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/NEMinMaxLocationKernel.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/IAccessWindow.h"
30 #include "arm_compute/core/ITensor.h"
31 #include "arm_compute/core/TensorInfo.h"
32 #include "arm_compute/core/Types.h"
33 #include "arm_compute/core/Validate.h"
34 #include "arm_compute/core/Window.h"
35 #include "arm_compute/core/utils/misc/Utility.h"
36 #include "src/core/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38
39 #include <algorithm>
40 #include <arm_neon.h>
41 #include <climits>
42 #include <cstddef>
43
44 namespace arm_compute
45 {
NEMinMaxKernel()46 NEMinMaxKernel::NEMinMaxKernel()
47 : _func(), _input(nullptr), _min(), _max(), _mtx()
48 {
49 }
50
configure(const IImage * input,void * min,void * max)51 void NEMinMaxKernel::configure(const IImage *input, void *min, void *max)
52 {
53 ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input);
54 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16, DataType::F32);
55 ARM_COMPUTE_ERROR_ON(nullptr == min);
56 ARM_COMPUTE_ERROR_ON(nullptr == max);
57
58 _input = input;
59 _min = min;
60 _max = max;
61
62 switch(_input->info()->data_type())
63 {
64 case DataType::U8:
65 _func = &NEMinMaxKernel::minmax_U8;
66 break;
67 case DataType::S16:
68 _func = &NEMinMaxKernel::minmax_S16;
69 break;
70 case DataType::F32:
71 _func = &NEMinMaxKernel::minmax_F32;
72 break;
73 default:
74 ARM_COMPUTE_ERROR("Unsupported data type");
75 break;
76 }
77
78 // Configure kernel window
79 constexpr unsigned int num_elems_processed_per_iteration = 1;
80
81 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
82
83 INEKernel::configure(win);
84 }
85
run(const Window & window,const ThreadInfo & info)86 void NEMinMaxKernel::run(const Window &window, const ThreadInfo &info)
87 {
88 ARM_COMPUTE_UNUSED(info);
89 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
90 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
91 ARM_COMPUTE_ERROR_ON(_func == nullptr);
92
93 (this->*_func)(window);
94 }
95
reset()96 void NEMinMaxKernel::reset()
97 {
98 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
99 switch(_input->info()->data_type())
100 {
101 case DataType::U8:
102 *static_cast<int32_t *>(_min) = UCHAR_MAX;
103 *static_cast<int32_t *>(_max) = 0;
104 break;
105 case DataType::S16:
106 *static_cast<int32_t *>(_min) = SHRT_MAX;
107 *static_cast<int32_t *>(_max) = SHRT_MIN;
108 break;
109 case DataType::F32:
110 *static_cast<float *>(_min) = std::numeric_limits<float>::max();
111 *static_cast<float *>(_max) = std::numeric_limits<float>::lowest();
112 break;
113 default:
114 ARM_COMPUTE_ERROR("Unsupported data type");
115 break;
116 }
117 }
118
119 template <typename T>
update_min_max(const T min,const T max)120 void NEMinMaxKernel::update_min_max(const T min, const T max)
121 {
122 arm_compute::lock_guard<arm_compute::Mutex> lock(_mtx);
123
124 using type = typename std::conditional<std::is_same<T, float>::value, float, int32_t>::type;
125
126 auto min_ptr = static_cast<type *>(_min);
127 auto max_ptr = static_cast<type *>(_max);
128
129 if(min < *min_ptr)
130 {
131 *min_ptr = min;
132 }
133
134 if(max > *max_ptr)
135 {
136 *max_ptr = max;
137 }
138 }
139
minmax_U8(Window win)140 void NEMinMaxKernel::minmax_U8(Window win)
141 {
142 uint8x8_t carry_min = vdup_n_u8(UCHAR_MAX);
143 uint8x8_t carry_max = vdup_n_u8(0);
144
145 uint8_t carry_max_scalar = 0;
