From f5f34bb068565bf9435ba5561aae1c9280db8bbe Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Tue, 22 Aug 2017 13:34:13 +0100 Subject: COMPMID-470: Neon Deconvolution. Implemented by up-sampling the input with zeros insertions between the input samples and convolving the Deconvolution kernels on the up-sampled result. The upsampling is performed by the function NEDeconvolutionLayerUpsample. Convolving is done by NEDirectConvolutionLayer. Change-Id: I25f7ba7c6b99cd9310797972ede40aeff4a54900 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/85319 Tested-by: Kaizen Reviewed-by: Anthony Barbier --- .../kernels/NEDeconvolutionLayerUpsampleKernel.cpp | 165 +++++++++++++++++++++ src/core/NEON/kernels/NEScaleKernel.cpp | 6 +- src/core/Utils.cpp | 37 +++++ 3 files changed, 206 insertions(+), 2 deletions(-) create mode 100644 src/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.cpp (limited to 'src/core') diff --git a/src/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.cpp b/src/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.cpp new file mode 100644 index 0000000000..71db2e9782 --- /dev/null +++ b/src/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.cpp @@ -0,0 +1,165 @@ +/* + * Copyright (c) 2016, 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h" + +#include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/Coordinates.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 +#include +#include + +using namespace arm_compute; + +NEDeconvolutionLayerUpsampleKernel::NEDeconvolutionLayerUpsampleKernel() + : _offsets(nullptr), _input(nullptr), _output(nullptr) +{ +} + +BorderSize NEDeconvolutionLayerUpsampleKernel::border_size() const +{ + return BorderSize(1); +} + +void NEDeconvolutionLayerUpsampleKernel::configure(const ITensor *input, const ITensor *offsets, ITensor *output) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) == 0); + ARM_COMPUTE_ERROR_ON(output->info()->dimension(1) == 0); + + for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) + { + ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i)); + } + + _input = input; + _output = output; + _offsets = offsets; + + constexpr unsigned int num_elems_processed_per_iteration = 16; + const int border_offset = border_size().left; + + // Configure kernel window + Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); + + AccessWindowRectangle input_access(input->info(), -border_offset, -border_offset, input->info()->dimension(0) + border_offset, input->info()->dimension(1) + border_offset); + AccessWindowHorizontal offsets_access(offsets->info(), 0, num_elems_processed_per_iteration); + AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); + + update_window_and_padding(win, input_access, offsets_access, output_access); + + output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); + + INEKernel::configure(win); +} + +void NEDeconvolutionLayerUpsampleKernel::scale_nearest(const Window &window) +{ + const size_t input_stride = _input->info()->strides_in_bytes()[1]; + + // Compute the ratio between source height and destination height + const auto hr = static_cast(_input->info()->dimension(1)) / static_cast(_output->info()->dimension(1)); + + // Don't increment in X and Y direction for the input tensor + // A pointer to the start of this plane is needed as base for the precomputed offsets + Window win_in(window); + win_in.set(Window::DimX, Window::Dimension(0, 0, 0)); + win_in.set(Window::DimY, Window::Dimension(0, 0, 0)); + + Window win_off; + win_off.set(Window::DimX, window[Window::DimX]); + win_off.set(Window::DimY, window[Window::DimY]); + + for(size_t d = Window::DimZ; d < _offsets->info()->num_dimensions(); ++d) + { + win_off.set(d, Window::Dimension(0, 0, 0)); + } + + Iterator in(_input, win_in); + Iterator out(_output, window); + Iterator offsets(_offsets, win_off); + + switch(_input->info()->data_type()) + { + case DataType::F32: + { + float32x4x4_t tmp = + { + { + vdupq_n_f32(0), + vdupq_n_f32(0) + } + }; + execute_window_loop(window, [&](const Coordinates & id) + { + const auto offsets_ptr = reinterpret_cast(offsets.ptr()); + + const size_t in_yi = (id.y() + 0.5f) * hr; + const size_t offset_row = in_yi * input_stride; + + tmp.val[0] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[0] + offset_row), tmp.val[0], 0); + tmp.val[0] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[4] + offset_row), tmp.val[0], 