1 /*
2 * Copyright (c) 2019-2021 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/CL/functions/CLDirectDeconvolutionLayer.h"
25
26 #include "arm_compute/core/CL/CLKernelLibrary.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/Utils.h"
29 #include "arm_compute/core/Validate.h"
30 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
31 #include "arm_compute/runtime/CL/CLScheduler.h"
32 #include "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h"
33 #include "src/core/CL/kernels/CLFillBorderKernel.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35
36 #include "src/common/utils/Log.h"
37
38 #include <memory>
39 #include <tuple>
40
41 namespace arm_compute
42 {
43 using namespace arm_compute::misc::shape_calculator;
44
CLDirectDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)45 CLDirectDeconvolutionLayer::CLDirectDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
46 : _memory_group(std::move(memory_manager)),
47 _scale_f(),
48 _conv_f(),
49 _flip_weights(),
50 _scaled_output(),
51 _original_weights(nullptr),
52 _weights_flipped(),
53 _flip_axis(),
54 _is_prepared(false)
55 {
56 }
57
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * bias,ITensorInfo * output,const PadStrideInfo & info,const WeightsInfo & weights_info)58 Status CLDirectDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, ITensorInfo *output, const PadStrideInfo &info,
59 const WeightsInfo &weights_info)
60 {
61 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
62 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8_SIGNED, DataType::QASYMM8, DataType::F16, DataType::F32);
63 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights);
64 const DataLayout data_layout = input->data_layout();
65
66 const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
67 const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
68 const size_t idx_c = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
69
70 ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) != weights->dimension(idx_h));
71 ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_w) < 1);
72
73 auto out_dims = deconvolution_output_dimensions(input->dimension(idx_w), input->dimension(idx_h), weights->dimension(idx_w), weights->dimension(idx_h), info);
74
75 const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input, *weights);
76
77 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
78
79 if(input->data_type() != weights->data_type())
80 {
81 ARM_COMPUTE_RETURN_ERROR_ON(weights->data_type() != DataType::QSYMM8_PER_CHANNEL || !is_data_type_quantized_asymmetric(input->data_type()));
82 }
83
84 if(bias != nullptr)
85 {
86 if(is_data_type_quantized_asymmetric(input->data_type()))
87 {
88 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
89 }
90 else
91 {
92 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
93 }
94 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, bias);
95 }
96
97 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_w) != output_shape[idx_w], "Output's width is invalid.");
98 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_h) != output_shape[idx_h], "Output's height is invalid.");
99 ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->dimension(idx_c) != output_shape[idx_c], "Output's depth is invalid.");
100
101 unsigned int deconv_pad_x = 0;
102 unsigned int deconv_pad_y = 0;
103 const unsigned int stride_x = info.stride().first;
104 const unsigned int stride_y = info.stride().second;
105 const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input, *weights, stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y);
106 TensorInfo scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(scale_out_shape).set_data_layout(data_layout));
107 const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
108
109 ARM_COMPUTE_RETURN_ON_ERROR(CLDeconvolutionLayerUpsample::validate(input, &scale_out_info, info));
110 ARM_COMPUTE_RETURN_ON_ERROR(CLConvolutionLayer::validate(&scale_out_info, weights, bias, output, conv_info, weights_info));
111
112 return Status{};
113 }
114
configure(ICLTensor * input,ICLTensor * weights,const ICLTensor * bias,ICLTensor * output,const PadStrideInfo & info,const WeightsInfo & weights_info)115 void CLDirectDeconvolutionLayer::configure(ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
116 const WeightsInfo &weights_info)
117 {
118 configure(CLKernelLibrary::get().get_compile_context(), input, weights, bias, output, info, weights_info);
119 }
120
configure(const CLCompileContext & compile_context,ICLTensor * input,ICLTensor * weights,const ICLTensor * bias,ICLTensor * output,const PadStrideInfo & info,const WeightsInfo & weights_info)121 void CLDirectDeconvolutionLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
