xref: /aosp_15_r20/external/ComputeLibrary/src/runtime/CL/functions/CLGenerateProposalsLayer.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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