From 96f977e43f452a75f2658b820791cb3d3da9c0a3 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Thu, 1 Jul 2021 12:20:56 +0100 Subject: Port NEWinogradConvolutionLayer Rename to CpuWinogradConv2d Allow memory to be injected externally Change-Id: I1f0a26ea533e326a7c63df86e708895c31752a39 Signed-off-by: Michalis Spyrou Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5926 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio --- src/core/NEON/NEKernels.h | 1 - .../kernels/NEWinogradConvolutionLayerKernel.cpp | 548 ------------------- .../kernels/NEWinogradConvolutionLayerKernel.h | 597 --------------------- src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp | 552 +++++++++++++++++++ src/core/cpu/kernels/CpuWinogradConv2dKernel.h | 575 ++++++++++++++++++++ 5 files changed, 1127 insertions(+), 1146 deletions(-) delete mode 100644 src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp delete mode 100644 src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h create mode 100644 src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp create mode 100644 src/core/cpu/kernels/CpuWinogradConv2dKernel.h (limited to 'src/core') diff --git a/src/core/NEON/NEKernels.h b/src/core/NEON/NEKernels.h index cd09544d31..6c6c51dd87 100644 --- a/src/core/NEON/NEKernels.h +++ b/src/core/NEON/NEKernels.h @@ -66,6 +66,5 @@ #include "src/core/NEON/kernels/NEStridedSliceKernel.h" #include "src/core/NEON/kernels/NETileKernel.h" #include "src/core/NEON/kernels/NEWeightsReshapeKernel.h" -#include "src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h" #endif /* ARM_COMPUTE_NEKERNELS_H */ diff --git a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp deleted file mode 100644 index be34980663..0000000000 --- a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp +++ /dev/null @@ -1,548 +0,0 @@ -/* - * Copyright (c) 2017-2021 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 "src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h" - -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.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 "src/core/NEON/kernels/convolution/common/utils.hpp" -#include "src/core/NEON/kernels/convolution/winograd/winograd_layer.hpp" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/WindowHelpers.h" - -#include - -namespace arm_compute -{ -//Batched Gemms - -namespace -{ -inline bool is_kernel_size_supported(DataType data_type, Size2D size) -{ - const std::array f32_support = { { Size2D(1, 3), Size2D(3, 1), Size2D(5, 5), Size2D(3, 3), Size2D(1, 5), Size2D(5, 1), Size2D(7, 1), Size2D(1, 7) } }; - const std::array f16_support = { { Size2D(3, 3) } }; - - switch(data_type) - { - case DataType::F16: - return std::end(f16_support) != std::find(std::begin(f16_support), std::end(f16_support), size); - case DataType::F32: - return std::end(f32_support) != std::find(std::begin(f32_support), std::end(f32_support), size); - default: - return false; - } -} - -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::F16, 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); - const auto input_width = input->dimension(idx_width); - const auto input_height = input->dimension(idx_height); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(input->data_type(), Size2D(input_width, input_height)), - "Only 1x3, 3x1, 1x5, 5x1, 7x1, 1x7, 3x3 and 5x5 kernels are supported"); - ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); - const Size2D &output_tile = winograd_info.output_tile_size; - const std::array supported_tile_sizes = { { Size2D(2U, 2U), Size2D(4U, 4U), Size2D(1U, 6U), Size2D(6U, 1U), Size2D(4, 1), Size2D(1, 4), Size2D(2, 1), Size2D(1, 2) } }; - ARM_COMPUTE_RETURN_ERROR_ON(std::end(supported_tile_sizes) == std::find(std::begin(supported_tile_sizes), std::end(supported_tile_sizes), output_tile)); - - // 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) -{ - // 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))); - const Window win = calculate_max_window(*input, Steps(), true /* skip border*/); - return std::make_pair(Status{}, win); -} - -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::F16, 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(!is_kernel_size_supported(input->data_type(), Size2D(kernel_dims.width, kernel_dims.height)), - "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); - - // 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 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)); - return std::make_pair(Status{}, calculate_max_window(*input, Steps(), true)); -} - -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()) / static_cast - (winograd_info.output_tile_size.width)); - 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()) / static_cast - (winograd_info.output_tile_size.height)); - 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::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area()); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(input->data_type(), Size2D(kernel_dims.width, kernel_dims.height)), - "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); - - const std::array supported_gemm_sizes = { { 8U, 16U, 36U } }; - ARM_COMPUTE_RETURN_ERROR_ON(std::end(supported_gemm_sizes) == std::find(std::begin(supported_gemm_sizes), std::end(supported_gemm_sizes), input->dimension(2))); - 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 *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))); - - return std::make_pair(Status{}, calculate_max_window(*input, Steps(), true)); -} -} // namespace - -Status INEWinogradLayerTransformWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *weights) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); - const DataLayout data_layout = input->data_layout(); - const unsigned int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(input->data_type(), Size2D(weights->dimension(width_idx), weights->dimension(height_idx))), - "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); - ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); - return Status{}; -} - -template -unsigned int NEWinogradLayerTransformWeightsKernel::get_weight_storage_size(int num_output_channels, int num_input_channels) const -{ - const KernelShape shape(num_output_channels, KernelRows, KernelCols, num_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(num_input_channels, num_output_channels) / sizeof(T)); -} - -template -NEWinogradLayerTransformWeightsKernel::NEWinogradLayerTransformWeightsKernel() - : _transform(nullptr), _weights_hwio(nullptr), _output(nullptr), _matrix_stride(0), _num_output_channels(0), _num_input_channels(0) -{ -} - -template -int NEWinogradLayerTransformWeightsKernel::get_matrix_stride(int num_output_channels, int num_input_channels) const -{ - return WinogradConv::get_kernel_matrix_stride(num_input_channels, num_output_channels); -} - -#ifndef DOXYGEN_SKIP_THIS -template -void NEWinogradLayerTransformWeightsKernel::configure( - const ITensor *weights_hwio, - ITensor *output, - const int matrix_stride, /** Stride across matrices in the output. */ - const int num_output_channels, /** Number of filters. */ - const int num_input_channels) /** Number of channels in each filter. */ -{ - _weights_hwio = weights_hwio; - _output = output; - _matrix_stride = matrix_stride; - _num_output_channels = num_output_channels; - _num_input_channels = num_input_channels; - _transform = std::make_unique(num_output_channels, num_input_channels); - - Window win; - auto win_last = _transform->get_window(); - win.set(Window::DimX, Window::Dimension(0, win_last, 1)); - INEKernel::configure(win); -} -#endif /* DOXYGEN_SKIP_THIS */ - -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->set_weight_tensor(_weights_hwio->buffer()); - const int matrix_row_stride = roundup(_num_output_channels, WinogradConv::N_BLOCK); - _transform->set_output_matrices(_output->buffer(), _matrix_stride, matrix_row_stride); - _transform->set_working_space(_output->buffer()); - - _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; -template class NEWinogradLayerTransformWeightsKernel; -template class NEWinogradLayerTransformWeightsKernel; -template class NEWinogradLayerTransformWeightsKernel; - -template class NEWinogradLayerTransformWeightsKernel; -template class NEWinogradLayerTransformWeightsKernel; -template class NEWinogradLayerTransformWeightsKernel; -template class NEWinogradLayerTransformWeightsKernel; - -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template class NEWinogradLayerTransformWeightsKernel<__fp16, 4, 4, 3, 3>; -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - -// Input transform - -template -unsigned int NEWinogradLayerTransformInputKernel::get_input_storage_size( - int num_batches, /* Number of batches in the input tensor. */ - int num_channels, /* Number of feature maps in the input tensor. */ - int num_rows, /* Number of rows in each feature map. */ - int num_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(num_batches, num_rows, num_cols, num_channels); - const KernelShape kern_shape(1, KernelRows, KernelCols, num_channels); - // Return the size, converted into units of TIn - return static_cast(WinogradConv::get_input_storage_size(num_batches, num_rows, num_cols, num_channels, same_padding) / sizeof(T)); -} - -template -unsigned int NEWinogradLayerTransformInputKernel::get_working_space_size(unsigned int num_threads) const -{ - return _transform->get_working_space_size(num_threads) / sizeof(T); -} - -template -int NEWinogradLayerTransformInputKernel::get_matrix_stride( - int num_batches, /* Number of batches in the input tensor. */ - int num_channels, /* Number of feature maps in the input tensor. */ - int num_rows, /* Number of rows in each feature map. */ - int num_cols, /* Number of columns in each feature map. */ - bool same_padding /* Use "SAME" padding, otherwise use "VALID". */) const -{ - return WinogradConv::get_input_matrix_stride(num_batches, num_rows, num_cols, num_channels, same_padding); -} - -template -NEWinogradLayerTransformInputKernel::NEWinogradLayerTransformInputKernel() - : _transform(nullptr), _input_nhwc(nullptr), _num_batches(0), _num_rows(0), _num_cols(0), _num_channels(0), _padding(), _output(nullptr), _matrix_stride(0), _padding_top(), _padding_left(), - _padding_right(), _padding_bottom(), _workspace(nullptr) -{ -} - -template -void NEWinogradLayerTransformInputKernel::configure( - const ITensor *input_nhwc, - const int num_batches, /* Number of batches in input tensor. */ - const int num_rows, /* Number