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/NEROIAlignLayerKernel.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 "arm_compute/core/utils/misc/ShapeCalculator.h"
31 #include "arm_compute/core/utils/misc/Utility.h"
32 #include "src/core/AccessWindowStatic.h"
33 #include "src/core/CPP/Validate.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/WindowHelpers.h"
36
37 #include <arm_neon.h>
38
39 using namespace arm_compute::misc::shape_calculator;
40
41 namespace arm_compute
42 {
43 namespace
44 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * rois,ITensorInfo * output,const ROIPoolingLayerInfo & pool_info)45 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info)
46 {
47 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, rois, output);
48 ARM_COMPUTE_RETURN_ERROR_ON(rois->dimension(0) != 5);
49 ARM_COMPUTE_RETURN_ERROR_ON(rois->num_dimensions() > 2);
50 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F32, DataType::F16);
51 ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC, DataLayout::NCHW);
52 ARM_COMPUTE_RETURN_ERROR_ON((pool_info.pooled_width() == 0) || (pool_info.pooled_height() == 0));
53 ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
54
55 if(output->total_size() != 0)
56 {
57 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
58 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
59 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(compute_roi_align_shape(*input, *rois, pool_info), output->tensor_shape());
60 }
61
62 if(input->data_type() == DataType::QASYMM8 || input->data_type() == DataType::QASYMM8_SIGNED)
63 {
64 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(rois, 1, DataType::QASYMM16);
65
66 const UniformQuantizationInfo rois_qinfo = rois->quantization_info().uniform();
67 ARM_COMPUTE_RETURN_ERROR_ON(rois_qinfo.scale != 0.125f);
68 ARM_COMPUTE_RETURN_ERROR_ON(rois_qinfo.offset != 0);
69 }
70 else
71 {
72 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, rois);
73 }
74
75 return Status{};
76 }
77 } // namespace
78
NEROIAlignLayerKernel()79 NEROIAlignLayerKernel::NEROIAlignLayerKernel()
80 : _input(nullptr), _output(nullptr), _rois(nullptr), _pool_info(0, 0, 0.f)
81 {
82 }
83
configure(const ITensor * input,const ITensor * rois,ITensor * output,const ROIPoolingLayerInfo & pool_info)84 void NEROIAlignLayerKernel::configure(const ITensor *input, const ITensor *rois, ITensor *output, const ROIPoolingLayerInfo &pool_info)
85 {
86 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, rois);
87 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), rois->info(), output->info(), pool_info));
88 // Output auto inizialitation if not yet initialized
89 const TensorShape output_shape = compute_roi_align_shape(*input->info(), *rois->info(), pool_info);
90 auto_init_if_empty((*output->info()), output_shape, 1, input->info()->data_type(), input->info()->quantization_info());
91 output->info()->set_data_layout(input->info()->data_layout());
92
93 // Configure kernel window
94 const unsigned int num_rois = rois->info()->dimension(1);
95 Window window;
96 window.set(Window::DimX, Window::Dimension(0, num_rois));
97 window.set(Window::DimY, Window::Dimension(0, 1));
98
99 Coordinates coord;
100 coord.set_num_dimensions(output->info()->num_dimensions());
101 output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
102
103 // Set instance variables
104 _input = input;
105 _rois = rois;
106 _output = output;
107 _pool_info = pool_info;
108
109 INEKernel::configure(window);
110 }
111
validate(const ITensorInfo * input,const ITensorInfo * rois,ITensorInfo * output,const ROIPoolingLayerInfo & pool_info)112 Status NEROIAlignLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info)
113 {
114 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, rois, output, pool_info));
115 return Status{};
116 }
117
118 /** Average pooling over an aligned window */
119 template <typename input_data_type>
roi_align_1x1(const ITensor * input,unsigned int roi_batch,float region_start_x,float bin_size_x,int grid_size_x,float region_end_x,float region_start_y,float bin_size_y,int grid_size_y,float region_end_y,int pz)120 inline input_data_type roi_align_1x1(const ITensor *input,
