1 /*
2 * Copyright (c) 2019-2020 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/NEBoundingBoxTransformKernel.h"
25
26 #include "arm_compute/core/Helpers.h"
27 #include "arm_compute/core/TensorInfo.h"
28 #include "arm_compute/core/Utils.h"
29 #include "arm_compute/core/Window.h"
30 #include "src/core/AccessWindowStatic.h"
31 #include "src/core/CPP/Validate.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34
35 #include <arm_neon.h>
36
37 namespace arm_compute
38 {
39 namespace
40 {
validate_arguments(const ITensorInfo * boxes,const ITensorInfo * pred_boxes,const ITensorInfo * deltas,const BoundingBoxTransformInfo & info)41 Status validate_arguments(const ITensorInfo *boxes, const ITensorInfo *pred_boxes, const ITensorInfo *deltas, const BoundingBoxTransformInfo &info)
42 {
43 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(boxes, pred_boxes, deltas);
44 ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(boxes);
45 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(boxes, DataType::QASYMM16, DataType::F32, DataType::F16);
46 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(deltas, DataType::QASYMM8, DataType::F32, DataType::F16);
47 ARM_COMPUTE_RETURN_ERROR_ON(deltas->tensor_shape()[1] != boxes->tensor_shape()[1]);
48 ARM_COMPUTE_RETURN_ERROR_ON(deltas->tensor_shape()[0] % 4 != 0);
49 ARM_COMPUTE_RETURN_ERROR_ON(boxes->tensor_shape()[0] != 4);
50 ARM_COMPUTE_RETURN_ERROR_ON(deltas->num_dimensions() > 2);
51 ARM_COMPUTE_RETURN_ERROR_ON(boxes->num_dimensions() > 2);
52 ARM_COMPUTE_RETURN_ERROR_ON(info.scale() <= 0);
53
54 if(boxes->data_type() == DataType::QASYMM16)
55 {
56 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(deltas, 1, DataType::QASYMM8);
57 const UniformQuantizationInfo deltas_qinfo = deltas->quantization_info().uniform();
58 ARM_COMPUTE_RETURN_ERROR_ON(deltas_qinfo.scale != 0.125f);
59 ARM_COMPUTE_RETURN_ERROR_ON(deltas_qinfo.offset != 0);
60 }
61 else
62 {
63 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(boxes, deltas);
64 }
65
66 if(pred_boxes->total_size() > 0)
67 {
68 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(pred_boxes->tensor_shape(), deltas->tensor_shape());
69 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(pred_boxes, deltas);
70 ARM_COMPUTE_RETURN_ERROR_ON(pred_boxes->num_dimensions() > 2);
71 if(pred_boxes->data_type() == DataType::QASYMM16)
72 {
73 const UniformQuantizationInfo pred_qinfo = pred_boxes->quantization_info().uniform();
74 ARM_COMPUTE_RETURN_ERROR_ON(pred_qinfo.scale != 0.125f);
75 ARM_COMPUTE_RETURN_ERROR_ON(pred_qinfo.offset != 0);
76 }
77 }
78
79 return Status{};
80 }
81 } // namespace
82
NEBoundingBoxTransformKernel()83 NEBoundingBoxTransformKernel::NEBoundingBoxTransformKernel()
84 : _boxes(nullptr), _pred_boxes(nullptr), _deltas(nullptr), _bbinfo(0, 0, 0)
85 {
86 }
87
configure(const ITensor * boxes,ITensor * pred_boxes,const ITensor * deltas,const BoundingBoxTransformInfo & info)88 void NEBoundingBoxTransformKernel::configure(const ITensor *boxes, ITensor *pred_boxes, const ITensor *deltas, const BoundingBoxTransformInfo &info)
89 {
90 ARM_COMPUTE_ERROR_ON_NULLPTR(boxes, pred_boxes, deltas);
91 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(boxes->info(), pred_boxes->info(), deltas->info(), info));
92
93 // Configure kernel window
94 auto_init_if_empty(*pred_boxes->info(), deltas->info()->clone()->set_data_type(boxes->info()->data_type()).set_quantization_info(boxes->info()->quantization_info()));
95
96 // Set instance variables
97 _boxes = boxes;
98 _pred_boxes = pred_boxes;
99 _deltas = deltas;
100 _bbinfo = info;
101
102 const unsigned int num_boxes = boxes->info()->dimension(1);
103 Window win = calculate_max_window(*pred_boxes->info(), Steps());
104 Coordinates coord;
105 coord.set_num_dimensions(pred_boxes->info()->num_dimensions());
