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/CLGenerateProposalsLayer.h"
25
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Types.h"
28 #include "arm_compute/runtime/CL/functions/CLDequantizationLayer.h"
29 #include "arm_compute/runtime/CL/functions/CLQuantizationLayer.h"
30 #include "src/core/CL/kernels/CLBoundingBoxTransformKernel.h"
31 #include "src/core/CL/kernels/CLGenerateProposalsLayerKernel.h"
32 #include "src/core/CL/kernels/CLPadLayerKernel.h"
33 #include "src/core/helpers/AutoConfiguration.h"
34
35 #include "src/common/utils/Log.h"
36
37 namespace arm_compute
38 {
CLGenerateProposalsLayer(std::shared_ptr<IMemoryManager> memory_manager)39 CLGenerateProposalsLayer::CLGenerateProposalsLayer(std::shared_ptr<IMemoryManager> memory_manager)
40 : _memory_group(memory_manager),
41 _permute_deltas(),
42 _flatten_deltas(),
43 _permute_scores(),
44 _flatten_scores(),
45 _compute_anchors_kernel(std::make_unique<CLComputeAllAnchorsKernel>()),
46 _bounding_box_kernel(std::make_unique<CLBoundingBoxTransformKernel>()),
47 _pad_kernel(std::make_unique<CLPadLayerKernel>()),
48 _dequantize_anchors(std::make_unique<CLDequantizationLayer>()),
49 _dequantize_deltas(std::make_unique<CLDequantizationLayer>()),
50 _quantize_all_proposals(std::make_unique<CLQuantizationLayer>()),
51 _cpp_nms(memory_manager),
52 _is_nhwc(false),
53 _is_qasymm8(false),
54 _deltas_permuted(),
55 _deltas_flattened(),
56 _deltas_flattened_f32(),
57 _scores_permuted(),
58 _scores_flattened(),
59 _all_anchors(),
60 _all_anchors_f32(),
61 _all_proposals(),
62 _all_proposals_quantized(),
63 _keeps_nms_unused(),
64 _classes_nms_unused(),
65 _proposals_4_roi_values(),
66 _all_proposals_to_use(nullptr),
67 _num_valid_proposals(nullptr),
68 _scores_out(nullptr)
69 {
70 }
71
72 CLGenerateProposalsLayer::~CLGenerateProposalsLayer() = default;
73
configure(const ICLTensor * scores,const ICLTensor * deltas,const ICLTensor * anchors,ICLTensor * proposals,ICLTensor * scores_out,ICLTensor * num_valid_proposals,const GenerateProposalsInfo & info)74 void CLGenerateProposalsLayer::configure(const ICLTensor *scores, const ICLTensor *deltas, const ICLTensor *anchors, ICLTensor *proposals, ICLTensor *scores_out, ICLTensor *num_valid_proposals,
75 const GenerateProposalsInfo &info)
76 {
77 configure(CLKernelLibrary::get().get_compile_context(), scores, deltas, anchors, proposals, scores_out, num_valid_proposals, info);
78 }
79
configure(const CLCompileContext & compile_context,const ICLTensor * scores,const ICLTensor * deltas,const ICLTensor * anchors,ICLTensor * proposals,ICLTensor * scores_out,ICLTensor * num_valid_proposals,const GenerateProposalsInfo & info)80 void CLGenerateProposalsLayer::configure(const CLCompileContext &compile_context, const ICLTensor *scores, const ICLTensor *deltas, const ICLTensor *anchors, ICLTensor *proposals,
81 ICLTensor *scores_out,
82 ICLTensor *num_valid_proposals, const GenerateProposalsInfo &info)
83 {
84 ARM_COMPUTE_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals);
85 ARM_COMPUTE_ERROR_THROW_ON(CLGenerateProposalsLayer::validate(scores->info(), deltas->info(), anchors->info(), proposals->info(), scores_out->info(), num_valid_proposals->info(), info));
86 ARM_COMPUTE_LOG_PARAMS(scores, deltas, anchors, proposals, scores_out, num_valid_proposals, info);
87
88 _is_nhwc = scores->info()->data_layout() == DataLayout::NHWC;
89 const DataType scores_data_type = scores->info()->data_type();
90 _is_qasymm8 = scores_data_type == DataType::QASYMM8;
91 const int num_anchors = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::CHANNEL));
