diff options
Diffstat (limited to 'src/cpu')
-rw-r--r-- | src/cpu/kernels/CpuWinogradConv2dKernel.cpp | 568 | ||||
-rw-r--r-- | src/cpu/kernels/CpuWinogradConv2dKernel.h | 533 | ||||
-rw-r--r-- | src/cpu/kernels/assembly/arm_gemm.hpp | 8 | ||||
-rw-r--r-- | src/cpu/operators/CpuWinogradConv2d.cpp | 914 | ||||
-rw-r--r-- | src/cpu/operators/CpuWinogradConv2d.h | 54 |
5 files changed, 374 insertions, 1703 deletions
diff --git a/src/cpu/kernels/CpuWinogradConv2dKernel.cpp b/src/cpu/kernels/CpuWinogradConv2dKernel.cpp index 803af09a67..818d878119 100644 --- a/src/cpu/kernels/CpuWinogradConv2dKernel.cpp +++ b/src/cpu/kernels/CpuWinogradConv2dKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,531 +21,95 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "src/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 <memory> +#include "src/cpu/kernels/CpuWinogradConv2dKernel.h" namespace arm_compute { namespace cpu { -//Batched Gemms - -namespace -{ -inline bool is_kernel_size_supported(DataType data_type, Size2D size) +CpuWinogradConv2dTransformInputKernel::CpuWinogradConv2dTransformInputKernel(arm_conv::winograd::WinogradImpl &w_impl, arm_conv::ConvolutionArgs &_c_args, uint32_t nthreads) + : _winograd_impl{ w_impl }, _conv_args{ _c_args }, _nthreads{ nthreads } { - const std::array<Size2D, 8> 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<Size2D, 8> 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) +void CpuWinogradConv2dTransformInputKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &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_UNUSED(window); + const ITensor *input_nhwc = tensors.get_const_tensor(TensorType::ACL_SRC); + const ITensor *winograd_input_transform = tensors.get_const_tensor(TensorType::ACL_DST); + const ITensor *workspace = tensors.get_const_tensor(TensorType::ACL_INT); - 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<Size2D, 8> 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)); + const unsigned int width_idx = 1; + const unsigned int height_idx = 2; + const unsigned int batch_idx = 3; + int element_size_in_bytes = input_nhwc->info()->element_size(); + const auto src_strides = input_nhwc->info()->strides_in_bytes(); - // 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)); + const size_t input_row_stride = src_strides[height_idx] / element_size_in_bytes; + const size_t input_col_stride = src_strides[width_idx] / element_size_in_bytes; + const size_t input_batch_stride = src_strides[batch_idx] / element_size_in_bytes; + const auto input_nhwc_ptr = reinterpret_cast<const void *>(input_nhwc->buffer() + input_nhwc->info()->offset_first_element_in_bytes()); + auto win_transf_ptr = reinterpret_cast<void *>(winograd_input_transform->buffer() + winograd_input_transform->info()->offset_first_element_in_bytes()); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - } - - return Status{}; + _winograd_impl.input_transform->execute( + _conv_args, + input_nhwc_ptr, + input_batch_stride, + input_row_stride, + input_col_stride, + win_transf_ptr, + _winograd_impl.winograd_spec, + workspace->buffer(), + info.thread_id, + _nthreads); } -std::pair<Status, Window> validate_and_configure_window_winograd_weight_trans(ITensorInfo *input, ITensorInfo *output, const WinogradInfo &winograd_info) +CpuWinogradConv2dTransformOutputKernel::CpuWinogradConv2dTransformOutputKernel(arm_conv::winograd::WinogradImpl &w_impl, arm_conv::ConvolutionArgs &_c_args, uint32_t nthreads) + : _winograd_impl{ w_impl }, _conv_args{ _c_args }, _nthreads{ nthreads } { - // 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) +// Inherited methods overridden: +void CpuWinogradConv2dTransformOutputKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &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_UNUSED(window); + const ITensor *dst_nhwc = tensors.get_const_tensor(TensorType::ACL_DST); + const ITensor *winograd_output_transform = tensors.get_const_tensor(TensorType::ACL_SRC_0); + const ITensor *biases = tensors.get_const_tensor(TensorType::ACL_SRC_1); + const ITensor *workspace = tensors.get_tensor(TensorType::ACL_INT); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - } - - return Status{}; -} + const unsigned int width_idx = 1; + const unsigned int height_idx = 2; + const unsigned int batch_idx = 3; + const int element_size_in_bytes = dst_nhwc->info()->element_size(); + const auto dst_strides = dst_nhwc->info()->strides_in_bytes(); -std::pair<Status, Window> 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<float> - (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<float> - (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<unsigned int, 3> 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 size_t out_row_stride = dst_strides[height_idx] / element_size_in_bytes; + const size_t out_col_stride = dst_strides[width_idx] / element_size_in_bytes; + const size_t out_batch_stride = dst_strides[batch_idx] / element_size_in_bytes; + const auto wout_transf_ptr = reinterpret_cast<const void *>(winograd_output_transform->buffer() + winograd_output_transform->info()->offset_first_element_in_bytes()); + auto dst_nhwc_ptr = reinterpret_cast<void *>(dst_nhwc->buffer() + dst_nhwc->info()->offset_first_element_in_bytes()); + void *biases_data_ptr = nullptr; + if(biases != nullptr) { - 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); + biases_data_ptr = reinterpret_cast<void *>(biases->buffer() + biases->info()->offset_first_element_in_bytes()); } - return Status{}; -} - -std::pair<Status, Window> 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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_weight_storage_size(int num_output_channels, int num_input_channels) const -{ - const KernelShape shape(num_output_channels, KernelRows, KernelCols, num_input_channels); - // WinogradConv returns the size in bytes, we divide by `sizeof(T)` to express that in units of T - return static_cast<unsigned int>(WinogradConv::get_kernel_storage_size(num_input_channels, num_output_channels) / sizeof(T)); -} - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::CpuWinogradConv2dTransformWeightsKernel() - : _transform(nullptr), _num_output_channels(0), _matrix_stride(0) -{ -} - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -int CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<WeightsTransform>(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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -bool CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::is_parallelisable() const -{ - return false; -} - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -Status CpuWinogradConv2dTransformWeightsKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<float, 2, 2, 3, 3>; -template class CpuWinogradConv2dTransformWeightsKernel<float, 4, 4, 3, 3>; -template class CpuWinogradConv2dTransformWeightsKernel<float, 2, 2, 5, 5>; -template class CpuWinogradConv2dTransformWeightsKernel<float, 1, 6, 1, 3>; -template class CpuWinogradConv2dTransformWeightsKernel<float, 6, 1, 3, 1>; - -template class CpuWinogradConv2dTransformWeightsKernel<float, 1, 4, 1, 5>; -template class CpuWinogradConv2dTransformWeightsKernel<float, 4, 1, 5, 1>; -template class CpuWinogradConv2dTransformWeightsKernel<float, 1, 2, 1, 7>; -template class CpuWinogradConv2dTransformWeightsKernel<float, 2, 1, 7, 1>; - -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template class CpuWinogradConv2dTransformWeightsKernel<__fp16, 4, 4, 3, 3>; -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - -// Input transform - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<unsigned int>(WinogradConv::get_input_storage_size(num_batches, num_rows, num_cols, num_channels, same_padding) / sizeof(T)); -} - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_working_space_size(unsigned int num_threads) const -{ - return _transform->get_working_space_size(num_threads); -} - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -int CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::CpuWinogradConv2dTransformInputKernel() - : _transform(nullptr), _num_channels(0), _matrix_stride(0) -{ -} - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<InputTransform>( - 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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<const