1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include "config/aom_config.h"
13 #include "config/aom_scale_rtcd.h"
14
15 #include "aom_dsp/aom_dsp_common.h"
16 #include "aom_mem/aom_mem.h"
17 #include "av1/common/av1_loopfilter.h"
18 #include "av1/common/entropymode.h"
19 #include "av1/common/thread_common.h"
20 #include "av1/common/reconinter.h"
21
22 // Set up nsync by width.
get_sync_range(int width)23 static INLINE int get_sync_range(int width) {
24 // nsync numbers are picked by testing. For example, for 4k
25 // video, using 4 gives best performance.
26 if (width < 640)
27 return 1;
28 else if (width <= 1280)
29 return 2;
30 else if (width <= 4096)
31 return 4;
32 else
33 return 8;
34 }
35
get_lr_sync_range(int width)36 static INLINE int get_lr_sync_range(int width) {
37 #if 0
38 // nsync numbers are picked by testing. For example, for 4k
39 // video, using 4 gives best performance.
40 if (width < 640)
41 return 1;
42 else if (width <= 1280)
43 return 2;
44 else if (width <= 4096)
45 return 4;
46 else
47 return 8;
48 #else
49 (void)width;
50 return 1;
51 #endif
52 }
53
54 // Allocate memory for lf row synchronization
loop_filter_alloc(AV1LfSync * lf_sync,AV1_COMMON * cm,int rows,int width,int num_workers)55 static void loop_filter_alloc(AV1LfSync *lf_sync, AV1_COMMON *cm, int rows,
56 int width, int num_workers) {
57 lf_sync->rows = rows;
58 #if CONFIG_MULTITHREAD
59 {
60 int i, j;
61
62 for (j = 0; j < MAX_MB_PLANE; j++) {
63 CHECK_MEM_ERROR(cm, lf_sync->mutex_[j],
64 aom_malloc(sizeof(*(lf_sync->mutex_[j])) * rows));
65 if (lf_sync->mutex_[j]) {
66 for (i = 0; i < rows; ++i) {
67 pthread_mutex_init(&lf_sync->mutex_[j][i], NULL);
68 }
69 }
70
71 CHECK_MEM_ERROR(cm, lf_sync->cond_[j],
72 aom_malloc(sizeof(*(lf_sync->cond_[j])) * rows));
73 if (lf_sync->cond_[j]) {
74 for (i = 0; i < rows; ++i) {
75 pthread_cond_init(&lf_sync->cond_[j][i], NULL);
76 }
77 }
78 }
79
80 CHECK_MEM_ERROR(cm, lf_sync->job_mutex,
81 aom_malloc(sizeof(*(lf_sync->job_mutex))));
82 if (lf_sync->job_mutex) {
83 pthread_mutex_init(lf_sync->job_mutex, NULL);
84 }
85 }
86 #endif // CONFIG_MULTITHREAD
87 CHECK_MEM_ERROR(cm, lf_sync->lfdata,
88 aom_malloc(num_workers * sizeof(*(lf_sync->lfdata))));
89 lf_sync->num_workers = num_workers;
90
91 for (int j = 0; j < MAX_MB_PLANE; j++) {
92 CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col[j],
93 aom_malloc(sizeof(*(lf_sync->cur_sb_col[j])) * rows));
94 }
95 CHECK_MEM_ERROR(
96 cm, lf_sync->job_queue,
97 aom_malloc(sizeof(*(lf_sync->job_queue)) * rows * MAX_MB_PLANE * 2));
98 // Set up nsync.
99 lf_sync->sync_range = get_sync_range(width);
100 }
101
102 // Deallocate lf synchronization related mutex and data
av1_loop_filter_dealloc(AV1LfSync * lf_sync)103 void av1_loop_filter_dealloc(AV1LfSync *lf_sync) {
104 if (lf_sync != NULL) {
105 int j;
106 #if CONFIG_MULTITHREAD
107 int i;
108 for (j = 0; j < MAX_MB_PLANE; j++) {
109 if (lf_sync->mutex_[j] != NULL) {
110 for (i = 0; i < lf_sync->rows; ++i) {
111 pthread_mutex_destroy(&lf_sync->mutex_[j][i]);
112 }
113 aom_free(lf_sync->mutex_[j]);
114 }
115 if (lf_sync->cond_[j] != NULL) {
116 for (i = 0; i < lf_sync->rows; ++i) {
117 pthread_cond_destroy(&lf_sync->cond_[j][i]);
118 }
119 aom_free(lf_sync->cond_[j]);
120 }
121 }
122 if (lf_sync->job_mutex != NULL) {
123 pthread_mutex_destroy(lf_sync->job_mutex);
124 aom_free(lf_sync->job_mutex);
125 }
126 #endif // CONFIG_MULTITHREAD
127 aom_free(lf_sync->lfdata);
128 for (j = 0; j < MAX_MB_PLANE; j++) {
129 aom_free(lf_sync->cur_sb_col[j]);
130 }
131
132 aom_free(lf_sync->job_queue);
133 // clear the structure as the source of this call may be a resize in which
134 // case this call will be followed by an _alloc() which may fail.
