arm_compute v17.04
[platform/upstream/armcl.git] / src / runtime / NEON / functions / NEOpticalFlow.cpp
1 /*
2  * Copyright (c) 2016, 2017 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/NEOpticalFlow.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/NEON/kernels/NELKTrackerKernel.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "arm_compute/core/Window.h"
32 #include "arm_compute/runtime/NEON/NEScheduler.h"
33 #include "arm_compute/runtime/NEON/functions/NEScharr3x3.h"
34 #include "arm_compute/runtime/Pyramid.h"
35 #include "arm_compute/runtime/Tensor.h"
36 #include "arm_compute/runtime/TensorAllocator.h"
37
38 using namespace arm_compute;
39
40 NEOpticalFlow::NEOpticalFlow()
41     : _func_scharr(), _kernel_tracker(), _scharr_gx(), _scharr_gy(), _new_points(nullptr), _new_points_estimates(nullptr), _old_points(nullptr), _new_points_internal(), _old_points_internal(),
42       _num_levels(0)
43 {
44 }
45
46 void NEOpticalFlow::configure(const Pyramid *old_pyramid, const Pyramid *new_pyramid, const IKeyPointArray *old_points, const IKeyPointArray *new_points_estimates,
47                               IKeyPointArray *new_points, Termination termination, float epsilon, unsigned int num_iterations, size_t window_dimension,
48                               bool use_initial_estimate, BorderMode border_mode, uint8_t constant_border_value)
49 {
50     ARM_COMPUTE_ERROR_ON(nullptr == old_pyramid);
51     ARM_COMPUTE_ERROR_ON(nullptr == new_pyramid);
52     ARM_COMPUTE_ERROR_ON(nullptr == old_points);
53     ARM_COMPUTE_ERROR_ON(nullptr == new_points_estimates);
54     ARM_COMPUTE_ERROR_ON(nullptr == new_points);
55     ARM_COMPUTE_ERROR_ON(old_pyramid->info()->num_levels() != new_pyramid->info()->num_levels());
56     ARM_COMPUTE_ERROR_ON(0 == old_pyramid->info()->num_levels());
57     ARM_COMPUTE_ERROR_ON(old_pyramid->info()->width() != new_pyramid->info()->width());
58     ARM_COMPUTE_ERROR_ON(old_pyramid->info()->height() != new_pyramid->info()->height());
59     ARM_COMPUTE_ERROR_ON(use_initial_estimate && old_points->num_values() != new_points_estimates->num_values());
60
61     _num_levels           = old_pyramid->info()->num_levels();
62     _old_points           = old_points;
63     _new_points           = new_points;
64     _new_points_estimates = new_points_estimates;
65
66     const float pyr_scale = old_pyramid->info()->scale();
67
68     _func_scharr    = arm_compute::cpp14::make_unique<NEScharr3x3[]>(_num_levels);
69     _kernel_tracker = arm_compute::cpp14::make_unique<NELKTrackerKernel[]>(_num_levels);
70     _scharr_gx      = arm_compute::cpp14::make_unique<Tensor[]>(_num_levels);
71     _scharr_gy      = arm_compute::cpp14::make_unique<Tensor[]>(_num_levels);
72
73     _old_points_internal = LKInternalKeypointArray(old_points->num_values());
74     _new_points_internal = LKInternalKeypointArray(old_points->num_values());
75     _new_points->resize(old_points->num_values());
76
77     const int border_offset = BorderMode::UNDEFINED == border_mode ? 1 : 0;
78
79     for(size_t i = 0; i < _num_levels; ++i)
80     {
81         /* Get images from the ith level of old and right pyramid */
82         IImage *old_ith_input = old_pyramid->get_pyramid_level(i);
83         IImage *new_ith_input = new_pyramid->get_pyramid_level(i);
84
85         /* Get width and height of images */
86         const unsigned int width_ith  = old_ith_input->info()->dimension(0);
87         const unsigned int height_ith = new_ith_input->info()->dimension(1);
88
89         TensorInfo tensor_info(TensorShape(width_ith, height_ith), Format::S16);
90
91         _scharr_gx[i].allocator()->init(tensor_info);
92         _scharr_gy[i].allocator()->init(tensor_info);
93
94         /* Init Scharr kernel */
95         _func_scharr[i].configure(old_ith_input, _scharr_gx.get() + i, _scharr_gy.get() + i, border_mode, constant_border_value);
96
97         /* Init Lucas-Kanade kernel */
98         _kernel_tracker[i].configure(old_ith_input, new_ith_input, _scharr_gx.get() + i, _scharr_gy.get() + i,
99                                      old_points, new_points_estimates, new_points,
100                                      &_old_points_internal, &_new_points_internal,
101                                      termination, use_initial_estimate, epsilon, num_iterations, window_dimension,
102                                      i, _num_levels, pyr_scale, border_offset);
103
104         _scharr_gx[i].allocator()->allocate();
105         _scharr_gy[i].allocator()->allocate();
106     }
107 }
108
109 void NEOpticalFlow::run()
110 {
111     ARM_COMPUTE_ERROR_ON_MSG(_num_levels == 0, "Unconfigured function");
112
113     for(unsigned int level = _num_levels; level > 0; --level)
114     {
115         /* Run Scharr kernel */
116         _func_scharr[level - 1].run();
117
118         /* Run Lucas-Kanade kernel */
119         NEScheduler::get().multithread(_kernel_tracker.get() + level - 1, Window::DimX);
120     }
121 }