1 /*
2 * Copyright (c) 2019-2020 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 "src/core/CL/kernels/CLBoundingBoxTransformKernel.h"
29 #include "src/core/CL/kernels/CLDequantizationLayerKernel.h"
30 #include "src/core/CL/kernels/CLGenerateProposalsLayerKernel.h"
31 #include "src/core/CL/kernels/CLPadLayerKernel.h"
32 #include "src/core/CL/kernels/CLPermuteKernel.h"
33 #include "src/core/CL/kernels/CLQuantizationLayerKernel.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "support/MemorySupport.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_kernel(support::cpp14::make_unique<CLPermuteKernel>()),
42 _flatten_deltas(),
43 _permute_scores_kernel(support::cpp14::make_unique<CLPermuteKernel>()),
44 _flatten_scores(),
45 _compute_anchors_kernel(support::cpp14::make_unique<CLComputeAllAnchorsKernel>()),
46 _bounding_box_kernel(support::cpp14::make_unique<CLBoundingBoxTransformKernel>()),
47 _pad_kernel(support::cpp14::make_unique<CLPadLayerKernel>()),
48 _dequantize_anchors(support::cpp14::make_unique<CLDequantizationLayerKernel>()),
49 _dequantize_deltas(support::cpp14::make_unique<CLDequantizationLayerKernel>()),
50 _quantize_all_proposals(support::cpp14::make_unique<CLQuantizationLayerKernel>()),
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
87 _is_nhwc = scores->info()->data_layout() == DataLayout::NHWC;
88 const DataType scores_data_type = scores->info()->data_type();
89 _is_qasymm8 = scores_data_type == DataType::QASYMM8;
90 const int num_anchors = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::CHANNEL));
91 const int feat_width = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::WIDTH));
92 const int feat_height = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::HEIGHT));
93 const int total_num_anchors = num_anchors * feat_width * feat_height;
94 const int pre_nms_topN = info.pre_nms_topN();
95 const int post_nms_topN = info.post_nms_topN();
96 const size_t values_per_roi = info.values_per_roi();
97
98 const QuantizationInfo scores_qinfo = scores->info()->quantization_info();
99 const DataType rois_data_type = (_is_qasymm8) ? DataType::QASYMM16 : scores_data_type;
100 const QuantizationInfo rois_qinfo = (_is_qasymm8) ? QuantizationInfo(0.125f, 0) : scores->info()->quantization_info();
101
102 // Compute all the anchors
103 _memory_group.manage(&_all_anchors);
104 _compute_anchors_kernel->configure(compile_context, anchors, &_all_anchors, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale()));
105
106 const TensorShape flatten_shape_deltas(values_per_roi, total_num_anchors);
107 _deltas_flattened.allocator()->init(TensorInfo(flatten_shape_deltas, 1, scores_data_type, deltas->info()->quantization_info()));
108
109 // Permute and reshape deltas
110 _memory_group.manage(&_deltas_flattened);
111 if(!_is_nhwc)
112 {
113 _memory_group.manage(&_deltas_permuted);
114 _permute_deltas_kernel->configure(compile_context, deltas, &_deltas_permuted, PermutationVector{ 2, 0, 1 });
115 _flatten_deltas.configure(compile_context, &_deltas_permuted, &_deltas_flattened);
116 _deltas_permuted.allocator()->allocate();
117 }
118 else
119 {
120 _flatten_deltas.configure(compile_context, deltas, &_deltas_flattened);
121 }
122
123 const TensorShape flatten_shape_scores(1, total_num_anchors);
124 _scores_flattened.allocator()->init(TensorInfo(flatten_shape_scores, 1, scores_data_type, scores_qinfo));
125
126 // Permute and reshape scores
127 _memory_group.manage(&_scores_flattened);
128 if(!_is_nhwc)
129 {
130 _memory_group.manage(&_scores_permuted);
131 _permute_scores_kernel->configure(compile_context, scores, &_scores_permuted, PermutationVector{ 2, 0, 1 });
132 _flatten_scores.configure(compile_context, &_scores_permuted, &_scores_flattened);
133 _scores_permuted.allocator()->allocate();
134 }
135 else
136 {
137 _flatten_scores.configure(compile_context, scores, &_scores_flattened);
138 }
139
140 CLTensor *anchors_to_use = &_all_anchors;
