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 "src/core/NEON/kernels/NEROIPoolingLayerKernel.h"
25
26 #include "arm_compute/core/TensorInfo.h"
27 #include "arm_compute/core/Validate.h"
28 #include "arm_compute/core/Window.h"
29 #include "src/core/CPP/Validate.h"
30 #include "src/core/helpers/AutoConfiguration.h"
31 #include "src/core/helpers/WindowHelpers.h"
32 #include "support/ToolchainSupport.h"
33
34 #include <cfloat>
35
36 namespace arm_compute
37 {
NEROIPoolingLayerKernel()38 NEROIPoolingLayerKernel::NEROIPoolingLayerKernel()
39 : _input(nullptr), _rois(nullptr), _output(nullptr), _pool_info(0, 0, 0.f)
40 {
41 }
42
configure(const ITensor * input,const ITensor * rois,ITensor * output,const ROIPoolingLayerInfo & pool_info)43 void NEROIPoolingLayerKernel::configure(const ITensor *input, const ITensor *rois, ITensor *output, const ROIPoolingLayerInfo &pool_info)
44 {
45 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, rois);
46
47 //Validate arguments
48 ARM_COMPUTE_ERROR_ON_NULLPTR(input->info(), rois->info(), output->info());
49 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(rois, 1, DataType::U16);
50 ARM_COMPUTE_ERROR_ON(rois->info()->dimension(0) != 5);
51 ARM_COMPUTE_ERROR_ON(rois->info()->num_dimensions() > 2);
52 ARM_COMPUTE_ERROR_ON_CPU_F16_UNSUPPORTED(input);
53 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
54 ARM_COMPUTE_ERROR_ON((pool_info.pooled_width() == 0) || (pool_info.pooled_height() == 0));
55
56 if(output->info()->total_size() != 0)
57 {
58 ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
59 ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pool_info.pooled_width()) || (output->info()->dimension(1) != pool_info.pooled_height()));
60 ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2));
61 ARM_COMPUTE_ERROR_ON(rois->info()->dimension(1) != output->info()->dimension(3));
62 }
63
64 // Output auto initialization if not yet initialized
65 TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), input->info()->dimension(2), rois->info()->dimension(1));
66 auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type());
67
68 ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
69 ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pool_info.pooled_width()) || (output->info()->dimension(1) != pool_info.pooled_height()));
70
71 // Set instance variables
72 _input = input;
73 _rois = rois;
74 _output = output;
75 _pool_info = pool_info;
76
77 // Configure kernel window
78 Window window;
79 window.set(Window::DimX, Window::Dimension(0, rois->info()->dimension(1)));
80 window.set(Window::DimY, Window::Dimension(0, 1));
81
82 Coordinates coord;
83 coord.set_num_dimensions(output->info()->num_dimensions());
84 output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
85
86 INEKernel::configure(window);
87 }
88
run(const Window & window,const ThreadInfo & info)89 void NEROIPoolingLayerKernel::run(const Window &window, const ThreadInfo &info)
90 {
91 ARM_COMPUTE_UNUSED(info);
92 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
93 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
94
95 const size_t values_per_roi = _rois->info()->dimension(0);
96
97 const int roi_list_start = window.x().start();
98 const int roi_list_end = window.x().end();
99 const int width = _input->info()->dimension(Window::DimX);
100 const int height = _input->info()->dimension(Window::DimY);
101 const int fms = _input->info()->dimension(Window::DimZ);
102 const int pooled_w = _pool_info.pooled_width();
103 const int pooled_h = _pool_info.pooled_height();
104 const float spatial_scale = _pool_info.spatial_scale();
105
106 const auto *rois_ptr = reinterpret_cast<const uint16_t *>(_rois->buffer());
107
108 for(int roi_indx = roi_list_start; roi_indx < roi_list_end; ++roi_indx)
109 {
110 const unsigned int roi_batch = rois_ptr[values_per_roi * roi_indx];
111 const auto x1 = rois_ptr[values_per_roi * roi_indx + 1];
112 const auto y1 = rois_ptr[values_per_roi * roi_indx + 2];
113 const auto x2 = rois_ptr[values_per_roi * roi_indx + 3];
114 const auto y2 = rois_ptr[values_per_roi * roi_indx + 4];
115
116 // Scale ROI
117 const int roi_anchor_x = support::cpp11::round(x1 * spatial_scale);
118 const int roi_anchor_y = support::cpp11::round(y1 * spatial_scale);
119 const int roi_width = std::max(support::cpp11::round((x2 - x1) * spatial_scale), 1.f);
120 const int roi_height = std::max(support::cpp11::round((y2 - y1) * spatial_scale), 1.f);
121
122 // Iterate through all feature maps
123 for(int fm = 0; fm < fms; ++fm)
124 {
125 // Iterate through all output pixels
126 for(int py = 0; py < pooled_h; ++py)
127 {
128 for(int px = 0; px < pooled_w; ++px)
129 {
130 auto region_start_x = static_cast<int>(std::floor((static_cast<float>(px) / pooled_w) * roi_width));
131 auto region_end_x = static_cast<int>(std::floor((static_cast<float>(px + 1) / pooled_w) * roi_width));
132 auto region_start_y = static_cast<int>(std::floor((static_cast<float>(py) / pooled_h) * roi_height));
133 auto region_end_y = static_cast<int>(std::floor((static_cast<float>(py + 1) / pooled_h) * roi_height));
134
135 region_start_x = std::min(std::max(region_start_x + roi_anchor_x, 0), width);
136 region_end_x = std::min(std::max(region_end_x + roi_anchor_x, 0), width);
137 region_start_y = std::min(std::max(region_start_y + roi_anchor_y, 0), height);
138 region_end_y = std::min(std::max(region_end_y + roi_anchor_y, 0), height);
139
140 // Iterate through the pooling region
141 if((region_end_x <= region_start_x) || (region_end_y <= region_start_y))
142 {
143 *reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(px, py, fm, roi_indx))) = 0;
144 }
145 else
146 {
147 float curr_max = -FLT_MAX;
148 for(int j = region_start_y; j < region_end_y; ++j)
149 {
150 for(int i = region_start_x; i < region_end_x; ++i)
151 {
152 const auto val = *reinterpret_cast<const float *>(_input->ptr_to_element(Coordinates(i, j, fm, roi_batch)));
153 curr_max = std::max(val, curr_max);
154 }
155 }
156 *reinterpret_cast<float *>(_output->ptr_to_element(Coordinates(px, py, fm, roi_indx))) = curr_max;
157 }
158 }
159 }
160 }
161 }
162 }
163 } // namespace arm_compute