121 unsigned int roi_batch,
122 float region_start_x,
123 float bin_size_x,
124 int grid_size_x,
125 float region_end_x,
126 float region_start_y,
127 float bin_size_y,
128 int grid_size_y,
129 float region_end_y,
130 int pz)
131 {
132 if((region_end_x <= region_start_x) || (region_end_y <= region_start_y))
133 {
134 return input_data_type(0);
135 }
136 else
137 {
138 const DataLayout data_layout = input->info()->data_layout();
139 float avg = 0;
140 // Iterate through the aligned pooling region
141 for(int iy = 0; iy < grid_size_y; ++iy)
142 {
143 for(int ix = 0; ix < grid_size_x; ++ix)
144 {
145 // Align the window in the middle of every bin
146 float y = region_start_y + (iy + 0.5) * bin_size_y / float(grid_size_y);
147 float x = region_start_x + (ix + 0.5) * bin_size_x / float(grid_size_x);
148
149 // Interpolation in the [0,0] [0,1] [1,0] [1,1] square
150 const int y_low = y;
151 const int x_low = x;
152 const int y_high = y_low + 1;
153 const int x_high = x_low + 1;
154
155 const float ly = y - y_low;
156 const float lx = x - x_low;
157 const float hy = 1. - ly;
158 const float hx = 1. - lx;
159
160 const float w1 = hy * hx;
161 const float w2 = hy * lx;
162 const float w3 = ly * hx;
163 const float w4 = ly * lx;
164 if(data_layout == DataLayout::NCHW)
165 {
166 const auto data1 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_low, pz, roi_batch)));
167 const auto data2 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_low, pz, roi_batch)));
168 const auto data3 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_high, pz, roi_batch)));
169 const auto data4 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_high, pz, roi_batch)));
170 avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4;
171 }
172 else
173 {
174 const auto data1 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_low, roi_batch)));
175 const auto data2 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_low, roi_batch)));
176 const auto data3 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_high, roi_batch)));
177 const auto data4 = *reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_high, roi_batch)));
178 avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4;
179 }
180 }
181 }
182
183 avg /= grid_size_x * grid_size_y;
184 return input_data_type(avg);
185 }
186 }
187
188 /** Average pooling over an aligned window */
189 template <typename input_data_type>
roi_align_1x1_qasymm8(const ITensor * input,unsigned int roi_batch,float region_start_x,float bin_size_x,int grid_size_x,float region_end_x,float region_start_y,float bin_size_y,int grid_size_y,float region_end_y,int pz,const QuantizationInfo & out_qinfo)190 inline input_data_type roi_align_1x1_qasymm8(const ITensor *input,
191 unsigned int roi_batch,
192 float region_start_x,
193 float bin_size_x,
194 int grid_size_x,
195 float region_end_x,
196 float region_start_y,
197 float bin_size_y,
198 int grid_size_y,
199 float region_end_y,
200 int pz,
201 const QuantizationInfo &out_qinfo)
202 {
203 if((region_end_x <= region_start_x) || (region_end_y <= region_start_y))
204 {
205 return input_data_type(out_qinfo.uniform().offset);
206 }
207 else
208 {
209 float avg = 0;
210 const UniformQuantizationInfo input_qinfo = input->info()->quantization_info().uniform();
211 const bool is_qasymm_signed = is_data_type_quantized_asymmetric_signed(input->info()->data_type());
212 const DataLayout data_layout = input->info()->data_layout();
213
214 // Iterate through the aligned pooling region
215 for(int iy = 0; iy < grid_size_y; ++iy)
216 {
217 for(int ix = 0; ix < grid_size_x; ++ix)
218 {
219 // Align the window in the middle of every bin
220 float y = region_start_y + (iy + 0.5) * bin_size_y / float(grid_size_y);
221 float x = region_start_x + (ix + 0.5) * bin_size_x / float(grid_size_x);
222
223 // Interpolation in the [0,0] [0,1] [1,0] [1,1] square
224 const int y_low = y;
225 const int x_low = x;
226 const int y_high = y_low + 1;
227 const int x_high = x_low + 1;
228
229 const float ly = y - y_low;
230 const float lx = x - x_low;
231 const float hy = 1. - ly;
232 const float hx = 1. - lx;
233
234 const float w1 = hy * hx;
235 const float w2 = hy * lx;
236 const float w3 = ly * hx;
