{}; +} + +template class NEWinogradLayerBatchedGEMMKernel; +template class NEWinogradLayerBatchedGEMMKernel; + +// Weights transform + +template +unsigned int NEWinogradLayerTransformWeightsKernel::get_weight_storage_size(int n_output_channels, int n_input_channels) const +{ + const KernelShape shape(n_output_channels, KernelRows, KernelCols, n_input_channels); + return static_cast( + // WinogradConv returns the size in bytes, we divide by `sizeof(T)` to express that in units of T + WinogradConv::get_kernel_storage_size(shape) / sizeof(T)); +} + +template +NEWinogradLayerTransformWeightsKernel::NEWinogradLayerTransformWeightsKernel() + : _transform() +{ +} + +template +int NEWinogradLayerTransformWeightsKernel::get_matrix_stride(const KernelShape &kernel_shape) const +{ + return WinogradConv::get_kernel_matrix_stride(kernel_shape); +} + +template +void NEWinogradLayerTransformWeightsKernel::configure( + const ITensor *weights_hwio, + T *const output, + const int matrix_stride, /** Stride across matrices in the output. */ + const int n_output_channels, /** Number of filters. */ + const int n_input_channels) /** Number of channels in each filter. */ +{ + const int matrix_row_stride = roundup(n_output_channels, WinogradConv::N_BLOCK); + _transform = support::cpp14::make_unique(reinterpret_cast(weights_hwio->buffer()), output, matrix_stride, matrix_row_stride, n_output_channels, + n_input_channels); + Window win; + auto win_last = _transform->get_window(); + win.set(Window::DimX, Window::Dimension(0, win_last, 1)); + INEKernel::configure(win); +} + +template +void NEWinogradLayerTransformWeightsKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + const size_t fst = window.x().start(); + const size_t lst = window.x().end(); + _transform->run(fst, lst); +} + +template +bool NEWinogradLayerTransformWeightsKernel::is_parallelisable() const +{ + return false; +} + +template +Status NEWinogradLayerTransformWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, + const WinogradInfo &winograd_info) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_weight_trans(input, output, winograd_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_weight_trans(input->clone().get(), output->clone().get(), winograd_info).first); + return Status{}; +} + +template class NEWinogradLayerTransformWeightsKernel; +template class NEWinogradLayerTransformWeightsKernel; + +// Input transform + +template +unsigned int NEWinogradLayerTransformInputKernel::get_input_storage_size( + int n_batches, /** Number of batches in the input tensor. */ + int n_channels, /** Number of feature maps in the input tensor. */ + int n_rows, /** Number of rows in each feature map. */ + int n_cols, /** Number of columns in each feature map. */ + bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ +) const +{ + // Construct shapes for the input and kernel tensors. + const Tensor4DShape input_shape(n_batches, n_rows, n_cols, n_channels); + const KernelShape kern_shape(1, KernelRows, KernelCols, n_channels); + const PaddingType padding = (same_padding) ? PADDING_SAME : PADDING_VALID; + // Return the size, converted into units of TIn + return static_cast(WinogradConv::get_input_storage_size(kern_shape, input_shape, padding) / sizeof(T)); +} + +template +int NEWinogradLayerTransformInputKernel::get_matrix_stride( + const KernelShape &kernel_shape, const Tensor4DShape &input_shape, const PaddingType padding_type) const +{ + return WinogradConv::get_input_matrix_stride(kernel_shape, input_shape, padding_type); +} + +template +NEWinogradLayerTransformInputKernel::NEWinogradLayerTransformInputKernel() + : _transform() +{ +} + +template +void NEWinogradLayerTransformInputKernel::configure( + 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. */ +{ + // _input_matrix_row_stride(n_input_channels), + _transform = support::cpp14::make_unique(input, n_batches, n_rows, n_cols, n_channels, padding, output, matrix_stride, n_channels); + Window win; + auto win_last = _transform->get_window(); + win.set(Window::DimX, Window::Dimension(0, win_last, 1)); + INEKernel::configure(win); +} + +template +void NEWinogradLayerTransformInputKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + const size_t fst = window.x().start(); + const size_t lst = window.x().end(); + _transform->run(fst, lst); +} + +template +bool NEWinogradLayerTransformInputKernel::is_parallelisable() const +{ + return false; +} + +template +Status NEWinogradLayerTransformInputKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_input_trans(input, output, winograd_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_input_trans(input->clone().get(), output->clone().get(), winograd_info).first); + + return Status{}; +} + +template class NEWinogradLayerTransformInputKernel; +template class NEWinogradLayerTransformInputKernel; + +// Output transform + +template +unsigned int NEWinogradLayerTransformOutputKernel::get_output_storage_size( + int n_batches, /** Number of batches in the output tensor. */ + int n_rows, /** Number of rows in each feature map of the input tensor. */ + int n_cols, /** Number of columns in each feature map of the input tensor. */ + int n_output_channels, /** Number of feature maps in the output tensor. */ + bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ +) const +{ + // Construct shapes for the input and kernel tensors. + const Tensor4DShape input_shape(n_batches, n_rows, n_cols, 1); + const KernelShape kern_shape(n_output_channels, KernelRows, KernelCols, 1); + const PaddingType padding = (same_padding) ? PADDING_SAME : PADDING_VALID; + + // Return the size, converted into units of TOut + return static_cast( + WinogradConv::get_output_storage_size(kern_shape, input_shape, padding) / sizeof(T)); +} + +template +NEWinogradLayerTransformOutputKernel::NEWinogradLayerTransformOutputKernel() + : _biases(nullptr), _output_workspace(nullptr), _matrix_stride(0), _matrix_row_stride(0), _output(nullptr), _n_batches(0), _n_rows(0), _n_cols(0), _n_channels(0) +{ +} + +template +int NEWinogradLayerTransformOutputKernel::get_matrix_stride( + const KernelShape &kernel_shape, const Tensor4DShape &input_shape, const PaddingType padding_type) const +{ + return WinogradConv::get_output_matrix_stride(kernel_shape, input_shape, padding_type); +} +template +Tensor4DShape NEWinogradLayerTransformOutputKernel::get_output_shape( + const KernelShape &kernel_shape, const Tensor4DShape &in_shape, const PaddingType padding) const +{ + return WinogradConv::get_output_shape(kernel_shape, in_shape, padding); +} + +template +void NEWinogradLayerTransformOutputKernel::configure( + const ITensor *biases, + const T *const output_workingspace, + const int matrix_stride, + T *const output, + const int n_batches, + const int n_rows, + const int n_cols, + const int n_channels) +{ + _biases = biases; + _output_workspace = output_workingspace; + _matrix_stride = matrix_stride; + _matrix_row_stride = roundup(n_channels, WinogradConv::N_BLOCK); + _output = output; + _n_batches = n_batches; + _n_rows = n_rows; + _n_cols = n_cols; + _n_channels = n_channels; + + // We don't have the biases buffer at this stage as it hasn't been allocated, we pass in nullptr OutputTransform is only used here to compute the window + OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride, nullptr, _output, _n_batches, _n_rows, _n_cols, _n_channels); + Window win; + auto win_last = output_transform.get_window(); + win.set(Window::DimX, Window::Dimension(0, win_last, 1)); + INEKernel::configure(win); +} + +template +void NEWinogradLayerTransformOutputKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_NULLPTR(_output_workspace); + ARM_COMPUTE_ERROR_ON_NULLPTR(_output); + + OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride, + (_biases ? reinterpret_cast(_biases->buffer()) : nullptr), _output, + _n_batches, _n_rows, _n_cols, _n_channels); + + // The code below cannot be moved to configure because biases hasn't been allocated at that point + const size_t fst = window.x().start(); + const size_t lst = window.x().end(); + output_transform.run(fst, lst); +} + +template +bool NEWinogradLayerTransformOutputKernel::is_parallelisable() const +{ + return false; +} + +template +Status NEWinogradLayerTransformOutputKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, + const WinogradInfo &winograd_info) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_output_trans(input, (bias != nullptr ? bias->clone().get() : nullptr), output, winograd_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_output_trans(input->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), output->clone().get(), + winograd_info) + .first); + + return Status{}; +} + +template class NEWinogradLayerTransformOutputKernel; +template class NEWinogradLayerTransformOutputKernel; + +} // namespace arm_compute diff --git a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradLayerKernel.cpp deleted file mode 100644 index 3cfe2af470..0000000000 --- a/src/core/NEON/kernels/NEWinogradLayerKernel.cpp +++ /dev/null @@ -1,561 +0,0 @@ -/* - * Copyright (c) 2017-2018 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/NEWinogradLayerKernel.h" - -#include "arm_compute/core/AccessWindowStatic.h" -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/IAccessWindow.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/Window.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "support/ToolchainSupport.h" - -namespace arm_compute -{ -//Batched Gemms - -namespace -{ -Status validate_arguments_winograd_gemm(const ITensorInfo *a, const ITensorInfo *b, const ITensor *c, const ITensorInfo *output, const float alpha, const float beta, - const GEMMInfo &gemm_info = GEMMInfo()) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(a); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(b); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); - - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported"); - - if(c != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, c->info()); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != c->info()->dimension(1), "The matrix C must have the same number of rows as the matrix A"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != c->info()->dimension(0), "The matrix C must have the same number of columns as the matrix B"); - } - - if(output->total_size() != 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, output); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != output->dimension(0), "The output matrix must have the same number of columns as the matrix B"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != output->dimension(1), "The output matrix must have the same number of rows as the matrix A"); - ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() != a->num_dimensions()); - } - - ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(0) != b->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); - ARM_COMPUTE_UNUSED(alpha, beta); - return Status{}; -} - -Status validate_arguments_winograd_weight_trans(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - - const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); - const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != 3 && input->dimension(idx_width) != 5); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_width) != input->dimension(idx_height)); - ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); - const Size2D &output_tile = winograd_info.output_tile_size; - ARM_COMPUTE_RETURN_ERROR_ON(output_tile != Size2D(2U, 2U) && output_tile != Size2D(4U, 4U)); - - // Checks performed when output is configured - if(output->total_size() != 0) - { - const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_filter_transform_shape(*input, winograd_info)); - - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - } - - return Status{}; -} - -std::pair validate_and_configure_window_winograd_weight_trans(ITensorInfo *input, ITensorInfo *output, const WinogradInfo &winograd_info) -{ - const Size2D kernel_dims = winograd_info.kernel_size; - // Output tensor auto inizialitation if not yet initialized - auto_init_if_empty(*output, input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_filter_transform_shape(*input, winograd_info))); - - unsigned int num_elems_processed_per_iteration_x = kernel_dims.width; - unsigned int num_elems_processed_per_iteration_y = kernel_dims.height; - - Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); - bool window_changed = false; - - AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); - AccessWindowStatic output_access(output, 0, 0, output->dimension(0), output->dimension(1)); - window_changed = update_window_and_padding(win, input_access, output_access); - output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape())); - - Window win_collapsed = win.collapse(win, Window::DimZ); - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - - return std::make_pair(err, win_collapsed); -} - -Status validate_arguments_winograd_input_trans(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) -{ - const Size2D &kernel_dims = winograd_info.kernel_size; - const PadStrideInfo &conv_info = winograd_info.convolution_info; - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd input transform only supports unit strides"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd input transform only supports 3x3 and 5x5 kernels"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd input transform only supports 3x3 and 5x5 kernels"); - - // Validate configured output - if(output->total_size() != 0) - { - const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); - - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - } - - return Status{}; -} - -std::pair validate_and_configure_window_winograd_input_trans(ITensorInfo *input, ITensorInfo *output, const WinogradInfo &winograd_info) -{ - const PadStrideInfo conv_info = winograd_info.convolution_info; - const Size2D output_tile_size = winograd_info.output_tile_size; - const Size2D kernel_dims = winograd_info.kernel_size; - const TensorShape output_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); - // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape)); - - unsigned int num_elems_read_per_iteration_x = (output_tile_size.width + kernel_dims.width - 1); - unsigned int num_elems_read_per_iteration_y = (output_tile_size.height + kernel_dims.height - 1); - - Window win = calculate_max_window(*input, Steps(1, 1)); - - AccessWindowRectangle input_access(input, -conv_info.pad_left(), -conv_info.pad_top(), num_elems_read_per_iteration_x, num_elems_read_per_iteration_y); - - bool window_changed = update_window_and_padding(win, input_access); - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, win); -} - -Status validate_arguments_winograd_output_trans(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info) -{ - const PadStrideInfo &conv_info = winograd_info.convolution_info; - const Size2D kernel_dims = winograd_info.kernel_size; - - // Number of tiles along the X and Y direction - const unsigned int num_tiles_x = std::ceil((winograd_info.input_dimensions.x() - (kernel_dims.width - 1) + conv_info.pad_left() + conv_info.pad_right()) / 2.f); - const unsigned int num_tiles_y = std::ceil((winograd_info.input_dimensions.y() - (kernel_dims.height - 1) + conv_info.pad_top() + conv_info.pad_bottom()) / 2.f); - const Size2D num_tiles = Size2D(num_tiles_x, num_tiles_y); - - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON(winograd_info.output_data_layout != DataLayout::NCHW); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area()); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd output transform only supports 3x3 and 5x5 kernels"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd output transform only supports 3x3 and 5x5 kernels"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(((input->dimension(2) != size_t(16U)) && (input->dimension(2) != size_t(36U))), "Only 2x2 and 4x4 output tile is supported"); - ARM_COMPUTE_UNUSED(kernel_dims); - if(bias != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0)); - ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() != size_t(1)); - } - - // Checks performed when output is configured - if(output->total_size() != 0) - { - const TensorInfo tensor_info_output = input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_output_transform_shape(*input, winograd_info)); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - } - return Status{}; -} - -std::pair validate_and_configure_window_winograd_output_trans(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output, const WinogradInfo &winograd_info) -{ - // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, input->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_winograd_output_transform_shape(*input, winograd_info))); - - constexpr unsigned int num_elems_processed_per_iteration = 1; - - Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); - bool window_changed = false; - - AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration); - AccessWindowStatic output_access(output, 0, 0, ceil_to_multiple(output->dimension(0), 2), ceil_to_multiple(output->dimension(1), 2)); - - if(bias != nullptr) - { - AccessWindowStatic bias_access(bias, 0, 0, bias->dimension(0), bias->dimension(1)); - window_changed = update_window_and_padding(win, input_access, bias_access, output_access); - } - else - { - window_changed = update_window_and_padding(win, input_access, output_access); - } - output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape())); - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, win); -} -} // namespace -template -NEWinogradLayerBatchedGEMMKernel::NEWinogradLayerBatchedGEMMKernel() - : _gemms() -{ -} - -template -void NEWinogradLayerBatchedGEMMKernel::configure( - const unsigned int n_gemms, - const int M, const int K, const int N, - const int a_matrix_stride, - const int a_row_stride, - const int b_matrix_stride, - const int b_row_stride, - const int c_matrix_stride, - const int c_row_stride, - const TIn *const a_ptr, - const TIn *const b_ptr, - TOut *const c_ptr) -{ - _gemms = support::cpp14::make_unique(n_gemms, M, K, N, a_matrix_stride, a_row_stride, b_matrix_stride, b_row_stride, c_matrix_stride, c_row_stride, a_ptr, b_ptr, c_ptr); - Window win; - auto win_last = _gemms->get_window(); - win.set(Window::DimX, Window::Dimension(0, win_last, 1)); - INEKernel::configure(win); -} - -template -void NEWinogradLayerBatchedGEMMKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - const size_t first_gemm = window.x().start(); - const size_t last_gemm = window.x().end(); - _gemms->run(first_gemm, last_gemm); -} - -template -unsigned int NEWinogradLayerBatchedGEMMKernel::get_number_gemms() const -{ - return WinogradBase::N_GEMMS; -} - -template -int NEWinogradLayerBatchedGEMMKernel::get_output_tile_rows() const -{ - return _output_tile_rows; -} - -template -int NEWinogradLayerBatchedGEMMKernel::get_output_tile_cols() const -{ - return _output_tile_cols; -} - -template -int NEWinogradLayerBatchedGEMMKernel::get_number_blocks() const -{ - return WinogradConv::N_BLOCK; -} - -template -Status NEWinogradLayerBatchedGEMMKernel::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensor *c, - const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_gemm(a, b, c, output, alpha, beta, gemm_info)); - return Status{}; -} - -template class NEWinogradLayerBatchedGEMMKernel; -template class NEWinogradLayerBatchedGEMMKernel; - -// Weights transform - -template -unsigned int NEWinogradLayerTransformWeightsKernel::get_weight_storage_size(int n_output_channels, int n_input_channels) const -{ - const KernelShape shape(n_output_channels, KernelRows, KernelCols, n_input_channels); - return static_cast( - // WinogradConv returns the size in bytes, we divide by `sizeof(T)` to express that in units of T - WinogradConv::get_kernel_storage_size(shape) / sizeof(T)); -} - -template -NEWinogradLayerTransformWeightsKernel::NEWinogradLayerTransformWeightsKernel() - : _transform() -{ -} - -template -int NEWinogradLayerTransformWeightsKernel::get_matrix_stride(const KernelShape &kernel_shape) const -{ - return WinogradConv::get_kernel_matrix_stride(kernel_shape); -} - -template -void NEWinogradLayerTransformWeightsKernel::configure( - const ITensor *weights_hwio, - T *const output, - const int matrix_stride, /** Stride across matrices in the output. */ - const int n_output_channels, /** Number of filters. */ - const int n_input_channels) /** Number of channels in each filter. */ -{ - const int matrix_row_stride = roundup(n_output_channels, WinogradConv::N_BLOCK); - _transform = support::cpp14::make_unique(reinterpret_cast(weights_hwio->buffer()), output, matrix_stride, matrix_row_stride, n_output_channels, - n_input_channels); - Window win; - auto win_last = _transform->get_window(); - win.set(Window::DimX, Window::Dimension(0, win_last, 1)); - INEKernel::configure(win); -} - -template -void NEWinogradLayerTransformWeightsKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - const size_t fst = window.x().start(); - const size_t lst = window.x().end(); - _transform->run(fst, lst); -} - -template -bool NEWinogradLayerTransformWeightsKernel::is_parallelisable() const -{ - return false; -} - -template -Status NEWinogradLayerTransformWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, - const WinogradInfo &winograd_info) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_weight_trans(input, output, winograd_info)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_weight_trans(input->clone().get(), output->clone().get(), winograd_info).first); - return Status{}; -} - -template class NEWinogradLayerTransformWeightsKernel; -template class NEWinogradLayerTransformWeightsKernel; - -// Input transform - -template -unsigned int NEWinogradLayerTransformInputKernel::get_input_storage_size( - int n_batches, /** Number of batches in the input tensor. */ - int