Manuel Bottini | 5209be5 | 2019-02-13 16:34:56 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019 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 "support/ToolchainSupport.h" |
| 29 | |
| 30 | namespace arm_compute |
| 31 | { |
| 32 | CLGenerateProposalsLayer::CLGenerateProposalsLayer(std::shared_ptr<IMemoryManager> memory_manager) |
| 33 | : _memory_group(std::move(memory_manager)), |
| 34 | _permute_deltas_kernel(), |
| 35 | _flatten_deltas_kernel(), |
| 36 | _permute_scores_kernel(), |
| 37 | _flatten_scores_kernel(), |
| 38 | _compute_anchors_kernel(), |
| 39 | _bounding_box_kernel(), |
| 40 | _memset_kernel(), |
| 41 | _padded_copy_kernel(), |
| 42 | _cpp_nms_kernel(), |
| 43 | _is_nhwc(false), |
| 44 | _deltas_permuted(), |
| 45 | _deltas_flattened(), |
| 46 | _scores_permuted(), |
| 47 | _scores_flattened(), |
| 48 | _all_anchors(), |
| 49 | _all_proposals(), |
| 50 | _keeps_nms_unused(), |
| 51 | _classes_nms_unused(), |
| 52 | _proposals_4_roi_values(), |
| 53 | _num_valid_proposals(nullptr), |
| 54 | _scores_out(nullptr) |
| 55 | { |
| 56 | } |
| 57 | |
| 58 | void CLGenerateProposalsLayer::configure(const ICLTensor *scores, const ICLTensor *deltas, const ICLTensor *anchors, ICLTensor *proposals, ICLTensor *scores_out, ICLTensor *num_valid_proposals, |
| 59 | const GenerateProposalsInfo &info) |
| 60 | { |
| 61 | ARM_COMPUTE_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals); |
| 62 | ARM_COMPUTE_ERROR_THROW_ON(CLGenerateProposalsLayer::validate(scores->info(), deltas->info(), anchors->info(), proposals->info(), scores_out->info(), num_valid_proposals->info(), info)); |
| 63 | |
| 64 | _is_nhwc = scores->info()->data_layout() == DataLayout::NHWC; |
| 65 | const DataType data_type = deltas->info()->data_type(); |
| 66 | const int num_anchors = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::CHANNEL)); |
| 67 | const int feat_width = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::WIDTH)); |
| 68 | const int feat_height = scores->info()->dimension(get_data_layout_dimension_index(scores->info()->data_layout(), DataLayoutDimension::HEIGHT)); |
| 69 | const int total_num_anchors = num_anchors * feat_width * feat_height; |
| 70 | const int pre_nms_topN = info.pre_nms_topN(); |
| 71 | const int post_nms_topN = info.post_nms_topN(); |
| 72 | const size_t values_per_roi = info.values_per_roi(); |
| 73 | |
| 74 | // Compute all the anchors |
| 75 | _memory_group.manage(&_all_anchors); |
| 76 | _compute_anchors_kernel.configure(anchors, &_all_anchors, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale())); |
| 77 | |
| 78 | const TensorShape flatten_shape_deltas(values_per_roi, total_num_anchors); |
| 79 | _deltas_flattened.allocator()->init(TensorInfo(flatten_shape_deltas, 1, data_type)); |
| 80 | |
| 81 | // Permute and reshape deltas |
| 82 | if(!_is_nhwc) |
| 83 | { |
| 84 | _memory_group.manage(&_deltas_permuted); |
| 85 | _memory_group.manage(&_deltas_flattened); |
| 86 | _permute_deltas_kernel.configure(deltas, &_deltas_permuted, PermutationVector{ 2, 0, 1 }); |
| 87 | _flatten_deltas_kernel.configure(&_deltas_permuted, &_deltas_flattened); |
| 88 | _deltas_permuted.allocator()->allocate(); |
| 89 | } |
| 90 | else |
| 91 | { |
| 92 | _memory_group.manage(&_deltas_flattened); |
| 93 | _flatten_deltas_kernel.configure(deltas, &_deltas_flattened); |
| 94 | } |
| 95 | |
| 96 | const TensorShape flatten_shape_scores(1, total_num_anchors); |
