/src/aom/av1/encoder/ethread.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * Copyright (c) 2016, Alliance for Open Media. All rights reserved |
3 | | * |
4 | | * This source code is subject to the terms of the BSD 2 Clause License and |
5 | | * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License |
6 | | * was not distributed with this source code in the LICENSE file, you can |
7 | | * obtain it at www.aomedia.org/license/software. If the Alliance for Open |
8 | | * Media Patent License 1.0 was not distributed with this source code in the |
9 | | * PATENTS file, you can obtain it at www.aomedia.org/license/patent. |
10 | | */ |
11 | | |
12 | | #include "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/intra_mode_search_utils.h" |
27 | | #include "av1/encoder/rdopt.h" |
28 | | #include "aom_dsp/aom_dsp_common.h" |
29 | | #include "av1/encoder/temporal_filter.h" |
30 | | #include "av1/encoder/tpl_model.h" |
31 | | |
32 | 0 | static AOM_INLINE void accumulate_rd_opt(ThreadData *td, ThreadData *td_t) { |
33 | 0 | td->rd_counts.compound_ref_used_flag |= |
34 | 0 | td_t->rd_counts.compound_ref_used_flag; |
35 | 0 | td->rd_counts.skip_mode_used_flag |= td_t->rd_counts.skip_mode_used_flag; |
36 | |
|
37 | 0 | for (int i = 0; i < TX_SIZES_ALL; i++) { |
38 | 0 | for (int j = 0; j < TX_TYPES; j++) |
39 | 0 | td->rd_counts.tx_type_used[i][j] += td_t->rd_counts.tx_type_used[i][j]; |
40 | 0 | } |
41 | |
|
42 | 0 | for (int i = 0; i < BLOCK_SIZES_ALL; i++) { |
43 | 0 | for (int j = 0; j < 2; j++) { |
44 | 0 | td->rd_counts.obmc_used[i][j] += td_t->rd_counts.obmc_used[i][j]; |
45 | 0 | } |
46 | 0 | } |
47 | |
|
48 | 0 | for (int i = 0; i < 2; i++) { |
49 | 0 | td->rd_counts.warped_used[i] += td_t->rd_counts.warped_used[i]; |
50 | 0 | } |
51 | |
|
52 | 0 | td->rd_counts.newmv_or_intra_blocks += td_t->rd_counts.newmv_or_intra_blocks; |
53 | 0 | } |
54 | | |
55 | 0 | static AOM_INLINE void update_delta_lf_for_row_mt(AV1_COMP *cpi) { |
56 | 0 | AV1_COMMON *cm = &cpi->common; |
57 | 0 | MACROBLOCKD *xd = &cpi->td.mb.e_mbd; |
58 | 0 | const int mib_size = cm->seq_params->mib_size; |
59 | 0 | const int frame_lf_count = |
60 | 0 | av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2; |
61 | 0 | for (int row = 0; row < cm->tiles.rows; row++) { |
62 | 0 | for (int col = 0; col < cm->tiles.cols; col++) { |
63 | 0 | TileDataEnc *tile_data = &cpi->tile_data[row * cm->tiles.cols + col]; |
64 | 0 | const TileInfo *const tile_info = &tile_data->tile_info; |
65 | 0 | for (int mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end; |
66 | 0 | mi_row += mib_size) { |
67 | 0 | if (mi_row == tile_info->mi_row_start) |
68 | 0 | av1_reset_loop_filter_delta(xd, av1_num_planes(cm)); |
69 | 0 | for (int mi_col = tile_info->mi_col_start; |
70 | 0 | mi_col < tile_info->mi_col_end; mi_col += mib_size) { |
71 | 0 | const int idx_str = cm->mi_params.mi_stride * mi_row + mi_col; |
72 | 0 | MB_MODE_INFO **mi = cm->mi_params.mi_grid_base + idx_str; |
73 | 0 | MB_MODE_INFO *mbmi = mi[0]; |
74 | 0 | if (mbmi->skip_txfm == 1 && |
75 | 0 | (mbmi->bsize == cm->seq_params->sb_size)) { |
76 | 0 | for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) |
77 | 0 | mbmi->delta_lf[lf_id] = xd->delta_lf[lf_id]; |
78 | 0 | mbmi->delta_lf_from_base = xd->delta_lf_from_base; |
79 | 0 | } else { |
80 | 0 | if (cm->delta_q_info.delta_lf_multi) { |
81 | 0 | for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) |
82 | 0 | xd->delta_lf[lf_id] = mbmi->delta_lf[lf_id]; |
83 | 0 | } else { |
84 | 0 | xd->delta_lf_from_base = mbmi->delta_lf_from_base; |
85 | 0 | } |
86 | 0 | } |
87 | 0 | } |
88 | 0 | } |
89 | 0 | } |
90 | 0 | } |
91 | 0 | } |
92 | | |
93 | | void av1_row_mt_sync_read_dummy(AV1EncRowMultiThreadSync *row_mt_sync, int r, |
94 | 0 | int c) { |
95 | 0 | (void)row_mt_sync; |
96 | 0 | (void)r; |
97 | 0 | (void)c; |
98 | 0 | return; |
99 | 0 | } |
100 | | |
101 | | void av1_row_mt_sync_write_dummy(AV1EncRowMultiThreadSync *row_mt_sync, int r, |
102 | 0 | int c, int cols) { |
103 | 0 | (void)row_mt_sync; |
104 | 0 | (void)r; |
105 | 0 | (void)c; |
106 | 0 | (void)cols; |
107 | 0 | return; |
108 | 0 | } |
109 | | |
110 | 0 | void av1_row_mt_sync_read(AV1EncRowMultiThreadSync *row_mt_sync, int r, int c) { |
111 | 0 | #if CONFIG_MULTITHREAD |
112 | 0 | const int nsync = row_mt_sync->sync_range; |
113 | |
|
114 | 0 | if (r) { |
115 | 0 | pthread_mutex_t *const mutex = &row_mt_sync->mutex_[r - 1]; |
116 | 0 | pthread_mutex_lock(mutex); |
117 | |
|
118 | 0 | while (c > row_mt_sync->num_finished_cols[r - 1] - nsync) { |
119 | 0 | pthread_cond_wait(&row_mt_sync->cond_[r - 1], mutex); |
120 | 0 | } |
121 | 0 | pthread_mutex_unlock(mutex); |
122 | 0 | } |
123 | | #else |
124 | | (void)row_mt_sync; |
125 | | (void)r; |
126 | | (void)c; |
127 | | #endif // CONFIG_MULTITHREAD |
128 | 0 | } |
129 | | |
130 | | void av1_row_mt_sync_write(AV1EncRowMultiThreadSync *row_mt_sync, int r, int c, |
131 | 0 | int cols) { |
132 | 0 | #if CONFIG_MULTITHREAD |
133 | 0 | const int nsync = row_mt_sync->sync_range; |
134 | 0 | int cur; |
135 | | // Only signal when there are enough encoded blocks for next row to run. |
136 | 0 | int sig = 1; |
137 | |
|
138 | 0 | if (c < cols - 1) { |
139 | 0 | cur = c; |
140 | 0 | if (c % nsync) sig = 0; |
141 | 0 | } else { |
142 | 0 | cur = cols + nsync; |
143 | 0 | } |
144 | |
|
145 | 0 | if (sig) { |
146 | 0 | pthread_mutex_lock(&row_mt_sync->mutex_[r]); |
147 | |
|
148 | 0 | row_mt_sync->num_finished_cols[r] = cur; |
149 | |
|
150 | 0 | pthread_cond_signal(&row_mt_sync->cond_[r]); |
151 | 0 | pthread_mutex_unlock(&row_mt_sync->mutex_[r]); |
152 | 0 | } |
153 | | #else |
154 | | (void)row_mt_sync; |
155 | | (void)r; |
156 | | (void)c; |
157 | | (void)cols; |
158 | | #endif // CONFIG_MULTITHREAD |
159 | 0 | } |
160 | | |
161 | | // Allocate memory for row synchronization |
162 | | static void row_mt_sync_mem_alloc(AV1EncRowMultiThreadSync *row_mt_sync, |
163 | 0 | AV1_COMMON *cm, int rows) { |
164 | 0 | #if CONFIG_MULTITHREAD |
165 | 0 | int i; |
166 | |
|
167 | 0 | CHECK_MEM_ERROR(cm, row_mt_sync->mutex_, |
168 | 0 | aom_malloc(sizeof(*row_mt_sync->mutex_) * rows)); |
169 | 0 | if (row_mt_sync->mutex_) { |
170 | 0 | for (i = 0; i < rows; ++i) { |
171 | 0 | pthread_mutex_init(&row_mt_sync->mutex_[i], NULL); |
172 | 0 | } |
173 | 0 | } |
174 | |
|
175 | 0 | CHECK_MEM_ERROR(cm, row_mt_sync->cond_, |
176 | 0 | aom_malloc(sizeof(*row_mt_sync->cond_) * rows)); |
177 | 0 | if (row_mt_sync->cond_) { |
178 | 0 | for (i = 0; i < rows; ++i) { |
179 | 0 | pthread_cond_init(&row_mt_sync->cond_[i], NULL); |
180 | 0 | } |
181 | 0 | } |
182 | 0 | #endif // CONFIG_MULTITHREAD |
183 | |
|
184 | 0 | CHECK_MEM_ERROR(cm, row_mt_sync->num_finished_cols, |
185 | 0 | aom_malloc(sizeof(*row_mt_sync->num_finished_cols) * rows)); |
186 | |
|
187 | 0 | row_mt_sync->rows = rows; |
188 | | // Set up nsync. |
189 | 0 | row_mt_sync->sync_range = 1; |
190 | 0 | } |
191 | | |
192 | | // Deallocate row based multi-threading synchronization related mutex and data |
193 | 0 | static void row_mt_sync_mem_dealloc(AV1EncRowMultiThreadSync *row_mt_sync) { |
194 | 0 | if (row_mt_sync != NULL) { |
195 | 0 | #if CONFIG_MULTITHREAD |
196 | 0 | int i; |
197 | |
|
198 | 0 | if (row_mt_sync->mutex_ != NULL) { |
199 | 0 | for (i = 0; i < row_mt_sync->rows; ++i) { |
200 | 0 | pthread_mutex_destroy(&row_mt_sync->mutex_[i]); |
201 | 0 | } |
202 | 0 | aom_free(row_mt_sync->mutex_); |
203 | 0 | } |
204 | 0 | if (row_mt_sync->cond_ != NULL) { |
205 | 0 | for (i = 0; i < row_mt_sync->rows; ++i) { |
206 | 0 | pthread_cond_destroy(&row_mt_sync->cond_[i]); |
207 | 0 | } |
208 | 0 | aom_free(row_mt_sync->cond_); |
209 | 0 | } |
210 | 0 | #endif // CONFIG_MULTITHREAD |
211 | 0 | 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 | 0 | av1_zero(*row_mt_sync); |
217 | 0 | } |
218 | 0 | } |
219 | | |
220 | | static void row_mt_mem_alloc(AV1_COMP *cpi, int max_rows, int max_cols, |
221 | 0 | int alloc_row_ctx) { |
222 | 0 | struct AV1Common *cm = &cpi->common; |
223 | 0 | AV1EncRowMultiThreadInfo *const enc_row_mt = &cpi->mt_info.enc_row_mt; |
224 | 0 | const int tile_cols = cm->tiles.cols; |
225 | 0 | const int tile_rows = cm->tiles.rows; |
226 | 0 | int tile_col, tile_row; |
227 | | |
228 | | // Allocate memory for row based multi-threading |
229 | 0 | for (tile_row = 0; tile_row < tile_rows; tile_row++) { |
230 | 0 | for (tile_col = 0; tile_col < tile_cols; tile_col++) { |
231 | 0 | int tile_index = tile_row * tile_cols + tile_col; |
232 | 0 | TileDataEnc *const this_tile = &cpi->tile_data[tile_index]; |
233 | |
|
234 | 0 | row_mt_sync_mem_alloc(&this_tile->row_mt_sync, cm, max_rows); |
235 | |
|
236 | 0 | this_tile->row_ctx = NULL; |
237 | 0 | if (alloc_row_ctx) { |
238 | 0 | assert(max_cols > 0); |
239 | 0 | const int num_row_ctx = AOMMAX(1, (max_cols - 1)); |
240 | 0 | CHECK_MEM_ERROR(cm, this_tile->row_ctx, |
241 | 0 | (FRAME_CONTEXT *)aom_memalign( |
242 | 0 | 16, num_row_ctx * sizeof(*this_tile->row_ctx))); |
243 | 0 | } |
244 | 0 | } |
245 | 0 | } |
246 | 0 | enc_row_mt->allocated_tile_cols = tile_cols; |
247 | 0 | enc_row_mt->allocated_tile_rows = tile_rows; |
248 | 0 | enc_row_mt->allocated_rows = max_rows; |
249 | 0 | enc_row_mt->allocated_cols = max_cols - 1; |
250 | 0 | } |
251 | | |
252 | 0 | void av1_row_mt_mem_dealloc(AV1_COMP *cpi) { |
253 | 0 | AV1EncRowMultiThreadInfo *const enc_row_mt = &cpi->mt_info.enc_row_mt; |
254 | 0 | const int tile_cols = enc_row_mt->allocated_tile_cols; |
255 | 0 | const int tile_rows = enc_row_mt->allocated_tile_rows; |
256 | 0 | int tile_col, tile_row; |
257 | | |
258 | | // Free row based multi-threading sync memory |
259 | 0 | for (tile_row = 0; tile_row < tile_rows; tile_row++) { |
260 | 0 | for (tile_col = 0; tile_col < tile_cols; tile_col++) { |
261 | 0 | int tile_index = tile_row * tile_cols + tile_col; |
262 | 0 | TileDataEnc *const this_tile = &cpi->tile_data[tile_index]; |
263 | |
|
264 | 0 | row_mt_sync_mem_dealloc(&this_tile->row_mt_sync); |
265 | |
|
266 | 0 | if (cpi->oxcf.algo_cfg.cdf_update_mode) aom_free(this_tile->row_ctx); |
267 | 0 | } |
268 | 0 | } |
269 | 0 | enc_row_mt->allocated_rows = 0; |
270 | 0 | enc_row_mt->allocated_cols = 0; |
271 | 0 | enc_row_mt->allocated_tile_cols = 0; |
272 | 0 | enc_row_mt->allocated_tile_rows = 0; |
273 | 0 | } |
274 | | |
275 | | static AOM_INLINE void assign_tile_to_thread(int *thread_id_to_tile_id, |
276 | 0 | int num_tiles, int num_workers) { |
277 | 0 | int tile_id = 0; |
278 | 0 | int i; |
279 | |
|
280 | 0 | for (i = 0; i < num_workers; i++) { |
281 | 0 | thread_id_to_tile_id[i] = tile_id++; |
282 | 0 | if (tile_id == num_tiles) tile_id = 0; |
283 | 0 | } |
284 | 0 | } |
285 | | |
286 | | static AOM_INLINE int get_next_job(TileDataEnc *const tile_data, |
287 | 0 | int *current_mi_row, int mib_size) { |
288 | 0 | AV1EncRowMultiThreadSync *const row_mt_sync = &tile_data->row_mt_sync; |
289 | 0 | const int mi_row_end = tile_data->tile_info.mi_row_end; |
290 | |
|
291 | 0 | if (row_mt_sync->next_mi_row < mi_row_end) { |
292 | 0 | *current_mi_row = row_mt_sync->next_mi_row; |
293 | 0 | row_mt_sync->num_threads_working++; |
294 | 0 | row_mt_sync->next_mi_row += mib_size; |
295 | 0 | return 1; |
296 | 0 | } |
297 | 0 | return 0; |
298 | 0 | } |
299 | | |
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 | 0 | const BLOCK_SIZE fp_block_size) { |
304 | 0 | const int tile_cols = cm->tiles.cols; |
305 | 0 | const int tile_rows = cm->tiles.rows; |
306 | |
|
307 | 0 | int tile_id = -1; // Stores the tile ID with minimum proc done |
308 | 0 | int max_mis_to_encode = 0; |
309 | 0 | int min_num_threads_working = INT_MAX; |
310 | |
|
311 | 0 | for (int tile_row = 0; tile_row < tile_rows; tile_row++) { |
312 | 0 | for (int tile_col = 0; tile_col < tile_cols; tile_col++) { |
313 | 0 | int tile_index = tile_row * tile_cols + tile_col; |
314 | 0 | TileDataEnc *const this_tile = &tile_data[tile_index]; |
315 | 0 | 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 | 0 | int num_b_rows_in_tile = |
324 | 0 | is_firstpass |
325 | 0 | ? av1_get_unit_rows_in_tile(this_tile->tile_info, fp_block_size) |
326 | 0 | : av1_get_sb_rows_in_tile(cm, this_tile->tile_info); |
327 | 0 | int num_b_cols_in_tile = |
328 | 0 | is_firstpass |
329 | 0 | ? av1_get_unit_cols_in_tile(this_tile->tile_info, fp_block_size) |
330 | 0 | : av1_get_sb_cols_in_tile(cm, this_tile->tile_info); |
331 | 0 | #endif |
332 | 0 | int theoretical_limit_on_threads = |
333 | 0 | AOMMIN((num_b_cols_in_tile + 1) >> 1, num_b_rows_in_tile); |
334 | 0 | int num_threads_working = row_mt_sync->num_threads_working; |
335 | |
|
336 | 0 | if (num_threads_working < theoretical_limit_on_threads) { |
337 | 0 | int num_mis_to_encode = |
338 | 0 | 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 | 0 | if (num_mis_to_encode > 0) { |
348 | 0 | if (num_threads_working < min_num_threads_working) { |
349 | 0 | min_num_threads_working = num_threads_working; |
350 | 0 | max_mis_to_encode = 0; |
351 | 0 | } |
352 | 0 | if (num_threads_working == min_num_threads_working && |
353 | 0 | num_mis_to_encode > max_mis_to_encode) { |
354 | 0 | tile_id = tile_index; |
355 | 0 | max_mis_to_encode = num_mis_to_encode; |
356 | 0 | } |
357 | 0 | } |
358 | 0 | } |
359 | 0 | } |
360 | 0 | } |
361 | 0 | if (tile_id == -1) { |
362 | 0 | *end_of_frame = 1; |
363 | 0 | } 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 | 0 | *cur_tile_id = tile_id; |
367 | 0 | const int unit_height = mi_size_high[fp_block_size]; |
368 | 0 | get_next_job(&tile_data[tile_id], current_mi_row, |
369 | 0 | is_firstpass ? unit_height : cm->seq_params->mib_size); |
370 | 0 | } |
371 | 0 | } |
372 | | |
373 | | #if !CONFIG_REALTIME_ONLY |
374 | 0 | static int fp_enc_row_mt_worker_hook(void *arg1, void *unused) { |
375 | 0 | EncWorkerData *const thread_data = (EncWorkerData *)arg1; |
376 | 0 | AV1_COMP *const cpi = thread_data->cpi; |
377 | 0 | AV1_COMMON *const cm = &cpi->common; |
378 | 0 | int thread_id = thread_data->thread_id; |
379 | 0 | AV1EncRowMultiThreadInfo *const enc_row_mt = &cpi->mt_info.enc_row_mt; |
380 | 0 | int cur_tile_id = enc_row_mt->thread_id_to_tile_id[thread_id]; |
381 | 0 | #if CONFIG_MULTITHREAD |
382 | 0 | pthread_mutex_t *enc_row_mt_mutex_ = enc_row_mt->mutex_; |
383 | 0 | #endif |
384 | 0 | (void)unused; |
385 | |
|
386 | 0 | assert(cur_tile_id != -1); |
387 | |
|
388 | 0 | const BLOCK_SIZE fp_block_size = cpi->fp_block_size; |
389 | 0 | const int unit_height = mi_size_high[fp_block_size]; |
390 | 0 | int end_of_frame = 0; |
391 | 0 | while (1) { |
392 | 0 | int current_mi_row = -1; |
393 | 0 | #if CONFIG_MULTITHREAD |
394 | 0 | pthread_mutex_lock(enc_row_mt_mutex_); |
395 | 0 | #endif |
396 | 0 | if (!get_next_job(&cpi->tile_data[cur_tile_id], ¤t_mi_row, |
397 | 0 | 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 | 0 | switch_tile_and_get_next_job(cm, cpi->tile_data, &cur_tile_id, |
401 | 0 | ¤t_mi_row, &end_of_frame, 1, |
402 | 0 | fp_block_size); |
403 | 0 | } |
404 | 0 | #if CONFIG_MULTITHREAD |
405 | 0 | pthread_mutex_unlock(enc_row_mt_mutex_); |
406 | 0 | #endif |
407 | 0 | if (end_of_frame == 1) break; |
