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 "av1/common/warped_motion.h"
13 #include "av1/common/thread_common.h"
14
15 #include "av1/encoder/bitstream.h"
16 #include "av1/encoder/encodeframe.h"
17 #include "av1/encoder/encoder.h"
18 #include "av1/encoder/encoder_alloc.h"
19 #include "av1/encoder/encodeframe_utils.h"
20 #include "av1/encoder/ethread.h"
21 #if !CONFIG_REALTIME_ONLY
22 #include "av1/encoder/firstpass.h"
23 #endif
24 #include "av1/encoder/global_motion.h"
25 #include "av1/encoder/global_motion_facade.h"
26 #include "av1/encoder/rdopt.h"
27 #include "aom_dsp/aom_dsp_common.h"
28 #include "av1/encoder/temporal_filter.h"
29 #include "av1/encoder/tpl_model.h"
30
accumulate_rd_opt(ThreadData * td,ThreadData * td_t)31 static AOM_INLINE void accumulate_rd_opt(ThreadData *td, ThreadData *td_t) {
32 for (int i = 0; i < REFERENCE_MODES; i++)
33 td->rd_counts.comp_pred_diff[i] += td_t->rd_counts.comp_pred_diff[i];
34
35 td->rd_counts.compound_ref_used_flag |=
36 td_t->rd_counts.compound_ref_used_flag;
37 td->rd_counts.skip_mode_used_flag |= td_t->rd_counts.skip_mode_used_flag;
38
39 for (int i = 0; i < TX_SIZES_ALL; i++) {
40 for (int j = 0; j < TX_TYPES; j++)
41 td->rd_counts.tx_type_used[i][j] += td_t->rd_counts.tx_type_used[i][j];
42 }
43
44 for (int i = 0; i < BLOCK_SIZES_ALL; i++) {
45 for (int j = 0; j < 2; j++) {
46 td->rd_counts.obmc_used[i][j] += td_t->rd_counts.obmc_used[i][j];
47 }
48 }
49
50 for (int i = 0; i < 2; i++) {
51 td->rd_counts.warped_used[i] += td_t->rd_counts.warped_used[i];
52 }
53 }
54
update_delta_lf_for_row_mt(AV1_COMP * cpi)55 static AOM_INLINE void update_delta_lf_for_row_mt(AV1_COMP *cpi) {
56 AV1_COMMON *cm = &cpi->common;
57 MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
58 const int mib_size = cm->seq_params->mib_size;
59 const int frame_lf_count =
60 av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2;
61 for (int row = 0; row < cm->tiles.rows; row++) {
62 for (int col = 0; col < cm->tiles.cols; col++) {
63 TileDataEnc *tile_data = &cpi->tile_data[row * cm->tiles.cols + col];
64 const TileInfo *const tile_info = &tile_data->tile_info;
65 for (int mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end;
66 mi_row += mib_size) {
67 if (mi_row == tile_info->mi_row_start)
68 av1_reset_loop_filter_delta(xd, av1_num_planes(cm));
69 for (int mi_col = tile_info->mi_col_start;
70 mi_col < tile_info->mi_col_end; mi_col += mib_size) {
71 const int idx_str = cm->mi_params.mi_stride * mi_row + mi_col;
72 MB_MODE_INFO **mi = cm->mi_params.mi_grid_base + idx_str;
73 MB_MODE_INFO *mbmi = mi[0];
74 if (mbmi->skip_txfm == 1 &&
75 (mbmi->bsize == cm->seq_params->sb_size)) {
76 for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id)
77 mbmi->delta_lf[lf_id] = xd->delta_lf[lf_id];
78 mbmi->delta_lf_from_base = xd->delta_lf_from_base;
79 } else {
80 if (cm->delta_q_info.delta_lf_multi) {
81 for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id)
82 xd->delta_lf[lf_id] = mbmi->delta_lf[lf_id];
83 } else {
84 xd->delta_lf_from_base = mbmi->delta_lf_from_base;
85 }
86 }
87 }
88 }
89 }
90 }
91 }
92
av1_row_mt_sync_read_dummy(AV1EncRowMultiThreadSync * row_mt_sync,int r,int c)93 void av1_row_mt_sync_read_dummy(AV1EncRowMultiThreadSync *row_mt_sync, int r,
94 int c) {
95 (void)row_mt_sync;
96 (void)r;
97 (void)c;
98 return;
99 }
100
av1_row_mt_sync_write_dummy(AV1EncRowMultiThreadSync * row_mt_sync,int r,int c,int cols)101 void av1_row_mt_sync_write_dummy(AV1EncRowMultiThreadSync *row_mt_sync, int r,
102 int c, int cols) {
103 (void)row_mt_sync;
104 (void)r;
105 (void)c;
106 (void)cols;
107 return;
108 }
109
av1_row_mt_sync_read(AV1EncRowMultiThreadSync * row_mt_sync,int r,int c)110 void av1_row_mt_sync_read(AV1EncRowMultiThreadSync *row_mt_sync, int r, int c) {
111 #if CONFIG_MULTITHREAD
112 const int nsync = row_mt_sync->sync_range;
113
114 if (r) {
115 pthread_mutex_t *const mutex = &row_mt_sync->mutex_[r - 1];
116 pthread_mutex_lock(mutex);
117
118 while (c > row_mt_sync->num_finished_cols[r - 1] - nsync) {
119 pthread_cond_wait(&row_mt_sync->cond_[r - 1], mutex);
120 }
121 pthread_mutex_unlock(mutex);
122 }
123 #else
124 (void)row_mt_sync;
125 (void)r;
126 (void)c;
127 #endif // CONFIG_MULTITHREAD
128 }
129
av1_row_mt_sync_write(AV1EncRowMultiThreadSync * row_mt_sync,int r,int c,int cols)130 void av1_row_mt_sync_write(AV1EncRowMultiThreadSync *row_mt_sync, int r, int c,
131 int cols) {
132 #if CONFIG_MULTITHREAD
133 const int nsync = row_mt_sync->sync_range;
134 int cur;
135 // Only signal when there are enough encoded blocks for next row to run.
136 int sig = 1;
137
138 if (c < cols - 1) {
139 cur = c;
140 if (c % nsync) sig = 0;
141 } else {
142 cur = cols + nsync;
143 }
144
145 if (sig) {
146 pthread_mutex_lock(&row_mt_sync->mutex_[r]);
147
148 row_mt_sync->num_finished_cols[r] = cur;
149
150 pthread_cond_signal(&row_mt_sync->cond_[r]);
151 pthread_mutex_unlock(&row_mt_sync->mutex_[r]);
152 }
153 #else
154 (void)row_mt_sync;
155 (void)r;
156 (void)c;
157 (void)cols;
158 #endif // CONFIG_MULTITHREAD
159 }
160
161 // Allocate memory for row synchronization
row_mt_sync_mem_alloc(AV1EncRowMultiThreadSync * row_mt_sync,AV1_COMMON * cm,int rows)162 static void row_mt_sync_mem_alloc(AV1EncRowMultiThreadSync *row_mt_sync,
163 AV1_COMMON *cm, int rows) {
164 #if CONFIG_MULTITHREAD
165 int i;
166
167 CHECK_MEM_ERROR(cm, row_mt_sync->mutex_,
168 aom_malloc(sizeof(*row_mt_sync->mutex_) * rows));
169 if (row_mt_sync->mutex_) {
170 for (i = 0; i < rows; ++i) {
171 pthread_mutex_init(&row_mt_sync->mutex_[i], NULL);
172 }
173 }
174
175 CHECK_MEM_ERROR(cm, row_mt_sync->cond_,
176 aom_malloc(sizeof(*row_mt_sync->cond_) * rows));
177 if (row_mt_sync->cond_) {
178 for (i = 0; i < rows; ++i) {
179 pthread_cond_init(&row_mt_sync->cond_[i], NULL);
180 }
181 }
182 #endif // CONFIG_MULTITHREAD
183
184 CHECK_MEM_ERROR(cm, row_mt_sync->num_finished_cols,
185 aom_malloc(sizeof(*row_mt_sync->num_finished_cols) * rows));
186
187 row_mt_sync->rows = rows;
188 // Set up nsync.
189 row_mt_sync->sync_range = 1;
190 }
191
192 // Deallocate row based multi-threading synchronization related mutex and data
row_mt_sync_mem_dealloc(AV1EncRowMultiThreadSync * row_mt_sync)193 static void row_mt_sync_mem_dealloc(AV1EncRowMultiThreadSync *row_mt_sync) {
194 if (row_mt_sync != NULL) {
195 #if CONFIG_MULTITHREAD
196 int i;
197
198 if (row_mt_sync->mutex_ != NULL) {
199 for (i = 0; i < row_mt_sync->rows; ++i) {
200 pthread_mutex_destroy(&row_mt_sync->mutex_[i]);
201 }
202 aom_free(row_mt_sync->mutex_);
203 }
204 if (row_mt_sync->cond_ != NULL) {
205 for (i = 0; i < row_mt_sync->rows; ++i) {
206 pthread_cond_destroy(&row_mt_sync->cond_[i]);
207 }
208 aom_free(row_mt_sync->cond_);
209 }
210 #endif // CONFIG_MULTITHREAD
211 aom_free(row_mt_sync->num_finished_cols);
212
213 // clear the structure as the source of this call may be dynamic change
214 // in tiles in which case this call will be followed by an _alloc()
215 // which may fail.
216 av1_zero(*row_mt_sync);
217 }
218 }
219
row_mt_mem_alloc(AV1_COMP * cpi,int max_rows,int max_cols,int alloc_row_ctx)220 static void row_mt_mem_alloc(AV1_COMP *cpi, int max_rows, int max_cols,
221 int alloc_row_ctx) {
222 struct AV1Common *cm = &cpi->common;
223 AV1EncRowMultiThreadInfo *const enc_row_mt = &cpi->mt_info.enc_row_mt;
224 const int tile_cols = cm->tiles.cols;
225 const int tile_rows = cm->tiles.rows;
226 int tile_col, tile_row;
227
228 // Allocate memory for row based multi-threading
229 for (tile_row = 0; tile_row < tile_rows; tile_row++) {
230 for (tile_col = 0; tile_col < tile_cols; tile_col++) {
231 int tile_index = tile_row * tile_cols + tile_col;
232 TileDataEnc *const this_tile = &cpi->tile_data[tile_index];
233
234 row_mt_sync_mem_alloc(&this_tile->row_mt_sync, cm, max_rows);
235
236 this_tile->row_ctx = NULL;
237 if (alloc_row_ctx) {
238 assert(max_cols > 0);
239 const int num_row_ctx = AOMMAX(1, (max_cols - 1));
240 CHECK_MEM_ERROR(cm, this_tile->row_ctx,
241 (FRAME_CONTEXT *)aom_memalign(
242 16, num_row_ctx * sizeof(*this_tile->row_ctx)));
243 }
244 }
245 }
246 enc_row_mt->allocated_tile_cols = tile_cols;
247 enc_row_mt->allocated_tile_rows = tile_rows;
248 enc_row_mt->allocated_rows = max_rows;
249 enc_row_mt->allocated_cols = max_cols - 1;
250 }
251
av1_row_mt_mem_dealloc(AV1_COMP * cpi)252 void av1_row_mt_mem_dealloc(AV1_COMP *cpi) {
253 AV1EncRowMultiThreadInfo *const enc_row_mt = &cpi->mt_info.enc_row_mt;
254 const int tile_cols = enc_row_mt->allocated_tile_cols;
255 const int tile_rows = enc_row_mt->allocated_tile_rows;
256 int tile_col, tile_row;
257
258 // Free row based multi-threading sync memory
259 for (tile_row = 0; tile_row < tile_rows; tile_row++) {
260 for (tile_col = 0; tile_col < tile_cols; tile_col++) {
261 int tile_index = tile_row * tile_cols + tile_col;
262 TileDataEnc *const this_tile = &cpi->tile_data[tile_index];
263
264 row_mt_sync_mem_dealloc(&this_tile->row_mt_sync);
265
266 if (cpi->oxcf.algo_cfg.cdf_update_mode) aom_free(this_tile->row_ctx);
267 }
268 }
269 enc_row_mt->allocated_rows = 0;
270 enc_row_mt->allocated_cols = 0;
271 enc_row_mt->allocated_tile_cols = 0;
272 enc_row_mt->allocated_tile_rows = 0;
273 }
274
assign_tile_to_thread(int * thread_id_to_tile_id,int num_tiles,int num_workers)275 static AOM_INLINE void assign_tile_to_thread(int *thread_id_to_tile_id,
276 int num_tiles, int num_workers) {
277 int tile_id = 0;
278 int i;
279
280 for (i = 0; i < num_workers; i++) {
281 thread_id_to_tile_id[i] = tile_id++;
282 if (tile_id == num_tiles) tile_id = 0;
283 }
284 }
285
get_next_job(TileDataEnc * const tile_data,int * current_mi_row,int mib_size)286 static AOM_INLINE int get_next_job(TileDataEnc *const tile_data,
287 int *current_mi_row, int mib_size) {
288 AV1EncRowMultiThreadSync *const row_mt_sync = &tile_data->row_mt_sync;
289 const int mi_row_end = tile_data->tile_info.mi_row_end;
290
291 if (row_mt_sync->next_mi_row < mi_row_end) {
292 *current_mi_row = row_mt_sync->next_mi_row;
293 row_mt_sync->num_threads_working++;
294 row_mt_sync->next_mi_row += mib_size;
295 return 1;
296 }
297 return 0;
298 }
299
switch_tile_and_get_next_job(AV1_COMMON * const cm,TileDataEnc * const tile_data,int * cur_tile_id,int * current_mi_row,int * end_of_frame,int is_firstpass,const BLOCK_SIZE fp_block_size)300 static AOM_INLINE void switch_tile_and_get_next_job(
301 AV1_COMMON *const cm, TileDataEnc *const tile_data, int *cur_tile_id,
302 int *current_mi_row, int *end_of_frame, int is_firstpass,
303 const BLOCK_SIZE fp_block_size) {
304 const int tile_cols = cm->tiles.cols;
305 const int tile_rows = cm->tiles.rows;
306
307 int tile_id = -1; // Stores the tile ID with minimum proc done
308 int max_mis_to_encode = 0;
309 int min_num_threads_working = INT_MAX;
310
311 for (int tile_row = 0; tile_row < tile_rows; tile_row++) {
312 for (int tile_col = 0; tile_col < tile_cols; tile_col++) {
313 int tile_index = tile_row * tile_cols + tile_col;
314 TileDataEnc *const this_tile = &tile_data[tile_index];
315 AV1EncRowMultiThreadSync *const row_mt_sync = &this_tile->row_mt_sync;
316
317 #if CONFIG_REALTIME_ONLY
318 int num_b_rows_in_tile =
319 av1_get_sb_rows_in_tile(cm, this_tile->tile_info);
320 int num_b_cols_in_tile =
321 av1_get_sb_cols_in_tile(cm, this_tile->tile_info);
322 #else
323 int num_b_rows_in_tile =
324 is_firstpass
325 ? av1_get_unit_rows_in_tile(this_tile->tile_info, fp_block_size)
326 : av1_get_sb_rows_in_tile(cm, this_tile->tile_info);
327 int num_b_cols_in_tile =
328 is_firstpass
329 ? av1_get_unit_cols_in_tile(this_tile->tile_info, fp_block_size)
330 : av1_get_sb_cols_in_tile(cm, this_tile->tile_info);
331 #endif
332 int theoretical_limit_on_threads =
333 AOMMIN((num_b_cols_in_tile + 1) >> 1, num_b_rows_in_tile);
334 int num_threads_working = row_mt_sync->num_threads_working;
335
336 if (num_threads_working < theoretical_limit_on_threads) {
337 int num_mis_to_encode =
338 this_tile->tile_info.mi_row_end - row_mt_sync->next_mi_row;
339
340 // Tile to be processed by this thread is selected on the basis of
341 // availability of jobs:
342 // 1) If jobs are available, tile to be processed is chosen on the
343 // basis of minimum number of threads working for that tile. If two or
344 // more tiles have same number of threads working for them, then the
345 // tile with maximum number of jobs available will be chosen.
346 // 2) If no jobs are available, then end_of_frame is reached.
347 if (num_mis_to_encode > 0) {
348 if (num_threads_working < min_num_threads_working) {
349 min_num_threads_working = num_threads_working;
350 max_mis_to_encode = 0;
351 }
352 if (num_threads_working == min_num_threads_working &&
353 num_mis_to_encode > max_mis_to_encode) {
354 tile_id = tile_index;
355 max_mis_to_encode = num_mis_to_encode;
356 }
357 }
358 }
359 }
360 }
361 if (tile_id == -1) {
362 *end_of_frame = 1;
363 } else {
364 // Update the current tile id to the tile id that will be processed next,
365 // which will be the least processed tile.