237 const float w4 = ly * lx;
238
239 if(data_layout == DataLayout::NCHW)
240 {
241 if(is_qasymm_signed)
242 {
243 float data1 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_low, pz, roi_batch))), input_qinfo);
244 float data2 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_low, pz, roi_batch))), input_qinfo);
245 float data3 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_high, pz, roi_batch))), input_qinfo);
246 float data4 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_high, pz, roi_batch))), input_qinfo);
247 avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4;
248 }
249 else
250 {
251 float data1 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_low, pz, roi_batch))), input_qinfo);
252 float data2 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_low, pz, roi_batch))), input_qinfo);
253 float data3 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_low, y_high, pz, roi_batch))), input_qinfo);
254 float data4 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(x_high, y_high, pz, roi_batch))), input_qinfo);
255 avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4;
256 }
257 }
258 else
259 {
260 if(is_qasymm_signed)
261 {
262 const auto data1 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_low, roi_batch))), input_qinfo);
263 const auto data2 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_low, roi_batch))), input_qinfo);
264 const auto data3 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_high, roi_batch))), input_qinfo);
265 const auto data4 = dequantize_qasymm8_signed(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_high, roi_batch))), input_qinfo);
266 avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4;
267 }
268 else
269 {
270 const auto data1 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_low, roi_batch))), input_qinfo);
271 const auto data2 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_low, roi_batch))), input_qinfo);
272 const auto data3 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_low, y_high, roi_batch))), input_qinfo);
273 const auto data4 = dequantize_qasymm8(*reinterpret_cast<const input_data_type *>(input->ptr_to_element(Coordinates(pz, x_high, y_high, roi_batch))), input_qinfo);
274 avg += w1 * data1 + w2 * data2 + w3 * data3 + w4 * data4;
275 }
276 }
277 }
278 }
279
280 avg /= grid_size_x * grid_size_y;
281
282 input_data_type res = 0;
283 if(is_qasymm_signed)
284 {
285 res = quantize_qasymm8_signed(avg, out_qinfo);
286 }
287 else
288 {
289 res = quantize_qasymm8(avg, out_qinfo);
290 }
291 return res;
292 }
293 }
294
compute_region_coordinate(int p,float bin_size,float roi_anchor,float max_value)295 inline float compute_region_coordinate(int p, float bin_size, float roi_anchor, float max_value)
296 {
297 const float region_start = p * bin_size + roi_anchor;
298 return utility::clamp(region_start, 0.0f, max_value);
299 }
300
run(const Window & window,const ThreadInfo & info)301 void NEROIAlignLayerKernel::run(const Window &window, const ThreadInfo &info)
302 {
303 const DataLayout data_layout = _input->info()->data_layout();
304 if(data_layout == DataLayout::NCHW || data_layout == DataLayout::NHWC)
305 {
306 switch(_input->info()->data_type())
307 {
308 case DataType::QASYMM8:
309 {
310 NEROIAlignLayerKernel::internal_run<uint8_t, uint16_t>(window, info);
311 break;
312 }
313 case DataType::QASYMM8_SIGNED:
314 {
315 NEROIAlignLayerKernel::internal_run<int8_t, uint16_t>(window, info);
316 break;
317 }
318 case DataType::F32:
319 {
320 NEROIAlignLayerKernel::internal_run<float>(window, info);
321 break;
322 }
323 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
324 case DataType::F16:
325 {
326 NEROIAlignLayerKernel::internal_run<float16_t>(window, info);
327 break;
328 }
329 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
330 default:
331 {
332 ARM_COMPUTE_ERROR("DataType not supported");
333 break;
334 }
335 }
336 }
337 else
338 {
339 ARM_COMPUTE_ERROR("Invalid layout");
340 }
341 }
342
343 template <typename input_data_type, typename roi_data_type>
internal_run(const Window & window,const ThreadInfo & info)344 void NEROIAlignLayerKernel::internal_run(const Window &window, const ThreadInfo &info)
345 {