408 | | |
409 | 0 | TileDataEnc *const this_tile = &cpi->tile_data[cur_tile_id]; |
410 | 0 | AV1EncRowMultiThreadSync *const row_mt_sync = &this_tile->row_mt_sync; |
411 | 0 | ThreadData *td = thread_data->td; |
412 | |
|
413 | 0 | assert(current_mi_row != -1 && |
414 | 0 | current_mi_row <= this_tile->tile_info.mi_row_end); |
415 | |
|
416 | 0 | const int unit_height_log2 = mi_size_high_log2[fp_block_size]; |
417 | 0 | av1_first_pass_row(cpi, td, this_tile, current_mi_row >> unit_height_log2, |
418 | 0 | fp_block_size); |
419 | 0 | #if CONFIG_MULTITHREAD |
420 | 0 | pthread_mutex_lock(enc_row_mt_mutex_); |
421 | 0 | #endif |
422 | 0 | row_mt_sync->num_threads_working--; |
423 | 0 | #if CONFIG_MULTITHREAD |
424 | 0 | pthread_mutex_unlock(enc_row_mt_mutex_); |
425 | 0 | #endif |
426 | 0 | } |
427 | |
|
428 | 0 | return 1; |
429 | 0 | } |
430 | | #endif |
431 | | |
432 | 0 | static int enc_row_mt_worker_hook(void *arg1, void *unused) { |
433 | 0 | EncWorkerData *const thread_data = (EncWorkerData *)arg1; |
434 | 0 | AV1_COMP *const cpi = thread_data->cpi; |
435 | 0 | AV1_COMMON *const cm = &cpi->common; |
436 | 0 | int thread_id = thread_data->thread_id; |
437 | 0 | AV1EncRowMultiThreadInfo *const enc_row_mt = &cpi->mt_info.enc_row_mt; |
438 | 0 | int cur_tile_id = enc_row_mt->thread_id_to_tile_id[thread_id]; |
439 | 0 | #if CONFIG_MULTITHREAD |
440 | 0 | pthread_mutex_t *enc_row_mt_mutex_ = enc_row_mt->mutex_; |
441 | 0 | #endif |
442 | 0 | (void)unused; |
443 | |
|
444 | 0 | assert(cur_tile_id != -1); |
445 | |
|
446 | 0 | const BLOCK_SIZE fp_block_size = cpi->fp_block_size; |
447 | 0 | 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 | 0 | thread_data->td->mb.e_mbd.tile_ctx = cm->fc; |
455 | 0 | while (1) { |
456 | 0 | int current_mi_row = -1; |
457 | 0 | #if CONFIG_MULTITHREAD |
458 | 0 | pthread_mutex_lock(enc_row_mt_mutex_); |
459 | 0 | #endif |
460 | 0 | if (!get_next_job(&cpi->tile_data[cur_tile_id], ¤t_mi_row, |
461 | 0 | 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 | 0 | switch_tile_and_get_next_job(cm, cpi->tile_data, &cur_tile_id, |
465 | 0 | ¤t_mi_row, &end_of_frame, 0, |
466 | 0 | fp_block_size); |
467 | 0 | } |
468 | 0 | #if CONFIG_MULTITHREAD |
469 | 0 | pthread_mutex_unlock(enc_row_mt_mutex_); |
470 | 0 | #endif |
471 | 0 | if (end_of_frame == 1) break; |
472 | | |
473 | 0 | TileDataEnc *const this_tile = &cpi->tile_data[cur_tile_id]; |
474 | 0 | AV1EncRowMultiThreadSync *const row_mt_sync = &this_tile->row_mt_sync; |
475 | 0 | const TileInfo *const tile_info = &this_tile->tile_info; |
476 | 0 | const int tile_row = tile_info->tile_row; |
477 | 0 | const int tile_col = tile_info->tile_col; |
478 | 0 | ThreadData *td = thread_data->td; |
479 | |
|
480 | 0 | assert(current_mi_row != -1 && current_mi_row <= tile_info->mi_row_end); |
481 | |
|
482 | 0 | td->mb.e_mbd.tile_ctx = td->tctx; |
483 | 0 | td->mb.tile_pb_ctx = &this_tile->tctx; |
484 | 0 | td->abs_sum_level = 0; |
485 | |
|
486 | 0 | if (this_tile->allow_update_cdf) { |
487 | 0 | td->mb.row_ctx = this_tile->row_ctx; |
488 | 0 | if (current_mi_row == tile_info->mi_row_start) |
489 | 0 | memcpy(td->mb.e_mbd.tile_ctx, &this_tile->tctx, sizeof(FRAME_CONTEXT)); |
490 | 0 | } else { |
491 | 0 | memcpy(td->mb.e_mbd.tile_ctx, &this_tile->tctx, sizeof(FRAME_CONTEXT)); |
492 | 0 | } |
493 | |
|
494 | 0 | av1_init_above_context(&cm->above_contexts, av1_num_planes(cm), tile_row, |
495 | 0 | &td->mb.e_mbd); |
496 | |
|
497 | 0 | cfl_init(&td->mb.e_mbd.cfl, cm->seq_params); |
498 | 0 | if (td->mb.txfm_search_info.mb_rd_record != NULL) { |
499 | 0 | av1_crc32c_calculator_init( |
500 | 0 | &td->mb.txfm_search_info.mb_rd_record->crc_calculator); |
501 | 0 | } |
502 | |
|
503 | 0 | av1_encode_sb_row(cpi, td, tile_row, tile_col, current_mi_row); |
504 | 0 | #if CONFIG_MULTITHREAD |
505 | 0 | pthread_mutex_lock(enc_row_mt_mutex_); |
506 | 0 | #endif |
507 | 0 | this_tile->abs_sum_level += td->abs_sum_level; |
508 | 0 | row_mt_sync->num_threads_working--; |
509 | 0 | #if CONFIG_MULTITHREAD |
510 | 0 | pthread_mutex_unlock(enc_row_mt_mutex_); |
511 | 0 | #endif |
512 | 0 | } |
513 | |
|
514 | 0 | return 1; |
515 | 0 | } |
516 | | |
517 | 0 | static int enc_worker_hook(void *arg1, void *unused) { |
518 | 0 | EncWorkerData *const thread_data = (EncWorkerData *)arg1; |
519 | 0 | AV1_COMP *const cpi = thread_data->cpi; |
520 | 0 | const AV1_COMMON *const cm = &cpi->common; |
521 | 0 | const int tile_cols = cm->tiles.cols; |
522 | 0 | const int tile_rows = cm->tiles.rows; |
523 | 0 | int t; |
524 | |
|
525 | 0 | (void)unused; |
526 | |
|
527 | 0 | for (t = thread_data->start; t < tile_rows * tile_cols; |
528 | 0 | t += cpi->mt_info.num_workers) { |
529 | 0 | int tile_row = t / tile_cols; |
530 | 0 | int tile_col = t % tile_cols; |
531 | |
|
532 | 0 | TileDataEnc *const this_tile = |
533 | 0 | &cpi->tile_data[tile_row * cm->tiles.cols + tile_col]; |
534 | 0 | thread_data->td->mb.e_mbd.tile_ctx = &this_tile->tctx; |
535 | 0 | thread_data->td->mb.tile_pb_ctx = &this_tile->tctx; |
536 | 0 | av1_encode_tile(cpi, thread_data->td, tile_row, tile_col); |
537 | 0 | } |
538 | |
|
539 | 0 | return 1; |
540 | 0 | } |
541 | | |
542 | 0 | void av1_init_frame_mt(AV1_PRIMARY *ppi, AV1_COMP *cpi) { |
543 | 0 | cpi->mt_info.workers = ppi->p_mt_info.workers; |
544 | 0 | cpi->mt_info.num_workers = ppi->p_mt_info.num_workers; |
545 | 0 | cpi->mt_info.tile_thr_data = ppi->p_mt_info.tile_thr_data; |
546 | 0 | int i; |
547 | 0 | for (i = MOD_FP; i < NUM_MT_MODULES; i++) { |
548 | 0 | cpi->mt_info.num_mod_workers[i] = |
549 | 0 | AOMMIN(cpi->mt_info.num_workers, ppi->p_mt_info.num_mod_workers[i]); |
550 | 0 | } |
551 | 0 | } |
552 | | |
553 | 0 | 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 | 0 | PrimaryMultiThreadInfo *const p_mt_info = &cpi->ppi->p_mt_info; |
561 | 0 | int num_cdef_workers = av1_get_num_mod_workers_for_alloc(p_mt_info, MOD_CDEF); |
562 | |
|
563 | 0 | av1_alloc_cdef_buffers(&cpi->common, &p_mt_info->cdef_worker, |
564 | 0 | &cpi->mt_info.cdef_sync, num_cdef_workers, 1); |
565 | 0 | cpi->mt_info.cdef_worker = &p_mt_info->cdef_worker[0]; |
566 | 0 | } |
567 | | |
568 | | #if !CONFIG_REALTIME_ONLY |
569 | 0 | void av1_init_lr_mt_buffers(AV1_COMP *cpi) { |
570 | 0 | AV1_COMMON *const cm = &cpi->common; |
571 | 0 | AV1LrSync *lr_sync = &cpi->mt_info.lr_row_sync; |
572 | 0 | if (lr_sync->sync_range) { |
573 | 0 | int num_lr_workers = |
574 | 0 | 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 | 0 | lr_sync->lrworkerdata[num_lr_workers - 1].rst_tmpbuf = cm->rst_tmpbuf; |
580 | 0 | lr_sync->lrworkerdata[num_lr_workers - 1].rlbs = cm->rlbs; |
581 | 0 | } |
582 | 0 | } |
583 | | #endif |
584 | | |
585 | | #if CONFIG_MULTITHREAD |
586 | 0 | void av1_init_mt_sync(AV1_COMP *cpi, int is_first_pass) { |
587 | 0 | AV1_COMMON *const cm = &cpi->common; |
588 | 0 | MultiThreadInfo *const mt_info = &cpi->mt_info; |
589 | | |
590 | | // Initialize enc row MT object. |
591 | 0 | if (is_first_pass || cpi->oxcf.row_mt == 1) { |
592 | 0 | AV1EncRowMultiThreadInfo *enc_row_mt = &mt_info->enc_row_mt; |
593 | 0 | if (enc_row_mt->mutex_ == NULL) { |
594 | 0 | CHECK_MEM_ERROR(cm, enc_row_mt->mutex_, |
595 | 0 | aom_malloc(sizeof(*(enc_row_mt->mutex_)))); |
596 | 0 | if (enc_row_mt->mutex_) pthread_mutex_init(enc_row_mt->mutex_, NULL); |
597 | 0 | } |
598 | 0 | } |
599 | |
|
600 | 0 | if (!is_first_pass) { |
601 | | // Initialize global motion MT object. |
602 | 0 | AV1GlobalMotionSync *gm_sync = &mt_info->gm_sync; |
603 | 0 | if (gm_sync->mutex_ == NULL) { |
604 | 0 | CHECK_MEM_ERROR(cm, gm_sync->mutex_, |
605 | 0 | aom_malloc(sizeof(*(gm_sync->mutex_)))); |
606 | 0 | if (gm_sync->mutex_) pthread_mutex_init(gm_sync->mutex_, NULL); |
607 | 0 | } |
608 | 0 | #if !CONFIG_REALTIME_ONLY |
609 | | // Initialize temporal filtering MT object. |
610 | 0 | AV1TemporalFilterSync *tf_sync = &mt_info->tf_sync; |
611 | 0 | if (tf_sync->mutex_ == NULL) { |
612 | 0 | CHECK_MEM_ERROR(cm, tf_sync->mutex_, |
613 | 0 | aom_malloc(sizeof(*tf_sync->mutex_))); |
614 | 0 | if (tf_sync->mutex_) pthread_mutex_init(tf_sync->mutex_, NULL); |
615 | 0 | } |
616 | 0 | #endif // !CONFIG_REALTIME_ONLY |
617 | | // Initialize CDEF MT object. |
618 | 0 | AV1CdefSync *cdef_sync = &mt_info->cdef_sync; |
619 | 0 | if (cdef_sync->mutex_ == NULL) { |
620 | 0 | CHECK_MEM_ERROR(cm, cdef_sync->mutex_, |
621 | 0 | aom_malloc(sizeof(*(cdef_sync->mutex_)))); |
622 | 0 | if (cdef_sync->mutex_) pthread_mutex_init(cdef_sync->mutex_, NULL); |
623 | 0 | } |
624 | | |
625 | | // Initialize loop filter MT object. |
626 | 0 | AV1LfSync *lf_sync = &mt_info->lf_row_sync; |
627 | | // Number of superblock rows |
628 | 0 | const int sb_rows = |
629 | 0 | ALIGN_POWER_OF_TWO(cm->height >> MI_SIZE_LOG2, MAX_MIB_SIZE_LOG2) >> |
630 | 0 | MAX_MIB_SIZE_LOG2; |
631 | 0 | PrimaryMultiThreadInfo *const p_mt_info = &cpi->ppi->p_mt_info; |
632 | 0 | int num_lf_workers = av1_get_num_mod_workers_for_alloc(p_mt_info, MOD_LPF); |
633 | |
|
634 | 0 | if (!lf_sync->sync_range || sb_rows != lf_sync->rows || |
635 | 0 | num_lf_workers > lf_sync->num_workers) { |
636 | 0 | av1_loop_filter_dealloc(lf_sync); |
637 | 0 | av1_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_lf_workers); |
638 | 0 | } |
639 | |
|
640 | 0 | #if !CONFIG_REALTIME_ONLY |
641 | 0 | if (is_restoration_used(cm)) { |
642 | | // Initialize loop restoration MT object. |
643 | 0 | AV1LrSync *lr_sync = &mt_info->lr_row_sync; |
644 | 0 | int rst_unit_size; |
645 | 0 | if (cm->width * cm->height > 352 * 288) |
646 | 0 | rst_unit_size = RESTORATION_UNITSIZE_MAX; |
647 | 0 | else |
648 | 0 | rst_unit_size = (RESTORATION_UNITSIZE_MAX >> 1); |
649 | 0 | int num_rows_lr = av1_lr_count_units_in_tile(rst_unit_size, cm->height); |
650 | 0 | int num_lr_workers = av1_get_num_mod_workers_for_alloc(p_mt_info, MOD_LR); |
651 | 0 | if (!lr_sync->sync_range || num_rows_lr > lr_sync->rows || |
652 | 0 | num_lr_workers > lr_sync->num_workers || |
653 | 0 | MAX_MB_PLANE > lr_sync->num_planes) { |
654 | 0 | av1_loop_restoration_dealloc(lr_sync, num_lr_workers); |
655 | 0 | av1_loop_restoration_alloc(lr_sync, cm, num_lr_workers, num_rows_lr, |
656 | 0 | MAX_MB_PLANE, cm->width); |
657 | 0 | } |
658 | 0 | } |
659 | 0 | #endif |
660 | | |
661 | | // Initialization of pack bitstream MT object. |
662 | 0 | AV1EncPackBSSync *pack_bs_sync = &mt_info->pack_bs_sync; |
663 | 0 | if (pack_bs_sync->mutex_ == NULL) { |
664 | 0 | CHECK_MEM_ERROR(cm, pack_bs_sync->mutex_, |
665 | 0 | aom_malloc(sizeof(*pack_bs_sync->mutex_))); |
666 | 0 | if (pack_bs_sync->mutex_) pthread_mutex_init(pack_bs_sync->mutex_, NULL); |
667 | 0 | } |
668 | 0 | } |
669 | 0 | } |
670 | | #endif // CONFIG_MULTITHREAD |
671 | | |
672 | | // Computes the number of workers to be considered while allocating memory for a |
673 | | // multi-threaded module under FPMT. |
674 | | int av1_get_num_mod_workers_for_alloc(PrimaryMultiThreadInfo *const p_mt_info, |
675 | 0 | MULTI_THREADED_MODULES mod_name) { |
676 | 0 | int num_mod_workers = p_mt_info->num_mod_workers[mod_name]; |
677 | 0 | if (p_mt_info->num_mod_workers[MOD_FRAME_ENC] > 1) { |
678 | | // TODO(anyone): Change num_mod_workers to num_mod_workers[MOD_FRAME_ENC]. |
679 | | // As frame parallel jobs will only perform multi-threading for the encode |
680 | | // stage, we can limit the allocations according to num_enc_workers per |
681 | | // frame parallel encode(a.k.a num_mod_workers[MOD_FRAME_ENC]). |
682 | 0 | num_mod_workers = p_mt_info->num_workers; |
683 | 0 | } |
684 | 0 | return num_mod_workers; |
685 | 0 | } |
686 | | |
687 | 0 | void av1_init_tile_thread_data(AV1_PRIMARY *ppi, int is_first_pass) { |
688 | 0 | PrimaryMultiThreadInfo *const p_mt_info = &ppi->p_mt_info; |
689 | |
|
690 | 0 | assert(p_mt_info->workers != NULL); |
691 | 0 | assert(p_mt_info->tile_thr_data != NULL); |
692 | |
|
693 | 0 | int num_workers = p_mt_info->num_workers; |
694 | 0 | int num_enc_workers = av1_get_num_mod_workers_for_alloc(p_mt_info, MOD_ENC); |
695 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
696 | 0 | EncWorkerData *const thread_data = &p_mt_info->tile_thr_data[i]; |
697 | |
|
698 | 0 | if (i > 0) { |
699 | | // Allocate thread data. |
700 | 0 | AOM_CHECK_MEM_ERROR(&ppi->error, thread_data->td, |
701 | 0 | aom_memalign(32, sizeof(*thread_data->td))); |
702 | 0 | av1_zero(*thread_data->td); |
703 | | #if CONFIG_FRAME_PARALLEL_ENCODE |
704 | | thread_data->original_td = thread_data->td; |
705 | | #endif |
706 | | |
707 | | // Set up shared coeff buffers. |
708 | 0 | av1_setup_shared_coeff_buffer( |
709 | 0 | &ppi->seq_params, &thread_data->td->shared_coeff_buf, &ppi->error); |
710 | 0 | AOM_CHECK_MEM_ERROR( |
711 | 0 | &ppi->error, thread_data->td->tmp_conv_dst, |
712 | 0 | aom_memalign(32, MAX_SB_SIZE * MAX_SB_SIZE * |
713 | 0 | sizeof(*thread_data->td->tmp_conv_dst))); |
714 | |
|
715 | 0 | if (i < p_mt_info->num_mod_workers[MOD_FP]) { |
716 | | // Set up firstpass PICK_MODE_CONTEXT. |
717 | 0 | thread_data->td->firstpass_ctx = av1_alloc_pmc( |
718 | 0 | ppi->cpi, BLOCK_16X16, &thread_data->td->shared_coeff_buf); |
719 | 0 | } |
720 | |
|
721 | 0 | if (!is_first_pass && i < num_enc_workers) { |
722 | | // Set up sms_tree. |
723 | 0 | av1_setup_sms_tree(ppi->cpi, thread_data->td); |
724 | |
|
725 | 0 | for (int x = 0; x < 2; x++) |
726 | 0 | for (int y = 0; y < 2; y++) |
727 | 0 | AOM_CHECK_MEM_ERROR( |
728 | 0 | &ppi->error, thread_data->td->hash_value_buffer[x][y], |
729 | 0 | (uint32_t *)aom_malloc( |
730 | 0 | AOM_BUFFER_SIZE_FOR_BLOCK_HASH * |
731 | 0 | sizeof(*thread_data->td->hash_value_buffer[0][0]))); |
732 | | |
733 | | // Allocate frame counters in thread data. |
734 | 0 | AOM_CHECK_MEM_ERROR(&ppi->error, thread_data->td->counts, |
735 | 0 | aom_calloc(1, sizeof(*thread_data->td->counts))); |
736 | | |
737 | | // Allocate buffers used by palette coding mode. |
738 | 0 | AOM_CHECK_MEM_ERROR( |
739 | 0 | &ppi->error, thread_data->td->palette_buffer, |
740 | 0 | aom_memalign(16, sizeof(*thread_data->td->palette_buffer))); |
741 | | |
742 | | // The buffers 'tmp_pred_bufs[]', 'comp_rd_buffer' and 'obmc_buffer' are |
743 | | // used in inter frames to store intermediate inter mode prediction |
744 | | // results and are not required for allintra encoding mode. Hence, the |
745 | | // memory allocations for these buffers are avoided for allintra |
746 | | // encoding mode. |
747 | 0 | if (ppi->cpi->oxcf.kf_cfg.key_freq_max != 0) { |
748 | 0 | alloc_obmc_buffers(&thread_data->td->obmc_buffer, &ppi->error); |
749 | |
|
750 | 0 | alloc_compound_type_rd_buffers(&ppi->error, |
751 | 0 | &thread_data->td->comp_rd_buffer); |
752 | |
|
753 | 0 | for (int j = 0; j < 2; ++j) { |
754 | 0 | AOM_CHECK_MEM_ERROR( |
755 | 0 | &ppi->error, thread_data->td->tmp_pred_bufs[j], |
756 | 0 | aom_memalign(32, |
757 | 0 | 2 * MAX_MB_PLANE * MAX_SB_SQUARE * |
758 | 0 | sizeof(*thread_data->td->tmp_pred_bufs[j]))); |
759 | 0 | } |
760 | 0 | } |
761 | |
|
762 | 0 | if (is_gradient_caching_for_hog_enabled(ppi->cpi)) { |
763 | 0 | const int plane_types = PLANE_TYPES >> ppi->seq_params.monochrome; |
764 | 0 | AOM_CHECK_MEM_ERROR( |