366 *cur_tile_id = tile_id;
367 const int unit_height = mi_size_high[fp_block_size];
368 get_next_job(&tile_data[tile_id], current_mi_row,
369 is_firstpass ? unit_height : cm->seq_params->mib_size);
370 }
371 }
372
373 #if !CONFIG_REALTIME_ONLY
fp_enc_row_mt_worker_hook(void * arg1,void * unused)374 static int fp_enc_row_mt_worker_hook(void *arg1, void *unused) {
375 EncWorkerData *const thread_data = (EncWorkerData *)arg1;
376 AV1_COMP *const cpi = thread_data->cpi;
377 AV1_COMMON *const cm = &cpi->common;
378 int thread_id = thread_data->thread_id;
379 AV1EncRowMultiThreadInfo *const enc_row_mt = &cpi->mt_info.enc_row_mt;
380 int cur_tile_id = enc_row_mt->thread_id_to_tile_id[thread_id];
381 #if CONFIG_MULTITHREAD
382 pthread_mutex_t *enc_row_mt_mutex_ = enc_row_mt->mutex_;
383 #endif
384 (void)unused;
385
386 assert(cur_tile_id != -1);
387
388 const BLOCK_SIZE fp_block_size = cpi->fp_block_size;
389 const int unit_height = mi_size_high[fp_block_size];
390 int end_of_frame = 0;
391 while (1) {
392 int current_mi_row = -1;
393 #if CONFIG_MULTITHREAD
394 pthread_mutex_lock(enc_row_mt_mutex_);
395 #endif
396 if (!get_next_job(&cpi->tile_data[cur_tile_id], ¤t_mi_row,
397 unit_height)) {
398 // No jobs are available for the current tile. Query for the status of
399 // other tiles and get the next job if available
400 switch_tile_and_get_next_job(cm, cpi->tile_data, &cur_tile_id,
401 ¤t_mi_row, &end_of_frame, 1,
402 fp_block_size);
403 }
404 #if CONFIG_MULTITHREAD
405 pthread_mutex_unlock(enc_row_mt_mutex_);
406 #endif
407 if (end_of_frame == 1) break;
408
409 TileDataEnc *const this_tile = &cpi->tile_data[cur_tile_id];
410 AV1EncRowMultiThreadSync *const row_mt_sync = &this_tile->row_mt_sync;
411 ThreadData *td = thread_data->td;
412
413 assert(current_mi_row != -1 &&
414 current_mi_row <= this_tile->tile_info.mi_row_end);
415
416 const int unit_height_log2 = mi_size_high_log2[fp_block_size];
417 av1_first_pass_row(cpi, td, this_tile, current_mi_row >> unit_height_log2,
418 fp_block_size);
419 #if CONFIG_MULTITHREAD
420 pthread_mutex_lock(enc_row_mt_mutex_);
421 #endif
422 row_mt_sync->num_threads_working--;
423 #if CONFIG_MULTITHREAD
424 pthread_mutex_unlock(enc_row_mt_mutex_);
425 #endif
426 }
427
428 return 1;
429 }
430 #endif
431
enc_row_mt_worker_hook(void * arg1,void * unused)432 static int enc_row_mt_worker_hook(void *arg1, void *unused) {
433 EncWorkerData *const thread_data = (EncWorkerData *)arg1;
434 AV1_COMP *const cpi = thread_data->cpi;
435 AV1_COMMON *const cm = &cpi->common;
436 int thread_id = thread_data->thread_id;
437 AV1EncRowMultiThreadInfo *const enc_row_mt = &cpi->mt_info.enc_row_mt;
438 int cur_tile_id = enc_row_mt->thread_id_to_tile_id[thread_id];
439 #if CONFIG_MULTITHREAD
440 pthread_mutex_t *enc_row_mt_mutex_ = enc_row_mt->mutex_;
441 #endif
442 (void)unused;
443
444 assert(cur_tile_id != -1);
445
446 const BLOCK_SIZE fp_block_size = cpi->fp_block_size;
447 int end_of_frame = 0;
448
449 // When master thread does not have a valid job to process, xd->tile_ctx
450 // is not set and it contains NULL pointer. This can result in NULL pointer
451 // access violation if accessed beyond the encode stage. Hence, updating
452 // thread_data->td->mb.e_mbd.tile_ctx is initialized with common frame
453 // context to avoid NULL pointer access in subsequent stages.
454 thread_data->td->mb.e_mbd.tile_ctx = cm->fc;
455 while (1) {
456 int current_mi_row = -1;
457 #if CONFIG_MULTITHREAD
458 pthread_mutex_lock(enc_row_mt_mutex_);
459 #endif
460 if (!get_next_job(&cpi->tile_data[cur_tile_id], ¤t_mi_row,
461 cm->seq_params->mib_size)) {
462 // No jobs are available for the current tile. Query for the status of
463 // other tiles and get the next job if available
464 switch_tile_and_get_next_job(cm, cpi->tile_data, &cur_tile_id,
465 ¤t_mi_row, &end_of_frame, 0,
466 fp_block_size);
467 }
468 #if CONFIG_MULTITHREAD
469 pthread_mutex_unlock(enc_row_mt_mutex_);
470 #endif
471 if (end_of_frame == 1) break;
472
473 TileDataEnc *const this_tile = &cpi->tile_data[cur_tile_id];
474 AV1EncRowMultiThreadSync *const row_mt_sync = &this_tile->row_mt_sync;
475 const TileInfo *const tile_info = &this_tile->tile_info;
476 const int tile_row = tile_info->tile_row;
477 const int tile_col = tile_info->tile_col;
478 ThreadData *td = thread_data->td;
479
480 assert(current_mi_row != -1 && current_mi_row <= tile_info->mi_row_end);
481
482 td->mb.e_mbd.tile_ctx = td->tctx;
483 td->mb.tile_pb_ctx = &this_tile->tctx;
484 td->abs_sum_level = 0;
485
486 if (this_tile->allow_update_cdf) {
487 td->mb.row_ctx = this_tile->row_ctx;
488 if (current_mi_row == tile_info->mi_row_start)
489 memcpy(td->mb.e_mbd.tile_ctx, &this_tile->tctx, sizeof(FRAME_CONTEXT));
490 } else {
491 memcpy(td->mb.e_mbd.tile_ctx, &this_tile->tctx, sizeof(FRAME_CONTEXT));
492 }
493
494 av1_init_above_context(&cm->above_contexts, av1_num_planes(cm), tile_row,
495 &td->mb.e_mbd);
496
497 cfl_init(&td->mb.e_mbd.cfl, cm->seq_params);
498 if (td->mb.txfm_search_info.txb_rd_records != NULL) {
499 av1_crc32c_calculator_init(
500 &td->mb.txfm_search_info.txb_rd_records->mb_rd_record.crc_calculator);
501 }
502
503 av1_encode_sb_row(cpi, td, tile_row, tile_col, current_mi_row);
504 #if CONFIG_MULTITHREAD
505 pthread_mutex_lock(enc_row_mt_mutex_);
506 #endif
507 this_tile->abs_sum_level += td->abs_sum_level;
508 row_mt_sync->num_threads_working--;
509 #if CONFIG_MULTITHREAD
510 pthread_mutex_unlock(enc_row_mt_mutex_);
511 #endif
512 }
513
514 return 1;
515 }
516
enc_worker_hook(void * arg1,void * unused)517 static int enc_worker_hook(void *arg1, void *unused) {
518 EncWorkerData *const thread_data = (EncWorkerData *)arg1;
519 AV1_COMP *const cpi = thread_data->cpi;
520 const AV1_COMMON *const cm = &cpi->common;
521 const int tile_cols = cm->tiles.cols;
522 const int tile_rows = cm->tiles.rows;
523 int t;
524
525 (void)unused;
526
527 for (t = thread_data->start; t < tile_rows * tile_cols;
528 t += cpi->mt_info.num_workers) {
529 int tile_row = t / tile_cols;
530 int tile_col = t % tile_cols;
531
532 TileDataEnc *const this_tile =
533 &cpi->tile_data[tile_row * cm->tiles.cols + tile_col];
534 thread_data->td->mb.e_mbd.tile_ctx = &this_tile->tctx;
535 thread_data->td->mb.tile_pb_ctx = &this_tile->tctx;
536 av1_encode_tile(cpi, thread_data->td, tile_row, tile_col);
537 }
538
539 return 1;
540 }
541
av1_init_frame_mt(AV1_PRIMARY * ppi,AV1_COMP * cpi)542 void av1_init_frame_mt(AV1_PRIMARY *ppi, AV1_COMP *cpi) {
543 cpi->mt_info.workers = ppi->p_mt_info.workers;
544 cpi->mt_info.num_workers = ppi->p_mt_info.num_workers;
545 cpi->mt_info.tile_thr_data = ppi->p_mt_info.tile_thr_data;
546 int i;
547 for (i = MOD_FP; i < NUM_MT_MODULES; i++) {
548 cpi->mt_info.num_mod_workers[i] =
549 AOMMIN(cpi->mt_info.num_workers, ppi->p_mt_info.num_mod_workers[i]);
550 }
551 }
552
av1_init_cdef_worker(AV1_COMP * cpi)553 void av1_init_cdef_worker(AV1_COMP *cpi) {
554 #if CONFIG_FRAME_PARALLEL_ENCODE
555 // The allocation is done only for level 0 parallel frames. No change
556 // in config is supported in the middle of a parallel encode set, since the
557 // rest of the MT modules also do not support dynamic change of config.
558 if (cpi->ppi->gf_group.frame_parallel_level[cpi->gf_frame_index] > 0) return;
559 #endif // CONFIG_FRAME_PARALLEL_ENCODE
560 PrimaryMultiThreadInfo *const p_mt_info = &cpi->ppi->p_mt_info;
561 int num_cdef_workers = av1_get_num_mod_workers_for_alloc(p_mt_info, MOD_CDEF);
562
563 av1_alloc_cdef_buffers(&cpi->common, &p_mt_info->cdef_worker,
564 &cpi->mt_info.cdef_sync, num_cdef_workers, 1);
565 cpi->mt_info.cdef_worker = &p_mt_info->cdef_worker[0];
566 }
567
568 #if !CONFIG_REALTIME_ONLY
av1_init_lr_mt_buffers(AV1_COMP * cpi)569 void av1_init_lr_mt_buffers(AV1_COMP *cpi) {
570 AV1_COMMON *const cm = &cpi->common;
571 AV1LrSync *lr_sync = &cpi->mt_info.lr_row_sync;
572 if (lr_sync->sync_range) {
573 int num_lr_workers =
574 av1_get_num_mod_workers_for_alloc(&cpi->ppi->p_mt_info, MOD_LR);
575 #if CONFIG_FRAME_PARALLEL_ENCODE
576 if (cpi->ppi->gf_group.frame_parallel_level[cpi->gf_frame_index] > 0)
577 return;
578 #endif // CONFIG_FRAME_PARALLEL_ENCODE
579 lr_sync->lrworkerdata[num_lr_workers - 1].rst_tmpbuf = cm->rst_tmpbuf;
580 lr_sync->lrworkerdata[num_lr_workers - 1].rlbs = cm->rlbs;
581 }
582 }
583 #endif
584
585 #if CONFIG_MULTITHREAD
av1_init_mt_sync(AV1_COMP * cpi,int is_first_pass)586 void av1_init_mt_sync(AV1_COMP *cpi, int is_first_pass) {
587 AV1_COMMON *const cm = &cpi->common;
588 MultiThreadInfo *const mt_info = &cpi->mt_info;
589
590 // Initialize enc row MT object.
591 if (is_first_pass || cpi->oxcf.row_mt == 1) {
592 AV1EncRowMultiThreadInfo *enc_row_mt = &mt_info->enc_row_mt;
593 if (enc_row_mt->mutex_ == NULL) {
594 CHECK_MEM_ERROR(cm, enc_row_mt->mutex_,
595 aom_malloc(sizeof(*(enc_row_mt->mutex_))));
596 if (enc_row_mt->mutex_) pthread_mutex_init(enc_row_mt->mutex_, NULL);
597 }
598 }
599
600 if (!is_first_pass) {
601 // Initialize global motion MT object.
602 AV1GlobalMotionSync *gm_sync = &mt_info->gm_sync;
603 if (gm_sync->mutex_ == NULL) {
604 CHECK_MEM_ERROR(cm, gm_sync->mutex_,
605 aom_malloc(sizeof(*(gm_sync->mutex_))));
606 if (gm_sync->mutex_) pthread_mutex_init(gm_sync->mutex_, NULL);
607 }
608 #if !CONFIG_REALTIME_ONLY
609 // Initialize temporal filtering MT object.
610 AV1TemporalFilterSync *tf_sync = &mt_info->tf_sync;
611 if (tf_sync->mutex_ == NULL) {
612 CHECK_MEM_ERROR(cm, tf_sync->mutex_,
613 aom_malloc(sizeof(*tf_sync->mutex_)));
614 if (tf_sync->mutex_) pthread_mutex_init(tf_sync->mutex_, NULL);
615 }
616 #endif // !CONFIG_REALTIME_ONLY
617 // Initialize CDEF MT object.
618 AV1CdefSync *cdef_sync = &mt_info->cdef_sync;
619 if (cdef_sync->mutex_ == NULL) {
620 CHECK_MEM_ERROR(cm, cdef_sync->mutex_,
621 aom_malloc(sizeof(*(cdef_sync->mutex_))));
622 if (cdef_sync->mutex_) pthread_mutex_init(cdef_sync->mutex_, NULL);
623 }
624
625 // Initialize loop filter MT object.
626 AV1LfSync *lf_sync = &mt_info->lf_row_sync;
627 // Number of superblock rows
628 const int sb_rows =
629 ALIGN_POWER_OF_TWO(cm->height >> MI_SIZE_LOG2, MAX_MIB_SIZE_LOG2) >>
630 MAX_MIB_SIZE_LOG2;
631 PrimaryMultiThreadInfo *const p_mt_info = &cpi->ppi->p_mt_info;
632 int num_lf_workers = av1_get_num_mod_workers_for_alloc(p_mt_info, MOD_LPF);
633
634 if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
635 num_lf_workers > lf_sync->num_workers) {
636 av1_loop_filter_dealloc(lf_sync);
637 av1_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_lf_workers);
638 }
639
640 #if !CONFIG_REALTIME_ONLY
641 // Initialize loop restoration MT object.
642 AV1LrSync *lr_sync = &mt_info->lr_row_sync;
643 int rst_unit_size;
644 if (cm->width * cm->height > 352 * 288)
645 rst_unit_size = RESTORATION_UNITSIZE_MAX;
646 else
647 rst_unit_size = (RESTORATION_UNITSIZE_MAX >> 1);
648 int num_rows_lr = av1_lr_count_units_in_tile(rst_unit_size, cm->height);
649 int num_lr_workers = av1_get_num_mod_workers_for_alloc(p_mt_info, MOD_LR);
650 if (!lr_sync->sync_range || num_rows_lr > lr_sync->rows ||
651 num_lr_workers > lr_sync->num_workers ||
652 MAX_MB_PLANE > lr_sync->num_planes) {
653 av1_loop_restoration_dealloc(lr_sync, num_lr_workers);
654 av1_loop_restoration_alloc(lr_sync, cm, num_lr_workers, num_rows_lr,
655 MAX_MB_PLANE, cm->width);
656 }
657 #endif
658
659 // Initialization of pack bitstream MT object.
660 AV1EncPackBSSync *pack_bs_sync = &mt_info->pack_bs_sync;
661 if (pack_bs_sync->mutex_ == NULL) {
662 CHECK_MEM_ERROR(cm, pack_bs_sync->mutex_,
663 aom_malloc(sizeof(*pack_bs_sync->mutex_)));
664 if (pack_bs_sync->mutex_) pthread_mutex_init(pack_bs_sync->mutex_, NULL);
665 }
666 }
667 }
668 #endif // CONFIG_MULTITHREAD
669
670 // Computes the number of workers to be considered while allocating memory for a
671 // multi-threaded module under FPMT.
av1_get_num_mod_workers_for_alloc(PrimaryMultiThreadInfo * const p_mt_info,MULTI_THREADED_MODULES mod_name)672 int av1_get_num_mod_workers_for_alloc(PrimaryMultiThreadInfo *const p_mt_info,
673 MULTI_THREADED_MODULES mod_name) {
674 int num_mod_workers = p_mt_info->num_mod_workers[mod_name];
675 if (p_mt_info->num_mod_workers[MOD_FRAME_ENC] > 1) {
676 // TODO(anyone): Change num_mod_workers to num_mod_workers[MOD_FRAME_ENC].
677 // As frame parallel jobs will only perform multi-threading for the encode
678 // stage, we can limit the allocations according to num_enc_workers per
679 // frame parallel encode(a.k.a num_mod_workers[MOD_FRAME_ENC]).
680 num_mod_workers = p_mt_info->num_workers;
681 }
682 return num_mod_workers;
683 }
684
av1_init_tile_thread_data(AV1_PRIMARY * ppi,int is_first_pass)685 void av1_init_tile_thread_data(AV1_PRIMARY *ppi, int is_first_pass) {
686 PrimaryMultiThreadInfo *const p_mt_info = &ppi->p_mt_info;
687
688 assert(p_mt_info->workers != NULL);
689 assert(p_mt_info->tile_thr_data != NULL);
690
691 int num_workers = p_mt_info->num_workers;
692 int num_enc_workers = av1_get_num_mod_workers_for_alloc(p_mt_info, MOD_ENC);
693 for (int i = num_workers - 1; i >= 0; i--) {
694 EncWorkerData *const thread_data = &p_mt_info->tile_thr_data[i];
695
696 if (i > 0) {
697 // Allocate thread data.
698 AOM_CHECK_MEM_ERROR(&ppi->error, thread_data->td,
699 aom_memalign(32, sizeof(*thread_data->td)));
700 av1_zero(*thread_data->td);
701 #if CONFIG_FRAME_PARALLEL_ENCODE
702 thread_data->original_td = thread_data->td;
703 #endif
704
705 // Set up shared coeff buffers.