T *>(input_nhwc->buffer() + input_nhwc->info()->offset_first_element_in_bytes()); - auto output_ptr = reinterpret_cast<T *>(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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -Status CpuWinogradConv2dTransformInputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<float, 2, 2, 3, 3>; -template class CpuWinogradConv2dTransformInputKernel<float, 4, 4, 3, 3>; -template class CpuWinogradConv2dTransformInputKernel<float, 2, 2, 5, 5>; -template class CpuWinogradConv2dTransformInputKernel<float, 1, 6, 1, 3>; -template class CpuWinogradConv2dTransformInputKernel<float, 6, 1, 3, 1>; - -template class CpuWinogradConv2dTransformInputKernel<float, 1, 4, 1, 5>; -template class CpuWinogradConv2dTransformInputKernel<float, 4, 1, 5, 1>; -template class CpuWinogradConv2dTransformInputKernel<float, 1, 2, 1, 7>; -template class CpuWinogradConv2dTransformInputKernel<float, 2, 1, 7, 1>; - -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -template class CpuWinogradConv2dTransformInputKernel<__fp16, 4, 4, 3, 3>; -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -// Output transform - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<unsigned int>( - WinogradConv::get_output_storage_size(num_batches, num_rows, num_cols, num_output_channels) / sizeof(T)); -} - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::CpuWinogradConv2dTransformOutputKernel() - : _transform(nullptr), _matrix_stride(0), _matrix_row_stride(0) -{ + // Output transform + _winograd_impl.output_transform->execute( + _conv_args, + wout_transf_ptr, + _winograd_impl.winograd_spec, + biases_data_ptr, + dst_nhwc_ptr, + out_batch_stride, + out_row_stride, + out_col_stride, + workspace->buffer(), + info.thread_id, + _nthreads); } -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -unsigned int CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::get_working_space_size(unsigned int num_threads) const -{ - return _transform->get_working_space_size(num_threads); -} - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -int CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -std::pair<unsigned int, unsigned int> CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<unsigned int, unsigned int>(num_rows, num_cols), padding_same); -} - -template <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<OutputTransform>(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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -void CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<T *>(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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -Status CpuWinogradConv2dTransformOutputKernel<T, OutputTileRows, OutputTileCols, KernelRows, KernelCols>::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<float, 2, 2, 3, 3>; -template class CpuWinogradConv2dTransformOutputKernel<float, 4, 4, 3, 3>; -template class CpuWinogradConv2dTransformOutputKernel<float, 2, 2, 5, 5>; -template class CpuWinogradConv2dTransformOutputKernel<float, 1, 6, 1, 3>; -template class CpuWinogradConv2dTransformOutputKernel<float, 6, 1, 3, 1>; - -template class CpuWinogradConv2dTransformOutputKernel<float, 1, 4, 1, 5>; -template class CpuWinogradConv2dTransformOutputKernel<float, 4, 1, 5, 1>; -template class CpuWinogradConv2dTransformOutputKernel<float, 1, 2, 1, 7>; -template class CpuWinogradConv2dTransformOutputKernel<float, 2, 1, 7, 1>; - -#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 +} // namespace arm_compute
\ No newline at end of file diff --git a/src/cpu/kernels/CpuWinogradConv2dKernel.h b/src/cpu/kernels/CpuWinogradConv2dKernel.h index 6909216d94..0170dcae22 100644 --- a/src/cpu/kernels/CpuWinogradConv2dKernel.h +++ b/src/cpu/kernels/CpuWinogradConv2dKernel.h @@ -24,550 +24,79 @@ #ifndef ARM_COMPUTE_CPUWINOGRADCONV2DKERNEL_H #define ARM_COMPUTE_CPUWINOGRADCONV2DKERNEL_H -#include "src/core/NEON/kernels/convolution/common/convolution.hpp" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Steps.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/runtime/Tensor.h" +#include "src/core/NEON/kernels/assembly/winograd.hpp" #include "src/core/NEON/kernels/convolution/common/tensor.hpp" #include "src/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<ICpuWinogradConv2dTransformInputKernel> -{ -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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -class CpuWinogradConv2dTransformInputKernel : public ICpuWinogradConv2dTransformInputKernel +class CpuWinogradConv2dTransformInputKernel final : public ICpuKernel<CpuWinogradConv2dTransformInputKernel> { 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; + /** Prevent instances of this class from being moved it contains references.*/ + CpuWinogradConv2dTransformInputKernel(CpuWinogradConv2dTransformInputKernel &&) = delete; - /** 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; + /** Prevent instances of this class from being moved it contains references.*/ + CpuWinogradConv2dTransformInputKernel &operator=(CpuWinogradConv2dTransformInputKernel &&) = delete; - /** 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; + CpuWinogradConv2dTransformInputKernel(arm_conv::winograd::WinogradImpl &w_impl, arm_conv::ConvolutionArgs &_c_args, uint32_t nthreads); - /** Default constructor */ - CpuWinogradConv2dTransformInputKernel(); + // Inherited methods overridden: + void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override; 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<OutputTileRows, OutputTileCols, KernelRows, KernelCols, winograd::WinogradRoots::Integers>; - /** Winograd convolution kernel */ - using WinogradConv = typename WinogradBase::template Convolution<T, T>; - - /** 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<T, T>; - - std::unique_ptr<InputTransform> _transform{ nullptr }; - int _num_channels; /**< Number of channels in input tensor. */ - int _matrix_stride; /**< Stride between output matrices. */ + arm_conv::winograd::WinogradImpl &_winograd_impl; + arm_conv::ConvolutionArgs &_conv_args; + uint32_t _nthreads; }; - -/** Interface for the kernel to perform Winograd output transform. */ -class ICpuWinogradConv2dTransformOutputKernel : public ICpuKernel<ICpuWinogradConv2dTransformOutputKernel> -{ -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<unsigned int, unsigned int> 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<float, float>::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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -class CpuWinogradConv2dTransformOutputKernel : public ICpuWinogradConv2dTransformOutputKernel +class CpuWinogradConv2dTransformOutputKernel : public ICpuKernel<CpuWinogradConv2dTransformOutputKernel> { 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<unsigned int, unsigned int> 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; + /** Prevent instances of this class from being moved it contains references.*/ + CpuWinogradConv2dTransformOutputKernel(CpuWinogradConv2dTransformOutputKernel &&) = delete; - /** 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; + /** Prevent instances of this class from being moved it contains references.