765 | 0 | &ppi->error, thread_data->td->pixel_gradient_info, |
766 | 0 | aom_malloc(sizeof(*thread_data->td->pixel_gradient_info) * |
767 | 0 | plane_types * MAX_SB_SQUARE)); |
768 | 0 | } |
769 | |
|
770 | 0 | if (ppi->cpi->sf.part_sf.partition_search_type == VAR_BASED_PARTITION) { |
771 | 0 | const int num_64x64_blocks = |
772 | 0 | (ppi->seq_params.sb_size == BLOCK_64X64) ? 1 : 4; |
773 | 0 | AOM_CHECK_MEM_ERROR( |
774 | 0 | &ppi->error, thread_data->td->vt64x64, |
775 | 0 | aom_malloc(sizeof(*thread_data->td->vt64x64) * num_64x64_blocks)); |
776 | 0 | } |
777 | 0 | } |
778 | 0 | } |
779 | |
|
780 | 0 | if (!is_first_pass && ppi->cpi->oxcf.row_mt == 1 && i < num_enc_workers) { |
781 | 0 | if (i == 0) { |
782 | | #if CONFIG_FRAME_PARALLEL_ENCODE |
783 | | for (int j = 0; j < ppi->num_fp_contexts; j++) { |
784 | | AOM_CHECK_MEM_ERROR(&ppi->error, ppi->parallel_cpi[j]->td.tctx, |
785 | | (FRAME_CONTEXT *)aom_memalign( |
786 | | 16, sizeof(*ppi->parallel_cpi[j]->td.tctx))); |
787 | | } |
788 | | #else |
789 | 0 | AOM_CHECK_MEM_ERROR( |
790 | 0 | &ppi->error, ppi->cpi->td.tctx, |
791 | 0 | (FRAME_CONTEXT *)aom_memalign(16, sizeof(*ppi->cpi->td.tctx))); |
792 | 0 | #endif |
793 | 0 | } else { |
794 | 0 | AOM_CHECK_MEM_ERROR( |
795 | 0 | &ppi->error, thread_data->td->tctx, |
796 | 0 | (FRAME_CONTEXT *)aom_memalign(16, sizeof(*thread_data->td->tctx))); |
797 | 0 | } |
798 | 0 | } |
799 | 0 | } |
800 | 0 | } |
801 | | |
802 | 0 | void av1_create_workers(AV1_PRIMARY *ppi, int num_workers) { |
803 | 0 | PrimaryMultiThreadInfo *const p_mt_info = &ppi->p_mt_info; |
804 | 0 | const AVxWorkerInterface *const winterface = aom_get_worker_interface(); |
805 | |
|
806 | 0 | AOM_CHECK_MEM_ERROR(&ppi->error, p_mt_info->workers, |
807 | 0 | aom_malloc(num_workers * sizeof(*p_mt_info->workers))); |
808 | |
|
809 | 0 | AOM_CHECK_MEM_ERROR( |
810 | 0 | &ppi->error, p_mt_info->tile_thr_data, |
811 | 0 | aom_calloc(num_workers, sizeof(*p_mt_info->tile_thr_data))); |
812 | |
|
813 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
814 | 0 | AVxWorker *const worker = &p_mt_info->workers[i]; |
815 | 0 | EncWorkerData *const thread_data = &p_mt_info->tile_thr_data[i]; |
816 | |
|
817 | 0 | winterface->init(worker); |
818 | 0 | worker->thread_name = "aom enc worker"; |
819 | |
|
820 | 0 | thread_data->thread_id = i; |
821 | | // Set the starting tile for each thread. |
822 | 0 | thread_data->start = i; |
823 | |
|
824 | 0 | if (i > 0) { |
825 | | // Create threads |
826 | 0 | if (!winterface->reset(worker)) |
827 | 0 | aom_internal_error(&ppi->error, AOM_CODEC_ERROR, |
828 | 0 | "Tile encoder thread creation failed"); |
829 | 0 | } |
830 | 0 | winterface->sync(worker); |
831 | |
|
832 | 0 | ++p_mt_info->num_workers; |
833 | 0 | } |
834 | 0 | } |
835 | | |
836 | | #if CONFIG_FRAME_PARALLEL_ENCODE |
837 | | // This function returns 1 if frame parallel encode is supported for |
838 | | // the current configuration. Returns 0 otherwise. |
839 | | static AOM_INLINE int is_fpmt_config(AV1_PRIMARY *ppi, AV1EncoderConfig *oxcf) { |
840 | | // FPMT is enabled for AOM_Q and AOM_VBR. |
841 | | // TODO(Tarun): Test and enable resize config. |
842 | | if (oxcf->rc_cfg.mode == AOM_CBR || oxcf->rc_cfg.mode == AOM_CQ) { |
843 | | return 0; |
844 | | } |
845 | | if (ppi->use_svc) { |
846 | | return 0; |
847 | | } |
848 | | if (oxcf->tile_cfg.enable_large_scale_tile) { |
849 | | return 0; |
850 | | } |
851 | | if (oxcf->dec_model_cfg.timing_info_present) { |
852 | | return 0; |
853 | | } |
854 | | if (oxcf->mode != GOOD) { |
855 | | return 0; |
856 | | } |
857 | | if (oxcf->tool_cfg.error_resilient_mode) { |
858 | | return 0; |
859 | | } |
860 | | if (oxcf->resize_cfg.resize_mode) { |
861 | | return 0; |
862 | | } |
863 | | if (oxcf->pass != AOM_RC_SECOND_PASS) { |
864 | | return 0; |
865 | | } |
866 | | if (oxcf->max_threads < 2) { |
867 | | return 0; |
868 | | } |
869 | | if (!oxcf->fp_mt) { |
870 | | return 0; |
871 | | } |
872 | | |
873 | | return 1; |
874 | | } |
875 | | |
876 | | int av1_check_fpmt_config(AV1_PRIMARY *const ppi, |
877 | | AV1EncoderConfig *const oxcf) { |
878 | | if (is_fpmt_config(ppi, oxcf)) return 1; |
879 | | // Reset frame parallel configuration for unsupported config |
880 | | if (ppi->num_fp_contexts > 1) { |
881 | | for (int i = 1; i < ppi->num_fp_contexts; i++) { |
882 | | // Release the previously-used frame-buffer |
883 | | if (ppi->parallel_cpi[i]->common.cur_frame != NULL) { |
884 | | --ppi->parallel_cpi[i]->common.cur_frame->ref_count; |
885 | | ppi->parallel_cpi[i]->common.cur_frame = NULL; |
886 | | } |
887 | | } |
888 | | |
889 | | int cur_gf_index = ppi->cpi->gf_frame_index; |
890 | | int reset_size = AOMMAX(0, ppi->gf_group.size - cur_gf_index); |
891 | | av1_zero_array(&ppi->gf_group.frame_parallel_level[cur_gf_index], |
892 | | reset_size); |
893 | | av1_zero_array(&ppi->gf_group.is_frame_non_ref[cur_gf_index], reset_size); |
894 | | av1_zero_array(&ppi->gf_group.src_offset[cur_gf_index], reset_size); |
895 | | #if CONFIG_FRAME_PARALLEL_ENCODE_2 |
896 | | memset(&ppi->gf_group.skip_frame_refresh[cur_gf_index][0], INVALID_IDX, |
897 | | sizeof(ppi->gf_group.skip_frame_refresh[cur_gf_index][0]) * |
898 | | reset_size * REF_FRAMES); |
899 | | memset(&ppi->gf_group.skip_frame_as_ref[cur_gf_index], INVALID_IDX, |
900 | | sizeof(ppi->gf_group.skip_frame_as_ref[cur_gf_index]) * reset_size); |
901 | | #endif |
902 | | ppi->num_fp_contexts = 1; |
903 | | } |
904 | | return 0; |
905 | | } |
906 | | |
907 | | // A large value for threads used to compute the max num_enc_workers |
908 | | // possible for each resolution. |
909 | | #define MAX_THREADS 100 |
910 | | |
911 | | // Computes the max number of enc workers possible for each resolution. |
912 | | static AOM_INLINE int compute_max_num_enc_workers( |
913 | | CommonModeInfoParams *const mi_params, int mib_size_log2) { |
914 | | int num_sb_rows = |
915 | | ALIGN_POWER_OF_TWO(mi_params->mi_rows, mib_size_log2) >> mib_size_log2; |
916 | | int num_sb_cols = |
917 | | ALIGN_POWER_OF_TWO(mi_params->mi_cols, mib_size_log2) >> mib_size_log2; |
918 | | |
919 | | return AOMMIN((num_sb_cols + 1) >> 1, num_sb_rows); |
920 | | } |
921 | | |
922 | | // Computes the number of frame parallel(fp) contexts to be created |
923 | | // based on the number of max_enc_workers. |
924 | | int av1_compute_num_fp_contexts(AV1_PRIMARY *ppi, AV1EncoderConfig *oxcf) { |
925 | | ppi->p_mt_info.num_mod_workers[MOD_FRAME_ENC] = 0; |
926 | | if (!av1_check_fpmt_config(ppi, oxcf)) { |
927 | | return 1; |
928 | | } |
929 | | int max_num_enc_workers = compute_max_num_enc_workers( |
930 | | &ppi->cpi->common.mi_params, ppi->cpi->common.seq_params->mib_size_log2); |
931 | | // Scaling factors and rounding factors used to tune worker_per_frame |
932 | | // computation. |
933 | | int rounding_factor[2] = { 2, 4 }; |
934 | | int scaling_factor[2] = { 4, 8 }; |
935 | | int is_480p_or_lesser = |
936 | | AOMMIN(oxcf->frm_dim_cfg.width, oxcf->frm_dim_cfg.height) <= 480; |
937 | | int is_sb_64 = 0; |
938 | | if (ppi->cpi != NULL) |
939 | | is_sb_64 = ppi->cpi->common.seq_params->sb_size == BLOCK_64X64; |
940 | | // A parallel frame encode has at least 1/4th the |
941 | | // theoretical limit of max enc workers in default case. For resolutions |
942 | | // larger than 480p, if SB size is 64x64, optimal performance is obtained with |
943 | | // limit of 1/8. |
944 | | int index = (!is_480p_or_lesser && is_sb_64) ? 1 : 0; |
945 | | int workers_per_frame = |
946 | | AOMMAX(1, (max_num_enc_workers + rounding_factor[index]) / |
947 | | scaling_factor[index]); |
948 | | int max_threads = oxcf->max_threads; |
949 | | int num_fp_contexts = max_threads / workers_per_frame; |
950 | | // Based on empirical results, FPMT gains with multi-tile are significant when |
951 | | // more parallel frames are available. Use FPMT with multi-tile encode only |
952 | | // when sufficient threads are available for parallel encode of |
953 | | // MAX_PARALLEL_FRAMES frames. |
954 | | if (oxcf->tile_cfg.tile_columns > 0 || oxcf->tile_cfg.tile_rows > 0) { |
955 | | if (num_fp_contexts < MAX_PARALLEL_FRAMES) num_fp_contexts = 1; |
956 | | } |
957 | | |
958 | | num_fp_contexts = AOMMAX(1, AOMMIN(num_fp_contexts, MAX_PARALLEL_FRAMES)); |
959 | | // Limit recalculated num_fp_contexts to ppi->num_fp_contexts. |
960 | | num_fp_contexts = (ppi->num_fp_contexts == 1) |
961 | | ? num_fp_contexts |
962 | | : AOMMIN(num_fp_contexts, ppi->num_fp_contexts); |
963 | | if (num_fp_contexts > 1) { |
964 | | ppi->p_mt_info.num_mod_workers[MOD_FRAME_ENC] = |
965 | | AOMMIN(max_num_enc_workers * num_fp_contexts, oxcf->max_threads); |
966 | | } |
967 | | return num_fp_contexts; |
968 | | } |
969 | | |
970 | | // Computes the number of workers to process each of the parallel frames. |
971 | | static AOM_INLINE int compute_num_workers_per_frame( |
972 | | const int num_workers, const int parallel_frame_count) { |
973 | | // Number of level 2 workers per frame context (floor division). |
974 | | int workers_per_frame = (num_workers / parallel_frame_count); |
975 | | return workers_per_frame; |
976 | | } |
977 | | |
978 | | // Prepare level 1 workers. This function is only called for |
979 | | // parallel_frame_count > 1. This function populates the mt_info structure of |
980 | | // frame level contexts appropriately by dividing the total number of available |
981 | | // workers amongst the frames as level 2 workers. It also populates the hook and |
982 | | // data members of level 1 workers. |
983 | | static AOM_INLINE void prepare_fpmt_workers(AV1_PRIMARY *ppi, |
984 | | AV1_COMP_DATA *first_cpi_data, |
985 | | AVxWorkerHook hook, |
986 | | int parallel_frame_count) { |
987 | | assert(parallel_frame_count <= ppi->num_fp_contexts && |
988 | | parallel_frame_count > 1); |
989 | | |
990 | | PrimaryMultiThreadInfo *const p_mt_info = &ppi->p_mt_info; |
991 | | int num_workers = p_mt_info->num_workers; |
992 | | |
993 | | int frame_idx = 0; |
994 | | int i = 0; |
995 | | while (i < num_workers) { |
996 | | // Assign level 1 worker |
997 | | AVxWorker *frame_worker = p_mt_info->p_workers[frame_idx] = |
998 | | &p_mt_info->workers[i]; |
999 | | AV1_COMP *cur_cpi = ppi->parallel_cpi[frame_idx]; |
1000 | | MultiThreadInfo *mt_info = &cur_cpi->mt_info; |
1001 | | AV1_COMMON *const cm = &cur_cpi->common; |
1002 | | const int num_planes = av1_num_planes(cm); |
1003 | | |
1004 | | // Assign start of level 2 worker pool |
1005 | | mt_info->workers = &p_mt_info->workers[i]; |
1006 | | mt_info->tile_thr_data = &p_mt_info->tile_thr_data[i]; |
1007 | | // Assign number of workers for each frame in the parallel encode set. |
1008 | | mt_info->num_workers = compute_num_workers_per_frame( |
1009 | | num_workers - i, parallel_frame_count - frame_idx); |
1010 | | for (int j = MOD_FP; j < NUM_MT_MODULES; j++) { |
1011 | | mt_info->num_mod_workers[j] = |
1012 | | AOMMIN(mt_info->num_workers, ppi->p_mt_info.num_mod_workers[j]); |
1013 | | } |
1014 | | if (ppi->p_mt_info.cdef_worker != NULL) { |
1015 | | mt_info->cdef_worker = &ppi->p_mt_info.cdef_worker[i]; |
1016 | | |
1017 | | // Back up the original cdef_worker pointers. |
1018 | | mt_info->restore_state_buf.cdef_srcbuf = mt_info->cdef_worker->srcbuf; |
1019 | | for (int plane = 0; plane < num_planes; plane++) |
1020 | | mt_info->restore_state_buf.cdef_colbuf[plane] = |
1021 | | mt_info->cdef_worker->colbuf[plane]; |
1022 | | } |
1023 | | #if !CONFIG_REALTIME_ONLY |
1024 | | if (is_restoration_used(cm)) { |
1025 | | // Back up the original LR buffers before update. |
1026 | | int idx = i + mt_info->num_workers - 1; |
1027 | | mt_info->restore_state_buf.rst_tmpbuf = |
1028 | | mt_info->lr_row_sync.lrworkerdata[idx].rst_tmpbuf; |
1029 | | mt_info->restore_state_buf.rlbs = |
1030 | | mt_info->lr_row_sync.lrworkerdata[idx].rlbs; |
1031 | | |
1032 | | // Update LR buffers. |
1033 | | mt_info->lr_row_sync.lrworkerdata[idx].rst_tmpbuf = cm->rst_tmpbuf; |
1034 | | mt_info->lr_row_sync.lrworkerdata[idx].rlbs = cm->rlbs; |
1035 | | } |
1036 | | #endif |
1037 | | |
1038 | | // At this stage, the thread specific CDEF buffers for the current frame's |
1039 | | // 'common' and 'cdef_sync' only need to be allocated. 'cdef_worker' has |
1040 | | // already been allocated across parallel frames. |
1041 | | av1_alloc_cdef_buffers(cm, &p_mt_info->cdef_worker, &mt_info->cdef_sync, |
1042 | | p_mt_info->num_workers, 0); |
1043 | | |
1044 | | frame_worker->hook = hook; |
1045 | | frame_worker->data1 = cur_cpi; |
1046 | | frame_worker->data2 = (frame_idx == 0) |
1047 | | ? first_cpi_data |
1048 | | : &ppi->parallel_frames_data[frame_idx - 1]; |
1049 | | frame_idx++; |
1050 | | i += mt_info->num_workers; |
1051 | | } |
1052 | | p_mt_info->p_num_workers = parallel_frame_count; |
1053 | | } |
1054 | | |
1055 | | // Launch level 1 workers to perform frame parallel encode. |
1056 | | static AOM_INLINE void launch_fpmt_workers(AV1_PRIMARY *ppi) { |
1057 | | const AVxWorkerInterface *const winterface = aom_get_worker_interface(); |
1058 | | int num_workers = ppi->p_mt_info.p_num_workers; |
1059 | | |
1060 | | for (int i = num_workers - 1; i >= 0; i--) { |
1061 | | AVxWorker *const worker = ppi->p_mt_info.p_workers[i]; |
1062 | | if (i == 0) |
1063 | | winterface->execute(worker); |
1064 | | else |
1065 | | winterface->launch(worker); |
1066 | | } |
1067 | | } |
1068 | | |
1069 | | // Synchronize level 1 workers. |
1070 | | static AOM_INLINE void sync_fpmt_workers(AV1_PRIMARY *ppi) { |
1071 | | const AVxWorkerInterface *const winterface = aom_get_worker_interface(); |
1072 | | int num_workers = ppi->p_mt_info.p_num_workers; |
1073 | | int had_error = 0; |
1074 | | // Points to error in the earliest display order frame in the parallel set. |
1075 | | const struct aom_internal_error_info *error; |
1076 | | |
1077 | | // Encoding ends. |
1078 | | for (int i = num_workers - 1; i >= 0; i--) { |
1079 | | AVxWorker *const worker = ppi->p_mt_info.p_workers[i]; |
1080 | | if (!winterface->sync(worker)) { |
1081 | | had_error = 1; |
1082 | | error = ((AV1_COMP *)worker->data1)->common.error; |
1083 | | } |
1084 | | } |
1085 | | |
1086 | | if (had_error) |
1087 | | aom_internal_error(&ppi->error, error->error_code, "%s", error->detail); |
1088 | | } |
1089 | | |
1090 | | // Restore worker states after parallel encode. |
1091 | | static AOM_INLINE void restore_workers_after_fpmt(AV1_PRIMARY *ppi, |
1092 | | int parallel_frame_count) { |
1093 | | assert(parallel_frame_count <= ppi->num_fp_contexts && |
1094 | | parallel_frame_count > 1); |
1095 | | (void)parallel_frame_count; |
1096 | | |
1097 | | PrimaryMultiThreadInfo *const p_mt_info = &ppi->p_mt_info; |
1098 | | int num_workers = p_mt_info->num_workers; |
1099 | | |
1100 | | int frame_idx = 0; |
1101 | | int i = 0; |
1102 | | while (i < num_workers) { |
1103 | | AV1_COMP *cur_cpi = ppi->parallel_cpi[frame_idx]; |
1104 | | MultiThreadInfo *mt_info = &cur_cpi->mt_info; |
1105 | | const AV1_COMMON *const cm = &cur_cpi->common; |
1106 | | const int num_planes = av1_num_planes(cm); |
1107 | | |
1108 | | // Restore the original cdef_worker pointers. |
1109 | | if (ppi->p_mt_info.cdef_worker != NULL) { |