706 av1_setup_shared_coeff_buffer(
707 &ppi->seq_params, &thread_data->td->shared_coeff_buf, &ppi->error);
708 AOM_CHECK_MEM_ERROR(
709 &ppi->error, thread_data->td->tmp_conv_dst,
710 aom_memalign(32, MAX_SB_SIZE * MAX_SB_SIZE *
711 sizeof(*thread_data->td->tmp_conv_dst)));
712
713 if (i < p_mt_info->num_mod_workers[MOD_FP]) {
714 // Set up firstpass PICK_MODE_CONTEXT.
715 thread_data->td->firstpass_ctx = av1_alloc_pmc(
716 ppi->cpi, BLOCK_16X16, &thread_data->td->shared_coeff_buf);
717 }
718
719 if (!is_first_pass && i < num_enc_workers) {
720 // Set up sms_tree.
721 av1_setup_sms_tree(ppi->cpi, thread_data->td);
722
723 alloc_obmc_buffers(&thread_data->td->obmc_buffer, &ppi->error);
724
725 for (int x = 0; x < 2; x++)
726 for (int y = 0; y < 2; y++)
727 AOM_CHECK_MEM_ERROR(
728 &ppi->error, thread_data->td->hash_value_buffer[x][y],
729 (uint32_t *)aom_malloc(
730 AOM_BUFFER_SIZE_FOR_BLOCK_HASH *
731 sizeof(*thread_data->td->hash_value_buffer[0][0])));
732
733 // Allocate frame counters in thread data.
734 AOM_CHECK_MEM_ERROR(&ppi->error, thread_data->td->counts,
735 aom_calloc(1, sizeof(*thread_data->td->counts)));
736
737 // Allocate buffers used by palette coding mode.
738 AOM_CHECK_MEM_ERROR(
739 &ppi->error, thread_data->td->palette_buffer,
740 aom_memalign(16, sizeof(*thread_data->td->palette_buffer)));
741
742 alloc_compound_type_rd_buffers(&ppi->error,
743 &thread_data->td->comp_rd_buffer);
744
745 for (int j = 0; j < 2; ++j) {
746 AOM_CHECK_MEM_ERROR(
747 &ppi->error, thread_data->td->tmp_pred_bufs[j],
748 aom_memalign(32, 2 * MAX_MB_PLANE * MAX_SB_SQUARE *
749 sizeof(*thread_data->td->tmp_pred_bufs[j])));
750 }
751
752 const SPEED_FEATURES *sf = &ppi->cpi->sf;
753 if (sf->intra_sf.intra_pruning_with_hog ||
754 sf->intra_sf.chroma_intra_pruning_with_hog) {
755 const int plane_types = PLANE_TYPES >> ppi->seq_params.monochrome;
756 AOM_CHECK_MEM_ERROR(
757 &ppi->error, thread_data->td->pixel_gradient_info,
758 aom_malloc(sizeof(*thread_data->td->pixel_gradient_info) *
759 plane_types * MAX_SB_SQUARE));
760 }
761
762 if (ppi->cpi->sf.part_sf.partition_search_type == VAR_BASED_PARTITION) {
763 const int num_64x64_blocks =
764 (ppi->seq_params.sb_size == BLOCK_64X64) ? 1 : 4;
765 AOM_CHECK_MEM_ERROR(
766 &ppi->error, thread_data->td->vt64x64,
767 aom_malloc(sizeof(*thread_data->td->vt64x64) * num_64x64_blocks));
768 }
769 }
770 }
771
772 if (!is_first_pass && ppi->cpi->oxcf.row_mt == 1 && i < num_enc_workers) {
773 if (i == 0) {
774 #if CONFIG_FRAME_PARALLEL_ENCODE
775 for (int j = 0; j < ppi->num_fp_contexts; j++) {
776 AOM_CHECK_MEM_ERROR(&ppi->error, ppi->parallel_cpi[j]->td.tctx,
777 (FRAME_CONTEXT *)aom_memalign(
778 16, sizeof(*ppi->parallel_cpi[j]->td.tctx)));
779 }
780 #else
781 AOM_CHECK_MEM_ERROR(
782 &ppi->error, ppi->cpi->td.tctx,
783 (FRAME_CONTEXT *)aom_memalign(16, sizeof(*ppi->cpi->td.tctx)));
784 #endif
785 } else {
786 AOM_CHECK_MEM_ERROR(
787 &ppi->error, thread_data->td->tctx,
788 (FRAME_CONTEXT *)aom_memalign(16, sizeof(*thread_data->td->tctx)));
789 }
790 }
791 }
792 }
793
av1_create_workers(AV1_PRIMARY * ppi,int num_workers)794 void av1_create_workers(AV1_PRIMARY *ppi, int num_workers) {
795 PrimaryMultiThreadInfo *const p_mt_info = &ppi->p_mt_info;
796 const AVxWorkerInterface *const winterface = aom_get_worker_interface();
797
798 AOM_CHECK_MEM_ERROR(&ppi->error, p_mt_info->workers,
799 aom_malloc(num_workers * sizeof(*p_mt_info->workers)));
800
801 AOM_CHECK_MEM_ERROR(
802 &ppi->error, p_mt_info->tile_thr_data,
803 aom_calloc(num_workers, sizeof(*p_mt_info->tile_thr_data)));
804
805 for (int i = num_workers - 1; i >= 0; i--) {
806 AVxWorker *const worker = &p_mt_info->workers[i];
807 EncWorkerData *const thread_data = &p_mt_info->tile_thr_data[i];
808
809 winterface->init(worker);
810 worker->thread_name = "aom enc worker";
811
812 thread_data->thread_id = i;
813 // Set the starting tile for each thread.
814 thread_data->start = i;
815
816 if (i > 0) {
817 // Create threads
818 if (!winterface->reset(worker))
819 aom_internal_error(&ppi->error, AOM_CODEC_ERROR,
820 "Tile encoder thread creation failed");
821 }
822 winterface->sync(worker);
823
824 ++p_mt_info->num_workers;
825 }
826 }
827
828 #if CONFIG_FRAME_PARALLEL_ENCODE
829 // This function returns 1 if frame parallel encode is supported for
830 // the current configuration. Returns 0 otherwise.
is_fp_config(AV1_PRIMARY * ppi,AV1EncoderConfig * oxcf)831 static AOM_INLINE int is_fp_config(AV1_PRIMARY *ppi, AV1EncoderConfig *oxcf) {
832 // FPMT is enabled for AOM_Q and AOM_VBR.
833 // TODO(Mufaddal, Aasaipriya): Test and enable multi-tile and resize config.
834 if (oxcf->rc_cfg.mode == AOM_CBR || oxcf->rc_cfg.mode == AOM_CQ) {
835 return 0;
836 }
837 if (ppi->use_svc) {
838 return 0;
839 }
840 if (oxcf->tile_cfg.tile_columns > 0 || oxcf->tile_cfg.tile_rows > 0) {
841 return 0;
842 }
843 if (oxcf->dec_model_cfg.timing_info_present) {
844 return 0;
845 }
846 if (oxcf->mode != GOOD) {
847 return 0;
848 }
849 if (oxcf->tool_cfg.error_resilient_mode) {
850 return 0;
851 }
852 if (oxcf->resize_cfg.resize_mode) {
853 return 0;
854 }
855 if (oxcf->passes == 1) {
856 return 0;
857 }
858
859 return 1;
860 }
861
862 // A large value for threads used to compute the max num_enc_workers
863 // possible for each resolution.
864 #define MAX_THREADS 100
865
866 // Computes the number of frame parallel(fp) contexts to be created
867 // based on the number of max_enc_workers.
av1_compute_num_fp_contexts(AV1_PRIMARY * ppi,AV1EncoderConfig * oxcf)868 int av1_compute_num_fp_contexts(AV1_PRIMARY *ppi, AV1EncoderConfig *oxcf) {
869 ppi->p_mt_info.num_mod_workers[MOD_FRAME_ENC] = 0;
870 if (!is_fp_config(ppi, oxcf)) {
871 return 1;
872 }
873 int max_num_enc_workers =
874 av1_compute_num_enc_workers(ppi->parallel_cpi[0], MAX_THREADS);
875
876 // A parallel frame encode must have at least 1/4th the theoretical limit of
877 // max enc workers. TODO(Remya) : Tune this value for multi-tile case.
878 int workers_per_frame = AOMMAX(1, (max_num_enc_workers + 2) / 4);
879 int max_threads = oxcf->max_threads;
880 int num_fp_contexts = max_threads / workers_per_frame;
881
882 num_fp_contexts = AOMMAX(1, AOMMIN(num_fp_contexts, MAX_PARALLEL_FRAMES));
883 if (num_fp_contexts > 1) {
884 assert(max_threads >= 2);
885 ppi->p_mt_info.num_mod_workers[MOD_FRAME_ENC] =
886 AOMMIN(max_num_enc_workers * num_fp_contexts, oxcf->max_threads);
887 }
888 return num_fp_contexts;
889 }
890
891 // Computes the number of workers to process each of the parallel frames.
compute_num_workers_per_frame(const int num_workers,const int parallel_frame_count)892 static AOM_INLINE int compute_num_workers_per_frame(
893 const int num_workers, const int parallel_frame_count) {
894 // Number of level 2 workers per frame context (floor division).
895 int workers_per_frame = (num_workers / parallel_frame_count);
896 return workers_per_frame;
897 }
898
899 // Prepare level 1 workers. This function is only called for
900 // parallel_frame_count > 1. This function populates the mt_info structure of
901 // frame level contexts appropriately by dividing the total number of available
902 // workers amongst the frames as level 2 workers. It also populates the hook and
903 // data members of level 1 workers.
prepare_fpmt_workers(AV1_PRIMARY * ppi,AV1_COMP_DATA * first_cpi_data,AVxWorkerHook hook,int parallel_frame_count)904 static AOM_INLINE void prepare_fpmt_workers(AV1_PRIMARY *ppi,
905 AV1_COMP_DATA *first_cpi_data,
906 AVxWorkerHook hook,
907 int parallel_frame_count) {
908 assert(parallel_frame_count <= ppi->num_fp_contexts &&
909 parallel_frame_count > 1);
910
911 PrimaryMultiThreadInfo *const p_mt_info = &ppi->p_mt_info;
912 int num_workers = p_mt_info->num_workers;
913
914 int frame_idx = 0;
915 int i = 0;
916 while (i < num_workers) {
917 // Assign level 1 worker
918 AVxWorker *frame_worker = p_mt_info->p_workers[frame_idx] =
919 &p_mt_info->workers[i];
920 AV1_COMP *cur_cpi = ppi->parallel_cpi[frame_idx];
921 MultiThreadInfo *mt_info = &cur_cpi->mt_info;
922 const int num_planes = av1_num_planes(&cur_cpi->common);
923
924 // Assign start of level 2 worker pool
925 mt_info->workers = &p_mt_info->workers[i];
926 mt_info->tile_thr_data = &p_mt_info->tile_thr_data[i];
927 // Assign number of workers for each frame in the parallel encode set.
928 mt_info->num_workers = compute_num_workers_per_frame(
929 num_workers - i, parallel_frame_count - frame_idx);
930 for (int j = MOD_FP; j < NUM_MT_MODULES; j++) {
931 mt_info->num_mod_workers[j] =
932 AOMMIN(mt_info->num_workers, ppi->p_mt_info.num_mod_workers[j]);
933 }
934 if (ppi->p_mt_info.cdef_worker != NULL) {
935 mt_info->cdef_worker = &ppi->p_mt_info.cdef_worker[i];
936
937 // Back up the original cdef_worker pointers.
938 mt_info->restore_state_buf.cdef_srcbuf = mt_info->cdef_worker->srcbuf;
939 for (int plane = 0; plane < num_planes; plane++)
940 mt_info->restore_state_buf.cdef_colbuf[plane] =
941 mt_info->cdef_worker->colbuf[plane];
942 }
943 #if !CONFIG_REALTIME_ONLY
944 // Back up the original LR buffers before update.
945 int idx = i + mt_info->num_workers - 1;
946 mt_info->restore_state_buf.rst_tmpbuf =
947 mt_info->lr_row_sync.lrworkerdata[idx].rst_tmpbuf;
948 mt_info->restore_state_buf.rlbs =
949 mt_info->lr_row_sync.lrworkerdata[idx].rlbs;
950
951 // Update LR buffers.
952 mt_info->lr_row_sync.lrworkerdata[idx].rst_tmpbuf =
953 cur_cpi->common.rst_tmpbuf;
954 mt_info->lr_row_sync.lrworkerdata[idx].rlbs = cur_cpi->common.rlbs;
955 #endif
956
957 // At this stage, the thread specific CDEF buffers for the current frame's
958 // 'common' and 'cdef_sync' only need to be allocated. 'cdef_worker' has
959 // already been allocated across parallel frames.
960 av1_alloc_cdef_buffers(&cur_cpi->common, &p_mt_info->cdef_worker,
961 &mt_info->cdef_sync, p_mt_info->num_workers, 0);
962
963 frame_worker->hook = hook;
964 frame_worker->data1 = cur_cpi;
965 frame_worker->data2 = (frame_idx == 0)
966 ? first_cpi_data
967 : &ppi->parallel_frames_data[frame_idx - 1];
968 frame_idx++;
969 i += mt_info->num_workers;
970 }
971 p_mt_info->p_num_workers = parallel_frame_count;
972 }
973
974 // Launch level 1 workers to perform frame parallel encode.
launch_fpmt_workers(AV1_PRIMARY * ppi)975 static AOM_INLINE void launch_fpmt_workers(AV1_PRIMARY *ppi) {
976 const AVxWorkerInterface *const winterface = aom_get_worker_interface();
977 int num_workers = ppi->p_mt_info.p_num_workers;
978
979 for (int i = num_workers - 1; i >= 0; i--) {
980 AVxWorker *const worker = ppi->p_mt_info.p_workers[i];
981 if (i == 0)
982 winterface->execute(worker);
983 else
984 winterface->launch(worker);
985 }
986 }
987
988 // Synchronize level 1 workers.
sync_fpmt_workers(AV1_PRIMARY * ppi)989 static AOM_INLINE void sync_fpmt_workers(AV1_PRIMARY *ppi) {
990 const AVxWorkerInterface *const winterface = aom_get_worker_interface();
991 int num_workers = ppi->p_mt_info.p_num_workers;
992 int had_error = 0;
993 // Points to error in the earliest display order frame in the parallel set.
994 const struct aom_internal_error_info *error;
995
996 // Encoding ends.
997 for (int i = num_workers - 1; i >= 0; i--) {
998 AVxWorker *const worker = ppi->p_mt_info.p_workers[i];
999 if (!winterface->sync(worker)) {
1000 had_error = 1;
1001 error = ((AV1_COMP *)worker->data1)->common.error;
1002 }
1003 }
1004
1005 if (had_error)
1006 aom_internal_error(&ppi->error, error->error_code, error->detail);
1007 }
1008
1009 // Restore worker states after parallel encode.
restore_workers_after_fpmt(AV1_PRIMARY * ppi,int parallel_frame_count)1010 static AOM_INLINE void restore_workers_after_fpmt(AV1_PRIMARY *ppi,
1011 int parallel_frame_count) {
1012 assert(parallel_frame_count <= ppi->num_fp_contexts &&
1013 parallel_frame_count > 1);
1014 (void)parallel_frame_count;
1015
1016 PrimaryMultiThreadInfo *const p_mt_info = &ppi->p_mt_info;
1017 int num_workers = p_mt_info->num_workers;
1018
1019 int frame_idx = 0;
1020 int i = 0;
1021 while (i < num_workers) {
1022 AV1_COMP *cur_cpi = ppi->parallel_cpi[frame_idx];
1023 MultiThreadInfo *mt_info = &cur_cpi->mt_info;
1024 const int num_planes = av1_num_planes(&cur_cpi->common);
1025
1026 // Restore the original cdef_worker pointers.
1027 if (ppi->p_mt_info.cdef_worker != NULL) {
1028 mt_info->cdef_worker->srcbuf = mt_info->restore_state_buf.cdef_srcbuf;
1029 for (int plane = 0; plane < num_planes; plane++)
1030 mt_info->cdef_worker->colbuf[plane] =
1031 mt_info->restore_state_buf.cdef_colbuf[plane];
1032 }
1033 #if !CONFIG_REALTIME_ONLY
1034 // Restore the original LR buffers.
1035 int idx = i + mt_info->num_workers - 1;
1036 mt_info->lr_row_sync.lrworkerdata[idx].rst_tmpbuf =
1037 mt_info->restore_state_buf.rst_tmpbuf;
1038 mt_info->lr_row_sync.lrworkerdata[idx].rlbs =
1039 mt_info->restore_state_buf.rlbs;
1040 #endif
1041
1042 frame_idx++;
1043 i += mt_info->num_workers;
1044 }
1045 }
1046
get_compressed_data_hook(void * arg1,void * arg2)1047 static int get_compressed_data_hook(void *arg1, void *arg2) {
1048 AV1_COMP *cpi = (AV1_COMP *)arg1;
1049 AV1_COMP_DATA *cpi_data = (AV1_COMP_DATA *)arg2;
1050 int status = av1_get_compressed_data(cpi, cpi_data);
1051
1052 // AOM_CODEC_OK(0) means no error.
1053 return !status;
1054 }
1055
1056 // This function encodes the raw frame data for each frame in parallel encode
1057 // set, and outputs the frame bit stream to the designated buffers.
av1_compress_parallel_frames(AV1_PRIMARY * const ppi,AV1_COMP_DATA * const first_cpi_data)1058 int av1_compress_parallel_frames(AV1_PRIMARY *const ppi,
1059 AV1_COMP_DATA *const first_cpi_data) {
1060 // Bitmask for the frame buffers referenced by cpi->scaled_ref_buf
1061 // corresponding to frames in the current parallel encode set.
