/src/aom/av1/common/thread_common.c
Line | Count | Source (jump to first uncovered line) |
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. |
23 | 0 | 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 | 0 | if (width < 640) |
27 | 0 | return 1; |
28 | 0 | else if (width <= 1280) |
29 | 0 | return 2; |
30 | 0 | else if (width <= 4096) |
31 | 0 | return 4; |
32 | 0 | else |
33 | 0 | return 8; |
34 | 0 | } |
35 | | |
36 | | #if !CONFIG_REALTIME_ONLY |
37 | 0 | static INLINE int get_lr_sync_range(int width) { |
38 | | #if 0 |
39 | | // nsync numbers are picked by testing. For example, for 4k |
40 | | // video, using 4 gives best performance. |
41 | | if (width < 640) |
42 | | return 1; |
43 | | else if (width <= 1280) |
44 | | return 2; |
45 | | else if (width <= 4096) |
46 | | return 4; |
47 | | else |
48 | | return 8; |
49 | | #else |
50 | 0 | (void)width; |
51 | 0 | return 1; |
52 | 0 | #endif |
53 | 0 | } |
54 | | #endif |
55 | | |
56 | | // Allocate memory for lf row synchronization |
57 | | void av1_loop_filter_alloc(AV1LfSync *lf_sync, AV1_COMMON *cm, int rows, |
58 | 0 | int width, int num_workers) { |
59 | 0 | lf_sync->rows = rows; |
60 | 0 | #if CONFIG_MULTITHREAD |
61 | 0 | { |
62 | 0 | int i, j; |
63 | |
|
64 | 0 | for (j = 0; j < MAX_MB_PLANE; j++) { |
65 | 0 | CHECK_MEM_ERROR(cm, lf_sync->mutex_[j], |
66 | 0 | aom_malloc(sizeof(*(lf_sync->mutex_[j])) * rows)); |
67 | 0 | if (lf_sync->mutex_[j]) { |
68 | 0 | for (i = 0; i < rows; ++i) { |
69 | 0 | pthread_mutex_init(&lf_sync->mutex_[j][i], NULL); |
70 | 0 | } |
71 | 0 | } |
72 | |
|
73 | 0 | CHECK_MEM_ERROR(cm, lf_sync->cond_[j], |
74 | 0 | aom_malloc(sizeof(*(lf_sync->cond_[j])) * rows)); |
75 | 0 | if (lf_sync->cond_[j]) { |
76 | 0 | for (i = 0; i < rows; ++i) { |
77 | 0 | pthread_cond_init(&lf_sync->cond_[j][i], NULL); |
78 | 0 | } |
79 | 0 | } |
80 | 0 | } |
81 | |
|
82 | 0 | CHECK_MEM_ERROR(cm, lf_sync->job_mutex, |
83 | 0 | aom_malloc(sizeof(*(lf_sync->job_mutex)))); |
84 | 0 | if (lf_sync->job_mutex) { |
85 | 0 | pthread_mutex_init(lf_sync->job_mutex, NULL); |
86 | 0 | } |
87 | 0 | } |
88 | 0 | #endif // CONFIG_MULTITHREAD |
89 | 0 | CHECK_MEM_ERROR(cm, lf_sync->lfdata, |
90 | 0 | aom_malloc(num_workers * sizeof(*(lf_sync->lfdata)))); |
91 | 0 | lf_sync->num_workers = num_workers; |
92 | |
|
93 | 0 | for (int j = 0; j < MAX_MB_PLANE; j++) { |
94 | 0 | CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col[j], |
95 | 0 | aom_malloc(sizeof(*(lf_sync->cur_sb_col[j])) * rows)); |
96 | 0 | } |
97 | 0 | CHECK_MEM_ERROR( |
98 | 0 | cm, lf_sync->job_queue, |
99 | 0 | aom_malloc(sizeof(*(lf_sync->job_queue)) * rows * MAX_MB_PLANE * 2)); |
100 | | // Set up nsync. |
101 | 0 | lf_sync->sync_range = get_sync_range(width); |
102 | 0 | } |
103 | | |
104 | | // Deallocate lf synchronization related mutex and data |
105 | 0 | void av1_loop_filter_dealloc(AV1LfSync *lf_sync) { |
106 | 0 | if (lf_sync != NULL) { |
107 | 0 | int j; |
108 | 0 | #if CONFIG_MULTITHREAD |
109 | 0 | int i; |
110 | 0 | for (j = 0; j < MAX_MB_PLANE; j++) { |
111 | 0 | if (lf_sync->mutex_[j] != NULL) { |
112 | 0 | for (i = 0; i < lf_sync->rows; ++i) { |
113 | 0 | pthread_mutex_destroy(&lf_sync->mutex_[j][i]); |
114 | 0 | } |
115 | 0 | aom_free(lf_sync->mutex_[j]); |
116 | 0 | } |
117 | 0 | if (lf_sync->cond_[j] != NULL) { |
118 | 0 | for (i = 0; i < lf_sync->rows; ++i) { |
119 | 0 | pthread_cond_destroy(&lf_sync->cond_[j][i]); |
120 | 0 | } |
121 | 0 | aom_free(lf_sync->cond_[j]); |
122 | 0 | } |
123 | 0 | } |
124 | 0 | if (lf_sync->job_mutex != NULL) { |
125 | 0 | pthread_mutex_destroy(lf_sync->job_mutex); |
126 | 0 | aom_free(lf_sync->job_mutex); |
127 | 0 | } |
128 | 0 | #endif // CONFIG_MULTITHREAD |
129 | 0 | aom_free(lf_sync->lfdata); |
130 | 0 | for (j = 0; j < MAX_MB_PLANE; j++) { |
131 | 0 | aom_free(lf_sync->cur_sb_col[j]); |
132 | 0 | } |
133 | |
|
134 | 0 | aom_free(lf_sync->job_queue); |
135 | | // clear the structure as the source of this call may be a resize in which |
136 | | // case this call will be followed by an _alloc() which may fail. |
137 | 0 | av1_zero(*lf_sync); |
138 | 0 | } |
139 | 0 | } |
140 | | |
141 | | static void loop_filter_data_reset(LFWorkerData *lf_data, |
142 | | YV12_BUFFER_CONFIG *frame_buffer, |
143 | 0 | struct AV1Common *cm, MACROBLOCKD *xd) { |
144 | 0 | struct macroblockd_plane *pd = xd->plane; |
145 | 0 | lf_data->frame_buffer = frame_buffer; |
146 | 0 | lf_data->cm = cm; |
147 | 0 | lf_data->xd = xd; |
148 | 0 | for (int i = 0; i < MAX_MB_PLANE; i++) { |
149 | 0 | memcpy(&lf_data->planes[i].dst, &pd[i].dst, sizeof(lf_data->planes[i].dst)); |
150 | 0 | lf_data->planes[i].subsampling_x = pd[i].subsampling_x; |
151 | 0 | lf_data->planes[i].subsampling_y = pd[i].subsampling_y; |
152 | 0 | } |
153 | 0 | } |
154 | | |
155 | | void av1_alloc_cdef_sync(AV1_COMMON *const cm, AV1CdefSync *cdef_sync, |
156 | 0 | int num_workers) { |
157 | 0 | if (num_workers < 1) return; |
158 | 0 | #if CONFIG_MULTITHREAD |
159 | 0 | if (cdef_sync->mutex_ == NULL) { |
160 | 0 | CHECK_MEM_ERROR(cm, cdef_sync->mutex_, |
161 | 0 | aom_malloc(sizeof(*(cdef_sync->mutex_)))); |
162 | 0 | if (cdef_sync->mutex_) pthread_mutex_init(cdef_sync->mutex_, NULL); |
163 | 0 | } |
164 | | #else |
165 | | (void)cm; |
166 | | (void)cdef_sync; |
167 | | #endif // CONFIG_MULTITHREAD |
168 | 0 | } |
169 | | |
170 | 0 | void av1_free_cdef_sync(AV1CdefSync *cdef_sync) { |
171 | 0 | if (cdef_sync == NULL) return; |
172 | 0 | #if CONFIG_MULTITHREAD |