1110 | | mt_info->cdef_worker->srcbuf = mt_info->restore_state_buf.cdef_srcbuf; |
1111 | | for (int plane = 0; plane < num_planes; plane++) |
1112 | | mt_info->cdef_worker->colbuf[plane] = |
1113 | | mt_info->restore_state_buf.cdef_colbuf[plane]; |
1114 | | } |
1115 | | #if !CONFIG_REALTIME_ONLY |
1116 | | if (is_restoration_used(cm)) { |
1117 | | // Restore the original LR buffers. |
1118 | | int idx = i + mt_info->num_workers - 1; |
1119 | | mt_info->lr_row_sync.lrworkerdata[idx].rst_tmpbuf = |
1120 | | mt_info->restore_state_buf.rst_tmpbuf; |
1121 | | mt_info->lr_row_sync.lrworkerdata[idx].rlbs = |
1122 | | mt_info->restore_state_buf.rlbs; |
1123 | | } |
1124 | | #endif |
1125 | | |
1126 | | frame_idx++; |
1127 | | i += mt_info->num_workers; |
1128 | | } |
1129 | | } |
1130 | | |
1131 | | static int get_compressed_data_hook(void *arg1, void *arg2) { |
1132 | | AV1_COMP *cpi = (AV1_COMP *)arg1; |
1133 | | AV1_COMP_DATA *cpi_data = (AV1_COMP_DATA *)arg2; |
1134 | | int status = av1_get_compressed_data(cpi, cpi_data); |
1135 | | |
1136 | | // AOM_CODEC_OK(0) means no error. |
1137 | | return !status; |
1138 | | } |
1139 | | |
1140 | | // This function encodes the raw frame data for each frame in parallel encode |
1141 | | // set, and outputs the frame bit stream to the designated buffers. |
1142 | | int av1_compress_parallel_frames(AV1_PRIMARY *const ppi, |
1143 | | AV1_COMP_DATA *const first_cpi_data) { |
1144 | | // Bitmask for the frame buffers referenced by cpi->scaled_ref_buf |
1145 | | // corresponding to frames in the current parallel encode set. |
1146 | | int ref_buffers_used_map = 0; |
1147 | | int frames_in_parallel_set = av1_init_parallel_frame_context( |
1148 | | first_cpi_data, ppi, &ref_buffers_used_map); |
1149 | | prepare_fpmt_workers(ppi, first_cpi_data, get_compressed_data_hook, |
1150 | | frames_in_parallel_set); |
1151 | | launch_fpmt_workers(ppi); |
1152 | | sync_fpmt_workers(ppi); |
1153 | | restore_workers_after_fpmt(ppi, frames_in_parallel_set); |
1154 | | |
1155 | | // Release cpi->scaled_ref_buf corresponding to frames in the current parallel |
1156 | | // encode set. |
1157 | | for (int i = 0; i < frames_in_parallel_set; ++i) { |
1158 | | av1_release_scaled_references_fpmt(ppi->parallel_cpi[i]); |
1159 | | } |
1160 | | av1_decrement_ref_counts_fpmt(ppi->cpi->common.buffer_pool, |
1161 | | ref_buffers_used_map); |
1162 | | return AOM_CODEC_OK; |
1163 | | } |
1164 | | #endif // CONFIG_FRAME_PARALLEL_ENCODE |
1165 | | |
1166 | | static AOM_INLINE void launch_workers(MultiThreadInfo *const mt_info, |
1167 | 0 | int num_workers) { |
1168 | 0 | const AVxWorkerInterface *const winterface = aom_get_worker_interface(); |
1169 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
1170 | 0 | AVxWorker *const worker = &mt_info->workers[i]; |
1171 | 0 | if (i == 0) |
1172 | 0 | winterface->execute(worker); |
1173 | 0 | else |
1174 | 0 | winterface->launch(worker); |
1175 | 0 | } |
1176 | 0 | } |
1177 | | |
1178 | | static AOM_INLINE void sync_enc_workers(MultiThreadInfo *const mt_info, |
1179 | 0 | AV1_COMMON *const cm, int num_workers) { |
1180 | 0 | const AVxWorkerInterface *const winterface = aom_get_worker_interface(); |
1181 | 0 | int had_error = 0; |
1182 | | |
1183 | | // Encoding ends. |
1184 | 0 | for (int i = num_workers - 1; i > 0; i--) { |
1185 | 0 | AVxWorker *const worker = &mt_info->workers[i]; |
1186 | 0 | had_error |= !winterface->sync(worker); |
1187 | 0 | } |
1188 | |
|
1189 | 0 | if (had_error) |
1190 | 0 | aom_internal_error(cm->error, AOM_CODEC_ERROR, |
1191 | 0 | "Failed to encode tile data"); |
1192 | 0 | } |
1193 | | |
1194 | | static AOM_INLINE void accumulate_counters_enc_workers(AV1_COMP *cpi, |
1195 | 0 | int num_workers) { |
1196 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
1197 | 0 | AVxWorker *const worker = &cpi->mt_info.workers[i]; |
1198 | 0 | EncWorkerData *const thread_data = (EncWorkerData *)worker->data1; |
1199 | 0 | cpi->intrabc_used |= thread_data->td->intrabc_used; |
1200 | 0 | cpi->deltaq_used |= thread_data->td->deltaq_used; |
1201 | | // Accumulate cyclic refresh params. |
1202 | 0 | if (cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ && |
1203 | 0 | !frame_is_intra_only(&cpi->common)) |
1204 | 0 | av1_accumulate_cyclic_refresh_counters(cpi->cyclic_refresh, |
1205 | 0 | &thread_data->td->mb); |
1206 | 0 | if (thread_data->td != &cpi->td) { |
1207 | 0 | if (cpi->sf.inter_sf.mv_cost_upd_level != INTERNAL_COST_UPD_OFF) { |
1208 | 0 | aom_free(thread_data->td->mb.mv_costs); |
1209 | 0 | } |
1210 | 0 | if (cpi->sf.intra_sf.dv_cost_upd_level != INTERNAL_COST_UPD_OFF) { |
1211 | 0 | aom_free(thread_data->td->mb.dv_costs); |
1212 | 0 | } |
1213 | 0 | } |
1214 | 0 | av1_dealloc_mb_data(&cpi->common, &thread_data->td->mb); |
1215 | | |
1216 | | // Accumulate counters. |
1217 | 0 | if (i > 0) { |
1218 | 0 | av1_accumulate_frame_counts(&cpi->counts, thread_data->td->counts); |
1219 | 0 | accumulate_rd_opt(&cpi->td, thread_data->td); |
1220 | 0 | cpi->td.mb.txfm_search_info.txb_split_count += |
1221 | 0 | thread_data->td->mb.txfm_search_info.txb_split_count; |
1222 | | #if CONFIG_SPEED_STATS |
1223 | | cpi->td.mb.txfm_search_info.tx_search_count += |
1224 | | thread_data->td->mb.txfm_search_info.tx_search_count; |
1225 | | #endif // CONFIG_SPEED_STATS |
1226 | 0 | } |
1227 | 0 | } |
1228 | 0 | } |
1229 | | |
1230 | | static AOM_INLINE void prepare_enc_workers(AV1_COMP *cpi, AVxWorkerHook hook, |
1231 | 0 | int num_workers) { |
1232 | 0 | MultiThreadInfo *const mt_info = &cpi->mt_info; |
1233 | 0 | AV1_COMMON *const cm = &cpi->common; |
1234 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
1235 | 0 | AVxWorker *const worker = &mt_info->workers[i]; |
1236 | 0 | EncWorkerData *const thread_data = &mt_info->tile_thr_data[i]; |
1237 | |
|
1238 | 0 | worker->hook = hook; |
1239 | 0 | worker->data1 = thread_data; |
1240 | 0 | worker->data2 = NULL; |
1241 | |
|
1242 | 0 | thread_data->thread_id = i; |
1243 | | // Set the starting tile for each thread. |
1244 | 0 | thread_data->start = i; |
1245 | |
|
1246 | 0 | thread_data->cpi = cpi; |
1247 | 0 | if (i == 0) { |
1248 | 0 | thread_data->td = &cpi->td; |
1249 | 0 | #if !CONFIG_FRAME_PARALLEL_ENCODE |
1250 | 0 | } |
1251 | | #else |
1252 | | } else { |
1253 | | thread_data->td = thread_data->original_td; |
1254 | | } |
1255 | | #endif // CONFIG_FRAME_PARALLEL_ENCODE |
1256 | |
|
1257 | 0 | thread_data->td->intrabc_used = 0; |
1258 | 0 | thread_data->td->deltaq_used = 0; |
1259 | 0 | thread_data->td->abs_sum_level = 0; |
1260 | | |
1261 | | // Before encoding a frame, copy the thread data from cpi. |
1262 | 0 | if (thread_data->td != &cpi->td) { |
1263 | 0 | thread_data->td->mb = cpi->td.mb; |
1264 | 0 | thread_data->td->rd_counts = cpi->td.rd_counts; |
1265 | 0 | thread_data->td->mb.obmc_buffer = thread_data->td->obmc_buffer; |
1266 | |
|
1267 | 0 | for (int x = 0; x < 2; x++) { |
1268 | 0 | for (int y = 0; y < 2; y++) { |
1269 | 0 | memcpy(thread_data->td->hash_value_buffer[x][y], |
1270 | 0 | cpi->td.mb.intrabc_hash_info.hash_value_buffer[x][y], |
1271 | 0 | AOM_BUFFER_SIZE_FOR_BLOCK_HASH * |
1272 | 0 | sizeof(*thread_data->td->hash_value_buffer[0][0])); |
1273 | 0 | thread_data->td->mb.intrabc_hash_info.hash_value_buffer[x][y] = |
1274 | 0 | thread_data->td->hash_value_buffer[x][y]; |
1275 | 0 | } |
1276 | 0 | } |
1277 | 0 | if (cpi->sf.inter_sf.mv_cost_upd_level != INTERNAL_COST_UPD_OFF) { |
1278 | 0 | CHECK_MEM_ERROR(cm, thread_data->td->mb.mv_costs, |
1279 | 0 | (MvCosts *)aom_malloc(sizeof(MvCosts))); |
1280 | 0 | memcpy(thread_data->td->mb.mv_costs, cpi->td.mb.mv_costs, |
1281 | 0 | sizeof(MvCosts)); |
1282 | 0 | } |
1283 | 0 | if (cpi->sf.intra_sf.dv_cost_upd_level != INTERNAL_COST_UPD_OFF) { |
1284 | 0 | CHECK_MEM_ERROR(cm, thread_data->td->mb.dv_costs, |
1285 | 0 | (IntraBCMVCosts *)aom_malloc(sizeof(IntraBCMVCosts))); |
1286 | 0 | memcpy(thread_data->td->mb.dv_costs, cpi->td.mb.dv_costs, |
1287 | 0 | sizeof(IntraBCMVCosts)); |
1288 | 0 | } |
1289 | 0 | } |
1290 | 0 | av1_alloc_mb_data(cm, &thread_data->td->mb, |
1291 | 0 | cpi->sf.rt_sf.use_nonrd_pick_mode, |
1292 | 0 | cpi->sf.rd_sf.use_mb_rd_hash); |
1293 | | |
1294 | | // Reset cyclic refresh counters. |
1295 | 0 | av1_init_cyclic_refresh_counters(&thread_data->td->mb); |
1296 | |
|
1297 | 0 | if (thread_data->td->counts != &cpi->counts) { |
1298 | 0 | memcpy(thread_data->td->counts, &cpi->counts, sizeof(cpi->counts)); |
1299 | 0 | } |
1300 | |
|
1301 | 0 | if (i > 0) { |
1302 | 0 | thread_data->td->mb.palette_buffer = thread_data->td->palette_buffer; |
1303 | 0 | thread_data->td->mb.comp_rd_buffer = thread_data->td->comp_rd_buffer; |
1304 | 0 | thread_data->td->mb.tmp_conv_dst = thread_data->td->tmp_conv_dst; |
1305 | 0 | for (int j = 0; j < 2; ++j) { |
1306 | 0 | thread_data->td->mb.tmp_pred_bufs[j] = |
1307 | 0 | thread_data->td->tmp_pred_bufs[j]; |
1308 | 0 | } |
1309 | 0 | thread_data->td->mb.pixel_gradient_info = |
1310 | 0 | thread_data->td->pixel_gradient_info; |
1311 | |
|
1312 | 0 | thread_data->td->mb.e_mbd.tmp_conv_dst = thread_data->td->mb.tmp_conv_dst; |
1313 | 0 | for (int j = 0; j < 2; ++j) { |
1314 | 0 | thread_data->td->mb.e_mbd.tmp_obmc_bufs[j] = |
1315 | 0 | thread_data->td->mb.tmp_pred_bufs[j]; |
1316 | 0 | } |
1317 | 0 | } |
1318 | 0 | } |
1319 | 0 | } |
1320 | | |
1321 | | #if !CONFIG_REALTIME_ONLY |
1322 | | static AOM_INLINE void fp_prepare_enc_workers(AV1_COMP *cpi, AVxWorkerHook hook, |
1323 | 0 | int num_workers) { |
1324 | 0 | AV1_COMMON *const cm = &cpi->common; |
1325 | 0 | MultiThreadInfo *const mt_info = &cpi->mt_info; |
1326 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
1327 | 0 | AVxWorker *const worker = &mt_info->workers[i]; |
1328 | 0 | EncWorkerData *const thread_data = &mt_info->tile_thr_data[i]; |
1329 | |
|
1330 | 0 | worker->hook = hook; |
1331 | 0 | worker->data1 = thread_data; |
1332 | 0 | worker->data2 = NULL; |
1333 | |
|
1334 | 0 | thread_data->thread_id = i; |
1335 | | // Set the starting tile for each thread. |
1336 | 0 | thread_data->start = i; |
1337 | |
|
1338 | 0 | thread_data->cpi = cpi; |
1339 | 0 | if (i == 0) { |
1340 | 0 | thread_data->td = &cpi->td; |
1341 | 0 | #if !CONFIG_FRAME_PARALLEL_ENCODE |
1342 | 0 | } |
1343 | | #else |
1344 | | } else { |
1345 | | thread_data->td = thread_data->original_td; |
1346 | | } |
1347 | | #endif // CONFIG_FRAME_PARALLEL_ENCODE |
1348 | | |
1349 | | // Before encoding a frame, copy the thread data from cpi. |
1350 | 0 | if (thread_data->td != &cpi->td) { |
1351 | 0 | thread_data->td->mb = cpi->td.mb; |
1352 | 0 | if (cpi->sf.inter_sf.mv_cost_upd_level != INTERNAL_COST_UPD_OFF) { |
1353 | 0 | CHECK_MEM_ERROR(cm, thread_data->td->mb.mv_costs, |
1354 | 0 | (MvCosts *)aom_malloc(sizeof(MvCosts))); |
1355 | 0 | memcpy(thread_data->td->mb.mv_costs, cpi->td.mb.mv_costs, |
1356 | 0 | sizeof(MvCosts)); |
1357 | 0 | } |
1358 | 0 | if (cpi->sf.intra_sf.dv_cost_upd_level != INTERNAL_COST_UPD_OFF) { |
1359 | 0 | CHECK_MEM_ERROR(cm, thread_data->td->mb.dv_costs, |
1360 | 0 | (IntraBCMVCosts *)aom_malloc(sizeof(IntraBCMVCosts))); |
1361 | 0 | memcpy(thread_data->td->mb.dv_costs, cpi->td.mb.dv_costs, |
1362 | 0 | sizeof(IntraBCMVCosts)); |
1363 | 0 | } |
1364 | 0 | } |
1365 | |
|
1366 | 0 | av1_alloc_mb_data(cm, &thread_data->td->mb, |
1367 | 0 | cpi->sf.rt_sf.use_nonrd_pick_mode, |
1368 | 0 | cpi->sf.rd_sf.use_mb_rd_hash); |
1369 | 0 | } |
1370 | 0 | } |
1371 | | #endif |
1372 | | |
1373 | | // Computes the number of workers for row multi-threading of encoding stage |
1374 | | static AOM_INLINE int compute_num_enc_row_mt_workers(AV1_COMMON *const cm, |
1375 | 0 | int max_threads) { |
1376 | 0 | TileInfo tile_info; |
1377 | 0 | const int tile_cols = cm->tiles.cols; |
1378 | 0 | const int tile_rows = cm->tiles.rows; |
1379 | 0 | int total_num_threads_row_mt = 0; |
1380 | 0 | for (int row = 0; row < tile_rows; row++) { |
1381 | 0 | for (int col = 0; col < tile_cols; col++) { |
1382 | 0 | av1_tile_init(&tile_info, cm, row, col); |
1383 | 0 | const int num_sb_rows_in_tile = av1_get_sb_rows_in_tile(cm, tile_info); |
1384 | 0 | const int num_sb_cols_in_tile = av1_get_sb_cols_in_tile(cm, tile_info); |
1385 | 0 | total_num_threads_row_mt += |
1386 | 0 | AOMMIN((num_sb_cols_in_tile + 1) >> 1, num_sb_rows_in_tile); |
1387 | 0 | } |
1388 | 0 | } |
1389 | 0 | return AOMMIN(max_threads, total_num_threads_row_mt); |
1390 | 0 | } |
1391 | | |
1392 | | // Computes the number of workers for tile multi-threading of encoding stage |
1393 | | static AOM_INLINE int compute_num_enc_tile_mt_workers(AV1_COMMON *const cm, |
1394 | 0 | int max_threads) { |
1395 | 0 | const int tile_cols = cm->tiles.cols; |
1396 | 0 | const int tile_rows = cm->tiles.rows; |
1397 | 0 | return AOMMIN(max_threads, tile_cols * tile_rows); |
1398 | 0 | } |
1399 | | |
1400 | | // Find max worker of all MT stages |
1401 | 0 | int av1_get_max_num_workers(AV1_COMP *cpi) { |
1402 | 0 | int max_num_workers = 0; |
1403 | 0 | for (int i = MOD_FP; i < NUM_MT_MODULES; i++) |
1404 | 0 | max_num_workers = |
1405 | 0 | AOMMAX(cpi->ppi->p_mt_info.num_mod_workers[i], max_num_workers); |
1406 | 0 | assert(max_num_workers >= 1); |
1407 | 0 | return AOMMIN(max_num_workers, cpi->oxcf.max_threads); |
1408 | 0 | } |
1409 | | |
1410 | | // Computes the number of workers for encoding stage (row/tile multi-threading) |
1411 | 0 | int av1_compute_num_enc_workers(AV1_COMP *cpi, int max_workers) { |
1412 | 0 | if (max_workers <= 1) return 1; |
1413 | 0 | if (cpi->oxcf.row_mt) |
1414 | 0 | return compute_num_enc_row_mt_workers(&cpi->common, max_workers); |
1415 | 0 | else |
1416 | 0 | return compute_num_enc_tile_mt_workers(&cpi->common, max_workers); |
1417 | 0 | } |
1418 | | |
1419 | 0 | void av1_encode_tiles_mt(AV1_COMP *cpi) { |
1420 | 0 | AV1_COMMON *const cm = &cpi->common; |
1421 | 0 | MultiThreadInfo *const mt_info = &cpi->mt_info; |
1422 | 0 | const int tile_cols = cm->tiles.cols; |
1423 | 0 | const int tile_rows = cm->tiles.rows; |
1424 | 0 | int num_workers = mt_info->num_mod_workers[MOD_ENC]; |
1425 | |
|
1426 | 0 | assert(IMPLIES(cpi->tile_data == NULL, |
1427 | 0 | cpi->allocated_tiles < tile_cols * tile_rows)); |
1428 | 0 | if (cpi->allocated_tiles < tile_cols * tile_rows) av1_alloc_tile_data(cpi); |
1429 | |
|
1430 | 0 | av1_init_tile_data(cpi); |
1431 | 0 | num_workers = AOMMIN(num_workers, mt_info->num_workers); |
1432 | |
|
1433 | 0 | prepare_enc_workers(cpi, enc_worker_hook, num_workers); |
1434 | 0 | launch_workers(&cpi->mt_info, num_workers); |
1435 | 0 | sync_enc_workers(&cpi->mt_info, cm, num_workers); |
1436 | 0 | accumulate_counters_enc_workers(cpi, num_workers); |
1437 | 0 | } |
1438 | | |
1439 | | // Accumulate frame counts. FRAME_COUNTS consist solely of 'unsigned int' |
1440 | | // members, so we treat it as an array, and sum over the whole length. |
1441 | | void av1_accumulate_frame_counts(FRAME_COUNTS *acc_counts, |
1442 | 0 | const FRAME_COUNTS *counts) { |
1443 | 0 | unsigned int *const acc = (unsigned int *)acc_counts; |
1444 | 0 | const unsigned int *const cnt = (const unsigned int *)counts; |
1445 | |
|
