1 /*
2 * Copyright (c) 2016-2020 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/NEWarpKernel.h"
25
26 #include "arm_compute/core/Coordinates.h"
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/Helpers.h"
29 #include "arm_compute/core/ITensor.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "arm_compute/core/Validate.h"
32 #include "arm_compute/core/Window.h"
33 #include "src/core/AccessWindowStatic.h"
34 #include "src/core/helpers/AutoConfiguration.h"
35 #include "src/core/helpers/ScaleHelpers.h"
36 #include "src/core/helpers/WindowHelpers.h"
37
38 #include <cstddef>
39
40 using namespace arm_compute;
41
42 namespace
43 {
nearest_interpolation(const uint8_t * in_ptr,int x,int y,size_t stride)44 inline uint8_t nearest_interpolation(const uint8_t *in_ptr, int x, int y, size_t stride)
45 {
46 return in_ptr[x + y * stride];
47 }
48 } // namespace
49
INEWarpKernel()50 INEWarpKernel::INEWarpKernel()
51 : _func(nullptr), _input(nullptr), _output(nullptr), _constant_border_value(0), _matrix()
52 {
53 }
54
border_size() const55 BorderSize INEWarpKernel::border_size() const
56 {
57 return BorderSize(1);
58 }
59
run(const Window & window,const ThreadInfo & info)60 void INEWarpKernel::run(const Window &window, const ThreadInfo &info)
61 {
62 ARM_COMPUTE_UNUSED(info);
63 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
64 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
65 ARM_COMPUTE_ERROR_ON(_func == nullptr);
66
67 (this->*_func)(window);
68 }
69
configure(const ITensor * input,ITensor * output,const std::array<float,9> & matrix,BorderMode border_mode,uint8_t constant_border_value)70 void INEWarpKernel::configure(const ITensor *input, ITensor *output, const std::array<float, 9> &matrix, BorderMode border_mode, uint8_t constant_border_value)
71 {
72 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
73 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
74
75 _matrix = matrix;
76 _constant_border_value = constant_border_value;
77
78 switch(border_mode)
79 {
80 case BorderMode::UNDEFINED:
81 _func = &INEWarpKernel::warp_undefined;
82 break;
83 case BorderMode::CONSTANT:
84 _func = &INEWarpKernel::warp_constant;
85 break;
86 case BorderMode::REPLICATE:
87 _func = &INEWarpKernel::warp_replicate;
88 break;
89 default:
90 ARM_COMPUTE_ERROR("Border mode not supported");
91 break;
92 }
93
94 _input = input;
95 _output = output;
96
97 // Configure kernel window
98 Window win = calculate_max_window(*output->info(), Steps(1U));
99
100 const ValidRegion &input_valid_region = input->info()->valid_region();
101
102 // Reads can occur within the valid region of the input
103 AccessWindowStatic input_access(input->info(),
104 input_valid_region.anchor[0] - border_size().left, input_valid_region.anchor[1] - border_size().top,
105 input_valid_region.anchor[0] + input_valid_region.shape[0] + border_size().right,
106 input_valid_region.anchor[1] + input_valid_region.shape[1] + border_size().bottom);
107 AccessWindowHorizontal output_access(output->info(), 0, 1);
108
109 update_window_and_padding(win, input_access, output_access);
110
111 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
112
113 INEKernel::configure(win);
114 }
115
116 template <InterpolationPolicy interpolation>
warp_undefined(const Window & window)117 void NEWarpAffineKernel<interpolation>::warp_undefined(const Window &window)
118 {
119 // Don't increment in X and Y direction for the input tensor