173 | 0 | if (cdef_sync->mutex_ != NULL) { |
174 | 0 | pthread_mutex_destroy(cdef_sync->mutex_); |
175 | 0 | aom_free(cdef_sync->mutex_); |
176 | 0 | } |
177 | 0 | #endif // CONFIG_MULTITHREAD |
178 | 0 | } |
179 | | |
180 | | static INLINE void cdef_row_mt_sync_read(AV1CdefSync *const cdef_sync, |
181 | 0 | int row) { |
182 | 0 | if (!row) return; |
183 | 0 | #if CONFIG_MULTITHREAD |
184 | 0 | AV1CdefRowSync *const cdef_row_mt = cdef_sync->cdef_row_mt; |
185 | 0 | pthread_mutex_lock(cdef_row_mt[row - 1].row_mutex_); |
186 | 0 | while (cdef_row_mt[row - 1].is_row_done != 1) |
187 | 0 | pthread_cond_wait(cdef_row_mt[row - 1].row_cond_, |
188 | 0 | cdef_row_mt[row - 1].row_mutex_); |
189 | 0 | cdef_row_mt[row - 1].is_row_done = 0; |
190 | 0 | pthread_mutex_unlock(cdef_row_mt[row - 1].row_mutex_); |
191 | | #else |
192 | | (void)cdef_sync; |
193 | | #endif // CONFIG_MULTITHREAD |
194 | 0 | } |
195 | | |
196 | | static INLINE void cdef_row_mt_sync_write(AV1CdefSync *const cdef_sync, |
197 | 0 | int row) { |
198 | 0 | #if CONFIG_MULTITHREAD |
199 | 0 | AV1CdefRowSync *const cdef_row_mt = cdef_sync->cdef_row_mt; |
200 | 0 | pthread_mutex_lock(cdef_row_mt[row].row_mutex_); |
201 | 0 | pthread_cond_signal(cdef_row_mt[row].row_cond_); |
202 | 0 | cdef_row_mt[row].is_row_done = 1; |
203 | 0 | pthread_mutex_unlock(cdef_row_mt[row].row_mutex_); |
204 | | #else |
205 | | (void)cdef_sync; |
206 | | (void)row; |
207 | | #endif // CONFIG_MULTITHREAD |
208 | 0 | } |
209 | | |
210 | | static INLINE void sync_read(AV1LfSync *const lf_sync, int r, int c, |
211 | 0 | int plane) { |
212 | 0 | #if CONFIG_MULTITHREAD |
213 | 0 | const int nsync = lf_sync->sync_range; |
214 | |
|
215 | 0 | if (r && !(c & (nsync - 1))) { |
216 | 0 | pthread_mutex_t *const mutex = &lf_sync->mutex_[plane][r - 1]; |
217 | 0 | pthread_mutex_lock(mutex); |
218 | |
|
219 | 0 | while (c > lf_sync->cur_sb_col[plane][r - 1] - nsync) { |
220 | 0 | pthread_cond_wait(&lf_sync->cond_[plane][r - 1], mutex); |
221 | 0 | } |
222 | 0 | pthread_mutex_unlock(mutex); |
223 | 0 | } |
224 | | #else |
225 | | (void)lf_sync; |
226 | | (void)r; |
227 | | (void)c; |
228 | | (void)plane; |
229 | | #endif // CONFIG_MULTITHREAD |
230 | 0 | } |
231 | | |
232 | | static INLINE void sync_write(AV1LfSync *const lf_sync, int r, int c, |
233 | 0 | const int sb_cols, int plane) { |
234 | 0 | #if CONFIG_MULTITHREAD |
235 | 0 | const int nsync = lf_sync->sync_range; |
236 | 0 | int cur; |
237 | | // Only signal when there are enough filtered SB for next row to run. |
238 | 0 | int sig = 1; |
239 | |
|
240 | 0 | if (c < sb_cols - 1) { |
241 | 0 | cur = c; |
242 | 0 | if (c % nsync) sig = 0; |
243 | 0 | } else { |
244 | 0 | cur = sb_cols + nsync; |
245 | 0 | } |
246 | |
|
247 | 0 | if (sig) { |
248 | 0 | pthread_mutex_lock(&lf_sync->mutex_[plane][r]); |
249 | |
|
250 | 0 | lf_sync->cur_sb_col[plane][r] = cur; |
251 | |
|
252 | 0 | pthread_cond_broadcast(&lf_sync->cond_[plane][r]); |
253 | 0 | pthread_mutex_unlock(&lf_sync->mutex_[plane][r]); |
254 | 0 | } |
255 | | #else |
256 | | (void)lf_sync; |
257 | | (void)r; |
258 | | (void)c; |
259 | | (void)sb_cols; |
260 | | (void)plane; |
261 | | #endif // CONFIG_MULTITHREAD |
262 | 0 | } |
263 | | |
264 | | static void enqueue_lf_jobs(AV1LfSync *lf_sync, int start, int stop, |
265 | 0 | const int planes_to_lf[3], int is_realtime) { |
266 | 0 | int mi_row, plane, dir; |
267 | 0 | AV1LfMTInfo *lf_job_queue = lf_sync->job_queue; |
268 | 0 | lf_sync->jobs_enqueued = 0; |
269 | 0 | lf_sync->jobs_dequeued = 0; |
270 | | |
271 | | // Launch all vertical jobs first, as they are blocking the horizontal ones. |
272 | | // Launch top row jobs for all planes first, in case the output can be |
273 | | // partially reconstructed row by row. |
274 | 0 | for (dir = 0; dir < 2; ++dir) { |
275 | 0 | for (mi_row = start; mi_row < stop; mi_row += MAX_MIB_SIZE) { |
276 | 0 | for (plane = 0; plane < 3; ++plane) { |
277 | 0 | if (!planes_to_lf[plane]) continue; |
278 | 0 | lf_job_queue->mi_row = mi_row; |
279 | 0 | lf_job_queue->plane = plane; |
280 | 0 | lf_job_queue->dir = dir; |
281 | 0 | lf_job_queue->is_realtime = is_realtime; |
282 | 0 | lf_job_queue++; |
283 | 0 | lf_sync->jobs_enqueued++; |
284 | 0 | } |
285 | 0 | } |
286 | 0 | } |
287 | 0 | } |
288 | | |
289 | 0 | static AV1LfMTInfo *get_lf_job_info(AV1LfSync *lf_sync) { |
290 | 0 | AV1LfMTInfo *cur_job_info = NULL; |
291 | |
|
292 | 0 | #if CONFIG_MULTITHREAD |
293 | 0 | pthread_mutex_lock(lf_sync->job_mutex); |
294 | |
|
295 | 0 | if (lf_sync->jobs_dequeued < lf_sync->jobs_enqueued) { |
296 | 0 | cur_job_info = lf_sync->job_queue + lf_sync->jobs_dequeued; |
297 | 0 | lf_sync->jobs_dequeued++; |
298 | 0 | } |
299 | |
|
300 | 0 | pthread_mutex_unlock(lf_sync->job_mutex); |
301 | | #else |
302 | | (void)lf_sync; |
303 | | #endif |
304 | |
|
305 | 0 | return cur_job_info; |
306 | 0 | } |
307 | | |
308 | | // One job of row loopfiltering. |
309 | | static INLINE void thread_loop_filter_rows( |
310 | | const YV12_BUFFER_CONFIG *const frame_buffer, AV1_COMMON *const cm, |
311 | | struct macroblockd_plane *planes, MACROBLOCKD *xd, int mi_row, int plane, |
312 | 0 | int dir, int is_realtime, AV1LfSync *const lf_sync) { |
313 | 0 | const int sb_cols = |
314 | 0 | ALIGN_POWER_OF_TWO(cm->mi_params.mi_cols, MAX_MIB_SIZE_LOG2) >> |
315 | 0 | MAX_MIB_SIZE_LOG2; |
316 | 0 | const int r = mi_row >> MAX_MIB_SIZE_LOG2; |
317 | 0 | int mi_col, c; |
318 | |
|
319 | 0 | if (dir == 0) { |