106 pred_boxes->info()->set_valid_region(ValidRegion(coord, pred_boxes->info()->tensor_shape()));
107 win.set(Window::DimX, Window::Dimension(0, 1u));
108 win.set(Window::DimY, Window::Dimension(0, num_boxes));
109
110 INEKernel::configure(win);
111 }
112
validate(const ITensorInfo * boxes,const ITensorInfo * pred_boxes,const ITensorInfo * deltas,const BoundingBoxTransformInfo & info)113 Status NEBoundingBoxTransformKernel::validate(const ITensorInfo *boxes, const ITensorInfo *pred_boxes, const ITensorInfo *deltas, const BoundingBoxTransformInfo &info)
114 {
115 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(boxes, pred_boxes, deltas, info));
116 return Status{};
117 }
118
119 template <>
internal_run(const Window & window)120 void NEBoundingBoxTransformKernel::internal_run<uint16_t>(const Window &window)
121 {
122 const size_t num_classes = _deltas->info()->tensor_shape()[0] >> 2;
123 const size_t deltas_width = _deltas->info()->tensor_shape()[0];
124 const int img_h = std::floor(_bbinfo.img_height() / _bbinfo.scale() + 0.5f);
125 const int img_w = std::floor(_bbinfo.img_width() / _bbinfo.scale() + 0.5f);
126
127 const auto scale_after = (_bbinfo.apply_scale() ? _bbinfo.scale() : 1.f);
128 const auto scale_before = _bbinfo.scale();
129 const auto offset = (_bbinfo.correct_transform_coords() ? 1.f : 0.f);
130
131 auto pred_ptr = reinterpret_cast<uint16_t *>(_pred_boxes->buffer() + _pred_boxes->info()->offset_first_element_in_bytes());
132 auto delta_ptr = reinterpret_cast<uint8_t *>(_deltas->buffer() + _deltas->info()->offset_first_element_in_bytes());
133
134 const auto boxes_qinfo = _boxes->info()->quantization_info().uniform();
135 const auto deltas_qinfo = _deltas->info()->quantization_info().uniform();
136 const auto pred_qinfo = _pred_boxes->info()->quantization_info().uniform();
137
138 Iterator box_it(_boxes, window);
139 execute_window_loop(window, [&](const Coordinates & id)
140 {
141 const auto ptr = reinterpret_cast<uint16_t *>(box_it.ptr());
142 const auto b0 = dequantize_qasymm16(*ptr, boxes_qinfo);
143 const auto b1 = dequantize_qasymm16(*(ptr + 1), boxes_qinfo);
144 const auto b2 = dequantize_qasymm16(*(ptr + 2), boxes_qinfo);
145 const auto b3 = dequantize_qasymm16(*(ptr + 3), boxes_qinfo);
146 const float width = (b2 / scale_before) - (b0 / scale_before) + 1.f;
147 const float height = (b3 / scale_before) - (b1 / scale_before) + 1.f;
148 const float ctr_x = (b0 / scale_before) + 0.5f * width;
149 const float ctr_y = (b1 / scale_before) + 0.5f * height;
150 for(size_t j = 0; j < num_classes; ++j)
151 {
152 // Extract deltas
153 const size_t delta_id = id.y() * deltas_width + 4u * j;
154 const float dx = dequantize_qasymm8(delta_ptr[delta_id], deltas_qinfo) / _bbinfo.weights()[0];
155 const float dy = dequantize_qasymm8(delta_ptr[delta_id + 1], deltas_qinfo) / _bbinfo.weights()[1];
156 float dw = dequantize_qasymm8(delta_ptr[delta_id + 2], deltas_qinfo) / _bbinfo.weights()[2];
157 float dh = dequantize_qasymm8(delta_ptr[delta_id + 3], deltas_qinfo) / _bbinfo.weights()[3];
158 // Clip dw and dh
159 dw = std::min(dw, _bbinfo.bbox_xform_clip());
160 dh = std::min(dh, _bbinfo.bbox_xform_clip());
161 // Determine the predictions
162 const float pred_ctr_x = dx * width + ctr_x;
163 const float pred_ctr_y = dy * height + ctr_y;
164 const float pred_w = std::exp(dw) * width;
165 const float pred_h = std::exp(dh) * height;
166 // Store the prediction into the output tensor
167 pred_ptr[delta_id] = quantize_qasymm16(scale_after * utility::clamp<float>(pred_ctr_x - 0.5f * pred_w, 0.f, img_w - 1.f), pred_qinfo);
168 pred_ptr[delta_id + 1] = quantize_qasymm16(scale_after * utility::clamp<float>(pred_ctr_y - 0.5f * pred_h, 0.f, img_h - 1.f), pred_qinfo);
169 pred_ptr[delta_id + 2] = quantize_qasymm16(scale_after * utility::clamp<float>(pred_ctr_x + 0.5f * pred_w - offset, 0.f, img_w - 1.f), pred_qinfo);
