1 /*
2 * Copyright (c) 2019-2022 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/cpu/kernels/genproposals/generic/neon/impl.h"
25 namespace arm_compute
26 {
27 class ITensor;
28 class Window;
29 namespace cpu
30 {
31 template <typename T>
compute_all_anchors(const ITensor * anchors,ITensor * all_anchors,ComputeAnchorsInfo anchors_info,const Window & window)32 void compute_all_anchors(const ITensor *anchors, ITensor *all_anchors, ComputeAnchorsInfo anchors_info, const Window &window)
33 {
34 Iterator all_anchors_it(all_anchors, window);
35 Iterator anchors_it(all_anchors, window);
36
37 const size_t num_anchors = anchors->info()->dimension(1);
38 const T stride = 1.f / anchors_info.spatial_scale();
39 const size_t feat_width = anchors_info.feat_width();
40
41 execute_window_loop(window, [&](const Coordinates & id)
42 {
43 const size_t anchor_offset = id.y() % num_anchors;
44
45 const auto out_anchor_ptr = reinterpret_cast<T *>(all_anchors_it.ptr());
46 const auto anchor_ptr = reinterpret_cast<T *>(anchors->ptr_to_element(Coordinates(0, anchor_offset)));
47
48 const size_t shift_idy = id.y() / num_anchors;
49 const T shiftx = (shift_idy % feat_width) * stride;
50 const T shifty = (shift_idy / feat_width) * stride;
51
52 *out_anchor_ptr = *anchor_ptr + shiftx;
53 *(out_anchor_ptr + 1) = *(1 + anchor_ptr) + shifty;
54 *(out_anchor_ptr + 2) = *(2 + anchor_ptr) + shiftx;
55 *(out_anchor_ptr + 3) = *(3 + anchor_ptr) + shifty;
56 },
57 all_anchors_it);
58 }
59
60 template void compute_all_anchors<float>(const ITensor *anchors, ITensor *all_anchors, ComputeAnchorsInfo anchors_info, const Window &window);
61 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
62 template void compute_all_anchors<float16_t>(const ITensor *anchors, ITensor *all_anchors, ComputeAnchorsInfo anchors_info, const Window &window);
63 #endif //defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
64
compute_all_anchors_qasymm16(const ITensor * anchors,ITensor * all_anchors,ComputeAnchorsInfo anchors_info,const Window & window)65 void compute_all_anchors_qasymm16(const ITensor *anchors, ITensor *all_anchors, ComputeAnchorsInfo anchors_info, const Window &window)
66 {
67 Iterator all_anchors_it(all_anchors, window);
68 Iterator anchors_it(all_anchors, window);
69
70 const size_t num_anchors = anchors->info()->dimension(1);
71 const float stride = 1.f / anchors_info.spatial_scale();
72 const size_t feat_width = anchors_info.feat_width();
73
74 const UniformQuantizationInfo qinfo = anchors->info()->quantization_info().uniform();
75
76 execute_window_loop(window, [&](const Coordinates & id)
77 {
78 const size_t anchor_offset = id.y() % num_anchors;
79
80 const auto out_anchor_ptr = reinterpret_cast<int16_t *>(all_anchors_it.ptr());
81 const auto anchor_ptr = reinterpret_cast<int16_t *>(anchors->ptr_to_element(Coordinates(0, anchor_offset)));
82
83 const size_t shift_idy = id.y() / num_anchors;
84 const float shiftx = (shift_idy % feat_width) * stride;
85 const float shifty = (shift_idy / feat_width) * stride;
86
87 const float new_anchor_x1 = dequantize_qsymm16(*anchor_ptr, qinfo.scale) + shiftx;
88 const float new_anchor_y1 = dequantize_qsymm16(*(1 + anchor_ptr), qinfo.scale) + shifty;
89 const float new_anchor_x2 = dequantize_qsymm16(*(2 + anchor_ptr), qinfo.scale) + shiftx;
90 const float new_anchor_y2 = dequantize_qsymm16(*(3 + anchor_ptr), qinfo.scale) + shifty;
91
92 *out_anchor_ptr = quantize_qsymm16(new_anchor_x1, qinfo.scale);
93 *(out_anchor_ptr + 1) = quantize_qsymm16(new_anchor_y1, qinfo.scale);
94 *(out_anchor_ptr + 2) = quantize_qsymm16(new_anchor_x2, qinfo.scale);
95 *(out_anchor_ptr + 3) = quantize_qsymm16(new_anchor_y2, qinfo.scale);
96 },
97 all_anchors_it);
98 }
99 } // namespace cpu
100 } // namespace arm_compute
101