• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017-2020 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/CLHOGMultiDetection.h"
25 
26 #include "arm_compute/core/CL/OpenCL.h"
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/TensorInfo.h"
29 #include "arm_compute/runtime/CL/CLArray.h"
30 #include "arm_compute/runtime/CL/CLScheduler.h"
31 #include "arm_compute/runtime/CL/CLTensor.h"
32 #include "arm_compute/runtime/Scheduler.h"
33 #include "src/core/CL/kernels/CLFillBorderKernel.h"
34 #include "src/core/CL/kernels/CLHOGDescriptorKernel.h"
35 #include "src/core/CL/kernels/CLHOGDetectorKernel.h"
36 #include "src/core/CL/kernels/CLMagnitudePhaseKernel.h"
37 #include "support/MemorySupport.h"
38 
39 using namespace arm_compute;
40 
CLHOGMultiDetection(std::shared_ptr<IMemoryManager> memory_manager)41 CLHOGMultiDetection::CLHOGMultiDetection(std::shared_ptr<IMemoryManager> memory_manager) // NOLINT
42     : _memory_group(std::move(memory_manager)),
43       _gradient_kernel(),
44       _orient_bin_kernel(),
45       _block_norm_kernel(),
46       _hog_detect_kernel(),
47       _non_maxima_kernel(),
48       _hog_space(),
49       _hog_norm_space(),
50       _detection_windows(),
51       _mag(),
52       _phase(),
53       _non_maxima_suppression(false),
54       _num_orient_bin_kernel(0),
55       _num_block_norm_kernel(0),
56       _num_hog_detect_kernel(0)
57 {
58 }
59 
60 CLHOGMultiDetection::~CLHOGMultiDetection() = default;
61 
configure(ICLTensor * input,const ICLMultiHOG * multi_hog,ICLDetectionWindowArray * detection_windows,ICLSize2DArray * detection_window_strides,BorderMode border_mode,uint8_t constant_border_value,float threshold,bool non_maxima_suppression,float min_distance)62 void CLHOGMultiDetection::configure(ICLTensor *input, const ICLMultiHOG *multi_hog, ICLDetectionWindowArray *detection_windows, ICLSize2DArray *detection_window_strides, BorderMode border_mode,
63                                     uint8_t constant_border_value, float threshold, bool non_maxima_suppression, float min_distance)
64 {
65     configure(CLKernelLibrary::get().get_compile_context(), input, multi_hog, detection_windows, detection_window_strides, border_mode, constant_border_value, threshold, non_maxima_suppression,
66               min_distance);
67 }
68 
configure(const CLCompileContext & compile_context,ICLTensor * input,const ICLMultiHOG * multi_hog,ICLDetectionWindowArray * detection_windows,ICLSize2DArray * detection_window_strides,BorderMode border_mode,uint8_t constant_border_value,float threshold,bool non_maxima_suppression,float min_distance)69 void CLHOGMultiDetection::configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLMultiHOG *multi_hog, ICLDetectionWindowArray *detection_windows,
70                                     ICLSize2DArray *detection_window_strides, BorderMode border_mode,
71                                     uint8_t constant_border_value, float threshold, bool non_maxima_suppression, float min_distance)
72 {
73     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
74     ARM_COMPUTE_ERROR_ON_INVALID_MULTI_HOG(multi_hog);
75     ARM_COMPUTE_ERROR_ON(nullptr == detection_windows);
76     ARM_COMPUTE_ERROR_ON(detection_window_strides->num_values() != multi_hog->num_models());
77 
78     const size_t       width      = input->info()->dimension(Window::DimX);
79     const size_t       height     = input->info()->dimension(Window::DimY);
80     const TensorShape &shape_img  = input->info()->tensor_shape();
81     const size_t       num_models = multi_hog->num_models();
82     PhaseType          phase_type = multi_hog->model(0)->info()->phase_type();
83 
84     size_t prev_num_bins     = multi_hog->model(0)->info()->num_bins();
85     Size2D prev_cell_size    = multi_hog->model(0)->info()->cell_size();
86     Size2D prev_block_size   = multi_hog->model(0)->info()->block_size();
87     Size2D prev_block_stride = multi_hog->model(0)->info()->block_stride();
88 
89     /* Check if CLHOGOrientationBinningKernel and CLHOGBlockNormalizationKernel kernels can be skipped for a specific HOG data-object
90      *
91      * 1) CLHOGOrientationBinningKernel and CLHOGBlockNormalizationKernel are skipped if the cell size and the number of bins don't change.