320 | 0 | for (mi_col = 0; mi_col < cm->mi_params.mi_cols; mi_col += MAX_MIB_SIZE) { |
321 | 0 | c = mi_col >> MAX_MIB_SIZE_LOG2; |
322 | |
|
323 | 0 | av1_setup_dst_planes(planes, cm->seq_params->sb_size, frame_buffer, |
324 | 0 | mi_row, mi_col, plane, plane + 1); |
325 | 0 | #if CONFIG_AV1_HIGHBITDEPTH |
326 | 0 | (void)is_realtime; |
327 | 0 | av1_filter_block_plane_vert(cm, xd, plane, &planes[plane], mi_row, |
328 | 0 | mi_col); |
329 | | #else |
330 | | if (is_realtime) { |
331 | | av1_filter_block_plane_vert_rt(cm, xd, plane, &planes[plane], mi_row, |
332 | | mi_col); |
333 | | |
334 | | } else { |
335 | | av1_filter_block_plane_vert(cm, xd, plane, &planes[plane], mi_row, |
336 | | mi_col); |
337 | | } |
338 | | #endif |
339 | 0 | if (lf_sync != NULL) sync_write(lf_sync, r, c, sb_cols, plane); |
340 | 0 | } |
341 | 0 | } else if (dir == 1) { |
342 | 0 | for (mi_col = 0; mi_col < cm->mi_params.mi_cols; mi_col += MAX_MIB_SIZE) { |
343 | 0 | c = mi_col >> MAX_MIB_SIZE_LOG2; |
344 | |
|
345 | 0 | if (lf_sync != NULL) { |
346 | | // Wait for vertical edge filtering of the top-right block to be |
347 | | // completed |
348 | 0 | sync_read(lf_sync, r, c, plane); |
349 | | |
350 | | // Wait for vertical edge filtering of the right block to be completed |
351 | 0 | sync_read(lf_sync, r + 1, c, plane); |
352 | 0 | } |
353 | |
|
354 | 0 | av1_setup_dst_planes(planes, cm->seq_params->sb_size, frame_buffer, |
355 | 0 | mi_row, mi_col, plane, plane + 1); |
356 | 0 | #if CONFIG_AV1_HIGHBITDEPTH |
357 | 0 | (void)is_realtime; |
358 | 0 | av1_filter_block_plane_horz(cm, xd, plane, &planes[plane], mi_row, |
359 | 0 | mi_col); |
360 | | #else |
361 | | if (is_realtime) { |
362 | | av1_filter_block_plane_horz_rt(cm, xd, plane, &planes[plane], mi_row, |
363 | | mi_col); |
364 | | } else { |
365 | | av1_filter_block_plane_horz(cm, xd, plane, &planes[plane], mi_row, |
366 | | mi_col); |
367 | | } |
368 | | #endif |
369 | 0 | } |
370 | 0 | } |
371 | 0 | } |
372 | | |
373 | | // Row-based multi-threaded loopfilter hook |
374 | 0 | static int loop_filter_row_worker(void *arg1, void *arg2) { |
375 | 0 | AV1LfSync *const lf_sync = (AV1LfSync *)arg1; |
376 | 0 | LFWorkerData *const lf_data = (LFWorkerData *)arg2; |
377 | 0 | AV1LfMTInfo *cur_job_info; |
378 | 0 | while ((cur_job_info = get_lf_job_info(lf_sync)) != NULL) { |
379 | 0 | const int is_realtime = cur_job_info->is_realtime && !cur_job_info->plane; |
380 | 0 | thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes, |
381 | 0 | lf_data->xd, cur_job_info->mi_row, |
382 | 0 | cur_job_info->plane, cur_job_info->dir, is_realtime, |
383 | 0 | lf_sync); |
384 | 0 | } |
385 | 0 | return 1; |
386 | 0 | } |
387 | | |
388 | | static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm, |
389 | | MACROBLOCKD *xd, int start, int stop, |
390 | | const int planes_to_lf[3], AVxWorker *workers, |
391 | | int num_workers, AV1LfSync *lf_sync, |
392 | 0 | int is_realtime) { |
393 | 0 | const AVxWorkerInterface *const winterface = aom_get_worker_interface(); |
394 | | // Number of superblock rows and cols |
395 | 0 | const int sb_rows = |
396 | 0 | ALIGN_POWER_OF_TWO(cm->mi_params.mi_rows, MAX_MIB_SIZE_LOG2) >> |
397 | 0 | MAX_MIB_SIZE_LOG2; |
398 | 0 | int i; |
399 | |
|
400 | 0 | if (!lf_sync->sync_range || sb_rows != lf_sync->rows || |
401 | 0 | num_workers > lf_sync->num_workers) { |
402 | 0 | av1_loop_filter_dealloc(lf_sync); |
403 | 0 | av1_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers); |
404 | 0 | } |
405 | | |
406 | | // Initialize cur_sb_col to -1 for all SB rows. |
407 | 0 | for (i = 0; i < MAX_MB_PLANE; i++) { |
408 | 0 | memset(lf_sync->cur_sb_col[i], -1, |
409 | 0 | sizeof(*(lf_sync->cur_sb_col[i])) * sb_rows); |
410 | 0 | } |
411 | |
|
412 | 0 | enqueue_lf_jobs(lf_sync, start, stop, planes_to_lf, is_realtime); |
413 | | |
414 | | // Set up loopfilter thread data. |
415 | 0 | for (i = num_workers - 1; i >= 0; --i) { |
416 | 0 | AVxWorker *const worker = &workers[i]; |
417 | 0 | LFWorkerData *const lf_data = &lf_sync->lfdata[i]; |
418 | |
|
419 | 0 | worker->hook = loop_filter_row_worker; |
420 | 0 | worker->data1 = lf_sync; |
421 | 0 | worker->data2 = lf_data; |
422 | | |
423 | | // Loopfilter data |
424 | 0 | loop_filter_data_reset(lf_data, frame, cm, xd); |
425 | | |
426 | | // Start loopfiltering |
427 | 0 | if (i == 0) { |
428 | 0 | winterface->execute(worker); |
429 | 0 | } else { |
430 | 0 | winterface->launch(worker); |
431 | 0 | } |
432 | 0 | } |
433 | | |
434 | | // Wait till all rows are finished |
435 | 0 | for (i = 1; i < num_workers; ++i) { |
436 | 0 | winterface->sync(&workers[i]); |
437 | 0 | } |
438 | 0 | } |
439 | | |
440 | | static void loop_filter_rows(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm, |
441 | | MACROBLOCKD *xd, int start, int stop, |
442 | 0 | const int planes_to_lf[3], int is_realtime) { |
443 | | // Filter top rows of all planes first, in case the output can be partially |
444 | | // reconstructed row by row. |
445 | 0 | int mi_row, plane, dir; |
446 | 0 | for (mi_row = start; mi_row < stop; mi_row += MAX_MIB_SIZE) { |
447 | 0 | for (plane = 0; plane < 3; ++plane) { |
448 | 0 | if (!planes_to_lf[plane]) continue; |
449 | 0 | for (dir = 0; dir < 2; ++dir) { |
450 | 0 | thread_loop_filter_rows(frame, cm, xd->plane, xd, mi_row, plane, dir, |
451 | 0 | is_realtime && !plane, /*lf_sync=*/NULL); |
452 | 0 | } |
453 | 0 | } |
454 | 0 | } |
455 | 0 | } |
456 | | |
457 | | void av1_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm, |