*/ + CpuWinogradConv2dTransformOutputKernel &operator=(CpuWinogradConv2dTransformOutputKernel &&) = delete; - /** 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<float, float>::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; + CpuWinogradConv2dTransformOutputKernel(arm_conv::winograd::WinogradImpl &w_impl, arm_conv::ConvolutionArgs &_c_args, uint32_t nthreads); + // Inherited methods overridden: 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<OutputTileRows, OutputTileCols, KernelRows, KernelCols, winograd::WinogradRoots::Integers>; - using WinogradConv = typename WinogradBase::template Convolution<T, T>; - using OutputTransform = typename WinogradBase::template OutputTransform<T, T>; - - std::unique_ptr<OutputTransform> _transform{ nullptr }; - int _matrix_stride; - int _matrix_row_stride; -}; - -/** Interface for the kernel to perform Winograd weights transform. */ -class ICpuWinogradConv2dTransformWeightsKernel : public ICpuKernel<ICpuWinogradConv2dTransformWeightsKernel> -{ -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 <typename T, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -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"; + return "CpuWinogradConv2dTransformOutputKernel"; } - /** 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<OutputTileRows, OutputTileCols, KernelRows, KernelCols, winograd::WinogradRoots::Integers>; - using WinogradConv = typename WinogradBase::template Convolution<T, T>; - using WeightsTransform = typename WinogradBase::template WeightsTransform<T, T>; - - std::unique_ptr<WeightsTransform> _transform{ nullptr }; - int _num_output_channels; - int _matrix_stride; -}; - -/** Kernel to perform Winograd. */ -template <typename TIn, typename TOut, int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols> -class CpuWinogradConv2dConfiguration -{ -public: - /** Winograd base kernel */ - using WinogradBase = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols, winograd::WinogradRoots::Integers>; - /** Winograd convolution kernel */ - - using WinogradConv = typename WinogradBase::template Convolution<TIn, TOut>; - - using TransformInputKernel = CpuWinogradConv2dTransformInputKernel<TIn, OutputTileRows, OutputTileCols, KernelRows, KernelCols>; - using TransformWeightsKernel = CpuWinogradConv2dTransformWeightsKernel<TIn, OutputTileRows, OutputTileCols, KernelRows, KernelCols>; - using TransformOutputKernel = CpuWinogradConv2dTransformOutputKernel<TOut, OutputTileRows, OutputTileCols, KernelRows, KernelCols>; + arm_conv::winograd::WinogradImpl &_winograd_impl; + const arm_conv::ConvolutionArgs &_conv_args; + uint32_t _nthreads; }; } // namespace cpu diff --git a/src/cpu/kernels/assembly/arm_gemm.hpp b/src/cpu/kernels/assembly/arm_gemm.hpp index 9920b863d9..247cb1d470 100644 --- a/src/cpu/kernels/assembly/arm_gemm.hpp +++ b/src/cpu/kernels/assembly/arm_gemm.hpp @@ -143,12 +143,12 @@ struct GemmArgs { public: const CPUInfo *_ci; - unsigned int _Msize; - unsigned int _Nsize; - unsigned int _Ksize; + unsigned int _Msize; // num of tiles + unsigned int _Nsize; // output channels + unsigned int _Ksize; // input channels unsigned int _Ksections; unsigned int _nbatches; - unsigned int _nmulti; + unsigned int _nmulti; // n_gemms to be performed bool _indirect_input; Activation _act; int _maxthreads; diff --git a/src/cpu/operators/CpuWinogradConv2d.cpp b/src/cpu/operators/CpuWinogradConv2d.cpp index dcc18ce8fa..7be2d6d230 100644 --- a/src/cpu/operators/CpuWinogradConv2d.cpp +++ b/src/cpu/operators/CpuWinogradConv2d.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -31,19 +31,19 @@ #include "arm_compute/runtime/NEON/NEScheduler.h" #include "src/common/utils/Log.h" #include "src/core/CPP/Validate.h" +#include "src/core/NEON/kernels/assembly/winograd.hpp" +#include "src/core/NEON/kernels/convolution/common/tensor.hpp" #include "src/core/NEON/kernels/convolution/common/utils.hpp" -#include "src/core/NEON/kernels/convolution/winograd/winograd.hpp" #include "src/core/helpers/MemoryHelpers.h" +#include "src/core/helpers/WindowHelpers.h" +#include "src/core/utils/AssemblyUtils.h" #include "src/cpu/kernels/CpuWinogradConv2dKernel.h" +#include "src/cpu/kernels/assembly/arm_gemm.hpp" #include "src/cpu/operators/CpuActivation.h" #include "src/cpu/operators/CpuPermute.h" -#include "src/cpu/operators/CpuWinogradConv2d.h" #include "src/cpu/utils/CpuAuxTensorHandler.h" - #include "support/Cast.h" -#include <set> - namespace arm_compute { namespace cpu @@ -53,174 +53,20 @@ using namespace arm_compute::utils::cast; namespace { -arm_gemm::Activation arm_gemm_activation_from_acl_activation(const ActivationLayerInfo &act_info) -{ - switch(act_info.activation()) - { - case ActivationLayerInfo::ActivationFunction::RELU: - { - return arm_gemm::Activation(arm_gemm::Activation::Type::ReLU, act_info.a(), act_info.b()); - } - case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: - { - return arm_gemm::Activation(arm_gemm::Activation::Type::BoundedReLU, act_info.a(), act_info.b()); - } - default: - { - return arm_gemm::Activation(arm_gemm::Activation::Type::None); - } - } -} - -inline Status validate_kernel_3x3(const Size2D input_dims, const ITensorInfo *src, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); - - if(src->data_type() == DataType::F32) - { - if(input_dims.width > 4 && input_dims.height > 4) - { - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 4, 4, 3, 3>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 4, 4, 3, 3>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 4, 4, 3, 3>::validate(batched_mm_output, biases, dst, winograd_info))); - } - else - { - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 2, 2, 3, 3>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 2, 2, 3, 3>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 2, 2, 3, 3>::validate(batched_mm_output, biases, dst, winograd_info))); - } - } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - else if(src->data_type() == DataType::F16) - { - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<__fp16, 4, 4, 3, 3>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<__fp16, 4, 4, 3, 3>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<__fp16, 4, 4, 3, 3>::validate(batched_mm_output, biases, dst, winograd_info))); - } -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ - - if(act_info.enabled()) - { - CpuActivation::validate(dst, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_5x5(const ITensorInfo *src, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 2, 2, 5, 5>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 2, 2, 5, 5>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 2, 2, 5, 5>::validate(batched_mm_output, biases, dst, winograd_info))); - if(act_info.enabled()) - { - CpuActivation::validate(dst, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_3x1(const ITensorInfo *src, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 1, 6, 1, 3>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 1, 6, 1, 3>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 1, 6, 1, 3>::validate(batched_mm_output, biases, dst, winograd_info))); - if(act_info.enabled()) - { - CpuActivation::validate(dst, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_1x3(const ITensorInfo *src, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 6, 1, 3, 1>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 6, 1, 3, 1>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 6, 1, 3, 1>::validate(batched_mm_output, biases, dst, winograd_info))); - - if(act_info.enabled()) - { - CpuActivation::validate(dst, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_5x1(const ITensorInfo *src, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 1, 4, 1, 5>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 1, 4, 1, 5>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 1, 4, 1, 5>::validate(batched_mm_output, biases, dst, winograd_info))); - if(act_info.enabled()) - { - CpuActivation::validate(dst, nullptr, act_info); - } - return Status{}; -} -inline Status validate_kernel_1x5(const ITensorInfo *src, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 4, 1, 5, 1>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 4, 1, 5, 1>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 4, 1, 5, 1>::validate(batched_mm_output, biases, dst, winograd_info))); - if(act_info.enabled()) - { - CpuActivation::validate(dst, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_7x1(const ITensorInfo *src, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 1, 2, 1, 7>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 1, 2, 1, 7>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 1, 2, 1, 7>::validate(batched_mm_output, biases, dst, winograd_info))); - if(act_info.enabled()) - { - CpuActivation::validate(dst, nullptr, act_info); - } - return Status{}; -} - -inline Status validate_kernel_1x7(const ITensorInfo *src, const TensorInfo *input0, const TensorInfo *input1, const TensorInfo *batched_mm_output, - const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info) +inline Tensor4DShape internal_get_shape(const ITensorInfo *in) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformInputKernel<float, 2, 1, 7, 1>::validate(src, input0, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformWeightsKernel<float, 2, 1, 7, 1>::validate(weights, input1, winograd_info))); - ARM_COMPUTE_RETURN_ON_ERROR((CpuWinogradConv2dTransformOutputKernel<float, 