of rows in input tensor. */ - const int num_cols, /* Number of columns in input tensor. */ - const int num_channels, /* Number of channels in input tensor. */ - const PaddingType padding, /* Padding type. */ - ITensor *output, /* Base of output matrices. */ - const int matrix_stride, /* Stride between output matrices. */ - ITensor *workspace) -{ - _input_nhwc = input_nhwc; - _num_batches = num_batches; - _num_rows = num_rows; - _num_cols = num_cols; - _num_channels = num_channels; - _padding = padding; - _output = output; - _matrix_stride = matrix_stride; - _workspace = workspace; - - _padding_top = (padding == PADDING_SAME) ? (KernelRows - 1) / 2 : 0; - _padding_left = (padding == PADDING_SAME) ? (KernelCols - 1) / 2 : 0; - _padding_bottom = (padding == PADDING_SAME) ? iceildiv(KernelRows - 1, 2) : 0; - _padding_right = (padding == PADDING_SAME) ? iceildiv(KernelCols - 1, 2) : 0; - - _transform = std::make_unique( - KernelRows, - KernelCols, - num_batches, - num_rows, - num_cols, - num_channels, - _padding_top, /**< Padding to apply to the top of the image. */ - _padding_left, /**< Padding to apply to the left of the image. */ - _padding_bottom, /**< Padding to apply to the bottom of the image. */ - _padding_right /**< Padding to apply to the right of the image. */ - ); - - 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); - ARM_COMPUTE_ERROR_ON_NULLPTR(_workspace); - - const int element_size_in_bytes = _input_nhwc->info()->element_size(); - const int input_col_stride = _input_nhwc->info()->strides_in_bytes().y() / element_size_in_bytes; - const int input_row_stride = _input_nhwc->info()->strides_in_bytes().z() / element_size_in_bytes; - const int input_batch_stride = _input_nhwc->info()->strides_in_bytes()[3] / element_size_in_bytes; - const auto input_nhwc_ptr = reinterpret_cast(_input_nhwc->buffer() + _input_nhwc->info()->offset_first_element_in_bytes()); - auto output_ptr = reinterpret_cast(_output->buffer() + _output->info()->offset_first_element_in_bytes()); - ARM_COMPUTE_ERROR_ON_NULLPTR(output_ptr); - - _transform->set_input_tensor(input_nhwc_ptr, input_batch_stride, input_row_stride, input_col_stride); - _transform->set_output_matrices(output_ptr, _matrix_stride, _num_channels); - - _transform->set_working_space(_workspace->buffer()); - - // 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(); - _transform->run(fst, lst, info.thread_id); -} - -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; -template class NEWinogradLayerTransformInputKernel; -template class NEWinogradLayerTransformInputKernel; -template class NEWinogradLayerTransformInputKernel; - -template class NEWinogradLayerTransformInputKernel; -template class NEWinogradLayerTransformInputKernel; -template class NEWinogradLayerTransformInputKernel; -template class NEWinogradLayerTransformInputKernel; - -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template class NEWinogradLayerTransformInputKernel<__fp16, 4, 4, 3, 3>; -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - -// Output transform - -template -unsigned int NEWinogradLayerTransformOutputKernel::get_output_storage_size( - int num_batches, /* Number of batches in the output tensor. */ - int num_rows, /* Number of rows in each feature map of the input tensor. */ - int num_cols, /* Number of columns in each feature map of the input tensor. */ - int num_output_channels /* Number of feature maps in the output tensor. */ -) const -{ - // Construct shapes for the input and kernel tensors. - const Tensor4DShape input_shape(num_batches, num_rows, num_cols, 1); - const KernelShape kern_shape(num_output_channels, KernelRows, KernelCols, 1); - // Return the size, converted into units of TOut - return static_cast( - WinogradConv::get_output_storage_size(num_batches, num_rows, num_cols, num_output_channels) / sizeof(T)); -} - -template -NEWinogradLayerTransformOutputKernel::NEWinogradLayerTransformOutputKernel() - : _transform(nullptr), _biases(nullptr), _transformed_output(nullptr), _workspace(nullptr), _matrix_stride(0), _matrix_row_stride(0), _output_nhwc(nullptr), _num_batches(0), _num_rows(0), - _num_cols(0), _num_channels(0) -{ -} - -template -unsigned int NEWinogradLayerTransformOutputKernel::get_working_space_size(unsigned int num_threads) const -{ - return _transform->get_working_space_size(num_threads) / sizeof(T); -} - -template -int NEWinogradLayerTransformOutputKernel::get_matrix_stride( - int num_batches, /* Number of batches in the output tensor. */ - int num_rows, /* Number of rows in each feature map of the input tensor. */ - int num_cols, /* Number of columns in each feature map of the input tensor. */ - int num_output_channels /* Number of feature maps in the output tensor. */ -) const -{ - return WinogradConv::get_output_matrix_stride(num_batches, num_rows, num_cols, num_output_channels); -} - -template -std::pair NEWinogradLayerTransformOutputKernel::get_output_shape( - int num_rows, /* Number of rows in each feature map of the input tensor. */ - int num_cols, /* Number of columns in each feature map of the input tensor. */ - bool padding_same) const -{ - return WinogradConv::get_output_shape(std::make_pair(num_rows, num_cols), padding_same); -} - -template -void NEWinogradLayerTransformOutputKernel::configure( - const ITensor *biases, - const ITensor *transformed_output, - const int matrix_stride, - ITensor *output_nhwc, - const int num_batches, - const int num_rows, - const int num_cols, - const int num_channels, - ITensor *workspace, - const arm_gemm::Activation &activation) -{ - _biases = biases; - _workspace = workspace; - _transformed_output = transformed_output; - _matrix_stride = matrix_stride; - _matrix_row_stride = roundup(num_channels, WinogradConv::N_BLOCK); - _output_nhwc = output_nhwc; - _num_batches = num_batches; - _num_rows = num_rows; - _num_cols = num_cols; - _num_channels = num_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 - _transform = std::make_unique(num_batches, num_rows, num_cols, num_channels, activation); - Window win; - auto win_last = _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(_workspace); - ARM_COMPUTE_ERROR_ON_NULLPTR(_transformed_output); - ARM_COMPUTE_ERROR_ON_NULLPTR(_output_nhwc); - - const int out_batch_stride = _output_nhwc->info()->strides_in_bytes()[3] / sizeof(T); - const int out_row_stride = _output_nhwc->info()->strides_in_bytes()[2] / sizeof(T); - const int out_col_stride = _output_nhwc->info()->strides_in_bytes()[1] / sizeof(T); - - _transform->set_input_matrices(_transformed_output->buffer(), _matrix_stride, _matrix_row_stride); - _transform->set_bias((_biases ? reinterpret_cast(_biases->buffer() + _biases->info()->offset_first_element_in_bytes()) : nullptr)); - _transform->set_output_tensor(_output_nhwc->buffer() + _output_nhwc->info()->offset_first_element_in_bytes(), out_batch_stride, out_row_stride, out_col_stride); - _transform->set_working_space(_workspace->buffer()); - // 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(); - _transform->run(fst, lst, info.thread_id); -} - -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(), output->clone().get(), winograd_info).first); - - return Status{}; -} - -template class NEWinogradLayerTransformOutputKernel; -template class NEWinogradLayerTransformOutputKernel; -template class NEWinogradLayerTransformOutputKernel; -template class NEWinogradLayerTransformOutputKernel; -template class NEWinogradLayerTransformOutputKernel; - -template class NEWinogradLayerTransformOutputKernel; -template class NEWinogradLayerTransformOutputKernel; -template class NEWinogradLayerTransformOutputKernel; -template class NEWinogradLayerTransformOutputKernel; - -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template class NEWinogradLayerTransformOutputKernel<__fp16, 4, 4, 3, 3>; -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -} // namespace arm_compute diff --git a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h deleted file mode 100644 index 75d257de4b..0000000000 --- a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.h +++ /dev/null @@ -1,597 +0,0 @@ -/* - * Copyright (c) 2017-2021 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. - */ -#ifndef ARM_COMPUTE_NEGEMMWINOGRADCONVOLUTIONLAYERKERNEL_H -#define ARM_COMPUTE_NEGEMMWINOGRADCONVOLUTIONLAYERKERNEL_H - -#include "src/core/NEON/INEKernel.h" -#include "src/core/NEON/kernels/convolution/common/convolution.hpp" -#include "src/core/NEON/kernels/convolution/common/tensor.hpp" - -#include "src/core/NEON/kernels/convolution/winograd/winograd_layer.hpp" - -namespace arm_compute -{ -// Forward declarations -class ITensor; - -/** Interface for the kernel to perform Winograd input transform. */ -class INEWinogradLayerTransformInputKernel : public INEKernel -{ -public: - /** Get the working space required to perform the transformation. - * - * Note, the working space is only required when performing the - * transformation - hence it can be reused whenever the transformation is - * not running. - * - * @param num_threads The greatest number of threads that will be used to execute the transform. - * @return Size of working space required in bytes. - */ - virtual unsigned int get_working_space_size(unsigned int num_threads) const = 0; - - /** Determine how much memory (in units of TIn) to allocate for the - * transformed input. - * - * @param[in] num_batches Number of batches in the input tensor. - * @param[in] num_channels Number of feature maps in the input tensor. - * @param[in] num_rows Number of rows in each feature map. - * @param[in] num_cols Number of columns in each feature map. - * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". - * - * @return Storage size (in units of TIn) required. - */ - virtual unsigned int get_input_storage_size(int num_batches, int num_channels, int num_rows, int num_cols, bool same_padding) const = 0; - - /** Gets the stride between matrices in the input worspace - * - * @param[in] num_batches Number of batches in the input tensor. - * @param[in] num_channels Number of feature maps in the input tensor. - * @param[in] num_rows Number of rows in each feature map. - * @param[in] num_cols Number of columns in each feature map. - * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". - * - * @return Stride expressed in bytes. - */ - virtual int get_matrix_stride(int num_batches, int num_channels, int num_rows, int num_cols, bool same_padding) const = 0; - - /** Configure the output transform