135 av1_zero(*lf_sync);
136 }
137 }
138
loop_filter_data_reset(LFWorkerData * lf_data,YV12_BUFFER_CONFIG * frame_buffer,struct AV1Common * cm,MACROBLOCKD * xd)139 static void loop_filter_data_reset(LFWorkerData *lf_data,
140 YV12_BUFFER_CONFIG *frame_buffer,
141 struct AV1Common *cm, MACROBLOCKD *xd) {
142 struct macroblockd_plane *pd = xd->plane;
143 lf_data->frame_buffer = frame_buffer;
144 lf_data->cm = cm;
145 lf_data->xd = xd;
146 for (int i = 0; i < MAX_MB_PLANE; i++) {
147 memcpy(&lf_data->planes[i].dst, &pd[i].dst, sizeof(lf_data->planes[i].dst));
148 lf_data->planes[i].subsampling_x = pd[i].subsampling_x;
149 lf_data->planes[i].subsampling_y = pd[i].subsampling_y;
150 }
151 }
152
sync_read(AV1LfSync * const lf_sync,int r,int c,int plane)153 static INLINE void sync_read(AV1LfSync *const lf_sync, int r, int c,
154 int plane) {
155 #if CONFIG_MULTITHREAD
156 const int nsync = lf_sync->sync_range;
157
158 if (r && !(c & (nsync - 1))) {
159 pthread_mutex_t *const mutex = &lf_sync->mutex_[plane][r - 1];
160 pthread_mutex_lock(mutex);
161
162 while (c > lf_sync->cur_sb_col[plane][r - 1] - nsync) {
163 pthread_cond_wait(&lf_sync->cond_[plane][r - 1], mutex);
164 }
165 pthread_mutex_unlock(mutex);
166 }
167 #else
168 (void)lf_sync;
169 (void)r;
170 (void)c;
171 (void)plane;
172 #endif // CONFIG_MULTITHREAD
173 }
174
sync_write(AV1LfSync * const lf_sync,int r,int c,const int sb_cols,int plane)175 static INLINE void sync_write(AV1LfSync *const lf_sync, int r, int c,
176 const int sb_cols, int plane) {
177 #if CONFIG_MULTITHREAD
178 const int nsync = lf_sync->sync_range;
179 int cur;
180 // Only signal when there are enough filtered SB for next row to run.
181 int sig = 1;
182
183 if (c < sb_cols - 1) {
184 cur = c;
185 if (c % nsync) sig = 0;
186 } else {
187 cur = sb_cols + nsync;
188 }
189
190 if (sig) {
191 pthread_mutex_lock(&lf_sync->mutex_[plane][r]);
192
193 lf_sync->cur_sb_col[plane][r] = cur;
194
195 pthread_cond_broadcast(&lf_sync->cond_[plane][r]);
196 pthread_mutex_unlock(&lf_sync->mutex_[plane][r]);
197 }
198 #else
199 (void)lf_sync;
200 (void)r;
201 (void)c;
202 (void)sb_cols;
203 (void)plane;
204 #endif // CONFIG_MULTITHREAD
205 }
206
enqueue_lf_jobs(AV1LfSync * lf_sync,AV1_COMMON * cm,int start,int stop,int is_decoding,int plane_start,int plane_end)207 static void enqueue_lf_jobs(AV1LfSync *lf_sync, AV1_COMMON *cm, int start,
208 int stop,
209 #if LOOP_FILTER_BITMASK
210 int is_decoding,
211 #endif
212 int plane_start, int plane_end) {
213 int mi_row, plane, dir;
214 AV1LfMTInfo *lf_job_queue = lf_sync->job_queue;
215 lf_sync->jobs_enqueued = 0;
216 lf_sync->jobs_dequeued = 0;
217
218 for (dir = 0; dir < 2; dir++) {
219 for (plane = plane_start; plane < plane_end; plane++) {
220 if (plane == 0 && !(cm->lf.filter_level[0]) && !(cm->lf.filter_level[1]))
221 break;
222 else if (plane == 1 && !(cm->lf.filter_level_u))
223 continue;
224 else if (plane == 2 && !(cm->lf.filter_level_v))
225 continue;
226 #if LOOP_FILTER_BITMASK
227 int step = MAX_MIB_SIZE;
228 if (is_decoding) {
229 step = MI_SIZE_64X64;
230 }
231 for (mi_row = start; mi_row < stop; mi_row += step)
232 #else
233 for (mi_row = start; mi_row < stop; mi_row += MAX_MIB_SIZE)
234 #endif
235 {
236 lf_job_queue->mi_row = mi_row;
237 lf_job_queue->plane = plane;
238 lf_job_queue->dir = dir;
239 lf_job_queue++;
240 lf_sync->jobs_enqueued++;
241 }
242 }
243 }
244 }
245
get_lf_job_info(AV1LfSync * lf_sync)246 static AV1LfMTInfo *get_lf_job_info(AV1LfSync *lf_sync) {
247 AV1LfMTInfo *cur_job_info = NULL;
248
249 #if CONFIG_MULTITHREAD
250 pthread_mutex_lock(lf_sync->job_mutex);
251
252 if (lf_sync->jobs_dequeued < lf_sync->jobs_enqueued) {
253 cur_job_info = lf_sync->job_queue + lf_sync->jobs_dequeued;
254 lf_sync->jobs_dequeued++;
255 }
256
257 pthread_mutex_unlock(lf_sync->job_mutex);
258 #else
259 (void)lf_sync;
260 #endif
261
262 return cur_job_info;
263 }
264
265 // Implement row loopfiltering for each thread.