n_channels, /** Number of feature maps in the input tensor. */ - int n_rows, /** Number of rows in each feature map. */ - int n_cols, /** Number of columns in each feature map. */ - bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ -) const -{ - // Construct shapes for the input and kernel tensors. - const Tensor4DShape input_shape(n_batches, n_rows, n_cols, n_channels); - const KernelShape kern_shape(1, KernelRows, KernelCols, n_channels); - const PaddingType padding = (same_padding) ? PADDING_SAME : PADDING_VALID; - // Return the size, converted into units of TIn - return static_cast(WinogradConv::get_input_storage_size(kern_shape, input_shape, padding) / sizeof(T)); -} - -template -int NEWinogradLayerTransformInputKernel::get_matrix_stride( - const KernelShape &kernel_shape, const Tensor4DShape &input_shape, const PaddingType padding_type) const -{ - return WinogradConv::get_input_matrix_stride(kernel_shape, input_shape, padding_type); -} - -template -NEWinogradLayerTransformInputKernel::NEWinogradLayerTransformInputKernel() - : _transform() -{ -} - -template -void NEWinogradLayerTransformInputKernel::configure( - 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. */ -{ - // _input_matrix_row_stride(n_input_channels), - _transform = support::cpp14::make_unique(input, n_batches, n_rows, n_cols, n_channels, padding, output, matrix_stride, n_channels); - Window win; - auto win_last = _transform->get_window(); - win.set(Window::DimX, Window::Dimension(0, win_last, 1)); - INEKernel::configure(win); -} - -template -void NEWinogradLayerTransformInputKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - const size_t fst = window.x().start(); - const size_t lst = window.x().end(); - _transform->run(fst, lst); -} - -template -bool NEWinogradLayerTransformInputKernel::is_parallelisable() const -{ - return false; -} - -template -Status NEWinogradLayerTransformInputKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_input_trans(input, output, winograd_info)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_input_trans(input->clone().get(), output->clone().get(), winograd_info).first); - - return Status{}; -} - -template class NEWinogradLayerTransformInputKernel; -template class NEWinogradLayerTransformInputKernel; - -// Output transform - -template -unsigned int NEWinogradLayerTransformOutputKernel::get_output_storage_size( - int n_batches, /** Number of batches in the output tensor. */ - int n_rows, /** Number of rows in each feature map of the input tensor. */ - int n_cols, /** Number of columns in each feature map of the input tensor. */ - int n_output_channels, /** Number of feature maps in the output tensor. */ - bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ -) const -{ - // Construct shapes for the input and kernel tensors. - const Tensor4DShape input_shape(n_batches, n_rows, n_cols, 1); - const KernelShape kern_shape(n_output_channels, KernelRows, KernelCols, 1); - const PaddingType padding = (same_padding) ? PADDING_SAME : PADDING_VALID; - - // Return the size, converted into units of TOut - return static_cast( - WinogradConv::get_output_storage_size(kern_shape, input_shape, padding) / sizeof(T)); -} - -template -NEWinogradLayerTransformOutputKernel::NEWinogradLayerTransformOutputKernel() - : _biases(nullptr), _output_workspace(nullptr), _matrix_stride(0), _matrix_row_stride(0), _output(nullptr), _n_batches(0), _n_rows(0), _n_cols(0), _n_channels(0) -{ -} - -template -int NEWinogradLayerTransformOutputKernel::get_matrix_stride( - const KernelShape &kernel_shape, const Tensor4DShape &input_shape, const PaddingType padding_type) const -{ - return WinogradConv::get_output_matrix_stride(kernel_shape, input_shape, padding_type); -} -template -Tensor4DShape NEWinogradLayerTransformOutputKernel::get_output_shape( - const KernelShape &kernel_shape, const Tensor4DShape &in_shape, const PaddingType padding) const -{ - return WinogradConv::get_output_shape(kernel_shape, in_shape, padding); -} - -template -void NEWinogradLayerTransformOutputKernel::configure( - const ITensor *biases, - const T *const output_workingspace, - const int matrix_stride, - T *const output, - const int n_batches, - const int n_rows, - const int n_cols, - const int n_channels) -{ - _biases = biases; - _output_workspace = output_workingspace; - _matrix_stride = matrix_stride; - _matrix_row_stride = roundup(n_channels, WinogradConv::N_BLOCK); - _output = output; - _n_batches = n_batches; - _n_rows = n_rows; - _n_cols = n_cols; - _n_channels = n_channels; - - // We don't have the biases buffer at this stage as it hasn't been allocated, we pass in nullptr OutputTransform is only used here to compute the window - OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride, nullptr, _output, _n_batches, _n_rows, _n_cols, _n_channels); - Window win; - auto win_last = output_transform.get_window(); - win.set(Window::DimX, Window::Dimension(0, win_last, 1)); - INEKernel::configure(win); -} - -template -void NEWinogradLayerTransformOutputKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_NULLPTR(_output_workspace); - ARM_COMPUTE_ERROR_ON_NULLPTR(_output); - - OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride, - (_biases ? reinterpret_cast(_biases->buffer()) : nullptr), _output, - _n_batches, _n_rows, _n_cols, _n_channels); - - // The code below cannot be moved to configure because biases hasn't been allocated at that point - const size_t fst = window.x().start(); - const size_t lst = window.x().end(); - output_transform.run(fst, lst); -} - -template -bool NEWinogradLayerTransformOutputKernel::is_parallelisable() const -{ - return false; -} - -template -Status NEWinogradLayerTransformOutputKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, - const WinogradInfo &winograd_info) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_winograd_output_trans(input, (bias != nullptr ? bias->clone().get() : nullptr), output, winograd_info)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_winograd_output_trans(input->clone().get(), (bias != nullptr ? bias->clone().get() : nullptr), output->clone().get(), - winograd_info) - .first); - - return Status{}; -} - -template class NEWinogradLayerTransformOutputKernel; -template class NEWinogradLayerTransformOutputKernel; - -} // namespace arm_compute diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp index 906378c565..7a37dfa39d 100644 --- a/src/graph/backends/NEON/NEFunctionFactory.cpp +++ b/src/graph/backends/NEON/NEFunctionFactory.cpp @@ -169,8 +169,8 @@ std::unique_ptr create_convolution_layer(ConvolutionLayerNode &node, } else if(conv_algorithm == ConvolutionMethod::WINOGRAD) { - std::tie(func, func_name) = create_named_memory_managed_function(std::string("NEWinogradLayer"), mm, - input, weights, biases, output, conv_info); + std::tie(func, func_name) = create_named_memory_managed_function(std::string("NEWinogradConvolutionLayer"), mm, + input, weights, biases, output, conv_info); } else { diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp index 074f03580f..e438e79c76 100644 --- a/src/graph/backends/NEON/NENodeValidator.cpp +++ b/src/graph/backends/NEON/NENodeValidator.cpp @@ -51,7 +51,7 @@ Status NENodeValidator::validate(INode *node) return detail::validate_convolution_layer(*polymorphic_downcast(node)); + NEWinogradConvolutionLayer>(*polymorphic_downcast(node)); case