1062 int ref_buffers_used_map = 0;
1063 int frames_in_parallel_set = av1_init_parallel_frame_context(
1064 first_cpi_data, ppi, &ref_buffers_used_map);
1065 prepare_fpmt_workers(ppi, first_cpi_data, get_compressed_data_hook,
1066 frames_in_parallel_set);
1067 launch_fpmt_workers(ppi);
1068 sync_fpmt_workers(ppi);
1069 restore_workers_after_fpmt(ppi, frames_in_parallel_set);
1070
1071 // Release cpi->scaled_ref_buf corresponding to frames in the current parallel
1072 // encode set.
1073 for (int i = 0; i < frames_in_parallel_set; ++i) {
1074 av1_release_scaled_references_fpmt(ppi->parallel_cpi[i]);
1075 }
1076 av1_decrement_ref_counts_fpmt(ppi->cpi->common.buffer_pool,
1077 ref_buffers_used_map);
1078 return AOM_CODEC_OK;
1079 }
1080 #endif // CONFIG_FRAME_PARALLEL_ENCODE
1081
launch_workers(MultiThreadInfo * const mt_info,int num_workers)1082 static AOM_INLINE void launch_workers(MultiThreadInfo *const mt_info,
1083 int num_workers) {
1084 const AVxWorkerInterface *const winterface = aom_get_worker_interface();
1085 for (int i = num_workers - 1; i >= 0; i--) {
1086 AVxWorker *const worker = &mt_info->workers[i];
1087 if (i == 0)
1088 winterface->execute(worker);
1089 else
1090 winterface->launch(worker);
1091 }
1092 }
1093
sync_enc_workers(MultiThreadInfo * const mt_info,AV1_COMMON * const cm,int num_workers)1094 static AOM_INLINE void sync_enc_workers(MultiThreadInfo *const mt_info,
1095 AV1_COMMON *const cm, int num_workers) {
1096 const AVxWorkerInterface *const winterface = aom_get_worker_interface();
1097 int had_error = 0;
1098
1099 // Encoding ends.
1100 for (int i = num_workers - 1; i > 0; i--) {
1101 AVxWorker *const worker = &mt_info->workers[i];
1102 had_error |= !winterface->sync(worker);
1103 }
1104
1105 if (had_error)
1106 aom_internal_error(cm->error, AOM_CODEC_ERROR,
1107 "Failed to encode tile data");
1108 }
1109
accumulate_counters_enc_workers(AV1_COMP * cpi,int num_workers)1110 static AOM_INLINE void accumulate_counters_enc_workers(AV1_COMP *cpi,
1111 int num_workers) {
1112 for (int i = num_workers - 1; i >= 0; i--) {
1113 AVxWorker *const worker = &cpi->mt_info.workers[i];
1114 EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
1115 cpi->intrabc_used |= thread_data->td->intrabc_used;
1116 cpi->deltaq_used |= thread_data->td->deltaq_used;
1117 // Accumulate cyclic refresh params.
1118 if (cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ &&
1119 !frame_is_intra_only(&cpi->common))
1120 av1_accumulate_cyclic_refresh_counters(cpi->cyclic_refresh,
1121 &thread_data->td->mb);
1122 if (thread_data->td != &cpi->td) {
1123 if (cpi->oxcf.cost_upd_freq.mv < COST_UPD_OFF) {
1124 aom_free(thread_data->td->mb.mv_costs);
1125 }
1126 if (cpi->oxcf.cost_upd_freq.dv < COST_UPD_OFF) {
1127 aom_free(thread_data->td->mb.dv_costs);
1128 }
1129 }
1130 av1_dealloc_mb_data(&cpi->common, &thread_data->td->mb);
1131
1132 // Accumulate counters.
1133 if (i > 0) {
1134 av1_accumulate_frame_counts(&cpi->counts, thread_data->td->counts);
1135 accumulate_rd_opt(&cpi->td, thread_data->td);
1136 cpi->td.mb.txfm_search_info.txb_split_count +=
1137 thread_data->td->mb.txfm_search_info.txb_split_count;
1138 #if CONFIG_SPEED_STATS
1139 cpi->td.mb.txfm_search_info.tx_search_count +=
1140 thread_data->td->mb.txfm_search_info.tx_search_count;
1141 #endif // CONFIG_SPEED_STATS
1142 }
1143 }
1144 }
1145
prepare_enc_workers(AV1_COMP * cpi,AVxWorkerHook hook,int num_workers)1146 static AOM_INLINE void prepare_enc_workers(AV1_COMP *cpi, AVxWorkerHook hook,
1147 int num_workers) {
1148 MultiThreadInfo *const mt_info = &cpi->mt_info;
1149 AV1_COMMON *const cm = &cpi->common;
1150 for (int i = num_workers - 1; i >= 0; i--) {
1151 AVxWorker *const worker = &mt_info->workers[i];
1152 EncWorkerData *const thread_data = &mt_info->tile_thr_data[i];
1153
1154 worker->hook = hook;
1155 worker->data1 = thread_data;
1156 worker->data2 = NULL;
1157
1158 thread_data->thread_id = i;
1159 // Set the starting tile for each thread.
1160 thread_data->start = i;
1161
1162 thread_data->cpi = cpi;
1163 if (i == 0) {
1164 thread_data->td = &cpi->td;
1165 #if !CONFIG_FRAME_PARALLEL_ENCODE
1166 }
1167 #else
1168 } else {
1169 thread_data->td = thread_data->original_td;
1170 }
1171 #endif // CONFIG_FRAME_PARALLEL_ENCODE
1172
1173 thread_data->td->intrabc_used = 0;
1174 thread_data->td->deltaq_used = 0;
1175 thread_data->td->abs_sum_level = 0;
1176
1177 // Before encoding a frame, copy the thread data from cpi.
1178 if (thread_data->td != &cpi->td) {
1179 thread_data->td->mb = cpi->td.mb;
1180 thread_data->td->rd_counts = cpi->td.rd_counts;
1181 thread_data->td->mb.obmc_buffer = thread_data->td->obmc_buffer;
1182
1183 for (int x = 0; x < 2; x++) {
1184 for (int y = 0; y < 2; y++) {
1185 memcpy(thread_data->td->hash_value_buffer[x][y],
1186 cpi->td.mb.intrabc_hash_info.hash_value_buffer[x][y],
1187 AOM_BUFFER_SIZE_FOR_BLOCK_HASH *
1188 sizeof(*thread_data->td->hash_value_buffer[0][0]));
1189 thread_data->td->mb.intrabc_hash_info.hash_value_buffer[x][y] =
1190 thread_data->td->hash_value_buffer[x][y];
1191 }
1192 }
1193 if (cpi->oxcf.cost_upd_freq.mv < COST_UPD_OFF) {
1194 CHECK_MEM_ERROR(cm, thread_data->td->mb.mv_costs,
1195 (MvCosts *)aom_malloc(sizeof(MvCosts)));
1196 memcpy(thread_data->td->mb.mv_costs, cpi->td.mb.mv_costs,
1197 sizeof(MvCosts));
1198 }
1199 if (cpi->oxcf.cost_upd_freq.dv < COST_UPD_OFF) {
1200 CHECK_MEM_ERROR(cm, thread_data->td->mb.dv_costs,
1201 (IntraBCMVCosts *)aom_malloc(sizeof(IntraBCMVCosts)));
1202 memcpy(thread_data->td->mb.dv_costs, cpi->td.mb.dv_costs,
1203 sizeof(IntraBCMVCosts));
1204 }
1205 }
1206 av1_alloc_mb_data(cm, &thread_data->td->mb,
1207 cpi->sf.rt_sf.use_nonrd_pick_mode);
1208
1209 // Reset cyclic refresh counters.
1210 av1_init_cyclic_refresh_counters(&thread_data->td->mb);
1211
1212 if (thread_data->td->counts != &cpi->counts) {
1213 memcpy(thread_data->td->counts, &cpi->counts, sizeof(cpi->counts));
1214 }
1215
1216 if (i > 0) {
1217 thread_data->td->mb.palette_buffer = thread_data->td->palette_buffer;
1218 thread_data->td->mb.comp_rd_buffer = thread_data->td->comp_rd_buffer;
1219 thread_data->td->mb.tmp_conv_dst = thread_data->td->tmp_conv_dst;
1220 for (int j = 0; j < 2; ++j) {
1221 thread_data->td->mb.tmp_pred_bufs[j] =
1222 thread_data->td->tmp_pred_bufs[j];
1223 }
1224 thread_data->td->mb.pixel_gradient_info =
1225 thread_data->td->pixel_gradient_info;
1226
1227 thread_data->td->mb.e_mbd.tmp_conv_dst = thread_data->td->mb.tmp_conv_dst;
1228 for (int j = 0; j < 2; ++j) {
1229 thread_data->td->mb.e_mbd.tmp_obmc_bufs[j] =
1230 thread_data->td->mb.tmp_pred_bufs[j];
1231 }
1232 }
1233 }
1234 }
1235
1236 #if !CONFIG_REALTIME_ONLY
fp_prepare_enc_workers(AV1_COMP * cpi,AVxWorkerHook hook,int num_workers)1237 static AOM_INLINE void fp_prepare_enc_workers(AV1_COMP *cpi, AVxWorkerHook hook,
1238 int num_workers) {
1239 AV1_COMMON *const cm = &cpi->common;
1240 MultiThreadInfo *const mt_info = &cpi->mt_info;
1241 for (int i = num_workers - 1; i >= 0; i--) {
1242 AVxWorker *const worker = &mt_info->workers[i];
1243 EncWorkerData *const thread_data = &mt_info->tile_thr_data[i];
1244
1245 worker->hook = hook;
1246 worker->data1 = thread_data;
1247 worker->data2 = NULL;
1248
1249 thread_data->thread_id = i;
1250 // Set the starting tile for each thread.
1251 thread_data->start = i;
1252
1253 thread_data->cpi = cpi;
1254 if (i == 0) {
1255 thread_data->td = &cpi->td;
1256 #if !CONFIG_FRAME_PARALLEL_ENCODE
1257 }
1258 #else
1259 } else {
1260 thread_data->td = thread_data->original_td;
1261 }
1262 #endif // CONFIG_FRAME_PARALLEL_ENCODE
1263
1264 // Before encoding a frame, copy the thread data from cpi.
1265 if (thread_data->td != &cpi->td) {
1266 thread_data->td->mb = cpi->td.mb;
1267 if (cpi->oxcf.cost_upd_freq.mv < COST_UPD_OFF) {
1268 CHECK_MEM_ERROR(cm, thread_data->td->mb.mv_costs,
1269 (MvCosts *)aom_malloc(sizeof(MvCosts)));
1270 memcpy(thread_data->td->mb.mv_costs, cpi->td.mb.mv_costs,
1271 sizeof(MvCosts));
1272 }
1273 if (cpi->oxcf.cost_upd_freq.dv < COST_UPD_OFF) {
1274 CHECK_MEM_ERROR(cm, thread_data->td->mb.dv_costs,
1275 (IntraBCMVCosts *)aom_malloc(sizeof(IntraBCMVCosts)));
1276 memcpy(thread_data->td->mb.dv_costs, cpi->td.mb.dv_costs,
1277 sizeof(IntraBCMVCosts));
1278 }
1279 }
1280
1281 av1_alloc_mb_data(cm, &thread_data->td->mb,
1282 cpi->sf.rt_sf.use_nonrd_pick_mode);
1283 }
1284 }
1285 #endif
1286
1287 // Computes the number of workers for row multi-threading of encoding stage
compute_num_enc_row_mt_workers(AV1_COMMON * const cm,int max_threads)1288 static AOM_INLINE int compute_num_enc_row_mt_workers(AV1_COMMON *const cm,
1289 int max_threads) {
1290 TileInfo tile_info;
1291 const int tile_cols = cm->tiles.cols;
1292 const int tile_rows = cm->tiles.rows;
1293 int total_num_threads_row_mt = 0;
1294 for (int row = 0; row < tile_rows; row++) {
1295 for (int col = 0; col < tile_cols; col++) {
1296 av1_tile_init(&tile_info, cm, row, col);
1297 const int num_sb_rows_in_tile = av1_get_sb_rows_in_tile(cm, tile_info);
1298 const int num_sb_cols_in_tile = av1_get_sb_cols_in_tile(cm, tile_info);
1299 total_num_threads_row_mt +=
1300 AOMMIN((num_sb_cols_in_tile + 1) >> 1, num_sb_rows_in_tile);
1301 }
1302 }
1303 return AOMMIN(max_threads, total_num_threads_row_mt);
1304 }
1305
1306 // Computes the number of workers for tile multi-threading of encoding stage
compute_num_enc_tile_mt_workers(AV1_COMMON * const cm,int max_threads)1307 static AOM_INLINE int compute_num_enc_tile_mt_workers(AV1_COMMON *const cm,
1308 int max_threads) {
1309 const int tile_cols = cm->tiles.cols;
1310 const int tile_rows = cm->tiles.rows;
1311 return AOMMIN(max_threads, tile_cols * tile_rows);
1312 }
1313
1314 // Find max worker of all MT stages
av1_get_max_num_workers(AV1_COMP * cpi)1315 int av1_get_max_num_workers(AV1_COMP *cpi) {
1316 int max_num_workers = 0;
1317 for (int i = MOD_FP; i < NUM_MT_MODULES; i++)
1318 max_num_workers =
1319 AOMMAX(cpi->ppi->p_mt_info.num_mod_workers[i], max_num_workers);
1320 assert(max_num_workers >= 1);
1321 return AOMMIN(max_num_workers, cpi->oxcf.max_threads);
1322 }
1323
1324 // Computes the number of workers for encoding stage (row/tile multi-threading)
av1_compute_num_enc_workers(AV1_COMP * cpi,int max_workers)1325 int av1_compute_num_enc_workers(AV1_COMP *cpi, int max_workers) {
1326 if (max_workers <= 1) return 1;
1327 if (cpi->oxcf.row_mt)
1328 return compute_num_enc_row_mt_workers(&cpi->common, max_workers);
1329 else
1330 return compute_num_enc_tile_mt_workers(&cpi->common, max_workers);
1331 }
1332
av1_encode_tiles_mt(AV1_COMP * cpi)1333 void av1_encode_tiles_mt(AV1_COMP *cpi) {
1334 AV1_COMMON *const cm = &cpi->common;
1335 MultiThreadInfo *const mt_info = &cpi->mt_info;
1336 const int tile_cols = cm->tiles.cols;
1337 const int tile_rows = cm->tiles.rows;
1338 int num_workers = mt_info->num_mod_workers[MOD_ENC];
1339
1340 assert(IMPLIES(cpi->tile_data == NULL,
1341 cpi->allocated_tiles < tile_cols * tile_rows));
1342 if (cpi->allocated_tiles < tile_cols * tile_rows) av1_alloc_tile_data(cpi);
1343
1344 av1_init_tile_data(cpi);
1345 num_workers = AOMMIN(num_workers, mt_info->num_workers);
1346
1347 prepare_enc_workers(cpi, enc_worker_hook, num_workers);
1348 launch_workers(&cpi->mt_info, num_workers);
1349 sync_enc_workers(&cpi->mt_info, cm, num_workers);
1350 accumulate_counters_enc_workers(cpi, num_workers);
1351 }
1352
1353 // Accumulate frame counts. FRAME_COUNTS consist solely of 'unsigned int'
1354 // members, so we treat it as an array, and sum over the whole length.