146 uint8_t carry_min_scalar = UCHAR_MAX;
147
148 const int x_start = win.x().start();
149 const int x_end = win.x().end();
150
151 // Handle X dimension manually to split into two loops
152 // First one will use vector operations, second one processes the left over pixels
153 win.set(Window::DimX, Window::Dimension(0, 1, 1));
154
155 Iterator input(_input, win);
156
157 execute_window_loop(win, [&](const Coordinates &)
158 {
159 int x = x_start;
160
161 // Vector loop
162 for(; x <= x_end - 16; x += 16)
163 {
164 const uint8x16_t pixels = vld1q_u8(input.ptr() + x);
165 const uint8x8_t tmp_min = vmin_u8(vget_high_u8(pixels), vget_low_u8(pixels));
166 const uint8x8_t tmp_max = vmax_u8(vget_high_u8(pixels), vget_low_u8(pixels));
167 carry_min = vmin_u8(tmp_min, carry_min);
168 carry_max = vmax_u8(tmp_max, carry_max);
169 }
170
171 // Process leftover pixels
172 for(; x < x_end; ++x)
173 {
174 const uint8_t pixel = input.ptr()[x];
175 carry_min_scalar = std::min(pixel, carry_min_scalar);
176 carry_max_scalar = std::max(pixel, carry_max_scalar);
177 }
178 },
179 input);
180
181 // Reduce result
182 carry_min = vpmin_u8(carry_min, carry_min);
183 carry_max = vpmax_u8(carry_max, carry_max);
184 carry_min = vpmin_u8(carry_min, carry_min);
185 carry_max = vpmax_u8(carry_max, carry_max);
186 carry_min = vpmin_u8(carry_min, carry_min);
187 carry_max = vpmax_u8(carry_max, carry_max);
188
189 // Extract max/min values
190 const uint8_t min_i = std::min(vget_lane_u8(carry_min, 0), carry_min_scalar);
191 const uint8_t max_i = std::max(vget_lane_u8(carry_max, 0), carry_max_scalar);
192
193 // Perform reduction of local min/max values
194 update_min_max(min_i, max_i);
195 }
196
minmax_S16(Window win)197 void NEMinMaxKernel::minmax_S16(Window win)
198 {
199 int16x4_t carry_min = vdup_n_s16(SHRT_MAX);
200 int16x4_t carry_max = vdup_n_s16(SHRT_MIN);
201
202 int16_t carry_max_scalar = SHRT_MIN;
203 int16_t carry_min_scalar = SHRT_MAX;
204
205 const int x_start = win.x().start();
206 const int x_end = win.x().end();
207
208 // Handle X dimension manually to split into two loops
209 // First one will use vector operations, second one processes the left over pixels
210 win.set(Window::DimX, Window::Dimension(0, 1, 1));
211
212 Iterator input(_input, win);
213
214 execute_window_loop(win, [&](const Coordinates &)
215 {
216 int x = x_start;
217 const auto in_ptr = reinterpret_cast<const int16_t *>(input.ptr());
218
219 // Vector loop
220 for(; x <= x_end - 16; x += 16)
221 {
222 const int16x8x2_t pixels = vld2q_s16(in_ptr + x);
223 const int16x8_t tmp_min1 = vminq_s16(pixels.val[0], pixels.val[1]);
224 const int16x8_t tmp_max1 = vmaxq_s16(pixels.val[0], pixels.val[1]);
225 const int16x4_t tmp_min2 = vmin_s16(vget_high_s16(tmp_min1), vget_low_s16(tmp_min1));
226 const int16x4_t tmp_max2 = vmax_s16(vget_high_s16(tmp_max1), vget_low_s16(tmp_max1));
227 carry_min = vmin_s16(tmp_min2, carry_min);
228 carry_max = vmax_s16(tmp_max2, carry_max);
229 }
230
231 // Process leftover pixels
232 for(; x < x_end; ++x)
233 {
234 const int16_t pixel = in_ptr[x];
235 carry_min_scalar = std::min(pixel, carry_min_scalar);
236 carry_max_scalar = std::max(pixel, carry_max_scalar);
237 }
238
239 },
240 input);
241
242 // Reduce result
243 carry_min = vpmin_s16(carry_min, carry_min);
244 carry_max = vpmax_s16(carry_max, carry_max);
245 carry_min = vpmin_s16(carry_min, carry_min);