458 | | MACROBLOCKD *xd, int plane_start, int plane_end, |
459 | | int partial_frame, AVxWorker *workers, |
460 | | int num_workers, AV1LfSync *lf_sync, |
461 | 0 | int is_realtime) { |
462 | 0 | int start_mi_row, end_mi_row, mi_rows_to_filter; |
463 | 0 | int planes_to_lf[3]; |
464 | | |
465 | | // For each luma and chroma plane, whether to filter it or not. |
466 | 0 | planes_to_lf[0] = (cm->lf.filter_level[0] || cm->lf.filter_level[1]) && |
467 | 0 | plane_start <= 0 && 0 < plane_end; |
468 | 0 | planes_to_lf[1] = cm->lf.filter_level_u && plane_start <= 1 && 1 < plane_end; |
469 | 0 | planes_to_lf[2] = cm->lf.filter_level_v && plane_start <= 2 && 2 < plane_end; |
470 | | // If the luma plane is purposely not filtered, neither are the chroma planes. |
471 | 0 | if (!planes_to_lf[0] && plane_start <= 0 && 0 < plane_end) return; |
472 | | // Early exit. |
473 | 0 | if (!planes_to_lf[0] && !planes_to_lf[1] && !planes_to_lf[2]) return; |
474 | | |
475 | 0 | start_mi_row = 0; |
476 | 0 | mi_rows_to_filter = cm->mi_params.mi_rows; |
477 | 0 | if (partial_frame && cm->mi_params.mi_rows > 8) { |
478 | 0 | start_mi_row = cm->mi_params.mi_rows >> 1; |
479 | 0 | start_mi_row &= 0xfffffff8; |
480 | 0 | mi_rows_to_filter = AOMMAX(cm->mi_params.mi_rows / 8, 8); |
481 | 0 | } |
482 | 0 | end_mi_row = start_mi_row + mi_rows_to_filter; |
483 | 0 | av1_loop_filter_frame_init(cm, plane_start, plane_end); |
484 | |
|
485 | 0 | if (num_workers > 1) { |
486 | | // Enqueue and execute loopfiltering jobs. |
487 | 0 | loop_filter_rows_mt(frame, cm, xd, start_mi_row, end_mi_row, planes_to_lf, |
488 | 0 | workers, num_workers, lf_sync, is_realtime); |
489 | 0 | } else { |
490 | | // Directly filter in the main thread. |
491 | 0 | loop_filter_rows(frame, cm, xd, start_mi_row, end_mi_row, planes_to_lf, |
492 | 0 | is_realtime); |
493 | 0 | } |
494 | 0 | } |
495 | | |
496 | | #if !CONFIG_REALTIME_ONLY |
497 | 0 | static INLINE void lr_sync_read(void *const lr_sync, int r, int c, int plane) { |
498 | 0 | #if CONFIG_MULTITHREAD |
499 | 0 | AV1LrSync *const loop_res_sync = (AV1LrSync *)lr_sync; |
500 | 0 | const int nsync = loop_res_sync->sync_range; |
501 | |
|
502 | 0 | if (r && !(c & (nsync - 1))) { |
503 | 0 | pthread_mutex_t *const mutex = &loop_res_sync->mutex_[plane][r - 1]; |
504 | 0 | pthread_mutex_lock(mutex); |
505 | |
|
506 | 0 | while (c > loop_res_sync->cur_sb_col[plane][r - 1] - nsync) { |
507 | 0 | pthread_cond_wait(&loop_res_sync->cond_[plane][r - 1], mutex); |
508 | 0 | } |
509 | 0 | pthread_mutex_unlock(mutex); |
510 | 0 | } |
511 | | #else |
512 | | (void)lr_sync; |
513 | | (void)r; |
514 | | (void)c; |
515 | | (void)plane; |
516 | | #endif // CONFIG_MULTITHREAD |
517 | 0 | } |
518 | | |
519 | | static INLINE void lr_sync_write(void *const lr_sync, int r, int c, |
520 | 0 | const int sb_cols, int plane) { |
521 | 0 | #if CONFIG_MULTITHREAD |
522 | 0 | AV1LrSync *const loop_res_sync = (AV1LrSync *)lr_sync; |
523 | 0 | const int nsync = loop_res_sync->sync_range; |
524 | 0 | int cur; |
525 | | // Only signal when there are enough filtered SB for next row to run. |
526 | 0 | int sig = 1; |
527 | |
|
528 | 0 | if (c < sb_cols - 1) { |
529 | 0 | cur = c; |
530 | 0 | if (c % nsync) sig = 0; |
531 | 0 | } else { |
532 | 0 | cur = sb_cols + nsync; |
533 | 0 | } |
534 | |
|
535 | 0 | if (sig) { |
536 | 0 | pthread_mutex_lock(&loop_res_sync->mutex_[plane][r]); |
537 | |
|
538 | 0 | loop_res_sync->cur_sb_col[plane][r] = cur; |
539 | |
|
540 | 0 | pthread_cond_broadcast(&loop_res_sync->cond_[plane][r]); |
541 | 0 | pthread_mutex_unlock(&loop_res_sync->mutex_[plane][r]); |
542 | 0 | } |
543 | | #else |
544 | | (void)lr_sync; |
545 | | (void)r; |
546 | | (void)c; |
547 | | (void)sb_cols; |
548 | | (void)plane; |
549 | | #endif // CONFIG_MULTITHREAD |
550 | 0 | } |
551 | | |
552 | | // Allocate memory for loop restoration row synchronization |
553 | | void av1_loop_restoration_alloc(AV1LrSync *lr_sync, AV1_COMMON *cm, |
554 | | int num_workers, int num_rows_lr, |
555 | 0 | int num_planes, int width) { |
556 | 0 | lr_sync->rows = num_rows_lr; |
557 | 0 | lr_sync->num_planes = num_planes; |
558 | 0 | #if CONFIG_MULTITHREAD |
559 | 0 | { |
560 | 0 | int i, j; |
561 | |
|
562 | 0 | for (j = 0; j < num_planes; j++) { |
563 | 0 | CHECK_MEM_ERROR(cm, lr_sync->mutex_[j], |
564 | 0 | aom_malloc(sizeof(*(lr_sync->mutex_[j])) * num_rows_lr)); |
565 | 0 | if (lr_sync->mutex_[j]) { |
566 | 0 | for (i = 0; i < num_rows_lr; ++i) { |
567 | 0 | pthread_mutex_init(&lr_sync->mutex_[j][i], NULL); |
568 | 0 | } |
569 | 0 | } |
570 | |
|
571 | 0 | CHECK_MEM_ERROR(cm, lr_sync->cond_[j], |
572 | 0 | aom_malloc(sizeof(*(lr_sync->cond_[j])) * num_rows_lr)); |
573 | 0 | if (lr_sync->cond_[j]) { |
574 | 0 | for (i = 0; i < num_rows_lr; ++i) { |
575 | 0 | pthread_cond_init(&lr_sync->cond_[j][i], NULL); |
576 | 0 | } |
577 | 0 | } |
578 | 0 | } |
579 | |
|
580 | 0 | CHECK_MEM_ERROR(cm, lr_sync->job_mutex, |
581 | 0 | aom_malloc(sizeof(*(lr_sync->job_mutex)))); |
582 | 0 | if (lr_sync->job_mutex) { |
583 | 0 | pthread_mutex_init(lr_sync->job_mutex, NULL); |
584 | 0 | } |
585 | 0 | } |
586 | 0 | #endif // CONFIG_MULTITHREAD |
587 | 0 | CHECK_MEM_ERROR(cm, lr_sync->lrworkerdata, |
588 | 0 | aom_malloc(num_workers * sizeof(*(lr_sync->lrworkerdata)))); |
589 | |
|
590 | 0 | for (int worker_idx = 0; worker_idx < num_workers; ++worker_idx) { |
591 | 0 | if (worker_idx < num_workers - 1) { |
592 | 0 | CHECK_MEM_ERROR(cm, lr_sync->lrworkerdata[worker_idx].rst_tmpbuf, |