av1_accumulate_frame_counts(FRAME_COUNTS * acc_counts,const FRAME_COUNTS * counts)1355 void av1_accumulate_frame_counts(FRAME_COUNTS *acc_counts,
1356 const FRAME_COUNTS *counts) {
1357 unsigned int *const acc = (unsigned int *)acc_counts;
1358 const unsigned int *const cnt = (const unsigned int *)counts;
1359
1360 const unsigned int n_counts = sizeof(FRAME_COUNTS) / sizeof(unsigned int);
1361
1362 for (unsigned int i = 0; i < n_counts; i++) acc[i] += cnt[i];
1363 }
1364
1365 // Computes the maximum number of sb_rows for row multi-threading of encoding
1366 // stage
compute_max_sb_rows_cols(AV1_COMP * cpi,int * max_sb_rows,int * max_sb_cols)1367 static AOM_INLINE void compute_max_sb_rows_cols(AV1_COMP *cpi, int *max_sb_rows,
1368 int *max_sb_cols) {
1369 AV1_COMMON *const cm = &cpi->common;
1370 const int tile_cols = cm->tiles.cols;
1371 const int tile_rows = cm->tiles.rows;
1372 for (int row = 0; row < tile_rows; row++) {
1373 for (int col = 0; col < tile_cols; col++) {
1374 const int tile_index = row * cm->tiles.cols + col;
1375 TileInfo tile_info = cpi->tile_data[tile_index].tile_info;
1376 const int num_sb_rows_in_tile = av1_get_sb_rows_in_tile(cm, tile_info);
1377 const int num_sb_cols_in_tile = av1_get_sb_cols_in_tile(cm, tile_info);
1378 *max_sb_rows = AOMMAX(*max_sb_rows, num_sb_rows_in_tile);
1379 *max_sb_cols = AOMMAX(*max_sb_cols, num_sb_cols_in_tile);
1380 }
1381 }
1382 }
1383
1384 #if !CONFIG_REALTIME_ONLY
1385 // Computes the number of workers for firstpass stage (row/tile multi-threading)
av1_fp_compute_num_enc_workers(AV1_COMP * cpi)1386 int av1_fp_compute_num_enc_workers(AV1_COMP *cpi) {
1387 AV1_COMMON *cm = &cpi->common;
1388 const int tile_cols = cm->tiles.cols;
1389 const int tile_rows = cm->tiles.rows;
1390 int total_num_threads_row_mt = 0;
1391 TileInfo tile_info;
1392
1393 if (cpi->oxcf.max_threads <= 1) return 1;
1394
1395 for (int row = 0; row < tile_rows; row++) {
1396 for (int col = 0; col < tile_cols; col++) {
1397 av1_tile_init(&tile_info, cm, row, col);
1398 const int num_mb_rows_in_tile =
1399 av1_get_unit_rows_in_tile(tile_info, cpi->fp_block_size);
1400 const int num_mb_cols_in_tile =
1401 av1_get_unit_cols_in_tile(tile_info, cpi->fp_block_size);
1402 total_num_threads_row_mt +=
1403 AOMMIN((num_mb_cols_in_tile + 1) >> 1, num_mb_rows_in_tile);
1404 }
1405 }
1406 return AOMMIN(cpi->oxcf.max_threads, total_num_threads_row_mt);
1407 }
1408
1409 // Computes the maximum number of mb_rows for row multi-threading of firstpass
1410 // stage
fp_compute_max_mb_rows(const AV1_COMMON * const cm,const TileDataEnc * const tile_data,const BLOCK_SIZE fp_block_size)1411 static AOM_INLINE int fp_compute_max_mb_rows(const AV1_COMMON *const cm,
1412 const TileDataEnc *const tile_data,
1413 const BLOCK_SIZE fp_block_size) {
1414 const int tile_cols = cm->tiles.cols;
1415 const int tile_rows = cm->tiles.rows;
1416 int max_mb_rows = 0;
1417 for (int row = 0; row < tile_rows; row++) {
1418 for (int col = 0; col < tile_cols; col++) {
1419 const int tile_index = row * cm->tiles.cols + col;
1420 TileInfo tile_info = tile_data[tile_index].tile_info;
1421 const int num_mb_rows_in_tile =
1422 av1_get_unit_rows_in_tile(tile_info, fp_block_size);
1423 max_mb_rows = AOMMAX(max_mb_rows, num_mb_rows_in_tile);
1424 }
1425 }
1426 return max_mb_rows;
1427 }
1428 #endif
1429
av1_encode_tiles_row_mt(AV1_COMP * cpi)1430 void av1_encode_tiles_row_mt(AV1_COMP *cpi) {
1431 AV1_COMMON *const cm = &cpi->common;
1432 MultiThreadInfo *const mt_info = &cpi->mt_info;
1433 AV1EncRowMultiThreadInfo *const enc_row_mt = &mt_info->enc_row_mt;
1434 const int tile_cols = cm->tiles.cols;
1435 const int tile_rows = cm->tiles.rows;
1436 int *thread_id_to_tile_id = enc_row_mt->thread_id_to_tile_id;
1437 int max_sb_rows = 0, max_sb_cols = 0;
1438 int num_workers = mt_info->num_mod_workers[MOD_ENC];
1439
1440 assert(IMPLIES(cpi->tile_data == NULL,
1441 cpi->allocated_tiles < tile_cols * tile_rows));
1442 if (cpi->allocated_tiles < tile_cols * tile_rows) {
1443 av1_row_mt_mem_dealloc(cpi);
1444 av1_alloc_tile_data(cpi);
1445 }
1446
1447 av1_init_tile_data(cpi);
1448
1449 compute_max_sb_rows_cols(cpi, &max_sb_rows, &max_sb_cols);
1450
1451 if (enc_row_mt->allocated_tile_cols != tile_cols ||
1452 enc_row_mt->allocated_tile_rows != tile_rows ||
1453 enc_row_mt->allocated_rows != max_sb_rows ||
1454 enc_row_mt->allocated_cols != (max_sb_cols - 1)) {
1455 av1_row_mt_mem_dealloc(cpi);
1456 row_mt_mem_alloc(cpi, max_sb_rows, max_sb_cols,
1457 cpi->oxcf.algo_cfg.cdf_update_mode);
1458 }
1459
1460 memset(thread_id_to_tile_id, -1,
1461 sizeof(*thread_id_to_tile_id) * MAX_NUM_THREADS);
1462
1463 for (int tile_row = 0; tile_row < tile_rows; tile_row++) {
1464 for (int tile_col = 0; tile_col < tile_cols; tile_col++) {
1465 int tile_index = tile_row * tile_cols + tile_col;
1466 TileDataEnc *const this_tile = &cpi->tile_data[tile_index];
1467 AV1EncRowMultiThreadSync *const row_mt_sync = &this_tile->row_mt_sync;
1468
1469 // Initialize num_finished_cols to -1 for all rows.
1470 memset(row_mt_sync->num_finished_cols, -1,
1471 sizeof(*row_mt_sync->num_finished_cols) * max_sb_rows);
1472 row_mt_sync->next_mi_row = this_tile->tile_info.mi_row_start;
1473 row_mt_sync->num_threads_working = 0;
1474
1475 av1_inter_mode_data_init(this_tile);
1476 av1_zero_above_context(cm, &cpi->td.mb.e_mbd,
1477 this_tile->tile_info.mi_col_start,
1478 this_tile->tile_info.mi_col_end, tile_row);
1479 }
1480 }
1481
1482 num_workers = AOMMIN(num_workers, mt_info->num_workers);
1483
1484 assign_tile_to_thread(thread_id_to_tile_id, tile_cols * tile_rows,
1485 num_workers);
1486 prepare_enc_workers(cpi, enc_row_mt_worker_hook, num_workers);
1487 launch_workers(&cpi->mt_info, num_workers);
1488 sync_enc_workers(&cpi->mt_info, cm, num_workers);
1489 if (cm->delta_q_info.delta_lf_present_flag) update_delta_lf_for_row_mt(cpi);
1490 accumulate_counters_enc_workers(cpi, num_workers);
1491 }
1492
1493 #if !CONFIG_REALTIME_ONLY
av1_fp_encode_tiles_row_mt(AV1_COMP * cpi)1494 void av1_fp_encode_tiles_row_mt(AV1_COMP *cpi) {
1495 AV1_COMMON *const cm = &cpi->common;
1496 MultiThreadInfo *const mt_info = &cpi->mt_info;
1497 AV1EncRowMultiThreadInfo *const enc_row_mt = &mt_info->enc_row_mt;
1498 const int tile_cols = cm->tiles.cols;
1499 const int tile_rows = cm->tiles.rows;
1500 int *thread_id_to_tile_id = enc_row_mt->thread_id_to_tile_id;
1501 int num_workers = 0;
1502 int max_mb_rows = 0;
1503
1504 assert(IMPLIES(cpi->tile_data == NULL,
1505 cpi->allocated_tiles < tile_cols * tile_rows));
1506 if (cpi->allocated_tiles < tile_cols * tile_rows) {
1507 av1_row_mt_mem_dealloc(cpi);
1508 av1_alloc_tile_data(cpi);
1509 }
1510
1511 av1_init_tile_data(cpi);
1512
1513 const BLOCK_SIZE fp_block_size = cpi->fp_block_size;
1514 max_mb_rows = fp_compute_max_mb_rows(cm, cpi->tile_data, fp_block_size);
1515
1516 // For pass = 1, compute the no. of workers needed. For single-pass encode
1517 // (pass = 0), no. of workers are already computed.
1518 if (mt_info->num_mod_workers[MOD_FP] == 0)
1519 num_workers = av1_fp_compute_num_enc_workers(cpi);
1520 else
1521 num_workers = mt_info->num_mod_workers[MOD_FP];
1522
1523 if (enc_row_mt->allocated_tile_cols != tile_cols ||
1524 enc_row_mt->allocated_tile_rows != tile_rows ||
1525 enc_row_mt->allocated_rows != max_mb_rows) {
1526 av1_row_mt_mem_dealloc(cpi);
1527 row_mt_mem_alloc(cpi, max_mb_rows, -1, 0);
1528 }
1529
1530 memset(thread_id_to_tile_id, -1,
1531 sizeof(*thread_id_to_tile_id) * MAX_NUM_THREADS);
1532
1533 for (int tile_row = 0; tile_row < tile_rows; tile_row++) {
1534 for (int tile_col = 0; tile_col < tile_cols; tile_col++) {
1535 int tile_index = tile_row * tile_cols + tile_col;
1536 TileDataEnc *const this_tile = &cpi->tile_data[tile_index];
1537 AV1EncRowMultiThreadSync *const row_mt_sync = &this_tile->row_mt_sync;
1538
1539 // Initialize num_finished_cols to -1 for all rows.
1540 memset(row_mt_sync->num_finished_cols, -1,
1541 sizeof(*row_mt_sync->num_finished_cols) * max_mb_rows);
1542 row_mt_sync->next_mi_row = this_tile->tile_info.mi_row_start;
1543 row_mt_sync->num_threads_working = 0;
1544 }
1545 }
1546
1547 num_workers = AOMMIN(num_workers, mt_info->num_workers);
1548 assign_tile_to_thread(thread_id_to_tile_id, tile_cols * tile_rows,
1549 num_workers);
1550 fp_prepare_enc_workers(cpi, fp_enc_row_mt_worker_hook, num_workers);
1551 launch_workers(&cpi->mt_info, num_workers);
1552 sync_enc_workers(&cpi->mt_info, cm, num_workers);
1553 for (int i = num_workers - 1; i >= 0; i--) {
1554 EncWorkerData *const thread_data = &cpi->mt_info.tile_thr_data[i];
1555 if (thread_data->td != &cpi->td) {
1556 if (cpi->oxcf.cost_upd_freq.mv < COST_UPD_OFF) {
1557 aom_free(thread_data->td->mb.mv_costs);
1558 }
1559 if (cpi->oxcf.cost_upd_freq.dv < COST_UPD_OFF) {
1560 aom_free(thread_data->td->mb.dv_costs);
1561 }
1562 }
1563 av1_dealloc_mb_data(cm, &thread_data->td->mb);
1564 }
1565 }
1566
av1_tpl_row_mt_sync_read_dummy(AV1TplRowMultiThreadSync * tpl_mt_sync,int r,int c)1567 void av1_tpl_row_mt_sync_read_dummy(AV1TplRowMultiThreadSync *tpl_mt_sync,
1568 int r, int c) {
1569 (void)tpl_mt_sync;
1570 (void)r;
1571 (void)c;
1572 return;
1573 }
1574
av1_tpl_row_mt_sync_write_dummy(AV1TplRowMultiThreadSync * tpl_mt_sync,int r,int c,int cols)1575 void av1_tpl_row_mt_sync_write_dummy(AV1TplRowMultiThreadSync *tpl_mt_sync,
1576 int r, int c, int cols) {
1577 (void)tpl_mt_sync;
1578 (void)r;
1579 (void)c;
1580 (void)cols;
1581 return;
1582 }
1583
av1_tpl_row_mt_sync_read(AV1TplRowMultiThreadSync * tpl_row_mt_sync,int r,int c)1584 void av1_tpl_row_mt_sync_read(AV1TplRowMultiThreadSync *tpl_row_mt_sync, int r,
1585 int c) {
1586 #if CONFIG_MULTITHREAD
1587 int nsync = tpl_row_mt_sync->sync_range;
1588
1589 if (r) {
1590 pthread_mutex_t *const mutex = &tpl_row_mt_sync->mutex_[r - 1];
1591 pthread_mutex_lock(mutex);
1592
1593 while (c > tpl_row_mt_sync->num_finished_cols[r - 1] - nsync)
1594 pthread_cond_wait(&tpl_row_mt_sync->cond_[r - 1], mutex);
1595 pthread_mutex_unlock(mutex);
1596 }
1597 #else
1598 (void)tpl_row_mt_sync;
1599 (void)r;
1600 (void)c;
1601 #endif // CONFIG_MULTITHREAD
1602 }
1603
av1_tpl_row_mt_sync_write(AV1TplRowMultiThreadSync * tpl_row_mt_sync,int r,int c,int cols)1604 void av1_tpl_row_mt_sync_write(AV1TplRowMultiThreadSync *tpl_row_mt_sync, int r,
1605 int c, int cols) {
1606 #if CONFIG_MULTITHREAD
1607 int nsync = tpl_row_mt_sync->sync_range;
1608 int cur;
1609 // Only signal when there are enough encoded blocks for next row to run.
1610 int sig = 1;
1611
1612 if (c < cols - 1) {
1613 cur = c;
1614 if (c % nsync) sig = 0;
1615 } else {
1616 cur = cols + nsync;
1617 }
1618
1619 if (sig) {
1620 pthread_mutex_lock(&tpl_row_mt_sync->mutex_[r]);
1621
1622 tpl_row_mt_sync->num_finished_cols[r] = cur;
1623
1624 pthread_cond_signal(&tpl_row_mt_sync->cond_[r]);
1625 pthread_mutex_unlock(&tpl_row_mt_sync->mutex_[r]);
1626 }
1627 #else
1628 (void)tpl_row_mt_sync;
1629 (void)r;
1630 (void)c;
1631 (void)cols;
1632 #endif // CONFIG_MULTITHREAD
1633 }
1634
1635 // Each worker calls tpl_worker_hook() and computes the tpl data.
tpl_worker_hook(void * arg1,void * unused)1636 static int tpl_worker_hook(void *arg1, void *unused) {
1637 (void)unused;
1638 EncWorkerData *thread_data = (EncWorkerData *)arg1;
1639 AV1_COMP *cpi = thread_data->cpi;
1640 AV1_COMMON *cm = &cpi->common;
1641 MACROBLOCK *x = &thread_data->td->mb;
1642 MACROBLOCKD *xd = &x->e_mbd;
1643 TplTxfmStats *tpl_txfm_stats = &thread_data->td->tpl_txfm_stats;
1644 CommonModeInfoParams *mi_params = &cm->mi_params;
1645 BLOCK_SIZE bsize = convert_length_to_bsize(cpi->ppi->tpl_data.tpl_bsize_1d);
1646 TX_SIZE tx_size = max_txsize_lookup[bsize];
1647 int mi_height = mi_size_high[bsize];
1648 int num_active_workers = cpi->ppi->tpl_data.tpl_mt_sync.num_threads_working;
1649
1650 av1_init_tpl_txfm_stats(tpl_txfm_stats);
1651
1652 for (int mi_row = thread_data->start * mi_height; mi_row < mi_params->mi_rows;
1653 mi_row += num_active_workers * mi_height) {
1654 // Motion estimation row boundary
1655 av1_set_mv_row_limits(mi_params, &x->mv_limits, mi_row, mi_height,
1656 cpi->oxcf.border_in_pixels);
1657 xd->mb_to_top_edge = -GET_MV_SUBPEL(mi_row * MI_SIZE);
1658 xd->mb_to_bottom_edge =
1659 GET_MV_SUBPEL((mi_params->mi_rows - mi_height - mi_row) * MI_SIZE);
1660 av1_mc_flow_dispenser_row(cpi, tpl_txfm_stats, x, mi_row, bsize, tx_size);
1661 }
1662 return 1;
1663 }
1664
1665 // Deallocate tpl synchronization related mutex and data.
av1_tpl_dealloc(AV1TplRowMultiThreadSync * tpl_sync)1666 void av1_tpl_dealloc(AV1TplRowMultiThreadSync *tpl_sync) {
1667 assert(tpl_sync != NULL);
1668
1669 #if CONFIG_MULTITHREAD
1670 if (tpl_sync->mutex_ != NULL) {
1671 for (int i = 0; i < tpl_sync->rows; ++i)
1672 pthread_mutex_destroy(&tpl_sync->mutex_[i]);
1673 aom_free(tpl_sync->mutex_);
1674 }
1675 if (tpl_sync->cond_ != NULL) {
1676 for (int i = 0; i < tpl_sync->rows; ++i)
1677 pthread_cond_destroy(&tpl_sync->cond_[i]);
1678 aom_free(tpl_sync->cond_);
1679 }
1680 #endif // CONFIG_MULTITHREAD
1681
1682 aom_free(tpl_sync->num_finished_cols);
1683 // clear the structure as the source of this call may be a resize in which
1684 // case this call will be followed by an _alloc() which may fail.
1685 av1_zero(*tpl_sync);
1686 }
1687
1688 // Allocate memory for tpl row synchronization.
av1_tpl_alloc(AV1TplRowMultiThreadSync * tpl_sync,AV1_COMMON * cm,int mb_rows)1689 void av1_tpl_alloc(AV1TplRowMultiThreadSync *tpl_sync, AV1_COMMON *cm,
1690 int mb_rows) {
1691 tpl_sync->rows = mb_rows;
1692 #if CONFIG_MULTITHREAD
1693 {
1694 CHECK_MEM_ERROR(cm, tpl_sync->mutex_,
1695 aom_malloc(sizeof(*tpl_sync->mutex_) * mb_rows));
1696 if (tpl_sync->mutex_) {
1697 for (int i = 0; i < mb_rows; ++i)
1698 pthread_mutex_init(&tpl_sync->mutex_[i], NULL);
1699 }
1700
1701 CHECK_MEM_ERROR(cm, tpl_sync->cond_,
1702 aom_malloc(sizeof(*tpl_sync->cond_) * mb_rows));
1703 if (tpl_sync->cond_) {
1704 for (int i = 0; i < mb_rows; ++i)
1705 pthread_cond_init(&tpl_sync->cond_[i], NULL);
1706 }
1707 }
1708 #endif // CONFIG_MULTITHREAD
1709 CHECK_MEM_ERROR(cm, tpl_sync->num_finished_cols,
1710 aom_malloc(sizeof(*tpl_sync->num_finished_cols) * mb_rows));
1711
1712 // Set up nsync.