kernel. - * - * @param[in] input_nhwc Input tensor in NHWC data layout format. - * @param[in] num_batches Number of batches in input tensor. - * @param[in] num_rows Number of rows in input tensor. - * @param[in] num_cols Number of columns in input tensor. - * @param[in] num_channels Number of channels in input tensor. - * @param[in] padding Padding type. - * @param[out] output Base of output matrices. - * @param[in] matrix_stride Stride between output matrices. - * @param[in] workspace Tensor to be used as the working space during the computation. - */ - virtual void configure(const ITensor *input_nhwc, const int num_batches, const int num_rows, const int num_cols, const int num_channels, - const PaddingType padding, ITensor *output, const int matrix_stride, ITensor *workspace) = 0; - - /** Destructor */ - virtual ~INEWinogradLayerTransformInputKernel() - { - } -}; - -/** Kernel to perform Winograd input transform. */ -template -class NEWinogradLayerTransformInputKernel : public INEWinogradLayerTransformInputKernel -{ -public: - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformInputKernel(const NEWinogradLayerTransformInputKernel &) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformInputKernel &operator=(const NEWinogradLayerTransformInputKernel &) = delete; - /** Allow instances of this class to be moved */ - NEWinogradLayerTransformInputKernel(NEWinogradLayerTransformInputKernel &&) = default; - /** Allow instances of this class to be moved */ - NEWinogradLayerTransformInputKernel &operator=(NEWinogradLayerTransformInputKernel &&) = default; - /** Default destructor */ - ~NEWinogradLayerTransformInputKernel() = default; - - /** Determine how much memory (in units of TIn) to allocate for the - * transformed input. - * - * @param[in] num_batches Number of batches in the input tensor. - * @param[in] num_channels Number of feature maps in the input tensor. - * @param[in] num_rows Number of rows in each feature map. - * @param[in] num_cols Number of columns in each feature map. - * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". - * - * @return Storage size (in units of TIn) required. - */ - unsigned int get_input_storage_size( - int num_batches, - int num_channels, - int num_rows, - int num_cols, - bool same_padding) const override; - - /** Get the working space required to perform the transformation. - * - * Note, the working space is only required when performing the - * transformation - hence it can be reused whenever the transformation is - * not running. - * - * @param[in] num_threads The greatest number of threads that will be used to execute the transform. - * - * @return Size of working space required in bytes. - */ - unsigned int get_working_space_size(unsigned int num_threads) const override; - - /** Gets the stride between matrices in the input worspace - * - * @param[in] num_batches Number of batches in the input tensor. - * @param[in] num_channels Number of feature maps in the input tensor. - * @param[in] num_rows Number of rows in each feature map. - * @param[in] num_cols Number of columns in each feature map. - * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". - * - * @return Stride expressed in bytes. - */ - int get_matrix_stride( - int num_batches, - int num_channels, - int num_rows, - int num_cols, - bool same_padding) const override; - - /** Default constructor */ - NEWinogradLayerTransformInputKernel(); - - const char *name() const override - { - return "NEWinogradLayerTransformInputKernel"; - } - - /** Configure the output transform kernel. - * - * @param[in] input_nhwc Input tensor. Data types supported: F16/F32. Layout supported NHWC. - * @param[in] num_batches Number of batches in input tensor. - * @param[in] num_rows Number of rows in input tensor. - * @param[in] num_cols Number of columns in input tensor. - * @param[in] num_channels Number of channels in input tensor. - * @param[in] padding Padding type. - * @param[out] output Base of output matrices. - * @param[in] matrix_stride Stride between output matrices. - * @param[in] workspace Tensor to be used as the working space during the computation. - */ - void configure( - const ITensor *input_nhwc, - const int num_batches, - const int num_rows, - const int num_cols, - const int num_channels, - const PaddingType padding, - ITensor *output, - const int matrix_stride, - ITensor *workspace) override; - - // Inherited methods overridden: - void run(const Window &window, const ThreadInfo &info) override; - - /** Winograd base kernel */ - using WinogradBase = winograd::WinogradGEMM; - /** Winograd convolution kernel */ - using WinogradConv = typename WinogradBase::template Convolution; - - /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformInputKernel - * - * @param[in] input First tensor input info. Data types supported: F16/F32. - * @param[in] output Output tensor info. Data types supported: same as @p input. - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info); - -private: - using InputTransform = typename WinogradBase::template InputTransform; - - std::unique_ptr _transform{ nullptr }; - const ITensor *_input_nhwc; - int _num_batches; /**< Number of batches in input tensor. */ - int _num_rows; /**< Number of rows in input tensor. */ - int _num_cols; /**< Number of columns in input tensor. */ - int _num_channels; /**< Number of channels in input tensor. */ - PaddingType _padding; /**< Padding type. */ - ITensor *_output; /**< Base of output matrices. */ - int _matrix_stride; /**< Stride between output matrices. */ - int _padding_top; /**< Padding to apply to the top of the image. */ - int _padding_left; /**< Padding to apply to the left of the image. */ - int _padding_right; /**< Padding to apply to the right of the image. */ - int _padding_bottom; /**< Padding to apply to the bottom of the image. */ - ITensor *_workspace; -}; - -/** Interface for the kernel to perform Winograd output transform. */ -class INEWinogradLayerTransformOutputKernel : public INEKernel -{ -public: - /** Get the working space required to perform the transformation. - * - * Note, the working space is only required when performing the - * transformation - hence it can be reused whenever the transformation is - * not running. - * - * @param[in] num_threads The greatest number of threads that will be used to execute the transform. - * - * @return Size of working space required in bytes. - */ - virtual unsigned int get_working_space_size(unsigned int num_threads) const = 0; - - /** Determine how much memory (in units of TOut) to allocate for the - * (Winograd domain) output. - * - * @param[in] num_batches Number of batches in the output tensor. - * @param[in] num_rows Number of rows in each feature map of the input tensor. - * @param[in] num_cols Number of columns in each feature map of the input tensor. - * @param[in] num_output_channels Number of feature maps in the output tensor. - * - * @return Storage size (in units of TOut) required. - */ - virtual unsigned int get_output_storage_size(int num_batches, int num_rows, int num_cols, int num_output_channels) const = 0; - - /** Gets the stride between matrices in the output worspace - * - * @param[in] num_batches Number of batches in the output tensor. - * @param[in] num_rows Number of rows in each feature map of the input tensor. - * @param[in] num_cols Number of columns in each feature map of the input tensor. - * @param[in] num_output_channels Number of feature maps in the output tensor. - * - * @return Stride expressed in bytes. - */ - virtual int get_matrix_stride(int num_batches, int num_rows, int num_cols, int num_output_channels) const = 0; - - /** Get the output shape of a convolution. - * - * @param[in] num_rows Number of rows in each feature map of the input tensor. - * @param[in] num_cols Number of columns in each feature map of the input tensor. - * @param[in] padding_same True if padding is SAME, false otherwise - * - * @return Shape of the output tensor - */ - virtual std::pair get_output_shape( - int num_rows, /* Number of rows in each feature map of the input tensor. */ - int num_cols, /* Number of columns in each feature map of the input tensor. */ - bool padding_same /* True if padding is SAME, false otherwise */ - ) const = 0; - - /** Configure the output transform kernel. - * - * @param[in] biases Pointer to the biases tensor. - * @param[in] transformed_output Pointer to working space for the output tensor in the Winograd domain. - * @param[in] matrix_stride Output matrix stride, can be computed with winograd::WinogradGEMM<2, 2, 3, 3>::Convolution::get_output_matrix_stride() - * @param[out] output_nhwc Pointer to a tensor in NHWC data layout ordered output tensor, in the spatial domain. - * @param[in] num_batches Number of batches in the input tensor. - * @param[in] num_rows Number of rows in output tensor. - * @param[in] num_cols Number of columns in output tensor. - * @param[in] num_channels Number of feature maps in the output tensor. - * @param[in] workspace Tensor to be used as the working space during the computation. - * @param[in] activation Activation to be used - */ - virtual void configure( - const ITensor *biases, - const ITensor *transformed_output, - const int matrix_stride, - ITensor *output_nhwc, - const int num_batches, - const int num_rows, - const int num_cols, - const int num_channels, - ITensor *workspace, - const arm_gemm::Activation &activation) = 0; - - virtual ~INEWinogradLayerTransformOutputKernel() - { - } -}; - -/** Kernel to perform Winograd output transform. */ -template -class NEWinogradLayerTransformOutputKernel : public INEWinogradLayerTransformOutputKernel -{ -public: - const char *name() const override - { - return "NEWinogradLayerTransformOutputKernel"; - } - /** Constructor */ - NEWinogradLayerTransformOutputKernel(); - - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformOutputKernel(const NEWinogradLayerTransformOutputKernel &) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformOutputKernel &operator=(const NEWinogradLayerTransformOutputKernel &) = delete; - /** Allow instances of this class to be moved */ - NEWinogradLayerTransformOutputKernel(NEWinogradLayerTransformOutputKernel &&) = default; - /** Allow instances of this class to be moved */ - NEWinogradLayerTransformOutputKernel &operator=(NEWinogradLayerTransformOutputKernel &&) = default; - /** Default destructor */ - ~NEWinogradLayerTransformOutputKernel() = default; - - // Inherited methods overridden: - /** Determine how much memory (in units of TOut) to allocate for the - * (Winograd domain) output. - * - * @param[in] num_batches Number of batches in the output tensor. - * @param[in] num_rows Number of rows in each feature map of the input tensor. - * @param[in] num_cols Number of columns in each feature map of the input tensor. - * @param[in] num_output_channels Number of feature maps in the output tensor. - * - * @return Storage size (in units of TOut) required. - */ - unsigned int get_output_storage_size(int num_batches, int num_rows, int num_cols, int num_output_channels) const override; - - /** Gets the stride between matrices in the output worspace - * - * @param[in] num_batches Number of batches in the output tensor. - * @param[in] num_rows Number of rows in each feature map of the input tensor. - * @param[in] num_cols Number of columns in each feature map of the input tensor. - * @param[in] num_output_channels Number of feature maps in the output tensor. - * - * @return Stride expressed in bytes. - */ - int get_matrix_stride(int num_batches, int num_rows, int num_cols, int num_output_channels) const override; - /** Get the output shape of a convolution. - * - * @param[in] num_rows Number of rows in each feature map of the input tensor. - * @param[in] num_cols Number of columns in each feature map of the input tensor. - * @param[in] padding_same True if padding is SAME, false otherwise - * - * @return Shape of the output tensor - */ - std::pair get_output_shape( - int num_rows, /* Number of rows in each feature map of the input tensor. */ - int num_cols, /* Number of columns in each feature map of the input tensor. */ - bool padding_same) const override; - - /** Get the working space required to perform the transformation. - * - * Note, the working space is only required when performing the - * transformation - hence it can be reused whenever the transformation is - * not running. - * - * @param[in] num_threads The greatest number of threads that will be used to execute the transform. - * - * @return Size of working space required in bytes. - */ - unsigned int get_working_space_size(unsigned int num_threads) const override; - - /** Configure the output transform kernel. - * - * @param[in] biases Pointer to the biases tensor. - * @param[in] transformed_output Pointer to working space for the output tensor in the Winograd domain. - * @param[in] matrix_stride Output matrix stride, can be computed with winograd::WinogradGEMM<2, 2, 3, 3>::Convolution::get_output_matrix_stride() - * @param[out] output_nhwc Pointer to a tensor with NHWC data layout, in the spatial domain. - * @param[in] num_batches Number of batches in the input tensor. - * @param[in] num_rows Number of rows in output tensor. - * @param[in] num_cols Number of columns in output tensor. - * @param[in] num_channels Number of feature maps in the output tensor. - * @param[in] workspace Tensor to be used as the working space during the computation. - * @param[in] activation Activation to be used - */ - void configure( - const ITensor *biases, - const ITensor *transformed_output, - const int matrix_stride, - ITensor *output_nhwc, - const int num_batches, - const int num_rows, - const int num_cols, - const int num_channels, - ITensor *workspace, - const arm_gemm::Activation &activation) override; - - void run(const Window &window, const ThreadInfo &info) override; - - /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformOutputKernel - * - * @param[in] input Source tensor info with shape [C, N, 16, batches] or [C, N, 36, batches]. Data types supported: F16/F32. - * @param[in] bias Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input - * @param[in] output Destination tensor info with shape [output_convolved_dims.width, output_convolved_dims.height, C, batches]. Data type supported: same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info); - -private: - using WinogradBase = winograd::WinogradGEMM; - using WinogradConv = typename WinogradBase::template Convolution; - using OutputTransform = typename WinogradBase::template OutputTransform; - - std::unique_ptr _transform{ nullptr }; - const ITensor *_biases; - const ITensor *_transformed_output; - ITensor *_workspace; - int _matrix_stride; - int _matrix_row_stride; - ITensor *_output_nhwc; - int _num_batches; - int _num_rows; - int _num_cols; - int _num_channels; -}; - -/** Interface for the kernel to perform Winograd weights transform. */ -class INEWinogradLayerTransformWeightsKernel : public INEKernel -{ -public: - /** Prevent instances of this class from being copied (As this class contains pointers) */ - INEWinogradLayerTransformWeightsKernel(const INEWinogradLayerTransformWeightsKernel &) = default; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - INEWinogradLayerTransformWeightsKernel &operator=(const INEWinogradLayerTransformWeightsKernel &) = default; - /** Allow instances of this class to be moved */ - INEWinogradLayerTransformWeightsKernel(INEWinogradLayerTransformWeightsKernel &&) = default; - /** Allow instances of this class to be moved */ - INEWinogradLayerTransformWeightsKernel &operator=(INEWinogradLayerTransformWeightsKernel &&) = default; - - INEWinogradLayerTransformWeightsKernel() - { - } - virtual ~INEWinogradLayerTransformWeightsKernel() - { - } - /** Determine how much memory (in units of T) to allocate for the - * transformed weights. - * - * @param[in] num_output_channels Number of output feature maps. - * @param[in] num_input_channels Number of input feature maps. - * - * @return Storage size (in units of T) required. - */ - virtual unsigned int get_weight_storage_size(int num_output_channels, int num_input_channels) const = 0; - /** Gets the stride between matrices in the kernel worspace - * - * @param[in] num_output_channels Number of output feature maps. - * @param[in] num_input_channels Number of input feature maps. - * - * @return Stride expressed in bytes. - */ - virtual int get_matrix_stride(int num_output_channels, int num_input_channels) const = 0; - - /** Configure the weights transform kernel. - * - * @param[in] weights_hwio Pointer to the weights tensor - * @param[out] output Pointer to working space for the output tensor in the Winograd domain. - * @param[in] matrix_stride Stride across matrices in the output workspace. - * @param[in] num_output_channels Number of filters. - * @param[in] num_input_channels Number of channels in each filter. - */ - - virtual void configure(const ITensor *weights_hwio, ITensor *output, const int matrix_stride, const int num_output_channels, const int num_input_channels) = 0; - - /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformWeightsKernel - * - * @param[in] input First tensor input info. Data types supported: F16/F32. - * @param[in] weights Weights tensor info. Data types supported: same as @p input. - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *weights); -}; - -/** Kernel to perform Winograd weights transform. */ -template -class NEWinogradLayerTransformWeightsKernel final : public INEWinogradLayerTransformWeightsKernel -{ -public: - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformWeightsKernel(const NEWinogradLayerTransformWeightsKernel &) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEWinogradLayerTransformWeightsKernel &operator=(const NEWinogradLayerTransformWeightsKernel &) = delete; - /** Allow instances of this class to be moved */ - NEWinogradLayerTransformWeightsKernel(NEWinogradLayerTransformWeightsKernel &&) = default; - /** Allow instances of this class to be moved */ - NEWinogradLayerTransformWeightsKernel &operator=(NEWinogradLayerTransformWeightsKernel &&) = default; - /** Default destructor */ - ~NEWinogradLayerTransformWeightsKernel() = default; - - /** Default constructor. */ - NEWinogradLayerTransformWeightsKernel(); - const char *name() const override - { - return "NEWinogradLayerTransformWeightsKernel"; - } - - /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradLayerTransformWeightsKernel - * - * @param[in] input Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout). - * kernel_x must be 3 and equal to kernel_y. Data types supported: F16/F32. - * @param[in] output Destination tensor info. The output is a 3D tensor with dimensions [OFM, IFM, 16] or [OFM, IFM, 36]. Data type supported: same as @p input - * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info); - - // Inherited methods overridden: - -#ifndef DOXYGEN_SKIP_THIS - /** Configure the weights transform kernel. - * - * @param[in] weights_hwio Pointer to the weights tensor - * @param[out] output Pointer to working space for the output tensor in the Winograd domain. - * @param[in] matrix_stride Stride across matrices in the output workspace. - * @param[in] num_output_channels Number of filters. - * @param[in] num_input_channels Number of channels in each filter. - */ - void configure(const ITensor *weights_hwio, ITensor *output, const int matrix_stride, const int num_output_channels, const int num_input_channels) override; -#endif /* DOXYGEN_SKIP_THIS */ - - /** Determine how much memory (in units of T) to allocate for the - * transformed weights. - * - * @param[in] num_output_channels Number of output feature maps. - * @param[in] num_input_channels Number of input feature maps. - * - * @return Storage size (in units of T) required. - */ - unsigned int get_weight_storage_size(int num_output_channels, int num_input_channels) const override; - - /** Gets the stride between matrices in the input worspace - * - * @param[in] num_output_channels Number of output feature maps. - * @param[in] num_input_channels Number of input feature maps. - * - * @return Stride expressed in bytes. - */ - int get_matrix_stride(int num_output_channels, int num_input_channels) const override; - void run(const Window &window, const ThreadInfo &info) override; - bool is_parallelisable() const override; - -private: - using WinogradBase = winograd::WinogradGEMM; - using WinogradConv = typename WinogradBase::template Convolution; - using WeightsTransform = typename WinogradBase::template WeightsTransform; - - std::unique_ptr _transform{ nullptr }; - const ITensor *_weights_hwio; - ITensor *_output; - int _matrix_stride; - int _num_output_channels; - int _num_input_channels; -}; - -/** Kernel to perform Winograd. */ -template -class NEWinogradLayerConfiguration -{ -public: - /** Winograd base kernel */ - using WinogradBase = winograd::WinogradGEMM; - /** Winograd convolution kernel */ - - using WinogradConv = typename WinogradBase::template Convolution; - - using TransformInputKernel = NEWinogradLayerTransformInputKernel; - using TransformWeightsKernel = NEWinogradLayerTransformWeightsKernel; - using TransformOutputKernel = NEWinogradLayerTransformOutputKernel; -}; - -} // namespace arm_compute -#endif /*ARM_COMPUTE_NEGEMMWINOGRADCONVOLUTIONLAYERKERNEL_H*/ diff --git a/src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp b/src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp new file mode 100644 index 0000000000..74b031b226 --- /dev/null +++ b/src/core/cpu/kernels/CpuWinogradConv2dKernel.cpp @@ -0,0 +1,552 @@ +/* + * Copyright (c) 2017-2021 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 "src/core/cpu/kernels/CpuWinogradConv2dKernel.