xref: /aosp_15_r20/external/ComputeLibrary/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2021, 2023 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/common/utils/Log.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       _do_upsampling(true)
82 {
83 }
84 
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * bias,const ITensorInfo * output,const PadStrideInfo & info,bool enable_fast_math)85 Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &info, bool enable_fast_math)
86 {
87     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
88     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
89     const unsigned int width_idx  = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::WIDTH);
90     const unsigned int height_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::HEIGHT);
91     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(width_idx) != weights->dimension(height_idx));
92     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(width_idx) < 1);
93     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(weights, input);
94     if(is_data_type_quantized_per_channel(weights->data_type()) && is_data_type_quantized(input->data_type()))
95     {
96         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QSYMM8_PER_CHANNEL);
97     }
98     else
99     {
100         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
101     }
102 
103     auto out_dims = deconvolution_output_dimensions(input->dimension(width_idx), input->dimension(height_idx), weights->dimension(width_idx), weights->dimension(height_idx), info);
104 
105     if(bias != nullptr)
106     {
107         if(is_data_type_quantized_asymmetric(input->data_type()))
108         {
109             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
110         }
111         else
112         {
113             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
114         }
115     }
116 
117     if(output->tensor_shape().total_size() > 0)
118     {
119         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
120 
121         const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input, *weights);
122 
123         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimX) != output_shape.x(), "Output's width is invalid.");
124         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimY) != output_shape.y(), "Output's height is invalid.");
125         ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(Window::DimZ) != output_shape.z(), "Output's depth is invalid.");
126     }
127 
128     uint32_t           deconv_pad_x = 0;
129     uint32_t           deconv_pad_y = 0;
130     const unsigned int stride_x     = info.stride().first;
131     const unsigned int stride_y     = info.stride().second;
132     // Guard against overflows in compute_deconvolution_upsampled_shape()
133     const DataLayout   data_layout = input->data_layout();
134     const size_t       idx_w       = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
135     const size_t       idx_h       = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
136     const unsigned int out_x       = (input->dimension(idx_w) - 1) * stride_x + 1;
137     const unsigned int out_y       = (input->dimension(idx_h) - 1) * stride_y + 1;
138     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) > out_x);
139     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_h) > out_y);
140     ARM_COMPUTE_RETURN_ERROR_ON((out_x - weights->dimension(idx_w) + 1) > out_dims.first);
141     ARM_COMPUTE_RETURN_ERROR_ON((out_y - weights->dimension(idx_h) + 1) > out_dims.second);
142 
143     const TensorShape   scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y);
144     TensorInfo          scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(scale_out_shape));
145     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
146 
147     const unsigned int batches_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::BATCHES);
148     const unsigned int channel_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::CHANNEL);
149     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(batches_idx) != scale_out_info.dimension(batches_idx));
150     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(channel_idx) != scale_out_info.dimension(channel_idx));
151 
152     ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayer::validate(&scale_out_info, weights, bias, output, conv_info, WeightsInfo(), Size2D(1U, 1U), ActivationLayerInfo(), enable_fast_math));
153 
154     return Status{};
155 }
156 
configure(ITensor * input,const ITensor * weights,const ITensor * bias,ITensor * output,const PadStrideInfo & info,bool enable_fast_math)157 void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info, bool enable_fast_math)
158 {
159     // Perform validation step
160     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
161     ARM_COMPUTE_ERROR_THROW_ON(NEDeconvolutionLayer::validate(input->info(), weights->info(), (bias == nullptr) ? nullptr : bias->info(), output->info(), info, enable_fast_math));
162     ARM_COMPUTE_LOG_PARAMS(input, weights, bias, output, info, enable_fast_math);
163 
164     const DataLayout   data_layout = input->info()->data_layout();
165     const unsigned int width_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
166     const unsigned int height_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
167     auto               out_dims    = deconvolution_output_dimensions(input->info()->dimension(width_idx), input->info()->dimension(height_idx),
168                                                                      weights->info()->dimension(width_idx), weights->info()->dimension(height_idx), info);
169 
170     const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info());
171 
172     _input            = input;
173     _original_weights = weights;
174     _info             = info;
175     _is_prepared      = false;
176 
177     const unsigned int stride_x = info.stride().first;
178     const unsigned int stride_y = info.stride().second;
179 
180     // Do not perform upsampling when input is unit stride and weight shape is 1x1
181     _do_upsampling = stride_x != 1 || stride_y != 1 || weights->info()->dimension(width_idx) != 1 || weights->info()->dimension(height_idx) != 1;
182 
183     // Output auto initialization if not yet initialized
184     auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->quantization_info());
185 
186     _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32));
187 
188     _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
189     _flip_weights.configure(weights, &_weights_flipped, &_flip_axis);
190 
191     // setup the function to convolve the upscaled output
192     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
193     uint32_t            deconv_pad_x = 0;
194     uint32_t            deconv_pad_y = 0;
195 
196     // Setup flip axis data
197     _flip_axis.allocator()->allocate();
198     auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer());
199     axis_data[0]   = static_cast<uint32_t>(width_idx);
200     axis_data[1]   = static_cast<uint32_t>(height_idx);
201 
202     // Setup convolution and upsampling, if needed
203     if (_do_upsampling)
204     {
205         _memory_group.manage(&_scaled_output);
206         const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(),
207                                                                                   stride_x, stride_y,
208                                                                                   out_dims, deconv_pad_x, deconv_pad_y);
209 
210         const PadStrideInfo upsample_info = compute_upsample_info(info, deconv_pad_x, deconv_pad_y);
211 
212         TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
213         scale_out_info.set_data_layout(data_layout);
214         _scaled_output.allocator()->init(scale_out_info);
215 
216         _upsample_f.configure(input, &_scaled_output, upsample_info);
217 
218         _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info, WeightsInfo(), Size2D(1U, 1U), ActivationLayerInfo(), enable_fast_math);
219 
220         _scaled_output.allocator()->allocate();
221     }
222     else
223     {
224         _conv_f.configure(input, &_weights_flipped, bias, output, conv_info, WeightsInfo(), Size2D(1U, 1U), ActivationLayerInfo(), enable_fast_math);
225     }
226 }
227 
run()228 void NEDeconvolutionLayer::run()
229 {
230     prepare();
231 
232     MemoryGroupResourceScope scope_mg(_memory_group);
233 
234     if(_do_upsampling)
235     {
236         _upsample_f.run();
237     }
238     _conv_f.run();
239 }
240 
prepare()241 void NEDeconvolutionLayer::prepare()
242 {
243     if(!_is_prepared)
244     {
245         ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
246 
247         // Run weights flipping and mark original weights tensor as unused
248         _weights_flipped.allocator()->allocate();
249         _flip_weights.run();
250         _original_weights->mark_as_unused();
251 
252         // Prepare convolution
253         _conv_f.prepare();
254 
255         _is_prepared = true;
256     }
257 }
258 } // namespace arm_compute
259