1446 | 0 | const unsigned int n_counts = sizeof(FRAME_COUNTS) / sizeof(unsigned int); |
1447 | |
|
1448 | 0 | for (unsigned int i = 0; i < n_counts; i++) acc[i] += cnt[i]; |
1449 | 0 | } |
1450 | | |
1451 | | // Computes the maximum number of sb_rows for row multi-threading of encoding |
1452 | | // stage |
1453 | | static AOM_INLINE void compute_max_sb_rows_cols(AV1_COMP *cpi, int *max_sb_rows, |
1454 | 0 | int *max_sb_cols) { |
1455 | 0 | AV1_COMMON *const cm = &cpi->common; |
1456 | 0 | const int tile_cols = cm->tiles.cols; |
1457 | 0 | const int tile_rows = cm->tiles.rows; |
1458 | 0 | for (int row = 0; row < tile_rows; row++) { |
1459 | 0 | for (int col = 0; col < tile_cols; col++) { |
1460 | 0 | const int tile_index = row * cm->tiles.cols + col; |
1461 | 0 | TileInfo tile_info = cpi->tile_data[tile_index].tile_info; |
1462 | 0 | const int num_sb_rows_in_tile = av1_get_sb_rows_in_tile(cm, tile_info); |
1463 | 0 | const int num_sb_cols_in_tile = av1_get_sb_cols_in_tile(cm, tile_info); |
1464 | 0 | *max_sb_rows = AOMMAX(*max_sb_rows, num_sb_rows_in_tile); |
1465 | 0 | *max_sb_cols = AOMMAX(*max_sb_cols, num_sb_cols_in_tile); |
1466 | 0 | } |
1467 | 0 | } |
1468 | 0 | } |
1469 | | |
1470 | | #if !CONFIG_REALTIME_ONLY |
1471 | | // Computes the number of workers for firstpass stage (row/tile multi-threading) |
1472 | 0 | int av1_fp_compute_num_enc_workers(AV1_COMP *cpi) { |
1473 | 0 | AV1_COMMON *cm = &cpi->common; |
1474 | 0 | const int tile_cols = cm->tiles.cols; |
1475 | 0 | const int tile_rows = cm->tiles.rows; |
1476 | 0 | int total_num_threads_row_mt = 0; |
1477 | 0 | TileInfo tile_info; |
1478 | |
|
1479 | 0 | if (cpi->oxcf.max_threads <= 1) return 1; |
1480 | | |
1481 | 0 | for (int row = 0; row < tile_rows; row++) { |
1482 | 0 | for (int col = 0; col < tile_cols; col++) { |
1483 | 0 | av1_tile_init(&tile_info, cm, row, col); |
1484 | 0 | const int num_mb_rows_in_tile = |
1485 | 0 | av1_get_unit_rows_in_tile(tile_info, cpi->fp_block_size); |
1486 | 0 | const int num_mb_cols_in_tile = |
1487 | 0 | av1_get_unit_cols_in_tile(tile_info, cpi->fp_block_size); |
1488 | 0 | total_num_threads_row_mt += |
1489 | 0 | AOMMIN((num_mb_cols_in_tile + 1) >> 1, num_mb_rows_in_tile); |
1490 | 0 | } |
1491 | 0 | } |
1492 | 0 | return AOMMIN(cpi->oxcf.max_threads, total_num_threads_row_mt); |
1493 | 0 | } |
1494 | | |
1495 | | // Computes the maximum number of mb_rows for row multi-threading of firstpass |
1496 | | // stage |
1497 | | static AOM_INLINE int fp_compute_max_mb_rows(const AV1_COMMON *const cm, |
1498 | | const TileDataEnc *const tile_data, |
1499 | 0 | const BLOCK_SIZE fp_block_size) { |
1500 | 0 | const int tile_cols = cm->tiles.cols; |
1501 | 0 | const int tile_rows = cm->tiles.rows; |
1502 | 0 | int max_mb_rows = 0; |
1503 | 0 | for (int row = 0; row < tile_rows; row++) { |
1504 | 0 | for (int col = 0; col < tile_cols; col++) { |
1505 | 0 | const int tile_index = row * cm->tiles.cols + col; |
1506 | 0 | TileInfo tile_info = tile_data[tile_index].tile_info; |
1507 | 0 | const int num_mb_rows_in_tile = |
1508 | 0 | av1_get_unit_rows_in_tile(tile_info, fp_block_size); |
1509 | 0 | max_mb_rows = AOMMAX(max_mb_rows, num_mb_rows_in_tile); |
1510 | 0 | } |
1511 | 0 | } |
1512 | 0 | return max_mb_rows; |
1513 | 0 | } |
1514 | | #endif |
1515 | | |
1516 | 0 | void av1_encode_tiles_row_mt(AV1_COMP *cpi) { |
1517 | 0 | AV1_COMMON *const cm = &cpi->common; |
1518 | 0 | MultiThreadInfo *const mt_info = &cpi->mt_info; |
1519 | 0 | AV1EncRowMultiThreadInfo *const enc_row_mt = &mt_info->enc_row_mt; |
1520 | 0 | const int tile_cols = cm->tiles.cols; |
1521 | 0 | const int tile_rows = cm->tiles.rows; |
1522 | 0 | int *thread_id_to_tile_id = enc_row_mt->thread_id_to_tile_id; |
1523 | 0 | int max_sb_rows = 0, max_sb_cols = 0; |
1524 | 0 | int num_workers = mt_info->num_mod_workers[MOD_ENC]; |
1525 | |
|
1526 | 0 | assert(IMPLIES(cpi->tile_data == NULL, |
1527 | 0 | cpi->allocated_tiles < tile_cols * tile_rows)); |
1528 | 0 | if (cpi->allocated_tiles < tile_cols * tile_rows) { |
1529 | 0 | av1_row_mt_mem_dealloc(cpi); |
1530 | 0 | av1_alloc_tile_data(cpi); |
1531 | 0 | } |
1532 | |
|
1533 | 0 | av1_init_tile_data(cpi); |
1534 | |
|
1535 | 0 | compute_max_sb_rows_cols(cpi, &max_sb_rows, &max_sb_cols); |
1536 | |
|
1537 | 0 | if (enc_row_mt->allocated_tile_cols != tile_cols || |
1538 | 0 | enc_row_mt->allocated_tile_rows != tile_rows || |
1539 | 0 | enc_row_mt->allocated_rows != max_sb_rows || |
1540 | 0 | enc_row_mt->allocated_cols != (max_sb_cols - 1)) { |
1541 | 0 | av1_row_mt_mem_dealloc(cpi); |
1542 | 0 | row_mt_mem_alloc(cpi, max_sb_rows, max_sb_cols, |
1543 | 0 | cpi->oxcf.algo_cfg.cdf_update_mode); |
1544 | 0 | } |
1545 | |
|
1546 | 0 | memset(thread_id_to_tile_id, -1, |
1547 | 0 | sizeof(*thread_id_to_tile_id) * MAX_NUM_THREADS); |
1548 | |
|
1549 | 0 | for (int tile_row = 0; tile_row < tile_rows; tile_row++) { |
1550 | 0 | for (int tile_col = 0; tile_col < tile_cols; tile_col++) { |
1551 | 0 | int tile_index = tile_row * tile_cols + tile_col; |
1552 | 0 | TileDataEnc *const this_tile = &cpi->tile_data[tile_index]; |
1553 | 0 | AV1EncRowMultiThreadSync *const row_mt_sync = &this_tile->row_mt_sync; |
1554 | | |
1555 | | // Initialize num_finished_cols to -1 for all rows. |
1556 | 0 | memset(row_mt_sync->num_finished_cols, -1, |
1557 | 0 | sizeof(*row_mt_sync->num_finished_cols) * max_sb_rows); |
1558 | 0 | row_mt_sync->next_mi_row = this_tile->tile_info.mi_row_start; |
1559 | 0 | row_mt_sync->num_threads_working = 0; |
1560 | |
|
1561 | 0 | av1_inter_mode_data_init(this_tile); |
1562 | 0 | av1_zero_above_context(cm, &cpi->td.mb.e_mbd, |
1563 | 0 | this_tile->tile_info.mi_col_start, |
1564 | 0 | this_tile->tile_info.mi_col_end, tile_row); |
1565 | 0 | } |
1566 | 0 | } |
1567 | |
|
1568 | 0 | num_workers = AOMMIN(num_workers, mt_info->num_workers); |
1569 | |
|
1570 | 0 | assign_tile_to_thread(thread_id_to_tile_id, tile_cols * tile_rows, |
1571 | 0 | num_workers); |
1572 | 0 | prepare_enc_workers(cpi, enc_row_mt_worker_hook, num_workers); |
1573 | 0 | launch_workers(&cpi->mt_info, num_workers); |
1574 | 0 | sync_enc_workers(&cpi->mt_info, cm, num_workers); |
1575 | 0 | if (cm->delta_q_info.delta_lf_present_flag) update_delta_lf_for_row_mt(cpi); |
1576 | 0 | accumulate_counters_enc_workers(cpi, num_workers); |
1577 | 0 | } |
1578 | | |
1579 | | #if !CONFIG_REALTIME_ONLY |
1580 | 0 | void av1_fp_encode_tiles_row_mt(AV1_COMP *cpi) { |
1581 | 0 | AV1_COMMON *const cm = &cpi->common; |
1582 | 0 | MultiThreadInfo *const mt_info = &cpi->mt_info; |
1583 | 0 | AV1EncRowMultiThreadInfo *const enc_row_mt = &mt_info->enc_row_mt; |
1584 | 0 | const int tile_cols = cm->tiles.cols; |
1585 | 0 | const int tile_rows = cm->tiles.rows; |
1586 | 0 | int *thread_id_to_tile_id = enc_row_mt->thread_id_to_tile_id; |
1587 | 0 | int num_workers = 0; |
1588 | 0 | int max_mb_rows = 0; |
1589 | |
|
1590 | 0 | assert(IMPLIES(cpi->tile_data == NULL, |
1591 | 0 | cpi->allocated_tiles < tile_cols * tile_rows)); |
1592 | 0 | if (cpi->allocated_tiles < tile_cols * tile_rows) { |
1593 | 0 | av1_row_mt_mem_dealloc(cpi); |
1594 | 0 | av1_alloc_tile_data(cpi); |
1595 | 0 | } |
1596 | |
|
1597 | 0 | av1_init_tile_data(cpi); |
1598 | |
|
1599 | 0 | const BLOCK_SIZE fp_block_size = cpi->fp_block_size; |
1600 | 0 | max_mb_rows = fp_compute_max_mb_rows(cm, cpi->tile_data, fp_block_size); |
1601 | | |
1602 | | // For pass = 1, compute the no. of workers needed. For single-pass encode |
1603 | | // (pass = 0), no. of workers are already computed. |
1604 | 0 | if (mt_info->num_mod_workers[MOD_FP] == 0) |
1605 | 0 | num_workers = av1_fp_compute_num_enc_workers(cpi); |
1606 | 0 | else |
1607 | 0 | num_workers = mt_info->num_mod_workers[MOD_FP]; |
1608 | |
|
1609 | 0 | if (enc_row_mt->allocated_tile_cols != tile_cols || |
1610 | 0 | enc_row_mt->allocated_tile_rows != tile_rows || |
1611 | 0 | enc_row_mt->allocated_rows != max_mb_rows) { |
1612 | 0 | av1_row_mt_mem_dealloc(cpi); |
1613 | 0 | row_mt_mem_alloc(cpi, max_mb_rows, -1, 0); |
1614 | 0 | } |
1615 | |
|
1616 | 0 | memset(thread_id_to_tile_id, -1, |
1617 | 0 | sizeof(*thread_id_to_tile_id) * MAX_NUM_THREADS); |
1618 | |
|
1619 | 0 | for (int tile_row = 0; tile_row < tile_rows; tile_row++) { |
1620 | 0 | for (int tile_col = 0; tile_col < tile_cols; tile_col++) { |
1621 | 0 | int tile_index = tile_row * tile_cols + tile_col; |
1622 | 0 | TileDataEnc *const this_tile = &cpi->tile_data[tile_index]; |
1623 | 0 | AV1EncRowMultiThreadSync *const row_mt_sync = &this_tile->row_mt_sync; |
1624 | | |
1625 | | // Initialize num_finished_cols to -1 for all rows. |
1626 | 0 | memset(row_mt_sync->num_finished_cols, -1, |
1627 | 0 | sizeof(*row_mt_sync->num_finished_cols) * max_mb_rows); |
1628 | 0 | row_mt_sync->next_mi_row = this_tile->tile_info.mi_row_start; |
1629 | 0 | row_mt_sync->num_threads_working = 0; |
1630 | 0 | } |
1631 | 0 | } |
1632 | |
|
1633 | 0 | num_workers = AOMMIN(num_workers, mt_info->num_workers); |
1634 | 0 | assign_tile_to_thread(thread_id_to_tile_id, tile_cols * tile_rows, |
1635 | 0 | num_workers); |
1636 | 0 | fp_prepare_enc_workers(cpi, fp_enc_row_mt_worker_hook, num_workers); |
1637 | 0 | launch_workers(&cpi->mt_info, num_workers); |
1638 | 0 | sync_enc_workers(&cpi->mt_info, cm, num_workers); |
1639 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
1640 | 0 | EncWorkerData *const thread_data = &cpi->mt_info.tile_thr_data[i]; |
1641 | 0 | if (thread_data->td != &cpi->td) { |
1642 | 0 | if (cpi->sf.inter_sf.mv_cost_upd_level != INTERNAL_COST_UPD_OFF) { |
1643 | 0 | aom_free(thread_data->td->mb.mv_costs); |
1644 | 0 | } |
1645 | 0 | if (cpi->sf.intra_sf.dv_cost_upd_level != INTERNAL_COST_UPD_OFF) { |
1646 | 0 | aom_free(thread_data->td->mb.dv_costs); |
1647 | 0 | } |
1648 | 0 | } |
1649 | 0 | av1_dealloc_mb_data(cm, &thread_data->td->mb); |
1650 | 0 | } |
1651 | 0 | } |
1652 | | |
1653 | | void av1_tpl_row_mt_sync_read_dummy(AV1TplRowMultiThreadSync *tpl_mt_sync, |
1654 | 0 | int r, int c) { |
1655 | 0 | (void)tpl_mt_sync; |
1656 | 0 | (void)r; |
1657 | 0 | (void)c; |
1658 | 0 | return; |
1659 | 0 | } |
1660 | | |
1661 | | void av1_tpl_row_mt_sync_write_dummy(AV1TplRowMultiThreadSync *tpl_mt_sync, |
1662 | 0 | int r, int c, int cols) { |
1663 | 0 | (void)tpl_mt_sync; |
1664 | 0 | (void)r; |
1665 | 0 | (void)c; |
1666 | 0 | (void)cols; |
1667 | 0 | return; |
1668 | 0 | } |
1669 | | |
1670 | | void av1_tpl_row_mt_sync_read(AV1TplRowMultiThreadSync *tpl_row_mt_sync, int r, |
1671 | 0 | int c) { |
1672 | 0 | #if CONFIG_MULTITHREAD |
1673 | 0 | int nsync = tpl_row_mt_sync->sync_range; |
1674 | |
|
1675 | 0 | if (r) { |
1676 | 0 | pthread_mutex_t *const mutex = &tpl_row_mt_sync->mutex_[r - 1]; |
1677 | 0 | pthread_mutex_lock(mutex); |
1678 | |
|
1679 | 0 | while (c > tpl_row_mt_sync->num_finished_cols[r - 1] - nsync) |
1680 | 0 | pthread_cond_wait(&tpl_row_mt_sync->cond_[r - 1], mutex); |
1681 | 0 | pthread_mutex_unlock(mutex); |
1682 | 0 | } |
1683 | | #else |
1684 | | (void)tpl_row_mt_sync; |
1685 | | (void)r; |
1686 | | (void)c; |
1687 | | #endif // CONFIG_MULTITHREAD |
1688 | 0 | } |
1689 | | |
1690 | | void av1_tpl_row_mt_sync_write(AV1TplRowMultiThreadSync *tpl_row_mt_sync, int r, |
1691 | 0 | int c, int cols) { |
1692 | 0 | #if CONFIG_MULTITHREAD |
1693 | 0 | int nsync = tpl_row_mt_sync->sync_range; |
1694 | 0 | int cur; |
1695 | | // Only signal when there are enough encoded blocks for next row to run. |
1696 | 0 | int sig = 1; |
1697 | |
|
1698 | 0 | if (c < cols - 1) { |
1699 | 0 | cur = c; |
1700 | 0 | if (c % nsync) sig = 0; |
1701 | 0 | } else { |
1702 | 0 | cur = cols + nsync; |
1703 | 0 | } |
1704 | |
|
1705 | 0 | if (sig) { |
1706 | 0 | pthread_mutex_lock(&tpl_row_mt_sync->mutex_[r]); |
1707 | |
|
1708 | 0 | tpl_row_mt_sync->num_finished_cols[r] = cur; |
1709 | |
|
1710 | 0 | pthread_cond_signal(&tpl_row_mt_sync->cond_[r]); |
1711 | 0 | pthread_mutex_unlock(&tpl_row_mt_sync->mutex_[r]); |
1712 | 0 | } |
1713 | | #else |
1714 | | (void)tpl_row_mt_sync; |
1715 | | (void)r; |
1716 | | (void)c; |
1717 | | (void)cols; |
1718 | | #endif // CONFIG_MULTITHREAD |
1719 | 0 | } |
1720 | | |
1721 | | // Each worker calls tpl_worker_hook() and computes the tpl data. |
1722 | 0 | static int tpl_worker_hook(void *arg1, void *unused) { |
1723 | 0 | (void)unused; |
1724 | 0 | EncWorkerData *thread_data = (EncWorkerData *)arg1; |
1725 | 0 | AV1_COMP *cpi = thread_data->cpi; |
1726 | 0 | AV1_COMMON *cm = &cpi->common; |
1727 | 0 | MACROBLOCK *x = &thread_data->td->mb; |
1728 | 0 | MACROBLOCKD *xd = &x->e_mbd; |
1729 | 0 | TplTxfmStats *tpl_txfm_stats = &thread_data->td->tpl_txfm_stats; |
1730 | 0 | CommonModeInfoParams *mi_params = &cm->mi_params; |
1731 | 0 | BLOCK_SIZE bsize = convert_length_to_bsize(cpi->ppi->tpl_data.tpl_bsize_1d); |
1732 | 0 | TX_SIZE tx_size = max_txsize_lookup[bsize]; |
1733 | 0 | int mi_height = mi_size_high[bsize]; |
1734 | 0 | int num_active_workers = cpi->ppi->tpl_data.tpl_mt_sync.num_threads_working; |
1735 | |
|
1736 | 0 | av1_init_tpl_txfm_stats(tpl_txfm_stats); |
1737 | |
|
1738 | 0 | for (int mi_row = thread_data->start * mi_height; mi_row < mi_params->mi_rows; |
1739 | 0 | mi_row += num_active_workers * mi_height) { |
1740 | | // Motion estimation row boundary |
1741 | 0 | av1_set_mv_row_limits(mi_params, &x->mv_limits, mi_row, mi_height, |
1742 | 0 | cpi->oxcf.border_in_pixels); |
1743 | 0 | xd->mb_to_top_edge = -GET_MV_SUBPEL(mi_row * MI_SIZE); |
1744 | 0 | xd->mb_to_bottom_edge = |
1745 | 0 | GET_MV_SUBPEL((mi_params->mi_rows - mi_height - mi_row) * MI_SIZE); |
1746 | 0 | av1_mc_flow_dispenser_row(cpi, tpl_txfm_stats, x, mi_row, bsize, tx_size); |
1747 | 0 | } |
1748 | 0 | return 1; |
1749 | 0 | } |
1750 | | |
1751 | | // Deallocate tpl synchronization related mutex and data. |
1752 | 0 | void av1_tpl_dealloc(AV1TplRowMultiThreadSync *tpl_sync) { |
1753 | 0 | assert(tpl_sync != NULL); |
1754 | |
|
1755 | 0 | #if CONFIG_MULTITHREAD |
1756 | 0 | if (tpl_sync->mutex_ != NULL) { |
1757 | 0 | for (int i = 0; i < tpl_sync->rows; ++i) |
1758 | 0 | pthread_mutex_destroy(&tpl_sync->mutex_[i]); |
1759 | 0 | aom_free(tpl_sync->mutex_); |
1760 | 0 | } |
1761 | 0 | if (tpl_sync->cond_ != NULL) { |
1762 | 0 | for (int i = 0; i < tpl_sync->rows; ++i) |
1763 | 0 | pthread_cond_destroy(&tpl_sync->cond_[i]); |
1764 | 0 | aom_free(tpl_sync->cond_); |
1765 | 0 | } |
1766 | 0 | #endif // CONFIG_MULTITHREAD |
1767 | |
|
1768 | 0 | aom_free(tpl_sync->num_finished_cols); |
1769 | | // clear the structure as the source of this call may be a resize in which |
1770 | | // case this call will be followed by an _alloc() which may fail. |
1771 | 0 | av1_zero(*tpl_sync); |
1772 | 0 | } |
1773 | | |
1774 | | // Allocate memory for tpl row synchronization. |
1775 | | void av1_tpl_alloc(AV1TplRowMultiThreadSync *tpl_sync, AV1_COMMON *cm, |
1776 | 0 | int mb_rows) { |
1777 | 0 | tpl_sync->rows = mb_rows; |
1778 | 0 | #if CONFIG_MULTITHREAD |
1779 | 0 | { |
1780 | 0 | CHECK_MEM_ERROR(cm, tpl_sync->mutex_, |
1781 | 0 | aom_malloc(sizeof(*tpl_sync->mutex_) * mb_rows)); |
1782 | 0 | if (tpl_sync->mutex_) { |
1783 | 0 | for (int i = 0; i < mb_rows; ++i) |