120 // A pointer to the start of this plane is needed as base for the precomputed offsets
121 Window win_in(window);
122 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
123 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
124
125 Iterator in(_input, win_in);
126 Iterator out(_output, window);
127
128 const int min_x = _input->info()->valid_region().anchor[0];
129 const int max_x = min_x + _input->info()->valid_region().shape[0];
130 const int min_y = _input->info()->valid_region().anchor[1];
131 const int max_y = min_y + _input->info()->valid_region().shape[1];
132 const size_t stride = _input->info()->strides_in_bytes()[1];
133
134 // x0 = M01 * x + M01 * y + M02
135 // y0 = M11 * x + M11 * y + M12
136 const float M00 = _matrix[0];
137 const float M10 = _matrix[1];
138 const float M01 = _matrix[0 + 1 * 2];
139 const float M11 = _matrix[1 + 1 * 2];
140 const float M02 = _matrix[0 + 2 * 2];
141 const float M12 = _matrix[1 + 2 * 2];
142
143 // "M00 * x" and "M10 * x", when x = window.x.start
144 const float start_x0 = M00 * window.x().start();
145 const float start_y0 = M10 * window.x().start();
146
147 // Current row
148 int y_cur = window.y().start();
149 int z_cur = window.z().start();
150 int d3_cur = window[3].start();
151 int d4_cur = window[4].start();
152 int d5_cur = window[5].start();
153
154 // const_x0 and const_y0 are the constant parts of x0 and y0 during the row processing
155 float const_x0 = M01 * y_cur + M02;
156 float const_y0 = M11 * y_cur + M12;
157
158 // Affine warp coordinates
159 float x0 = start_x0 + const_x0;
160 float y0 = start_y0 + const_y0;
161
162 execute_window_loop(window, [&](const Coordinates & id)
163 {
164 // Check if we are processing a new row. If so, update the current processed row (y_cur), x0, y0 and z0
165 if((y_cur != id.y()) || (z_cur != id.z()) || (d3_cur != id[3]) || (d4_cur != id[4]) || (d5_cur != id[5]))
166 {
167 y_cur = id.y();
168 z_cur = id.z();
169 d3_cur = id[3];
170 d4_cur = id[4];
171 d5_cur = id[5];
172
173 const_x0 = M01 * y_cur + M02;
174 const_y0 = M11 * y_cur + M12;
175
176 x0 = start_x0 + const_x0;
177 y0 = start_y0 + const_y0;
178 }
179
180 // Only write to output if x0 and y0 are within the valid region.
181 // Otherwise the read value would be undefined.
182 if((min_y <= y0) && (y0 < max_y) && (min_x <= x0) && (x0 < max_x))
183 {
184 switch(interpolation)
185 {
186 case InterpolationPolicy::NEAREST_NEIGHBOR:
187 *out.ptr() = nearest_interpolation(in.ptr(), x0, y0, stride);
188 break;
189 case InterpolationPolicy::BILINEAR:
190 *out.ptr() = scale_helpers::pixel_bilinear_c1(in.ptr(), stride, x0, y0);
191 break;
192 default:
193 ARM_COMPUTE_ERROR("Interpolation not supported");
194 }
195 }
196
197 x0 += M00;
198 y0 += M10;
199 },
200 in, out);
201 }
202
203 template <InterpolationPolicy interpolation>
warp_constant(const Window & window)204 void NEWarpAffineKernel<interpolation>::warp_constant(const Window &window)
205 {
206 // Don't increment in X and Y direction for the input tensor
207 // A pointer to the start of this plane is needed as base for the precomputed offsets
208 Window win_in(window);
209 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
210 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
211
212 Iterator in(_input, win_in);
213 Iterator out(_output, window);
214
215 const int min_x = _input->info()->valid_region().anchor[0];
216 const int max_x = min_x + _input->info()->valid_region().shape[0];
217 const int min_y = _input->info()->valid_region().anchor[1];