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.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 "src/core/NEON/kernels/convolution/common/utils.hpp" +#include "src/core/NEON/kernels/convolution/winograd/winograd_layer.hpp" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" + +#include + +namespace arm_compute +{ +namespace cpu +{ +//Batched Gemms + +namespace +{ +inline bool is_kernel_size_supported(DataType data_type, Size2D size) +{ + const std::array f32_support = { { Size2D(1, 3), Size2D(3, 1), Size2D(5, 5), Size2D(3, 3), Size2D(1, 5), Size2D(5, 1), Size2D(7, 1), Size2D(1, 7) } }; + const std::array f16_support = { { Size2D(3, 3) } }; + + switch(data_type) + { + case DataType::F16: + return std::end(f16_support) != std::find(std::begin(f16_support), std::end(f16_support), size); + case DataType::F32: + return std::end(f32_support) != std::find(std::begin(f32_support), std::end(f32_support), size); + default: + return false; + } +} + +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::F16, 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); + const auto input_width = input->dimension(idx_width); + const auto input_height = input->dimension(idx_height); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(input->data_type(), Size2D(input_width, input_height)), + "Only 1x3, 3x1, 1x5, 5x1, 7x1, 1x7, 3x3 and 5x5 kernels are supported"); + ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4); + const Size2D &output_tile = winograd_info.output_tile_size; + const std::array supported_tile_sizes = { { Size2D(2U, 2U), Size2D(4U, 4U), Size2D(1U, 6U), Size2D(6U, 1U), Size2D(4, 1), Size2D(1, 4), Size2D(2, 1), Size2D(1, 2) } }; + ARM_COMPUTE_RETURN_ERROR_ON(std::end(supported_tile_sizes) == std::find(std::begin(supported_tile_sizes), std::end(supported_tile_sizes), output_tile)); + + // 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) +{ + // 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))); + const Window win = calculate_max_window(*input, Steps(), true /* skip border*/); + return std::make_pair(Status{}, win); +} + +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::F16, 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(!is_kernel_size_supported(input->data_type(), Size2D(kernel_dims.width, kernel_dims.height)), + "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); + + // 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 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)); + return std::make_pair(Status{}, calculate_max_window(*input, Steps(), true)); +} + +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()) / static_cast + (winograd_info.output_tile_size.width)); + 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()) / static_cast + (winograd_info.output_tile_size.height)); + 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::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area()); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(input->data_type(), Size2D(kernel_dims.width, kernel_dims.height)), + "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); + + const std::array supported_gemm_sizes = { { 8U, 16U, 36U } }; + ARM_COMPUTE_RETURN_ERROR_ON(std::end(supported_gemm_sizes) == std::find(std::begin(supported_gemm_sizes), std::end(supported_gemm_sizes), input->dimension(2))); + 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 *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))); + + return std::make_pair(Status{}, calculate_max_window(*input, Steps(), true)); +} +} // namespace + +Status ICpuWinogradConv2dTransformWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *weights) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + const DataLayout data_layout = input->data_layout(); + const unsigned int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!is_kernel_size_supported(input->data_type(), Size2D(weights->dimension(width_idx), weights->dimension(height_idx))), + "Only 1x3, 3x1, 3x3 and 5x5 kernels are supported"); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); + return Status{}; +} + +template +unsigned int CpuWinogradConv2dTransformWeightsKernel::get_weight_storage_size(int num_output_channels, int num_input_channels) const +{ + const KernelShape shape(num_output_channels, KernelRows, KernelCols, num_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(num_input_channels, num_output_channels) / sizeof(T)); +} + +template +CpuWinogradConv2dTransformWeightsKernel::CpuWinogradConv2dTransformWeightsKernel() + : _transform(nullptr), _num_output_channels(0), _matrix_stride(0) +{ +} + +template +int CpuWinogradConv2dTransformWeightsKernel::get_matrix_stride(int num_output_channels, int num_input_channels) const +{ + return WinogradConv::get_kernel_matrix_stride(num_input_channels, num_output_channels); +} + +#ifndef DOXYGEN_SKIP_THIS +template +void CpuWinogradConv2dTransformWeightsKernel::configure( + const ITensorInfo *weights_hwio, + ITensorInfo *output, + const int matrix_stride, /** Stride across matrices in the output. */ + const int num_output_channels, /** Number of filters. */ + const int num_input_channels) /** Number of channels in each filter. */ +{ + ARM_COMPUTE_UNUSED(weights_hwio, output); + + _transform = std::make_unique(num_output_channels, num_input_channels); + _num_output_channels = num_output_channels; + _matrix_stride = matrix_stride; + + Window win; + auto win_last = _transform->get_window(); + win.set(Window::DimX, Window::Dimension(0, win_last, 1)); + ICpuKernel::configure(win); +} +#endif /* DOXYGEN_SKIP_THIS */ + +template +void CpuWinogradConv2dTransformWeightsKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON(tensors.empty()); + + const size_t fst = window.x().start(); + const size_t lst = window.x().end(); + + const ITensor *weights_hwio = tensors.get_const_tensor(TensorType::ACL_SRC); + ITensor *output = tensors.get_tensor(TensorType::ACL_DST); + + _transform->set_weight_tensor(weights_hwio->buffer()); + const int matrix_row_stride = roundup(_num_output_channels, WinogradConv::N_BLOCK); + _transform->set_output_matrices(output->buffer(), _matrix_stride, matrix_row_stride); + _transform->set_working_space(output->buffer()); + + _transform->run(fst, lst); +} + +template +bool CpuWinogradConv2dTransformWeightsKernel::is_parallelisable() const +{ + return false; +} + +template +Status CpuWinogradConv2dTransformWeightsKernel::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 CpuWinogradConv2dTransformWeightsKernel; +template class CpuWinogradConv2dTransformWeightsKernel; +template class CpuWinogradConv2dTransformWeightsKernel; +template class CpuWinogradConv2dTransformWeightsKernel; +template class CpuWinogradConv2dTransformWeightsKernel; + +template class CpuWinogradConv2dTransformWeightsKernel; +template class CpuWinogradConv2dTransformWeightsKernel; +template class CpuWinogradConv2dTransformWeightsKernel; +template class CpuWinogradConv2dTransformWeightsKernel; + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +template class CpuWinogradConv2dTransformWeightsKernel<__fp16, 4, 4, 3, 3>; +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +// Input transform + +template +unsigned int CpuWinogradConv2dTransformInputKernel::get_input_storage_size( + int num_batches, /* Number of batches in the input tensor. */ + int num_channels, /* Number of feature maps in the input tensor. */ + int num_rows, /* Number of rows in each feature map. */ + int num_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(num_batches, num_rows, num_cols, num_channels); + const KernelShape kern_shape(1, KernelRows, KernelCols, num_channels); + // Return the size, converted into units of TIn + return static_cast(WinogradConv::get_input_storage_size(num_batches, num_rows, num_cols, num_channels, same_padding) / sizeof(T)); +} + +template +unsigned int CpuWinogradConv2dTransformInputKernel::get_working_space_size(unsigned int num_threads) const +{ + return _transform->get_working_space_size(num_threads) / sizeof(T); +} + +template +int CpuWinogradConv2dTransformInputKernel::get_matrix_stride( + int num_batches, /* Number of batches in the input tensor. */ + int num_channels, /* Number of feature maps in the input tensor. */ + int num_rows, /* Number of rows in each feature map. */ + int num_cols, /* Number of columns in each feature map. */ + bool same_padding /* Use "SAME" padding, otherwise use "VALID". */) const +{ + return WinogradConv::get_input_matrix_stride(num_batches, num_rows, num_cols, num_channels, same_padding); +} + +template +CpuWinogradConv2dTransformInputKernel::CpuWinogradConv2dTransformInputKernel() + : _transform(nullptr), _num_channels(0), _matrix_stride(0) +{ +} + +template +void CpuWinogradConv2dTransformInputKernel::configure( + const ITensorInfo *input_nhwc, + const int num_batches, /* Number of batches in input tensor. */ + const int num_rows, /* Number of rows in input tensor. */ + const int num_cols, /* Number of columns in input tensor. */ + const int num_channels, /* Number of channels in input tensor. */ + const PaddingType padding, /* Padding type. */ + ITensorInfo *output, /* Base of output matrices. */ + const int matrix_stride, /* Stride between output matrices. */ + ITensorInfo *workspace) +{ + ARM_COMPUTE_UNUSED(input_nhwc, output, matrix_stride, workspace); + + _num_channels = num_channels; + _matrix_stride = matrix_stride; + + const int padding_top = (padding == PADDING_SAME) ? (KernelRows - 1) / 2 : 0; + const int padding_left = (padding == PADDING_SAME) ? (KernelCols - 1) / 2 : 0; + const int padding_bottom = (padding == PADDING_SAME) ? iceildiv(KernelRows - 1, 2) : 0; + const int padding_right = (padding == PADDING_SAME) ? iceildiv(KernelCols - 1, 2) : 0; + + _transform = std::make_unique( + KernelRows, + KernelCols, + num_batches, + num_rows, + num_cols, + num_channels, + padding_top, /**< Padding to apply to the top of the image. */ + padding_left, /**< Padding to apply to the left of the image. */ + padding_bottom, /**< Padding to apply to the bottom of the image. */ + padding_right /**< Padding to apply to the right of the image. */ + ); + + Window win; + auto win_last = _transform->get_window(); + win.set(Window::DimX, Window::Dimension(0, win_last, 1)); + ICpuKernel::configure(win); +} + +template +void CpuWinogradConv2dTransformInputKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON(tensors.empty()); + + const ITensor *input_nhwc = tensors.get_const_tensor(TensorType::ACL_SRC); + const ITensor *workspace = tensors.get_const_tensor(TensorType::ACL_INT); + ITensor *output = tensors.get_tensor(TensorType::ACL_DST); + + const int element_size_in_bytes = input_nhwc->info()->element_size(); + const int input_col_stride = input_nhwc->info()->strides_in_bytes().y() / element_size_in_bytes; + const int input_row_stride = input_nhwc->info()->strides_in_bytes().z() / element_size_in_bytes; + const int input_batch_stride = input_nhwc->info()->strides_in_bytes()[3] / element_size_in_bytes; + const auto input_nhwc_ptr = reinterpret_cast(input_nhwc->buffer() + input_nhwc->info()->offset_first_element_in_bytes()); + auto output_ptr = reinterpret_cast(output->buffer() + output->info()->offset_first_element_in_bytes()); + ARM_COMPUTE_ERROR_ON_NULLPTR(output_ptr); + + _transform->set_input_tensor(input_nhwc_ptr, input_batch_stride, input_row_stride, input_col_stride); + _transform->set_output_matrices(output_ptr, _matrix_stride, _num_channels); + + _transform->set_working_space(workspace->buffer()); + + // 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(); + _transform->run(fst, lst, info.thread_id); +} + +template +Status CpuWinogradConv2dTransformInputKernel::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 CpuWinogradConv2dTransformInputKernel; +template class CpuWinogradConv2dTransformInputKernel; +template class CpuWinogradConv2dTransformInputKernel; +template class CpuWinogradConv2dTransformInputKernel; +template class CpuWinogradConv2dTransformInputKernel; + +template class CpuWinogradConv2dTransformInputKernel; +template class CpuWinogradConv2dTransformInputKernel; +template class CpuWinogradConv2dTransformInputKernel; +template class CpuWinogradConv2dTransformInputKernel; + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +template class CpuWinogradConv2dTransformInputKernel<__fp16, 4, 4, 3, 3>; +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +// Output transform + +template +unsigned int CpuWinogradConv2dTransformOutputKernel::get_output_storage_size( + int num_batches, /* Number of batches in the output tensor. */ + int num_rows, /* Number of rows in each feature map of the input tensor. */ + int num_cols, /* Number of columns in each feature map of the input tensor. */ + int num_output_channels /* Number of feature maps in the output tensor. */ +) const +{ + // Construct shapes for the input and kernel tensors. + const Tensor4DShape input_shape(num_batches, num_rows, num_cols, 1); + const KernelShape kern_shape(num_output_channels, KernelRows, KernelCols, 1); + // Return the size, converted into units of TOut + return static_cast( + WinogradConv::get_output_storage_size(num_batches, num_rows, num_cols, num_output_channels) / sizeof(T)); +} + +template +CpuWinogradConv2dTransformOutputKernel::CpuWinogradConv2dTransformOutputKernel() + : _transform(nullptr), _matrix_stride(0), _matrix_row_stride(0) +{ +} + +template +unsigned int CpuWinogradConv2dTransformOutputKernel::get_working_space_size(unsigned int num_threads) const +{ + return _transform->get_working_space_size(num_threads) / sizeof(T); +} + +template +int CpuWinogradConv2dTransformOutputKernel::get_matrix_stride( + int num_batches, /* Number of batches in the output tensor. */ + int num_rows, /* Number of rows in each feature map of the input tensor. */ + int num_cols, /* Number of columns in each feature map of the input tensor. */ + int num_output_channels /* Number of feature maps in the output tensor. */ +) const +{ + return WinogradConv::get_output_matrix_stride(num_batches, num_rows, num_cols, num_output_channels); +} + +template +std::pair CpuWinogradConv2dTransformOutputKernel::get_output_shape( + int num_rows, /* Number of rows in each feature map of the input tensor. */ + int num_cols, /* Number of columns in each feature map of the input tensor. */ + bool padding_same) const +{ + return WinogradConv::get_output_shape(std::make_pair(num_rows, num_cols), padding_same); +} + +template +void CpuWinogradConv2dTransformOutputKernel::configure( + const ITensorInfo *biases, + const ITensorInfo *transformed_output, + const int matrix_stride, + ITensorInfo *output_nhwc, + const int num_batches, + const int num_rows, + const int num_cols, + const int num_channels, + ITensorInfo *workspace, + const arm_gemm::Activation &activation) +{ + ARM_COMPUTE_UNUSED(biases, transformed_output, output_nhwc, num_batches, num_rows, num_cols, workspace, activation); + + _matrix_stride = matrix_stride; + _matrix_row_stride = roundup(num_channels, WinogradConv::N_BLOCK); + + // 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 + _transform = std::make_unique(num_batches, num_rows, num_cols, num_channels, activation); + Window win; + auto win_last = _transform->get_window(); + win.set(Window::DimX, Window::Dimension(0, win_last, 1)); + + ICpuKernel::configure(win); +} + +template +void CpuWinogradConv2dTransformOutputKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON(tensors.empty()); + + const ITensor *biases = tensors.get_const_tensor(TensorType::ACL_SRC_0); + const ITensor *transformed_output = tensors.get_const_tensor(TensorType::ACL_SRC_1); + ITensor *workspace = tensors.get_tensor(TensorType::ACL_INT); + ITensor *dst_nhwc = tensors.get_tensor(TensorType::ACL_DST); + + const int out_batch_stride = dst_nhwc->info()->strides_in_bytes()[3] / sizeof(T); + const int out_row_stride = dst_nhwc->info()->strides_in_bytes()[2] / sizeof(T); + const int out_col_stride = dst_nhwc->info()->strides_in_bytes()[1] / sizeof(T); + + _transform->set_input_matrices(transformed_output->buffer(), _matrix_stride, _matrix_row_stride); + _transform->set_bias((biases ? reinterpret_cast(biases->buffer() + biases->info()->offset_first_element_in_bytes()) : nullptr)); + _transform->set_output_tensor(dst_nhwc->buffer() + dst_nhwc->info()->offset_first_element_in_bytes(), out_batch_stride, out_row_stride, out_col_stride); + _transform->set_working_space(workspace->buffer()); + + // 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(); + _transform->run(fst, lst, info.thread_id); +} + +template +Status CpuWinogradConv2dTransformOutputKernel::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(), output->clone().get(), winograd_info).first); + + return Status{}; +} + +template class CpuWinogradConv2dTransformOutputKernel; +template class CpuWinogradConv2dTransformOutputKernel; +template class CpuWinogradConv2dTransformOutputKernel; +template class CpuWinogradConv2dTransformOutputKernel; +template class CpuWinogradConv2dTransformOutputKernel; + +template class CpuWinogradConv2dTransformOutputKernel; +template class CpuWinogradConv2dTransformOutputKernel; +template class CpuWinogradConv2dTransformOutputKernel; +template class CpuWinogradConv2dTransformOutputKernel; + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +template class CpuWinogradConv2dTransformOutputKernel<__fp16, 4, 4, 3, 3>; +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +} // namespace cpu +} // namespace arm_compute diff --git a/src/core/cpu/kernels/CpuWinogradConv2dKernel.h b/src/core/cpu/kernels/CpuWinogradConv2dKernel.h new file mode 100644 index 0000000000..b5a29ffd02 --- /dev/null +++ b/src/core/cpu/kernels/CpuWinogradConv2dKernel.h @@ -0,0 +1,575 @@ +/* + * Copyright (c) 2017-2021 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. + */ +#ifndef ARM_COMPUTE_CPUWINOGRADCONV2DKERNEL_H +#define ARM_COMPUTE_CPUWINOGRADCONV2DKERNEL_H + +#include "src/core/NEON/kernels/convolution/common/convolution.hpp" +#include "src/core/NEON/kernels/convolution/common/tensor.hpp" +#include "src/core/cpu/ICpuKernel.h" + +#include "src/core/NEON/kernels/convolution/winograd/winograd_layer.hpp" + +namespace arm_compute +{ +namespace cpu +{ +/** Interface for the kernel to perform Winograd input transform. */ +class ICpuWinogradConv2dTransformInputKernel : public ICpuKernel +{ +public: + /** Get the working space required to perform the transformation. + * + * Note, the working space is only required when performing the + * transformation - hence it can be reused whenever the transformation is + * not running. + * + * @param num_threads The greatest number of threads that will be used to execute the transform. + * @return Size of working space required in bytes. + */ + virtual unsigned int get_working_space_size(unsigned int num_threads) const = 0; + + /** Determine how much memory (in units of TIn) to allocate for the + * transformed input. + * + * @param[in] num_batches Number of batches in the input tensor. + * @param[in] num_channels Number of feature maps in the input tensor. + * @param[in] num_rows Number of rows in each feature map. + * @param[in] num_cols Number of columns in each feature map. + * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". + * + * @return Storage size (in units of TIn) required. + */ + virtual unsigned int get_input_storage_size(int num_batches, int num_channels, int num_rows, int num_cols, bool same_padding) const = 0; + + /** Gets the stride between matrices in the input worspace + * + * @param[in] num_batches Number of batches in the input tensor. + * @param[in] num_channels Number of feature maps in the input tensor. + * @param[in] num_rows Number of rows in each feature map. + * @param[in] num_cols Number of columns in each feature map. + * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". + * + * @return Stride expressed in bytes. + */ + virtual int get_matrix_stride(int num_batches, int num_channels, int num_rows, int num_cols, bool same_padding) const = 0; + + /** Configure the output transform kernel. + * + * @param[in] input_nhwc Input tensor in NHWC data layout format. + * @param[in] num_batches Number of batches in input