| 97 | _scores_flattened.allocator()->init(TensorInfo(flatten_shape_scores, 1, data_type)); |
| 98 | |
| 99 | // Permute and reshape scores |
| 100 | if(!_is_nhwc) |
| 101 | { |
| 102 | _memory_group.manage(&_scores_permuted); |
| 103 | _memory_group.manage(&_scores_flattened); |
| 104 | _permute_scores_kernel.configure(scores, &_scores_permuted, PermutationVector{ 2, 0, 1 }); |
| 105 | _flatten_scores_kernel.configure(&_scores_permuted, &_scores_flattened); |
| 106 | _scores_permuted.allocator()->allocate(); |
| 107 | } |
| 108 | else |
| 109 | { |
| 110 | _memory_group.manage(&_scores_flattened); |
| 111 | _flatten_scores_kernel.configure(scores, &_scores_flattened); |
| 112 | } |
| 113 | |
| 114 | // Bounding box transform |
| 115 | _memory_group.manage(&_all_proposals); |
| 116 | BoundingBoxTransformInfo bbox_info(info.im_width(), info.im_height(), 1.f); |
| 117 | _bounding_box_kernel.configure(&_all_anchors, &_all_proposals, &_deltas_flattened, bbox_info); |
| 118 | _deltas_flattened.allocator()->allocate(); |
| 119 | _all_anchors.allocator()->allocate(); |
| 120 | |
| 121 | // The original layer implementation first selects the best pre_nms_topN anchors (thus having a lightweight sort) |
| 122 | // that are then transformed by bbox_transform. The boxes generated are then fed into a non-sorting NMS operation. |
| 123 | // 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) |
| 124 | // and the filtering |
| 125 | const int scores_nms_size = std::min<int>(std::min<int>(post_nms_topN, pre_nms_topN), total_num_anchors); |
| 126 | const float min_size_scaled = info.min_size() * info.im_scale(); |
| 127 | _memory_group.manage(&_classes_nms_unused); |
| 128 | _memory_group.manage(&_keeps_nms_unused); |
| 129 | |
| 130 | // Note that NMS needs outputs preinitialized. |
| 131 | auto_init_if_empty(*scores_out->info(), TensorShape(scores_nms_size), 1, data_type); |
| 132 | auto_init_if_empty(*_proposals_4_roi_values.info(), TensorShape(values_per_roi, scores_nms_size), 1, data_type); |
| 133 | auto_init_if_empty(*num_valid_proposals->info(), TensorShape(1), 1, DataType::U32); |
| 134 | |
| 135 | // Initialize temporaries (unused) outputs |
| 136 | _classes_nms_unused.allocator()->init(TensorInfo(TensorShape(1, 1), 1, data_type)); |
| 137 | _keeps_nms_unused.allocator()->init(*scores_out->info()); |
| 138 | |
| 139 | // Save the output (to map and unmap them at run) |
| 140 | _scores_out = scores_out; |
| 141 | _num_valid_proposals = num_valid_proposals; |
| 142 | |
| 143 | _memory_group.manage(&_proposals_4_roi_values); |
| 144 | _cpp_nms_kernel.configure(&_scores_flattened, &_all_proposals, nullptr, scores_out, &_proposals_4_roi_values, &_classes_nms_unused, nullptr, &_keeps_nms_unused, num_valid_proposals, |
| 145 | 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())); |
| 146 | _keeps_nms_unused.allocator()->allocate(); |
| 147 | _classes_nms_unused.allocator()->allocate(); |
| 148 | _all_proposals.allocator()->allocate(); |
| 149 | _scores_flattened.allocator()->allocate(); |
| 150 | |
| 151 | // Add the first column that represents the batch id. This will be all zeros, as we don't support multiple images |
| 152 | _padded_copy_kernel.configure(&_proposals_4_roi_values, proposals, PaddingList{ { 1, 0 } }); |
| 153 | _proposals_4_roi_values.allocator()->allocate(); |
| 154 | |
| 155 | _memset_kernel.configure(proposals, PixelValue()); |
| 156 | } |
| 157 | |