218 const int max_y = min_y + _input->info()->valid_region().shape[1];
219 const size_t stride = _input->info()->strides_in_bytes()[1];
220
221 // x0 = M01 * x + M01 * y + M02
222 // y0 = M11 * x + M11 * y + M12
223 const float M00 = _matrix[0];
224 const float M10 = _matrix[1];
225 const float M01 = _matrix[0 + 1 * 2];
226 const float M11 = _matrix[1 + 1 * 2];
227 const float M02 = _matrix[0 + 2 * 2];
228 const float M12 = _matrix[1 + 2 * 2];
229
230 // "M00 * x" and "M10 * x", when x = window.x.start
231 const float start_x0 = M00 * window.x().start();
232 const float start_y0 = M10 * window.x().start();
233
234 // Current row
235 int y_cur = window.y().start();
236 int z_cur = window.z().start();
237 int d3_cur = window[3].start();
238 int d4_cur = window[4].start();
239 int d5_cur = window[5].start();
240
241 // const_x0 and const_y0 are the constant parts of x0 and y0 during the row processing
242 float const_x0 = M01 * y_cur + M02;
243 float const_y0 = M11 * y_cur + M12;
244
245 // Affine warp coordinates
246 float x0 = start_x0 + const_x0;
247 float y0 = start_y0 + const_y0;
248
249 execute_window_loop(window, [&](const Coordinates & id)
250 {
251 // Check if we are processing a new row. If so, update the current processed row (y_cur), x0, y0 and z0
252 if((y_cur != id.y()) || (z_cur != id.z()) || (d3_cur != id[3]) || (d4_cur != id[4]) || (d5_cur != id[5]))
253 {
254 y_cur = id.y();
255 z_cur = id.z();
256 d3_cur = id[3];
257 d4_cur = id[4];
258 d5_cur = id[5];
259
260 const_x0 = M01 * y_cur + M02;
261 const_y0 = M11 * y_cur + M12;
262
263 x0 = start_x0 + const_x0;
264 y0 = start_y0 + const_y0;
265 }
266
267 // Only use input values if x0 and y0 are within the valid region.
268 // Otherwise write the constant border value.
269 if((min_y <= y0) && (y0 < max_y) && (min_x <= x0) && (x0 < max_x))
270 {
271 switch(interpolation)
272 {
273 case InterpolationPolicy::NEAREST_NEIGHBOR:
274 *out.ptr() = nearest_interpolation(in.ptr(), x0, y0, stride);
275 break;
276 case InterpolationPolicy::BILINEAR:
277 *out.ptr() = scale_helpers::pixel_bilinear_c1(in.ptr(), stride, x0, y0);
278 break;
279 default:
280 ARM_COMPUTE_ERROR("Interpolation not supported");
281 }
282 }
283 else
284 {
285 switch(interpolation)
286 {
287 case InterpolationPolicy::NEAREST_NEIGHBOR:
288 *out.ptr() = _constant_border_value;
289 break;
290 case InterpolationPolicy::BILINEAR:
291 {
292 const auto xi = utility::clamp<int>(std::floor(x0), min_x - 1, max_x);
293 const auto yi = utility::clamp<int>(std::floor(y0), min_y - 1, max_y);
294 const auto xi_1 = utility::clamp<int>(std::floor(x0 + 1), min_x - 1, max_x);
295 const auto yi_1 = utility::clamp<int>(std::floor(y0 + 1), min_y - 1, max_y);
296
297 const float dx = x0 - std::floor(x0);
298 const float dy = y0 - std::floor(y0);
299 const float dx1 = 1.0f - dx;
300 const float dy1 = 1.0f - dy;
301
302 const float a00 = *(in.ptr() + xi + yi * stride);
303 const float a01 = *(in.ptr() + xi_1 + yi * stride);
304 const float a10 = *(in.ptr() + xi + yi_1 * stride);
305 const float a11 = *(in.ptr() + xi_1 + yi_1 * stride);
306
307 *out.ptr() = a00 * (dx1 * dy1) + a01 * (dx * dy1) + a10 * (dx1 * dy) + a11 * (dx * dy);
308 }
309 break;
310 default:
311 ARM_COMPUTE_ERROR("Interpolation not supported");
312 }
313 }
314
315 x0 += M00;
316 y0 += M10;
317 },
318 in, out);
319 }
320
321 template <InterpolationPolicy interpolation>