122 const WeightsInfo &weights_info)
123 {
124 ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
125 ARM_COMPUTE_LOG_PARAMS(input, weights, bias, output, info, weights_info);
126
127 const unsigned int pad_left = info.pad_left();
128 const unsigned int pad_right = info.pad_right();
129 const unsigned int pad_top = info.pad_top();
130 const unsigned int pad_bottom = info.pad_bottom();
131 const unsigned int stride_x = info.stride().first;
132 const unsigned int stride_y = info.stride().second;
133
134 const DataLayout data_layout = input->info()->data_layout();
135
136 const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
137 const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
138
139 _original_weights = weights;
140 _flip_axis.allocator()->init(TensorInfo(TensorShape(2U), 1, DataType::U32));
141 _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
142 _flip_weights.configure(compile_context, weights, &_weights_flipped, &_flip_axis);
143
144 auto out_dims = deconvolution_output_dimensions(input->info()->dimension(idx_w), input->info()->dimension(idx_h), weights->info()->dimension(idx_w), weights->info()->dimension(idx_h), info);
145
146 const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info());
147
148 // Output auto initialization if not yet initialized
149 auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_layout(data_layout));
150
151 // Perform validation step
152 ARM_COMPUTE_ERROR_THROW_ON(CLDirectDeconvolutionLayer::validate(input->info(), weights->info(), bias == nullptr ? nullptr : bias->info(), output->info(), info));
153
154 _is_prepared = weights_info.retain_internal_weights();
155
156 _memory_group.manage(&_scaled_output);
157
158 // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape
159 unsigned int deconv_pad_x = 0;
160 unsigned int deconv_pad_y = 0;
161 const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y, out_dims, deconv_pad_x, deconv_pad_y);
162
163 unsigned int deconv_pad_left = pad_right > pad_left ? pad_right - pad_left : 0;
164 unsigned int deconv_pad_right = pad_left > pad_right ? pad_left - pad_right : 0;
165 deconv_pad_x -= deconv_pad_left + deconv_pad_right;
166 ARM_COMPUTE_ERROR_ON((deconv_pad_x % 2) != 0);
167 deconv_pad_left += deconv_pad_x / 2;
168 deconv_pad_right += deconv_pad_x / 2;
169
170 unsigned int deconv_pad_top = pad_bottom > pad_top ? pad_bottom - pad_top : 0;
171 unsigned int deconv_pad_bottom = pad_top > pad_bottom ? pad_top - pad_bottom : 0;
172 deconv_pad_y -= deconv_pad_top + deconv_pad_bottom;
173 ARM_COMPUTE_ERROR_ON((deconv_pad_y % 2) != 0);
174 deconv_pad_top += deconv_pad_y / 2;
175 deconv_pad_bottom += deconv_pad_y / 2;
176
177 TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
178 scale_out_info.set_data_layout(data_layout);
179 _scaled_output.allocator()->init(scale_out_info);
180
181 // configure scale function
182 const PadStrideInfo upsample_info(stride_x, stride_y, deconv_pad_left, deconv_pad_right, deconv_pad_top, deconv_pad_bottom, DimensionRoundingType::FLOOR);
183 _scale_f.configure(compile_context, input, &_scaled_output, upsample_info);
184
185 // Setup the function to convolve the upscaled output
186 const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
187 _conv_f.configure(compile_context, &_scaled_output, &_weights_flipped, bias, output, conv_info, weights_info);
188 _scaled_output.allocator()->allocate();
189
190 // Setup flip axis data
191 _flip_axis.allocator()->allocate();
192 _flip_axis.map(true);
193 auto axis_data = reinterpret_cast<uint32_t *>(_flip_axis.buffer());
194 if(weights->info()->data_layout() == DataLayout::NHWC)
195 {
196 axis_data[0] = 1;
197 axis_data[1] = 2;
198 }
199 else
200 {
201 axis_data[0] = 0;
202 axis_data[1] = 1;
203 }
204 _flip_axis.unmap();
205 }
206
run()207 void CLDirectDeconvolutionLayer::run()
208 {
209 prepare();
210
211 MemoryGroupResourceScope scope_mg(_memory_group);
212
213 _scale_f.run();
214 _conv_f.run();
215 }
216
prepare()217 void CLDirectDeconvolutionLayer::prepare()
218 {
219 if(!_is_prepared)
220 {
221 ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
222
223 // Run weights flipping and mark original weights tensor as unused
224 _weights_flipped.allocator()->allocate();
225 _flip_weights.run();
226 _original_weights->mark_as_unused();
227
228 // Prepare convolution
229 _conv_f.prepare();
230
231 // Free flipped weights
232 if(!_weights_flipped.is_used())
233 {
234 _weights_flipped.allocator()->free();
235 }
236 _is_prepared = true;
237 }
238 }
239 } // namespace arm_compute
240