thread_loop_filter_rows(const YV12_BUFFER_CONFIG * const frame_buffer,AV1_COMMON * const cm,struct macroblockd_plane * planes,MACROBLOCKD * xd,AV1LfSync * const lf_sync)266 static INLINE void thread_loop_filter_rows(
267 const YV12_BUFFER_CONFIG *const frame_buffer, AV1_COMMON *const cm,
268 struct macroblockd_plane *planes, MACROBLOCKD *xd,
269 AV1LfSync *const lf_sync) {
270 const int sb_cols =
271 ALIGN_POWER_OF_TWO(cm->mi_cols, MAX_MIB_SIZE_LOG2) >> MAX_MIB_SIZE_LOG2;
272 int mi_row, mi_col, plane, dir;
273 int r, c;
274
275 while (1) {
276 AV1LfMTInfo *cur_job_info = get_lf_job_info(lf_sync);
277
278 if (cur_job_info != NULL) {
279 mi_row = cur_job_info->mi_row;
280 plane = cur_job_info->plane;
281 dir = cur_job_info->dir;
282 r = mi_row >> MAX_MIB_SIZE_LOG2;
283
284 if (dir == 0) {
285 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MAX_MIB_SIZE) {
286 c = mi_col >> MAX_MIB_SIZE_LOG2;
287
288 av1_setup_dst_planes(planes, cm->seq_params.sb_size, frame_buffer,
289 mi_row, mi_col, plane, plane + 1);
290
291 av1_filter_block_plane_vert(cm, xd, plane, &planes[plane], mi_row,
292 mi_col);
293 sync_write(lf_sync, r, c, sb_cols, plane);
294 }
295 } else if (dir == 1) {
296 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MAX_MIB_SIZE) {
297 c = mi_col >> MAX_MIB_SIZE_LOG2;
298
299 // Wait for vertical edge filtering of the top-right block to be
300 // completed
301 sync_read(lf_sync, r, c, plane);
302
303 // Wait for vertical edge filtering of the right block to be
304 // completed
305 sync_read(lf_sync, r + 1, c, plane);
306
307 av1_setup_dst_planes(planes, cm->seq_params.sb_size, frame_buffer,
308 mi_row, mi_col, plane, plane + 1);
309 av1_filter_block_plane_horz(cm, xd, plane, &planes[plane], mi_row,
310 mi_col);
311 }
312 }
313 } else {
314 break;
315 }
316 }
317 }
318
319 // Row-based multi-threaded loopfilter hook
loop_filter_row_worker(void * arg1,void * arg2)320 static int loop_filter_row_worker(void *arg1, void *arg2) {
321 AV1LfSync *const lf_sync = (AV1LfSync *)arg1;
322 LFWorkerData *const lf_data = (LFWorkerData *)arg2;
323 thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
324 lf_data->xd, lf_sync);
325 return 1;
326 }
327
328 #if LOOP_FILTER_BITMASK
thread_loop_filter_bitmask_rows(const YV12_BUFFER_CONFIG * const frame_buffer,AV1_COMMON * const cm,struct macroblockd_plane * planes,MACROBLOCKD * xd,AV1LfSync * const lf_sync)329 static INLINE void thread_loop_filter_bitmask_rows(
330 const YV12_BUFFER_CONFIG *const frame_buffer, AV1_COMMON *const cm,
331 struct macroblockd_plane *planes, MACROBLOCKD *xd,
332 AV1LfSync *const lf_sync) {
333 const int sb_cols =
334 ALIGN_POWER_OF_TWO(cm->mi_cols, MIN_MIB_SIZE_LOG2) >> MIN_MIB_SIZE_LOG2;
335 int mi_row, mi_col, plane, dir;
336 int r, c;
337 (void)xd;
338
339 while (1) {
340 AV1LfMTInfo *cur_job_info = get_lf_job_info(lf_sync);
341
342 if (cur_job_info != NULL) {
343 mi_row = cur_job_info->mi_row;
344 plane = cur_job_info->plane;
345 dir = cur_job_info->dir;
346 r = mi_row >> MIN_MIB_SIZE_LOG2;
347
348 if (dir == 0) {
349 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_SIZE_64X64) {
350 c = mi_col >> MIN_MIB_SIZE_LOG2;
351
352 av1_setup_dst_planes(planes, BLOCK_64X64, frame_buffer, mi_row,
353 mi_col, plane, plane + 1);
354
355 av1_filter_block_plane_bitmask_vert(cm, &planes[plane], plane, mi_row,
356 mi_col);
357 sync_write(lf_sync, r, c, sb_cols, plane);
358 }
359 } else if (dir == 1) {
360 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_SIZE_64X64) {
361 c = mi_col >> MIN_MIB_SIZE_LOG2;
362
363 // Wait for vertical edge filtering of the top-right block to be
364 // completed
365 sync_read(lf_sync, r, c, plane);
366
367 // Wait for vertical edge filtering of the right block to be
368 // completed
369 sync_read(lf_sync, r + 1, c, plane);
370
371 av1_setup_dst_planes(planes, BLOCK_64X64, frame_buffer, mi_row,
372 mi_col, plane, plane + 1);
373 av1_filter_block_plane_bitmask_horz(cm, &planes[plane], plane, mi_row,
374 mi_col);
375 }