warp_replicate(const Window & window)322 void NEWarpAffineKernel<interpolation>::warp_replicate(const Window &window)
323 {
324 // Don't increment in X and Y direction for the input tensor
325 // A pointer to the start of this plane is needed as base for the precomputed offsets
326 Window win_in(window);
327 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
328 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
329
330 Iterator in(_input, win_in);
331 Iterator out(_output, window);
332
333 const int min_x = _input->info()->valid_region().anchor[0];
334 const int max_x = min_x + _input->info()->valid_region().shape[0];
335 const int min_y = _input->info()->valid_region().anchor[1];
336 const int max_y = min_y + _input->info()->valid_region().shape[1];
337 const size_t stride = _input->info()->strides_in_bytes()[1];
338
339 // Current row
340 int y_cur = window.y().start();
341 int z_cur = window.z().start();
342 int d3_cur = window[3].start();
343 int d4_cur = window[4].start();
344 int d5_cur = window[5].start();
345
346 const float M00 = _matrix[0];
347 const float M10 = _matrix[1];
348 const float M01 = _matrix[0 + 1 * 2];
349 const float M11 = _matrix[1 + 1 * 2];
350 const float M02 = _matrix[0 + 2 * 2];
351 const float M12 = _matrix[1 + 2 * 2];
352
353 // "M00 * x" and "M10 * x", when x = window.x.start
354 const float start_x0 = M00 * window.x().start();
355 const float start_y0 = M10 * window.x().start();
356
357 // const_x0 and const_y0 are the constant parts of x0 and y0 during the row processing
358 float const_x0 = M01 * y_cur + M02;
359 float const_y0 = M11 * y_cur + M12;
360
361 float x0 = start_x0 + const_x0;
362 float y0 = start_y0 + const_y0;
363
364 execute_window_loop(window, [&](const Coordinates & id)
365 {
366 // Check if we are processing a new row. If so, update the current processed row (y_cur), x0, y0 and z0
367 if((y_cur != id.y()) || (z_cur != id.z()) || (d3_cur != id[3]) || (d4_cur != id[4]) || (d5_cur != id[5]))
368 {
369 y_cur = id.y();
370 z_cur = id.z();
371 d3_cur = id[3];
372 d4_cur = id[4];
373 d5_cur = id[5];
374
375 const_x0 = M01 * y_cur + M02;
376 const_y0 = M11 * y_cur + M12;
377
378 x0 = start_x0 + const_x0;
379 y0 = start_y0 + const_y0;
380 }
381
382 // Only load from (x0, y0) if the point is within the valid region.
383 // Otherwise load from the edge of the valid region.
384 if((min_y <= y0) && (y0 < max_y) && (min_x <= x0) && (x0 < max_x))
385 {
386 switch(interpolation)
387 {
388 case InterpolationPolicy::NEAREST_NEIGHBOR:
389 *out.ptr() = nearest_interpolation(in.ptr(), x0, y0, stride);
390 break;
391 case InterpolationPolicy::BILINEAR:
392 *out.ptr() = scale_helpers::pixel_bilinear_c1(in.ptr(), stride, x0, y0);
393 break;
394 default:
395 ARM_COMPUTE_ERROR("Interpolation not supported");
396 }
397 }
398 else
399 {
400 // Clamp coordinates
401 const auto xi = utility::clamp<int>(std::floor(x0), min_x, max_x - 1);
402 const auto yi = utility::clamp<int>(std::floor(y0), min_y, max_y - 1);
403 switch(interpolation)
404 {
405 case InterpolationPolicy::NEAREST_NEIGHBOR:
406 *out.ptr() = *(in.ptr() + xi + yi * stride);
407 break;
408 case InterpolationPolicy::BILINEAR:
409 {
410 const auto xi_1 = utility::clamp<int>(std::floor(x0 + 1), min_x, max_x - 1);
411 const auto yi_1 = utility::clamp<int>(std::floor(y0 + 1), min_y, max_y - 1);
412
413 const float dx = x0 - std::floor(x0);
414 const float dy = y0 - std::floor(y0);
415 const float dx1 = 1.0f - dx;
416 const float dy1 = 1.0f - dy;
417