| 158 | Status CLGenerateProposalsLayer::validate(const ITensorInfo *scores, const ITensorInfo *deltas, const ITensorInfo *anchors, const ITensorInfo *proposals, const ITensorInfo *scores_out, |
| 159 | const ITensorInfo *num_valid_proposals, const GenerateProposalsInfo &info) |
| 160 | { |
| 161 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores, deltas, anchors, proposals, scores_out, num_valid_proposals); |
| 162 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(scores, DataLayout::NCHW, DataLayout::NHWC); |
| 163 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(scores, deltas); |
| 164 | |
| 165 | const int num_anchors = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::CHANNEL)); |
| 166 | const int feat_width = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::WIDTH)); |
| 167 | const int feat_height = scores->dimension(get_data_layout_dimension_index(scores->data_layout(), DataLayoutDimension::HEIGHT)); |
| 168 | const int num_images = scores->dimension(3); |
| 169 | const int total_num_anchors = num_anchors * feat_width * feat_height; |
| 170 | const int values_per_roi = info.values_per_roi(); |
| 171 | |
| 172 | ARM_COMPUTE_RETURN_ERROR_ON(num_images > 1); |
| 173 | |
| 174 | TensorInfo all_anchors_info(anchors->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true)); |
| 175 | ARM_COMPUTE_RETURN_ON_ERROR(CLComputeAllAnchorsKernel::validate(anchors, &all_anchors_info, ComputeAnchorsInfo(feat_width, feat_height, info.spatial_scale()))); |
| 176 | |
| 177 | TensorInfo deltas_permuted_info = deltas->clone()->set_tensor_shape(TensorShape(values_per_roi * num_anchors, feat_width, feat_height)).set_is_resizable(true); |
| 178 | TensorInfo scores_permuted_info = scores->clone()->set_tensor_shape(TensorShape(num_anchors, feat_width, feat_height)).set_is_resizable(true); |
| 179 | if(scores->data_layout() == DataLayout::NHWC) |
| 180 | { |
| 181 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(deltas, &deltas_permuted_info); |
| 182 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(scores, &scores_permuted_info); |
| 183 | } |
| 184 | else |
| 185 | { |
| 186 | ARM_COMPUTE_RETURN_ON_ERROR(CLPermuteKernel::validate(deltas, &deltas_permuted_info, PermutationVector{ 2, 0, 1 })); |
| 187 | ARM_COMPUTE_RETURN_ON_ERROR(CLPermuteKernel::validate(scores, &scores_permuted_info, PermutationVector{ 2, 0, 1 })); |
| 188 | } |
| 189 | |
| 190 | TensorInfo deltas_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true)); |
| 191 | ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(&deltas_permuted_info, &deltas_flattened_info)); |
| 192 | |
| 193 | TensorInfo scores_flattened_info(deltas->clone()->set_tensor_shape(TensorShape(1, total_num_anchors)).set_is_resizable(true)); |
| 194 | TensorInfo proposals_4_roi_values(deltas->clone()->set_tensor_shape(TensorShape(values_per_roi, total_num_anchors)).set_is_resizable(true)); |
| 195 | |
| 196 | ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(&scores_permuted_info, &scores_flattened_info)); |
| 197 | ARM_COMPUTE_RETURN_ON_ERROR(CLBoundingBoxTransformKernel::validate(&all_anchors_info, &proposals_4_roi_values, &deltas_flattened_info, BoundingBoxTransformInfo(info.im_width(), info.im_height(), |
| 198 | 1.f))); |
| 199 | |
| 200 | ARM_COMPUTE_RETURN_ON_ERROR(CLCopyKernel::validate(&proposals_4_roi_values, proposals, PaddingList{ { 0, 1 } })); |
| 201 | ARM_COMPUTE_RETURN_ON_ERROR(CLMemsetKernel::validate(proposals, PixelValue())); |