376 }
377 } else {
378 break;
379 }
380 }
381 }
382
383 // Row-based multi-threaded loopfilter hook
loop_filter_bitmask_row_worker(void * arg1,void * arg2)384 static int loop_filter_bitmask_row_worker(void *arg1, void *arg2) {
385 AV1LfSync *const lf_sync = (AV1LfSync *)arg1;
386 LFWorkerData *const lf_data = (LFWorkerData *)arg2;
387 thread_loop_filter_bitmask_rows(lf_data->frame_buffer, lf_data->cm,
388 lf_data->planes, lf_data->xd, lf_sync);
389 return 1;
390 }
391 #endif // LOOP_FILTER_BITMASK
392
loop_filter_rows_mt(YV12_BUFFER_CONFIG * frame,AV1_COMMON * cm,MACROBLOCKD * xd,int start,int stop,int plane_start,int plane_end,int is_decoding,AVxWorker * workers,int nworkers,AV1LfSync * lf_sync)393 static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
394 MACROBLOCKD *xd, int start, int stop,
395 int plane_start, int plane_end,
396 #if LOOP_FILTER_BITMASK
397 int is_decoding,
398 #endif
399 AVxWorker *workers, int nworkers,
400 AV1LfSync *lf_sync) {
401 const AVxWorkerInterface *const winterface = aom_get_worker_interface();
402 #if LOOP_FILTER_BITMASK
403 int sb_rows;
404 if (is_decoding) {
405 sb_rows =
406 ALIGN_POWER_OF_TWO(cm->mi_rows, MIN_MIB_SIZE_LOG2) >> MIN_MIB_SIZE_LOG2;
407 } else {
408 sb_rows =
409 ALIGN_POWER_OF_TWO(cm->mi_rows, MAX_MIB_SIZE_LOG2) >> MAX_MIB_SIZE_LOG2;
410 }
411 #else
412 // Number of superblock rows and cols
413 const int sb_rows =
414 ALIGN_POWER_OF_TWO(cm->mi_rows, MAX_MIB_SIZE_LOG2) >> MAX_MIB_SIZE_LOG2;
415 #endif
416 const int num_workers = nworkers;
417 int i;
418
419 if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
420 num_workers > lf_sync->num_workers) {
421 av1_loop_filter_dealloc(lf_sync);
422 loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
423 }
424
425 // Initialize cur_sb_col to -1 for all SB rows.
426 for (i = 0; i < MAX_MB_PLANE; i++) {
427 memset(lf_sync->cur_sb_col[i], -1,
428 sizeof(*(lf_sync->cur_sb_col[i])) * sb_rows);
429 }
430
431 enqueue_lf_jobs(lf_sync, cm, start, stop,
432 #if LOOP_FILTER_BITMASK
433 is_decoding,
434 #endif
435 plane_start, plane_end);
436
437 // Set up loopfilter thread data.
438 for (i = 0; i < num_workers; ++i) {
439 AVxWorker *const worker = &workers[i];
440 LFWorkerData *const lf_data = &lf_sync->lfdata[i];
441
442 #if LOOP_FILTER_BITMASK
443 if (is_decoding) {
444 worker->hook = loop_filter_bitmask_row_worker;
445 } else {
446 worker->hook = loop_filter_row_worker;
447 }
448 #else
449 worker->hook = loop_filter_row_worker;
450 #endif
451 worker->data1 = lf_sync;
452 worker->data2 = lf_data;
453
454 // Loopfilter data
455 loop_filter_data_reset(lf_data, frame, cm, xd);
456
457 // Start loopfiltering
458 if (i == num_workers - 1) {
459 winterface->execute(worker);
460 } else {
461 winterface->launch(worker);
462 }
463 }
464
465 // Wait till all rows are finished
466 for (i = 0; i < num_workers; ++i) {
467 winterface->sync(&workers[i]);
468 }
469 }
470
av1_loop_filter_frame_mt(YV12_BUFFER_CONFIG * frame,AV1_COMMON * cm,MACROBLOCKD * xd,int plane_start,int plane_end,int partial_frame,int is_decoding,AVxWorker * workers,int num_workers,AV1LfSync * lf_sync)471 void av1_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
472 MACROBLOCKD *xd, int plane_start, int plane_end,
473 int partial_frame,
474 #if LOOP_FILTER_BITMASK
475 int is_decoding,
476 #endif
477 AVxWorker *workers, int num_workers,
478 AV1LfSync *lf_sync) {
479 int start_mi_row, end_mi_row, mi_rows_to_filter;
480
481 start_mi_row = 0;
482 mi_rows_to_filter = cm->mi_rows;
483 if (partial_frame && cm->mi_rows > 8) {
484 start_mi_row = cm->mi_rows >> 1;
485 start_mi_row &= 0xfffffff8;
486 mi_rows_to_filter = AOMMAX(cm->mi_rows / 8, 8);
487 }
488 end_mi_row = start_mi_row + mi_rows_to_filter;
489 av1_loop_filter_frame_init(cm, plane_start, plane_end);
490
491 #if LOOP_FILTER_BITMASK
492 if (is_decoding) {
493 cm->is_decoding = is_decoding;
494 // TODO(chengchen): currently use one thread to build bitmasks for the
495 // frame. Make it support multi-thread later.
