Line data Source code
1 : /*
2 : * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
3 : *
4 : * Use of this source code is governed by a BSD-style license
5 : * that can be found in the LICENSE file in the root of the source
6 : * tree. An additional intellectual property rights grant can be found
7 : * in the file PATENTS. All contributing project authors may
8 : * be found in the AUTHORS file in the root of the source tree.
9 : */
10 :
11 : #include "./vpx_config.h"
12 : #include "vpx_dsp/vpx_dsp_common.h"
13 : #include "vpx_mem/vpx_mem.h"
14 : #include "vp9/common/vp9_entropymode.h"
15 : #include "vp9/common/vp9_thread_common.h"
16 : #include "vp9/common/vp9_reconinter.h"
17 : #include "vp9/common/vp9_loopfilter.h"
18 :
19 : #if CONFIG_MULTITHREAD
20 0 : static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
21 0 : const int kMaxTryLocks = 4000;
22 0 : int locked = 0;
23 : int i;
24 :
25 0 : for (i = 0; i < kMaxTryLocks; ++i) {
26 0 : if (!pthread_mutex_trylock(mutex)) {
27 0 : locked = 1;
28 0 : break;
29 : }
30 : }
31 :
32 0 : if (!locked) pthread_mutex_lock(mutex);
33 0 : }
34 : #endif // CONFIG_MULTITHREAD
35 :
36 0 : static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
37 : #if CONFIG_MULTITHREAD
38 0 : const int nsync = lf_sync->sync_range;
39 :
40 0 : if (r && !(c & (nsync - 1))) {
41 0 : pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
42 0 : mutex_lock(mutex);
43 :
44 0 : while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
45 0 : pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
46 : }
47 0 : pthread_mutex_unlock(mutex);
48 : }
49 : #else
50 : (void)lf_sync;
51 : (void)r;
52 : (void)c;
53 : #endif // CONFIG_MULTITHREAD
54 0 : }
55 :
56 0 : static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
57 : const int sb_cols) {
58 : #if CONFIG_MULTITHREAD
59 0 : const int nsync = lf_sync->sync_range;
60 : int cur;
61 : // Only signal when there are enough filtered SB for next row to run.
62 0 : int sig = 1;
63 :
64 0 : if (c < sb_cols - 1) {
65 0 : cur = c;
66 0 : if (c % nsync) sig = 0;
67 : } else {
68 0 : cur = sb_cols + nsync;
69 : }
70 :
71 0 : if (sig) {
72 0 : mutex_lock(&lf_sync->mutex_[r]);
73 :
74 0 : lf_sync->cur_sb_col[r] = cur;
75 :
76 0 : pthread_cond_signal(&lf_sync->cond_[r]);
77 0 : pthread_mutex_unlock(&lf_sync->mutex_[r]);
78 : }
79 : #else
80 : (void)lf_sync;
81 : (void)r;
82 : (void)c;
83 : (void)sb_cols;
84 : #endif // CONFIG_MULTITHREAD
85 0 : }
86 :
87 : // Implement row loopfiltering for each thread.
88 0 : static INLINE void thread_loop_filter_rows(
89 : const YV12_BUFFER_CONFIG *const frame_buffer, VP9_COMMON *const cm,
90 : struct macroblockd_plane planes[MAX_MB_PLANE], int start, int stop,
91 : int y_only, VP9LfSync *const lf_sync) {
92 0 : const int num_planes = y_only ? 1 : MAX_MB_PLANE;
93 0 : const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
94 : int mi_row, mi_col;
95 : enum lf_path path;
96 0 : if (y_only)
97 0 : path = LF_PATH_444;
98 0 : else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
99 0 : path = LF_PATH_420;
100 0 : else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
101 0 : path = LF_PATH_444;
102 : else
103 0 : path = LF_PATH_SLOW;
104 :
105 0 : for (mi_row = start; mi_row < stop;
106 0 : mi_row += lf_sync->num_workers * MI_BLOCK_SIZE) {
107 0 : MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
108 0 : LOOP_FILTER_MASK *lfm = get_lfm(&cm->lf, mi_row, 0);
109 :
110 0 : for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE, ++lfm) {
111 0 : const int r = mi_row >> MI_BLOCK_SIZE_LOG2;
112 0 : const int c = mi_col >> MI_BLOCK_SIZE_LOG2;
113 : int plane;
114 :
115 0 : sync_read(lf_sync, r, c);
116 :
117 0 : vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
118 :
119 0 : vp9_adjust_mask(cm, mi_row, mi_col, lfm);
120 :
121 0 : vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, lfm);
122 0 : for (plane = 1; plane < num_planes; ++plane) {
123 0 : switch (path) {
124 : case LF_PATH_420:
125 0 : vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, lfm);
126 0 : break;
127 : case LF_PATH_444:
128 0 : vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, lfm);
129 0 : break;
130 : case LF_PATH_SLOW:
131 0 : vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
132 : mi_row, mi_col);
133 0 : break;
134 : }
135 : }
136 :
137 0 : sync_write(lf_sync, r, c, sb_cols);
138 : }
139 : }
140 0 : }
141 :
142 : // Row-based multi-threaded loopfilter hook
143 0 : static int loop_filter_row_worker(VP9LfSync *const lf_sync,
144 : LFWorkerData *const lf_data) {
145 0 : thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
146 : lf_data->start, lf_data->stop, lf_data->y_only,
147 : lf_sync);
148 0 : return 1;
149 : }
150 :
151 0 : static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame, VP9_COMMON *cm,
152 : struct macroblockd_plane planes[MAX_MB_PLANE],
153 : int start, int stop, int y_only,
154 : VPxWorker *workers, int nworkers,
155 : VP9LfSync *lf_sync) {
156 0 : const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
157 : // Number of superblock rows and cols
158 0 : const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
159 : // Decoder may allocate more threads than number of tiles based on user's
160 : // input.