1713 tpl_sync->sync_range = 1;
1714 }
1715
1716 // Each worker is prepared by assigning the hook function and individual thread
1717 // data.
prepare_tpl_workers(AV1_COMP * cpi,AVxWorkerHook hook,int num_workers)1718 static AOM_INLINE void prepare_tpl_workers(AV1_COMP *cpi, AVxWorkerHook hook,
1719 int num_workers) {
1720 MultiThreadInfo *mt_info = &cpi->mt_info;
1721 for (int i = num_workers - 1; i >= 0; i--) {
1722 AVxWorker *worker = &mt_info->workers[i];
1723 EncWorkerData *thread_data = &mt_info->tile_thr_data[i];
1724
1725 worker->hook = hook;
1726 worker->data1 = thread_data;
1727 worker->data2 = NULL;
1728
1729 thread_data->thread_id = i;
1730 // Set the starting tile for each thread.
1731 thread_data->start = i;
1732
1733 thread_data->cpi = cpi;
1734 if (i == 0) {
1735 thread_data->td = &cpi->td;
1736 #if !CONFIG_FRAME_PARALLEL_ENCODE
1737 }
1738 #else
1739 } else {
1740 thread_data->td = thread_data->original_td;
1741 }
1742 #endif // CONFIG_FRAME_PARALLEL_ENCODE
1743
1744 // Before encoding a frame, copy the thread data from cpi.
1745 if (thread_data->td != &cpi->td) {
1746 thread_data->td->mb = cpi->td.mb;
1747 // OBMC buffers are used only to init MS params and remain unused when
1748 // called from tpl, hence set the buffers to defaults.
1749 av1_init_obmc_buffer(&thread_data->td->mb.obmc_buffer);
1750 thread_data->td->mb.tmp_conv_dst = thread_data->td->tmp_conv_dst;
1751 thread_data->td->mb.e_mbd.tmp_conv_dst = thread_data->td->mb.tmp_conv_dst;
1752 }
1753 }
1754 }
1755
1756 // Accumulate transform stats after tpl.
tpl_accumulate_txfm_stats(ThreadData * main_td,const MultiThreadInfo * mt_info,int num_workers)1757 static void tpl_accumulate_txfm_stats(ThreadData *main_td,
1758 const MultiThreadInfo *mt_info,
1759 int num_workers) {
1760 TplTxfmStats *accumulated_stats = &main_td->tpl_txfm_stats;
1761 for (int i = num_workers - 1; i >= 0; i--) {
1762 AVxWorker *const worker = &mt_info->workers[i];
1763 EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
1764 ThreadData *td = thread_data->td;
1765 if (td != main_td) {
1766 const TplTxfmStats *tpl_txfm_stats = &td->tpl_txfm_stats;
1767 av1_accumulate_tpl_txfm_stats(tpl_txfm_stats, accumulated_stats);
1768 }
1769 }
1770 }
1771
1772 // Implements multi-threading for tpl.
av1_mc_flow_dispenser_mt(AV1_COMP * cpi)1773 void av1_mc_flow_dispenser_mt(AV1_COMP *cpi) {
1774 AV1_COMMON *cm = &cpi->common;
1775 CommonModeInfoParams *mi_params = &cm->mi_params;
1776 MultiThreadInfo *mt_info = &cpi->mt_info;
1777 TplParams *tpl_data = &cpi->ppi->tpl_data;
1778 AV1TplRowMultiThreadSync *tpl_sync = &tpl_data->tpl_mt_sync;
1779 int mb_rows = mi_params->mb_rows;
1780 int num_workers =
1781 AOMMIN(mt_info->num_mod_workers[MOD_TPL], mt_info->num_workers);
1782
1783 if (mb_rows != tpl_sync->rows) {
1784 av1_tpl_dealloc(tpl_sync);
1785 av1_tpl_alloc(tpl_sync, cm, mb_rows);
1786 }
1787 tpl_sync->num_threads_working = num_workers;
1788
1789 // Initialize cur_mb_col to -1 for all MB rows.
1790 memset(tpl_sync->num_finished_cols, -1,
1791 sizeof(*tpl_sync->num_finished_cols) * mb_rows);
1792
1793 prepare_tpl_workers(cpi, tpl_worker_hook, num_workers);
1794 launch_workers(&cpi->mt_info, num_workers);
1795 sync_enc_workers(&cpi->mt_info, cm, num_workers);
1796 tpl_accumulate_txfm_stats(&cpi->td, &cpi->mt_info, num_workers);
1797 }
1798
1799 // Deallocate memory for temporal filter multi-thread synchronization.
av1_tf_mt_dealloc(AV1TemporalFilterSync * tf_sync)1800 void av1_tf_mt_dealloc(AV1TemporalFilterSync *tf_sync) {
1801 assert(tf_sync != NULL);
1802 #if CONFIG_MULTITHREAD
1803 if (tf_sync->mutex_ != NULL) {
1804 pthread_mutex_destroy(tf_sync->mutex_);
1805 aom_free(tf_sync->mutex_);
1806 }
1807 #endif // CONFIG_MULTITHREAD
1808 tf_sync->next_tf_row = 0;
1809 }
1810
1811 // Checks if a job is available. If job is available,
1812 // populates next_tf_row and returns 1, else returns 0.
tf_get_next_job(AV1TemporalFilterSync * tf_mt_sync,int * current_mb_row,int mb_rows)1813 static AOM_INLINE int tf_get_next_job(AV1TemporalFilterSync *tf_mt_sync,
1814 int *current_mb_row, int mb_rows) {
1815 int do_next_row = 0;
1816 #if CONFIG_MULTITHREAD
1817 pthread_mutex_t *tf_mutex_ = tf_mt_sync->mutex_;
1818 pthread_mutex_lock(tf_mutex_);
1819 #endif
1820 if (tf_mt_sync->next_tf_row < mb_rows) {
1821 *current_mb_row = tf_mt_sync->next_tf_row;
1822 tf_mt_sync->next_tf_row++;
1823 do_next_row = 1;
1824 }
1825 #if CONFIG_MULTITHREAD
1826 pthread_mutex_unlock(tf_mutex_);
1827 #endif
1828 return do_next_row;
1829 }
1830
1831 // Hook function for each thread in temporal filter multi-threading.
tf_worker_hook(void * arg1,void * unused)1832 static int tf_worker_hook(void *arg1, void *unused) {
1833 (void)unused;
1834 EncWorkerData *thread_data = (EncWorkerData *)arg1;
1835 AV1_COMP *cpi = thread_data->cpi;
1836 ThreadData *td = thread_data->td;
1837 TemporalFilterCtx *tf_ctx = &cpi->tf_ctx;
1838 AV1TemporalFilterSync *tf_sync = &cpi->mt_info.tf_sync;
1839 const struct scale_factors *scale = &cpi->tf_ctx.sf;
1840 const int num_planes = av1_num_planes(&cpi->common);
1841 assert(num_planes >= 1 && num_planes <= MAX_MB_PLANE);
1842
1843 MACROBLOCKD *mbd = &td->mb.e_mbd;
1844 uint8_t *input_buffer[MAX_MB_PLANE];
1845 MB_MODE_INFO **input_mb_mode_info;
1846 tf_save_state(mbd, &input_mb_mode_info, input_buffer, num_planes);
1847 tf_setup_macroblockd(mbd, &td->tf_data, scale);
1848
1849 int current_mb_row = -1;
1850
1851 while (tf_get_next_job(tf_sync, ¤t_mb_row, tf_ctx->mb_rows))
1852 av1_tf_do_filtering_row(cpi, td, current_mb_row);
1853
1854 tf_restore_state(mbd, input_mb_mode_info, input_buffer, num_planes);
1855
1856 return 1;
1857 }
1858
1859 // Assigns temporal filter hook function and thread data to each worker.
prepare_tf_workers(AV1_COMP * cpi,AVxWorkerHook hook,int num_workers,int is_highbitdepth)1860 static void prepare_tf_workers(AV1_COMP *cpi, AVxWorkerHook hook,
1861 int num_workers, int is_highbitdepth) {
1862 MultiThreadInfo *mt_info = &cpi->mt_info;
1863 mt_info->tf_sync.next_tf_row = 0;
1864 for (int i = num_workers - 1; i >= 0; i--) {
1865 AVxWorker *worker = &mt_info->workers[i];
1866 EncWorkerData *thread_data = &mt_info->tile_thr_data[i];
1867
1868 worker->hook = hook;
1869 worker->data1 = thread_data;
1870 worker->data2 = NULL;
1871
1872 thread_data->thread_id = i;
1873 // Set the starting tile for each thread.
1874 thread_data->start = i;
1875
1876 thread_data->cpi = cpi;
1877 if (i == 0) {
1878 thread_data->td = &cpi->td;
1879 #if !CONFIG_FRAME_PARALLEL_ENCODE
1880 }
1881 #else
1882 } else {
1883 thread_data->td = thread_data->original_td;
1884 }
1885 #endif // CONFIG_FRAME_PARALLEL_ENCODE
1886
1887 // Before encoding a frame, copy the thread data from cpi.
1888 if (thread_data->td != &cpi->td) {
1889 thread_data->td->mb = cpi->td.mb;
1890 // OBMC buffers are used only to init MS params and remain unused when
1891 // called from tf, hence set the buffers to defaults.
1892 av1_init_obmc_buffer(&thread_data->td->mb.obmc_buffer);
1893 tf_alloc_and_reset_data(&thread_data->td->tf_data, cpi->tf_ctx.num_pels,
1894 is_highbitdepth);
1895 }
1896 }
1897 }
1898
1899 // Deallocate thread specific data for temporal filter.
tf_dealloc_thread_data(AV1_COMP * cpi,int num_workers,int is_highbitdepth)1900 static void tf_dealloc_thread_data(AV1_COMP *cpi, int num_workers,
1901 int is_highbitdepth) {
1902 MultiThreadInfo *mt_info = &cpi->mt_info;
1903 for (int i = num_workers - 1; i >= 0; i--) {
1904 EncWorkerData *thread_data = &mt_info->tile_thr_data[i];
1905 ThreadData *td = thread_data->td;
1906 if (td != &cpi->td) tf_dealloc_data(&td->tf_data, is_highbitdepth);
1907 }
1908 }
1909
1910 // Accumulate sse and sum after temporal filtering.
tf_accumulate_frame_diff(AV1_COMP * cpi,int num_workers)1911 static void tf_accumulate_frame_diff(AV1_COMP *cpi, int num_workers) {
1912 FRAME_DIFF *total_diff = &cpi->td.tf_data.diff;
1913 for (int i = num_workers - 1; i >= 0; i--) {
1914 AVxWorker *const worker = &cpi->mt_info.workers[i];
1915 EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
1916 ThreadData *td = thread_data->td;
1917 FRAME_DIFF *diff = &td->tf_data.diff;
1918 if (td != &cpi->td) {
1919 total_diff->sse += diff->sse;
1920 total_diff->sum += diff->sum;
1921 }
1922 }
1923 }
1924
1925 // Implements multi-threading for temporal filter.
av1_tf_do_filtering_mt(AV1_COMP * cpi)1926 void av1_tf_do_filtering_mt(AV1_COMP *cpi) {
1927 AV1_COMMON *cm = &cpi->common;
1928 MultiThreadInfo *mt_info = &cpi->mt_info;
1929 const int is_highbitdepth = cpi->tf_ctx.is_highbitdepth;
1930
1931 int num_workers =
1932 AOMMIN(mt_info->num_mod_workers[MOD_TF], mt_info->num_workers);
1933
1934 prepare_tf_workers(cpi, tf_worker_hook, num_workers, is_highbitdepth);
1935 launch_workers(mt_info, num_workers);
1936 sync_enc_workers(mt_info, cm, num_workers);
1937 tf_accumulate_frame_diff(cpi, num_workers);
1938 tf_dealloc_thread_data(cpi, num_workers, is_highbitdepth);
1939 }
1940
1941 // Checks if a job is available in the current direction. If a job is available,
1942 // frame_idx will be populated and returns 1, else returns 0.
get_next_gm_job(AV1_COMP * cpi,int * frame_idx,int cur_dir)1943 static AOM_INLINE int get_next_gm_job(AV1_COMP *cpi, int *frame_idx,
1944 int cur_dir) {
1945 GlobalMotionInfo *gm_info = &cpi->gm_info;
1946 JobInfo *job_info = &cpi->mt_info.gm_sync.job_info;
1947
1948 int total_refs = gm_info->num_ref_frames[cur_dir];
1949 int8_t cur_frame_to_process = job_info->next_frame_to_process[cur_dir];
1950
1951 if (cur_frame_to_process < total_refs && !job_info->early_exit[cur_dir]) {
1952 *frame_idx = gm_info->reference_frames[cur_dir][cur_frame_to_process].frame;
1953 job_info->next_frame_to_process[cur_dir] += 1;
1954 return 1;
1955 }
1956 return 0;
1957 }
1958
1959 // Switches the current direction and calls the function get_next_gm_job() if
1960 // the speed feature 'prune_ref_frame_for_gm_search' is not set.
switch_direction(AV1_COMP * cpi,int * frame_idx,int * cur_dir)1961 static AOM_INLINE void switch_direction(AV1_COMP *cpi, int *frame_idx,
1962 int *cur_dir) {
1963 if (cpi->sf.gm_sf.prune_ref_frame_for_gm_search) return;
1964 // Switch the direction and get next job
1965 *cur_dir = !(*cur_dir);
1966 get_next_gm_job(cpi, frame_idx, *(cur_dir));
1967 }
1968
1969 // Initializes inliers, num_inliers and segment_map.
init_gm_thread_data(const GlobalMotionInfo * gm_info,GlobalMotionThreadData * thread_data)1970 static AOM_INLINE void init_gm_thread_data(
1971 const GlobalMotionInfo *gm_info, GlobalMotionThreadData *thread_data) {
1972 for (int m = 0; m < RANSAC_NUM_MOTIONS; m++) {
1973 MotionModel motion_params = thread_data->params_by_motion[m];
1974 av1_zero(motion_params.params);
1975 motion_params.num_inliers = 0;
1976 }
1977
1978 av1_zero_array(thread_data->segment_map,
1979 gm_info->segment_map_w * gm_info->segment_map_h);
1980 }
1981
1982 // Hook function for each thread in global motion multi-threading.
gm_mt_worker_hook(void * arg1,void * unused)1983 static int gm_mt_worker_hook(void *arg1, void *unused) {
1984 (void)unused;
1985
1986 EncWorkerData *thread_data = (EncWorkerData *)arg1;
1987 AV1_COMP *cpi = thread_data->cpi;
1988 GlobalMotionInfo *gm_info = &cpi->gm_info;
1989 MultiThreadInfo *mt_info = &cpi->mt_info;
1990 JobInfo *job_info = &mt_info->gm_sync.job_info;
1991 int thread_id = thread_data->thread_id;
1992 GlobalMotionThreadData *gm_thread_data =
1993 &mt_info->gm_sync.thread_data[thread_id];
1994 int cur_dir = job_info->thread_id_to_dir[thread_id];
1995 #if CONFIG_MULTITHREAD
1996 pthread_mutex_t *gm_mt_mutex_ = mt_info->gm_sync.mutex_;
1997 #endif
1998
1999 while (1) {
2000 int ref_buf_idx = -1;
2001 int ref_frame_idx = -1;
2002
2003 #if CONFIG_MULTITHREAD
2004 pthread_mutex_lock(gm_mt_mutex_);
2005 #endif
2006
2007 // Populates ref_buf_idx(the reference frame type) for which global motion
2008 // estimation will be done.
2009 if (!get_next_gm_job(cpi, &ref_buf_idx, cur_dir)) {
2010 // No jobs are available for the current direction. Switch
2011 // to other direction and get the next job, if available.
2012 switch_direction(cpi, &ref_buf_idx, &cur_dir);
2013 }
2014
2015 // 'ref_frame_idx' holds the index of the current reference frame type in
2016 // gm_info->reference_frames. job_info->next_frame_to_process will be
2017 // incremented in get_next_gm_job() and hence subtracting by 1.
2018 ref_frame_idx = job_info->next_frame_to_process[cur_dir] - 1;
2019
2020 #if CONFIG_MULTITHREAD
2021 pthread_mutex_unlock(gm_mt_mutex_);
2022 #endif
2023
2024 if (ref_buf_idx == -1) break;
2025
2026 init_gm_thread_data(gm_info, gm_thread_data);
2027
2028 // Compute global motion for the given ref_buf_idx.
2029 av1_compute_gm_for_valid_ref_frames(
2030 cpi, gm_info->ref_buf, ref_buf_idx, gm_info->num_src_corners,
2031 gm_info->src_corners, gm_info->src_buffer,
2032 gm_thread_data->params_by_motion, gm_thread_data->segment_map,
2033 gm_info->segment_map_w, gm_info->segment_map_h);
2034
2035 #if CONFIG_MULTITHREAD
2036 pthread_mutex_lock(gm_mt_mutex_);
2037 #endif
2038 assert(ref_frame_idx != -1);
2039 // If global motion w.r.t. current ref frame is
2040 // INVALID/TRANSLATION/IDENTITY, skip the evaluation of global motion w.r.t
2041 // the remaining ref frames in that direction. The below exit is disabled
2042 // when ref frame distance w.r.t. current frame is zero. E.g.:
2043 // source_alt_ref_frame w.r.t. ARF frames.