496 for (int plane = plane_start; plane < plane_end; plane++) {
497 if (plane == 0 && !(cm->lf.filter_level[0]) && !(cm->lf.filter_level[1]))
498 break;
499 else if (plane == 1 && !(cm->lf.filter_level_u))
500 continue;
501 else if (plane == 2 && !(cm->lf.filter_level_v))
502 continue;
503
504 // TODO(chengchen): can we remove this?
505 struct macroblockd_plane *pd = xd->plane;
506 av1_setup_dst_planes(pd, cm->seq_params.sb_size, frame, 0, 0, plane,
507 plane + 1);
508
509 av1_build_bitmask_vert_info(cm, &pd[plane], plane);
510 av1_build_bitmask_horz_info(cm, &pd[plane], plane);
511 }
512 loop_filter_rows_mt(frame, cm, xd, start_mi_row, end_mi_row, plane_start,
513 plane_end, 1, workers, num_workers, lf_sync);
514 } else {
515 loop_filter_rows_mt(frame, cm, xd, start_mi_row, end_mi_row, plane_start,
516 plane_end, 0, workers, num_workers, lf_sync);
517 }
518 #else
519 loop_filter_rows_mt(frame, cm, xd, start_mi_row, end_mi_row, plane_start,
520 plane_end, workers, num_workers, lf_sync);
521 #endif
522 }
523
lr_sync_read(void * const lr_sync,int r,int c,int plane)524 static INLINE void lr_sync_read(void *const lr_sync, int r, int c, int plane) {
525 #if CONFIG_MULTITHREAD
526 AV1LrSync *const loop_res_sync = (AV1LrSync *)lr_sync;
527 const int nsync = loop_res_sync->sync_range;
528
529 if (r && !(c & (nsync - 1))) {
530 pthread_mutex_t *const mutex = &loop_res_sync->mutex_[plane][r - 1];
531 pthread_mutex_lock(mutex);
532
533 while (c > loop_res_sync->cur_sb_col[plane][r - 1] - nsync) {
534 pthread_cond_wait(&loop_res_sync->cond_[plane][r - 1], mutex);
535 }
536 pthread_mutex_unlock(mutex);
537 }
538 #else
539 (void)lr_sync;
540 (void)r;
541 (void)c;
542 (void)plane;
543 #endif // CONFIG_MULTITHREAD
544 }
545
lr_sync_write(void * const lr_sync,int r,int c,const int sb_cols,int plane)546 static INLINE void lr_sync_write(void *const lr_sync, int r, int c,
547 const int sb_cols, int plane) {
548 #if CONFIG_MULTITHREAD
549 AV1LrSync *const loop_res_sync = (AV1LrSync *)lr_sync;
550 const int nsync = loop_res_sync->sync_range;
551 int cur;
552 // Only signal when there are enough filtered SB for next row to run.
553 int sig = 1;
554
555 if (c < sb_cols - 1) {
556 cur = c;
557 if (c % nsync) sig = 0;
558 } else {
559 cur = sb_cols + nsync;
560 }
561
562 if (sig) {
563 pthread_mutex_lock(&loop_res_sync->mutex_[plane][r]);
564
565 loop_res_sync->cur_sb_col[plane][r] = cur;
566
567 pthread_cond_broadcast(&loop_res_sync->cond_[plane][r]);
568 pthread_mutex_unlock(&loop_res_sync->mutex_[plane][r]);
569 }
570 #else
571 (void)lr_sync;
572 (void)r;
573 (void)c;
574 (void)sb_cols;
575 (void)plane;
576 #endif // CONFIG_MULTITHREAD
577 }
578
579 // Allocate memory for loop restoration row synchronization
loop_restoration_alloc(AV1LrSync * lr_sync,AV1_COMMON * cm,int num_workers,int num_rows_lr,int num_planes,int width)580 static void loop_restoration_alloc(AV1LrSync *lr_sync, AV1_COMMON *cm,
581 int num_workers, int num_rows_lr,
582 int num_planes, int width) {
583 lr_sync->rows = num_rows_lr;
584 lr_sync->num_planes = num_planes;
585 #if CONFIG_MULTITHREAD
586 {
587 int i, j;
588
589 for (j = 0; j < num_planes; j++) {
590 CHECK_MEM_ERROR(cm, lr_sync->mutex_[j],
