Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/runtime/NEON/functions/NEDeconvolutionLayerUpsample.h" |
| 25 | |
| 26 | #include "arm_compute/core/Coordinates.h" |
| 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "arm_compute/core/Helpers.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
| 30 | #include "arm_compute/core/NEON/kernels/NEDeconvolutionLayerUpsampleKernel.h" |
| 31 | #include "arm_compute/core/PixelValue.h" |
| 32 | #include "arm_compute/core/TensorInfo.h" |
| 33 | #include "arm_compute/core/Window.h" |
| 34 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 35 | #include "arm_compute/runtime/TensorAllocator.h" |
| 36 | #include "support/ToolchainSupport.h" |
| 37 | |
| 38 | #include <cmath> |
| 39 | #include <cstddef> |
| 40 | #include <utility> |
| 41 | |
| 42 | using namespace arm_compute; |
| 43 | |
| 44 | namespace |
| 45 | { |
| 46 | inline void precompute_offsets(ITensor *offsets, float wr, size_t input_element_size, const std::pair<unsigned int, unsigned int> &a, |
| 47 | const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info) |
| 48 | { |
| 49 | ARM_COMPUTE_ERROR_ON(nullptr == offsets); |
| 50 | Window win; |
| 51 | const int padx = info.pad().first; |
| 52 | const int pady = info.pad().second; |
| 53 | const int ax = a.first; |
| 54 | const int ay = a.second; |
| 55 | const int offset_width = offsets->info()->dimension(0); |
| 56 | const int offset_height = offsets->info()->dimension(1); |
| 57 | // The values of ax and ay denote the number of ZEROS to be added on the top and right inner border of the image. |
| 58 | // Step value along the XY axis will depend on the number of zeros to be inserted between samples (number of zeros + 1). |
| 59 | // Pre-compute the X offset, Y's stride is unknown at this point so we can't precompute Y's offsets |
| 60 | for(int yi = ay; yi < (offset_height - pady); yi += (1 + iz.second)) |
| 61 | { |
| 62 | for(int xi = padx; xi < (offset_width - ax); xi += (1 + iz.first)) |
| 63 | { |
| 64 | int *ptr = reinterpret_cast<int *>(offsets->ptr_to_element(Coordinates(xi, yi))); |
| 65 | const size_t in_xi = (xi + 0.5f) * wr; |
| 66 | *reinterpret_cast<int32_t *>(ptr) = in_xi * input_element_size; |
| 67 | } |
| 68 | } |
| 69 | } |
| 70 | } // namespace |
| 71 | |
| 72 | NEDeconvolutionLayerUpsample::NEDeconvolutionLayerUpsample(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT |
| 73 | : _memory_group(std::move(memory_manager)), |
| 74 | _offsets(), |
| 75 | _border_handler(), |
| 76 | _upsample() |
| 77 | { |
| 78 | } |
| 79 | |
| 80 | void NEDeconvolutionLayerUpsample::configure(ITensor *input, ITensor *output, const std::pair<unsigned int, unsigned int> &a, |
| 81 | const std::pair<unsigned int, unsigned int> &iz, const PadStrideInfo &info) |
| 82 | { |
| 83 | ARM_COMPUTE_ERROR_ON(nullptr == input); |
| 84 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 85 | |
| 86 | for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) |
| 87 | { |
| 88 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i)); |
| 89 | } |
| 90 | |
| 91 | // Get the tensor shape |
| 92 | const TensorShape shape(output->info()->dimension(0), output->info()->dimension(1)); |
| 93 | |
| 94 | // Compute the ratio between source width/height and destination width/height |
| 95 | const auto wr = static_cast<float>(input->info()->dimension(0)) / static_cast<float>(output->info()->dimension(0)); |
| 96 | const auto hr = static_cast<float>(input->info()->dimension(1)) / static_cast<float>(output->info()->dimension(1)); |
| 97 | ARM_COMPUTE_UNUSED(hr); |
| 98 | // Get the element size of the input image |
| 99 | const size_t input_element_size = input->info()->element_size(); |
| 100 | |
| 101 | TensorInfo tensor_info_offsets(shape, Format::S32); |
| 102 | _offsets.allocator()->init(tensor_info_offsets); |
| 103 | |
| 104 | _upsample.configure(input, &_offsets, output); |
| 105 | |
| 106 | // Allocate once the configure methods have been called |
| 107 | _offsets.allocator()->allocate(); |
| 108 | // Pre-compute offsets for nearest interpolation |
| 109 | std::fill_n(reinterpret_cast<int32_t *>(_offsets.buffer()), _offsets.info()->total_size() / sizeof(int32_t), -1 * input_element_size); |
| 110 | precompute_offsets(&_offsets, wr, input_element_size, a, iz, info); |
| 111 | |
| 112 | _border_handler.configure(input, _upsample.border_size(), BorderMode::CONSTANT, PixelValue(0)); |
| 113 | } |
| 114 | |
| 115 | void NEDeconvolutionLayerUpsample::run() |
| 116 | { |
| 117 | NEScheduler::get().schedule(&_border_handler, Window::DimZ); |
| 118 | _memory_group.acquire(); |
| 119 | NEScheduler::get().schedule(&_upsample, Window::DimY); |
| 120 | _memory_group.release(); |
| 121 | } |