1784 | 0 | pthread_mutex_init(&tpl_sync->mutex_[i], NULL); |
1785 | 0 | } |
1786 | |
|
1787 | 0 | CHECK_MEM_ERROR(cm, tpl_sync->cond_, |
1788 | 0 | aom_malloc(sizeof(*tpl_sync->cond_) * mb_rows)); |
1789 | 0 | if (tpl_sync->cond_) { |
1790 | 0 | for (int i = 0; i < mb_rows; ++i) |
1791 | 0 | pthread_cond_init(&tpl_sync->cond_[i], NULL); |
1792 | 0 | } |
1793 | 0 | } |
1794 | 0 | #endif // CONFIG_MULTITHREAD |
1795 | 0 | CHECK_MEM_ERROR(cm, tpl_sync->num_finished_cols, |
1796 | 0 | aom_malloc(sizeof(*tpl_sync->num_finished_cols) * mb_rows)); |
1797 | | |
1798 | | // Set up nsync. |
1799 | 0 | tpl_sync->sync_range = 1; |
1800 | 0 | } |
1801 | | |
1802 | | // Each worker is prepared by assigning the hook function and individual thread |
1803 | | // data. |
1804 | | static AOM_INLINE void prepare_tpl_workers(AV1_COMP *cpi, AVxWorkerHook hook, |
1805 | 0 | int num_workers) { |
1806 | 0 | MultiThreadInfo *mt_info = &cpi->mt_info; |
1807 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
1808 | 0 | AVxWorker *worker = &mt_info->workers[i]; |
1809 | 0 | EncWorkerData *thread_data = &mt_info->tile_thr_data[i]; |
1810 | |
|
1811 | 0 | worker->hook = hook; |
1812 | 0 | worker->data1 = thread_data; |
1813 | 0 | worker->data2 = NULL; |
1814 | |
|
1815 | 0 | thread_data->thread_id = i; |
1816 | | // Set the starting tile for each thread. |
1817 | 0 | thread_data->start = i; |
1818 | |
|
1819 | 0 | thread_data->cpi = cpi; |
1820 | 0 | if (i == 0) { |
1821 | 0 | thread_data->td = &cpi->td; |
1822 | 0 | #if !CONFIG_FRAME_PARALLEL_ENCODE |
1823 | 0 | } |
1824 | | #else |
1825 | | } else { |
1826 | | thread_data->td = thread_data->original_td; |
1827 | | } |
1828 | | #endif // CONFIG_FRAME_PARALLEL_ENCODE |
1829 | | |
1830 | | // Before encoding a frame, copy the thread data from cpi. |
1831 | 0 | if (thread_data->td != &cpi->td) { |
1832 | 0 | thread_data->td->mb = cpi->td.mb; |
1833 | | // OBMC buffers are used only to init MS params and remain unused when |
1834 | | // called from tpl, hence set the buffers to defaults. |
1835 | 0 | av1_init_obmc_buffer(&thread_data->td->mb.obmc_buffer); |
1836 | 0 | thread_data->td->mb.tmp_conv_dst = thread_data->td->tmp_conv_dst; |
1837 | 0 | thread_data->td->mb.e_mbd.tmp_conv_dst = thread_data->td->mb.tmp_conv_dst; |
1838 | 0 | } |
1839 | 0 | } |
1840 | 0 | } |
1841 | | |
1842 | | // Accumulate transform stats after tpl. |
1843 | | static void tpl_accumulate_txfm_stats(ThreadData *main_td, |
1844 | | const MultiThreadInfo *mt_info, |
1845 | 0 | int num_workers) { |
1846 | 0 | TplTxfmStats *accumulated_stats = &main_td->tpl_txfm_stats; |
1847 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
1848 | 0 | AVxWorker *const worker = &mt_info->workers[i]; |
1849 | 0 | EncWorkerData *const thread_data = (EncWorkerData *)worker->data1; |
1850 | 0 | ThreadData *td = thread_data->td; |
1851 | 0 | if (td != main_td) { |
1852 | 0 | const TplTxfmStats *tpl_txfm_stats = &td->tpl_txfm_stats; |
1853 | 0 | av1_accumulate_tpl_txfm_stats(tpl_txfm_stats, accumulated_stats); |
1854 | 0 | } |
1855 | 0 | } |
1856 | 0 | } |
1857 | | |
1858 | | // Implements multi-threading for tpl. |
1859 | 0 | void av1_mc_flow_dispenser_mt(AV1_COMP *cpi) { |
1860 | 0 | AV1_COMMON *cm = &cpi->common; |
1861 | 0 | CommonModeInfoParams *mi_params = &cm->mi_params; |
1862 | 0 | MultiThreadInfo *mt_info = &cpi->mt_info; |
1863 | 0 | TplParams *tpl_data = &cpi->ppi->tpl_data; |
1864 | 0 | AV1TplRowMultiThreadSync *tpl_sync = &tpl_data->tpl_mt_sync; |
1865 | 0 | int mb_rows = mi_params->mb_rows; |
1866 | 0 | int num_workers = |
1867 | 0 | AOMMIN(mt_info->num_mod_workers[MOD_TPL], mt_info->num_workers); |
1868 | |
|
1869 | 0 | if (mb_rows != tpl_sync->rows) { |
1870 | 0 | av1_tpl_dealloc(tpl_sync); |
1871 | 0 | av1_tpl_alloc(tpl_sync, cm, mb_rows); |
1872 | 0 | } |
1873 | 0 | tpl_sync->num_threads_working = num_workers; |
1874 | | |
1875 | | // Initialize cur_mb_col to -1 for all MB rows. |
1876 | 0 | memset(tpl_sync->num_finished_cols, -1, |
1877 | 0 | sizeof(*tpl_sync->num_finished_cols) * mb_rows); |
1878 | |
|
1879 | 0 | prepare_tpl_workers(cpi, tpl_worker_hook, num_workers); |
1880 | 0 | launch_workers(&cpi->mt_info, num_workers); |
1881 | 0 | sync_enc_workers(&cpi->mt_info, cm, num_workers); |
1882 | 0 | tpl_accumulate_txfm_stats(&cpi->td, &cpi->mt_info, num_workers); |
1883 | 0 | } |
1884 | | |
1885 | | // Deallocate memory for temporal filter multi-thread synchronization. |
1886 | 0 | void av1_tf_mt_dealloc(AV1TemporalFilterSync *tf_sync) { |
1887 | 0 | assert(tf_sync != NULL); |
1888 | 0 | #if CONFIG_MULTITHREAD |
1889 | 0 | if (tf_sync->mutex_ != NULL) { |
1890 | 0 | pthread_mutex_destroy(tf_sync->mutex_); |
1891 | 0 | aom_free(tf_sync->mutex_); |
1892 | 0 | } |
1893 | 0 | #endif // CONFIG_MULTITHREAD |
1894 | 0 | tf_sync->next_tf_row = 0; |
1895 | 0 | } |
1896 | | |
1897 | | // Checks if a job is available. If job is available, |
1898 | | // populates next_tf_row and returns 1, else returns 0. |
1899 | | static AOM_INLINE int tf_get_next_job(AV1TemporalFilterSync *tf_mt_sync, |
1900 | 0 | int *current_mb_row, int mb_rows) { |
1901 | 0 | int do_next_row = 0; |
1902 | 0 | #if CONFIG_MULTITHREAD |
1903 | 0 | pthread_mutex_t *tf_mutex_ = tf_mt_sync->mutex_; |
1904 | 0 | pthread_mutex_lock(tf_mutex_); |
1905 | 0 | #endif |
1906 | 0 | if (tf_mt_sync->next_tf_row < mb_rows) { |
1907 | 0 | *current_mb_row = tf_mt_sync->next_tf_row; |
1908 | 0 | tf_mt_sync->next_tf_row++; |
1909 | 0 | do_next_row = 1; |
1910 | 0 | } |
1911 | 0 | #if CONFIG_MULTITHREAD |
1912 | 0 | pthread_mutex_unlock(tf_mutex_); |
1913 | 0 | #endif |
1914 | 0 | return do_next_row; |
1915 | 0 | } |
1916 | | |
1917 | | // Hook function for each thread in temporal filter multi-threading. |
1918 | 0 | static int tf_worker_hook(void *arg1, void *unused) { |
1919 | 0 | (void)unused; |
1920 | 0 | EncWorkerData *thread_data = (EncWorkerData *)arg1; |
1921 | 0 | AV1_COMP *cpi = thread_data->cpi; |
1922 | 0 | ThreadData *td = thread_data->td; |
1923 | 0 | TemporalFilterCtx *tf_ctx = &cpi->tf_ctx; |
1924 | 0 | AV1TemporalFilterSync *tf_sync = &cpi->mt_info.tf_sync; |
1925 | 0 | const struct scale_factors *scale = &cpi->tf_ctx.sf; |
1926 | 0 | const int num_planes = av1_num_planes(&cpi->common); |
1927 | 0 | assert(num_planes >= 1 && num_planes <= MAX_MB_PLANE); |
1928 | |
|
1929 | 0 | MACROBLOCKD *mbd = &td->mb.e_mbd; |
1930 | 0 | uint8_t *input_buffer[MAX_MB_PLANE]; |
1931 | 0 | MB_MODE_INFO **input_mb_mode_info; |
1932 | 0 | tf_save_state(mbd, &input_mb_mode_info, input_buffer, num_planes); |
1933 | 0 | tf_setup_macroblockd(mbd, &td->tf_data, scale); |
1934 | |
|
1935 | 0 | int current_mb_row = -1; |
1936 | |
|
1937 | 0 | while (tf_get_next_job(tf_sync, ¤t_mb_row, tf_ctx->mb_rows)) |
1938 | 0 | av1_tf_do_filtering_row(cpi, td, current_mb_row); |
1939 | |
|
1940 | 0 | tf_restore_state(mbd, input_mb_mode_info, input_buffer, num_planes); |
1941 | |
|
1942 | 0 | return 1; |
1943 | 0 | } |
1944 | | |
1945 | | // Assigns temporal filter hook function and thread data to each worker. |
1946 | | static void prepare_tf_workers(AV1_COMP *cpi, AVxWorkerHook hook, |
1947 | 0 | int num_workers, int is_highbitdepth) { |
1948 | 0 | MultiThreadInfo *mt_info = &cpi->mt_info; |
1949 | 0 | mt_info->tf_sync.next_tf_row = 0; |
1950 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
1951 | 0 | AVxWorker *worker = &mt_info->workers[i]; |
1952 | 0 | EncWorkerData *thread_data = &mt_info->tile_thr_data[i]; |
1953 | |
|
1954 | 0 | worker->hook = hook; |
1955 | 0 | worker->data1 = thread_data; |
1956 | 0 | worker->data2 = NULL; |
1957 | |
|
1958 | 0 | thread_data->thread_id = i; |
1959 | | // Set the starting tile for each thread. |
1960 | 0 | thread_data->start = i; |
1961 | |
|
1962 | 0 | thread_data->cpi = cpi; |
1963 | 0 | if (i == 0) { |
1964 | 0 | thread_data->td = &cpi->td; |
1965 | 0 | #if !CONFIG_FRAME_PARALLEL_ENCODE |
1966 | 0 | } |
1967 | | #else |
1968 | | } else { |
1969 | | thread_data->td = thread_data->original_td; |
1970 | | } |
1971 | | #endif // CONFIG_FRAME_PARALLEL_ENCODE |
1972 | | |
1973 | | // Before encoding a frame, copy the thread data from cpi. |
1974 | 0 | if (thread_data->td != &cpi->td) { |
1975 | 0 | thread_data->td->mb = cpi->td.mb; |
1976 | | // OBMC buffers are used only to init MS params and remain unused when |
1977 | | // called from tf, hence set the buffers to defaults. |
1978 | 0 | av1_init_obmc_buffer(&thread_data->td->mb.obmc_buffer); |
1979 | 0 | tf_alloc_and_reset_data(&thread_data->td->tf_data, cpi->tf_ctx.num_pels, |
1980 | 0 | is_highbitdepth); |
1981 | 0 | } |
1982 | 0 | } |
1983 | 0 | } |
1984 | | |
1985 | | // Deallocate thread specific data for temporal filter. |
1986 | | static void tf_dealloc_thread_data(AV1_COMP *cpi, int num_workers, |
1987 | 0 | int is_highbitdepth) { |
1988 | 0 | MultiThreadInfo *mt_info = &cpi->mt_info; |
1989 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
1990 | 0 | EncWorkerData *thread_data = &mt_info->tile_thr_data[i]; |
1991 | 0 | ThreadData *td = thread_data->td; |
1992 | 0 | if (td != &cpi->td) tf_dealloc_data(&td->tf_data, is_highbitdepth); |
1993 | 0 | } |
1994 | 0 | } |
1995 | | |
1996 | | // Accumulate sse and sum after temporal filtering. |
1997 | 0 | static void tf_accumulate_frame_diff(AV1_COMP *cpi, int num_workers) { |
1998 | 0 | FRAME_DIFF *total_diff = &cpi->td.tf_data.diff; |
1999 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
2000 | 0 | AVxWorker *const worker = &cpi->mt_info.workers[i]; |
2001 | 0 | EncWorkerData *const thread_data = (EncWorkerData *)worker->data1; |
2002 | 0 | ThreadData *td = thread_data->td; |
2003 | 0 | FRAME_DIFF *diff = &td->tf_data.diff; |
2004 | 0 | if (td != &cpi->td) { |
2005 | 0 | total_diff->sse += diff->sse; |
2006 | 0 | total_diff->sum += diff->sum; |
2007 | 0 | } |
2008 | 0 | } |
2009 | 0 | } |
2010 | | |
2011 | | // Implements multi-threading for temporal filter. |
2012 | 0 | void av1_tf_do_filtering_mt(AV1_COMP *cpi) { |
2013 | 0 | AV1_COMMON *cm = &cpi->common; |
2014 | 0 | MultiThreadInfo *mt_info = &cpi->mt_info; |
2015 | 0 | const int is_highbitdepth = cpi->tf_ctx.is_highbitdepth; |
2016 | |
|
2017 | 0 | int num_workers = |
2018 | 0 | AOMMIN(mt_info->num_mod_workers[MOD_TF], mt_info->num_workers); |
2019 | |
|
2020 | 0 | prepare_tf_workers(cpi, tf_worker_hook, num_workers, is_highbitdepth); |
2021 | 0 | launch_workers(mt_info, num_workers); |
2022 | 0 | sync_enc_workers(mt_info, cm, num_workers); |
2023 | 0 | tf_accumulate_frame_diff(cpi, num_workers); |
2024 | 0 | tf_dealloc_thread_data(cpi, num_workers, is_highbitdepth); |
2025 | 0 | } |
2026 | | |
2027 | | // Checks if a job is available in the current direction. If a job is available, |
2028 | | // frame_idx will be populated and returns 1, else returns 0. |
2029 | | static AOM_INLINE int get_next_gm_job(AV1_COMP *cpi, int *frame_idx, |
2030 | 0 | int cur_dir) { |
2031 | 0 | GlobalMotionInfo *gm_info = &cpi->gm_info; |
2032 | 0 | JobInfo *job_info = &cpi->mt_info.gm_sync.job_info; |
2033 | |
|
2034 | 0 | int total_refs = gm_info->num_ref_frames[cur_dir]; |
2035 | 0 | int8_t cur_frame_to_process = job_info->next_frame_to_process[cur_dir]; |
2036 | |
|
2037 | 0 | if (cur_frame_to_process < total_refs && !job_info->early_exit[cur_dir]) { |
2038 | 0 | *frame_idx = gm_info->reference_frames[cur_dir][cur_frame_to_process].frame; |
2039 | 0 | job_info->next_frame_to_process[cur_dir] += 1; |
2040 | 0 | return 1; |
2041 | 0 | } |
2042 | 0 | return 0; |
2043 | 0 | } |
2044 | | |
2045 | | // Switches the current direction and calls the function get_next_gm_job() if |
2046 | | // the speed feature 'prune_ref_frame_for_gm_search' is not set. |
2047 | | static AOM_INLINE void switch_direction(AV1_COMP *cpi, int *frame_idx, |
2048 | 0 | int *cur_dir) { |
2049 | 0 | if (cpi->sf.gm_sf.prune_ref_frame_for_gm_search) return; |
2050 | | // Switch the direction and get next job |
2051 | 0 | *cur_dir = !(*cur_dir); |
2052 | 0 | get_next_gm_job(cpi, frame_idx, *(cur_dir)); |
2053 | 0 | } |
2054 | | |
2055 | | // Initializes inliers, num_inliers and segment_map. |
2056 | | static AOM_INLINE void init_gm_thread_data( |
2057 | 0 | const GlobalMotionInfo *gm_info, GlobalMotionThreadData *thread_data) { |
2058 | 0 | for (int m = 0; m < RANSAC_NUM_MOTIONS; m++) { |
2059 | 0 | MotionModel motion_params = thread_data->params_by_motion[m]; |
2060 | 0 | av1_zero(motion_params.params); |
2061 | 0 | motion_params.num_inliers = 0; |
2062 | 0 | } |
2063 | |
|
2064 | 0 | av1_zero_array(thread_data->segment_map, |
2065 | 0 | gm_info->segment_map_w * gm_info->segment_map_h); |
2066 | 0 | } |
2067 | | |
2068 | | // Hook function for each thread in global motion multi-threading. |
2069 | 0 | static int gm_mt_worker_hook(void *arg1, void *unused) { |
2070 | 0 | (void)unused; |
2071 | |
|
2072 | 0 | EncWorkerData *thread_data = (EncWorkerData *)arg1; |
2073 | 0 | AV1_COMP *cpi = thread_data->cpi; |
2074 | 0 | GlobalMotionInfo *gm_info = &cpi->gm_info; |
2075 | 0 | MultiThreadInfo *mt_info = &cpi->mt_info; |
2076 | 0 | JobInfo *job_info = &mt_info->gm_sync.job_info; |
2077 | 0 | int thread_id = thread_data->thread_id; |
2078 | 0 | GlobalMotionThreadData *gm_thread_data = |
2079 | 0 | &mt_info->gm_sync.thread_data[thread_id]; |
2080 | 0 | int cur_dir = job_info->thread_id_to_dir[thread_id]; |
2081 | 0 | #if CONFIG_MULTITHREAD |
2082 | 0 | pthread_mutex_t *gm_mt_mutex_ = mt_info->gm_sync.mutex_; |
2083 | 0 | #endif |
2084 | |
|
2085 | 0 | while (1) { |
2086 | 0 | int ref_buf_idx = -1; |
2087 | 0 | int ref_frame_idx = -1; |
2088 | |
|
2089 | 0 | #if CONFIG_MULTITHREAD |
2090 | 0 | pthread_mutex_lock(gm_mt_mutex_); |
2091 | 0 | #endif |
2092 | | |
2093 | | // Populates ref_buf_idx(the reference frame type) for which global motion |
2094 | | // estimation will be done. |
2095 | 0 | if (!get_next_gm_job(cpi, &ref_buf_idx, cur_dir)) { |
2096 | | // No jobs are available for the current direction. Switch |
2097 | | // to other direction and get the next job, if available. |
2098 | 0 | switch_direction(cpi, &ref_buf_idx, &cur_dir); |
2099 | 0 | } |
2100 | | |
2101 | | // 'ref_frame_idx' holds the index of the current reference frame type in |
2102 | | // gm_info->reference_frames. job_info->next_frame_to_process will be |
2103 | | // incremented in get_next_gm_job() and hence subtracting by 1. |
2104 | 0 | ref_frame_idx = job_info->next_frame_to_process[cur_dir] - 1; |
2105 | |
|
2106 | 0 | #if CONFIG_MULTITHREAD |
2107 | 0 | pthread_mutex_unlock(gm_mt_mutex_); |
2108 | 0 | #endif |
2109 | |
|
2110 | 0 | if (ref_buf_idx == -1) break; |
2111 | | |
2112 | 0 | init_gm_thread_data(gm_info, gm_thread_data); |
2113 | | |
2114 | | // Compute global motion for the given ref_buf_idx. |
2115 | 0 | av1_compute_gm_for_valid_ref_frames( |
2116 | 0 | cpi, gm_info->ref_buf, ref_buf_idx, gm_info->num_src_corners, |
2117 | 0 | gm_info->src_corners, gm_info->src_buffer, |
2118 | 0 | gm_thread_data->params_by_motion, gm_thread_data->segment_map, |
2119 | 0 | gm_info->segment_map_w, gm_info->segment_map_h); |
2120 | |
|
2121 | 0 | #if CONFIG_MULTITHREAD |
2122 | 0 | pthread_mutex_lock(gm_mt_mutex_); |
2123 | 0 | #endif |