346 ARM_COMPUTE_UNUSED(info);
347 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
348 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
349
350 const DataLayout data_layout = _input->info()->data_layout();
351 const size_t values_per_roi = _rois->info()->dimension(0);
352
353 const int roi_list_start = window.x().start();
354 const int roi_list_end = window.x().end();
355
356 const unsigned int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
357 const unsigned int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
358 const unsigned int idx_depth = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
359
360 const int input_width = _input->info()->dimension(idx_width);
361 const int input_height = _input->info()->dimension(idx_height);
362 const int input_chanels = _input->info()->dimension(idx_depth);
363 const int pooled_w = _pool_info.pooled_width();
364 const int pooled_h = _pool_info.pooled_height();
365
366 const DataType data_type = _input->info()->data_type();
367 const bool is_qasymm = is_data_type_quantized_asymmetric(data_type);
368
369 const auto *rois_ptr = reinterpret_cast<const roi_data_type *>(_rois->buffer());
370 const QuantizationInfo &rois_qinfo = _rois->info()->quantization_info();
371 for(int roi_indx = roi_list_start; roi_indx < roi_list_end; ++roi_indx)
372 {
373 const unsigned int roi_batch = rois_ptr[values_per_roi * roi_indx];
374
375 roi_data_type qx1 = rois_ptr[values_per_roi * roi_indx + 1];
376 roi_data_type qy1 = rois_ptr[values_per_roi * roi_indx + 2];
377 roi_data_type qx2 = rois_ptr[values_per_roi * roi_indx + 3];
378 roi_data_type qy2 = rois_ptr[values_per_roi * roi_indx + 4];
379 float x1(qx1);
380 float x2(qx2);
381 float y1(qy1);
382 float y2(qy2);
383 if(is_qasymm)
384 {
385 x1 = dequantize_qasymm16(qx1, rois_qinfo);
386 x2 = dequantize_qasymm16(qx2, rois_qinfo);
387 y1 = dequantize_qasymm16(qy1, rois_qinfo);
388 y2 = dequantize_qasymm16(qy2, rois_qinfo);
389 }
390 const float roi_anchor_x = x1 * _pool_info.spatial_scale();
391 const float roi_anchor_y = y1 * _pool_info.spatial_scale();
392 const float roi_dims_x = std::max((x2 - x1) * _pool_info.spatial_scale(), 1.0f);
393 const float roi_dims_y = std::max((y2 - y1) * _pool_info.spatial_scale(), 1.0f);
394 float bin_size_x = roi_dims_x / _pool_info.pooled_width();
395 float bin_size_y = roi_dims_y / _pool_info.pooled_height();
396
397 // Iterate through all feature maps
398 for(int ch = 0; ch < input_chanels; ++ch)
399 {
400 // Iterate through all output pixels
401 for(int py = 0; py < pooled_h; ++py)
402 {
403 for(int px = 0; px < pooled_w; ++px)
404 {
405 const float region_start_x = compute_region_coordinate(px, bin_size_x, roi_anchor_x, input_width);
406 const float region_start_y = compute_region_coordinate(py, bin_size_y, roi_anchor_y, input_height);
407 const float region_end_x = compute_region_coordinate(px + 1, bin_size_x, roi_anchor_x, input_width);
408 const float region_end_y = compute_region_coordinate(py + 1, bin_size_y, roi_anchor_y, input_height);
409 const int roi_bin_grid_x = (_pool_info.sampling_ratio() > 0) ? _pool_info.sampling_ratio() : int(ceil(bin_size_x));
410 const int roi_bin_grid_y = (_pool_info.sampling_ratio() > 0) ? _pool_info.sampling_ratio() : int(ceil(bin_size_y));
411 input_data_type out_val(0);
412 if(is_qasymm)
413 {
414 out_val = roi_align_1x1_qasymm8<input_data_type>(
415 _input, roi_batch, region_start_x, bin_size_x,
416 roi_bin_grid_x, region_end_x, region_start_y, bin_size_y,
417 roi_bin_grid_y, region_end_y, ch, _output->info()->quantization_info());
418 }
419 else
420 {
421 out_val = roi_align_1x1<input_data_type>(
422 _input, roi_batch, region_start_x, bin_size_x,
423 roi_bin_grid_x, region_end_x, region_start_y, bin_size_y,
424 roi_bin_grid_y, region_end_y, ch);
425 }
426
427 if(data_layout == DataLayout::NCHW)
428 {
429 auto out_ptr = reinterpret_cast<input_data_type *>(_output->ptr_to_element(Coordinates(px, py, ch, roi_indx)));
430 *out_ptr = out_val;
431 }
432 else
433 {
434 auto out_ptr = reinterpret_cast<input_data_type *>(_output->ptr_to_element(Coordinates(ch, px, py, roi_indx)));
435 *out_ptr = out_val;
436 }
437 }
438 }
439 }
440 }
441 }
442 } // namespace arm_compute
443