593 | 0 | (int32_t *)aom_memalign(16, RESTORATION_TMPBUF_SIZE)); |
594 | 0 | CHECK_MEM_ERROR(cm, lr_sync->lrworkerdata[worker_idx].rlbs, |
595 | 0 | aom_malloc(sizeof(RestorationLineBuffers))); |
596 | |
|
597 | 0 | } else { |
598 | 0 | lr_sync->lrworkerdata[worker_idx].rst_tmpbuf = cm->rst_tmpbuf; |
599 | 0 | lr_sync->lrworkerdata[worker_idx].rlbs = cm->rlbs; |
600 | 0 | } |
601 | 0 | } |
602 | |
|
603 | 0 | lr_sync->num_workers = num_workers; |
604 | |
|
605 | 0 | for (int j = 0; j < num_planes; j++) { |
606 | 0 | CHECK_MEM_ERROR( |
607 | 0 | cm, lr_sync->cur_sb_col[j], |
608 | 0 | aom_malloc(sizeof(*(lr_sync->cur_sb_col[j])) * num_rows_lr)); |
609 | 0 | } |
610 | 0 | CHECK_MEM_ERROR( |
611 | 0 | cm, lr_sync->job_queue, |
612 | 0 | aom_malloc(sizeof(*(lr_sync->job_queue)) * num_rows_lr * num_planes)); |
613 | | // Set up nsync. |
614 | 0 | lr_sync->sync_range = get_lr_sync_range(width); |
615 | 0 | } |
616 | | |
617 | | // Deallocate loop restoration synchronization related mutex and data |
618 | 0 | void av1_loop_restoration_dealloc(AV1LrSync *lr_sync, int num_workers) { |
619 | 0 | if (lr_sync != NULL) { |
620 | 0 | int j; |
621 | 0 | #if CONFIG_MULTITHREAD |
622 | 0 | int i; |
623 | 0 | for (j = 0; j < MAX_MB_PLANE; j++) { |
624 | 0 | if (lr_sync->mutex_[j] != NULL) { |
625 | 0 | for (i = 0; i < lr_sync->rows; ++i) { |
626 | 0 | pthread_mutex_destroy(&lr_sync->mutex_[j][i]); |
627 | 0 | } |
628 | 0 | aom_free(lr_sync->mutex_[j]); |
629 | 0 | } |
630 | 0 | if (lr_sync->cond_[j] != NULL) { |
631 | 0 | for (i = 0; i < lr_sync->rows; ++i) { |
632 | 0 | pthread_cond_destroy(&lr_sync->cond_[j][i]); |
633 | 0 | } |
634 | 0 | aom_free(lr_sync->cond_[j]); |
635 | 0 | } |
636 | 0 | } |
637 | 0 | if (lr_sync->job_mutex != NULL) { |
638 | 0 | pthread_mutex_destroy(lr_sync->job_mutex); |
639 | 0 | aom_free(lr_sync->job_mutex); |
640 | 0 | } |
641 | 0 | #endif // CONFIG_MULTITHREAD |
642 | 0 | for (j = 0; j < MAX_MB_PLANE; j++) { |
643 | 0 | aom_free(lr_sync->cur_sb_col[j]); |
644 | 0 | } |
645 | |
|
646 | 0 | aom_free(lr_sync->job_queue); |
647 | |
|
648 | 0 | if (lr_sync->lrworkerdata) { |
649 | 0 | for (int worker_idx = 0; worker_idx < num_workers - 1; worker_idx++) { |
650 | 0 | LRWorkerData *const workerdata_data = |
651 | 0 | lr_sync->lrworkerdata + worker_idx; |
652 | |
|
653 | 0 | aom_free(workerdata_data->rst_tmpbuf); |
654 | 0 | aom_free(workerdata_data->rlbs); |
655 | 0 | } |
656 | 0 | aom_free(lr_sync->lrworkerdata); |
657 | 0 | } |
658 | | |
659 | | // clear the structure as the source of this call may be a resize in which |
660 | | // case this call will be followed by an _alloc() which may fail. |
661 | 0 | av1_zero(*lr_sync); |
662 | 0 | } |
663 | 0 | } |
664 | | |
665 | | static void enqueue_lr_jobs(AV1LrSync *lr_sync, AV1LrStruct *lr_ctxt, |
666 | 0 | AV1_COMMON *cm) { |
667 | 0 | FilterFrameCtxt *ctxt = lr_ctxt->ctxt; |
668 | |
|
669 | 0 | const int num_planes = av1_num_planes(cm); |
670 | 0 | AV1LrMTInfo *lr_job_queue = lr_sync->job_queue; |
671 | 0 | int32_t lr_job_counter[2], num_even_lr_jobs = 0; |
672 | 0 | lr_sync->jobs_enqueued = 0; |
673 | 0 | lr_sync->jobs_dequeued = 0; |
674 | |
|
675 | 0 | for (int plane = 0; plane < num_planes; plane++) { |
676 | 0 | if (cm->rst_info[plane].frame_restoration_type == RESTORE_NONE) continue; |
677 | 0 | num_even_lr_jobs = |
678 | 0 | num_even_lr_jobs + ((ctxt[plane].rsi->vert_units_per_tile + 1) >> 1); |
679 | 0 | } |
680 | 0 | lr_job_counter[0] = 0; |
681 | 0 | lr_job_counter[1] = num_even_lr_jobs; |
682 | |
|
683 | 0 | for (int plane = 0; plane < num_planes; plane++) { |
684 | 0 | if (cm->rst_info[plane].frame_restoration_type == RESTORE_NONE) continue; |
685 | 0 | const int is_uv = plane > 0; |
686 | 0 | const int ss_y = is_uv && cm->seq_params->subsampling_y; |
687 | |
|
688 | 0 | AV1PixelRect tile_rect = ctxt[plane].tile_rect; |
689 | 0 | const int unit_size = ctxt[plane].rsi->restoration_unit_size; |
690 | |
|
691 | 0 | const int tile_h = tile_rect.bottom - tile_rect.top; |
692 | 0 | const int ext_size = unit_size * 3 / 2; |
693 | |
|
694 | 0 | int y0 = 0, i = 0; |
695 | 0 | while (y0 < tile_h) { |
696 | 0 | int remaining_h = tile_h - y0; |
697 | 0 | int h = (remaining_h < ext_size) ? remaining_h : unit_size; |
698 | |
|
699 | 0 | RestorationTileLimits limits; |
700 | 0 | limits.v_start = tile_rect.top + y0; |
701 | 0 | limits.v_end = tile_rect.top + y0 + h; |
702 | 0 | assert(limits.v_end <= tile_rect.bottom); |
703 | | // Offset the tile upwards to align with the restoration processing stripe |
704 | 0 | const int voffset = RESTORATION_UNIT_OFFSET >> ss_y; |
705 | 0 | limits.v_start = AOMMAX(tile_rect.top, limits.v_start - voffset); |
706 | 0 | if (limits.v_end < tile_rect.bottom) limits.v_end -= voffset; |
707 | |
|
708 | 0 | assert(lr_job_counter[0] <= num_even_lr_jobs); |
709 | |
|
710 | 0 | lr_job_queue[lr_job_counter[i & 1]].lr_unit_row = i; |
711 | 0 | lr_job_queue[lr_job_counter[i & 1]].plane = plane; |
712 | 0 | lr_job_queue[lr_job_counter[i & 1]].v_start = limits.v_start; |
713 | 0 | lr_job_queue[lr_job_counter[i & 1]].v_end = limits.v_end; |
714 | 0 | lr_job_queue[lr_job_counter[i & 1]].sync_mode = i & 1; |
715 | 0 | if ((i & 1) == 0) { |
716 | 0 | lr_job_queue[lr_job_counter[i & 1]].v_copy_start = |
717 | 0 | limits.v_start + RESTORATION_BORDER; |
718 | 0 | lr_job_queue[lr_job_counter[i & 1]].v_copy_end = |
719 | 0 | limits.v_end - RESTORATION_BORDER; |
720 | 0 | if (i == 0) { |