tensor. + * @param[in] num_rows Number of rows in input tensor. + * @param[in] num_cols Number of columns in input tensor. + * @param[in] num_channels Number of channels in input tensor. + * @param[in] padding Padding type. + * @param[out] output Base of output matrices. + * @param[in] matrix_stride Stride between output matrices. + * @param[in] workspace Tensor to be used as the working space during the computation. + */ + virtual void configure(const ITensorInfo *input_nhwc, const int num_batches, const int num_rows, const int num_cols, const int num_channels, + const PaddingType padding, ITensorInfo *output, const int matrix_stride, ITensorInfo *workspace) = 0; + + /** Destructor */ + virtual ~ICpuWinogradConv2dTransformInputKernel() + { + } +}; + +/** Kernel to perform Winograd input transform. */ +template +class CpuWinogradConv2dTransformInputKernel : public ICpuWinogradConv2dTransformInputKernel +{ +public: + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuWinogradConv2dTransformInputKernel(const CpuWinogradConv2dTransformInputKernel &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuWinogradConv2dTransformInputKernel &operator=(const CpuWinogradConv2dTransformInputKernel &) = delete; + /** Allow instances of this class to be moved */ + CpuWinogradConv2dTransformInputKernel(CpuWinogradConv2dTransformInputKernel &&) = default; + /** Allow instances of this class to be moved */ + CpuWinogradConv2dTransformInputKernel &operator=(CpuWinogradConv2dTransformInputKernel &&) = default; + /** Default destructor */ + ~CpuWinogradConv2dTransformInputKernel() = default; + + /** Determine how much memory (in units of TIn) to allocate for the + * transformed input. + * + * @param[in] num_batches Number of batches in the input tensor. + * @param[in] num_channels Number of feature maps in the input tensor. + * @param[in] num_rows Number of rows in each feature map. + * @param[in] num_cols Number of columns in each feature map. + * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". + * + * @return Storage size (in units of TIn) required. + */ + unsigned int get_input_storage_size( + int num_batches, + int num_channels, + int num_rows, + int num_cols, + bool same_padding) const override; + + /** Get the working space required to perform the transformation. + * + * Note, the working space is only required when performing the + * transformation - hence it can be reused whenever the transformation is + * not running. + * + * @param[in] num_threads The greatest number of threads that will be used to execute the transform. + * + * @return Size of working space required in bytes. + */ + unsigned int get_working_space_size(unsigned int num_threads) const override; + + /** Gets the stride between matrices in the input worspace + * + * @param[in] num_batches Number of batches in the input tensor. + * @param[in] num_channels Number of feature maps in the input tensor. + * @param[in] num_rows Number of rows in each feature map. + * @param[in] num_cols Number of columns in each feature map. + * @param[in] same_padding Use "SAME" padding, otherwise use "VALID". + * + * @return Stride expressed in bytes. + */ + int get_matrix_stride( + int num_batches, + int num_channels, + int num_rows, + int num_cols, + bool same_padding) const override; + + /** Default constructor */ + CpuWinogradConv2dTransformInputKernel(); + + const char *name() const override + { + return "CpuWinogradConv2dTransformInputKernel"; + } + + /** Configure the output transform kernel. + * + * @param[in] input_nhwc Input tensor. Data types supported: F16/F32. Layout supported NHWC. + * @param[in] num_batches Number of batches in input tensor. + * @param[in] num_rows Number of rows in input tensor. + * @param[in] num_cols Number of columns in input tensor. + * @param[in] num_channels Number of channels in input tensor. + * @param[in] padding Padding type. + * @param[out] output Base of output matrices. + * @param[in] matrix_stride Stride between output matrices. + * @param[in] workspace Tensor to be used as the working space during the computation. + */ + void configure( + const ITensorInfo *input_nhwc, + const int num_batches, + const int num_rows, + const int num_cols, + const int num_channels, + const PaddingType padding, + ITensorInfo *output, + const int matrix_stride, + ITensorInfo *workspace) override; + + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; + + /** Winograd base kernel */ + using WinogradBase = winograd::WinogradGEMM; + /** Winograd convolution kernel */ + using WinogradConv = typename WinogradBase::template Convolution; + + /** Static function to check if given info will lead to a valid configuration of @ref CpuWinogradConv2dTransformInputKernel + * + * @param[in] input First tensor input info. Data types supported: F16/F32. + * @param[in] output Output tensor info. Data types supported: same as @p input. + * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info); + +private: + using InputTransform = typename WinogradBase::template InputTransform; + + std::unique_ptr _transform{ nullptr }; + int _num_channels; /**< Number of channels in input tensor. */ + int _matrix_stride; /**< Stride between output matrices. */ +}; + +/** Interface for the kernel to perform Winograd output transform. */ +class ICpuWinogradConv2dTransformOutputKernel : public ICpuKernel +{ +public: + /** Get the working space required to perform the transformation. + * + * Note, the working space is only required when performing the + * transformation - hence it can be reused whenever the transformation is + * not running. + * + * @param[in] num_threads The greatest number of threads that will be used to execute the transform. + * + * @return Size of working space required in bytes. + */ + virtual unsigned int get_working_space_size(unsigned int num_threads) const = 0; + + /** Determine how much memory (in units of TOut) to allocate for the + * (Winograd domain) output. + * + * @param[in] num_batches Number of batches in the output tensor. + * @param[in] num_rows Number of rows in each feature map of the input tensor. + * @param[in] num_cols Number of columns in each feature map of the input tensor. + * @param[in] num_output_channels Number of feature maps in the output tensor. + * + * @return Storage size (in units of TOut) required. + */ + virtual unsigned int get_output_storage_size(int num_batches, int num_rows, int num_cols, int num_output_channels) const = 0; + + /** Gets the stride between matrices in the output worspace + * + * @param[in] num_batches Number of batches in the output tensor. + * @param[in] num_rows Number of rows in each feature map of the input tensor. + * @param[in] num_cols Number of columns in each feature map of the input tensor. + * @param[in] num_output_channels Number of feature maps in the output tensor. + * + * @return Stride expressed in bytes. + */ + virtual int get_matrix_stride(int num_batches, int num_rows, int num_cols, int num_output_channels) const = 0; + + /** Get the output shape of a convolution. + * + * @param[in] num_rows Number of rows in each feature map of the input tensor. + * @param[in] num_cols Number of columns in each feature map of the input tensor. + * @param[in] padding_same True if padding is SAME, false otherwise + * + * @return Shape of the output tensor + */ + virtual std::pair get_output_shape( + int num_rows, /* Number of rows in each feature map of the input tensor. */ + int num_cols, /* Number of columns in each feature map of the input tensor. */ + bool padding_same /* True if padding is SAME, false otherwise */ + ) const = 0; + + /** Configure the output transform kernel. + * + * @param[in] biases Pointer to the biases tensor. + * @param[in] transformed_output Pointer to working space for the output tensor in the Winograd domain. + * @param[in] matrix_stride Output matrix stride, can be computed with winograd::WinogradGEMM<2, 2, 3, 3>::Convolution::get_output_matrix_stride() + * @param[out] output_nhwc Pointer to a tensor in NHWC data layout ordered output tensor, in the spatial domain. + * @param[in] num_batches Number of batches in the input tensor. + * @param[in] num_rows Number of rows in output tensor. + * @param[in] num_cols Number of columns in output tensor. + * @param[in] num_channels Number of feature maps in the output tensor. + * @param[in] workspace Tensor to be used as the working space during the computation. + * @param[in] activation Activation to be used + */ + virtual void configure( + const ITensorInfo *biases, + const ITensorInfo *transformed_output, + const int matrix_stride, + ITensorInfo *output_nhwc, + const int num_batches, + const int num_rows, + const int num_cols, + const int num_channels, + ITensorInfo *workspace, + const arm_gemm::Activation &activation) = 0; + + virtual ~ICpuWinogradConv2dTransformOutputKernel() + { + } +}; + +/** Kernel to perform Winograd output transform. */ +template +class CpuWinogradConv2dTransformOutputKernel : public ICpuWinogradConv2dTransformOutputKernel +{ +public: + const char *name() const override + { + return "CpuWinogradConv2dTransformOutputKernel"; + } + /** Constructor */ + CpuWinogradConv2dTransformOutputKernel(); + + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuWinogradConv2dTransformOutputKernel(const CpuWinogradConv2dTransformOutputKernel &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuWinogradConv2dTransformOutputKernel &operator=(const CpuWinogradConv2dTransformOutputKernel &) = delete; + /** Allow instances of this class to be moved */ + CpuWinogradConv2dTransformOutputKernel(CpuWinogradConv2dTransformOutputKernel &&) = default; + /** Allow instances of this class to be moved */ + CpuWinogradConv2dTransformOutputKernel &operator=(CpuWinogradConv2dTransformOutputKernel &&) = default; + /** Default destructor */ + ~CpuWinogradConv2dTransformOutputKernel() = default; + + // Inherited methods overridden: + /** Determine how much memory (in units of TOut) to allocate for the + * (Winograd domain) output. + * + * @param[in] num_batches Number of batches in the output tensor. + * @param[in] num_rows Number of rows in each feature map of the input tensor. + * @param[in] num_cols Number of columns in each feature map of the input tensor. + * @param[in] num_output_channels Number of feature maps in the output tensor. + * + * @return Storage size (in units of TOut) required. + */ + unsigned int get_output_storage_size(int num_batches, int num_rows, int num_cols, int