246 carry_max = vpmax_s16(carry_max, carry_max);
247
248 // Extract max/min values
249 const int16_t min_i = std::min(vget_lane_s16(carry_min, 0), carry_min_scalar);
250 const int16_t max_i = std::max(vget_lane_s16(carry_max, 0), carry_max_scalar);
251
252 // Perform reduction of local min/max values
253 update_min_max(min_i, max_i);
254 }
255
minmax_F32(Window win)256 void NEMinMaxKernel::minmax_F32(Window win)
257 {
258 float32x2_t carry_min = vdup_n_f32(std::numeric_limits<float>::max());
259 float32x2_t carry_max = vdup_n_f32(std::numeric_limits<float>::lowest());
260
261 float carry_min_scalar = std::numeric_limits<float>::max();
262 float carry_max_scalar = std::numeric_limits<float>::lowest();
263
264 const int x_start = win.x().start();
265 const int x_end = win.x().end();
266
267 // Handle X dimension manually to split into two loops
268 // First one will use vector operations, second one processes the left over pixels
269 win.set(Window::DimX, Window::Dimension(0, 1, 1));
270
271 Iterator input(_input, win);
272
273 execute_window_loop(win, [&](const Coordinates &)
274 {
275 int x = x_start;
276 const auto in_ptr = reinterpret_cast<const float *>(input.ptr());
277
278 // Vector loop
279 for(; x <= x_end - 8; x += 8)
280 {
281 const float32x4x2_t pixels = vld2q_f32(in_ptr + x);
282 const float32x4_t tmp_min1 = vminq_f32(pixels.val[0], pixels.val[1]);
283 const float32x4_t tmp_max1 = vmaxq_f32(pixels.val[0], pixels.val[1]);
284 const float32x2_t tmp_min2 = vmin_f32(vget_high_f32(tmp_min1), vget_low_f32(tmp_min1));
285 const float32x2_t tmp_max2 = vmax_f32(vget_high_f32(tmp_max1), vget_low_f32(tmp_max1));
286 carry_min = vmin_f32(tmp_min2, carry_min);
287 carry_max = vmax_f32(tmp_max2, carry_max);
288 }
289
290 // Process leftover pixels
291 for(; x < x_end; ++x)
292 {
293 const float pixel = in_ptr[x];
294 carry_min_scalar = std::min(pixel, carry_min_scalar);
295 carry_max_scalar = std::max(pixel, carry_max_scalar);
296 }
297
298 },
299 input);
300
301 // Reduce result
302 carry_min = vpmin_f32(carry_min, carry_min);
303 carry_max = vpmax_f32(carry_max, carry_max);
304 carry_min = vpmin_f32(carry_min, carry_min);
305 carry_max = vpmax_f32(carry_max, carry_max);
306
307 // Extract max/min values
308 const float min_i = std::min(vget_lane_f32(carry_min, 0), carry_min_scalar);
309 const float max_i = std::max(vget_lane_f32(carry_max, 0), carry_max_scalar);
310
311 // Perform reduction of local min/max values
312 update_min_max(min_i, max_i);
313 }
314
NEMinMaxLocationKernel()315 NEMinMaxLocationKernel::NEMinMaxLocationKernel()
316 : _func(nullptr), _input(nullptr), _min(nullptr), _max(nullptr), _min_count(nullptr), _max_count(nullptr), _min_loc(nullptr), _max_loc(nullptr)
317 {
318 }
319
is_parallelisable() const320 bool NEMinMaxLocationKernel::is_parallelisable() const
321 {
322 return false;
323 }
324
325 template <class T, std::size_t... N>
326 struct NEMinMaxLocationKernel::create_func_table<T, utility::index_sequence<N...>>
327 {
328 static const std::array<NEMinMaxLocationKernel::MinMaxLocFunction, sizeof...(N)> func_table;
329 };
330
331 template <class T, std::size_t... N>
332 const std::array<NEMinMaxLocationKernel::MinMaxLocFunction, sizeof...(N)> NEMinMaxLocationKernel::create_func_table<T, utility::index_sequence<N...>>::func_table
333 {
334 &NEMinMaxLocationKernel::minmax_loc<T, bool(N & 8), bool(N & 4), bool(N & 2), bool(N & 1)>...