721 | 0 | assert(limits.v_start == tile_rect.top); |
722 | 0 | lr_job_queue[lr_job_counter[i & 1]].v_copy_start = tile_rect.top; |
723 | 0 | } |
724 | 0 | if (i == (ctxt[plane].rsi->vert_units_per_tile - 1)) { |
725 | 0 | assert(limits.v_end == tile_rect.bottom); |
726 | 0 | lr_job_queue[lr_job_counter[i & 1]].v_copy_end = tile_rect.bottom; |
727 | 0 | } |
728 | 0 | } else { |
729 | 0 | lr_job_queue[lr_job_counter[i & 1]].v_copy_start = |
730 | 0 | AOMMAX(limits.v_start - RESTORATION_BORDER, tile_rect.top); |
731 | 0 | lr_job_queue[lr_job_counter[i & 1]].v_copy_end = |
732 | 0 | AOMMIN(limits.v_end + RESTORATION_BORDER, tile_rect.bottom); |
733 | 0 | } |
734 | 0 | lr_job_counter[i & 1]++; |
735 | 0 | lr_sync->jobs_enqueued++; |
736 | |
|
737 | 0 | y0 += h; |
738 | 0 | ++i; |
739 | 0 | } |
740 | 0 | } |
741 | 0 | } |
742 | | |
743 | 0 | static AV1LrMTInfo *get_lr_job_info(AV1LrSync *lr_sync) { |
744 | 0 | AV1LrMTInfo *cur_job_info = NULL; |
745 | |
|
746 | 0 | #if CONFIG_MULTITHREAD |
747 | 0 | pthread_mutex_lock(lr_sync->job_mutex); |
748 | |
|
749 | 0 | if (lr_sync->jobs_dequeued < lr_sync->jobs_enqueued) { |
750 | 0 | cur_job_info = lr_sync->job_queue + lr_sync->jobs_dequeued; |
751 | 0 | lr_sync->jobs_dequeued++; |
752 | 0 | } |
753 | |
|
754 | 0 | pthread_mutex_unlock(lr_sync->job_mutex); |
755 | | #else |
756 | | (void)lr_sync; |
757 | | #endif |
758 | |
|
759 | 0 | return cur_job_info; |
760 | 0 | } |
761 | | |
762 | | // Implement row loop restoration for each thread. |
763 | 0 | static int loop_restoration_row_worker(void *arg1, void *arg2) { |
764 | 0 | AV1LrSync *const lr_sync = (AV1LrSync *)arg1; |
765 | 0 | LRWorkerData *lrworkerdata = (LRWorkerData *)arg2; |
766 | 0 | AV1LrStruct *lr_ctxt = (AV1LrStruct *)lrworkerdata->lr_ctxt; |
767 | 0 | FilterFrameCtxt *ctxt = lr_ctxt->ctxt; |
768 | 0 | int lr_unit_row; |
769 | 0 | int plane; |
770 | 0 | const int tile_row = LR_TILE_ROW; |
771 | 0 | const int tile_col = LR_TILE_COL; |
772 | 0 | const int tile_cols = LR_TILE_COLS; |
773 | 0 | const int tile_idx = tile_col + tile_row * tile_cols; |
774 | 0 | typedef void (*copy_fun)(const YV12_BUFFER_CONFIG *src_ybc, |
775 | 0 | YV12_BUFFER_CONFIG *dst_ybc, int hstart, int hend, |
776 | 0 | int vstart, int vend); |
777 | 0 | static const copy_fun copy_funs[3] = { aom_yv12_partial_coloc_copy_y, |
778 | 0 | aom_yv12_partial_coloc_copy_u, |
779 | 0 | aom_yv12_partial_coloc_copy_v }; |
780 | |
|
781 | 0 | while (1) { |
782 | 0 | AV1LrMTInfo *cur_job_info = get_lr_job_info(lr_sync); |
783 | 0 | if (cur_job_info != NULL) { |
784 | 0 | RestorationTileLimits limits; |
785 | 0 | sync_read_fn_t on_sync_read; |
786 | 0 | sync_write_fn_t on_sync_write; |
787 | 0 | limits.v_start = cur_job_info->v_start; |
788 | 0 | limits.v_end = cur_job_info->v_end; |
789 | 0 | lr_unit_row = cur_job_info->lr_unit_row; |
790 | 0 | plane = cur_job_info->plane; |
791 | 0 | const int unit_idx0 = tile_idx * ctxt[plane].rsi->units_per_tile; |
792 | | |
793 | | // sync_mode == 1 implies only sync read is required in LR Multi-threading |
794 | | // sync_mode == 0 implies only sync write is required. |
795 | 0 | on_sync_read = |
796 | 0 | cur_job_info->sync_mode == 1 ? lr_sync_read : av1_lr_sync_read_dummy; |
797 | 0 | on_sync_write = cur_job_info->sync_mode == 0 ? lr_sync_write |
798 | 0 | : av1_lr_sync_write_dummy; |
799 | |
|
800 | 0 | av1_foreach_rest_unit_in_row( |
801 | 0 | &limits, &(ctxt[plane].tile_rect), lr_ctxt->on_rest_unit, lr_unit_row, |
802 | 0 | ctxt[plane].rsi->restoration_unit_size, unit_idx0, |
803 | 0 | ctxt[plane].rsi->horz_units_per_tile, |
804 | 0 | ctxt[plane].rsi->vert_units_per_tile, plane, &ctxt[plane], |
805 | 0 | lrworkerdata->rst_tmpbuf, lrworkerdata->rlbs, on_sync_read, |
806 | 0 | on_sync_write, lr_sync); |
807 | |
|
808 | 0 | copy_funs[plane](lr_ctxt->dst, lr_ctxt->frame, ctxt[plane].tile_rect.left, |
809 | 0 | ctxt[plane].tile_rect.right, cur_job_info->v_copy_start, |
810 | 0 | cur_job_info->v_copy_end); |
811 | 0 | } else { |
812 | 0 | break; |
813 | 0 | } |
814 | 0 | } |
815 | 0 | return 1; |
816 | 0 | } |
817 | | |
818 | | static void foreach_rest_unit_in_planes_mt(AV1LrStruct *lr_ctxt, |
819 | | AVxWorker *workers, int nworkers, |
820 | 0 | AV1LrSync *lr_sync, AV1_COMMON *cm) { |
821 | 0 | FilterFrameCtxt *ctxt = lr_ctxt->ctxt; |
822 | |
|
823 | 0 | const int num_planes = av1_num_planes(cm); |
824 | |
|
825 | 0 | const AVxWorkerInterface *const winterface = aom_get_worker_interface(); |
826 | 0 | int num_rows_lr = 0; |
827 | |
|
828 | 0 | for (int plane = 0; plane < num_planes; plane++) { |
829 | 0 | if (cm->rst_info[plane].frame_restoration_type == RESTORE_NONE) continue; |
830 | | |
831 | 0 | const AV1PixelRect tile_rect = ctxt[plane].tile_rect; |
832 | 0 | const int max_tile_h = tile_rect.bottom - tile_rect.top; |
833 | |
|
834 | 0 | const int unit_size = cm->rst_info[plane].restoration_unit_size; |
835 | |
|
836 | 0 | num_rows_lr = |
837 | 0 | AOMMAX(num_rows_lr, av1_lr_count_units_in_tile(unit_size, max_tile_h)); |
838 | 0 | } |
839 | |
|
840 | 0 | const int num_workers = nworkers; |
841 | 0 | int i; |
842 | 0 | assert(MAX_MB_PLANE == 3); |
843 | |
|
844 | 0 | if (!lr_sync->sync_range || num_rows_lr > lr_sync->rows || |
845 | 0 | num_workers > lr_sync->num_workers || num_planes > lr_sync->num_planes) { |
846 | 0 | av1_loop_restoration_dealloc(lr_sync, num_workers); |
847 | 0 | av1_loop_restoration_alloc(lr_sync, cm, num_workers, num_rows_lr, |
848 | 0 | num_planes, cm->width); |
849 | 0 | } |