2, 1, 7, 1>::validate(batched_mm_output, biases, dst, winograd_info))); - - if(act_info.enabled()) - { - CpuActivation::validate(dst, nullptr, act_info); - } - return Status{}; -} - -inline Tensor4DShape internal_get_input_shape(const ITensorInfo *src) -{ - const DataLayout data_layout = src->data_layout(); - const int in_width = src->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)); - const int in_height = src->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)); - const int in_channels = src->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL)); - const int in_batches = src->dimension(3); + const DataLayout data_layout = in->data_layout(); + const int in_width = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)); + const int in_height = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)); + const int in_channels = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL)); + const int in_batches = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES)); return Tensor4DShape{ in_batches, in_height, in_width, in_channels }; } Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info) { - ARM_COMPUTE_UNUSED(dst); + ARM_COMPUTE_UNUSED(dst, weights); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd layer only supports unit strides."); @@ -229,108 +75,85 @@ Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, co ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases); ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } - return ICpuWinogradConv2dTransformWeightsKernel::validate(src, weights); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights); + return Status{}; } -Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims, DataType data_type) + +bool get_winograd_kernel_implementation(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, + const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math, + arm_conv::winograd::WinogradImpl *winograd_impl, std::unique_ptr<arm_conv::ConvolutionArgs> &conv_args) { - Size2D output_tile = Size2D{}; - if(kernel_dims == Size2D(3U, 3U)) - { - output_tile = (input_dims.width <= 4 || input_dims.height <= 4) ? Size2D(2U, 2U) : Size2D(4U, 4U); - if(data_type == DataType::F16) - { - output_tile = Size2D(4U, 4U); - } - } - else if(kernel_dims == Size2D(5U, 5U)) - { - output_tile = Size2D(2U, 2U); - } - else if(kernel_dims == Size2D(1U, 3U)) - { - output_tile = Size2D(1U, 6U); - } - else if(kernel_dims == Size2D(3U, 1U)) - { - output_tile = Size2D(6U, 1U); - } - else if(kernel_dims == Size2D(1U, 5U)) - { - output_tile = Size2D(1U, 4U); - } - else if(kernel_dims == Size2D(5U, 1U)) - { - output_tile = Size2D(4U, 1U); - } - else if(kernel_dims == Size2D(7U, 1U)) + arm_conv::winograd::WinogradConfig winograd_cfg; + arm_gemm::GemmConfig cfg; + + const DataType data_type = src->data_type(); + Tensor4DShape in_shape{ internal_get_shape(src) }; + Tensor4DShape out_shape{ internal_get_shape(dst) }; + Tensor4DShape kernel_shape{ internal_get_shape(weights) }; + uint32_t nthreads = NEScheduler::get().num_threads(); + // Get configuration arguments for Winograd + winograd_cfg.output_rows = 0; + winograd_cfg.output_cols = 0; + conv_args = std::make_unique<arm_conv::ConvolutionArgs>( + in_shape.n_batches, + arm_conv::Shape2D{ static_cast<uint32_t>(in_shape.n_rows), static_cast<uint32_t>(in_shape.n_cols) }, + in_shape.n_channels, + conv_info.pad_top(), + conv_info.pad_left(), + arm_conv::Shape2D{ static_cast<uint32_t>(out_shape.n_rows), static_cast<uint32_t>(out_shape.n_cols) }, + out_shape.n_channels, + arm_conv::Shape2D{ static_cast<uint32_t>(kernel_shape.n_rows), static_cast<uint32_t>(kernel_shape.n_cols) }, + assembly_utils::map_to_arm_gemm_activation(act_info)); + + bool success = false; + if(data_type == DataType::F32) { - output_tile = Size2D(2U, 1U); + success = arm_conv::winograd::get_implementation<float>( + *winograd_impl, &CPUInfo::get(), *conv_args, nthreads, enable_fast_math, &winograd_cfg, nullptr); } - else if(kernel_dims == Size2D(1U, 7U)) +#if defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) + else if(data_type == DataType::F16) { - output_tile = Size2D(1U, 2U); + success = arm_conv::winograd::get_implementation<__fp16>( + *winograd_impl, &CPUInfo::get(), *conv_args, nthreads, enable_fast_math, &winograd_cfg, nullptr); } - return output_tile; -} - -bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_size, DataType data_type) -{ - // Check if we want to configure a Winograd configuration which requires fast math - using WinogradConfiguration = std::pair<std::pair<int, int>, std::pair<int, int>>; - - const std::vector<WinogradConfiguration> fast_math_winograd_f16 = - { - WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3)) - }; - - const std::vector<WinogradConfiguration> fast_math_winograd_f32 = - { - WinogradConfiguration(std::pair<int, int>(2, 2), std::pair<int, int>(5, 5)), - WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5)) - }; - - auto p = std::make_pair(std::pair<int, int>(output_tile.width, output_tile.height), - std::pair<int, int>(kernel_size.width, kernel_size.height)); - - switch(data_type) +#endif // defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) + else { - case DataType::F16: - return std::find(fast_math_winograd_f16.begin(), fast_math_winograd_f16.end(), p) != fast_math_winograd_f16.end(); - case DataType::F32: - return std::find(fast_math_winograd_f32.begin(), fast_math_winograd_f32.end(), p) != fast_math_winograd_f32.end(); - default: - return false; + success = false; } + return success; } - inline bool fuse_function_supported(const ActivationLayerInfo &act_info) { return act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU || act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU; } - } // namespace CpuWinogradConv2d::CpuWinogradConv2d() + : _gemm_function(std::make_unique<CpuGemm>()), _activation_func(std::make_unique<CpuActivation>()), + _transform_input_kernel(nullptr), + _transform_output_kernel(nullptr), _permute_input(std::make_unique<CpuPermute>()), _permute_output(std::make_unique<CpuPermute>()), _permute_weights(std::make_unique<CpuPermute>()), - _transform_input_kernel(nullptr), - _transform_weights_kernel(nullptr), - _transform_output_kernel(nullptr), - _data_layout(), _aux_mem(AuxTensorIdx::Count), - _input_nhwc(), - _output_nhwc(), + _conv_args{ nullptr }, + _winograd_impl{}, + _data_layout(), + _winograd_transformed_input{}, + _winograd_transformed_output{}, + _winograd_transformed_weights{}, _input_workspace(), - _kernel_storage(), _output_workspace(), - _input_transformed(), - _output_transformed(), _weights_hwio(), - _run_activation(false), - _is_prepared(false) + _input_nhwc(), + _output_nhwc(), + _is_prepared{ false }, + _run_activation{ false } { } @@ -342,464 +165,199 @@ void CpuWinogradConv2d::configure(const ITensorInfo *src, const ITensorInfo *wei ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, weights, biases, dst, conv_info)); ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, conv_info, act_info, enable_fast_math); - - // Get indices for the width and height - _data_layout = src->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); - const unsigned int channel_idx = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL); - - const Size2D input_dims = Size2D(src->dimension(width_idx), src->dimension(height_idx)); - const Size2D kernel_size = Size2D(weights->dimension(width_idx), weights->dimension(height_idx)); - const DataType data_type = src->data_type(); - const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, data_type); - - // Check if the Winograd configuration requires fast math - if(!enable_fast_math) - { - ARM_COMPUTE_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size, data_type), - "This Winograd configuration requires enable_fast_math=true"); - } - - _is_prepared = false; - - std::unique_ptr<ICpuWinogradConv2dTransformInputKernel> transform_input_kernel; - std::unique_ptr<ICpuWinogradConv2dTransformWeightsKernel> transform_weights_kernel; - std::unique_ptr<ICpuWinogradConv2dTransformOutputKernel> transform_output_kernel; - - int n_gemms = 1; - int N_BLOCK = 1; // Size of block used by GEMM. - if(data_type == DataType::F32) - { - if(kernel_size == Size2D(3, 3)) + ARM_COMPUTE_UNUSED(biases); + const DataType data_type = src->data_type(); + uint32_t nthreads = NEScheduler::get().num_threads(); + _data_layout = src->data_layout(); + const Tensor4DShape kernel_shape{ internal_get_shape(weights) }; + + bool success = get_winograd_kernel_implementation(src, weights, dst, conv_info, act_info, enable_fast_math, &_winograd_impl, _conv_args); + + ARM_COMPUTE_EXIT_ON_MSG_VAR(!success, "Unsupported kernel size: %d x %d.