591 aom_malloc(sizeof(*(lr_sync->mutex_[j])) * num_rows_lr));
592 if (lr_sync->mutex_[j]) {
593 for (i = 0; i < num_rows_lr; ++i) {
594 pthread_mutex_init(&lr_sync->mutex_[j][i], NULL);
595 }
596 }
597
598 CHECK_MEM_ERROR(cm, lr_sync->cond_[j],
599 aom_malloc(sizeof(*(lr_sync->cond_[j])) * num_rows_lr));
600 if (lr_sync->cond_[j]) {
601 for (i = 0; i < num_rows_lr; ++i) {
602 pthread_cond_init(&lr_sync->cond_[j][i], NULL);
603 }
604 }
605 }
606
607 CHECK_MEM_ERROR(cm, lr_sync->job_mutex,
608 aom_malloc(sizeof(*(lr_sync->job_mutex))));
609 if (lr_sync->job_mutex) {
610 pthread_mutex_init(lr_sync->job_mutex, NULL);
611 }
612 }
613 #endif // CONFIG_MULTITHREAD
614 CHECK_MEM_ERROR(cm, lr_sync->lrworkerdata,
615 aom_malloc(num_workers * sizeof(*(lr_sync->lrworkerdata))));
616
617 for (int worker_idx = 0; worker_idx < num_workers; ++worker_idx) {
618 if (worker_idx < num_workers - 1) {
619 CHECK_MEM_ERROR(cm, lr_sync->lrworkerdata[worker_idx].rst_tmpbuf,
620 (int32_t *)aom_memalign(16, RESTORATION_TMPBUF_SIZE));
621 CHECK_MEM_ERROR(cm, lr_sync->lrworkerdata[worker_idx].rlbs,
622 aom_malloc(sizeof(RestorationLineBuffers)));
623
624 } else {
625 lr_sync->lrworkerdata[worker_idx].rst_tmpbuf = cm->rst_tmpbuf;
626 lr_sync->lrworkerdata[worker_idx].rlbs = cm->rlbs;
627 }
628 }
629
630 lr_sync->num_workers = num_workers;
631
632 for (int j = 0; j < num_planes; j++) {
633 CHECK_MEM_ERROR(
634 cm, lr_sync->cur_sb_col[j],
635 aom_malloc(sizeof(*(lr_sync->cur_sb_col[j])) * num_rows_lr));
636 }
637 CHECK_MEM_ERROR(
638 cm, lr_sync->job_queue,
639 aom_malloc(sizeof(*(lr_sync->job_queue)) * num_rows_lr * num_planes));
640 // Set up nsync.
641 lr_sync->sync_range = get_lr_sync_range(width);
642 }
643
644 // Deallocate loop restoration synchronization related mutex and data
av1_loop_restoration_dealloc(AV1LrSync * lr_sync,int num_workers)645 void av1_loop_restoration_dealloc(AV1LrSync *lr_sync, int num_workers) {
646 if (lr_sync != NULL) {
647 int j;
648 #if CONFIG_MULTITHREAD
649 int i;
650 for (j = 0; j < MAX_MB_PLANE; j++) {
651 if (lr_sync->mutex_[j] != NULL) {
652 for (i = 0; i < lr_sync->rows; ++i) {
653 pthread_mutex_destroy(&lr_sync->mutex_[j][i]);
654 }
655 aom_free(lr_sync->mutex_[j]);
656 }
657 if (lr_sync->cond_[j] != NULL) {
658 for (i = 0; i < lr_sync->rows; ++i) {
659 pthread_cond_destroy(&lr_sync->cond_[j][i]);
660 }
661 aom_free(lr_sync->cond_[j]);
662 }
663 }
664 if (lr_sync->job_mutex != NULL) {
665 pthread_mutex_destroy(lr_sync->job_mutex);
666 aom_free(lr_sync->job_mutex);
667 }
668 #endif // CONFIG_MULTITHREAD
669 for (j = 0; j < MAX_MB_PLANE; j++) {
670 aom_free(lr_sync->cur_sb_col[j]);
671 }
672
673 aom_free(lr_sync->job_queue);
674
675 if (lr_sync->lrworkerdata) {
676 for (int worker_idx = 0; worker_idx < num_workers - 1; worker_idx++) {
677 LRWorkerData *const workerdata_data =
678 lr_sync->lrworkerdata + worker_idx;
679
680 aom_free(workerdata_data->rst_tmpbuf);
681 aom_free(workerdata_data->rlbs);
682 }
683 aom_free(lr_sync->lrworkerdata);
684 }
685
686 // clear the structure as the source of this call may be a resize in which
687 // case this call will be followed by an _alloc() which may fail.