92 const int feat_width = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::WIDTH));
93 const int feat_height = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::HEIGHT));
94 const int total_num_anchors = num_anchors * feat_width * feat_height;
95 const int pre_nms_topN = info.pre_nms_topN();
96 const int post_nms_topN = info.post_nms_topN();
97 const size_t values_per_roi = info.values_per_roi();
98
99 const QuantizationInfo scores_qinfo = scores->info()->quantization_info();
100 const DataType rois_data_type = (_is_qasymm8) ? DataType::QASYMM16 : scores_data_type;
101 const QuantizationInfo rois_qinfo = (_is_qasymm8) ? QuantizationInfo(0.125f, 0) : scores->info()->quantization_info();
102
103 // Compute all the anchors
104 _memory_group.manage(&_all_anchors);
105 _compute_anchors_kernel->configure(compile_context, anchors, &_all_anchors, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale()));
106
107 const TensorShape flatten_shape_deltas(values_per_roi, total_num_anchors);
108 _deltas_flattened.allocator()->init(TensorInfo(flatten_shape_deltas, 1, scores_data_type, deltas->info()->quantization_info()));
109
110 // Permute and reshape deltas
111 _memory_group.manage(&_deltas_flattened);
112 if(!_is_nhwc)
113 {
114 _memory_group.manage(&_deltas_permuted);
115 _permute_deltas.configure(compile_context, deltas, &_deltas_permuted, PermutationVector{ 2, 0, 1 });
116 _flatten_deltas.configure(compile_context, &_deltas_permuted, &_deltas_flattened);
117 _deltas_permuted.allocator()->allocate();
118 }
119 else
120 {
121 _flatten_deltas.configure(compile_context, deltas, &_deltas_flattened);
122 }
123
124 const TensorShape flatten_shape_scores(1, total_num_anchors);
125 _scores_flattened.allocator()->init(TensorInfo(flatten_shape_scores, 1, scores_data_type, scores_qinfo));
126
127 // Permute and reshape scores
128 _memory_group.manage(&_scores_flattened);
129 if(!_is_nhwc)
130 {
131 _memory_group.manage(&_scores_permuted);
132 _permute_scores.configure(compile_context, scores, &_scores_permuted, PermutationVector{ 2, 0, 1 });
133 _flatten_scores.configure(compile_context, &_scores_permuted, &_scores_flattened);
134 _scores_permuted.allocator()->allocate();
135 }
136 else
137 {
138 _flatten_scores.configure(compile_context, scores, &_scores_flattened);
139 }
140
141 CLTensor *anchors_to_use = &_all_anchors;
142 CLTensor *deltas_to_use = &_deltas_flattened;
143 if(_is_qasymm8)
144 {
145 _all_anchors_f32.allocator()->init(TensorInfo(_all_anchors.info()->tensor_shape(), 1, DataType::F32));
146 _deltas_flattened_f32.allocator()->init(TensorInfo(_deltas_flattened.info()->tensor_shape(), 1, DataType::F32));
147 _memory_group.manage(&_all_anchors_f32);
148 _memory_group.manage(&_deltas_flattened_f32);
149 // Dequantize anchors to float
150 _dequantize_anchors->configure(compile_context, &_all_anchors, &_all_anchors_f32);
151 _all_anchors.allocator()->allocate();
152 anchors_to_use = &_all_anchors_f32;
153 // Dequantize deltas to float
154 _dequantize_deltas->configure(compile_context, &_deltas_flattened, &_deltas_flattened_f32);
155 _deltas_flattened.allocator()->allocate();
156 deltas_to_use = &_deltas_flattened_f32;
157 }
158 // Bounding box transform
159 _memory_group.manage(&_all_proposals);
160 BoundingBoxTransformInfo bbox_info(info.im_width(), info.im_height(), 1.f);
161 _bounding_box_kernel->configure(compile_context, anchors_to_use, &_all_proposals, deltas_to_use, bbox_info);
162 deltas_to_use->allocator()->allocate();
163 anchors_to_use->allocator()->allocate();
164
165 _all_proposals_to_use = &_all_proposals;
166 if(_is_qasymm8)
167 {
168 _memory_group.manage(&_all_proposals_quantized);