\n", kernel_shape.n_rows, kernel_shape.n_cols); + ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using input transform: %s\n", _winograd_impl.input_transform->get_name().c_str()); + ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using weight transform: %s\n", _winograd_impl.input_transform->get_name().c_str()); + ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using output transform: %s\n", _winograd_impl.input_transform->get_name().c_str()); + + const bool has_impl = ((_winograd_impl.input_transform != nullptr) && (_winograd_impl.output_transform != nullptr) && (_winograd_impl.gemm_args != nullptr)); + if(has_impl) + { + // Determine how much working space is required, allocate it. + const size_t input_workspace_size = _winograd_impl.input_transform->get_working_space_size(*_conv_args, nthreads); + const size_t output_workspace_size = _winograd_impl.output_transform->get_working_space_size(*_conv_args, nthreads); + + TensorInfo input_workspace_info(TensorShape(input_workspace_size), 1, DataType::U8); + TensorInfo output_workspace_info(TensorShape(output_workspace_size), 1, DataType::U8); + _input_workspace = input_workspace_info; + _output_workspace = output_workspace_info; + + const auto &wds = _winograd_impl.winograd_spec; + + // Preparing winograd transformed input tensor + const size_t data_type_size = src->element_size(); + const uint32_t m = _winograd_impl.gemm_args->_Msize; // Total number of tiles + const uint32_t k = _winograd_impl.gemm_args->_Ksize; // Input channels + const uint32_t n = _winograd_impl.gemm_args->_Nsize; // Output channels + const uint32_t n_gemms = _winograd_impl.gemm_args->_nmulti; + const uint32_t n_batches = _winograd_impl.gemm_args->_nbatches; + constexpr size_t storage_alignment = 64; + + const TensorShape a_shape(k, m, n_batches, n_gemms); + Strides a_strides(data_type_size); + a_strides.set(1, data_type_size * _winograd_impl.winograd_spec.input_ld_row); + a_strides.set(2, data_type_size * _winograd_impl.winograd_spec.input_ld_batch); + a_strides.set(3, data_type_size * _winograd_impl.winograd_spec.input_ld_matrix); + + const TensorShape b_shape(n, k, n_gemms); + Strides b_strides(data_type_size); + b_strides.set(1, data_type_size * _winograd_impl.winograd_spec.weight_ld_row); + b_strides.set(2, data_type_size * _winograd_impl.winograd_spec.weight_ld_matrix); + + const TensorShape d_shape(n, m, n_batches, n_gemms); + Strides d_strides(data_type_size); + d_strides.set(1, data_type_size * _winograd_impl.winograd_spec.output_ld_row); + d_strides.set(2, data_type_size * _winograd_impl.winograd_spec.output_ld_batch); + d_strides.set(3, data_type_size * _winograd_impl.winograd_spec.output_ld_matrix); + + TensorInfo a_info{}; + TensorInfo b_info{}; + TensorInfo d_info{}; + a_info.init(a_shape, 1, data_type, a_strides, 0, wds.input_matrix_size_bytes); + b_info.init(b_shape, 1, data_type, b_strides, 0, wds.weight_matrix_size_bytes); + d_info.init(d_shape, 1, data_type, d_strides, 0, wds.output_matrix_size_bytes); + + _winograd_transformed_input = a_info; + _winograd_transformed_weights = b_info; + _winograd_transformed_output = d_info; + + PermutationVector weights_permutation_vector(3U, 0U, 1U, 2U); + + // Configure the kernel to transform the input tensor from NCHW -> NHWC + if(_data_layout == DataLayout::NCHW) { - if(src->dimension(width_idx) > 4 && src->dimension(height_idx) > 4) - { - using config = CpuWinogradConv2dConfiguration<float, float, 4, 4, 3, 3>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else - { - using config = CpuWinogradConv2dConfiguration<float, float, 2, 2, 3, 3>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } + _permute_input->configure(src, &_input_nhwc, PermutationVector(2U, 0U, 1U)); + weights_permutation_vector = PermutationVector(3U, 2U, 0U, 1U); } - else if(kernel_size == Size2D(5, 5)) - { - using config = CpuWinogradConv2dConfiguration<float, float, 2, 2, 5, 5>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(1, 3)) - { - using config = CpuWinogradConv2dConfiguration<float, float, 6, 1, 3, 1>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(3, 1)) - { - using config = CpuWinogradConv2dConfiguration<float, float, 1, 6, 1, 3>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(1, 5)) - { - using config = CpuWinogradConv2dConfiguration<float, float, 4, 1, 5, 1>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(5, 1)) - { - using config = CpuWinogradConv2dConfiguration<float, float, 1, 4, 1, 5>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(1, 7)) - { - using config = CpuWinogradConv2dConfiguration<float, float, 2, 1, 7, 1>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else if(kernel_size == Size2D(7, 1)) - { - using config = CpuWinogradConv2dConfiguration<float, float, 1, 2, 1, 7>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; - } - else + + // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map] + _permute_weights->configure(weights, &_weights_hwio, weights_permutation_vector); + + // Reorder the convoluted output to ACL's ordering NCHW + if(_data_layout == DataLayout::NCHW) { - ARM_COMPUTE_ERROR("Not supported."); + // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output() + TensorInfo info(TensorShape(dst->dimension(2), dst->dimension(0), + dst->dimension(1), dst->dimension(3)), + 1, dst->data_type()); + _output_nhwc = info; + _permute_output->configure(&_output_nhwc, dst, PermutationVector(1U, 2U, 0U)); } - } -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - else if(data_type == DataType::F16) - { - if(kernel_size == Size2D(3, 3)) + + // Configure GEMM function + _gemm_function->configure(&_winograd_transformed_input, &_winograd_transformed_weights, nullptr, &_winograd_transformed_output, 1.0f, 0.f); + + //Configure Activation Layer + _run_activation = act_info.enabled() && !fuse_function_supported(act_info); + if(_run_activation) { - using config = CpuWinogradConv2dConfiguration<__fp16, __fp16, 4, 4, 3, 3>; - transform_input_kernel = std::make_unique<config::TransformInputKernel>(); - transform_weights_kernel = std::make_unique<config::TransformWeightsKernel>(); - transform_output_kernel = std::make_unique<config::TransformOutputKernel>(); - n_gemms = config::WinogradBase::N_GEMMS; - N_BLOCK = config::WinogradConv::N_BLOCK; + _activation_func->configure(dst, nullptr, act_info); } - else + + auto asm_mem_req = _gemm_function->workspace(); + _aux_mem[GemmWorkspace] = asm_mem_req[GemmWorkspace]; + _aux_mem[Pretranspose] = asm_mem_req[Pretranspose]; + _aux_mem[InterleavedLHS] = asm_mem_req[InterleavedLHS]; + _aux_mem[TransposedRHS] = asm_mem_req[TransposedRHS]; + _aux_mem[TempResult] = asm_mem_req[TempResult]; + + // Request temporary memory. Overlap memory needed for Input/Output transformations as they run on different non-overlapping time-steps. + _aux_mem[TransformedInput] = MemoryInfo(offset_int_vec(TransformedInput), MemoryLifetime::Temporary, wds.input_matrix_size_bytes, storage_alignment); + _aux_mem[TransformedOutput] = MemoryInfo(offset_int_vec(TransformedOutput), MemoryLifetime::Temporary, wds.output_matrix_size_bytes, storage_alignment); + _aux_mem[WorkspaceIO] = MemoryInfo(offset_int_vec(WorkspaceIO), MemoryLifetime::Temporary, std::max(input_workspace_size, output_workspace_size)); + _aux_mem[PermutedWeights] = MemoryInfo(offset_int_vec(PermutedWeights), MemoryLifetime::Prepare, _weights_hwio.total_size()); + _aux_mem[TransformedWeights] = MemoryInfo(offset_int_vec(TransformedWeights), MemoryLifetime::Persistent, wds.weight_matrix_size_bytes, storage_alignment); + if(_data_layout == DataLayout::NCHW) { - ARM_COMPUTE_ERROR("Not supported."); + _aux_mem[PermutedInput].merge(offset_int_vec(PermutedInput), src->total_size()); + _aux_mem[PermutedOutput].merge(offset_int_vec(PermutedOutput), dst->total_size()); } } -#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - else - { - ARM_COMPUTE_ERROR("Not supported."); - } - - const PaddingType use_padding_type = (conv_info.pad_top() != 0u || conv_info.pad_left() != 0) ? PADDING_SAME : PADDING_VALID; - const bool use_same_padding = use_padding_type == PADDING_SAME; - - // Get convolved dimensions - const int in_channels = src->dimension(channel_idx); - const int out_channels = dst->dimension(channel_idx); - - const Tensor4DShape in_shape(internal_get_input_shape(src)); - const size_t data_type_size = src->element_size(); - // Get the memory required to instantiate a new Winograd operator. - constexpr size_t storage_alignment = 64; - - // Kernel Storage - const size_t kernel_storage_size = transform_weights_kernel->get_weight_storage_size(out_channels, in_channels) * data_type_size; - - // Input storage - const size_t input_storage_size = transform_input_kernel->get_input_storage_size(in_shape.n_batches, in_shape.n_channels, in_shape.n_rows, in_shape.n_cols, use_same_padding) * data_type_size; - - // Output storage - const size_t output_storage_size = transform_output_kernel->get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels) * data_type_size; - const int kernel_matrix_stride = transform_weights_kernel->get_matrix_stride(out_channels, in_channels); - const int output_matrix_stride = transform_output_kernel->get_matrix_stride(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels); - const auto output_shape = transform_output_kernel->get_output_shape(in_shape.n_rows, in_shape.n_cols, use_padding_type == PADDING_SAME); - const int input_matrix_stride = transform_input_kernel->get_matrix_stride(in_shape.n_batches, in_channels, in_shape.n_rows, in_shape.n_cols, use_padding_type == PADDING_SAME); - - // Configure GEMM - const int tile_rows = iceildiv(output_shape.first, output_tile.height); - const int tile_cols = iceildiv(output_shape.second, output_tile.width); - const int m = in_shape.n_batches * tile_rows * tile_cols; - const int k = in_shape.n_channels; - const int n = out_channels; - const int kernel_matrix_row_stride = roundup(out_channels, N_BLOCK); - const int output_matrix_row_stride = kernel_matrix_row_stride; - - TensorShape a_shape(k, m, 1, n_gemms); - Strides a_strides(data_type_size); - a_strides.set(1, a_strides[0] * k); - //a_strides.set(2, data_type_size * input_matrix_stride / n_gemms); FIXME: This is the real batch size, but RSH's code crashes if it's not 0. - a_strides.set(2, 0); - a_strides.set(3, data_type_size * input_matrix_stride); - - TensorShape b_shape(n, k, n_gemms); - Strides b_strides(data_type_size); - b_strides.set(1, data_type_size * kernel_matrix_row_stride); - b_strides.set(2, data_type_size * kernel_matrix_stride); - - TensorShape d_shape(n, m, 1, n_gemms); - Strides d_strides(data_type_size); - d_strides.set(1, data_type_size * output_matrix_row_stride); - //d_strides.set(2, data_type_size * output_matrix_stride / n_gemms); FIXME: This is the real batch size, but RSH's code crashes if it's not 0. - d_strides.set(2, 0); - d_strides.set(3, data_type_size * output_matrix_stride); - - TensorInfo a_info{}; - TensorInfo b_info{}; - TensorInfo d_info{}; - a_info.init(a_shape, 1, data_type, a_strides, 0, input_storage_size); - b_info.init(b_shape, 1, data_type, b_strides, 0, kernel_storage_size); - d_info.init(d_shape, 1, data_type, d_strides, 0, output_storage_size); - - _input_transformed = a_info; - _kernel_storage = b_info; - _output_transformed = d_info; - - const ITensorInfo *input_to_use = src; - ITensorInfo *output_to_use = dst; - PermutationVector weights_permutation_vector(3U, 0U, 1U, 2U); - const unsigned int max_num_threads = NEScheduler::get().num_threads(); - - // Configure the kernel to transform the input tensor from NCHW -> NHWC - if(_data_layout == DataLayout::NCHW) - { - _permute_input->configure(src, &_input_nhwc, PermutationVector(2U, 0U, 1U)); - input_to_use = &_input_nhwc; - weights_permutation_vector = PermutationVector(3U, 2U, 0U, 1U); - } - - // Configure input transform kernel - transform_input_kernel->configure(input_to_use, in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, in_shape.n_channels, use_padding_type, - &_input_transformed, input_matrix_stride, &_input_workspace); - const size_t input_workspace_size = transform_input_kernel->get_working_space_size(max_num_threads); - TensorInfo input_workspace_info(TensorShape(input_workspace_size), 1, DataType::U8); - _input_workspace = input_workspace_info; - - // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map] - _permute_weights->configure(weights, &_weights_hwio, weights_permutation_vector); - transform_weights_kernel->configure(&_weights_hwio, &_kernel_storage, kernel_matrix_stride, out_channels, in_channels); - - // Configure GEMM function - _gemm_function->configure(&_input_transformed, &_kernel_storage, nullptr, &_output_transformed, 1.0f, 0.f); - - // Configure output transform function - // The biases tensor has not been allocated at this point in time, the output transform will add the biases to the final result in the run() method - if(_data_layout == DataLayout::NCHW) - { - // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output() - TensorInfo info(TensorShape(dst->dimension(2), dst->dimension(0), - dst->dimension(1), dst->dimension(3)), - 1, dst->data_type()); - _output_nhwc = info; - output_to_use = &_output_nhwc; - } - const arm_gemm::Activation activation = arm_gemm_activation_from_acl_activation(act_info); - - transform_output_kernel->configure(biases, - &_output_transformed, - output_matrix_stride, - output_to_use, - in_shape.n_batches, - output_shape.first, - output_shape.second, - out_channels, - &_output_workspace, - activation); - - const size_t output_workspace_size = transform_output_kernel->get_working_space_size(max_num_threads); - TensorInfo output_workspace_info(TensorShape(output_workspace_size), 1, DataType::U8); - _output_workspace = output_workspace_info; - - // Reorder the convoluted output to ACL's ordering NCHW - if(_data_layout == DataLayout::NCHW) - { - _permute_output->configure(&_output_nhwc, dst, PermutationVector(1U, 2U, 0U)); - } - - _transform_input_kernel = std::move(transform_input_kernel); - _transform_weights_kernel = std::move(transform_weights_kernel); - _transform_output_kernel = std::move(transform_output_kernel); - - //Configure Activation Layer - _run_activation = act_info.enabled() && !fuse_function_supported(act_info); - if(_run_activation) - { - _activation_func->configure(dst, nullptr, act_info); - } - - auto asm_mem_req = _gemm_function->workspace(); - _aux_mem[GemmWorkspace] = asm_mem_req[GemmWorkspace]; - _aux_mem[Pretranspose] = asm_mem_req[Pretranspose]; - _aux_mem[InterleavedLHS] = asm_mem_req[InterleavedLHS]; - _aux_mem[TransposedRHS] = asm_mem_req[TransposedRHS]; - _aux_mem[TempResult] = asm_mem_req[TempResult]; - - // Request temporary memory. Overlap memory needed for Input/Output transformations as they run on different non-overlapping time-steps. - _aux_mem[TransformedInput] = MemoryInfo(offset_int_vec(TransformedInput), MemoryLifetime::Temporary, input_storage_size, storage_alignment); - _aux_mem[TransformedOutput] = MemoryInfo(offset_int_vec(TransformedOutput), MemoryLifetime::Temporary, output_storage_size, storage_alignment); - _aux_mem[WorkspaceIO] = MemoryInfo(offset_int_vec(WorkspaceIO), MemoryLifetime::Temporary, std::max(input_workspace_size, output_workspace_size)); - _aux_mem[PermutedWeights] = MemoryInfo(offset_int_vec(PermutedWeights), MemoryLifetime::Prepare, _weights_hwio.total_size()); - _aux_mem[TransformedWeights] = MemoryInfo(offset_int_vec(TransformedWeights), MemoryLifetime::Persistent, kernel_storage_size, storage_alignment); - if(_data_layout == DataLayout::NCHW) - { - _aux_mem[PermutedInput].merge(offset_int_vec(PermutedInput), src->total_size()); - _aux_mem[PermutedOutput].merge(offset_int_vec(PermutedOutput), dst->total_size()); - } } - Status CpuWinogradConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, conv_info)); - // Get indices for the width and height - const size_t idx_width = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::WIDTH); - const size_t idx_height = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::HEIGHT); + const Tensor4DShape kernel_shape{ internal_get_shape(weights) }; + arm_conv::winograd::WinogradImpl