688 av1_zero(*lr_sync);
689 }
690 }
691
enqueue_lr_jobs(AV1LrSync * lr_sync,AV1LrStruct * lr_ctxt,AV1_COMMON * cm)692 static void enqueue_lr_jobs(AV1LrSync *lr_sync, AV1LrStruct *lr_ctxt,
693 AV1_COMMON *cm) {
694 FilterFrameCtxt *ctxt = lr_ctxt->ctxt;
695
696 const int num_planes = av1_num_planes(cm);
697 AV1LrMTInfo *lr_job_queue = lr_sync->job_queue;
698 int32_t lr_job_counter[2], num_even_lr_jobs = 0;
699 lr_sync->jobs_enqueued = 0;
700 lr_sync->jobs_dequeued = 0;
701
702 for (int plane = 0; plane < num_planes; plane++) {
703 if (cm->rst_info[plane].frame_restoration_type == RESTORE_NONE) continue;
704 num_even_lr_jobs =
705 num_even_lr_jobs + ((ctxt[plane].rsi->vert_units_per_tile + 1) >> 1);
706 }
707 lr_job_counter[0] = 0;
708 lr_job_counter[1] = num_even_lr_jobs;
709
710 for (int plane = 0; plane < num_planes; plane++) {
711 if (cm->rst_info[plane].frame_restoration_type == RESTORE_NONE) continue;
712 const int is_uv = plane > 0;
713 const int ss_y = is_uv && cm->seq_params.subsampling_y;
714
715 AV1PixelRect tile_rect = ctxt[plane].tile_rect;
716 const int unit_size = ctxt[plane].rsi->restoration_unit_size;
717
718 const int tile_h = tile_rect.bottom - tile_rect.top;
719 const int ext_size = unit_size * 3 / 2;
720
721 int y0 = 0, i = 0;
722 while (y0 < tile_h) {
723 int remaining_h = tile_h - y0;
724 int h = (remaining_h < ext_size) ? remaining_h : unit_size;
725
726 RestorationTileLimits limits;
727 limits.v_start = tile_rect.top + y0;
728 limits.v_end = tile_rect.top + y0 + h;
729 assert(limits.v_end <= tile_rect.bottom);
730 // Offset the tile upwards to align with the restoration processing stripe
731 const int voffset = RESTORATION_UNIT_OFFSET >> ss_y;
732 limits.v_start = AOMMAX(tile_rect.top, limits.v_start - voffset);
733 if (limits.v_end < tile_rect.bottom) limits.v_end -= voffset;
734
735 assert(lr_job_counter[0] <= num_even_lr_jobs);
736
737 lr_job_queue[lr_job_counter[i & 1]].lr_unit_row = i;
738 lr_job_queue[lr_job_counter[i & 1]].plane = plane;
739 lr_job_queue[lr_job_counter[i & 1]].v_start = limits.v_start;
740 lr_job_queue[lr_job_counter[i & 1]].v_end = limits.v_end;
741 lr_job_queue[lr_job_counter[i & 1]].sync_mode = i & 1;
742 if ((i & 1) == 0) {
743 lr_job_queue[lr_job_counter[i & 1]].v_copy_start =
744 limits.v_start + RESTORATION_BORDER;
745 lr_job_queue[lr_job_counter[i & 1]].v_copy_end =
746 limits.v_end - RESTORATION_BORDER;
747 if (i == 0) {
748 assert(limits.v_start == tile_rect.top);
749 lr_job_queue[lr_job_counter[i & 1]].v_copy_start = tile_rect.top;
750 }
751 if (i == (ctxt[plane].rsi->vert_units_per_tile - 1)) {
752 assert(limits.v_end == tile_rect.bottom);
753 lr_job_queue[lr_job_counter[i & 1]].v_copy_end = tile_rect.bottom;
754 }
755 } else {
756 lr_job_queue[lr_job_counter[i & 1]].v_copy_start =
757 AOMMAX(limits.v_start - RESTORATION_BORDER, tile_rect.top);
758 lr_job_queue[lr_job_counter[i & 1]].v_copy_end =
759 AOMMIN(limits.v_end + RESTORATION_BORDER, tile_rect.bottom);
760 }
761 lr_job_counter[i & 1]++;
762 lr_sync->jobs_enqueued++;
763
764 y0 += h;
765 ++i;
766 }
767 }
768 }
769
get_lr_job_info(AV1LrSync * lr_sync)770 static AV1LrMTInfo *get_lr_job_info(AV1LrSync *lr_sync) {
771 AV1LrMTInfo *cur_job_info = NULL;
772
773 #if CONFIG_MULTITHREAD
774 pthread_mutex_lock(lr_sync->job_mutex);
775
776 if (lr_sync->jobs_dequeued < lr_sync->jobs_enqueued) {
777 cur_job_info = lr_sync->job_queue + lr_sync->jobs_dequeued;
778 lr_sync->jobs_dequeued++;
779 }
780
781 pthread_mutex_unlock(lr_sync->job_mutex);
782 #else
783 (void)lr_sync;
784 #endif
785
786 return cur_job_info;
787 }
788
789 // Implement row loop restoration for each thread.
loop_restoration_row_worker(void * arg1,void * arg2)790 static int loop_restoration_row_worker(void *arg1, void *arg2) {
791 AV1LrSync *const lr_sync = (AV1LrSync *)arg1;
792 LRWorkerData *lrworkerdata = (LRWorkerData *)arg2;
793 AV1LrStruct *lr_ctxt = (AV1LrStruct *)lrworkerdata->lr_ctxt;
794 FilterFrameCtxt *ctxt = lr_ctxt->ctxt;
795 int lr_unit_row;
796 int plane;
797 const int tile_row = LR_TILE_ROW;
798 const int tile_col = LR_TILE_COL;
799 const int tile_cols = LR_TILE_COLS;
800 const int tile_idx = tile_col + tile_row * tile_cols;
801 typedef void (*copy_fun)(const YV12_BUFFER_CONFIG *src_ybc,
802 YV12_BUFFER_CONFIG *dst_ybc, int hstart, int hend,
803 int vstart, int vend);
804 static const copy_fun copy_funs[3] = { aom_yv12_partial_coloc_copy_y,
805 aom_yv12_partial_coloc_copy_u,
806 aom_yv12_partial_coloc_copy_v };
807
808 while (1) {
809 AV1LrMTInfo *cur_job_info = get_lr_job_info(lr_sync);
810 if (cur_job_info != NULL) {
811 RestorationTileLimits limits;
812 sync_read_fn_t on_sync_read;
813 sync_write_fn_t on_sync_write;
814 limits.v_start = cur_job_info->v_start;
815 limits.v_end = cur_job_info->v_end;
816 lr_unit_row = cur_job_info->lr_unit_row;
817 plane = cur_job_info->plane;
818 const int unit_idx0 = tile_idx * ctxt[plane].rsi->units_per_tile;
819
820 // sync_mode == 1 implies only sync read is required in LR Multi-threading
821 // sync_mode == 0 implies only sync write is required.