169 // Requantize all_proposals to QASYMM16 with 0.125 scale and 0 offset
170 _all_proposals_quantized.allocator()->init(TensorInfo(_all_proposals.info()->tensor_shape(), 1, DataType::QASYMM16, QuantizationInfo(0.125f, 0)));
171 _quantize_all_proposals->configure(compile_context, &_all_proposals, &_all_proposals_quantized);
172 _all_proposals.allocator()->allocate();
173 _all_proposals_to_use = &_all_proposals_quantized;
174 }
175 // The original layer implementation first selects the best pre_nms_topN anchors (thus having a lightweight sort)
176 // that are then transformed by bbox_transform. The boxes generated are then fed into a non-sorting NMS operation.
177 // Since we are reusing the NMS layer and we don't implement any CL/sort, we let NMS do the sorting (of all the input)
178 // and the filtering
179 const int scores_nms_size = std::min<int>(std::min<int>(post_nms_topN, pre_nms_topN), total_num_anchors);
180 const float min_size_scaled = info.min_size() * info.im_scale();
181 _memory_group.manage(&_classes_nms_unused);
182 _memory_group.manage(&_keeps_nms_unused);
183
184 // Note that NMS needs outputs preinitialized.
185 auto_init_if_empty(*scores_out->info(), TensorShape(scores_nms_size), 1, scores_data_type, scores_qinfo);
186 auto_init_if_empty(*_proposals_4_roi_values.info(), TensorShape(values_per_roi, scores_nms_size), 1, rois_data_type, rois_qinfo);
187 auto_init_if_empty(*num_valid_proposals->info(), TensorShape(1), 1, DataType::U32);
188
189 // Initialize temporaries (unused) outputs
190 _classes_nms_unused.allocator()->init(TensorInfo(TensorShape(scores_nms_size), 1, scores_data_type, scores_qinfo));
191 _keeps_nms_unused.allocator()->init(*scores_out->info());
192
193 // Save the output (to map and unmap them at run)
194 _scores_out = scores_out;
195 _num_valid_proposals = num_valid_proposals;
196
197 _memory_group.manage(&_proposals_4_roi_values);
198 _cpp_nms.configure(&_scores_flattened, _all_proposals_to_use, nullptr, scores_out, &_proposals_4_roi_values, &_classes_nms_unused, nullptr, &_keeps_nms_unused, num_valid_proposals,
199 BoxNMSLimitInfo(0.0f, info.nms_thres(), scores_nms_size, false, NMSType::LINEAR, 0.5f, 0.001f, true, min_size_scaled, info.im_width(), info.im_height()));
200 _keeps_nms_unused.allocator()->allocate();
201 _classes_nms_unused.allocator()->allocate();
202 _all_proposals_to_use->allocator()->allocate();
203 _scores_flattened.allocator()->allocate();
204
205 // Add the first column that represents the batch id. This will be all zeros, as we don't support multiple images
206 _pad_kernel->configure(compile_context, &_proposals_4_roi_values, proposals, PaddingList{ { 1, 0 } });
207 _proposals_4_roi_values.allocator()->allocate();
208 }
209
validate(const ITensorInfo * scores,const ITensorInfo * deltas,const ITensorInfo * anchors,const ITensorInfo * proposals,const ITensorInfo * scores_out,const ITensorInfo * num_valid_proposals,const GenerateProposalsInfo & info)210 Status CLGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITensorInfo *deltas, const ITensorInfo *anchors, const ITensorInfo *proposals, const ITensorInfo *scores_out,
211 const ITensorInfo *num_valid_proposals, const GenerateProposalsInfo &info)
212 {
213 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals);
214 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(scores, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
215 ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(scores, DataLayout::NCHW, DataLayout::NHWC);
216 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(scores, deltas);
217 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(scores, deltas);