2124 | 0 | assert(ref_frame_idx != -1); |
2125 | | // If global motion w.r.t. current ref frame is |
2126 | | // INVALID/TRANSLATION/IDENTITY, skip the evaluation of global motion w.r.t |
2127 | | // the remaining ref frames in that direction. The below exit is disabled |
2128 | | // when ref frame distance w.r.t. current frame is zero. E.g.: |
2129 | | // source_alt_ref_frame w.r.t. ARF frames. |
2130 | 0 | if (cpi->sf.gm_sf.prune_ref_frame_for_gm_search && |
2131 | 0 | gm_info->reference_frames[cur_dir][ref_frame_idx].distance != 0 && |
2132 | 0 | cpi->common.global_motion[ref_buf_idx].wmtype != ROTZOOM) |
2133 | 0 | job_info->early_exit[cur_dir] = 1; |
2134 | |
|
2135 | 0 | #if CONFIG_MULTITHREAD |
2136 | 0 | pthread_mutex_unlock(gm_mt_mutex_); |
2137 | 0 | #endif |
2138 | 0 | } |
2139 | 0 | return 1; |
2140 | 0 | } |
2141 | | |
2142 | | // Assigns global motion hook function and thread data to each worker. |
2143 | | static AOM_INLINE void prepare_gm_workers(AV1_COMP *cpi, AVxWorkerHook hook, |
2144 | 0 | int num_workers) { |
2145 | 0 | MultiThreadInfo *mt_info = &cpi->mt_info; |
2146 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
2147 | 0 | AVxWorker *worker = &mt_info->workers[i]; |
2148 | 0 | EncWorkerData *thread_data = &mt_info->tile_thr_data[i]; |
2149 | |
|
2150 | 0 | worker->hook = hook; |
2151 | 0 | worker->data1 = thread_data; |
2152 | 0 | worker->data2 = NULL; |
2153 | |
|
2154 | 0 | thread_data->thread_id = i; |
2155 | | // Set the starting tile for each thread. |
2156 | 0 | thread_data->start = i; |
2157 | |
|
2158 | 0 | thread_data->cpi = cpi; |
2159 | 0 | if (i == 0) { |
2160 | 0 | thread_data->td = &cpi->td; |
2161 | 0 | #if !CONFIG_FRAME_PARALLEL_ENCODE |
2162 | 0 | } |
2163 | | #else |
2164 | | } else { |
2165 | | thread_data->td = thread_data->original_td; |
2166 | | } |
2167 | | #endif // CONFIG_FRAME_PARALLEL_ENCODE |
2168 | 0 | } |
2169 | 0 | } |
2170 | | |
2171 | | // Assigns available threads to past/future direction. |
2172 | | static AOM_INLINE void assign_thread_to_dir(int8_t *thread_id_to_dir, |
2173 | 0 | int num_workers) { |
2174 | 0 | int8_t frame_dir_idx = 0; |
2175 | |
|
2176 | 0 | for (int i = 0; i < num_workers; i++) { |
2177 | 0 | thread_id_to_dir[i] = frame_dir_idx++; |
2178 | 0 | if (frame_dir_idx == MAX_DIRECTIONS) frame_dir_idx = 0; |
2179 | 0 | } |
2180 | 0 | } |
2181 | | |
2182 | | // Computes number of workers for global motion multi-threading. |
2183 | 0 | static AOM_INLINE int compute_gm_workers(const AV1_COMP *cpi) { |
2184 | 0 | int total_refs = |
2185 | 0 | cpi->gm_info.num_ref_frames[0] + cpi->gm_info.num_ref_frames[1]; |
2186 | 0 | int num_gm_workers = cpi->sf.gm_sf.prune_ref_frame_for_gm_search |
2187 | 0 | ? AOMMIN(MAX_DIRECTIONS, total_refs) |
2188 | 0 | : total_refs; |
2189 | 0 | num_gm_workers = AOMMIN(num_gm_workers, cpi->mt_info.num_workers); |
2190 | 0 | return (num_gm_workers); |
2191 | 0 | } |
2192 | | |
2193 | | // Frees the memory allocated for each worker in global motion multi-threading. |
2194 | 0 | void av1_gm_dealloc(AV1GlobalMotionSync *gm_sync_data) { |
2195 | 0 | if (gm_sync_data->thread_data != NULL) { |
2196 | 0 | for (int j = 0; j < gm_sync_data->allocated_workers; j++) { |
2197 | 0 | GlobalMotionThreadData *thread_data = &gm_sync_data->thread_data[j]; |
2198 | 0 | aom_free(thread_data->segment_map); |
2199 | |
|
2200 | 0 | for (int m = 0; m < RANSAC_NUM_MOTIONS; m++) |
2201 | 0 | aom_free(thread_data->params_by_motion[m].inliers); |
2202 | 0 | } |
2203 | 0 | aom_free(gm_sync_data->thread_data); |
2204 | 0 | } |
2205 | 0 | } |
2206 | | |
2207 | | // Allocates memory for inliers and segment_map for each worker in global motion |
2208 | | // multi-threading. |
2209 | 0 | static AOM_INLINE void gm_alloc(AV1_COMP *cpi, int num_workers) { |
2210 | 0 | AV1_COMMON *cm = &cpi->common; |
2211 | 0 | AV1GlobalMotionSync *gm_sync = &cpi->mt_info.gm_sync; |
2212 | 0 | GlobalMotionInfo *gm_info = &cpi->gm_info; |
2213 | |
|
2214 | 0 | gm_sync->allocated_workers = num_workers; |
2215 | 0 | gm_sync->allocated_width = cpi->source->y_width; |
2216 | 0 | gm_sync->allocated_height = cpi->source->y_height; |
2217 | |
|
2218 | 0 | CHECK_MEM_ERROR(cm, gm_sync->thread_data, |
2219 | 0 | aom_malloc(sizeof(*gm_sync->thread_data) * num_workers)); |
2220 | |
|
2221 | 0 | for (int i = 0; i < num_workers; i++) { |
2222 | 0 | GlobalMotionThreadData *thread_data = &gm_sync->thread_data[i]; |
2223 | 0 | CHECK_MEM_ERROR( |
2224 | 0 | cm, thread_data->segment_map, |
2225 | 0 | aom_malloc(sizeof(*thread_data->segment_map) * gm_info->segment_map_w * |
2226 | 0 | gm_info->segment_map_h)); |
2227 | |
|
2228 | 0 | for (int m = 0; m < RANSAC_NUM_MOTIONS; m++) { |
2229 | 0 | CHECK_MEM_ERROR( |
2230 | 0 | cm, thread_data->params_by_motion[m].inliers, |
2231 | 0 | aom_malloc(sizeof(*thread_data->params_by_motion[m].inliers) * 2 * |
2232 | 0 | MAX_CORNERS)); |
2233 | 0 | } |
2234 | 0 | } |
2235 | 0 | } |
2236 | | |
2237 | | // Implements multi-threading for global motion. |
2238 | 0 | void av1_global_motion_estimation_mt(AV1_COMP *cpi) { |
2239 | 0 | AV1GlobalMotionSync *gm_sync = &cpi->mt_info.gm_sync; |
2240 | 0 | JobInfo *job_info = &gm_sync->job_info; |
2241 | |
|
2242 | 0 | av1_zero(*job_info); |
2243 | |
|
2244 | 0 | int num_workers = compute_gm_workers(cpi); |
2245 | |
|
2246 | 0 | if (num_workers > gm_sync->allocated_workers || |
2247 | 0 | cpi->source->y_width != gm_sync->allocated_width || |
2248 | 0 | cpi->source->y_height != gm_sync->allocated_height) { |
2249 | 0 | av1_gm_dealloc(gm_sync); |
2250 | 0 | gm_alloc(cpi, num_workers); |
2251 | 0 | } |
2252 | |
|
2253 | 0 | assign_thread_to_dir(job_info->thread_id_to_dir, num_workers); |
2254 | 0 | prepare_gm_workers(cpi, gm_mt_worker_hook, num_workers); |
2255 | 0 | launch_workers(&cpi->mt_info, num_workers); |
2256 | 0 | sync_enc_workers(&cpi->mt_info, &cpi->common, num_workers); |
2257 | 0 | } |
2258 | | #endif // !CONFIG_REALTIME_ONLY |
2259 | | |
2260 | | // Compare and order tiles based on absolute sum of tx coeffs. |
2261 | 0 | static int compare_tile_order(const void *a, const void *b) { |
2262 | 0 | const PackBSTileOrder *const tile_a = (const PackBSTileOrder *)a; |
2263 | 0 | const PackBSTileOrder *const tile_b = (const PackBSTileOrder *)b; |
2264 | |
|
2265 | 0 | if (tile_a->abs_sum_level > tile_b->abs_sum_level) |
2266 | 0 | return -1; |
2267 | 0 | else if (tile_a->abs_sum_level == tile_b->abs_sum_level) |
2268 | 0 | return (tile_a->tile_idx > tile_b->tile_idx ? 1 : -1); |
2269 | 0 | else |
2270 | 0 | return 1; |
2271 | 0 | } |
2272 | | |
2273 | | // Get next tile index to be processed for pack bitstream |
2274 | | static AOM_INLINE int get_next_pack_bs_tile_idx( |
2275 | 0 | AV1EncPackBSSync *const pack_bs_sync, const int num_tiles) { |
2276 | 0 | assert(pack_bs_sync->next_job_idx <= num_tiles); |
2277 | 0 | if (pack_bs_sync->next_job_idx == num_tiles) return -1; |
2278 | | |
2279 | 0 | return pack_bs_sync->pack_bs_tile_order[pack_bs_sync->next_job_idx++] |
2280 | 0 | .tile_idx; |
2281 | 0 | } |
2282 | | |
2283 | | // Calculates bitstream chunk size based on total buffer size and tile or tile |
2284 | | // group size. |
2285 | | static AOM_INLINE size_t get_bs_chunk_size(int tg_or_tile_size, |
2286 | | const int frame_or_tg_size, |
2287 | | size_t *remain_buf_size, |
2288 | | size_t max_buf_size, |
2289 | 0 | int is_last_chunk) { |
2290 | 0 | size_t this_chunk_size; |
2291 | 0 | assert(*remain_buf_size > 0); |
2292 | 0 | if (is_last_chunk) { |
2293 | 0 | this_chunk_size = *remain_buf_size; |
2294 | 0 | *remain_buf_size = 0; |
2295 | 0 | } else { |
2296 | 0 | const uint64_t size_scale = (uint64_t)max_buf_size * tg_or_tile_size; |
2297 | 0 | this_chunk_size = (size_t)(size_scale / frame_or_tg_size); |
2298 | 0 | *remain_buf_size -= this_chunk_size; |
2299 | 0 | assert(*remain_buf_size > 0); |
2300 | 0 | } |
2301 | 0 | assert(this_chunk_size > 0); |
2302 | 0 | return this_chunk_size; |
2303 | 0 | } |
2304 | | |
2305 | | // Initializes params required for pack bitstream tile. |
2306 | | static void init_tile_pack_bs_params(AV1_COMP *const cpi, uint8_t *const dst, |
2307 | | struct aom_write_bit_buffer *saved_wb, |
2308 | | PackBSParams *const pack_bs_params_arr, |
2309 | 0 | uint8_t obu_extn_header) { |
2310 | 0 | MACROBLOCKD *const xd = &cpi->td.mb.e_mbd; |
2311 | 0 | AV1_COMMON *const cm = &cpi->common; |
2312 | 0 | const CommonTileParams *const tiles = &cm->tiles; |
2313 | 0 | const int num_tiles = tiles->cols * tiles->rows; |
2314 | | // Fixed size tile groups for the moment |
2315 | 0 | const int num_tg_hdrs = cpi->num_tg; |
2316 | | // Tile group size in terms of number of tiles. |
2317 | 0 | const int tg_size_in_tiles = (num_tiles + num_tg_hdrs - 1) / num_tg_hdrs; |
2318 | 0 | uint8_t *tile_dst = dst; |
2319 | 0 | uint8_t *tile_data_curr = dst; |
2320 | | // Max tile group count can not be more than MAX_TILES. |
2321 | 0 | int tg_size_mi[MAX_TILES] = { 0 }; // Size of tile group in mi units |
2322 | 0 | int tile_idx; |
2323 | 0 | int tg_idx = 0; |
2324 | 0 | int tile_count_in_tg = 0; |
2325 | 0 | int new_tg = 1; |
2326 | | |
2327 | | // Populate pack bitstream params of all tiles. |
2328 | 0 | for (tile_idx = 0; tile_idx < num_tiles; tile_idx++) { |
2329 | 0 | const TileInfo *const tile_info = &cpi->tile_data[tile_idx].tile_info; |
2330 | 0 | PackBSParams *const pack_bs_params = &pack_bs_params_arr[tile_idx]; |
2331 | | // Calculate tile size in mi units. |
2332 | 0 | const int tile_size_mi = (tile_info->mi_col_end - tile_info->mi_col_start) * |
2333 | 0 | (tile_info->mi_row_end - tile_info->mi_row_start); |
2334 | 0 | int is_last_tile_in_tg = 0; |
2335 | 0 | tile_count_in_tg++; |
2336 | 0 | if (tile_count_in_tg == tg_size_in_tiles || tile_idx == (num_tiles - 1)) |
2337 | 0 | is_last_tile_in_tg = 1; |
2338 | | |
2339 | | // Populate pack bitstream params of this tile. |
2340 | 0 | pack_bs_params->curr_tg_hdr_size = 0; |
2341 | 0 | pack_bs_params->obu_extn_header = obu_extn_header; |
2342 | 0 | pack_bs_params->saved_wb = saved_wb; |
2343 | 0 | pack_bs_params->obu_header_size = 0; |
2344 | 0 | pack_bs_params->is_last_tile_in_tg = is_last_tile_in_tg; |
2345 | 0 | pack_bs_params->new_tg = new_tg; |
2346 | 0 | pack_bs_params->tile_col = tile_info->tile_col; |
2347 | 0 | pack_bs_params->tile_row = tile_info->tile_row; |
2348 | 0 | pack_bs_params->tile_size_mi = tile_size_mi; |
2349 | 0 | tg_size_mi[tg_idx] += tile_size_mi; |
2350 | |
|
2351 | 0 | if (new_tg) new_tg = 0; |
2352 | 0 | if (is_last_tile_in_tg) { |
2353 | 0 | tile_count_in_tg = 0; |
2354 | 0 | new_tg = 1; |
2355 | 0 | tg_idx++; |
2356 | 0 | } |
2357 | 0 | } |
2358 | |
|
2359 | 0 | assert(cpi->available_bs_size > 0); |
2360 | 0 | size_t tg_buf_size[MAX_TILES] = { 0 }; |
2361 | 0 | size_t max_buf_size = cpi->available_bs_size; |
2362 | 0 | size_t remain_buf_size = max_buf_size; |
2363 | 0 | const int frame_size_mi = cm->mi_params.mi_rows * cm->mi_params.mi_cols; |
2364 | |
|
2365 | 0 | tile_idx = 0; |
2366 | | // Prepare obu, tile group and frame header of each tile group. |
2367 | 0 | for (tg_idx = 0; tg_idx < cpi->num_tg; tg_idx++) { |
2368 | 0 | PackBSParams *const pack_bs_params = &pack_bs_params_arr[tile_idx]; |
2369 | 0 | int is_last_tg = tg_idx == cpi->num_tg - 1; |
2370 | | // Prorate bitstream buffer size based on tile group size and available |
2371 | | // buffer size. This buffer will be used to store headers and tile data. |
2372 | 0 | tg_buf_size[tg_idx] = |
2373 | 0 | get_bs_chunk_size(tg_size_mi[tg_idx], frame_size_mi, &remain_buf_size, |
2374 | 0 | max_buf_size, is_last_tg); |
2375 | |
|
2376 | 0 | pack_bs_params->dst = tile_dst; |
2377 | 0 | pack_bs_params->tile_data_curr = tile_dst; |
2378 | | |
2379 | | // Write obu, tile group and frame header at first tile in the tile |
2380 | | // group. |
2381 | 0 | av1_write_obu_tg_tile_headers(cpi, xd, pack_bs_params, tile_idx); |
2382 | 0 | tile_dst += tg_buf_size[tg_idx]; |
2383 | | |
2384 | | // Exclude headers from tile group buffer size. |
2385 | 0 | tg_buf_size[tg_idx] -= pack_bs_params->curr_tg_hdr_size; |
2386 | 0 | tile_idx += tg_size_in_tiles; |
2387 | 0 | } |
2388 | |
|
2389 | 0 | tg_idx = 0; |
2390 | | // Calculate bitstream buffer size of each tile in the tile group. |
2391 | 0 | for (tile_idx = 0; tile_idx < num_tiles; tile_idx++) { |
2392 | 0 | PackBSParams *const pack_bs_params = &pack_bs_params_arr[tile_idx]; |
2393 | |
|
2394 | 0 | if (pack_bs_params->new_tg) { |
2395 | 0 | max_buf_size = tg_buf_size[tg_idx]; |
2396 | 0 | remain_buf_size = max_buf_size; |
2397 | 0 | } |
2398 | | |
2399 | | // Prorate bitstream buffer size of this tile based on tile size and |
2400 | | // available buffer size. For this proration, header size is not accounted. |
2401 | 0 | const size_t tile_buf_size = get_bs_chunk_size( |
2402 | 0 | pack_bs_params->tile_size_mi, tg_size_mi[tg_idx], &remain_buf_size, |
2403 | 0 | max_buf_size, pack_bs_params->is_last_tile_in_tg); |
2404 | 0 | pack_bs_params->tile_buf_size = tile_buf_size; |
2405 | | |
2406 | | // Update base address of bitstream buffer for tile and tile group. |
2407 | 0 | if (pack_bs_params->new_tg) { |
2408 | 0 | tile_dst = pack_bs_params->dst; |
2409 | 0 | tile_data_curr = pack_bs_params->tile_data_curr; |
2410 | | // Account header size in first tile of a tile group. |
2411 | 0 | pack_bs_params->tile_buf_size += pack_bs_params->curr_tg_hdr_size; |
2412 | 0 | } else { |
2413 | 0 | pack_bs_params->dst = tile_dst; |
2414 | 0 | pack_bs_params->tile_data_curr = tile_data_curr; |
2415 | 0 | } |
2416 | |
|
2417 | 0 | if (pack_bs_params->is_last_tile_in_tg) tg_idx++; |
2418 | 0 | tile_dst += pack_bs_params->tile_buf_size; |
2419 | 0 | } |
2420 | 0 | } |
2421 | | |
2422 | | // Worker hook function of pack bitsteam multithreading. |
2423 | 0 | static int pack_bs_worker_hook(void *arg1, void *arg2) { |
2424 | 0 | EncWorkerData *const thread_data = (EncWorkerData *)arg1; |
2425 | 0 | PackBSParams *const pack_bs_params = (PackBSParams *)arg2; |
2426 | 0 | AV1_COMP *const cpi = thread_data->cpi; |
2427 | 0 | AV1_COMMON *const cm = &cpi->common; |
2428 | 0 | AV1EncPackBSSync *const pack_bs_sync = &cpi->mt_info.pack_bs_sync; |
2429 | 0 | const CommonTileParams *const tiles = &cm->tiles; |
2430 | 0 | const int num_tiles = tiles->cols * tiles->rows; |
2431 | |
|
2432 | 0 | while (1) { |
2433 | 0 | #if CONFIG_MULTITHREAD |
2434 | 0 | pthread_mutex_lock(pack_bs_sync->mutex_); |
2435 | 0 | #endif |
2436 | 0 | const int tile_idx = get_next_pack_bs_tile_idx(pack_bs_sync, num_tiles); |
2437 | 0 | #if CONFIG_MULTITHREAD |
2438 | 0 | pthread_mutex_unlock(pack_bs_sync->mutex_); |
2439 | 0 | #endif |
2440 | 0 | if (tile_idx == -1) break; |
2441 | 0 | TileDataEnc *this_tile = &cpi->tile_data[tile_idx]; |
2442 | 0 | thread_data->td->mb.e_mbd.tile_ctx = &this_tile->tctx; |
2443 | |
|
2444 | 0 | av1_pack_tile_info(cpi, thread_data->td, &pack_bs_params[tile_idx]); |
2445 | 0 | } |
2446 | |
|
2447 | 0 | return 1; |
2448 | 0 | } |
2449 | | |
2450 | | // Prepares thread data and workers of pack bitsteam multithreading. |
2451 | | static void prepare_pack_bs_workers(AV1_COMP *const cpi, |
2452 | | PackBSParams *const pack_bs_params, |