161 0 : const int tile_cols = 1 << cm->log2_tile_cols;
162 0 : const int num_workers = VPXMIN(nworkers, tile_cols);
163 : int i;
164 :
165 0 : if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
166 0 : num_workers > lf_sync->num_workers) {
167 0 : vp9_loop_filter_dealloc(lf_sync);
168 0 : vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
169 : }
170 :
171 : // Initialize cur_sb_col to -1 for all SB rows.
172 0 : memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
173 :
174 : // Set up loopfilter thread data.
175 : // The decoder is capping num_workers because it has been observed that using
176 : // more threads on the loopfilter than there are cores will hurt performance
177 : // on Android. This is because the system will only schedule the tile decode
178 : // workers on cores equal to the number of tile columns. Then if the decoder
179 : // tries to use more threads for the loopfilter, it will hurt performance
180 : // because of contention. If the multithreading code changes in the future
181 : // then the number of workers used by the loopfilter should be revisited.
182 0 : for (i = 0; i < num_workers; ++i) {
183 0 : VPxWorker *const worker = &workers[i];
184 0 : LFWorkerData *const lf_data = &lf_sync->lfdata[i];
185 :
186 0 : worker->hook = (VPxWorkerHook)loop_filter_row_worker;
187 0 : worker->data1 = lf_sync;
188 0 : worker->data2 = lf_data;
189 :
190 : // Loopfilter data
191 0 : vp9_loop_filter_data_reset(lf_data, frame, cm, planes);
192 0 : lf_data->start = start + i * MI_BLOCK_SIZE;
193 0 : lf_data->stop = stop;
194 0 : lf_data->y_only = y_only;
195 :
196 : // Start loopfiltering
197 0 : if (i == num_workers - 1) {
198 0 : winterface->execute(worker);
199 : } else {
200 0 : winterface->launch(worker);
201 : }
202 : }
203 :
204 : // Wait till all rows are finished
205 0 : for (i = 0; i < num_workers; ++i) {
206 0 : winterface->sync(&workers[i]);
207 : }
208 0 : }
209 :
210 0 : void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame, VP9_COMMON *cm,
211 : struct macroblockd_plane planes[MAX_MB_PLANE],
212 : int frame_filter_level, int y_only,
213 : int partial_frame, VPxWorker *workers,
214 : int num_workers, VP9LfSync *lf_sync) {
215 : int start_mi_row, end_mi_row, mi_rows_to_filter;
216 :
217 0 : if (!frame_filter_level) return;
218 :
219 0 : start_mi_row = 0;
220 0 : mi_rows_to_filter = cm->mi_rows;
221 0 : if (partial_frame && cm->mi_rows > 8) {
222 0 : start_mi_row = cm->mi_rows >> 1;
223 0 : start_mi_row &= 0xfffffff8;
224 0 : mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
225 : }
226 0 : end_mi_row = start_mi_row + mi_rows_to_filter;
227 0 : vp9_loop_filter_frame_init(cm, frame_filter_level);
228 :
229 0 : loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row, y_only,
230 : workers, num_workers, lf_sync);
231 : }
232 :
233 : // Set up nsync by width.
234 0 : static INLINE int get_sync_range(int width) {
235 : // nsync numbers are picked by testing. For example, for 4k
236 : // video, using 4 gives best performance.
237 0 : if (width < 640)
238 0 : return 1;
239 0 : else if (width <= 1280)
240 0 : return 2;
241 0 : else if (width <= 4096)
242 0 : return 4;
243 : else
244 0 : return 8;
245 : }
246 :
247 : // Allocate memory for lf row synchronization
248 0 : void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
249 : int width, int num_workers) {
250 0 : lf_sync->rows = rows;
251 : #if CONFIG_MULTITHREAD
252 : {
253 : int i;
254 :
255 0 : CHECK_MEM_ERROR(cm, lf_sync->mutex_,
256 : vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
257 0 : if (lf_sync->mutex_) {
258 0 : for (i = 0; i < rows; ++i) {
259 0 : pthread_mutex_init(&lf_sync->mutex_[i], NULL);
260 : }
261 : }
262 :
263 0 : CHECK_MEM_ERROR(cm, lf_sync->cond_,
264 : vpx_malloc(sizeof(*lf_sync->cond_) * rows));
265 0 : if (lf_sync->cond_) {
266 0 : for (i = 0; i < rows; ++i) {
267 0 : pthread_cond_init(&lf_sync->cond_[i], NULL);
268 : }
269 : }
270 : }
271 : #endif // CONFIG_MULTITHREAD
272 :
273 0 : CHECK_MEM_ERROR(cm, lf_sync->lfdata,
274 : vpx_malloc(num_workers * sizeof(*lf_sync->lfdata)));
275 0 : lf_sync->num_workers = num_workers;
276 :
277 0 : CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
278 : vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
279 :
280 : // Set up nsync.