92      *        Since "multi_hog" is sorted,it is enough to check the HOG descriptors at level "ith" and level "(i-1)th
93      * 2) CLHOGBlockNormalizationKernel is skipped if the cell size, the number of bins and block size do not change.
94      *         Since "multi_hog" is sorted,it is enough to check the HOG descriptors at level "ith" and level "(i-1)th
95      *
96      * @note Since the orientation binning and block normalization kernels can be skipped, we need to keep track of the input to process for each kernel
97      *       with "input_orient_bin", "input_hog_detect" and "input_block_norm"
98      */
99     std::vector<size_t> input_orient_bin;
100     std::vector<size_t> input_hog_detect;
101     std::vector<std::pair<size_t, size_t>> input_block_norm;
102 
103     input_orient_bin.push_back(0);
104     input_hog_detect.push_back(0);
105     input_block_norm.emplace_back(0, 0);
106 
107     for(size_t i = 1; i < num_models; ++i)
108     {
109         size_t cur_num_bins     = multi_hog->model(i)->info()->num_bins();
110         Size2D cur_cell_size    = multi_hog->model(i)->info()->cell_size();
111         Size2D cur_block_size   = multi_hog->model(i)->info()->block_size();
112         Size2D cur_block_stride = multi_hog->model(i)->info()->block_stride();
113 
114         if((cur_num_bins != prev_num_bins) || (cur_cell_size.width != prev_cell_size.width) || (cur_cell_size.height != prev_cell_size.height))
115         {
116             prev_num_bins     = cur_num_bins;
117             prev_cell_size    = cur_cell_size;
118             prev_block_size   = cur_block_size;
119             prev_block_stride = cur_block_stride;
120 
121             // Compute orientation binning and block normalization kernels. Update input to process
122             input_orient_bin.push_back(i);
123             input_block_norm.emplace_back(i, input_orient_bin.size() - 1);
124         }
125         else if((cur_block_size.width != prev_block_size.width) || (cur_block_size.height != prev_block_size.height) || (cur_block_stride.width != prev_block_stride.width)
126                 || (cur_block_stride.height != prev_block_stride.height))
127         {
128             prev_block_size   = cur_block_size;
129             prev_block_stride = cur_block_stride;
130 
131             // Compute block normalization kernel. Update input to process
132             input_block_norm.emplace_back(i, input_orient_bin.size() - 1);
133         }
134 
135         // Update input to process for hog detector kernel
136         input_hog_detect.push_back(input_block_norm.size() - 1);
137     }
138 
139     _detection_windows      = detection_windows;
140     _non_maxima_suppression = non_maxima_suppression;
141     _num_orient_bin_kernel  = input_orient_bin.size(); // Number of CLHOGOrientationBinningKernel kernels to compute
142     _num_block_norm_kernel  = input_block_norm.size(); // Number of CLHOGBlockNormalizationKernel kernels to compute
143     _num_hog_detect_kernel  = input_hog_detect.size(); // Number of CLHOGDetector functions to compute
144 
145     _orient_bin_kernel.reserve(_num_orient_bin_kernel);
146     _block_norm_kernel.reserve(_num_block_norm_kernel);
147     _hog_detect_kernel.resize(_num_hog_detect_kernel);
148     _hog_space.resize(_num_orient_bin_kernel);
149     _hog_norm_space.resize(_num_block_norm_kernel);
150 
151     // Allocate tensors for magnitude and phase
152     TensorInfo info_mag(shape_img, Format::S16);
153     _mag.allocator()->init(info_mag);
154 
155     TensorInfo info_phase(shape_img, Format::U8);
156     _phase.allocator()->init(info_phase);
157 
158     // Manage intermediate buffers
159     _memory_group.manage(&_mag);
160     _memory_group.manage(&_phase);
161 
162     // Initialise gradient kernel
163     _gradient_kernel.configure(compile_context, input, &_mag, &_phase, phase_type, border_mode, constant_border_value);
164 
165     // Configure NETensor for the HOG space and orientation binning kernel
166     for(size_t i = 0; i < _num_orient_bin_kernel; ++i)
167     {
168         const size_t idx_multi_hog = input_orient_bin[i];
169 
170         // Get the corresponding cell size and number of bins
171         const Size2D &cell     = multi_hog->model(idx_multi_hog)->info()->cell_size();
172         const size_t  num_bins = multi_hog->model(idx_multi_hog)->info()->num_bins();
173 
