1 /*
2 * Copyright (c) 2017-2020 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/NEDeconvolutionLayer.h"
25
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/Utils.h"
28 #include "arm_compute/core/Validate.h"
29 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
30 #include "arm_compute/runtime/NEON/NEScheduler.h"
31 #include "src/core/NEON/kernels/NEWeightsReshapeKernel.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33
34 using namespace arm_compute::misc::shape_calculator;
35
36 namespace arm_compute
37 {
38 namespace
39 {
compute_upsample_info(const PadStrideInfo & info,uint32_t deconv_pad_x,uint32_t deconv_pad_y)40 PadStrideInfo compute_upsample_info(const PadStrideInfo &info, uint32_t deconv_pad_x, uint32_t deconv_pad_y)
41 {
42 const unsigned int pad_left = info.pad_left();
43 const unsigned int pad_right = info.pad_right();
44 const unsigned int pad_top = info.pad_top();
45 const unsigned int pad_bottom = info.pad_bottom();
46 const unsigned int stride_x = info.stride().first;
47 const unsigned int stride_y = info.stride().second;
48
49 // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape
50 unsigned int deconv_pad_left = pad_right > pad_left ? pad_right - pad_left : 0;
51 unsigned int deconv_pad_right = pad_left > pad_right ? pad_left - pad_right : 0;
52 deconv_pad_x -= deconv_pad_left + deconv_pad_right;
53 ARM_COMPUTE_ERROR_ON((deconv_pad_x % 2) != 0);
54 deconv_pad_left += deconv_pad_x / 2;
55 deconv_pad_right += deconv_pad_x / 2;
56
57 unsigned int deconv_pad_top = pad_bottom > pad_top ? pad_bottom - pad_top : 0;
58 unsigned int deconv_pad_bottom = pad_top > pad_bottom ? pad_top - pad_bottom : 0;
59 deconv_pad_y -= deconv_pad_top + deconv_pad_bottom;
60 ARM_COMPUTE_ERROR_ON((deconv_pad_y % 2) != 0);
61 deconv_pad_top += deconv_pad_y / 2;
62 deconv_pad_bottom += deconv_pad_y / 2;
63
64 return PadStrideInfo(stride_x, stride_y, deconv_pad_left, deconv_pad_right, deconv_pad_top, deconv_pad_bottom, DimensionRoundingType::FLOOR);
65 }
66
67 } // namespace
68
NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)69 NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
70 : _memory_group(std::move(memory_manager)),
71 _conv_f(),
72 _upsample_f(),
73 _flip_weights(),
74 _scaled_output(),
75 _weights_flipped(),
76 _flip_axis(),
77 _original_weights(nullptr),
78 _input(nullptr),
79 _info(),
80 _is_prepared(false)
81 {
82 }
83
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * bias,const ITensorInfo * output,const PadStrideInfo & info)84 Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &info)
85 {
86 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
87 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
88 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, input);
89 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(weights, input);
90 const unsigned int width_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::WIDTH);
91 const unsigned int height_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::HEIGHT);
92 ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(width_idx) != weights->dimension(height_idx));
93 ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(width_idx) < 1);
94
95 auto out_dims = deconvolution_output_dimensions(input->dimension(width_idx), input->dimension(height_idx), weights->dimension(width_idx), weights->dimension(height_idx), info);
96
97 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
98 if(bias != nullptr)
99 {
100 if(is_data_type_quantized_asymmetric(input->data_type()))
101 {
102 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
103 }
104 else
105 {
106 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
107 }
108 }
109
110 if(output->tensor_shape().total_size() > 0)
111 {
112 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
113
114 const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input, *weights);
115
116 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
117 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
118 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
119 }
120
121 uint32_t deconv_pad_x = 0;
122 uint32_t deconv_pad_y = 0;
123 const unsigned int stride_x = info.stride().first;
124 const unsigned int stride_y = info.stride().second;
125 const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y);
126 TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(scale_out_shape));
127 const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
128
129 const unsigned int batches_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::BATCHES);
130 const unsigned int channel_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::CHANNEL);
131 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(batches_idx) != scale_out_info.dimension(batches_idx));
132 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(channel_idx) != scale_out_info.dimension(channel_idx));
133
134 ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayer::validate(&scale_out_info, weights, bias, output, conv_info, WeightsInfo()));
135
136 return Status{};
137 }
138
configure(ITensor * input,const ITensor * weights,const ITensor * bias,ITensor * output,const PadStrideInfo & info)139 void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info)
140 {
141 // Perform validation step
142 ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
143 ARM_COMPUTE_ERROR_THROW_ON(NEDeconvolutionLayer::validate(input->info(), weights->info(), (bias == nullptr) ? nullptr : bias->info(), output->info(), info));
144
145 const DataLayout data_layout = input->info()->data_layout();
146 const unsigned int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
147 const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
148 auto out_dims = deconvolution_output_dimensions(input->info()->dimension(width_idx), input->info()->dimension(height_idx),
149 weights->info()->dimension(width_idx), weights->info()->dimension(height_idx), info);
150
151 const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info());
152
153 _input = input;
154 _original_weights = weights;
155 _info = info;
156 _is_prepared = false;
157
158 const unsigned int stride_x = info.stride().first;
159 const unsigned int stride_y = info.stride().second;
160
161 // Output auto initialization if not yet initialized
162 auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->quantization_info());
163
164 _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32));
165 _memory_group.manage(&_scaled_output);
166
167 _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
168 _flip_weights.configure(weights, &_weights_flipped, &_flip_axis);
169
170 // setup the function to convolve the upscaled output
171 const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
172 uint32_t deconv_pad_x = 0;
173 uint32_t deconv_pad_y = 0;
174
175 const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(),
176 stride_x, stride_y,
177 out_dims, deconv_pad_x, deconv_pad_y);
178
179 const PadStrideInfo upsample_info = compute_upsample_info(info, deconv_pad_x, deconv_pad_y);
180
181 TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
182 scale_out_info.set_data_layout(data_layout);
183 _scaled_output.allocator()->init(scale_out_info);
184
185 _upsample_f.configure(input, &_scaled_output, upsample_info);
186
187 _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info);
188
189 // Setup flip axis data
190 _flip_axis.allocator()->allocate();
191 auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer());
192 axis_data[0] = static_cast<uint32_t>(width_idx);
193 axis_data[1] = static_cast<uint32_t>(height_idx);
194
195 _scaled_output.allocator()->allocate();
196 }
197
run()198 void NEDeconvolutionLayer::run()
199 {
200 prepare();
201
202 MemoryGroupResourceScope scope_mg(_memory_group);
203
204 _upsample_f.run();
205 _conv_f.run();
206 }
207
prepare()208 void NEDeconvolutionLayer::prepare()
209 {
210 if(!_is_prepared)
211 {
212 ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
213
214 // Run weights flipping and mark original weights tensor as unused
215 _weights_flipped.allocator()->allocate();
216 _flip_weights.run();
217 _original_weights->mark_as_unused();
218
219 // Prepare convolution
220 _conv_f.prepare();
221
222 _is_prepared = true;
223 }
224 }
225 } // namespace arm_compute
226