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