Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | ced7a8d | 2018-02-01 16:31:33 +0000 | [diff] [blame^] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 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/NEDeconvolutionLayer.h" |
| 25 | |
| 26 | #include "arm_compute/core/Helpers.h" |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/Utils.h" |
| 28 | #include "arm_compute/core/Validate.h" |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 29 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 30 | |
| 31 | using namespace arm_compute; |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 32 | using namespace arm_compute::misc::shape_calculator; |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 33 | |
| 34 | NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT |
| 35 | : _memory_group(std::move(memory_manager)), |
Georgios Pinitas | ced7a8d | 2018-02-01 16:31:33 +0000 | [diff] [blame^] | 36 | _direct_conv_f(), |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 37 | _conv_f(), |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 38 | _scaled_output(), |
| 39 | _input(nullptr), |
| 40 | _info(), |
Georgios Pinitas | ced7a8d | 2018-02-01 16:31:33 +0000 | [diff] [blame^] | 41 | _inner_border(), |
| 42 | _run_direct_convolution(false) |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 43 | { |
| 44 | } |
| 45 | |
| 46 | void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info, |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 47 | unsigned int inner_border_right, unsigned int inner_border_top) |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 48 | { |
| 49 | ARM_COMPUTE_ERROR_ON_NULLPTR(output); |
| 50 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); |
| 51 | ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != weights->info()->dimension(1)); |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 52 | |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 53 | _input = input; |
| 54 | _info = info; |
| 55 | _inner_border = std::make_pair(inner_border_right, inner_border_top); |
Georgios Pinitas | ced7a8d | 2018-02-01 16:31:33 +0000 | [diff] [blame^] | 56 | // FIXME: ConvolutionLayer Segfaults in GEMM assembly code for 1x1 convolutions |
| 57 | _run_direct_convolution = (weights->info()->dimension(0) == weights->info()->dimension(1)) && (weights->info()->dimension(0) == 1); |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 58 | |
| 59 | const unsigned int stride_x = info.stride().first; |
| 60 | const unsigned int stride_y = info.stride().second; |
| 61 | auto out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1), |
| 62 | info.pad().first, info.pad().second, inner_border_right, inner_border_top, stride_x, stride_y); |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 63 | |
| 64 | const TensorShape output_shape = deconvolution_output_shape(out_dims, input->info()->tensor_shape(), weights->info()->tensor_shape()); |
| 65 | |
Anthony Barbier | 1ff480b | 2018-01-11 09:23:37 +0000 | [diff] [blame] | 66 | ARM_COMPUTE_UNUSED(output_shape); |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 67 | ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid."); |
| 68 | ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid."); |
| 69 | ARM_COMPUTE_ERROR_ON_MSG(output->info()->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid."); |
| 70 | |
| 71 | _memory_group.manage(&_scaled_output); |
| 72 | |
| 73 | // configure scale function |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 74 | // Init and allocate intermmidiate tensor for output, same size as input but the first two axis are the same as the output tensor |
| 75 | const TensorInfo scale_out_info(compute_deconvolution_shape(*input->info(), stride_x, stride_y, inner_border_right, inner_border_top, info), 1, input->info()->data_type(), |
| 76 | input->info()->fixed_point_position()); |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 77 | _scaled_output.allocator()->init(scale_out_info); |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 78 | |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 79 | // setup the function to convolve the upscaled output |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 80 | const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL); |
Georgios Pinitas | ced7a8d | 2018-02-01 16:31:33 +0000 | [diff] [blame^] | 81 | (_run_direct_convolution) ? _direct_conv_f.configure(&_scaled_output, weights, bias, output, conv_info) : _conv_f.configure(&_scaled_output, weights, bias, output, conv_info); |
| 82 | |
| 83 | // Allocate auxiliary tensors |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 84 | _scaled_output.allocator()->allocate(); |
| 85 | } |
| 86 | |
| 87 | void NEDeconvolutionLayer::run() |
| 88 | { |
| 89 | _memory_group.acquire(); |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 90 | |
| 91 | // Initialize _scaled_output buffer |
| 92 | const int width_in = _input->info()->dimension(0); |
| 93 | const int height_in = _input->info()->dimension(1); |
| 94 | const int width_scaled = _scaled_output.info()->dimension(0); |
| 95 | const int height_scaled = _scaled_output.info()->dimension(1); |
| 96 | const int num_2d_slices = _input->info()->tensor_shape().total_size() / (width_in * height_in); |
| 97 | const int stride_x = _info.stride().first; |
| 98 | const int stride_y = _info.stride().second; |
| 99 | |
Georgios Pinitas | ced7a8d | 2018-02-01 16:31:33 +0000 | [diff] [blame^] | 100 | std::fill_n(_scaled_output.buffer(), _scaled_output.info()->total_size(), 0); |
Michalis Spyrou | 780db4e | 2017-11-23 09:49:51 +0000 | [diff] [blame] | 101 | |
| 102 | // scaled_output is the input for the forward convolution. We copy the input elements to scaled_output |
| 103 | // and insert rows and columns with zeroes depending on the stride values. |
| 104 | for(int slice = 0; slice < num_2d_slices; ++slice) |
| 105 | { |
| 106 | const int start_x = _info.pad().first; |
| 107 | const int start_y = _inner_border.second + _info.pad().second; |
| 108 | const int end_y = height_scaled - _info.pad().second; |
| 109 | const int end_x = width_scaled - _inner_border.first - _info.pad().first; |
| 110 | |
| 111 | for(int yi = start_y, in_y = 0; yi < end_y; yi += stride_y, in_y++) |
| 112 | { |
| 113 | for(int xi = start_x, in_x = 0; xi < end_x; xi += stride_x, in_x++) |
| 114 | { |
| 115 | const auto in = *(reinterpret_cast<float *>(_input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(in_x, in_y, slice)))); |
| 116 | *(reinterpret_cast<float *>(_scaled_output.buffer() + _scaled_output.info()->offset_element_in_bytes(Coordinates(xi, yi, slice)))) = in; |
| 117 | } |
| 118 | } |
| 119 | } |
| 120 | |
Georgios Pinitas | ced7a8d | 2018-02-01 16:31:33 +0000 | [diff] [blame^] | 121 | // Run convolution layer |
| 122 | (_run_direct_convolution) ? _direct_conv_f.run() : _conv_f.run(); |
| 123 | |
Pablo Tello | f5f34bb | 2017-08-22 13:34:13 +0100 | [diff] [blame] | 124 | _memory_group.release(); |
| 125 | } |