2044 if (cpi->sf.gm_sf.prune_ref_frame_for_gm_search &&
2045 gm_info->reference_frames[cur_dir][ref_frame_idx].distance != 0 &&
2046 cpi->common.global_motion[ref_buf_idx].wmtype != ROTZOOM)
2047 job_info->early_exit[cur_dir] = 1;
2048
2049 #if CONFIG_MULTITHREAD
2050 pthread_mutex_unlock(gm_mt_mutex_);
2051 #endif
2052 }
2053 return 1;
2054 }
2055
2056 // Assigns global motion hook function and thread data to each worker.
prepare_gm_workers(AV1_COMP * cpi,AVxWorkerHook hook,int num_workers)2057 static AOM_INLINE void prepare_gm_workers(AV1_COMP *cpi, AVxWorkerHook hook,
2058 int num_workers) {
2059 MultiThreadInfo *mt_info = &cpi->mt_info;
2060 for (int i = num_workers - 1; i >= 0; i--) {
2061 AVxWorker *worker = &mt_info->workers[i];
2062 EncWorkerData *thread_data = &mt_info->tile_thr_data[i];
2063
2064 worker->hook = hook;
2065 worker->data1 = thread_data;
2066 worker->data2 = NULL;
2067
2068 thread_data->thread_id = i;
2069 // Set the starting tile for each thread.
2070 thread_data->start = i;
2071
2072 thread_data->cpi = cpi;
2073 if (i == 0) {
2074 thread_data->td = &cpi->td;
2075 #if !CONFIG_FRAME_PARALLEL_ENCODE
2076 }
2077 #else
2078 } else {
2079 thread_data->td = thread_data->original_td;
2080 }
2081 #endif // CONFIG_FRAME_PARALLEL_ENCODE
2082 }
2083 }
2084
2085 // Assigns available threads to past/future direction.
assign_thread_to_dir(int8_t * thread_id_to_dir,int num_workers)2086 static AOM_INLINE void assign_thread_to_dir(int8_t *thread_id_to_dir,
2087 int num_workers) {
2088 int8_t frame_dir_idx = 0;
2089
2090 for (int i = 0; i < num_workers; i++) {
2091 thread_id_to_dir[i] = frame_dir_idx++;
2092 if (frame_dir_idx == MAX_DIRECTIONS) frame_dir_idx = 0;
2093 }
2094 }
2095
2096 // Computes number of workers for global motion multi-threading.
compute_gm_workers(const AV1_COMP * cpi)2097 static AOM_INLINE int compute_gm_workers(const AV1_COMP *cpi) {
2098 int total_refs =
2099 cpi->gm_info.num_ref_frames[0] + cpi->gm_info.num_ref_frames[1];
2100 int num_gm_workers = cpi->sf.gm_sf.prune_ref_frame_for_gm_search
2101 ? AOMMIN(MAX_DIRECTIONS, total_refs)
2102 : total_refs;
2103 num_gm_workers = AOMMIN(num_gm_workers, cpi->mt_info.num_workers);
2104 return (num_gm_workers);
2105 }
2106
2107 // Frees the memory allocated for each worker in global motion multi-threading.
av1_gm_dealloc(AV1GlobalMotionSync * gm_sync_data)2108 void av1_gm_dealloc(AV1GlobalMotionSync *gm_sync_data) {
2109 if (gm_sync_data->thread_data != NULL) {
2110 for (int j = 0; j < gm_sync_data->allocated_workers; j++) {
2111 GlobalMotionThreadData *thread_data = &gm_sync_data->thread_data[j];
2112 aom_free(thread_data->segment_map);
2113
2114 for (int m = 0; m < RANSAC_NUM_MOTIONS; m++)
2115 aom_free(thread_data->params_by_motion[m].inliers);
2116 }
2117 aom_free(gm_sync_data->thread_data);
2118 }
2119 }
2120
2121 // Allocates memory for inliers and segment_map for each worker in global motion
2122 // multi-threading.
gm_alloc(AV1_COMP * cpi,int num_workers)2123 static AOM_INLINE void gm_alloc(AV1_COMP *cpi, int num_workers) {
2124 AV1_COMMON *cm = &cpi->common;
2125 AV1GlobalMotionSync *gm_sync = &cpi->mt_info.gm_sync;
2126 GlobalMotionInfo *gm_info = &cpi->gm_info;
2127
2128 gm_sync->allocated_workers = num_workers;
2129 gm_sync->allocated_width = cpi->source->y_width;
2130 gm_sync->allocated_height = cpi->source->y_height;
2131
2132 CHECK_MEM_ERROR(cm, gm_sync->thread_data,
2133 aom_malloc(sizeof(*gm_sync->thread_data) * num_workers));
2134
2135 for (int i = 0; i < num_workers; i++) {
2136 GlobalMotionThreadData *thread_data = &gm_sync->thread_data[i];
2137 CHECK_MEM_ERROR(
2138 cm, thread_data->segment_map,
2139 aom_malloc(sizeof(*thread_data->segment_map) * gm_info->segment_map_w *
2140 gm_info->segment_map_h));
2141
2142 for (int m = 0; m < RANSAC_NUM_MOTIONS; m++) {
2143 CHECK_MEM_ERROR(
2144 cm, thread_data->params_by_motion[m].inliers,
2145 aom_malloc(sizeof(*thread_data->params_by_motion[m].inliers) * 2 *
2146 MAX_CORNERS));
2147 }
2148 }
2149 }
2150
2151 // Implements multi-threading for global motion.
av1_global_motion_estimation_mt(AV1_COMP * cpi)2152 void av1_global_motion_estimation_mt(AV1_COMP *cpi) {
2153 AV1GlobalMotionSync *gm_sync = &cpi->mt_info.gm_sync;
2154 JobInfo *job_info = &gm_sync->job_info;
2155
2156 av1_zero(*job_info);
2157
2158 int num_workers = compute_gm_workers(cpi);
2159
2160 if (num_workers > gm_sync->allocated_workers ||
2161 cpi->source->y_width != gm_sync->allocated_width ||
2162 cpi->source->y_height != gm_sync->allocated_height) {
2163 av1_gm_dealloc(gm_sync);
2164 gm_alloc(cpi, num_workers);
2165 }
2166
2167 assign_thread_to_dir(job_info->thread_id_to_dir, num_workers);
2168 prepare_gm_workers(cpi, gm_mt_worker_hook, num_workers);
2169 launch_workers(&cpi->mt_info, num_workers);
2170 sync_enc_workers(&cpi->mt_info, &cpi->common, num_workers);
2171 }
2172 #endif // !CONFIG_REALTIME_ONLY
2173
2174 // Compare and order tiles based on absolute sum of tx coeffs.
compare_tile_order(const void * a,const void * b)2175 static int compare_tile_order(const void *a, const void *b) {
2176 const PackBSTileOrder *const tile_a = (const PackBSTileOrder *)a;
2177 const PackBSTileOrder *const tile_b = (const PackBSTileOrder *)b;
2178
2179 if (tile_a->abs_sum_level > tile_b->abs_sum_level)
2180 return -1;
2181 else if (tile_a->abs_sum_level == tile_b->abs_sum_level)
2182 return (tile_a->tile_idx > tile_b->tile_idx ? 1 : -1);
2183 else
2184 return 1;
2185 }
2186
2187 // Get next tile index to be processed for pack bitstream
get_next_pack_bs_tile_idx(AV1EncPackBSSync * const pack_bs_sync,const int num_tiles)2188 static AOM_INLINE int get_next_pack_bs_tile_idx(
2189 AV1EncPackBSSync *const pack_bs_sync, const int num_tiles) {
2190 assert(pack_bs_sync->next_job_idx <= num_tiles);
2191 if (pack_bs_sync->next_job_idx == num_tiles) return -1;
2192
2193 return pack_bs_sync->pack_bs_tile_order[pack_bs_sync->next_job_idx++]
2194 .tile_idx;
2195 }
2196
2197 // Calculates bitstream chunk size based on total buffer size and tile or tile
2198 // group size.
get_bs_chunk_size(int tg_or_tile_size,const int frame_or_tg_size,size_t * remain_buf_size,size_t max_buf_size,int is_last_chunk)2199 static AOM_INLINE size_t get_bs_chunk_size(int tg_or_tile_size,
2200 const int frame_or_tg_size,
2201 size_t *remain_buf_size,
2202 size_t max_buf_size,
2203 int is_last_chunk) {
2204 size_t this_chunk_size;
2205 assert(*remain_buf_size > 0);
2206 if (is_last_chunk) {
2207 this_chunk_size = *remain_buf_size;
2208 *remain_buf_size = 0;
2209 } else {
2210 const uint64_t size_scale = (uint64_t)max_buf_size * tg_or_tile_size;
2211 this_chunk_size = (size_t)(size_scale / frame_or_tg_size);
2212 *remain_buf_size -= this_chunk_size;
2213 assert(*remain_buf_size > 0);
2214 }
2215 assert(this_chunk_size > 0);
2216 return this_chunk_size;
2217 }
2218
2219 // Initializes params required for pack bitstream tile.
init_tile_pack_bs_params(AV1_COMP * const cpi,uint8_t * const dst,struct aom_write_bit_buffer * saved_wb,PackBSParams * const pack_bs_params_arr,uint8_t obu_extn_header)2220 static void init_tile_pack_bs_params(AV1_COMP *const cpi, uint8_t *const dst,
2221 struct aom_write_bit_buffer *saved_wb,
2222 PackBSParams *const pack_bs_params_arr,
2223 uint8_t obu_extn_header) {
2224 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
2225 AV1_COMMON *const cm = &cpi->common;
2226 const CommonTileParams *const tiles = &cm->tiles;
2227 const int num_tiles = tiles->cols * tiles->rows;
2228 // Fixed size tile groups for the moment
2229 const int num_tg_hdrs = cpi->num_tg;
2230 // Tile group size in terms of number of tiles.
2231 const int tg_size_in_tiles = (num_tiles + num_tg_hdrs - 1) / num_tg_hdrs;
2232 uint8_t *tile_dst = dst;
2233 uint8_t *tile_data_curr = dst;
2234 // Max tile group count can not be more than MAX_TILES.
2235 int tg_size_mi[MAX_TILES] = { 0 }; // Size of tile group in mi units
2236 int tile_idx;
2237 int tg_idx = 0;
2238 int tile_count_in_tg = 0;
2239 int new_tg = 1;
2240
2241 // Populate pack bitstream params of all tiles.
2242 for (tile_idx = 0; tile_idx < num_tiles; tile_idx++) {
2243 const TileInfo *const tile_info = &cpi->tile_data[tile_idx].tile_info;
2244 PackBSParams *const pack_bs_params = &pack_bs_params_arr[tile_idx];
2245 // Calculate tile size in mi units.
2246 const int tile_size_mi = (tile_info->mi_col_end - tile_info->mi_col_start) *
2247 (tile_info->mi_row_end - tile_info->mi_row_start);
2248 int is_last_tile_in_tg = 0;
2249 tile_count_in_tg++;
2250 if (tile_count_in_tg == tg_size_in_tiles || tile_idx == (num_tiles - 1))
2251 is_last_tile_in_tg = 1;
2252
2253 // Populate pack bitstream params of this tile.
2254 pack_bs_params->curr_tg_hdr_size = 0;
2255 pack_bs_params->obu_extn_header = obu_extn_header;
2256 pack_bs_params->saved_wb = saved_wb;
2257 pack_bs_params->obu_header_size = 0;
2258 pack_bs_params->is_last_tile_in_tg = is_last_tile_in_tg;
2259 pack_bs_params->new_tg = new_tg;
2260 pack_bs_params->tile_col = tile_info->tile_col;
2261 pack_bs_params->tile_row = tile_info->tile_row;
2262 pack_bs_params->tile_size_mi = tile_size_mi;
2263 tg_size_mi[tg_idx] += tile_size_mi;
2264
2265 if (new_tg) new_tg = 0;
2266 if (is_last_tile_in_tg) {
2267 tile_count_in_tg = 0;
2268 new_tg = 1;
2269 tg_idx++;
2270 }
2271 }
2272
2273 assert(cpi->available_bs_size > 0);
2274 size_t tg_buf_size[MAX_TILES] = { 0 };
2275 size_t max_buf_size = cpi->available_bs_size;
2276 size_t remain_buf_size = max_buf_size;
2277 const int frame_size_mi = cm->mi_params.mi_rows * cm->mi_params.mi_cols;
2278
2279 tile_idx = 0;
2280 // Prepare obu, tile group and frame header of each tile group.
2281 for (tg_idx = 0; tg_idx < cpi->num_tg; tg_idx++) {
2282 PackBSParams *const pack_bs_params = &pack_bs_params_arr[tile_idx];
2283 int is_last_tg = tg_idx == cpi->num_tg - 1;
2284 // Prorate bitstream buffer size based on tile group size and available
2285 // buffer size. This buffer will be used to store headers and tile data.
2286 tg_buf_size[tg_idx] =
2287 get_bs_chunk_size(tg_size_mi[tg_idx], frame_size_mi, &remain_buf_size,
2288 max_buf_size, is_last_tg);
2289
2290 pack_bs_params->dst = tile_dst;
2291 pack_bs_params->tile_data_curr = tile_dst;
2292
2293 // Write obu, tile group and frame header at first tile in the tile
2294 // group.
2295 av1_write_obu_tg_tile_headers(cpi, xd, pack_bs_params, tile_idx);
2296 tile_dst += tg_buf_size[tg_idx];
2297
2298 // Exclude headers from tile group buffer size.
2299 tg_buf_size[tg_idx] -= pack_bs_params->curr_tg_hdr_size;
2300 tile_idx += tg_size_in_tiles;
2301 }
2302
2303 tg_idx = 0;
2304 // Calculate bitstream buffer size of each tile in the tile group.
2305 for (tile_idx = 0; tile_idx < num_tiles; tile_idx++) {
2306 PackBSParams *const pack_bs_params = &pack_bs_params_arr[tile_idx];
2307
2308 if (pack_bs_params->new_tg) {
2309 max_buf_size = tg_buf_size[tg_idx];
2310 remain_buf_size = max_buf_size;
2311 }
2312
2313 // Prorate bitstream buffer size of this tile based on tile size and
2314 // available buffer size. For this proration, header size is not accounted.
2315 const size_t tile_buf_size = get_bs_chunk_size(
2316 pack_bs_params->tile_size_mi, tg_size_mi[tg_idx], &remain_buf_size,
2317 max_buf_size, pack_bs_params->is_last_tile_in_tg);
2318 pack_bs_params->tile_buf_size = tile_buf_size;
2319
2320 // Update base address of bitstream buffer for tile and tile group.
2321 if (pack_bs_params->new_tg) {
2322 tile_dst = pack_bs_params->dst;
2323 tile_data_curr = pack_bs_params->tile_data_curr;
2324 // Account header size in first tile of a tile group.
2325 pack_bs_params->tile_buf_size += pack_bs_params->curr_tg_hdr_size;
2326 } else {
2327 pack_bs_params->dst = tile_dst;
2328 pack_bs_params->tile_data_curr = tile_data_curr;
2329 }
2330
2331 if (pack_bs_params->is_last_tile_in_tg) tg_idx++;
2332 tile_dst += pack_bs_params->tile_buf_size;
2333 }
2334 }
2335
2336 // Worker hook function of pack bitsteam multithreading.
pack_bs_worker_hook(void * arg1,void * arg2)2337 static int pack_bs_worker_hook(void *arg1, void *arg2) {
2338 EncWorkerData *const thread_data = (EncWorkerData *)arg1;
2339 PackBSParams *const pack_bs_params = (PackBSParams *)arg2;
2340 AV1_COMP *const cpi = thread_data->cpi;
2341 AV1_COMMON *const cm = &cpi->common;
2342 AV1EncPackBSSync *const pack_bs_sync = &cpi->mt_info.pack_bs_sync;
2343 const CommonTileParams *const tiles = &cm->tiles;
2344 const int num_tiles = tiles->cols * tiles->rows;
2345
2346 while (1) {
2347 #if CONFIG_MULTITHREAD
2348 pthread_mutex_lock(pack_bs_sync->mutex_);
2349 #endif
2350 const int tile_idx = get_next_pack_bs_tile_idx(pack_bs_sync, num_tiles);
2351 #if CONFIG_MULTITHREAD
2352 pthread_mutex_unlock(pack_bs_sync->mutex_);
2353 #endif
2354 if (tile_idx == -1) break;
2355 TileDataEnc *this_tile = &cpi->tile_data[tile_idx];
2356 thread_data->td->mb.e_mbd.tile_ctx = &this_tile->tctx;
2357
2358 av1_pack_tile_info(cpi, thread_data->td, &pack_bs_params[tile_idx]);
2359 }
2360
2361 return 1;
2362 }
2363
2364 // Prepares thread data and workers of pack bitsteam multithreading.
prepare_pack_bs_workers(AV1_COMP * const cpi,PackBSParams * const pack_bs_params,AVxWorkerHook hook,const int num_workers)2365 static void prepare_pack_bs_workers(AV1_COMP *const cpi,
2366 PackBSParams *const pack_bs_params,
2367 AVxWorkerHook hook, const int num_workers) {
2368 MultiThreadInfo *const mt_info = &cpi->mt_info;
2369 for (int i = num_workers - 1; i >= 0; i--) {
2370 AVxWorker *worker = &mt_info->workers[i];
2371 EncWorkerData *const thread_data = &mt_info->tile_thr_data[i];
2372 if (i == 0) {
2373 thread_data->td = &cpi->td;
2374 #if !CONFIG_FRAME_PARALLEL_ENCODE
2375 }
2376 #else
2377 } else {
2378 thread_data->td = thread_data->original_td;
2379 }
2380 #endif // CONFIG_FRAME_PARALLEL_ENCODE
2381
2382 if (thread_data->td != &cpi->td) thread_data->td->mb = cpi->td.mb;
2383
2384 thread_data->cpi = cpi;
2385 thread_data->start = i;
2386 thread_data->thread_id = i;
2387 av1_reset_pack_bs_thread_data(thread_data->td);
2388
2389 worker->hook = hook;
2390 worker->data1 = thread_data;
2391 worker->data2 = pack_bs_params;
2392 }
2393
2394 AV1_COMMON *const cm = &cpi->common;
2395 AV1EncPackBSSync *const pack_bs_sync = &mt_info->pack_bs_sync;
2396 const uint16_t num_tiles = cm->tiles.rows * cm->tiles.cols;
2397 pack_bs_sync->next_job_idx = 0;
2398
2399 PackBSTileOrder *const pack_bs_tile_order = pack_bs_sync->pack_bs_tile_order;
2400 // Reset tile order data of pack bitstream
2401 av1_zero_array(pack_bs_tile_order, num_tiles);
2402
2403 // Populate pack bitstream tile order structure
2404 for (uint16_t tile_idx = 0; tile_idx < num_tiles; tile_idx++) {
2405 pack_bs_tile_order[tile_idx].abs_sum_level =
2406 cpi->tile_data[tile_idx].abs_sum_level;
2407 pack_bs_tile_order[tile_idx].tile_idx = tile_idx;
2408 }
2409
2410 // Sort tiles in descending order based on tile area.