418 const float a00 = *(in.ptr() + xi + yi * stride);
419 const float a01 = *(in.ptr() + xi_1 + yi * stride);
420 const float a10 = *(in.ptr() + xi + yi_1 * stride);
421 const float a11 = *(in.ptr() + xi_1 + yi_1 * stride);
422
423 *out.ptr() = a00 * (dx1 * dy1) + a01 * (dx * dy1) + a10 * (dx1 * dy) + a11 * (dx * dy);
424 }
425 break;
426 default:
427 ARM_COMPUTE_ERROR("Interpolation not supported");
428 }
429 }
430
431 x0 += M00;
432 y0 += M10;
433 },
434 in, out);
435 }
436
437 template <InterpolationPolicy interpolation>
warp_undefined(const Window & window)438 void NEWarpPerspectiveKernel<interpolation>::warp_undefined(const Window &window)
439 {
440 // Don't increment in X and Y direction for the input tensor
441 // A pointer to the start of this plane is needed as base for the precomputed offsets
442 Window win_in(window);
443 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
444 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
445
446 Iterator in(_input, win_in);
447 Iterator out(_output, window);
448
449 const int min_x = _input->info()->valid_region().anchor[0];
450 const int max_x = min_x + _input->info()->valid_region().shape[0];
451 const int min_y = _input->info()->valid_region().anchor[1];
452 const int max_y = min_y + _input->info()->valid_region().shape[1];
453 const size_t stride = _input->info()->strides_in_bytes()[1];
454
455 // x0 = M00 * x + M01 * y + M02
456 // y0 = M10 * x + M11 * y + M12
457 // z0 = M20 * x + M21 * y + M22
458 // xn = x0 / z0
459 // yn = y0 / z0
460 const float M00 = _matrix[0];
461 const float M10 = _matrix[1];
462 const float M20 = _matrix[2];
463 const float M01 = _matrix[0 + 1 * 3];
464 const float M11 = _matrix[1 + 1 * 3];
465 const float M21 = _matrix[2 + 1 * 3];
466 const float M02 = _matrix[0 + 2 * 3];
467 const float M12 = _matrix[1 + 2 * 3];
468 const float M22 = _matrix[2 + 2 * 3];
469
470 // "M00 * x", "M10 * x" and "M20 * x", when x = window.x.start
471 const float start_x0 = M00 * window.x().start();
472 const float start_y0 = M10 * window.x().start();
473 const float start_z0 = M20 * window.x().start();
474
475 // Current row
476 int y_cur = window.y().start();
477 int z_cur = window.z().start();
478 int d3_cur = window[3].start();
479 int d4_cur = window[4].start();
480 int d5_cur = window[5].start();
481
482 // const_x0, const_y0 and const_z0 are the constant parts of x0, y0 and z0 during the row processing
483 float const_x0 = M01 * y_cur + M02;
484 float const_y0 = M11 * y_cur + M12;
485 float const_z0 = M21 * y_cur + M22;
486
487 // Perspective warp coordinates
488 float x0 = start_x0 + const_x0;
489 float y0 = start_y0 + const_y0;
490 float z0 = start_z0 + const_z0;
491
492 execute_window_loop(window, [&](const Coordinates & id)
493 {
494 // Check if we are processing a new row. If so, update the current processed row (y_cur), x0, y0 and z0
495 if((y_cur != id.y()) || (z_cur != id.z()) || (d3_cur != id[3]) || (d4_cur != id[4]) || (d5_cur != id[5]))
496 {
497 y_cur = id.y();
498 z_cur = id.z();
499 d3_cur = id[3];
500 d4_cur = id[4];
501 d5_cur = id[5];
502
503 const_x0 = M01 * y_cur + M02;
504 const_y0 = M11 * y_cur + M12;
505 const_z0 = M21 * y_cur + M22;
506
507 x0 = start_x0 + const_x0;
508 y0 = start_y0 + const_y0;
509 z0 = start_z0 + const_z0;
510 }
511
512 const float xn = x0 / z0;
513 const float yn = y0 / z0;
514
515 // Only write to output if xn and yn are within the valid region.
516 // Otherwise the read value would be undefined.
