From bd9097db81f229c2d7bbafc2bcf392b7c1c49b58 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Fri, 26 Jul 2019 15:31:02 +0100 Subject: COMPMID-2336: Rename the new generic depthwise convolution on NEON Change-Id: I45cacf75b08bb9d867343037507e56f200ad6ac0 Signed-off-by: Gian Marco Iodice Reviewed-on: https://review.mlplatform.org/c/1637 Tested-by: Arm Jenkins Reviewed-by: Giorgio Arena Comments-Addressed: Arm Jenkins --- .../kernels/NEDepthwiseConvolutionLayerKernel.cpp | 330 --------------------- .../NEDepthwiseConvolutionLayerNativeKernel.cpp | 330 +++++++++++++++++++++ .../NEON/functions/NEDepthwiseConvolutionLayer.cpp | 2 +- 3 files changed, 331 insertions(+), 331 deletions(-) delete mode 100644 src/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.cpp create mode 100644 src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp (limited to 'src') diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.cpp deleted file mode 100644 index feb2071d47..0000000000 --- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.cpp +++ /dev/null @@ -1,330 +0,0 @@ -/* - * Copyright (c) 2019 ARM Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerKernel.h" - -#include "arm_compute/core/AccessWindowStatic.h" -#include "arm_compute/core/NEON/wrapper/traits.h" -#include "arm_compute/core/NEON/wrapper/wrapper.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" - -namespace arm_compute -{ -namespace -{ -template -void depthwise_loop_multiplier1(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, - const Size2D &dilation, const Window &window) -{ - using VectorType = typename wrapper::traits::neon_vector::type; - using TagType = typename wrapper::traits::neon_vector::tag_type; - - const size_t input_stride_y = input->info()->strides_in_bytes().y(); - const size_t input_stride_z = input->info()->strides_in_bytes().z(); - const size_t input_max_offset = input->info()->strides_in_bytes().z() * input->info()->dimension(2) - (input->info()->padding().bottom + input->info()->padding().top) * - input->info()->strides_in_bytes().y(); - const size_t weights_width = weights->info()->dimension(1); - const size_t weights_height = weights->info()->dimension(2); - const size_t weights_stride_y = weights->info()->strides_in_bytes().y(); - const size_t weights_stride_z = weights->info()->strides_in_bytes().z(); - const size_t conv_stride_x = conv_info.stride().first; - const size_t conv_stride_y = conv_info.stride().second; - const size_t conv_pad_left = conv_info.pad_left(); - const size_t conv_pad_top = conv_info.pad_top(); - - Window win_input = window; - win_input.set(Window::DimY, Window::Dimension(0, 0, 0)); - win_input.set(Window::DimZ, Window::Dimension(0, 0, 0)); - - Window win_weights = win_input; - win_weights.set(3, Window::Dimension(0, 0, 0)); - - Iterator input_it(input, win_input); - Iterator weights_it(weights, win_weights); - Iterator output_it(output, window); - Iterator biases_it{}; - - if(has_biases) - { - biases_it = Iterator(biases, win_weights); - } - - execute_window_loop(window, [&](const Coordinates & id) - { - VectorType acc = wrapper::vdup_n(static_cast(0), TagType{}); - - const int input_y = id.y() * conv_stride_x - conv_pad_left; - const int input_z = id.z() * conv_stride_y - conv_pad_top; - int input_offset = input_y * input_stride_y + input_z * input_stride_z; - - auto weights_ptr = weights_it.ptr(); - for(size_t h = 0; h < weights_height; ++h) - { - int offs = input_offset; - for(size_t w = 0; w < weights_width; ++w) - { - const auto input_vals = wrapper::vload(reinterpret_cast(input_it.ptr() + std::min(static_cast(offs), input_max_offset))); - const auto weights_vals = wrapper::vload(reinterpret_cast(weights_ptr + w * weights_stride_y)); - - acc = wrapper::vmla(acc, weights_vals, input_vals); - offs += dilation.x() * input_stride_y; - } - - weights_ptr += weights_stride_z; - input_offset += dilation.y() * input_stride_z; - } - - if(has_biases) - { - const auto biases_vals = wrapper::vload(reinterpret_cast(biases_it.ptr())); - acc = wrapper::vadd(acc, biases_vals); - } - - wrapper::vstore(reinterpret_cast(output_it.ptr()), acc); - }, - input_it, weights_it, biases_it, output_it); -} - -template -void depthwise_loop_generic(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, - const Size2D &dilation, unsigned int depth_multiplier, const Window &window) -{ - const size_t input_stride_y = input->info()->strides_in_bytes().y(); - const size_t input_stride_z = input->info()->strides_in_bytes().z(); - const size_t input_max_offset = input->info()->strides_in_bytes().z() * input->info()->dimension(2) - (input->info()->padding().bottom + input->info()->padding().top) * - input->info()->strides_in_bytes().y(); - const size_t weights_width = weights->info()->dimension(1); - const size_t weights_height = weights->info()->dimension(2); - const size_t weights_stride_y = weights->info()->strides_in_bytes().y(); - const size_t weights_stride_z = weights->info()->strides_in_bytes().z(); - const size_t conv_stride_x = conv_info.stride().first; - const size_t conv_stride_y = conv_info.stride().second; - const size_t conv_pad_left = conv_info.pad_left(); - const size_t conv_pad_top = conv_info.pad_top(); - - Window win_input = window; - win_input.set(Window::DimY, Window::Dimension(0, 0, 0)); - win_input.set(Window::DimZ, Window::Dimension(0, 0, 0)); - - Window win_weights = win_input; - win_weights.set(3, Window::Dimension(0, 0, 0)); - - win_input.set_dimension_step(Window::DimX, 1); - - Iterator input_it(input, win_input); - Iterator weights_it(weights, win_weights); - Iterator output_it(output, window); - Iterator biases_it{}; - - if(has_biases) - { - biases_it = Iterator(biases, win_weights); - } - - execute_window_loop(window, [&](const Coordinates & id) - { - std::vector acc(depth_multiplier, static_cast(0)); - - const int input_y = id.y() * conv_stride_x - conv_pad_left; - const int input_z = id.z() * conv_stride_y - conv_pad_top; - int input_offset = input_y * input_stride_y + input_z * input_stride_z; - - auto weights_ptr = weights_it.ptr(); - for(size_t h = 0; h < weights_height; ++h) - { - int offs = input_offset; - for(size_t w = 0; w < weights_width; ++w) - { - const auto input_val = *(reinterpret_cast(input_it.ptr() + std::min(static_cast(offs), input_max_offset))); - - for(size_t m = 0; m < depth_multiplier; ++m) - { - const auto weights_val = *(reinterpret_cast(weights_ptr + m * sizeof(T) + w * weights_stride_y)); - acc.at(m) = std::fma(weights_val, input_val, acc.at(m)); - } - - offs += dilation.x() * input_stride_y; - } - - weights_ptr += weights_stride_z; - input_offset += dilation.y() * input_stride_z; - } - - if(has_biases) - { - for(size_t m = 0; m < depth_multiplier; ++m) - { - const auto biases_val = *(reinterpret_cast(biases_it.ptr() + m * sizeof(T))); - *(reinterpret_cast(output_it.ptr() + m * sizeof(T))) = acc.at(m) + biases_val; - } - } - else - { - for(size_t m = 0; m < depth_multiplier; ++m) - { - *(reinterpret_cast(output_it.ptr() + m * sizeof(T))) = acc.at(m); - } - } - }, - input_it, weights_it, biases_it, output_it); -} - -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, - const Size2D &dilation) -{ - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); - ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier == 0); - ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(0) * depth_multiplier) != weights->dimension(0)); - ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1)); - ARM_COMPUTE_RETURN_ERROR_ON((conv_info.stride().first < 1) || (conv_info.stride().second < 1)); - - if(biases != nullptr) - { - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0)); - } - - if(output->total_size() != 0) - { - const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); - } - - return Status{}; -} -} // namespace - -std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *biases, - ITensorInfo *output, const PadStrideInfo &conv_info, - unsigned int depth_multiplier, const Size2D &dilation) -{ - // Get convolved dimensions - const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); - - // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output, input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape)); - - // Configure kernel window (generic) - const unsigned int num_elems_read_per_iteration = (depth_multiplier == 1) ? 8 / element_size_from_data_type(input->data_type()) : 1; - const unsigned int num_elems_written_per_iteration = num_elems_read_per_iteration * depth_multiplier; - - // Configure kernel window - Window win = calculate_max_window(*output, Steps(num_elems_written_per_iteration)); - - AccessWindowStatic input_access(input, 0, -conv_info.pad_left(), ceil_to_multiple(num_elems_read_per_iteration, input->dimension(0)), - input->dimension(1) + std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top())); - AccessWindowHorizontal weights_access(weights, 0, num_elems_written_per_iteration); - AccessWindowHorizontal output_access(output, 0, num_elems_written_per_iteration); - - bool window_changed = update_window_and_padding(win, input_access, weights_access, output_access); - - if(biases != nullptr) - { - AccessWindowHorizontal biases_access(biases, 0, num_elems_written_per_iteration); - window_changed |= update_window_and_padding(win, biases_access); - } - - output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, win); -} - -NEDepthwiseConvolutionLayerKernel::NEDepthwiseConvolutionLayerKernel() - : _func(), _border_size(0), _input(), _weights(), _biases(), _output(), _conv_info(), _depth_multiplier(1), _dilation() -{ -} - -BorderSize NEDepthwiseConvolutionLayerKernel::border_size() const -{ - return _border_size; -} - -void NEDepthwiseConvolutionLayerKernel::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, - const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, dilation)); - - _input = input; - _weights = weights; - _biases = biases; - _output = output; - _conv_info = conv_info; - _depth_multiplier = depth_multiplier; - _border_size = BorderSize(_conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0); - _dilation = dilation; - - switch(_input->info()->data_type()) - { - case DataType::F32: - _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerKernel::run_depthwise : &NEDepthwiseConvolutionLayerKernel::run_depthwise; - break; - default: - ARM_COMPUTE_ERROR("Data type not supported"); - break; - } - - auto win_config = validate_and_configure_window(_input->info(), _weights->info(), (biases != nullptr) ? biases->info() : nullptr, _output->info(), _conv_info, _depth_multiplier, dilation); - ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - INEKernel::configure(win_config.second); -} - -Status NEDepthwiseConvolutionLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, - unsigned int depth_multiplier, - const Size2D &dilation) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, dilation)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), (biases != nullptr) ? biases->clone().get() : nullptr, output->clone().get(), conv_info, - depth_multiplier, dilation) - .first); - return Status{}; -} - -void NEDepthwiseConvolutionLayerKernel::run(const Window &window, const ThreadInfo &info) -{ - ARM_COMPUTE_UNUSED(info); - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - - (this->*_func)(window); -} - -template -void NEDepthwiseConvolutionLayerKernel::run_depthwise(const Window &window) -{ - ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); - ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); - - if(_depth_multiplier == 1) - { - depthwise_loop_multiplier1(_input, _weights, _biases, _output, _conv_info, _dilation, window); - } - else - { - depthwise_loop_generic(_input, _weights, _biases, _output, _conv_info, _dilation, _depth_multiplier, window); - } -} -} // namespace arm_compute diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp new file mode 100644 index 0000000000..aafdb2e8a4 --- /dev/null +++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.cpp @@ -0,0 +1,330 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h" + +#include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/NEON/wrapper/traits.h" +#include "arm_compute/core/NEON/wrapper/wrapper.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" + +namespace arm_compute +{ +namespace +{ +template +void depthwise_loop_multiplier1(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, + const Size2D &dilation, const Window &window) +{ + using VectorType = typename wrapper::traits::neon_vector::type; + using TagType = typename wrapper::traits::neon_vector::tag_type; + + const size_t input_stride_y = input->info()->strides_in_bytes().y(); + const size_t input_stride_z = input->info()->strides_in_bytes().z(); + const size_t input_max_offset = input->info()->strides_in_bytes().z() * input->info()->dimension(2) - (input->info()->padding().bottom + input->info()->padding().top) * + input->info()->strides_in_bytes().y(); + const size_t weights_width = weights->info()->dimension(1); + const size_t weights_height = weights->info()->dimension(2); + const size_t weights_stride_y = weights->info()->strides_in_bytes().y(); + const size_t weights_stride_z = weights->info()->strides_in_bytes().z(); + const size_t conv_stride_x = conv_info.stride().first; + const size_t conv_stride_y = conv_info.stride().second; + const size_t conv_pad_left = conv_info.pad_left(); + const size_t conv_pad_top = conv_info.pad_top(); + + Window win_input = window; + win_input.set(Window::DimY, Window::Dimension(0, 0, 0)); + win_input.set(Window::DimZ, Window::Dimension(0, 0, 0)); + + Window win_weights = win_input; + win_weights.set(3, Window::Dimension(0, 0, 0)); + + Iterator input_it(input, win_input); + Iterator weights_it(weights, win_weights); + Iterator output_it(output, window); + Iterator biases_it{}; + + if(has_biases) + { + biases_it = Iterator(biases, win_weights); + } + + execute_window_loop(window, [&](const Coordinates & id) + { + VectorType