141 CLTensor *deltas_to_use = &_deltas_flattened;
142 if(_is_qasymm8)
143 {
144 _all_anchors_f32.allocator()->init(TensorInfo(_all_anchors.info()->tensor_shape(), 1, DataType::F32));
145 _deltas_flattened_f32.allocator()->init(TensorInfo(_deltas_flattened.info()->tensor_shape(), 1, DataType::F32));
146 _memory_group.manage(&_all_anchors_f32);
147 _memory_group.manage(&_deltas_flattened_f32);
148 // Dequantize anchors to float
149 _dequantize_anchors->configure(compile_context, &_all_anchors, &_all_anchors_f32);
150 _all_anchors.allocator()->allocate();
151 anchors_to_use = &_all_anchors_f32;
152 // Dequantize deltas to float
153 _dequantize_deltas->configure(compile_context, &_deltas_flattened, &_deltas_flattened_f32);
154 _deltas_flattened.allocator()->allocate();
155 deltas_to_use = &_deltas_flattened_f32;
156 }
157 // Bounding box transform
158 _memory_group.manage(&_all_proposals);
159 BoundingBoxTransformInfo bbox_info(info.im_width(), info.im_height(), 1.f);
160 _bounding_box_kernel->configure(compile_context, anchors_to_use, &_all_proposals, deltas_to_use, bbox_info);
161 deltas_to_use->allocator()->allocate();
162 anchors_to_use->allocator()->allocate();
163
164 _all_proposals_to_use = &_all_proposals;
165 if(_is_qasymm8)
166 {
167 _memory_group.manage(&_all_proposals_quantized);
168 // Requantize all_proposals to QASYMM16 with 0.125 scale and 0 offset
169 _all_proposals_quantized.allocator()->init(TensorInfo(_all_proposals.info()->tensor_shape(), 1, DataType::QASYMM16, QuantizationInfo(0.125f, 0)));
170 _quantize_all_proposals->configure(compile_context, &_all_proposals, &_all_proposals_quantized);
171 _all_proposals.allocator()->allocate();
172 _all_proposals_to_use = &_all_proposals_quantized;
173 }
174 // The original layer implementation first selects the best pre_nms_topN anchors (thus having a lightweight sort)
175 // that are then transformed by bbox_transform. The boxes generated are then fed into a non-sorting NMS operation.
176 // 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)
177 // and the filtering
178 const int scores_nms_size = std::min<int>(std::min<int>(post_nms_topN, pre_nms_topN), total_num_anchors);
179 const float min_size_scaled = info.min_size() * info.im_scale();
180 _memory_group.manage(&_classes_nms_unused);
181 _memory_group.manage(&_keeps_nms_unused);
182
183 // Note that NMS needs outputs preinitialized.
184 auto_init_if_empty(*scores_out->info(), TensorShape(scores_nms_size), 1, scores_data_type, scores_qinfo);
185 auto_init_if_empty(*_proposals_4_roi_values.info(), TensorShape(values_per_roi, scores_nms_size), 1, rois_data_type, rois_qinfo);
186 auto_init_if_empty(*num_valid_proposals->info(), TensorShape(1), 1, DataType::U32);
187
188 // Initialize temporaries (unused) outputs
189 _classes_nms_unused.allocator()->init(TensorInfo(TensorShape(scores_nms_size), 1, scores_data_type, scores_qinfo));
190 _keeps_nms_unused.allocator()->init(*scores_out->info());
191
192 // Save the output (to map and unmap them at run)
193 _scores_out = scores_out;
194 _num_valid_proposals = num_valid_proposals;
195
196 _memory_group.manage(&_proposals_4_roi_values);
197 _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,
198 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()));
199 _keeps_nms_unused.allocator()->allocate();
200 _classes_nms_unused.allocator()->allocate();
201 _all_proposals_to_use->allocator()->allocate();
202 _scores_flattened.allocator()->allocate();
203
204 // Add the first column that represents the batch id. This will be all zeros, as we don't support multiple images
205 _pad_kernel->configure(compile_context, &_proposals_4_roi_values, proposals, PaddingList{ { 1, 0 } });
206 _proposals_4_roi_values.allocator()->allocate();
207 }
208
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)209 Status CLGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITensorInfo *deltas, const ITensorInfo *anchors, const ITensorInfo *proposals, const ITensorInfo *scores_out,