NodeType::DepthwiseConvolutionLayer: return detail::validate_depthwise_convolution_layer(*polymorphic_downcast(node)); diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp index 61ea2db15b..0ad4babedc 100644 --- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp @@ -51,7 +51,7 @@ void NEConvolutionLayer::configure(ITensor *input, const ITensor *weights, const { case ConvolutionMethod::WINOGRAD: { - auto f = arm_compute::support::cpp14::make_unique(_memory_manager); + auto f = arm_compute::support::cpp14::make_unique(_memory_manager); f->configure(input, weights, biases, output, conv_info, act_info); _function = std::move(f); break; @@ -83,7 +83,7 @@ Status NEConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo { case ConvolutionMethod::WINOGRAD: //Validate Winograd - NEWinogradLayer::validate(input, weights, biases, output, conv_info, act_info); + NEWinogradConvolutionLayer::validate(input, weights, biases, output, conv_info, act_info); break; case ConvolutionMethod::GEMM: //Validate Gemm-based Convolution diff --git a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp new file mode 100644 index 0000000000..a1256ac8cb --- /dev/null +++ b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp @@ -0,0 +1,390 @@ +/* + * Copyright (c) 2017-2018 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/runtime/NEON/functions/NEWinogradConvolutionLayer.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/runtime/NEON/AssemblyHelper.h" +#include "arm_compute/runtime/NEON/NEScheduler.h" +#include "support/ToolchainSupport.h" + +#include "arm_compute/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h" + +#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" + +namespace +{ +inline Tensor4DShape internal_get_input_shape(const arm_compute::ITensor *input) +{ + const int in_width = input->info()->dimension(0); + const int in_height = input->info()->dimension(1); + const int in_batches = input->info()->dimension(3); + const int in_channels = input->info()->dimension(2); + return Tensor4DShape({ in_batches, in_height, in_width, in_channels }); +} +} /* namespace */ + +namespace arm_compute +{ +namespace +{ +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(weights); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != 3 && weights->dimension(0) != 5, "Only 3 and 5 kernels are supported"); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); + + if(biases != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + // Get parameters from conv_info + unsigned int stride_x = 0; + unsigned int stride_y = 0; + std::tie(stride_x, stride_y) = conv_info.stride(); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(stride_y != 1 || stride_x != 1, "Winograd layer only supports unit strides."); + + ARM_COMPUTE_UNUSED(output); + return Status{}; +} +} //namespace + +NEWinogradConvolutionLayer::NEWinogradConvolutionLayer(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _arm_gemm(nullptr), _gemm_kernel(nullptr), _transform_input_kernel(nullptr), _transform_output_kernel(nullptr), _transform_weights_kernel(nullptr), + _activationlayer_function(), _permute_input(), _permute_weights(), _permute_output(), _input_workspace(), _output_workspace(), _kernel_storage(), _input_nhwc(), _output_nhwc(), _weights_hwio(), + _workspace(), _input(), _weights(), _output(), _reshaped_kernel(false), _is_activationlayer_enabled(false) +{ +} /* arm_compute */ + +void NEWinogradConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_UNUSED(conv_info); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info)); + + _weights = weights; + _input = input; + _output = output; + + std::unique_ptr> transform_input_kernel; + std::unique_ptr> transform_weights_kernel; + std::unique_ptr> transform_output_kernel; + + const int weights_width = weights->info()->dimension(0); + const int weights_height = weights->info()->dimension(1); + + int output_tile_rows = 0; + int output_tile_cols = 0; + int n_gemms = 0; + int N_BLOCK = 0; // Size of block used by GEMM. + + switch(weights_width) + { + case 3: + { + transform_input_kernel = support::cpp14::make_unique>(); + transform_weights_kernel = support::cpp14::make_unique>(); + transform_output_kernel = support::cpp14::make_unique>(); + output_tile_rows = 2; + output_tile_cols = 2; + n_gemms = NEWinogradLayerBatchedGEMMKernel::WinogradBase::N_GEMMS; + N_BLOCK = NEWinogradLayerBatchedGEMMKernel::WinogradConv::N_BLOCK; + break; + } + case 5: + { + transform_input_kernel = support::cpp14::make_unique>(); + transform_weights_kernel = support::cpp14::make_unique>(); + transform_output_kernel = support::cpp14::make_unique>(); + output_tile_rows = 2; + output_tile_cols = 2; + n_gemms = NEWinogradLayerBatchedGEMMKernel::WinogradBase::N_GEMMS; + N_BLOCK = NEWinogradLayerBatchedGEMMKernel::WinogradConv::N_BLOCK; + break; + } + default: + { + ARM_COMPUTE_ERROR("Not supported."); + break; + } + } + + const PaddingType use_padding_type = (conv_info.pad_left() != 0u) ? PADDING_SAME : PADDING_VALID; + const bool use_same_padding = use_padding_type == PADDING_SAME; + + // Get parameters from conv_info + unsigned int stride_x = 0; + unsigned int stride_y = 0; + std::tie(stride_x, stride_y) = conv_info.stride(); + ARM_COMPUTE_ERROR_ON_MSG(stride_y != 1 || stride_x != 1, "Winograd layer only supports unit strides."); + + // Get convolved dimensions + const int in_channels = input->info()->dimension(2); + const int out_channels = output->info()->dimension(2); + + const Tensor4DShape in_shape(internal_get_input_shape(input)); + const size_t data_type_size = input->info()->element_size(); + // Get the memory required to instantiate a new Winograd operator. + constexpr size_t storage_alignment = 64; + const size_t kernel_storage_size = transform_weights_kernel->get_weight_storage_size(out_channels, in_channels) * data_type_size; + _kernel_storage.allocator()->init(TensorInfo(TensorShape{ (kernel_storage_size + storage_alignment - 1) }, 1, DataType::U8)); + _kernel_storage.allocator()->allocate(); + // Input storage + const size_t input_storage_size = transform_input_kernel->get_input_storage_size(in_shape.n_batches, in_shape.n_channels, in_shape.n_rows, in_shape.n_cols, use_same_padding) * data_type_size; + _input_workspace.allocator()->init(TensorInfo(TensorShape{ (input_storage_size + storage_alignment - 1) }, 1, DataType::U8)); + _input_workspace.allocator()->allocate(); + + // Output storage + const size_t output_storage_size = transform_output_kernel->get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels, use_same_padding) * data_type_size; + _output_workspace.allocator()->init(TensorInfo(TensorShape{ (output_storage_size + storage_alignment - 1) }, 1, DataType::U8)); + _output_workspace.allocator()->allocate(); + + // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output() + TensorInfo info(TensorShape(_output->info()->dimension(2), _output->info()->dimension(0), + _output->info()->dimension(1), _output->info()->dimension(3)), + 1, _output->info()->data_type()); + _output_nhwc.allocator()->init(info); + _output_nhwc.allocator()->allocate(); + + // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map] + _permute_weights.configure(weights, &_weights_hwio, PermutationVector(3U, 2U, 0U, 