170 pred_ptr[delta_id + 3] = quantize_qasymm16(scale_after * utility::clamp<float>(pred_ctr_y + 0.5f * pred_h - offset, 0.f, img_h - 1.f), pred_qinfo);
171 }
172 },
173 box_it);
174 }
175
176 template <typename T>
internal_run(const Window & window)177 void NEBoundingBoxTransformKernel::internal_run(const Window &window)
178 {
179 const size_t num_classes = _deltas->info()->tensor_shape()[0] >> 2;
180 const size_t deltas_width = _deltas->info()->tensor_shape()[0];
181 const int img_h = std::floor(_bbinfo.img_height() / _bbinfo.scale() + 0.5f);
182 const int img_w = std::floor(_bbinfo.img_width() / _bbinfo.scale() + 0.5f);
183
184 const auto scale_after = (_bbinfo.apply_scale() ? T(_bbinfo.scale()) : T(1));
185 const auto scale_before = T(_bbinfo.scale());
186 ARM_COMPUTE_ERROR_ON(scale_before <= 0);
187 const auto offset = (_bbinfo.correct_transform_coords() ? T(1.f) : T(0.f));
188
189 auto pred_ptr = reinterpret_cast<T *>(_pred_boxes->buffer() + _pred_boxes->info()->offset_first_element_in_bytes());
190 auto delta_ptr = reinterpret_cast<T *>(_deltas->buffer() + _deltas->info()->offset_first_element_in_bytes());
191
192 Iterator box_it(_boxes, window);
193 execute_window_loop(window, [&](const Coordinates & id)
194 {
195 const auto ptr = reinterpret_cast<T *>(box_it.ptr());
196 const auto b0 = *ptr;
197 const auto b1 = *(ptr + 1);
198 const auto b2 = *(ptr + 2);
199 const auto b3 = *(ptr + 3);
200 const T width = (b2 / scale_before) - (b0 / scale_before) + T(1.f);
201 const T height = (b3 / scale_before) - (b1 / scale_before) + T(1.f);
202 const T ctr_x = (b0 / scale_before) + T(0.5f) * width;
203 const T ctr_y = (b1 / scale_before) + T(0.5f) * height;
204 for(size_t j = 0; j < num_classes; ++j)
205 {
206 // Extract deltas
207 const size_t delta_id = id.y() * deltas_width + 4u * j;
208 const T dx = delta_ptr[delta_id] / T(_bbinfo.weights()[0]);
209 const T dy = delta_ptr[delta_id + 1] / T(_bbinfo.weights()[1]);
210 T dw = delta_ptr[delta_id + 2] / T(_bbinfo.weights()[2]);
211 T dh = delta_ptr[delta_id + 3] / T(_bbinfo.weights()[3]);
212 // Clip dw and dh
213 dw = std::min(dw, T(_bbinfo.bbox_xform_clip()));
214 dh = std::min(dh, T(_bbinfo.bbox_xform_clip()));
215 // Determine the predictions
216 const T pred_ctr_x = dx * width + ctr_x;
217 const T pred_ctr_y = dy * height + ctr_y;
218 const T pred_w = std::exp(dw) * width;
219 const T pred_h = std::exp(dh) * height;
220 // Store the prediction into the output tensor
221 pred_ptr[delta_id] = scale_after * utility::clamp<T>(pred_ctr_x - T(0.5f) * pred_w, T(0), T(img_w - 1));
222 pred_ptr[delta_id + 1] = scale_after * utility::clamp<T>(pred_ctr_y - T(0.5f) * pred_h, T(0), T(img_h - 1));
223 pred_ptr[delta_id + 2] = scale_after * utility::clamp<T>(pred_ctr_x + T(0.5f) * pred_w - offset, T(0), T(img_w - 1));
224 pred_ptr[delta_id + 3] = scale_after * utility::clamp<T>(pred_ctr_y + T(0.5f) * pred_h - offset, T(0), T(img_h - 1));
225 }
226 },
227 box_it);
228 }
229
run(const Window & window,const ThreadInfo & info)230 void NEBoundingBoxTransformKernel::run(const Window &window, const ThreadInfo &info)
231 {
232 ARM_COMPUTE_UNUSED(info);
233 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
234 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
235 switch(_boxes->info()->data_type())
236 {
237 case DataType::F32:
238 {
239 internal_run<float>(window);
240 break;
241 }
242 case DataType::QASYMM16:
243 {
244 internal_run<uint16_t>(window);
245 break;
246 }
247 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
248 case DataType::F16:
249 {
250 internal_run<float16_t>(window);
251 break;
252 }
253 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
254 default:
255 {
256 ARM_COMPUTE_ERROR("Data type not supported");
257 }
258 }
259 }
260 } // namespace arm_compute
261