850 | | |
851 | | // Initialize cur_sb_col to -1 for all SB rows. |
852 | 0 | for (i = 0; i < num_planes; i++) { |
853 | 0 | memset(lr_sync->cur_sb_col[i], -1, |
854 | 0 | sizeof(*(lr_sync->cur_sb_col[i])) * num_rows_lr); |
855 | 0 | } |
856 | |
|
857 | 0 | enqueue_lr_jobs(lr_sync, lr_ctxt, cm); |
858 | | |
859 | | // Set up looprestoration thread data. |
860 | 0 | for (i = num_workers - 1; i >= 0; --i) { |
861 | 0 | AVxWorker *const worker = &workers[i]; |
862 | 0 | lr_sync->lrworkerdata[i].lr_ctxt = (void *)lr_ctxt; |
863 | 0 | worker->hook = loop_restoration_row_worker; |
864 | 0 | worker->data1 = lr_sync; |
865 | 0 | worker->data2 = &lr_sync->lrworkerdata[i]; |
866 | | |
867 | | // Start loop restoration |
868 | 0 | if (i == 0) { |
869 | 0 | winterface->execute(worker); |
870 | 0 | } else { |
871 | 0 | winterface->launch(worker); |
872 | 0 | } |
873 | 0 | } |
874 | | |
875 | | // Wait till all rows are finished |
876 | 0 | for (i = 1; i < num_workers; ++i) { |
877 | 0 | winterface->sync(&workers[i]); |
878 | 0 | } |
879 | 0 | } |
880 | | |
881 | | void av1_loop_restoration_filter_frame_mt(YV12_BUFFER_CONFIG *frame, |
882 | | AV1_COMMON *cm, int optimized_lr, |
883 | | AVxWorker *workers, int num_workers, |
884 | 0 | AV1LrSync *lr_sync, void *lr_ctxt) { |
885 | 0 | assert(!cm->features.all_lossless); |
886 | |
|
887 | 0 | const int num_planes = av1_num_planes(cm); |
888 | |
|
889 | 0 | AV1LrStruct *loop_rest_ctxt = (AV1LrStruct *)lr_ctxt; |
890 | |
|
891 | 0 | av1_loop_restoration_filter_frame_init(loop_rest_ctxt, frame, cm, |
892 | 0 | optimized_lr, num_planes); |
893 | |
|
894 | 0 | foreach_rest_unit_in_planes_mt(loop_rest_ctxt, workers, num_workers, lr_sync, |
895 | 0 | cm); |
896 | 0 | } |
897 | | #endif |
898 | | |
899 | | // Initializes cdef_sync parameters. |
900 | 0 | static AOM_INLINE void reset_cdef_job_info(AV1CdefSync *const cdef_sync) { |
901 | 0 | cdef_sync->end_of_frame = 0; |
902 | 0 | cdef_sync->fbr = 0; |
903 | 0 | cdef_sync->fbc = 0; |
904 | 0 | } |
905 | | |
906 | | static AOM_INLINE void launch_cdef_workers(AVxWorker *const workers, |
907 | 0 | int num_workers) { |
908 | 0 | const AVxWorkerInterface *const winterface = aom_get_worker_interface(); |
909 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
910 | 0 | AVxWorker *const worker = &workers[i]; |
911 | 0 | if (i == 0) |
912 | 0 | winterface->execute(worker); |
913 | 0 | else |
914 | 0 | winterface->launch(worker); |
915 | 0 | } |
916 | 0 | } |
917 | | |
918 | | static AOM_INLINE void sync_cdef_workers(AVxWorker *const workers, |
919 | | AV1_COMMON *const cm, |
920 | 0 | int num_workers) { |
921 | 0 | const AVxWorkerInterface *const winterface = aom_get_worker_interface(); |
922 | 0 | int had_error = 0; |
923 | | |
924 | | // Wait for completion of Cdef frame. |
925 | 0 | for (int i = num_workers - 1; i > 0; i--) { |
926 | 0 | AVxWorker *const worker = &workers[i]; |
927 | 0 | had_error |= !winterface->sync(worker); |
928 | 0 | } |
929 | 0 | if (had_error) |
930 | 0 | aom_internal_error(cm->error, AOM_CODEC_ERROR, |
931 | 0 | "Failed to process cdef frame"); |
932 | 0 | } |
933 | | |
934 | | // Updates the row index of the next job to be processed. |
935 | | // Also updates end_of_frame flag when the processing of all rows is complete. |
936 | | static void update_cdef_row_next_job_info(AV1CdefSync *const cdef_sync, |
937 | 0 | const int nvfb) { |
938 | 0 | cdef_sync->fbr++; |
939 | 0 | if (cdef_sync->fbr == nvfb) { |
940 | 0 | cdef_sync->end_of_frame = 1; |
941 | 0 | } |
942 | 0 | } |
943 | | |
944 | | // Checks if a job is available. If job is available, |
945 | | // populates next job information and returns 1, else returns 0. |
946 | | static AOM_INLINE int get_cdef_row_next_job(AV1CdefSync *const cdef_sync, |
947 | 0 | int *cur_fbr, const int nvfb) { |
948 | 0 | #if CONFIG_MULTITHREAD |
949 | 0 | pthread_mutex_lock(cdef_sync->mutex_); |
950 | 0 | #endif // CONFIG_MULTITHREAD |
951 | 0 | int do_next_row = 0; |
952 | | // Populates information needed for current job and update the row |
953 | | // index of the next row to be processed. |
954 | 0 | if (cdef_sync->end_of_frame == 0) { |
955 | 0 | do_next_row = 1; |
956 | 0 | *cur_fbr = cdef_sync->fbr; |
957 | 0 | update_cdef_row_next_job_info(cdef_sync, nvfb); |
958 | 0 | } |
959 | 0 | #if CONFIG_MULTITHREAD |
960 | 0 | pthread_mutex_unlock(cdef_sync->mutex_); |
961 | 0 | #endif // CONFIG_MULTITHREAD |
962 | 0 | return do_next_row; |
963 | 0 | } |
964 | | |
965 | | // Hook function for each thread in CDEF multi-threading. |
966 | 0 | static int cdef_sb_row_worker_hook(void *arg1, void *arg2) { |
967 | 0 | AV1CdefSync *const cdef_sync = (AV1CdefSync *)arg1; |
968 | 0 | AV1CdefWorkerData *const cdef_worker = (AV1CdefWorkerData *)arg2; |
969 | 0 | const int nvfb = |
970 | 0 | (cdef_worker->cm->mi_params.mi_rows + MI_SIZE_64X64 - 1) / MI_SIZE_64X64; |
971 | 0 | int cur_fbr; |
972 | 0 | while (get_cdef_row_next_job(cdef_sync, &cur_fbr, nvfb)) { |
973 | 0 | av1_cdef_fb_row(cdef_worker->cm, cdef_worker->xd, cdef_worker->linebuf, |
974 | 0 | cdef_worker->colbuf, cdef_worker->srcbuf, cur_fbr, |
975 | 0 | cdef_worker->cdef_init_fb_row_fn, cdef_sync); |
976 | 0 | } |
977 | 0 | return 1; |
978 | 0 | } |
979 | | |
980 | | // Assigns CDEF hook function and thread data to each worker. |
981 | | static void prepare_cdef_frame_workers( |
982 | | AV1_COMMON *const cm, MACROBLOCKD *xd, AV1CdefWorkerData *const cdef_worker, |
983 | | AVxWorkerHook hook, AVxWorker *const workers, AV1CdefSync *const cdef_sync, |