1U)); + _weights_hwio.allocator()->allocate(); + + // configure the kernel to transform the input tensor from NCHW -> NHWC + _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U)); + _input_nhwc.allocator()->allocate(); + + const KernelShape kernel_shape({ out_channels, weights_height, weights_width, in_channels }); + + // Configure the InputTransform + const int input_matrix_stride = transform_input_kernel->get_matrix_stride(kernel_shape, in_shape, use_padding_type); + transform_input_kernel->configure(reinterpret_cast(_input_nhwc.buffer()), in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type, + reinterpret_cast(_input_workspace.buffer()), input_matrix_stride); + + // Configure WeightsTransform + const int kernel_matrix_stride = transform_weights_kernel->get_matrix_stride(kernel_shape); + transform_weights_kernel->configure(&_weights_hwio, reinterpret_cast(_kernel_storage.buffer()), kernel_matrix_stride, out_channels, in_channels); + + // Configure OutputTransform + //The biases tensor has not been allocated at this point in time, the output transform will add the biases to the final result in the run() method + const int output_matrix_stride = transform_output_kernel->get_matrix_stride(kernel_shape, in_shape, use_padding_type); + const auto output_shape(transform_output_kernel->get_output_shape(kernel_shape, in_shape, use_padding_type)); + + transform_output_kernel->configure(biases, reinterpret_cast(_output_workspace.buffer()), + output_matrix_stride, reinterpret_cast(_output_nhwc.buffer()), + in_shape.n_batches, output_shape.n_rows, output_shape.n_cols, out_channels); + + // Configure GEMM + const int tile_rows = iceildiv(output_shape.n_rows, output_tile_rows); + const int tile_cols = iceildiv(output_shape.n_cols, output_tile_cols); + const int m = in_shape.n_batches * tile_rows * tile_cols; + const int k = in_shape.n_channels; + const int n = out_channels; + const int input_matrix_row_stride = in_shape.n_channels; + const int kernel_matrix_row_stride = roundup(out_channels, N_BLOCK); + const int output_matrix_row_stride = kernel_matrix_row_stride; + unsigned int num_threads = NEScheduler::get().num_threads(); + + _arm_gemm = arm_gemm::gemm(NEScheduler::get().cpu_info(), m, n, k, 1, n_gemms, false, false, 1.f, 0.f, num_threads, false); + _arm_gemm->set_arrays(reinterpret_cast(_input_workspace.buffer()), input_matrix_row_stride, 0, input_matrix_stride, reinterpret_cast(_kernel_storage.buffer()), + kernel_matrix_row_stride, kernel_matrix_stride, reinterpret_cast(_output_workspace.buffer()), output_matrix_row_stride, 0, output_matrix_stride); + + auto acl_gemm_wrapper = support::cpp14::make_unique>>(); + acl_gemm_wrapper->configure(_arm_gemm.get()); + const size_t workspace_size = _arm_gemm->get_working_size(); + + // Allocate workspace + if(workspace_size > 0) + { + const unsigned int alignment = 4096; + allocate_workspace(workspace_size, _workspace, _memory_group, alignment, 1); + _arm_gemm->set_working_space(reinterpret_cast(_workspace.buffer())); + } + + const unsigned int window_size = _arm_gemm->get_window_size(); + if(window_size < num_threads) + { + num_threads = window_size; + _arm_gemm->set_nthreads(num_threads); + } + + _gemm_kernel = std::move(acl_gemm_wrapper); + + // Reorder the convoluted output to ACL's ordering NCHW + _permute_output.configure(&_output_nhwc, _output, PermutationVector(1U, 2U, 0U)); + + _transform_input_kernel = std::move(transform_input_kernel); + _transform_weights_kernel = std::move(transform_weights_kernel); + _transform_output_kernel = std::move(transform_output_kernel); + + //Configure Activation Layer + _is_activationlayer_enabled = act_info.enabled(); + if(_is_activationlayer_enabled) + { + _activationlayer_function.configure(output, nullptr, act_info); + } +} + +void NEWinogradConvolutionLayer::run() +{ + _memory_group.acquire(); + if(!_reshaped_kernel) + { + _reshaped_kernel = true; + _permute_weights.run(); + NEScheduler::get().schedule(_transform_weights_kernel.get(), Window::DimX); + } + //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC + _permute_input.run(); + + // Transform input tensor to the winograd domain + NEScheduler::get().schedule(_transform_input_kernel.get(), Window::DimX); + + //Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs + NEScheduler::get().schedule(_gemm_kernel.get(), Window::DimX); + + // Transform output tensor to the spatial domain + NEScheduler::get().schedule(_transform_output_kernel.get(), Window::DimX); + + // Reorder the convoluted output to ACL's ordering NCHW + _permute_output.run(); + + if(_is_activationlayer_enabled) + { + _activationlayer_function.run(); + } + _memory_group.release(); +} + +Status NEWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const ActivationLayerInfo &act_info) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info)); + + // Get indices for the width and height + const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); + const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); + // Input shape + const TensorShape input_shape = input->tensor_shape(); + + // Kernel size + const unsigned int kernel_w = weights->tensor_shape()[idx_width]; + const unsigned int kernel_h = weights->tensor_shape()[idx_height]; + + const WinogradInfo winograd_info = WinogradInfo(Size2D(2, 2), + Size2D(kernel_w, kernel_h), + Size2D(input_shape[idx_width], input_shape[idx_height]), + conv_info, + input->data_layout()); + + // Validate input transform + const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); + const TensorInfo input0 = input->clone()->set_tensor_shape(input0_shape); + switch(weights->dimension(0)) + { + case 3: + { + ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel::validate(input, &input0, winograd_info))); + break; + } + case 5: + { + ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel::validate(input, &input0, winograd_info))); + break; + } + default: + { + ARM_COMPUTE_RETURN_ERROR_MSG("Only 3x3 and 5x5 kernels supported."); + break; + } + } + // Validate filter transform + const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, winograd_info); + const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); + + switch(weights->dimension(0)) + { + case 3: + { + ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel::validate(weights, &input1, winograd_info))); + break; + } + case 5: + { + ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel::validate(weights, &input1, winograd_info))); + break; + } + default: + { + ARM_COMPUTE_RETURN_ERROR_MSG("Only 3x3 and 5x5 kernels supported."); + break; + } + } + // Validate batched matrix multiply + TensorShape batched_mm_output_shape = input0.tensor_shape(); + batched_mm_output_shape[0] = input1.tensor_shape()[0]; + const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); + switch(weights->dimension(0)) + { + case 3: + { + ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerBatchedGEMMKernel::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, + true /* Reshape weights only for the first run*/)))); + // Validate output transform + ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel::validate(&batched_mm_output, biases, output, winograd_info))); + break; + } + case 5: + { + ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerBatchedGEMMKernel::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, + true /* Reshape weights only for the first run*/)))); + // Validate output transform + ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel::validate(&batched_mm_output, biases, output, winograd_info))); + break; + } + default: + { + ARM_COMPUTE_RETURN_ERROR_MSG("Only 3x3 and 5x5 kernels supported."); + break; + } + } + + // Validate Activation Layer + if(act_info.enabled()) + { + NEActivationLayer::validate(output, nullptr, act_info); + } + return Status{}; +} + +} // namespace arm_compute diff --git a/src/runtime/NEON/functions/NEWinogradLayer.cpp b/src/runtime/NEON/functions/NEWinogradLayer.cpp deleted file mode 100644 index 7d93bcff07..0000000000 --- a/src/runtime/NEON/functions/NEWinogradLayer.cpp +++ /dev/null @@ -1,390 +0,0 @@ -/* - * Copyright (c) 2017-2018 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/runtime/NEON/functions/NEWinogradLayer.h" - -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/runtime/NEON/AssemblyHelper.h" -#include "arm_compute/runtime/NEON/NEScheduler.h" -#include "support/ToolchainSupport.h" - -#include "arm_compute/core/NEON/kernels/NEWinogradLayerKernel.h" - -#include "arm_compute/core/NEON/kernels/convolution/winograd/winograd_gemm.hpp" - -namespace -{ -inline Tensor4DShape internal_get_input_shape(const arm_compute::ITensor *input) -{ - const int in_width = input->info()->dimension(0); - const int in_height = input->info()->dimension(1); - const int in_batches = input->info()->dimension(3); - const int in_channels = input->info()->dimension(2); - return Tensor4DShape({ in_batches, in_height, in_width, in_channels }); -} -} /* namespace */ - -namespace arm_compute -{ -namespace -{ -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(weights); - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != 3 && weights->dimension(0) != 5, "Only 3 and 5 kernels are supported"); - ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); - - if(biases != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - } - - // Get parameters from conv_info - unsigned int stride_x = 0; - unsigned int stride_y = 0; - std::tie(stride_x, stride_y) = conv_info.stride(); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(stride_y != 1 || stride_x != 1, "Winograd layer only supports unit strides."); - - ARM_COMPUTE_UNUSED(output); - return Status{}; -} -} //namespace - -NEWinogradLayer::NEWinogradLayer(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _arm_gemm(nullptr), _gemm_kernel(nullptr), _transform_input_kernel(nullptr), _transform_output_kernel(nullptr), _transform_weights_kernel(nullptr), - _activationlayer_function(), _permute_input(), _permute_weights(), _permute_output(), _input_workspace(), _output_workspace(), _kernel_storage(), _input_nhwc(), _output_nhwc(), _weights_hwio(), - _workspace(), _input(), _weights(), _output(), _reshaped_kernel(false), _is_activationlayer_enabled(false) -{ -} /* arm_compute */ - -void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); - ARM_COMPUTE_UNUSED(conv_info); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info)); - - _weights = weights; - _input = input; - _output = output; - - std::unique_ptr> transform_input_kernel; - std::unique_ptr> transform_weights_kernel; - std::unique_ptr> transform_output_kernel; - - const int weights_width = weights->info()->dimension(0); - const int weights_height = weights->info()->dimension(1); - - int output_tile_rows = 0; - int output_tile_cols = 0; - int n_gemms = 0; - int N_BLOCK = 0; // Size of block used by GEMM. - - switch(weights_width) - { - case 3: - { - transform_input_kernel = support::cpp14::make_unique>(); - transform_weights_kernel = support::cpp14::make_unique>(); - transform_output_kernel = support::cpp14::make_unique>(); - output_tile_rows = 2; - output_tile_cols = 2; - n_gemms = NEWinogradLayerBatchedGEMMKernel::WinogradBase::N_GEMMS; - N_BLOCK = NEWinogradLayerBatchedGEMMKernel::WinogradConv::N_BLOCK; - break; - } - case 5: - { - transform_input_kernel = support::cpp14::make_unique>(); - transform_weights_kernel = support::cpp14::make_unique>(); - transform_output_kernel = support::cpp14::make_unique>(); - output_tile_rows = 2; - output_tile_cols = 2; - n_gemms = NEWinogradLayerBatchedGEMMKernel::WinogradBase::N_GEMMS; - N_BLOCK = NEWinogradLayerBatchedGEMMKernel::WinogradConv::N_BLOCK; - break; - } - default: - { - ARM_COMPUTE_ERROR("Not supported."); - break; - } - } - - const PaddingType use_padding_type = (conv_info.pad_left() != 0u) ? PADDING_SAME : PADDING_VALID; - const bool use_same_padding = use_padding_type == PADDING_SAME; - - // Get parameters from conv_info - unsigned int stride_x = 0; - unsigned int stride_y = 0; - std::tie(stride_x, stride_y) = conv_info.stride(); - ARM_COMPUTE_ERROR_ON_MSG(stride_y != 1 || stride_x != 1, "Winograd layer only supports unit strides."); - - // Get convolved dimensions - const int in_channels = input->info()->dimension(2); - const int out_channels = output->info()->dimension(2); - - const Tensor4DShape in_shape(internal_get_input_shape(input)); - const size_t data_type_size = input->info()->element_size(); - // Get the memory required to instantiate a new Winograd operator. - constexpr size_t storage_alignment = 64; - const size_t kernel_storage_size = transform_weights_kernel->get_weight_storage_size(out_channels, in_channels) * data_type_size; - _kernel_storage.allocator()->init(TensorInfo(TensorShape{ (kernel_storage_size + storage_alignment - 1) }, 1, DataType::U8)); - _kernel_storage.allocator()->allocate(); - // Input storage - const size_t input_storage_size = transform_input_kernel->get_input_storage_size(in_shape.n_batches, in_shape.n_channels, in_shape.n_rows, in_shape.n_cols, use_same_padding) * data_type_size; - _input_workspace.allocator()->init(TensorInfo(TensorShape{ (input_storage_size + storage_alignment - 1) }, 1, DataType::U8)); - _input_workspace.allocator()->allocate(); - - // Output storage - const size_t output_storage_size = transform_output_kernel->get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels, use_same_padding) * data_type_size; - _output_workspace.allocator()->init(TensorInfo(TensorShape{ (output_storage_size + storage_alignment - 1) }, 1, DataType::U8)); - _output_workspace.allocator()->allocate(); - - // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output() - TensorInfo info(TensorShape(_output->info()->dimension(2), _output->info()->dimension(0), - _output->info()->dimension(1), _output->info()->dimension(3)), - 1, _output->info()->data_type()); - _output_nhwc.allocator()->init(info); - _output_nhwc.allocator()->allocate(); - - // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map] - _permute_weights.configure(weights, &_weights_hwio, PermutationVector(3U, 2U, 0U, 1U)); - _weights_hwio.allocator()->allocate(); - - // configure the kernel to transform the input tensor from NCHW -> NHWC - _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U)); - _input_nhwc.allocator()->allocate(); - - const KernelShape kernel_shape({ out_channels, weights_height, weights_width, in_channels }); - - // Configure the InputTransform - const int input_matrix_stride = transform_input_kernel->get_matrix_stride(kernel_shape, in_shape, use_padding_type); - transform_input_kernel->configure(reinterpret_cast(_input_nhwc.buffer()), in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type, - reinterpret_cast(_input_workspace.buffer()), input_matrix_stride); - - // Configure