210 const ITensorInfo *num_valid_proposals, const GenerateProposalsInfo &info)
211 {
212 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals);
213 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(scores, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
214 ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(scores, DataLayout::NCHW, DataLayout::NHWC);
215 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(scores, deltas);
216 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(scores, deltas);
217
218 const int num_anchors = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::CHANNEL));
219 const int feat_width = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::WIDTH));
220 const int feat_height = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::HEIGHT));
221 const int num_images = scores->dimension(3);
222 const int total_num_anchors = num_anchors * feat_width * feat_height;
223 const int values_per_roi = info.values_per_roi();
224
225 const bool is_qasymm8 = scores->data_type() == DataType::QASYMM8;
226
227 ARM_COMPUTE_RETURN_ERROR_ON(num_images > 1);
228
229 if(is_qasymm8)
230 {
231 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(anchors, 1, DataType::QSYMM16);
232 const UniformQuantizationInfo anchors_qinfo = anchors->quantization_info().uniform();
233 ARM_COMPUTE_RETURN_ERROR_ON(anchors_qinfo.scale != 0.125f);
234 }
235
236 TensorInfo all_anchors_info(anchors->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
237 ARM_COMPUTE_RETURN_ON_ERROR(CLComputeAllAnchorsKernel::validate(anchors, &all_anchors_info, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale())));
238
239 TensorInfo deltas_permuted_info = deltas->clone()->set_tensor_shape(TensorShape(values_per_roi * num_anchors, feat_width, feat_height)).set_is_resizable(true);
240 TensorInfo scores_permuted_info = scores->clone()->set_tensor_shape(TensorShape(num_anchors, feat_width, feat_height)).set_is_resizable(true);
241 if(scores->data_layout() == DataLayout::NHWC)
242 {
243 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(deltas, &deltas_permuted_info);
244 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(scores, &scores_permuted_info);
245 }
246 else
247 {
248 ARM_COMPUTE_RETURN_ON_ERROR(CLPermuteKernel::validate(deltas, &deltas_permuted_info, PermutationVector{ 2, 0, 1 }));
249 ARM_COMPUTE_RETURN_ON_ERROR(CLPermuteKernel::validate(scores, &scores_permuted_info, PermutationVector{ 2, 0, 1 }));
250 }
251
252 TensorInfo deltas_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
253 ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(&deltas_permuted_info, &deltas_flattened_info));
254
255 TensorInfo scores_flattened_info(scores->clone()->set_tensor_shape(TensorShape(1, total_num_anchors)).set_is_resizable(true));
256 TensorInfo proposals_4_roi_values(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
257
258 ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayer::validate(&scores_permuted_info, &scores_flattened_info));
259
260 TensorInfo *proposals_4_roi_values_to_use = &proposals_4_roi_values;
261 TensorInfo proposals_4_roi_values_quantized(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true));
262 proposals_4_roi_values_quantized.set_data_type(DataType::QASYMM16).set_quantization_info(QuantizationInfo(0.125f, 0));
263 if(is_qasymm8)
264 {
265 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));
266 ARM_COMPUTE_RETURN_ON_ERROR(CLDequantizationLayerKernel::validate(&all_anchors_info, &all_anchors_f32_info));
267
268 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));
269 ARM_COMPUTE_RETURN_ON_ERROR(CLDequantizationLayerKernel::validate(&deltas_flattened_info, &deltas_flattened_f32_info));
270
271 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));
272 ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_f32_info, &proposals_4_roi_values_f32, &deltas_flattened_f32_info,