984 | 0 | int num_workers, cdef_init_fb_row_t cdef_init_fb_row_fn) { |
985 | 0 | const int num_planes = av1_num_planes(cm); |
986 | |
|
987 | 0 | cdef_worker[0].srcbuf = cm->cdef_info.srcbuf; |
988 | 0 | for (int plane = 0; plane < num_planes; plane++) |
989 | 0 | cdef_worker[0].colbuf[plane] = cm->cdef_info.colbuf[plane]; |
990 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
991 | 0 | AVxWorker *const worker = &workers[i]; |
992 | 0 | cdef_worker[i].cm = cm; |
993 | 0 | cdef_worker[i].xd = xd; |
994 | 0 | cdef_worker[i].cdef_init_fb_row_fn = cdef_init_fb_row_fn; |
995 | 0 | for (int plane = 0; plane < num_planes; plane++) |
996 | 0 | cdef_worker[i].linebuf[plane] = cm->cdef_info.linebuf[plane]; |
997 | |
|
998 | 0 | worker->hook = hook; |
999 | 0 | worker->data1 = cdef_sync; |
1000 | 0 | worker->data2 = &cdef_worker[i]; |
1001 | 0 | } |
1002 | 0 | } |
1003 | | |
1004 | | // Initializes row-level parameters for CDEF frame. |
1005 | | void av1_cdef_init_fb_row_mt(const AV1_COMMON *const cm, |
1006 | | const MACROBLOCKD *const xd, |
1007 | | CdefBlockInfo *const fb_info, |
1008 | | uint16_t **const linebuf, uint16_t *const src, |
1009 | 0 | struct AV1CdefSyncData *const cdef_sync, int fbr) { |
1010 | 0 | const int num_planes = av1_num_planes(cm); |
1011 | 0 | const int nvfb = (cm->mi_params.mi_rows + MI_SIZE_64X64 - 1) / MI_SIZE_64X64; |
1012 | 0 | const int luma_stride = |
1013 | 0 | ALIGN_POWER_OF_TWO(cm->mi_params.mi_cols << MI_SIZE_LOG2, 4); |
1014 | | |
1015 | | // for the current filter block, it's top left corner mi structure (mi_tl) |
1016 | | // is first accessed to check whether the top and left boundaries are |
1017 | | // frame boundaries. Then bottom-left and top-right mi structures are |
1018 | | // accessed to check whether the bottom and right boundaries |
1019 | | // (respectively) are frame boundaries. |
1020 | | // |
1021 | | // Note that we can't just check the bottom-right mi structure - eg. if |
1022 | | // we're at the right-hand edge of the frame but not the bottom, then |
1023 | | // the bottom-right mi is NULL but the bottom-left is not. |
1024 | 0 | fb_info->frame_boundary[TOP] = (MI_SIZE_64X64 * fbr == 0) ? 1 : 0; |
1025 | 0 | if (fbr != nvfb - 1) |
1026 | 0 | fb_info->frame_boundary[BOTTOM] = |
1027 | 0 | (MI_SIZE_64X64 * (fbr + 1) == cm->mi_params.mi_rows) ? 1 : 0; |
1028 | 0 | else |
1029 | 0 | fb_info->frame_boundary[BOTTOM] = 1; |
1030 | |
|
1031 | 0 | fb_info->src = src; |
1032 | 0 | fb_info->damping = cm->cdef_info.cdef_damping; |
1033 | 0 | fb_info->coeff_shift = AOMMAX(cm->seq_params->bit_depth - 8, 0); |
1034 | 0 | av1_zero(fb_info->dir); |
1035 | 0 | av1_zero(fb_info->var); |
1036 | |
|
1037 | 0 | for (int plane = 0; plane < num_planes; plane++) { |
1038 | 0 | const int stride = luma_stride >> xd->plane[plane].subsampling_x; |
1039 | 0 | uint16_t *top_linebuf = &linebuf[plane][0]; |
1040 | 0 | uint16_t *bot_linebuf = &linebuf[plane][nvfb * CDEF_VBORDER * stride]; |
1041 | 0 | { |
1042 | 0 | const int mi_high_l2 = MI_SIZE_LOG2 - xd->plane[plane].subsampling_y; |
1043 | 0 | const int top_offset = MI_SIZE_64X64 * (fbr + 1) << mi_high_l2; |
1044 | 0 | const int bot_offset = MI_SIZE_64X64 * (fbr + 1) << mi_high_l2; |
1045 | |
|
1046 | 0 | if (fbr != nvfb - 1) // if (fbr != 0) // top line buffer copy |
1047 | 0 | av1_cdef_copy_sb8_16( |
1048 | 0 | cm, &top_linebuf[(fbr + 1) * CDEF_VBORDER * stride], stride, |
1049 | 0 | xd->plane[plane].dst.buf, top_offset - CDEF_VBORDER, 0, |
1050 | 0 | xd->plane[plane].dst.stride, CDEF_VBORDER, stride); |
1051 | 0 | if (fbr != nvfb - 1) // bottom line buffer copy |
1052 | 0 | av1_cdef_copy_sb8_16(cm, &bot_linebuf[fbr * CDEF_VBORDER * stride], |
1053 | 0 | stride, xd->plane[plane].dst.buf, bot_offset, 0, |
1054 | 0 | xd->plane[plane].dst.stride, CDEF_VBORDER, stride); |
1055 | 0 | } |
1056 | |
|
1057 | 0 | fb_info->top_linebuf[plane] = &linebuf[plane][fbr * CDEF_VBORDER * stride]; |
1058 | 0 | fb_info->bot_linebuf[plane] = |
1059 | 0 | &linebuf[plane] |
1060 | 0 | [nvfb * CDEF_VBORDER * stride + (fbr * CDEF_VBORDER * stride)]; |
1061 | 0 | } |
1062 | |
|
1063 | 0 | cdef_row_mt_sync_write(cdef_sync, fbr); |
1064 | 0 | cdef_row_mt_sync_read(cdef_sync, fbr); |
1065 | 0 | } |
1066 | | |
1067 | | // Implements multi-threading for CDEF. |
1068 | | // Perform CDEF on input frame. |
1069 | | // Inputs: |
1070 | | // frame: Pointer to input frame buffer. |
1071 | | // cm: Pointer to common structure. |
1072 | | // xd: Pointer to common current coding block structure. |
1073 | | // Returns: |
1074 | | // Nothing will be returned. |
1075 | | void av1_cdef_frame_mt(AV1_COMMON *const cm, MACROBLOCKD *const xd, |
1076 | | AV1CdefWorkerData *const cdef_worker, |
1077 | | AVxWorker *const workers, AV1CdefSync *const cdef_sync, |
1078 | | int num_workers, |
1079 | 0 | cdef_init_fb_row_t cdef_init_fb_row_fn) { |
1080 | 0 | YV12_BUFFER_CONFIG *frame = &cm->cur_frame->buf; |
1081 | 0 | const int num_planes = av1_num_planes(cm); |
1082 | |
|
1083 | 0 | av1_setup_dst_planes(xd->plane, cm->seq_params->sb_size, frame, 0, 0, 0, |
1084 | 0 | num_planes); |
1085 | |
|
1086 | 0 | reset_cdef_job_info(cdef_sync); |
1087 | 0 | prepare_cdef_frame_workers(cm, xd, cdef_worker, cdef_sb_row_worker_hook, |
1088 | 0 | workers, cdef_sync, num_workers, |
1089 | 0 | cdef_init_fb_row_fn); |
1090 | 0 | launch_cdef_workers(workers, num_workers); |
1091 | 0 | sync_cdef_workers(workers, cm, num_workers); |
1092 | 0 | } |