335 };
336
configure(const IImage * input,void * min,void * max,ICoordinates2DArray * min_loc,ICoordinates2DArray * max_loc,uint32_t * min_count,uint32_t * max_count)337 void NEMinMaxLocationKernel::configure(const IImage *input, void *min, void *max,
338 ICoordinates2DArray *min_loc, ICoordinates2DArray *max_loc,
339 uint32_t *min_count, uint32_t *max_count)
340 {
341 ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input);
342 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16, DataType::F32);
343 ARM_COMPUTE_ERROR_ON(nullptr == min);
344 ARM_COMPUTE_ERROR_ON(nullptr == max);
345
346 _input = input;
347 _min = min;
348 _max = max;
349 _min_count = min_count;
350 _max_count = max_count;
351 _min_loc = min_loc;
352 _max_loc = max_loc;
353
354 unsigned int count_min = (nullptr != min_count ? 1 : 0);
355 unsigned int count_max = (nullptr != max_count ? 1 : 0);
356 unsigned int loc_min = (nullptr != min_loc ? 1 : 0);
357 unsigned int loc_max = (nullptr != max_loc ? 1 : 0);
358
359 unsigned int table_idx = (count_min << 3) | (count_max << 2) | (loc_min << 1) | loc_max;
360
361 switch(input->info()->data_type())
362 {
363 case DataType::U8:
364 _func = create_func_table<uint8_t, utility::index_sequence_t<16>>::func_table[table_idx];
365 break;
366 case DataType::S16:
367 _func = create_func_table<int16_t, utility::index_sequence_t<16>>::func_table[table_idx];
368 break;
369 case DataType::F32:
370 _func = create_func_table<float, utility::index_sequence_t<16>>::func_table[table_idx];
371 break;
372 default:
373 ARM_COMPUTE_ERROR("Unsupported data type");
374 break;
375 }
376
377 constexpr unsigned int num_elems_processed_per_iteration = 1;
378
379 // Configure kernel window
380 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));
381
382 update_window_and_padding(win, AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration));
383
384 INEKernel::configure(win);
385 }
386
run(const Window & window,const ThreadInfo & info)387 void NEMinMaxLocationKernel::run(const Window &window, const ThreadInfo &info)
388 {
389 ARM_COMPUTE_UNUSED(info);
390 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
391 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
392 ARM_COMPUTE_ERROR_ON(_func == nullptr);
393
394 (this->*_func)(window);
395 }
396
397 template <class T, bool count_min, bool count_max, bool loc_min, bool loc_max>
minmax_loc(const Window & win)398 void NEMinMaxLocationKernel::minmax_loc(const Window &win)
399 {
400 if(count_min || count_max || loc_min || loc_max)
401 {
402 Iterator input(_input, win);
403
404 size_t min_count = 0;
405 size_t max_count = 0;
406
407 // Clear min location array
408 if(loc_min)
409 {
410 _min_loc->clear();
411 }
412
413 // Clear max location array
414 if(loc_max)
415 {
416 _max_loc->clear();
417 }
418
419 using type = typename std::conditional<std::is_same<T, float>::value, float, int32_t>::type;
420
421 auto min_ptr = static_cast<type *>(_min);
422 auto max_ptr = static_cast<type *>(_max);
423
424 execute_window_loop(win, [&](const Coordinates & id)
425 {
426 auto in_ptr = reinterpret_cast<const T *>(input.ptr());
427 int32_t idx = id.x();
428 int32_t idy = id.y();
429
430 const T pixel = *in_ptr;
431 Coordinates2D p{ idx, idy };
432
433 if(count_min || loc_min)
434 {
435 if(*min_ptr == pixel)
436 {
437 if(count_min)
438 {
439 ++min_count;
440 }
441
442 if(loc_min)
443 {
444 _min_loc->push_back(p);
445 }
446 }
447 }
448
449 if(count_max || loc_max)
450 {
451 if(*max_ptr == pixel)
452 {
453 if(count_max)
454 {
455 ++max_count;
456 }
457
458 if(loc_max)
459 {
460 _max_loc->push_back(p);
461 }
462 }
463 }
464 },
465 input);
466
467 if(count_min)
468 {
469 *_min_count = min_count;
470 }
471
472 if(count_max)
473 {
474 *_max_count = max_count;
475 }
476 }
477 }
478 } // namespace arm_compute
479