218
219 const int num_anchors = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::CHANNEL));
220 const int feat_width = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::WIDTH));
221 const int feat_height = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::HEIGHT));
222 const int num_images = scores->dimension(3);
223 const int total_num_anchors = num_anchors * feat_width * feat_height;
224 const int values_per_roi = info.values_per_roi();
225
226 const bool is_qasymm8 = scores->data_type() == DataType::QASYMM8;
227
228 ARM_COMPUTE_RETURN_ERROR_ON(num_images > 1);
229
230 if(is_qasymm8)
231 {
232 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(anchors, 1, DataType::QSYMM16);
233 const UniformQuantizationInfo anchors_qinfo = anchors->quantization_info().uniform();
234 ARM_COMPUTE_RETURN_ERROR_ON(anchors_qinfo.scale != 0.125f);
235 }
236
237 TensorInfo all_anchors_info(anchors->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
238 ARM_COMPUTE_RETURN_ON_ERROR(CLComputeAllAnchorsKernel::validate(anchors, &all_anchors_info, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale())));
239
240 TensorInfo deltas_permuted_info = deltas->clone()->set_tensor_shape(TensorShape(values_per_roi * num_anchors, feat_width, feat_height)).set_is_resizable(true);
241 TensorInfo scores_permuted_info = scores->clone()->set_tensor_shape(TensorShape(num_anchors, feat_width, feat_height)).set_is_resizable(true);
242 if(scores->data_layout() == DataLayout::NHWC)
243 {
244 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(deltas, &deltas_permuted_info);
245 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(scores, &scores_permuted_info);
246 }
247 else
248 {
249 ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(deltas, &deltas_permuted_info, PermutationVector{ 2, 0, 1 }));
250 ARM_COMPUTE_RETURN_ON_ERROR(CLPermute::validate(scores, &scores_permuted_info, PermutationVector{ 2, 0, 1 }));
251 }
252
253 TensorInfo deltas_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
254 ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(&deltas_permuted_info, &deltas_flattened_info));
255
256 TensorInfo scores_flattened_info(scores->clone()->set_tensor_shape(TensorShape(1, total_num_anchors)).set_is_resizable(true));
257 TensorInfo proposals_4_roi_values(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
258
259 ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(&scores_permuted_info, &scores_flattened_info));
260
261 TensorInfo *proposals_4_roi_values_to_use = &proposals_4_roi_values;
262 TensorInfo proposals_4_roi_values_quantized(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
263 proposals_4_roi_values_quantized.set_data_type(DataType::QASYMM16).set_quantization_info(QuantizationInfo(0.125f, 0));
264 if(is_qasymm8)
265 {
266 TensorInfo all_anchors_f32_info(anchors->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true).set_data_type(DataType::F32));
267 ARM_COMPUTE_RETURN_ON_ERROR(CLDequantizationLayer::validate(&all_anchors_info, &all_anchors_f32_info));
268
269 TensorInfo deltas_flattened_f32_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true).set_data_type(DataType::F32));
270 ARM_COMPUTE_RETURN_ON_ERROR(CLDequantizationLayer::validate(&deltas_flattened_info, &deltas_flattened_f32_info));
271
272 TensorInfo proposals_4_roi_values_f32(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true).set_data_type(DataType::F32));
273 ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_f32_info, &proposals_4_roi_values_f32, &deltas_flattened_f32_info,
274 BoundingBoxTransformInfo(info.im_width(), info.im_height(), 1.f)));
275