174         // Calculate number of cells along the x and y directions for the hog_space
175         const size_t num_cells_x = width / cell.width;
176         const size_t num_cells_y = height / cell.height;
177 
178         // TensorShape of hog space
179         TensorShape shape_hog_space = input->info()->tensor_shape();
180         shape_hog_space.set(Window::DimX, num_cells_x);
181         shape_hog_space.set(Window::DimY, num_cells_y);
182 
183         // Allocate HOG space
184         TensorInfo info_space(shape_hog_space, num_bins, DataType::F32);
185         _hog_space[i].allocator()->init(info_space);
186 
187         // Manage intermediate buffers
188         _memory_group.manage(&_hog_space[i]);
189 
190         // Initialise orientation binning kernel
191         _orient_bin_kernel.emplace_back(support::cpp14::make_unique<CLHOGOrientationBinningKernel>());
192         _orient_bin_kernel.back()->configure(compile_context, &_mag, &_phase, &_hog_space[i], multi_hog->model(idx_multi_hog)->info());
193     }
194 
195     // Allocate intermediate tensors
196     _mag.allocator()->allocate();
197     _phase.allocator()->allocate();
198 
199     // Configure CLTensor for the normalized HOG space and block normalization kernel
200     for(size_t i = 0; i < _num_block_norm_kernel; ++i)
201     {
202         const size_t idx_multi_hog  = input_block_norm[i].first;
203         const size_t idx_orient_bin = input_block_norm[i].second;
204 
205         // Allocate normalized HOG space
206         TensorInfo tensor_info(*(multi_hog->model(idx_multi_hog)->info()), width, height);
207         _hog_norm_space[i].allocator()->init(tensor_info);
208 
209         // Manage intermediate buffers
210         _memory_group.manage(&_hog_norm_space[i]);
211 
212         // Initialize block normalization kernel
213         _block_norm_kernel.emplace_back(support::cpp14::make_unique<CLHOGBlockNormalizationKernel>());
214         _block_norm_kernel.back()->configure(compile_context, &_hog_space[idx_orient_bin], &_hog_norm_space[i], multi_hog->model(idx_multi_hog)->info());
215     }
216 
217     // Allocate intermediate tensors
218     for(size_t i = 0; i < _num_orient_bin_kernel; ++i)
219     {
220         _hog_space[i].allocator()->allocate();
221     }
222 
223     detection_window_strides->map(CLScheduler::get().queue(), true);
224 
225     // Configure HOG detector kernel
226     for(size_t i = 0; i < _num_hog_detect_kernel; ++i)
227     {
228         const size_t idx_block_norm = input_hog_detect[i];
229 
230         _hog_detect_kernel[i].configure(compile_context, &_hog_norm_space[idx_block_norm], multi_hog->cl_model(i), detection_windows, detection_window_strides->at(i), threshold, i);
231     }
232 
233     detection_window_strides->unmap(CLScheduler::get().queue());
234 
235     // Configure non maxima suppression kernel
236     _non_maxima_kernel.configure(_detection_windows, min_distance);
237 
238     // Allocate intermediate tensors
239     for(size_t i = 0; i < _num_block_norm_kernel; ++i)
240     {
241         _hog_norm_space[i].allocator()->allocate();
242     }
243 }
244 
run()245 void CLHOGMultiDetection::run()
246 {
247     ARM_COMPUTE_ERROR_ON_MSG(_detection_windows == nullptr, "Unconfigured function");
248 
249     MemoryGroupResourceScope scope_mg(_memory_group);
250 
251     // Reset detection window
252     _detection_windows->clear();
253 
254     // Run gradient
255     _gradient_kernel.run();
256 
257     // Run orientation binning kernel
258     for(size_t i = 0; i < _num_orient_bin_kernel; ++i)
259     {
260         CLScheduler::get().enqueue(*_orient_bin_kernel[i], false);
261     }
262 
263     // Run block normalization kernel
264     for(size_t i = 0; i < _num_block_norm_kernel; ++i)
265     {
266         CLScheduler::get().enqueue(*_block_norm_kernel[i], false);
267     }
268 
269     // Run HOG detector kernel
270     for(size_t i = 0; i < _num_hog_detect_kernel; ++i)
271     {
272         _hog_detect_kernel[i].run();
273     }
274 
275     // Run non-maxima suppression kernel if enabled
276     if(_non_maxima_suppression)
277     {
278         // Map detection windows array before computing non maxima suppression
279         _detection_windows->map(CLScheduler::get().queue(), true);
280         Scheduler::get().schedule(&_non_maxima_kernel, Window::DimY);
281         _detection_windows->unmap(CLScheduler::get().queue());
282     }
283 }
284