517 if((min_y <= yn) && (yn < max_y) && (min_x <= xn) && (xn < max_x))
518 {
519 switch(interpolation)
520 {
521 case InterpolationPolicy::NEAREST_NEIGHBOR:
522 *out.ptr() = nearest_interpolation(in.ptr(), xn, yn, stride);
523 break;
524 case InterpolationPolicy::BILINEAR:
525 *out.ptr() = scale_helpers::pixel_bilinear_c1(in.ptr(), stride, xn, yn);
526 break;
527 default:
528 ARM_COMPUTE_ERROR("Interpolation not supported");
529 }
530 }
531
532 x0 += M00;
533 y0 += M10;
534 z0 += M20;
535 },
536 in, out);
537 }
538
539 template <InterpolationPolicy interpolation>
warp_constant(const Window & window)540 void NEWarpPerspectiveKernel<interpolation>::warp_constant(const Window &window)
541 {
542 // Don't increment in X and Y direction for the input tensor
543 // A pointer to the start of this plane is needed as base for the precomputed offsets
544 Window win_in(window);
545 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
546 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
547
548 Iterator in(_input, win_in);
549 Iterator out(_output, window);
550
551 const int min_x = _input->info()->valid_region().anchor[0];
552 const int max_x = min_x + _input->info()->valid_region().shape[0];
553 const int min_y = _input->info()->valid_region().anchor[1];
554 const int max_y = min_y + _input->info()->valid_region().shape[1];
555 const size_t stride = _input->info()->strides_in_bytes()[1];
556
557 // x0 = M00 * x + M01 * y + M02
558 // y0 = M10 * x + M11 * y + M12
559 // z0 = M20 * x + M21 * y + M22
560 // xn = x0 / z0
561 // yn = y0 / z0
562 const float M00 = _matrix[0];
563 const float M10 = _matrix[1];
564 const float M20 = _matrix[2];
565 const float M01 = _matrix[0 + 1 * 3];
566 const float M11 = _matrix[1 + 1 * 3];
567 const float M21 = _matrix[2 + 1 * 3];
568 const float M02 = _matrix[0 + 2 * 3];
569 const float M12 = _matrix[1 + 2 * 3];
570 const float M22 = _matrix[2 + 2 * 3];
571
572 // "M00 * x", "M10 * x" and "M20 * x", when x = window.x.start
573 const float start_x0 = M00 * window.x().start();
574 const float start_y0 = M10 * window.x().start();
575 const float start_z0 = M20 * window.x().start();
576
577 // Current row
578 int y_cur = window.y().start();
579 int z_cur = window.z().start();
580 int d3_cur = window[3].start();
581 int d4_cur = window[4].start();
582 int d5_cur = window[5].start();
583
584 // const_x0, const_y0 and const_z0 are the constant parts of x0, y0 and z0 during the row processing
585 float const_x0 = M01 * y_cur + M02;
586 float const_y0 = M11 * y_cur + M12;
587 float const_z0 = M21 * y_cur + M22;
588
589 // Perspective warp coordinates
590 float x0 = start_x0 + const_x0;
591 float y0 = start_y0 + const_y0;
592 float z0 = start_z0 + const_z0;
593
594 execute_window_loop(window, [&](const Coordinates & id)
595 {
596 // Check if we are processing a new row. If so, update the current processed row (y_cur), x0, y0 and z0
597 if((y_cur != id.y()) || (z_cur != id.z()) || (d3_cur != id[3]) || (d4_cur != id[4]) || (d5_cur != id[5]))
598 {
599 y_cur = id.y();
600 z_cur = id.z();
601 d3_cur = id[3];
602 d4_cur = id[4];
603 d5_cur = id[5];
604
605 const_x0 = M01 * y_cur + M02;
606 const_y0 = M11 * y_cur + M12;
607 const_z0 = M21 * y_cur + M22;
608
609 x0 = start_x0 + const_x0;
610 y0 = start_y0 + const_y0;
611 z0 = start_z0 + const_z0;
612 }
613
614 const float xn = x0 / z0;
615 const float yn = y0 / z0;
616
617 // Only use input values if xn and yn are within the valid region.