num_output_channels) const override; + + /** Gets the stride between matrices in the output worspace + * + * @param[in] num_batches Number of batches in the output tensor. + * @param[in] num_rows Number of rows in each feature map of the input tensor. + * @param[in] num_cols Number of columns in each feature map of the input tensor. + * @param[in] num_output_channels Number of feature maps in the output tensor. + * + * @return Stride expressed in bytes. + */ + int get_matrix_stride(int num_batches, int num_rows, int num_cols, int num_output_channels) const override; + /** Get the output shape of a convolution. + * + * @param[in] num_rows Number of rows in each feature map of the input tensor. + * @param[in] num_cols Number of columns in each feature map of the input tensor. + * @param[in] padding_same True if padding is SAME, false otherwise + * + * @return Shape of the output tensor + */ + std::pair get_output_shape( + int num_rows, /* Number of rows in each feature map of the input tensor. */ + int num_cols, /* Number of columns in each feature map of the input tensor. */ + bool padding_same) const override; + + /** Get the working space required to perform the transformation. + * + * Note, the working space is only required when performing the + * transformation - hence it can be reused whenever the transformation is + * not running. + * + * @param[in] num_threads The greatest number of threads that will be used to execute the transform. + * + * @return Size of working space required in bytes. + */ + unsigned int get_working_space_size(unsigned int num_threads) const override; + + /** Configure the output transform kernel. + * + * @param[in] biases Pointer to the biases tensor. + * @param[in] transformed_output Pointer to working space for the output tensor in the Winograd domain. + * @param[in] matrix_stride Output matrix stride, can be computed with winograd::WinogradGEMM<2, 2, 3, 3>::Convolution::get_output_matrix_stride() + * @param[out] output_nhwc Pointer to a tensor with NHWC data layout, in the spatial domain. + * @param[in] num_batches Number of batches in the input tensor. + * @param[in] num_rows Number of rows in output tensor. + * @param[in] num_cols Number of columns in output tensor. + * @param[in] num_channels Number of feature maps in the output tensor. + * @param[in] workspace Tensor to be used as the working space during the computation. + * @param[in] activation Activation to be used + */ + void configure( + const ITensorInfo *biases, + const ITensorInfo *transformed_output, + const int matrix_stride, + ITensorInfo *output_nhwc, + const int num_batches, + const int num_rows, + const int num_cols, + const int num_channels, + ITensorInfo *workspace, + const arm_gemm::Activation &activation) override; + + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; + + /** Static function to check if given info will lead to a valid configuration of @ref CpuWinogradConv2dTransformOutputKernel + * + * @param[in] input Source tensor info with shape [C, N, 16, batches] or [C, N, 36, batches]. Data types supported: F16/F32. + * @param[in] bias Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input + * @param[in] output Destination tensor info with shape [output_convolved_dims.width, output_convolved_dims.height, C, batches]. Data type supported: same as @p input + * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info); + +private: + using WinogradBase = winograd::WinogradGEMM; + using WinogradConv = typename WinogradBase::template Convolution; + using OutputTransform = typename WinogradBase::template OutputTransform; + + std::unique_ptr _transform{ nullptr }; + int _matrix_stride; + int _matrix_row_stride; +}; + +/** Interface for the kernel to perform Winograd weights transform. */ +class ICpuWinogradConv2dTransformWeightsKernel : public ICpuKernel +{ +public: + /** Prevent instances of this class from being copied (As this class contains pointers) */ + ICpuWinogradConv2dTransformWeightsKernel(const ICpuWinogradConv2dTransformWeightsKernel &) = default; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + ICpuWinogradConv2dTransformWeightsKernel &operator=(const ICpuWinogradConv2dTransformWeightsKernel &) = default; + /** Allow instances of this class to be moved */ + ICpuWinogradConv2dTransformWeightsKernel(ICpuWinogradConv2dTransformWeightsKernel &&) = default; + /** Allow instances of this class to be moved */ + ICpuWinogradConv2dTransformWeightsKernel &operator=(ICpuWinogradConv2dTransformWeightsKernel &&) = default; + + ICpuWinogradConv2dTransformWeightsKernel() + { + } + virtual ~ICpuWinogradConv2dTransformWeightsKernel() + { + } + /** Determine how much memory (in units of T) to allocate for the + * transformed weights. + * + * @param[in] num_output_channels Number of output feature maps. + * @param[in] num_input_channels Number of input feature maps. + * + * @return Storage size (in units of T) required. + */ + virtual unsigned int get_weight_storage_size(int num_output_channels, int num_input_channels) const = 0; + /** Gets the stride between matrices in the kernel worspace + * + * @param[in] num_output_channels Number of output feature maps. + * @param[in] num_input_channels Number of input feature maps. + * + * @return Stride expressed in bytes. + */ + virtual int get_matrix_stride(int num_output_channels, int num_input_channels) const = 0; + + /** Configure the weights transform kernel. + * + * @param[in] weights_hwio Pointer to the weights tensor info + * @param[out] output Pointer to working space for the output tensor in the Winograd domain. + * @param[in] matrix_stride Stride across matrices in the output workspace. + * @param[in] num_output_channels Number of filters. + * @param[in] num_input_channels Number of channels in each filter. + */ + + virtual void configure(const ITensorInfo *weights_hwio, ITensorInfo *output, const int matrix_stride, const int num_output_channels, const int num_input_channels) = 0; + + /** Static function to check if given info will lead to a valid configuration of @ref CpuWinogradConv2dTransformWeightsKernel + * + * @param[in] input First tensor input info. Data types supported: F16/F32. + * @param[in] weights Weights tensor info. Data types supported: same as @p input. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights); +}; + +/** Kernel to perform Winograd weights transform. */ +template +class CpuWinogradConv2dTransformWeightsKernel final : public ICpuWinogradConv2dTransformWeightsKernel +{ +public: + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuWinogradConv2dTransformWeightsKernel(const CpuWinogradConv2dTransformWeightsKernel &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CpuWinogradConv2dTransformWeightsKernel &operator=(const CpuWinogradConv2dTransformWeightsKernel &) = delete; + /** Allow instances of this class to be moved */ + CpuWinogradConv2dTransformWeightsKernel(CpuWinogradConv2dTransformWeightsKernel &&) = default; + /** Allow instances of this class to be moved */ + CpuWinogradConv2dTransformWeightsKernel &operator=(CpuWinogradConv2dTransformWeightsKernel &&) = default; + /** Default destructor */ + ~CpuWinogradConv2dTransformWeightsKernel() = default; + + /** Default constructor. */ + CpuWinogradConv2dTransformWeightsKernel(); + const char *name() const override + { + return "CpuWinogradConv2dTransformWeightsKernel"; + } + + /** Static function to check if given info will lead to a valid configuration of @ref CpuWinogradConv2dTransformWeightsKernel + * + * @param[in] input Source tensor info. The input is a 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] (NCHW data layout). + * kernel_x must be 3 and equal to kernel_y. Data types supported: F16/F32. + * @param[in] output Destination tensor info. The output is a 3D tensor with dimensions [OFM, IFM, 16] or [OFM, IFM, 36]. Data type supported: same as @p input + * @param[in] winograd_info Contains Winograd's information described in @ref WinogradInfo + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, const WinogradInfo &winograd_info); + + // Inherited methods overridden: + +#ifndef DOXYGEN_SKIP_THIS + /** Configure the weights transform kernel. + * + * @param[in] weights_hwio Pointer to the weights tensor info + * @param[out] output Pointer to working space for the output tensor in the Winograd domain. + * @param[in] matrix_stride Stride across matrices in the output workspace. + * @param[in] num_output_channels Number of filters. + * @param[in] num_input_channels Number of channels in each filter. + */ + void configure(const ITensorInfo *weights_hwio, ITensorInfo *output, const int matrix_stride, const int num_output_channels, const int num_input_channels) override; +#endif /* DOXYGEN_SKIP_THIS */ + + /** Determine how much memory (in units of T) to allocate for the + * transformed weights. + * + * @param[in] num_output_channels Number of output feature maps. + * @param[in] num_input_channels Number of input feature maps. + * + * @return Storage size (in units of T) required. + */ + unsigned int get_weight_storage_size(int num_output_channels, int num_input_channels) const override; + + /** Gets the stride between matrices in the input worspace + * + * @param[in] num_output_channels Number of output feature maps. + * @param[in] num_input_channels Number of input feature maps. + * + * @return Stride expressed in bytes. + */ + int get_matrix_stride(int num_output_channels, int num_input_channels) const override; + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; + bool is_parallelisable() const override; + +private: + using WinogradBase = winograd::WinogradGEMM; + using WinogradConv = typename WinogradBase::template Convolution; + using WeightsTransform = typename WinogradBase::template WeightsTransform; + + std::unique_ptr _transform{ nullptr }; + int _num_output_channels; + int _matrix_stride; +}; + +/** Kernel to perform Winograd. */ +template +class CpuWinogradConv2dConfiguration +{ +public: + /** Winograd base kernel */ + using WinogradBase = winograd::WinogradGEMM; + /** Winograd convolution kernel */ + + using WinogradConv = typename WinogradBase::template Convolution; + + using TransformInputKernel = CpuWinogradConv2dTransformInputKernel; + using TransformWeightsKernel = CpuWinogradConv2dTransformWeightsKernel; + using TransformOutputKernel = CpuWinogradConv2dTransformOutputKernel; +}; + +} // namespace cpu +} // namespace arm_compute +#endif /*ARM_COMPUTE_CPUWINOGRADCONV2DKERNEL_H*/ -- cgit v1.2.1