| 202 | |
| 203 | if(num_valid_proposals->total_size() > 0) |
| 204 | { |
| 205 | ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->num_dimensions() > 1); |
| 206 | ARM_COMPUTE_RETURN_ERROR_ON(num_valid_proposals->dimension(0) > 1); |
| 207 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(num_valid_proposals, 1, DataType::U32); |
| 208 | } |
| 209 | |
| 210 | if(proposals->total_size() > 0) |
| 211 | { |
| 212 | ARM_COMPUTE_RETURN_ERROR_ON(proposals->num_dimensions() > 2); |
| 213 | ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(0) != size_t(values_per_roi) + 1); |
| 214 | ARM_COMPUTE_RETURN_ERROR_ON(proposals->dimension(1) != size_t(total_num_anchors)); |
| 215 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(proposals, deltas); |
| 216 | } |
| 217 | |
| 218 | if(scores_out->total_size() > 0) |
| 219 | { |
| 220 | ARM_COMPUTE_RETURN_ERROR_ON(scores_out->num_dimensions() > 1); |
| 221 | ARM_COMPUTE_RETURN_ERROR_ON(scores_out->dimension(0) != size_t(total_num_anchors)); |
| 222 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(scores_out, scores); |
| 223 | } |
| 224 | |
| 225 | return Status{}; |
| 226 | } |
| 227 | |
| 228 | void CLGenerateProposalsLayer::run_cpp_nms_kernel() |
| 229 | { |
| 230 | // Map inputs |
| 231 | _scores_flattened.map(true); |
| 232 | _all_proposals.map(true); |
| 233 | |
| 234 | // Map outputs |
| 235 | _scores_out->map(CLScheduler::get().queue(), true); |
| 236 | _proposals_4_roi_values.map(CLScheduler::get().queue(), true); |
| 237 | _num_valid_proposals->map(CLScheduler::get().queue(), true); |
| 238 | _keeps_nms_unused.map(true); |
| 239 | _classes_nms_unused.map(true); |
| 240 | |
| 241 | // Run nms |
| 242 | CPPScheduler::get().schedule(&_cpp_nms_kernel, Window::DimX); |
| 243 | |
| 244 | // Unmap outputs |
| 245 | _keeps_nms_unused.unmap(); |
| 246 | _classes_nms_unused.unmap(); |
| 247 | _scores_out->unmap(CLScheduler::get().queue()); |
| 248 | _proposals_4_roi_values.unmap(CLScheduler::get().queue()); |
| 249 | _num_valid_proposals->unmap(CLScheduler::get().queue()); |
| 250 | |
| 251 | // Unmap inputs |
| 252 | _scores_flattened.unmap(); |
| 253 | _all_proposals.unmap(); |
| 254 | } |
| 255 | |
| 256 | void CLGenerateProposalsLayer::run() |
| 257 | { |
| 258 | // Acquire all the temporaries |
Georgios Pinitas | da953f2 | 2019-04-02 17:27:03 +0100 | [diff] [blame] | 259 | MemoryGroupResourceScope scope_mg(_memory_group); |
Manuel Bottini | 5209be5 | 2019-02-13 16:34:56 +0000 | [diff] [blame] | 260 | |
| 261 | // Compute all the anchors |
| 262 | CLScheduler::get().enqueue(_compute_anchors_kernel, false); |
| 263 | |
| 264 | // Transpose and reshape the inputs |
| 265 | if(!_is_nhwc) |
| 266 | { |
| 267 | CLScheduler::get().enqueue(_permute_deltas_kernel, false); |
| 268 | CLScheduler::get().enqueue(_permute_scores_kernel, false); |
| 269 | } |
| 270 | CLScheduler::get().enqueue(_flatten_deltas_kernel, false); |
| 271 | CLScheduler::get().enqueue(_flatten_scores_kernel, false); |
| 272 | |
| 273 | // Build the boxes |
| 274 | CLScheduler::get().enqueue(_bounding_box_kernel, false); |
| 275 | // Non maxima suppression |
| 276 | run_cpp_nms_kernel(); |
| 277 | // Add dummy batch indexes |
| 278 | CLScheduler::get().enqueue(_memset_kernel, true); |
| 279 | CLScheduler::get().enqueue(_padded_copy_kernel, true); |
Manuel Bottini | 5209be5 | 2019-02-13 16:34:56 +0000 | [diff] [blame] | 280 | } |
| 281 | } // namespace arm_compute |