273 BoundingBoxTransformInfo(info.im_width(), info.im_height(), 1.f)));
274
275 ARM_COMPUTE_RETURN_ON_ERROR(CLQuantizationLayerKernel::validate(&proposals_4_roi_values_f32, &proposals_4_roi_values_quantized));
276 proposals_4_roi_values_to_use = &proposals_4_roi_values_quantized;
277 }
278 else
279 {
280 ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_info, &proposals_4_roi_values, &deltas_flattened_info,
281 BoundingBoxTransformInfo(info.im_width(), info.im_height(), 1.f)));
282 }
283
284 ARM_COMPUTE_RETURN_ON_ERROR(CLPadLayerKernel::validate(proposals_4_roi_values_to_use, proposals, PaddingList{ { 1, 0 } }));
285
286 if(num_valid_proposals->total_size() > 0)
287 {
288 ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->num_dimensions() > 1);
289 ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->dimension(0) > 1);
290 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(num_valid_proposals, 1, DataType::U32);
291 }
292
293 if(proposals->total_size() > 0)
294 {
295 ARM_COMPUTE_RETURN_ERROR_ON(proposals->num_dimensions() > 2);
296 ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(0) != size_t(values_per_roi) + 1);
297 ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(1) != size_t(total_num_anchors));
298 if(is_qasymm8)
299 {
300 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(proposals, 1, DataType::QASYMM16);
301 const UniformQuantizationInfo proposals_qinfo = proposals->quantization_info().uniform();
302 ARM_COMPUTE_RETURN_ERROR_ON(proposals_qinfo.scale != 0.125f);
303 ARM_COMPUTE_RETURN_ERROR_ON(proposals_qinfo.offset != 0);
304 }
305 else
306 {
307 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(proposals, scores);
308 }
309 }
310
311 if(scores_out->total_size() > 0)
312 {
313 ARM_COMPUTE_RETURN_ERROR_ON(scores_out->num_dimensions() > 1);
314 ARM_COMPUTE_RETURN_ERROR_ON(scores_out->dimension(0) != size_t(total_num_anchors));
315 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(scores_out, scores);
316 }
317
318 return Status{};
319 }
320
run_cpp_nms_kernel()321 void CLGenerateProposalsLayer::run_cpp_nms_kernel()
322 {
323 // Map inputs
324 _scores_flattened.map(true);
325 _all_proposals_to_use->map(true);
326
327 // Map outputs
328 _scores_out->map(CLScheduler::get().queue(), true);
329 _proposals_4_roi_values.map(CLScheduler::get().queue(), true);
330 _num_valid_proposals->map(CLScheduler::get().queue(), true);
331 _keeps_nms_unused.map(true);
332 _classes_nms_unused.map(true);
333
334 // Run nms
335 _cpp_nms.run();
336
337 // Unmap outputs
338 _keeps_nms_unused.unmap();
339 _classes_nms_unused.unmap();
340 _scores_out->unmap(CLScheduler::get().queue());
341 _proposals_4_roi_values.unmap(CLScheduler::get().queue());
342 _num_valid_proposals->unmap(CLScheduler::get().queue());
343
344 // Unmap inputs
345 _scores_flattened.unmap();
346 _all_proposals_to_use->unmap();
347 }
348
run()349 void CLGenerateProposalsLayer::run()
350 {
351 // Acquire all the temporaries
352 MemoryGroupResourceScope scope_mg(_memory_group);
353
354 // Compute all the anchors
355 CLScheduler::get().enqueue(*_compute_anchors_kernel, false);
356
357 // Transpose and reshape the inputs
358 if(!_is_nhwc)
359 {
360 CLScheduler::get().enqueue(*_permute_deltas_kernel, false);
361 CLScheduler::get().enqueue(*_permute_scores_kernel, false);
362 }
363 _flatten_deltas.run();
364 _flatten_scores.run();
365
366 if(_is_qasymm8)
367 {
368 CLScheduler::get().enqueue(*_dequantize_anchors, false);
369 CLScheduler::get().enqueue(*_dequantize_deltas, false);
370 }
371
372 // Build the boxes
373 CLScheduler::get().enqueue(*_bounding_box_kernel, false);
374
375 if(_is_qasymm8)
376 {
377 CLScheduler::get().enqueue(*_quantize_all_proposals, false);
378 }
379
380 // Non maxima suppression
381 run_cpp_nms_kernel();
382 // Add dummy batch indexes
383 CLScheduler::get().enqueue(*_pad_kernel, true);
384 }
385 } // namespace arm_compute
386