1); + tmp.val[0] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[8] + offset_row), tmp.val[0], 2); + tmp.val[0] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[12] + offset_row), tmp.val[0], 3); + + tmp.val[1] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[1] + offset_row), tmp.val[1], 0); + tmp.val[1] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[5] + offset_row), tmp.val[1], 1); + tmp.val[1] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[9] + offset_row), tmp.val[1], 2); + tmp.val[1] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[13] + offset_row), tmp.val[1], 3); + + tmp.val[2] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[2] + offset_row), tmp.val[2], 0); + tmp.val[2] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[6] + offset_row), tmp.val[2], 1); + tmp.val[2] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[10] + offset_row), tmp.val[2], 2); + tmp.val[2] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[14] + offset_row), tmp.val[2], 3); + + tmp.val[3] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[3] + offset_row), tmp.val[3], 0); + tmp.val[3] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[7] + offset_row), tmp.val[3], 1); + tmp.val[3] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[11] + offset_row), tmp.val[3], 2); + tmp.val[3] = vsetq_lane_f32(*reinterpret_cast(in.ptr() + offsets_ptr[15] + offset_row), tmp.val[3], 3); + + vst4q_f32(reinterpret_cast(out.ptr()), tmp); + }, + in, offsets, out); + break; + } + default: + ARM_COMPUTE_ERROR("Not supported"); + break; + } +} + +void NEDeconvolutionLayerUpsampleKernel::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); + scale_nearest(window); +} diff --git a/src/core/NEON/kernels/NEScaleKernel.cpp b/src/core/NEON/kernels/NEScaleKernel.cpp index 6634d4b13c..b1ced7e38d 100644 --- a/src/core/NEON/kernels/NEScaleKernel.cpp +++ b/src/core/NEON/kernels/NEScaleKernel.cpp @@ -180,8 +180,10 @@ void NEScaleKernel::scale_nearest(const Window &window) const auto offsets_ptr = reinterpret_cast(offsets.ptr()); const uint8_t *const in_ptr = in.ptr(); - const int in_yi = std::floor((id.y() + 0.5f) * hr); - const int offset_row = in_yi * input_stride; + const int in_yi = std::floor((id.y() + 0.5f) * hr); + const int in_yi_clamped = std::min(static_cast(_input->info()->dimension(1)), std::max(in_yi, -1)); + ARM_COMPUTE_ERROR_ON(in_yi_clamped < -1 || in_yi_clamped > static_cast(_input->info()->dimension(1))); + const int offset_row = in_yi_clamped * input_stride; tmp = vsetq_lane_u8(in_ptr[offsets_ptr[0] + offset_row], tmp, 0); tmp = vsetq_lane_u8(in_ptr[offsets_ptr[1] + offset_row], tmp, 1); diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp index 99d39569c7..d5ce1ea027 100644 --- a/src/core/Utils.cpp +++ b/src/core/Utils.cpp @@ -247,6 +247,43 @@ std::string arm_compute::lower_string(const std::string &val) return res; } +TensorShape arm_compute::deconvolution_output_shape(const std::pair &out_dims, TensorShape input, TensorShape weights) +{ + TensorShape out_shape(input); + out_shape.set(0, out_dims.first); + out_shape.set(1, out_dims.second); + out_shape.set(2, weights[3]); + return out_shape; +} + +const std::pair arm_compute::deconvolution_output_dimensions( + unsigned int in_width, unsigned int in_height, unsigned int kernel_width, unsigned int kernel_height, unsigned int padx, unsigned int pady, + unsigned int ax, unsigned int ay, float upscalex, float upscaley, DimensionRoundingType round) +{ + ARM_COMPUTE_ERROR_ON(in_width < 1 || in_height < 1); + ARM_COMPUTE_ERROR_ON(((in_width - 1) * upscalex + kernel_width + ax) < 2.f * padx); + ARM_COMPUTE_ERROR_ON(((in_height - 1) * upscaley + kernel_height + ay) < 2.f * pady); + const float fw = (in_width - 1) * upscalex - 2.f * padx + kernel_width + ax; + const float fh = (in_height - 1) * upscaley - 2.f * pady + kernel_height + ay; + int w = 0; + int h = 0; + switch(round) + { + case DimensionRoundingType::FLOOR: + w = std::floor(fw); + h = std::floor(fh); + break; + case DimensionRoundingType::CEIL: + w = std::ceil(fw); + h = std::ceil(fh); + break; + default: + ARM_COMPUTE_ERROR("Not supported"); + break; + } + return std::make_pair(w, h); +} + const std::pair arm_compute::scaled_dimensions(unsigned int width, unsigned int height, unsigned int kernel_width, unsigned int kernel_height, const PadStrideInfo &pad_stride_info) -- cgit v1.2.1