618 if((min_y <= yn) && (yn < max_y) && (min_x <= xn) && (xn < max_x))
619 {
620 switch(interpolation)
621 {
622 case InterpolationPolicy::NEAREST_NEIGHBOR:
623 *out.ptr() = nearest_interpolation(in.ptr(), xn, yn, stride);
624 break;
625 case InterpolationPolicy::BILINEAR:
626 *out.ptr() = scale_helpers::pixel_bilinear_c1(in.ptr(), stride, xn, yn);
627 break;
628 default:
629 ARM_COMPUTE_ERROR("Interpolation not supported");
630 }
631 }
632 else
633 {
634 switch(interpolation)
635 {
636 case InterpolationPolicy::NEAREST_NEIGHBOR:
637 *out.ptr() = _constant_border_value;
638 break;
639 case InterpolationPolicy::BILINEAR:
640 {
641 const auto xi = utility::clamp<int>(std::floor(xn), min_x - 1, max_x);
642 const auto yi = utility::clamp<int>(std::floor(yn), min_y - 1, max_y);
643 const auto xi_1 = utility::clamp<int>(std::floor(xn + 1), min_x - 1, max_x);
644 const auto yi_1 = utility::clamp<int>(std::floor(yn + 1), min_y - 1, max_y);
645
646 const float dx = xn - std::floor(xn);
647 const float dy = yn - std::floor(yn);
648 const float dx1 = 1.0f - dx;
649 const float dy1 = 1.0f - dy;
650
651 const float a00 = *(in.ptr() + xi + yi * stride);
652 const float a01 = *(in.ptr() + xi_1 + yi * stride);
653 const float a10 = *(in.ptr() + xi + yi_1 * stride);
654 const float a11 = *(in.ptr() + xi_1 + yi_1 * stride);
655
656 *out.ptr() = a00 * (dx1 * dy1) + a01 * (dx * dy1) + a10 * (dx1 * dy) + a11 * (dx * dy);
657 }
658 break;
659 default:
660 ARM_COMPUTE_ERROR("Interpolation not supported");
661 }
662 }
663
664 x0 += M00;
665 y0 += M10;
666 z0 += M20;
667 },
668 in, out);
669 }
670
671 template <InterpolationPolicy interpolation>
warp_replicate(const Window & window)672 void NEWarpPerspectiveKernel<interpolation>::warp_replicate(const Window &window)
673 {
674 // Don't increment in X and Y direction for the input tensor
675 // A pointer to the start of this plane is needed as base for the precomputed offsets
676 Window win_in(window);
677 win_in.set(Window::DimX, Window::Dimension(0, 0, 0));
678 win_in.set(Window::DimY, Window::Dimension(0, 0, 0));
679
680 Iterator in(_input, win_in);
681 Iterator out(_output, window);
682
683 const int min_x = _input->info()->valid_region().anchor[0];
684 const int max_x = min_x + _input->info()->valid_region().shape[0];
685 const int min_y = _input->info()->valid_region().anchor[1];
686 const int max_y = min_y + _input->info()->valid_region().shape[1];
687 const size_t stride = _input->info()->strides_in_bytes()[1];
688
689 // Current row
690 int y_cur = window.y().start();
691 int z_cur = window.z().start();
692 int d3_cur = window[3].start();
693 int d4_cur = window[4].start();
694 int d5_cur = window[5].start();
695
696 // x0 = M00 * x + M01 * y + M02
697 // y0 = M10 * x + M11 * y + M12
698 // z0 = M20 * x + M21 * y + M22
699 // xn = x0 / z0
700 // yn = y0 / z0
701 const float M00 = _matrix[0];
702 const float M10 = _matrix[1];
703 const float M20 = _matrix[2];
704 const float M01 = _matrix[0 + 1 * 3];
705 const float M11 = _matrix[1 + 1 * 3];
706 const float M21 = _matrix[2 + 1 * 3];
707 const float M02 = _matrix[0 + 2 * 3];
708 const float M12 = _matrix[1 + 2 * 3];
709 const float M22 = _matrix[2 + 2 * 3];
710
711 // "M00 * x", "M10 * x" and "M20 * x", when x = window.x.start