2453 | 0 | AVxWorkerHook hook, const int num_workers) { |
2454 | 0 | MultiThreadInfo *const mt_info = &cpi->mt_info; |
2455 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
2456 | 0 | AVxWorker *worker = &mt_info->workers[i]; |
2457 | 0 | EncWorkerData *const thread_data = &mt_info->tile_thr_data[i]; |
2458 | 0 | if (i == 0) { |
2459 | 0 | thread_data->td = &cpi->td; |
2460 | 0 | #if !CONFIG_FRAME_PARALLEL_ENCODE |
2461 | 0 | } |
2462 | | #else |
2463 | | } else { |
2464 | | thread_data->td = thread_data->original_td; |
2465 | | } |
2466 | | #endif // CONFIG_FRAME_PARALLEL_ENCODE |
2467 | |
|
2468 | 0 | if (thread_data->td != &cpi->td) thread_data->td->mb = cpi->td.mb; |
2469 | |
|
2470 | 0 | thread_data->cpi = cpi; |
2471 | 0 | thread_data->start = i; |
2472 | 0 | thread_data->thread_id = i; |
2473 | 0 | av1_reset_pack_bs_thread_data(thread_data->td); |
2474 | |
|
2475 | 0 | worker->hook = hook; |
2476 | 0 | worker->data1 = thread_data; |
2477 | 0 | worker->data2 = pack_bs_params; |
2478 | 0 | } |
2479 | |
|
2480 | 0 | AV1_COMMON *const cm = &cpi->common; |
2481 | 0 | AV1EncPackBSSync *const pack_bs_sync = &mt_info->pack_bs_sync; |
2482 | 0 | const uint16_t num_tiles = cm->tiles.rows * cm->tiles.cols; |
2483 | 0 | pack_bs_sync->next_job_idx = 0; |
2484 | |
|
2485 | 0 | PackBSTileOrder *const pack_bs_tile_order = pack_bs_sync->pack_bs_tile_order; |
2486 | | // Reset tile order data of pack bitstream |
2487 | 0 | av1_zero_array(pack_bs_tile_order, num_tiles); |
2488 | | |
2489 | | // Populate pack bitstream tile order structure |
2490 | 0 | for (uint16_t tile_idx = 0; tile_idx < num_tiles; tile_idx++) { |
2491 | 0 | pack_bs_tile_order[tile_idx].abs_sum_level = |
2492 | 0 | cpi->tile_data[tile_idx].abs_sum_level; |
2493 | 0 | pack_bs_tile_order[tile_idx].tile_idx = tile_idx; |
2494 | 0 | } |
2495 | | |
2496 | | // Sort tiles in descending order based on tile area. |
2497 | 0 | qsort(pack_bs_tile_order, num_tiles, sizeof(*pack_bs_tile_order), |
2498 | 0 | compare_tile_order); |
2499 | 0 | } |
2500 | | |
2501 | | // Accumulates data after pack bitsteam processing. |
2502 | | static void accumulate_pack_bs_data( |
2503 | | AV1_COMP *const cpi, const PackBSParams *const pack_bs_params_arr, |
2504 | | uint8_t *const dst, uint32_t *total_size, const FrameHeaderInfo *fh_info, |
2505 | | int *const largest_tile_id, unsigned int *max_tile_size, |
2506 | | uint32_t *const obu_header_size, uint8_t **tile_data_start, |
2507 | 0 | const int num_workers) { |
2508 | 0 | const AV1_COMMON *const cm = &cpi->common; |
2509 | 0 | const CommonTileParams *const tiles = &cm->tiles; |
2510 | 0 | const int tile_count = tiles->cols * tiles->rows; |
2511 | | // Fixed size tile groups for the moment |
2512 | 0 | size_t curr_tg_data_size = 0; |
2513 | 0 | int is_first_tg = 1; |
2514 | 0 | uint8_t *curr_tg_start = dst; |
2515 | 0 | size_t src_offset = 0; |
2516 | 0 | size_t dst_offset = 0; |
2517 | |
|
2518 | 0 | for (int tile_idx = 0; tile_idx < tile_count; tile_idx++) { |
2519 | | // PackBSParams stores all parameters required to pack tile and header |
2520 | | // info. |
2521 | 0 | const PackBSParams *const pack_bs_params = &pack_bs_params_arr[tile_idx]; |
2522 | 0 | uint32_t tile_size = 0; |
2523 | |
|
2524 | 0 | if (pack_bs_params->new_tg) { |
2525 | 0 | curr_tg_start = dst + *total_size; |
2526 | 0 | curr_tg_data_size = pack_bs_params->curr_tg_hdr_size; |
2527 | 0 | *tile_data_start += pack_bs_params->curr_tg_hdr_size; |
2528 | 0 | *obu_header_size = pack_bs_params->obu_header_size; |
2529 | 0 | } |
2530 | 0 | curr_tg_data_size += |
2531 | 0 | pack_bs_params->buf.size + (pack_bs_params->is_last_tile_in_tg ? 0 : 4); |
2532 | |
|
2533 | 0 | if (pack_bs_params->buf.size > *max_tile_size) { |
2534 | 0 | *largest_tile_id = tile_idx; |
2535 | 0 | *max_tile_size = (unsigned int)pack_bs_params->buf.size; |
2536 | 0 | } |
2537 | 0 | tile_size += |
2538 | 0 | (uint32_t)pack_bs_params->buf.size + *pack_bs_params->total_size; |
2539 | | |
2540 | | // Pack all the chunks of tile bitstreams together |
2541 | 0 | if (tile_idx != 0) memmove(dst + dst_offset, dst + src_offset, tile_size); |
2542 | |
|
2543 | 0 | if (pack_bs_params->is_last_tile_in_tg) |
2544 | 0 | av1_write_last_tile_info( |
2545 | 0 | cpi, fh_info, pack_bs_params->saved_wb, &curr_tg_data_size, |
2546 | 0 | curr_tg_start, &tile_size, tile_data_start, largest_tile_id, |
2547 | 0 | &is_first_tg, *obu_header_size, pack_bs_params->obu_extn_header); |
2548 | 0 | src_offset += pack_bs_params->tile_buf_size; |
2549 | 0 | dst_offset += tile_size; |
2550 | 0 | *total_size += tile_size; |
2551 | 0 | } |
2552 | | |
2553 | | // Accumulate thread data |
2554 | 0 | MultiThreadInfo *const mt_info = &cpi->mt_info; |
2555 | 0 | for (int idx = num_workers - 1; idx >= 0; idx--) { |
2556 | 0 | ThreadData const *td = mt_info->tile_thr_data[idx].td; |
2557 | 0 | av1_accumulate_pack_bs_thread_data(cpi, td); |
2558 | 0 | } |
2559 | 0 | } |
2560 | | |
2561 | | void av1_write_tile_obu_mt( |
2562 | | AV1_COMP *const cpi, uint8_t *const dst, uint32_t *total_size, |
2563 | | struct aom_write_bit_buffer *saved_wb, uint8_t obu_extn_header, |
2564 | | const FrameHeaderInfo *fh_info, int *const largest_tile_id, |
2565 | | unsigned int *max_tile_size, uint32_t *const obu_header_size, |
2566 | 0 | uint8_t **tile_data_start, const int num_workers) { |
2567 | 0 | MultiThreadInfo *const mt_info = &cpi->mt_info; |
2568 | |
|
2569 | 0 | PackBSParams pack_bs_params[MAX_TILES]; |
2570 | 0 | uint32_t tile_size[MAX_TILES] = { 0 }; |
2571 | |
|
2572 | 0 | for (int tile_idx = 0; tile_idx < MAX_TILES; tile_idx++) |
2573 | 0 | pack_bs_params[tile_idx].total_size = &tile_size[tile_idx]; |
2574 | |
|
2575 | 0 | init_tile_pack_bs_params(cpi, dst, saved_wb, pack_bs_params, obu_extn_header); |
2576 | 0 | prepare_pack_bs_workers(cpi, pack_bs_params, pack_bs_worker_hook, |
2577 | 0 | num_workers); |
2578 | 0 | launch_workers(mt_info, num_workers); |
2579 | 0 | sync_enc_workers(mt_info, &cpi->common, num_workers); |
2580 | 0 | accumulate_pack_bs_data(cpi, pack_bs_params, dst, total_size, fh_info, |
2581 | 0 | largest_tile_id, max_tile_size, obu_header_size, |
2582 | 0 | tile_data_start, num_workers); |
2583 | 0 | } |
2584 | | |
2585 | | // Deallocate memory for CDEF search multi-thread synchronization. |
2586 | 0 | void av1_cdef_mt_dealloc(AV1CdefSync *cdef_sync) { |
2587 | 0 | (void)cdef_sync; |
2588 | 0 | assert(cdef_sync != NULL); |
2589 | 0 | #if CONFIG_MULTITHREAD |
2590 | 0 | if (cdef_sync->mutex_ != NULL) { |
2591 | 0 | pthread_mutex_destroy(cdef_sync->mutex_); |
2592 | 0 | aom_free(cdef_sync->mutex_); |
2593 | 0 | } |
2594 | 0 | #endif // CONFIG_MULTITHREAD |
2595 | 0 | } |
2596 | | |
2597 | | // Updates the row and column indices of the next job to be processed. |
2598 | | // Also updates end_of_frame flag when the processing of all blocks is complete. |
2599 | 0 | static void update_next_job_info(AV1CdefSync *cdef_sync, int nvfb, int nhfb) { |
2600 | 0 | cdef_sync->fbc++; |
2601 | 0 | if (cdef_sync->fbc == nhfb) { |
2602 | 0 | cdef_sync->fbr++; |
2603 | 0 | if (cdef_sync->fbr == nvfb) { |
2604 | 0 | cdef_sync->end_of_frame = 1; |
2605 | 0 | } else { |
2606 | 0 | cdef_sync->fbc = 0; |
2607 | 0 | } |
2608 | 0 | } |
2609 | 0 | } |
2610 | | |
2611 | | // Initializes cdef_sync parameters. |
2612 | 0 | static AOM_INLINE void cdef_reset_job_info(AV1CdefSync *cdef_sync) { |
2613 | 0 | #if CONFIG_MULTITHREAD |
2614 | 0 | if (cdef_sync->mutex_) pthread_mutex_init(cdef_sync->mutex_, NULL); |
2615 | 0 | #endif // CONFIG_MULTITHREAD |
2616 | 0 | cdef_sync->end_of_frame = 0; |
2617 | 0 | cdef_sync->fbr = 0; |
2618 | 0 | cdef_sync->fbc = 0; |
2619 | 0 | } |
2620 | | |
2621 | | // Checks if a job is available. If job is available, |
2622 | | // populates next job information and returns 1, else returns 0. |
2623 | | static AOM_INLINE int cdef_get_next_job(AV1CdefSync *cdef_sync, |
2624 | | CdefSearchCtx *cdef_search_ctx, |
2625 | | int *cur_fbr, int *cur_fbc, |
2626 | 0 | int *sb_count) { |
2627 | 0 | #if CONFIG_MULTITHREAD |
2628 | 0 | pthread_mutex_lock(cdef_sync->mutex_); |
2629 | 0 | #endif // CONFIG_MULTITHREAD |
2630 | 0 | int do_next_block = 0; |
2631 | 0 | const int nvfb = cdef_search_ctx->nvfb; |
2632 | 0 | const int nhfb = cdef_search_ctx->nhfb; |
2633 | | |
2634 | | // If a block is skip, do not process the block and |
2635 | | // check the skip condition for the next block. |
2636 | 0 | while ((!cdef_sync->end_of_frame) && |
2637 | 0 | (cdef_sb_skip(cdef_search_ctx->mi_params, cdef_sync->fbr, |
2638 | 0 | cdef_sync->fbc))) { |
2639 | 0 | update_next_job_info(cdef_sync, nvfb, nhfb); |
2640 | 0 | } |
2641 | | |
2642 | | // Populates information needed for current job and update the row, |
2643 | | // column indices of the next block to be processed. |
2644 | 0 | if (cdef_sync->end_of_frame == 0) { |
2645 | 0 | do_next_block = 1; |
2646 | 0 | *cur_fbr = cdef_sync->fbr; |
2647 | 0 | *cur_fbc = cdef_sync->fbc; |
2648 | 0 | *sb_count = cdef_search_ctx->sb_count; |
2649 | 0 | cdef_search_ctx->sb_count++; |
2650 | 0 | update_next_job_info(cdef_sync, nvfb, nhfb); |
2651 | 0 | } |
2652 | 0 | #if CONFIG_MULTITHREAD |
2653 | 0 | pthread_mutex_unlock(cdef_sync->mutex_); |
2654 | 0 | #endif // CONFIG_MULTITHREAD |
2655 | 0 | return do_next_block; |
2656 | 0 | } |
2657 | | |
2658 | | // Hook function for each thread in CDEF search multi-threading. |
2659 | 0 | static int cdef_filter_block_worker_hook(void *arg1, void *arg2) { |
2660 | 0 | AV1CdefSync *const cdef_sync = (AV1CdefSync *)arg1; |
2661 | 0 | CdefSearchCtx *cdef_search_ctx = (CdefSearchCtx *)arg2; |
2662 | 0 | int cur_fbr, cur_fbc, sb_count; |
2663 | 0 | while (cdef_get_next_job(cdef_sync, cdef_search_ctx, &cur_fbr, &cur_fbc, |
2664 | 0 | &sb_count)) { |
2665 | 0 | av1_cdef_mse_calc_block(cdef_search_ctx, cur_fbr, cur_fbc, sb_count); |
2666 | 0 | } |
2667 | 0 | return 1; |
2668 | 0 | } |
2669 | | |
2670 | | // Assigns CDEF search hook function and thread data to each worker. |
2671 | | static void prepare_cdef_workers(MultiThreadInfo *mt_info, |
2672 | | CdefSearchCtx *cdef_search_ctx, |
2673 | 0 | AVxWorkerHook hook, int num_workers) { |
2674 | 0 | for (int i = num_workers - 1; i >= 0; i--) { |
2675 | 0 | AVxWorker *worker = &mt_info->workers[i]; |
2676 | 0 | worker->hook = hook; |
2677 | 0 | worker->data1 = &mt_info->cdef_sync; |
2678 | 0 | worker->data2 = cdef_search_ctx; |
2679 | 0 | } |
2680 | 0 | } |
2681 | | |
2682 | | // Implements multi-threading for CDEF search. |
2683 | | void av1_cdef_mse_calc_frame_mt(AV1_COMMON *cm, MultiThreadInfo *mt_info, |
2684 | 0 | CdefSearchCtx *cdef_search_ctx) { |
2685 | 0 | AV1CdefSync *cdef_sync = &mt_info->cdef_sync; |
2686 | 0 | const int num_workers = mt_info->num_mod_workers[MOD_CDEF_SEARCH]; |
2687 | |
|
2688 | 0 | cdef_reset_job_info(cdef_sync); |
2689 | 0 | prepare_cdef_workers(mt_info, cdef_search_ctx, cdef_filter_block_worker_hook, |
2690 | 0 | num_workers); |
2691 | 0 | launch_workers(mt_info, num_workers); |
2692 | 0 | sync_enc_workers(mt_info, cm, num_workers); |
2693 | 0 | } |
2694 | | |
2695 | | // Computes num_workers for temporal filter multi-threading. |
2696 | 0 | static AOM_INLINE int compute_num_tf_workers(AV1_COMP *cpi) { |
2697 | | // For single-pass encode, using no. of workers as per tf block size was not |
2698 | | // found to improve speed. Hence the thread assignment for single-pass encode |
2699 | | // is kept based on compute_num_enc_workers(). |
2700 | 0 | if (cpi->oxcf.pass < AOM_RC_SECOND_PASS) |
2701 | 0 | return (av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads)); |
2702 | | |
2703 | 0 | if (cpi->oxcf.max_threads <= 1) return 1; |
2704 | | |
2705 | 0 | const int frame_height = cpi->common.height; |
2706 | 0 | const BLOCK_SIZE block_size = TF_BLOCK_SIZE; |
2707 | 0 | const int mb_height = block_size_high[block_size]; |
2708 | 0 | const int mb_rows = get_num_blocks(frame_height, mb_height); |
2709 | 0 | return AOMMIN(cpi->oxcf.max_threads, mb_rows); |
2710 | 0 | } |
2711 | | |
2712 | | // Computes num_workers for tpl multi-threading. |
2713 | 0 | static AOM_INLINE int compute_num_tpl_workers(AV1_COMP *cpi) { |
2714 | 0 | return av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads); |
2715 | 0 | } |
2716 | | |
2717 | | // Computes num_workers for loop filter multi-threading. |
2718 | 0 | static AOM_INLINE int compute_num_lf_workers(AV1_COMP *cpi) { |
2719 | 0 | return av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads); |
2720 | 0 | } |
2721 | | |
2722 | | // Computes num_workers for cdef multi-threading. |
2723 | 0 | static AOM_INLINE int compute_num_cdef_workers(AV1_COMP *cpi) { |
2724 | 0 | return av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads); |
2725 | 0 | } |
2726 | | |
2727 | | // Computes num_workers for loop-restoration multi-threading. |
2728 | 0 | static AOM_INLINE int compute_num_lr_workers(AV1_COMP *cpi) { |
2729 | 0 | return av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads); |
2730 | 0 | } |
2731 | | |
2732 | | // Computes num_workers for pack bitstream multi-threading. |
2733 | 0 | static AOM_INLINE int compute_num_pack_bs_workers(AV1_COMP *cpi) { |
2734 | 0 | if (cpi->oxcf.max_threads <= 1) return 1; |
2735 | 0 | return compute_num_enc_tile_mt_workers(&cpi->common, cpi->oxcf.max_threads); |
2736 | 0 | } |
2737 | | |
2738 | 0 | int compute_num_mod_workers(AV1_COMP *cpi, MULTI_THREADED_MODULES mod_name) { |
2739 | 0 | int num_mod_workers = 0; |
2740 | 0 | switch (mod_name) { |
2741 | 0 | case MOD_FP: |
2742 | 0 | if (cpi->oxcf.pass >= AOM_RC_SECOND_PASS) |
2743 | 0 | num_mod_workers = 0; |
2744 | 0 | else |
2745 | 0 | num_mod_workers = |
2746 | 0 | av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads); |
2747 | 0 | break; |
2748 | 0 | case MOD_TF: num_mod_workers = compute_num_tf_workers(cpi); break; |
2749 | 0 | case MOD_TPL: num_mod_workers = compute_num_tpl_workers(cpi); break; |
2750 | 0 | case MOD_GME: num_mod_workers = 1; break; |
2751 | 0 | case MOD_ENC: |
2752 | 0 | num_mod_workers = av1_compute_num_enc_workers(cpi, cpi->oxcf.max_threads); |
2753 | 0 | break; |
2754 | 0 | case MOD_LPF: num_mod_workers = compute_num_lf_workers(cpi); break; |
2755 | 0 | case MOD_CDEF_SEARCH: |
2756 | 0 | num_mod_workers = compute_num_cdef_workers(cpi); |
2757 | 0 | break; |
2758 | 0 | case MOD_CDEF: num_mod_workers = compute_num_cdef_workers(cpi); break; |
2759 | 0 | case MOD_LR: num_mod_workers = compute_num_lr_workers(cpi); break; |
2760 | 0 | case MOD_PACK_BS: num_mod_workers = compute_num_pack_bs_workers(cpi); break; |
2761 | 0 | case MOD_FRAME_ENC: |
2762 | 0 | num_mod_workers = cpi->ppi->p_mt_info.num_mod_workers[MOD_FRAME_ENC]; |
2763 | 0 | break; |
2764 | 0 | default: assert(0); break; |
2765 | 0 | } |
2766 | 0 | return (num_mod_workers); |
2767 | 0 | } |
2768 | | // Computes the number of workers for each MT modules in the encoder |
2769 | 0 | void av1_compute_num_workers_for_mt(AV1_COMP *cpi) { |
2770 | 0 | for (int i = MOD_FP; i < NUM_MT_MODULES; i++) |
2771 | 0 | cpi->ppi->p_mt_info.num_mod_workers[i] = |
2772 | 0 | compute_num_mod_workers(cpi, (MULTI_THREADED_MODULES)i); |
2773 | 0 | } |