WeightsTransform - const int kernel_matrix_stride = transform_weights_kernel->get_matrix_stride(kernel_shape); - transform_weights_kernel->configure(&_weights_hwio, reinterpret_cast(_kernel_storage.buffer()), kernel_matrix_stride, out_channels, in_channels); - - // Configure OutputTransform - //The biases tensor has not been allocated at this point in time, the output transform will add the biases to the final result in the run() method - const int output_matrix_stride = transform_output_kernel->get_matrix_stride(kernel_shape, in_shape, use_padding_type); - const auto output_shape(transform_output_kernel->get_output_shape(kernel_shape, in_shape, use_padding_type)); - - transform_output_kernel->configure(biases, reinterpret_cast(_output_workspace.buffer()), - output_matrix_stride, reinterpret_cast(_output_nhwc.buffer()), - in_shape.n_batches, output_shape.n_rows, output_shape.n_cols, out_channels); - - // Configure GEMM - const int tile_rows = iceildiv(output_shape.n_rows, output_tile_rows); - const int tile_cols = iceildiv(output_shape.n_cols, output_tile_cols); - const int m = in_shape.n_batches * tile_rows * tile_cols; - const int k = in_shape.n_channels; - const int n = out_channels; - const int input_matrix_row_stride = in_shape.n_channels; - const int kernel_matrix_row_stride = roundup(out_channels, N_BLOCK); - const int output_matrix_row_stride = kernel_matrix_row_stride; - unsigned int num_threads = NEScheduler::get().num_threads(); - - _arm_gemm = arm_gemm::gemm(NEScheduler::get().cpu_info(), m, n, k, 1, n_gemms, false, false, 1.f, 0.f, num_threads, false); - _arm_gemm->set_arrays(reinterpret_cast(_input_workspace.buffer()), input_matrix_row_stride, 0, input_matrix_stride, reinterpret_cast(_kernel_storage.buffer()), - kernel_matrix_row_stride, kernel_matrix_stride, reinterpret_cast(_output_workspace.buffer()), output_matrix_row_stride, 0, output_matrix_stride); - - auto acl_gemm_wrapper = support::cpp14::make_unique>>(); - acl_gemm_wrapper->configure(_arm_gemm.get()); - const size_t workspace_size = _arm_gemm->get_working_size(); - - // Allocate workspace - if(workspace_size > 0) - { - const unsigned int alignment = 4096; - allocate_workspace(workspace_size, _workspace, _memory_group, alignment, 1); - _arm_gemm->set_working_space(reinterpret_cast(_workspace.buffer())); - } - - const unsigned int window_size = _arm_gemm->get_window_size(); - if(window_size < num_threads) - { - num_threads = window_size; - _arm_gemm->set_nthreads(num_threads); - } - - _gemm_kernel = std::move(acl_gemm_wrapper); - - // Reorder the convoluted output to ACL's ordering NCHW - _permute_output.configure(&_output_nhwc, _output, PermutationVector(1U, 2U, 0U)); - - _transform_input_kernel = std::move(transform_input_kernel); - _transform_weights_kernel = std::move(transform_weights_kernel); - _transform_output_kernel = std::move(transform_output_kernel); - - //Configure Activation Layer - _is_activationlayer_enabled = act_info.enabled(); - if(_is_activationlayer_enabled) - { - _activationlayer_function.configure(output, nullptr, act_info); - } -} - -void NEWinogradLayer::run() -{ - _memory_group.acquire(); - if(!_reshaped_kernel) - { - _reshaped_kernel = true; - _permute_weights.run(); - NEScheduler::get().schedule(_transform_weights_kernel.get(), Window::DimX); - } - //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC - _permute_input.run(); - - // Transform input tensor to the winograd domain - NEScheduler::get().schedule(_transform_input_kernel.get(), Window::DimX); - - //Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs - NEScheduler::get().schedule(_gemm_kernel.get(), Window::DimX); - - // Transform output tensor to the spatial domain - NEScheduler::get().schedule(_transform_output_kernel.get(), Window::DimX); - - // Reorder the convoluted output to ACL's ordering NCHW - _permute_output.run(); - - if(_is_activationlayer_enabled) - { - _activationlayer_function.run(); - } - _memory_group.release(); -} - -Status NEWinogradLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, - const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info)); - - // Get indices for the width and height - const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); - const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); - // Input shape - const TensorShape input_shape = input->tensor_shape(); - - // Kernel size - const unsigned int kernel_w = weights->tensor_shape()[idx_width]; - const unsigned int kernel_h = weights->tensor_shape()[idx_height]; - - const WinogradInfo winograd_info = WinogradInfo(Size2D(2, 2), - Size2D(kernel_w, kernel_h), - Size2D(input_shape[idx_width], input_shape[idx_height]), - conv_info, - input->data_layout()); - - // Validate input transform - const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); - const TensorInfo input0 = input->clone()->set_tensor_shape(input0_shape); - switch(weights->dimension(0)) - { - case 3: - { - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel::validate(input, &input0, winograd_info))); - break; - } - case 5: - { - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformInputKernel::validate(input, &input0, winograd_info))); - break; - } - default: - { - ARM_COMPUTE_RETURN_ERROR_MSG("Only 3x3 and 5x5 kernels supported."); - break; - } - } - // Validate filter transform - const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, winograd_info); - const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); - - switch(weights->dimension(0)) - { - case 3: - { - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel::validate(weights, &input1, winograd_info))); - break; - } - case 5: - { - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformWeightsKernel::validate(weights, &input1, winograd_info))); - break; - } - default: - { - ARM_COMPUTE_RETURN_ERROR_MSG("Only 3x3 and 5x5 kernels supported."); - break; - } - } - // Validate batched matrix multiply - TensorShape batched_mm_output_shape = input0.tensor_shape(); - batched_mm_output_shape[0] = input1.tensor_shape()[0]; - const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); - switch(weights->dimension(0)) - { - case 3: - { - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerBatchedGEMMKernel::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, - true /* Reshape weights only for the first run*/)))); - // Validate output transform - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel::validate(&batched_mm_output, biases, output, winograd_info))); - break; - } - case 5: - { - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerBatchedGEMMKernel::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, - true /* Reshape weights only for the first run*/)))); - // Validate output transform - ARM_COMPUTE_RETURN_ON_ERROR((NEWinogradLayerTransformOutputKernel::validate(&batched_mm_output, biases, output, winograd_info))); - break; - } - default: - { - ARM_COMPUTE_RETURN_ERROR_MSG("Only 3x3 and 5x5 kernels supported."); - break; - } - } - - // Validate Activation Layer - if(act_info.enabled()) - { - NEActivationLayer::validate(output, nullptr, act_info); - } - return Status{}; -} - -} // namespace arm_compute -- cgit v1.2.1