276 ARM_COMPUTE_RETURN_ON_ERROR(CLQuantizationLayer::validate(&proposals_4_roi_values_f32, &proposals_4_roi_values_quantized));
277 proposals_4_roi_values_to_use = &proposals_4_roi_values_quantized;
278 }
279 else
280 {
281 ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_info, &proposals_4_roi_values, &deltas_flattened_info,
282 BoundingBoxTransformInfo(info.im_width(), info.im_height(), 1.f)));
283 }
284
285 ARM_COMPUTE_RETURN_ON_ERROR(CLPadLayerKernel::validate(proposals_4_roi_values_to_use, proposals, PaddingList{ { 1, 0 } }));
286
287 if(num_valid_proposals->total_size() > 0)
288 {
289 ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->num_dimensions() > 1);
290 ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->dimension(0) > 1);
291 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(num_valid_proposals, 1, DataType::U32);
292 }
293
294 if(proposals->total_size() > 0)
295 {
296 ARM_COMPUTE_RETURN_ERROR_ON(proposals->num_dimensions() > 2);
297 ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(0) != size_t(values_per_roi) + 1);
298 ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(1) != size_t(total_num_anchors));
299 if(is_qasymm8)
300 {
301 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(proposals, 1, DataType::QASYMM16);
302 const UniformQuantizationInfo proposals_qinfo = proposals->quantization_info().uniform();
303 ARM_COMPUTE_RETURN_ERROR_ON(proposals_qinfo.scale != 0.125f);
304 ARM_COMPUTE_RETURN_ERROR_ON(proposals_qinfo.offset != 0);
305 }
306 else
307 {
308 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(proposals, scores);
309 }
310 }
311
312 if(scores_out->total_size() > 0)
313 {
314 ARM_COMPUTE_RETURN_ERROR_ON(scores_out->num_dimensions() > 1);
315 ARM_COMPUTE_RETURN_ERROR_ON(scores_out->dimension(0) != size_t(total_num_anchors));
316 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(scores_out, scores);
317 }
318
319 return Status{};
320 }
321
run_cpp_nms_kernel()322 void CLGenerateProposalsLayer::run_cpp_nms_kernel()
323 {
324 // Map inputs
325 _scores_flattened.map(true);
326 _all_proposals_to_use->map(true);
327
328 // Map outputs
329 _scores_out->map(CLScheduler::get().queue(), true);
330 _proposals_4_roi_values.map(CLScheduler::get().queue(), true);
331 _num_valid_proposals->map(CLScheduler::get().queue(), true);
332 _keeps_nms_unused.map(true);
333 _classes_nms_unused.map(true);
334
335 // Run nms
336 _cpp_nms.run();
337
338 // Unmap outputs
339 _keeps_nms_unused.unmap();
340 _classes_nms_unused.unmap();
341 _scores_out->unmap(CLScheduler::get().queue());
342 _proposals_4_roi_values.unmap(CLScheduler::get().queue());
343 _num_valid_proposals->unmap(CLScheduler::get().queue());
344
345 // Unmap inputs
346 _scores_flattened.unmap();
347 _all_proposals_to_use->unmap();
348 }
349
run()350 void CLGenerateProposalsLayer::run()
351 {
352 // Acquire all the temporaries
353 MemoryGroupResourceScope scope_mg(_memory_group);
354
355 // Compute all the anchors
356 CLScheduler::get().enqueue(*_compute_anchors_kernel, false);
357
358 // Transpose and reshape the inputs
359 if(!_is_nhwc)
360 {
361 _permute_deltas.run();
362 _permute_scores.run();
363 }
364 _flatten_deltas.run();
365 _flatten_scores.run();
366
367 if(_is_qasymm8)
368 {
369 _dequantize_anchors->run();
370 _dequantize_deltas->run();
371 }
372
373 // Build the boxes
374 CLScheduler::get().enqueue(*_bounding_box_kernel, false);
375
376 if(_is_qasymm8)
377 {
378 _quantize_all_proposals->run();
379 }
380
381 // Non maxima suppression
382 run_cpp_nms_kernel();
383 // Add dummy batch indexes
384 CLScheduler::get().enqueue(*_pad_kernel, true);
385 }
386 } // namespace arm_compute
387