winograd_impl{}; - // Input shape, kernel size and output tile - const Size2D input_dims = Size2D(src->dimension(idx_width), src->dimension(idx_height)); - const Size2D kernel_size = Size2D(weights->dimension(idx_width), weights->dimension(idx_height)); - const DataType data_type = src->data_type(); - const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, data_type); + std::unique_ptr<arm_conv::ConvolutionArgs> conv_args; + const bool success = get_winograd_kernel_implementation(src, weights, dst, conv_info, act_info, enable_fast_math, &winograd_impl, conv_args); - // Check if the Winograd configuration requires fast math - if(!enable_fast_math) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size, data_type), - "This Winograd configuration requires enable_fast_math=true"); - } - - const WinogradInfo winograd_info = WinogradInfo(output_tile, - kernel_size, - input_dims, - conv_info, - src->data_layout()); - - // Validate input transform - const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*src, winograd_info); - const TensorInfo input0 = src->clone()->set_tensor_shape(input0_shape); - // Validate filter transform - const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, winograd_info); - const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); - // Validate batched matrix multiply - TensorShape batched_mm_output_shape = input0.tensor_shape(); - batched_mm_output_shape[0] = input1.tensor_shape()[0]; - const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); - - if(kernel_size == Size2D(3, 3)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != conv_info.pad_left(), "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_bottom(), "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_left(), "Only SAME or VALID padding supported"); - return validate_kernel_3x3(input_dims, src, &input0, &input1, &batched_mm_output, weights, biases, dst, winograd_info, act_info); - } - else if(kernel_size == Size2D(5, 5)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != conv_info.pad_left(), "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_bottom(), "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != conv_info.pad_left(), "Only SAME or VALID padding supported"); - return validate_kernel_5x5(src, &input0, &input1, &batched_mm_output, weights, biases, dst, winograd_info, act_info); - } - if(kernel_size == Size2D(3, 1)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_bottom() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_3x1(src, &input0, &input1, &batched_mm_output, weights, biases, dst, winograd_info, act_info); - } - else if(kernel_size == Size2D(1, 3)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 1, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_right() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_1x3(src, &input0, &input1, &batched_mm_output, weights, biases, dst, winograd_info, act_info); - } - else if(kernel_size == Size2D(5, 1)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_bottom() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_5x1(src, &input0, &input1, &batched_mm_output, weights, biases, dst, winograd_info, act_info); - } - else if(kernel_size == Size2D(1, 5)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 2, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_right() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_1x5(src, &input0, &input1, &batched_mm_output, weights, biases, dst, winograd_info, act_info); - } - else if(kernel_size == Size2D(7, 1)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_left() != 3, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_right() != 0u && conv_info.pad_right() != 3, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_bottom() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_7x1(src, &input0, &input1, &batched_mm_output, weights, biases, dst, winograd_info, act_info); - } - else if(kernel_size == Size2D(1, 7)) - { - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_top() != 0u && conv_info.pad_top() != 3, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_bottom() != 0u && conv_info.pad_bottom() != 3, "Only SAME or VALID padding supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.pad_left() != 0u && conv_info.pad_right() != 0, "Only SAME or VALID padding supported"); - return validate_kernel_1x7(src, &input0, &input1, &batched_mm_output, weights, biases, dst, winograd_info, act_info); - } - else - { - ARM_COMPUTE_RETURN_ERROR_MSG("Kernel shape not supported"); - } + ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR(success == false, "Unsupported kernel size: %d x %d.\n", kernel_shape.n_rows, kernel_shape.n_cols); + ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using input transform: %s\n", winograd_impl.input_transform->get_name().c_str()); + ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using weight transform: %s\n", winograd_impl.input_transform->get_name().c_str()); + ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using output transform: %s\n", winograd_impl.input_transform->get_name().c_str()); + return Status{}; } void CpuWinogradConv2d::run(ITensorPack &tensors) { prepare(tensors); + auto src = tensors.get_const_tensor(ACL_SRC_0); + auto biases = tensors.get_const_tensor(ACL_SRC_2); + auto output = tensors.get_tensor(ACL_DST); + Window win; - auto a = tensors.get_const_tensor(ACL_SRC_0); - auto c = tensors.get_const_tensor(ACL_SRC_2); - auto d = tensors.get_tensor(ACL_DST); + const uint32_t nthreads = NEScheduler::get().num_threads(); + // The Winograd transform implementation does fine-grain threading inside the transforms. Just pass thread_id and nthreads. + win.set(Window::DimX, Window::Dimension(0, nthreads, 1)); + + // Wrap the winograd-domain tensorInfos created in configuration in tensors and allocate the required memory. CpuAuxTensorHandler input_nhwc(offset_int_vec(PermutedInput), _input_nhwc, tensors, true); - CpuAuxTensorHandler input_transformed(offset_int_vec(TransformedInput), _input_transformed, tensors, true); + CpuAuxTensorHandler winograd_input_transformed(offset_int_vec(TransformedInput), _winograd_transformed_input, tensors, true); CpuAuxTensorHandler input_workspace(offset_int_vec(WorkspaceIO), _input_workspace, tensors, true); - - const bool is_nchw = _data_layout == DataLayout::NCHW; + const bool is_nchw = _data_layout == DataLayout::NCHW; if(is_nchw) { //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC - ITensorPack pack{ { ACL_SRC, a }, { ACL_DST, input_nhwc.get() } }; + ITensorPack pack{ { ACL_SRC, src }, { ACL_DST, input_nhwc.get() } }; _permute_input->run(pack); } - // Transform input tensor to the winograd domain - ITensorPack transform_input_pack{ { ACL_SRC, is_nchw ? input_nhwc.get() : a }, { ACL_DST, input_transformed.get() }, { ACL_INT, input_workspace.get() } }; - NEScheduler::get().schedule_op(_transform_input_kernel.get(), Window::DimX, _transform_input_kernel->window(), transform_input_pack); + CpuAuxTensorHandler winograd_output_transformed(offset_int_vec(TransformedOutput), _winograd_transformed_output, tensors, true); + CpuAuxTensorHandler output_workspace(offset_int_vec(WorkspaceIO), _output_workspace, tensors, true); + CpuAuxTensorHandler output_nhwc(offset_int_vec(PermutedOutput), _output_nhwc, tensors, true); + + ITensorPack transform_input_pack{ { ACL_SRC, is_nchw ? input_nhwc.get() : src }, { ACL_DST, winograd_input_transformed.get() }, { ACL_INT, input_workspace.get() } }; + _transform_input_kernel = std::make_unique<CpuWinogradConv2dTransformInputKernel>(_winograd_impl, *_conv_args, nthreads); - CpuAuxTensorHandler output_transformed(offset_int_vec(TransformedOutput), _output_transformed, tensors, true); - CpuAuxTensorHandler weights_transformed(offset_int_vec(TransformedWeights), _kernel_storage, tensors, true); + NEScheduler::get().schedule_op(_transform_input_kernel.get(), Window::DimX, win, transform_input_pack); + + CpuAuxTensorHandler winograd_weights_transformed(offset_int_vec(TransformedWeights), _winograd_transformed_weights, tensors, true); // Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs ITensorPack gemm_pack = tensors; - gemm_pack.add_const_tensor(ACL_SRC, input_transformed.get()); - gemm_pack.add_const_tensor(ACL_SRC_1, weights_transformed.get()); + gemm_pack.add_const_tensor(ACL_SRC, winograd_input_transformed.get()); + gemm_pack.add_const_tensor(ACL_SRC_1, winograd_weights_transformed.get()); gemm_pack.add_const_tensor(ACL_BIAS, nullptr); - gemm_pack.add_tensor(ACL_DST, output_transformed.get()); + gemm_pack.add_tensor(ACL_DST, winograd_output_transformed.get()); _gemm_function->run(gemm_pack); - // Transform output tensor to the spatial domain - CpuAuxTensorHandler output_workspace(offset_int_vec(WorkspaceIO), _output_workspace, tensors, true); - CpuAuxTensorHandler output_nhwc(offset_int_vec(PermutedOutput), _output_nhwc, tensors, true); - ITensorPack transform_output_pack{ { ACL_SRC_0, c }, { ACL_SRC_1, output_transformed.get() }, { ACL_DST, is_nchw ? output_nhwc.get() : d }, { ACL_INT, output_workspace.get() } }; - NEScheduler::get().schedule_op(_transform_output_kernel.get(), Window::DimX, _transform_output_kernel->window(), transform_output_pack); - + // Output transform + _transform_output_kernel = std::make_unique<CpuWinogradConv2dTransformOutputKernel>(_winograd_impl, *_conv_args, nthreads); + ITensorPack transform_output_pack{ { ACL_SRC_0, winograd_output_transformed.get() }, { ACL_DST, is_nchw ? output_nhwc.get() : output }, { ACL_SRC_1, biases }, { ACL_INT, output_workspace.get() } }; + NEScheduler::get().schedule_op(_transform_output_kernel.get(), Window::DimX, win, transform_output_pack); if(is_nchw) { // Reorder the convoluted output to ACL's ordering NCHW - ITensorPack pack{ { ACL_SRC, output_nhwc.get() }, { ACL_DST, d } }; + ITensorPack pack{ { ACL_SRC, output_nhwc.get() }, { ACL_DST, output } }; _permute_output->run(pack); } - if(_run_activation) { - ITensorPack pack{ { ACL_SRC, d }, { ACL_DST, d } }; + ITensorPack pack{ { ACL_SRC, output }, { ACL_DST, output } }; _activation_func->run(pack); } } @@ -808,34 +366,54 @@ void CpuWinogradConv2d::prepare(ITensorPack &tensors) { if(!_is_prepared) { - // Permute weights const ITensor *weights = tensors.get_const_tensor(ACL_SRC_1); ITensor *weights_aux = utils::cast::polymorphic_cast<ITensor *>(tensors.get_tensor(offset_int_vec(PermutedWeights))); - ARM_COMPUTE_ERROR_ON_NULLPTR(weights, weights_aux); CpuAuxTensorHandler permuted_weights(_weights_hwio, *weights_aux); ITensorPack permute_tensors{ { ACL_SRC, weights }, { ACL_DST, permuted_weights.get() } }; _permute_weights->run(permute_tensors); + const int element_size_in_bytes = permuted_weights.get()->info()->element_size(); + // Weights were in OHWI format, before being permuted "permuted_weights" to be in HWIO format. + const unsigned int height_idx = 3; // H in HWIO + const unsigned int width_idx = 2; // W in HWIO + const unsigned int channel_idx = 1; // I in HWIO - // Transform weights + const int permuted_weight_row_stride = permuted_weights.get()->info()->strides_in_bytes()[height_idx] / element_size_in_bytes; + const int permuted_weight_col_stride = permuted_weights.get()->info()->strides_in_bytes()[width_idx] / element_size_in_bytes; + const int permuted_weight_channel_stride = permuted_weights.get()->info()->strides_in_bytes()[channel_idx] / element_size_in_bytes; + + // Wrap the winograd-domain transformed weight TensorInfo in Auxiliary tensor and allocate the required memory. ITensor *weights_transf = utils::cast::polymorphic_cast<ITensor *>(tensors.get_tensor(offset_int_vec(TransformedWeights))); ARM_COMPUTE_ERROR_ON_NULLPTR(weights_transf); - - CpuAuxTensorHandler transformed_weights(_kernel_storage, *weights_transf); - ITensorPack transform_tensors{ { ACL_SRC, permuted_weights.get() }, { ACL_DST, transformed_weights.get() } }; - NEScheduler::get().schedule_op(_transform_weights_kernel.get(), Window::DimX, _transform_weights_kernel->window(), transform_tensors); - + CpuAuxTensorHandler winograd_transformed_weights(_winograd_transformed_weights, *weights_transf); + + const void *permuted_weights_ptr; + void *win_wght_transf_ptr; + + permuted_weights_ptr = reinterpret_cast<const void *>(permuted_weights.get()->buffer() + permuted_weights.get()->info()->offset_first_element_in_bytes()); + win_wght_transf_ptr = reinterpret_cast<void *>(winograd_transformed_weights.get()->buffer() + winograd_transformed_weights.get()->info()->offset_first_element_in_bytes()); + + // Prepare Weights + _winograd_impl.weight_transform->execute( + *_conv_args, + permuted_weights_ptr, + permuted_weight_row_stride, + permuted_weight_col_stride, + permuted_weight_channel_stride, + win_wght_transf_ptr, + _winograd_impl.winograd_spec, + 0, 1 // Thread 1 of 1 + ); ITensorPack gemm_pack = tensors; - gemm_pack.add_const_tensor(ACL_SRC_1, transformed_weights.get()); + gemm_pack.add_const_tensor(ACL_SRC_1, winograd_transformed_weights.get()); _gemm_function->prepare(gemm_pack); - - _is_prepared = true; + _is_prepared = 1; } } - experimental::MemoryRequirements CpuWinogradConv2d::workspace() const { return _aux_mem; } + } // namespace cpu -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute diff --git a/src/cpu/operators/CpuWinogradConv2d.h b/src/cpu/operators/CpuWinogradConv2d.h index 0abd110f73..e0df34e2db 100644 --- a/src/cpu/operators/CpuWinogradConv2d.h +++ b/src/cpu/operators/CpuWinogradConv2d.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -29,6 +29,7 @@ #include "src/core/common/Macros.h" #include "src/cpu/ICpuOperator.h" #include "src/cpu/kernels/CpuWinogradConv2dKernel.h" +#include "src/cpu/kernels/assembly/gemm_common.hpp" #include "src/cpu/operators/CpuActivation.h" #include "src/cpu/operators/CpuGemm.h" #include "src/cpu/operators/CpuPermute.h" @@ -59,13 +60,13 @@ public: * |F16 |F16 |F16 |F16 | * |F32 |F32 |F32 |F32 | * - * @param[in] src Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], + * @param[in] src Source tensor Info. 3 lower dimensions represent a single input [width, height, IFM], * while every optional dimension from 4 and above represent a batch of inputs. * Data types supported: F16/F32. - * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. + * @param[in] weights Weights tensor Info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. * Currently only 3x3 and 5x5 kernels are supported. - * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. - * @param[out] dst Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * @param[in] biases Biases tensor Info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. + * @param[out] dst Destination tensor Info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. * Data types supported: Same as @p input. * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. @@ -107,28 +108,27 @@ private: PermutedOutput = TransformedInput, Count = 10 }; - - std::unique_ptr<CpuGemm> _gemm_function; - std::unique_ptr<CpuActivation> _activation_func; - std::unique_ptr<CpuPermute> _permute_input; - std::unique_ptr<CpuPermute> _permute_output; - std::unique_ptr<CpuPermute> _permute_weights; - std::unique_ptr<ICPPKernel> _transform_input_kernel; - std::unique_ptr<ICPPKernel> _transform_weights_kernel; - std::unique_ptr<ICPPKernel> _transform_output_kernel; - - DataLayout _data_layout; - experimental::MemoryRequirements _aux_mem{ Count }; - TensorInfo _input_nhwc; - TensorInfo _output_nhwc; - TensorInfo _input_workspace; - TensorInfo _kernel_storage; - TensorInfo _output_workspace; - TensorInfo _input_transformed; - TensorInfo _output_transformed; - TensorInfo _weights_hwio; - bool _run_activation; - bool _is_prepared; + std::unique_ptr<CpuGemm> _gemm_function; + std::unique_ptr<CpuActivation> _activation_func; + std::unique_ptr<ICPPKernel> _transform_input_kernel; + std::unique_ptr<ICPPKernel> _transform_output_kernel; + std::unique_ptr<CpuPermute> _permute_input; + std::unique_ptr<CpuPermute> _permute_output; + std::unique_ptr<CpuPermute> _permute_weights; + experimental::MemoryRequirements _aux_mem{ Count }; + std::unique_ptr<arm_conv::ConvolutionArgs> _conv_args; // Make it unique ptr because this type does not have a default constructor + arm_conv::winograd::WinogradImpl _winograd_impl; + DataLayout _data_layout; + TensorInfo _winograd_transformed_input; + TensorInfo _winograd_transformed_output; + TensorInfo _winograd_transformed_weights; + TensorInfo _input_workspace; + TensorInfo _output_workspace; + TensorInfo _weights_hwio; + TensorInfo _input_nhwc; + TensorInfo _output_nhwc; + bool _is_prepared; + bool _run_activation; }; } // namespace cpu } // namespace arm_compute |