2411 qsort(pack_bs_tile_order, num_tiles, sizeof(*pack_bs_tile_order),
2412 compare_tile_order);
2413 }
2414
2415 // Accumulates data after pack bitsteam processing.
accumulate_pack_bs_data(AV1_COMP * const cpi,const PackBSParams * const pack_bs_params_arr,uint8_t * const dst,uint32_t * total_size,const FrameHeaderInfo * fh_info,int * const largest_tile_id,unsigned int * max_tile_size,uint32_t * const obu_header_size,uint8_t ** tile_data_start,const int num_workers)2416 static void accumulate_pack_bs_data(
2417 AV1_COMP *const cpi, const PackBSParams *const pack_bs_params_arr,
2418 uint8_t *const dst, uint32_t *total_size, const FrameHeaderInfo *fh_info,
2419 int *const largest_tile_id, unsigned int *max_tile_size,
2420 uint32_t *const obu_header_size, uint8_t **tile_data_start,
2421 const int num_workers) {
2422 const AV1_COMMON *const cm = &cpi->common;
2423 const CommonTileParams *const tiles = &cm->tiles;
2424 const int tile_count = tiles->cols * tiles->rows;
2425 // Fixed size tile groups for the moment
2426 size_t curr_tg_data_size = 0;
2427 int is_first_tg = 1;
2428 uint8_t *curr_tg_start = dst;
2429 size_t src_offset = 0;
2430 size_t dst_offset = 0;
2431
2432 for (int tile_idx = 0; tile_idx < tile_count; tile_idx++) {
2433 // PackBSParams stores all parameters required to pack tile and header
2434 // info.
2435 const PackBSParams *const pack_bs_params = &pack_bs_params_arr[tile_idx];
2436 uint32_t tile_size = 0;
2437
2438 if (pack_bs_params->new_tg) {
2439 curr_tg_start = dst + *total_size;
2440 curr_tg_data_size = pack_bs_params->curr_tg_hdr_size;
2441 *tile_data_start += pack_bs_params->curr_tg_hdr_size;
2442 *obu_header_size = pack_bs_params->obu_header_size;
2443 }
2444 curr_tg_data_size +=
2445 pack_bs_params->buf.size + (pack_bs_params->is_last_tile_in_tg ? 0 : 4);
2446
2447 if (pack_bs_params->buf.size > *max_tile_size) {
2448 *largest_tile_id = tile_idx;
2449 *max_tile_size = (unsigned int)pack_bs_params->buf.size;
2450 }
2451 tile_size +=
2452 (uint32_t)pack_bs_params->buf.size + *pack_bs_params->total_size;
2453
2454 // Pack all the chunks of tile bitstreams together
2455 if (tile_idx != 0) memmove(dst + dst_offset, dst + src_offset, tile_size);
2456
2457 if (pack_bs_params->is_last_tile_in_tg)
2458 av1_write_last_tile_info(
2459 cpi, fh_info, pack_bs_params->saved_wb, &curr_tg_data_size,
2460 curr_tg_start, &tile_size, tile_data_start, largest_tile_id,
2461 &is_first_tg, *obu_header_size, pack_bs_params->obu_extn_header);
2462 src_offset += pack_bs_params->tile_buf_size;
2463 dst_offset += tile_size;
2464 *total_size += tile_size;
2465 }
2466
2467 // Accumulate thread data
2468 MultiThreadInfo *const mt_info = &cpi->mt_info;
2469 for (int idx = num_workers - 1; idx >= 0; idx--) {
2470 ThreadData const *td = mt_info->tile_thr_data[idx].td;
2471 av1_accumulate_pack_bs_thread_data(cpi, td);
2472 }
2473 }
2474
av1_write_tile_obu_mt(AV1_COMP * const cpi,uint8_t * const dst,uint32_t * total_size,struct aom_write_bit_buffer * saved_wb,uint8_t obu_extn_header,const FrameHeaderInfo * fh_info,int * const largest_tile_id,unsigned int * max_tile_size,uint32_t * const obu_header_size,uint8_t ** tile_data_start,const int num_workers)2475 void av1_write_tile_obu_mt(
2476 AV1_COMP *const cpi, uint8_t *const dst, uint32_t *total_size,
2477 struct aom_write_bit_buffer *saved_wb, uint8_t obu_extn_header,
2478 const FrameHeaderInfo *fh_info, int *const largest_tile_id,
2479 unsigned int *max_tile_size, uint32_t *const obu_header_size,
2480 uint8_t **tile_data_start, const int num_workers) {
2481 MultiThreadInfo *const mt_info = &cpi->mt_info;
2482
2483 PackBSParams pack_bs_params[MAX_TILES];
2484 uint32_t tile_size[MAX_TILES] = { 0 };
2485
2486 for (int tile_idx = 0; tile_idx < MAX_TILES; tile_idx++)
2487 pack_bs_params[tile_idx].total_size = &tile_size[tile_idx];
2488
2489 init_tile_pack_bs_params(cpi, dst, saved_wb, pack_bs_params, obu_extn_header);
2490 prepare_pack_bs_workers(cpi, pack_bs_params, pack_bs_worker_hook,
2491 num_workers);
2492 launch_workers(mt_info, num_workers);
2493 sync_enc_workers(mt_info, &cpi->common, num_workers);
2494 accumulate_pack_bs_data(cpi, pack_bs_params, dst, total_size, fh_info,
2495 largest_tile_id, max_tile_size, obu_header_size,
2496 tile_data_start, num_workers);
2497 }
2498
2499 // Deallocate memory for CDEF search multi-thread synchronization.
av1_cdef_mt_dealloc(AV1CdefSync * cdef_sync)2500 void av1_cdef_mt_dealloc(AV1CdefSync *cdef_sync) {
2501 (void)cdef_sync;
2502 assert(cdef_sync != NULL);
2503 #if CONFIG_MULTITHREAD
2504 if (cdef_sync->mutex_ != NULL) {
2505 pthread_mutex_destroy(cdef_sync->mutex_);
2506 aom_free(cdef_sync->mutex_);
2507 }
2508 #endif // CONFIG_MULTITHREAD
2509 }
2510
2511 // Updates the row and column indices of the next job to be processed.
2512 // Also updates end_of_frame flag when the processing of all blocks is complete.
update_next_job_info(AV1CdefSync * cdef_sync,int nvfb,int nhfb)2513 static void update_next_job_info(AV1CdefSync *cdef_sync, int nvfb, int nhfb) {
2514 cdef_sync->fbc++;
2515 if (cdef_sync->fbc == nhfb) {
2516 cdef_sync->fbr++;
2517 if (cdef_sync->fbr == nvfb) {
2518 cdef_sync->end_of_frame = 1;
2519 } else {
2520 cdef_sync->fbc = 0;
2521 }
2522 }
2523 }
2524
2525 // Initializes cdef_sync parameters.
cdef_reset_job_info(AV1CdefSync * cdef_sync)2526 static AOM_INLINE void cdef_reset_job_info(AV1CdefSync *cdef_sync) {
2527 #if CONFIG_MULTITHREAD
2528 if (cdef_sync->mutex_) pthread_mutex_init(cdef_sync->mutex_, NULL);
2529 #endif // CONFIG_MULTITHREAD
2530 cdef_sync->end_of_frame = 0;
2531 cdef_sync->fbr = 0;
2532 cdef_sync->fbc = 0;
2533 }
2534
2535 // Checks if a job is available. If job is available,
2536 // populates next job information and returns 1, else returns 0.
cdef_get_next_job(AV1CdefSync * cdef_sync,CdefSearchCtx * cdef_search_ctx,int * cur_fbr,int * cur_fbc,int * sb_count)2537 static AOM_INLINE int cdef_get_next_job(AV1CdefSync *cdef_sync,
2538 CdefSearchCtx *cdef_search_ctx,
2539 int *cur_fbr, int *cur_fbc,
2540 int *sb_count) {
2541 #if CONFIG_MULTITHREAD
2542 pthread_mutex_lock(cdef_sync->mutex_);
2543 #endif // CONFIG_MULTITHREAD
2544 int do_next_block = 0;
2545 const int nvfb = cdef_search_ctx->nvfb;
2546 const int nhfb = cdef_search_ctx->nhfb;
2547
2548 // If a block is skip, do not process the block and
2549 // check the skip condition for the next block.
2550 while ((!cdef_sync->end_of_frame) &&
2551 (cdef_sb_skip(cdef_search_ctx->mi_params, cdef_sync->fbr,
2552 cdef_sync->fbc))) {
2553 update_next_job_info(cdef_sync, nvfb, nhfb);
2554 }
2555
2556 // Populates information needed for current job and update the row,
2557 // column indices of the next block to be processed.
2558 if (cdef_sync->end_of_frame == 0) {
2559 do_next_block = 1;
2560 *cur_fbr = cdef_sync->fbr;
2561 *cur_fbc = cdef_sync->fbc;
2562 *sb_count = cdef_search_ctx->sb_count;
2563 cdef_search_ctx->sb_count++;
2564 update_next_job_info(cdef_sync, nvfb, nhfb);
2565 }
2566 #if CONFIG_MULTITHREAD
2567 pthread_mutex_unlock(cdef_sync->mutex_);
2568 #endif // CONFIG_MULTITHREAD
2569 return do_next_block;
2570 }
2571
2572 // Hook function for each thread in CDEF search multi-threading.
cdef_filter_block_worker_hook(void * arg1,void * arg2)2573 static int cdef_filter_block_worker_hook(void *arg1, void *arg2) {
2574 AV1CdefSync *const cdef_sync = (AV1CdefSync *)arg1;
2575 CdefSearchCtx *cdef_search_ctx = (CdefSearchCtx *)arg2;
2576 int cur_fbr, cur_fbc, sb_count;
2577 while (cdef_get_next_job(cdef_sync, cdef_search_ctx, &cur_fbr, &cur_fbc,
2578 &sb_count)) {
2579 av1_cdef_mse_calc_block(cdef_search_ctx, cur_fbr, cur_fbc, sb_count);
2580 }
2581 return 1;
2582 }
2583
2584 // Assigns CDEF search hook function and thread data to each worker.
prepare_cdef_workers(MultiThreadInfo * mt_info,CdefSearchCtx * cdef_search_ctx,AVxWorkerHook hook,int num_workers)2585 static void prepare_cdef_workers(MultiThreadInfo *mt_info,
2586 CdefSearchCtx *cdef_search_ctx,
2587 AVxWorkerHook hook, int num_workers) {
2588 for (int i = num_workers - 1; i >= 0; i--) {
2589 AVxWorker *worker = &mt_info->workers[i];
2590 worker->hook = hook;
2591 worker->data1 = &mt_info->cdef_sync;
2592 worker->data2 = cdef_search_ctx;
2593 }
2594 }
2595
2596 // Implements multi-threading for CDEF search.
av1_cdef_mse_calc_frame_mt(AV1_COMMON * cm,MultiThreadInfo * mt_info,CdefSearchCtx * cdef_search_ctx)2597 void av1_cdef_mse_calc_frame_mt(AV1_COMMON *cm, MultiThreadInfo *mt_info,
2598 CdefSearchCtx *cdef_search_ctx) {
2599 AV1CdefSync *cdef_sync = &mt_info->cdef_sync;
2600 const int num_workers = mt_info->num_mod_workers[MOD_CDEF_SEARCH];
2601
2602 cdef_reset_job_info(cdef_sync);
2603 prepare_cdef_workers(mt_info, cdef_search_ctx, cdef_filter_block_worker_hook,
2604 num_workers);
2605 launch_workers(mt_info, num_workers);
2606 sync_enc_workers(mt_info, cm, num_workers);
2607 }
2608
2609 // Computes num_workers for temporal filter multi-threading.
compute_num_tf_workers(AV1_COMP * cpi)2610 static AOM_INLINE int compute_num_tf_workers(AV1_COMP *cpi) {
2611 // For single-pass encode, using no. of workers as per tf block size was not
2612 // found to improve speed. Hence the thread assignment for single-pass encode
2613 // is kept based on compute_num_enc_workers().
2614 if (cpi->oxcf.pass < AOM_RC_SECOND_PASS)
2615 return (av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads));
2616
2617 if (cpi->oxcf.max_threads <= 1) return 1;
2618
2619 const int frame_height = cpi->common.height;
2620 const BLOCK_SIZE block_size = TF_BLOCK_SIZE;
2621 const int mb_height = block_size_high[block_size];
2622 const int mb_rows = get_num_blocks(frame_height, mb_height);
2623 return AOMMIN(cpi->oxcf.max_threads, mb_rows);
2624 }
2625
2626 // Computes num_workers for tpl multi-threading.
compute_num_tpl_workers(AV1_COMP * cpi)2627 static AOM_INLINE int compute_num_tpl_workers(AV1_COMP *cpi) {
2628 return av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads);
2629 }
2630
2631 // Computes num_workers for loop filter multi-threading.
compute_num_lf_workers(AV1_COMP * cpi)2632 static AOM_INLINE int compute_num_lf_workers(AV1_COMP *cpi) {
2633 return av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads);
2634 }
2635
2636 // Computes num_workers for cdef multi-threading.
compute_num_cdef_workers(AV1_COMP * cpi)2637 static AOM_INLINE int compute_num_cdef_workers(AV1_COMP *cpi) {
2638 return av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads);
2639 }
2640
2641 // Computes num_workers for loop-restoration multi-threading.
compute_num_lr_workers(AV1_COMP * cpi)2642 static AOM_INLINE int compute_num_lr_workers(AV1_COMP *cpi) {
2643 return av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads);
2644 }
2645
2646 // Computes num_workers for pack bitstream multi-threading.
compute_num_pack_bs_workers(AV1_COMP * cpi)2647 static AOM_INLINE int compute_num_pack_bs_workers(AV1_COMP *cpi) {
2648 if (cpi->oxcf.max_threads <= 1) return 1;
2649 return compute_num_enc_tile_mt_workers(&cpi->common, cpi->oxcf.max_threads);
2650 }
2651
compute_num_mod_workers(AV1_COMP * cpi,MULTI_THREADED_MODULES mod_name)2652 int compute_num_mod_workers(AV1_COMP *cpi, MULTI_THREADED_MODULES mod_name) {
2653 int num_mod_workers = 0;
2654 switch (mod_name) {
2655 case MOD_FP:
2656 if (cpi->oxcf.pass >= AOM_RC_SECOND_PASS)
2657 num_mod_workers = 0;
2658 else
2659 num_mod_workers =
2660 av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads);
2661 break;
2662 case MOD_TF: num_mod_workers = compute_num_tf_workers(cpi); break;
2663 case MOD_TPL: num_mod_workers = compute_num_tpl_workers(cpi); break;
2664 case MOD_GME: num_mod_workers = 1; break;
2665 case MOD_ENC:
2666 num_mod_workers = av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads);
2667 break;
2668 case MOD_LPF: num_mod_workers = compute_num_lf_workers(cpi); break;
2669 case MOD_CDEF_SEARCH:
2670 num_mod_workers = compute_num_cdef_workers(cpi);
2671 break;
2672 case MOD_CDEF: num_mod_workers = compute_num_cdef_workers(cpi); break;
2673 case MOD_LR: num_mod_workers = compute_num_lr_workers(cpi); break;
2674 case MOD_PACK_BS: num_mod_workers = compute_num_pack_bs_workers(cpi); break;
2675 case MOD_FRAME_ENC:
2676 num_mod_workers = cpi->ppi->p_mt_info.num_mod_workers[MOD_FRAME_ENC];
2677 break;
2678 default: assert(0); break;
2679 }
2680 return (num_mod_workers);
2681 }
2682 // Computes the number of workers for each MT modules in the encoder
av1_compute_num_workers_for_mt(AV1_COMP * cpi)2683 void av1_compute_num_workers_for_mt(AV1_COMP *cpi) {
2684 for (int i = MOD_FP; i < NUM_MT_MODULES; i++)
2685 cpi->ppi->p_mt_info.num_mod_workers[i] =
2686 compute_num_mod_workers(cpi, (MULTI_THREADED_MODULES)i);
2687 }
2688