acc = wrapper::vdup_n(static_cast(0), TagType{}); + + const int input_y = id.y() * conv_stride_x - conv_pad_left; + const int input_z = id.z() * conv_stride_y - conv_pad_top; + int input_offset = input_y * input_stride_y + input_z * input_stride_z; + + auto weights_ptr = weights_it.ptr(); + for(size_t h = 0; h < weights_height; ++h) + { + int offs = input_offset; + for(size_t w = 0; w < weights_width; ++w) + { + const auto input_vals = wrapper::vload(reinterpret_cast(input_it.ptr() + std::min(static_cast(offs), input_max_offset))); + const auto weights_vals = wrapper::vload(reinterpret_cast(weights_ptr + w * weights_stride_y)); + + acc = wrapper::vmla(acc, weights_vals, input_vals); + offs += dilation.x() * input_stride_y; + } + + weights_ptr += weights_stride_z; + input_offset += dilation.y() * input_stride_z; + } + + if(has_biases) + { + const auto biases_vals = wrapper::vload(reinterpret_cast(biases_it.ptr())); + acc = wrapper::vadd(acc, biases_vals); + } + + wrapper::vstore(reinterpret_cast(output_it.ptr()), acc); + }, + input_it, weights_it, biases_it, output_it); +} + +template +void depthwise_loop_generic(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, + const Size2D &dilation, unsigned int depth_multiplier, const Window &window) +{ + const size_t input_stride_y = input->info()->strides_in_bytes().y(); + const size_t input_stride_z = input->info()->strides_in_bytes().z(); + const size_t input_max_offset = input->info()->strides_in_bytes().z() * input->info()->dimension(2) - (input->info()->padding().bottom + input->info()->padding().top) * + input->info()->strides_in_bytes().y(); + const size_t weights_width = weights->info()->dimension(1); + const size_t weights_height = weights->info()->dimension(2); + const size_t weights_stride_y = weights->info()->strides_in_bytes().y(); + const size_t weights_stride_z = weights->info()->strides_in_bytes().z(); + const size_t conv_stride_x = conv_info.stride().first; + const size_t conv_stride_y = conv_info.stride().second; + const size_t conv_pad_left = conv_info.pad_left(); + const size_t conv_pad_top = conv_info.pad_top(); + + Window win_input = window; + win_input.set(Window::DimY, Window::Dimension(0, 0, 0)); + win_input.set(Window::DimZ, Window::Dimension(0, 0, 0)); + + Window win_weights = win_input; + win_weights.set(3, Window::Dimension(0, 0, 0)); + + win_input.set_dimension_step(Window::DimX, 1); + + Iterator input_it(input, win_input); + Iterator weights_it(weights, win_weights); + Iterator output_it(output, window); + Iterator biases_it{}; + + if(has_biases) + { + biases_it = Iterator(biases, win_weights); + } + + execute_window_loop(window, [&](const Coordinates & id) + { + std::vector acc(depth_multiplier, static_cast(0)); + + const int input_y = id.y() * conv_stride_x - conv_pad_left; + const int input_z = id.z() * conv_stride_y - conv_pad_top; + int input_offset = input_y * input_stride_y + input_z * input_stride_z; + + auto weights_ptr = weights_it.ptr(); + for(size_t h = 0; h < weights_height; ++h) + { + int offs = input_offset; + for(size_t w = 0; w < weights_width; ++w) + { + const auto input_val = *(reinterpret_cast(input_it.ptr() + std::min(static_cast(offs), input_max_offset))); + + for(size_t m = 0; m < depth_multiplier; ++m) + { + const auto weights_val = *(reinterpret_cast(weights_ptr + m * sizeof(T) + w * weights_stride_y)); + acc.at(m) = std::fma(weights_val, input_val, acc.at(m)); + } + + offs += dilation.x() * input_stride_y; + } + + weights_ptr += weights_stride_z; + input_offset += dilation.y() * input_stride_z; + } + + if(has_biases) + { + for(size_t m = 0; m < depth_multiplier; ++m) + { + const auto biases_val = *(reinterpret_cast(biases_it.ptr() + m * sizeof(T))); + *(reinterpret_cast(output_it.ptr() + m * sizeof(T))) = acc.at(m) + biases_val; + } + } + else + { + for(size_t m = 0; m < depth_multiplier; ++m) + { + *(reinterpret_cast(output_it.ptr() + m * sizeof(T))) = acc.at(m); + } + } + }, + input_it, weights_it, biases_it, output_it); +} + +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, + const Size2D &dilation) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); + ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier == 0); + ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(0) * depth_multiplier) != weights->dimension(0)); + ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1)); + ARM_COMPUTE_RETURN_ERROR_ON((conv_info.stride().first < 1) || (conv_info.stride().second < 1)); + + if(biases != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0)); + } + + if(output->total_size() != 0) + { + const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); + } + + return Status{}; +} +} // namespace + +std::pair validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *biases, + ITensorInfo *output, const PadStrideInfo &conv_info, + unsigned int depth_multiplier, const Size2D &dilation) +{ + // Get convolved dimensions + const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); + + // Output auto inizialitation if not yet initialized + auto_init_if_empty(*output, input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape)); + + // Configure kernel window (generic) + const unsigned int num_elems_read_per_iteration = (depth_multiplier == 1) ? 8 / element_size_from_data_type(input->data_type()) : 1; + const unsigned int num_elems_written_per_iteration = num_elems_read_per_iteration * depth_multiplier; + + // Configure kernel window + Window win = calculate_max_window(*output, Steps(num_elems_written_per_iteration)); + + AccessWindowStatic input_access(input, 0, -conv_info.pad_left(), ceil_to_multiple(num_elems_read_per_iteration, input->dimension(0)), + input->dimension(1) + std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top())); + AccessWindowHorizontal weights_access(weights, 0, num_elems_written_per_iteration); + AccessWindowHorizontal output_access(output, 0, num_elems_written_per_iteration); + + bool window_changed = update_window_and_padding(win, input_access, weights_access, output_access); + + if(biases != nullptr) + { + AccessWindowHorizontal biases_access(biases, 0, num_elems_written_per_iteration); + window_changed |= update_window_and_padding(win, biases_access); + } + + output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} + +NEDepthwiseConvolutionLayerNativeKernel::NEDepthwiseConvolutionLayerNativeKernel() + : _func(), _border_size(0), _input(), _weights(), _biases(), _output(), _conv_info(), _depth_multiplier(1), _dilation() +{ +} + +BorderSize NEDepthwiseConvolutionLayerNativeKernel::border_size() const +{ + return _border_size; +} + +void NEDepthwiseConvolutionLayerNativeKernel::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, + const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, dilation)); + + _input = input; + _weights = weights; + _biases = biases; + _output = output; + _conv_info = conv_info; + _depth_multiplier = depth_multiplier; + _border_size = BorderSize(_conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0); + _dilation = dilation; + + switch(_input->info()->data_type()) + { + case DataType::F32: + _func = (biases != nullptr) ? &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise : &NEDepthwiseConvolutionLayerNativeKernel::run_depthwise; + break; + default: + ARM_COMPUTE_ERROR("Data type not supported"); + break; + } + + auto win_config = validate_and_configure_window(_input->info(), _weights->info(), (biases != nullptr) ? biases->info() : nullptr, _output->info(), _conv_info, _depth_multiplier, dilation); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + INEKernel::configure(win_config.second); +} + +Status NEDepthwiseConvolutionLayerNativeKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + unsigned int depth_multiplier, + const Size2D &dilation) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, dilation)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), (biases != nullptr) ? biases->clone().get() : nullptr, output->clone().get(), conv_info, + depth_multiplier, dilation) + .first); + return Status{}; +} + +void NEDepthwiseConvolutionLayerNativeKernel::run(const Window &window, const ThreadInfo &info) +{ + ARM_COMPUTE_UNUSED(info); + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + + (this->*_func)(window); +} + +template +void NEDepthwiseConvolutionLayerNativeKernel::run_depthwise(const Window &window) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); + + if(_depth_multiplier == 1) + { + depthwise_loop_multiplier1(_input, _weights, _biases, _output, _conv_info, _dilation, window); + } + else + { + depthwise_loop_generic(_input, _weights, _biases, _output, _conv_info, _dilation, _depth_multiplier, window); + } +} +} // namespace arm_compute diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp index c2ed901169..cdd278b2f1 100644 --- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp @@ -947,7 +947,7 @@ Status NEDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITe } else { - ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayerKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, dilation)); + ARM_COMPUTE_RETURN_ON_ERROR(NEDepthwiseConvolutionLayerNativeKernel::validate(input, weights, biases, output, conv_info, depth_multiplier, dilation)); } // Validate Activation Layer -- cgit v1.2.1