822 on_sync_read =
823 cur_job_info->sync_mode == 1 ? lr_sync_read : av1_lr_sync_read_dummy;
824 on_sync_write = cur_job_info->sync_mode == 0 ? lr_sync_write
825 : av1_lr_sync_write_dummy;
826
827 av1_foreach_rest_unit_in_row(
828 &limits, &(ctxt[plane].tile_rect), lr_ctxt->on_rest_unit, lr_unit_row,
829 ctxt[plane].rsi->restoration_unit_size, unit_idx0,
830 ctxt[plane].rsi->horz_units_per_tile,
831 ctxt[plane].rsi->vert_units_per_tile, plane, &ctxt[plane],
832 lrworkerdata->rst_tmpbuf, lrworkerdata->rlbs, on_sync_read,
833 on_sync_write, lr_sync);
834
835 copy_funs[plane](lr_ctxt->dst, lr_ctxt->frame, ctxt[plane].tile_rect.left,
836 ctxt[plane].tile_rect.right, cur_job_info->v_copy_start,
837 cur_job_info->v_copy_end);
838 } else {
839 break;
840 }
841 }
842 return 1;
843 }
844
foreach_rest_unit_in_planes_mt(AV1LrStruct * lr_ctxt,AVxWorker * workers,int nworkers,AV1LrSync * lr_sync,AV1_COMMON * cm)845 static void foreach_rest_unit_in_planes_mt(AV1LrStruct *lr_ctxt,
846 AVxWorker *workers, int nworkers,
847 AV1LrSync *lr_sync, AV1_COMMON *cm) {
848 FilterFrameCtxt *ctxt = lr_ctxt->ctxt;
849
850 const int num_planes = av1_num_planes(cm);
851
852 const AVxWorkerInterface *const winterface = aom_get_worker_interface();
853 int num_rows_lr = 0;
854
855 for (int plane = 0; plane < num_planes; plane++) {
856 if (cm->rst_info[plane].frame_restoration_type == RESTORE_NONE) continue;
857
858 const AV1PixelRect tile_rect = ctxt[plane].tile_rect;
859 const int max_tile_h = tile_rect.bottom - tile_rect.top;
860
861 const int unit_size = cm->rst_info[plane].restoration_unit_size;
862
863 num_rows_lr =
864 AOMMAX(num_rows_lr, av1_lr_count_units_in_tile(unit_size, max_tile_h));
865 }
866
867 const int num_workers = nworkers;
868 int i;
869 assert(MAX_MB_PLANE == 3);
870
871 if (!lr_sync->sync_range || num_rows_lr != lr_sync->rows ||
872 num_workers > lr_sync->num_workers || num_planes != lr_sync->num_planes) {
873 av1_loop_restoration_dealloc(lr_sync, num_workers);
874 loop_restoration_alloc(lr_sync, cm, num_workers, num_rows_lr, num_planes,
875 cm->width);
876 }
877
878 // Initialize cur_sb_col to -1 for all SB rows.
879 for (i = 0; i < num_planes; i++) {
880 memset(lr_sync->cur_sb_col[i], -1,
881 sizeof(*(lr_sync->cur_sb_col[i])) * num_rows_lr);
882 }
883
884 enqueue_lr_jobs(lr_sync, lr_ctxt, cm);
885
886 // Set up looprestoration thread data.
887 for (i = 0; i < num_workers; ++i) {
888 AVxWorker *const worker = &workers[i];
889 lr_sync->lrworkerdata[i].lr_ctxt = (void *)lr_ctxt;
890 worker->hook = loop_restoration_row_worker;
891 worker->data1 = lr_sync;
892 worker->data2 = &lr_sync->lrworkerdata[i];
893
894 // Start loopfiltering
895 if (i == num_workers - 1) {
896 winterface->execute(worker);
897 } else {
898 winterface->launch(worker);
899 }
900 }
901
902 // Wait till all rows are finished
903 for (i = 0; i < num_workers; ++i) {
904 winterface->sync(&workers[i]);
905 }
906 }
907
av1_loop_restoration_filter_frame_mt(YV12_BUFFER_CONFIG * frame,AV1_COMMON * cm,int optimized_lr,AVxWorker * workers,int num_workers,AV1LrSync * lr_sync,void * lr_ctxt)908 void av1_loop_restoration_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
909 AV1_COMMON *cm, int optimized_lr,
910 AVxWorker *workers, int num_workers,
911 AV1LrSync *lr_sync, void *lr_ctxt) {
912 assert(!cm->all_lossless);
913
914 const int num_planes = av1_num_planes(cm);
915
916 AV1LrStruct *loop_rest_ctxt = (AV1LrStruct *)lr_ctxt;
917
918 av1_loop_restoration_filter_frame_init(loop_rest_ctxt, frame, cm,
919 optimized_lr, num_planes);
920
921 foreach_rest_unit_in_planes_mt(loop_rest_ctxt, workers, num_workers, lr_sync,
922 cm);
923 }
924