712 const float start_x0 = M00 * window.x().start();
713 const float start_y0 = M10 * window.x().start();
714 const float start_z0 = M20 * window.x().start();
715
716 // const_x0, const_y0 and const_z0 are the constant parts of x0, y0 and z0 during the row processing
717 float const_x0 = M01 * y_cur + M02;
718 float const_y0 = M11 * y_cur + M12;
719 float const_z0 = M21 * y_cur + M22;
720
721 // Perspective warp coordinates
722 float x0 = start_x0 + const_x0;
723 float y0 = start_y0 + const_y0;
724 float z0 = start_z0 + const_z0;
725
726 execute_window_loop(window, [&](const Coordinates & id)
727 {
728 // Check if we are processing a new row. If so, update the current processed row (y_cur), x0, y0 and z0
729 if((y_cur != id.y()) || (z_cur != id.z()) || (d3_cur != id[3]) || (d4_cur != id[4]) || (d5_cur != id[5]))
730 {
731 y_cur = id.y();
732 z_cur = id.z();
733 d3_cur = id[3];
734 d4_cur = id[4];
735 d5_cur = id[5];
736
737 const_x0 = M01 * y_cur + M02;
738 const_y0 = M11 * y_cur + M12;
739 const_z0 = M21 * y_cur + M22;
740
741 x0 = start_x0 + const_x0;
742 y0 = start_y0 + const_y0;
743 z0 = start_z0 + const_z0;
744 }
745
746 const float xn = x0 / z0;
747 const float yn = y0 / z0;
748
749 // Only load from (x0, y0) if the point is within the valid region.
750 if((min_y <= yn) && (yn < max_y) && (min_x <= xn) && (xn < max_x))
751 {
752 switch(interpolation)
753 {
754 case InterpolationPolicy::NEAREST_NEIGHBOR:
755 *out.ptr() = nearest_interpolation(in.ptr(), xn, yn, stride);
756 break;
757 case InterpolationPolicy::BILINEAR:
758 *out.ptr() = scale_helpers::pixel_bilinear_c1(in.ptr(), stride, xn, yn);
759 break;
760 default:
761 ARM_COMPUTE_ERROR("Interpolation not supported");
762 }
763 }
764 else
765 {
766 // Clamp coordinates
767 const auto xi = utility::clamp<int>(std::floor(xn), min_x, max_x - 1);
768 const auto yi = utility::clamp<int>(std::floor(yn), min_y, max_y - 1);
769 switch(interpolation)
770 {
771 case InterpolationPolicy::NEAREST_NEIGHBOR:
772 *out.ptr() = *(in.ptr() + xi + yi * stride);
773 break;
774 case InterpolationPolicy::BILINEAR:
775 {
776 const auto xi_1 = utility::clamp<int>(std::floor(xn + 1), min_x, max_x - 1);
777 const auto yi_1 = utility::clamp<int>(std::floor(yn + 1), min_y, max_y - 1);
778
779 const float dx = xn - std::floor(xn);
780 const float dy = yn - std::floor(yn);
781 const float dx1 = 1.0f - dx;
782 const float dy1 = 1.0f - dy;
783
784 const float a00 = *(in.ptr() + xi + yi * stride);
785 const float a01 = *(in.ptr() + xi_1 + yi * stride);
786 const float a10 = *(in.ptr() + xi + yi_1 * stride);
787 const float a11 = *(in.ptr() + xi_1 + yi_1 * stride);
788
789 *out.ptr() = a00 * (dx1 * dy1) + a01 * (dx * dy1) + a10 * (dx1 * dy) + a11 * (dx * dy);
790 }
791 break;
792 default:
793 ARM_COMPUTE_ERROR("Interpolation not supported");
794 }
795 }
796
797 x0 += M00;
798 y0 += M10;
799 z0 += M20;
800 },
801 in, out);
802 }
803
804 template class arm_compute::NEWarpAffineKernel<InterpolationPolicy::NEAREST_NEIGHBOR>;
805 template class arm_compute::NEWarpAffineKernel<InterpolationPolicy::BILINEAR>;
806 template class arm_compute::NEWarpPerspectiveKernel<InterpolationPolicy::NEAREST_NEIGHBOR>;
807 template class arm_compute::NEWarpPerspectiveKernel<InterpolationPolicy::BILINEAR>;
808