281 0 : lf_sync->sync_range = get_sync_range(width);
282 0 : }
283 :
284 : // Deallocate lf synchronization related mutex and data
285 0 : void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) {
286 0 : if (lf_sync != NULL) {
287 : #if CONFIG_MULTITHREAD
288 : int i;
289 :
290 0 : if (lf_sync->mutex_ != NULL) {
291 0 : for (i = 0; i < lf_sync->rows; ++i) {
292 0 : pthread_mutex_destroy(&lf_sync->mutex_[i]);
293 : }
294 0 : vpx_free(lf_sync->mutex_);
295 : }
296 0 : if (lf_sync->cond_ != NULL) {
297 0 : for (i = 0; i < lf_sync->rows; ++i) {
298 0 : pthread_cond_destroy(&lf_sync->cond_[i]);
299 : }
300 0 : vpx_free(lf_sync->cond_);
301 : }
302 : #endif // CONFIG_MULTITHREAD
303 0 : vpx_free(lf_sync->lfdata);
304 0 : vpx_free(lf_sync->cur_sb_col);
305 : // clear the structure as the source of this call may be a resize in which
306 : // case this call will be followed by an _alloc() which may fail.
307 0 : vp9_zero(*lf_sync);
308 : }
309 0 : }
310 :
311 : // Accumulate frame counts.
312 0 : void vp9_accumulate_frame_counts(FRAME_COUNTS *accum,
313 : const FRAME_COUNTS *counts, int is_dec) {
314 : int i, j, k, l, m;
315 :
316 0 : for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
317 0 : for (j = 0; j < INTRA_MODES; j++)
318 0 : accum->y_mode[i][j] += counts->y_mode[i][j];
319 :
320 0 : for (i = 0; i < INTRA_MODES; i++)
321 0 : for (j = 0; j < INTRA_MODES; j++)
322 0 : accum->uv_mode[i][j] += counts->uv_mode[i][j];
323 :
324 0 : for (i = 0; i < PARTITION_CONTEXTS; i++)
325 0 : for (j = 0; j < PARTITION_TYPES; j++)
326 0 : accum->partition[i][j] += counts->partition[i][j];
327 :
328 0 : if (is_dec) {
329 : int n;
330 0 : for (i = 0; i < TX_SIZES; i++)
331 0 : for (j = 0; j < PLANE_TYPES; j++)
332 0 : for (k = 0; k < REF_TYPES; k++)
333 0 : for (l = 0; l < COEF_BANDS; l++)
334 0 : for (m = 0; m < COEFF_CONTEXTS; m++) {
335 0 : accum->eob_branch[i][j][k][l][m] +=
336 0 : counts->eob_branch[i][j][k][l][m];
337 0 : for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
338 0 : accum->coef[i][j][k][l][m][n] += counts->coef[i][j][k][l][m][n];
339 : }
340 : } else {
341 0 : for (i = 0; i < TX_SIZES; i++)
342 0 : for (j = 0; j < PLANE_TYPES; j++)
343 0 : for (k = 0; k < REF_TYPES; k++)
344 0 : for (l = 0; l < COEF_BANDS; l++)
345 0 : for (m = 0; m < COEFF_CONTEXTS; m++)
346 0 : accum->eob_branch[i][j][k][l][m] +=
347 0 : counts->eob_branch[i][j][k][l][m];
348 : // In the encoder, coef is only updated at frame
349 : // level, so not need to accumulate it here.
350 : // for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
351 : // accum->coef[i][j][k][l][m][n] +=
352 : // counts->coef[i][j][k][l][m][n];
353 : }
354 :
355 0 : for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
356 0 : for (j = 0; j < SWITCHABLE_FILTERS; j++)
357 0 : accum->switchable_interp[i][j] += counts->switchable_interp[i][j];
358 :
359 0 : for (i = 0; i < INTER_MODE_CONTEXTS; i++)
360 0 : for (j = 0; j < INTER_MODES; j++)
361 0 : accum->inter_mode[i][j] += counts->inter_mode[i][j];
362 :
363 0 : for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
364 0 : for (j = 0; j < 2; j++)
365 0 : accum->intra_inter[i][j] += counts->intra_inter[i][j];
366 :
367 0 : for (i = 0; i < COMP_INTER_CONTEXTS; i++)
368 0 : for (j = 0; j < 2; j++) accum->comp_inter[i][j] += counts->comp_inter[i][j];
369 :
370 0 : for (i = 0; i < REF_CONTEXTS; i++)
371 0 : for (j = 0; j < 2; j++)
372 0 : for (k = 0; k < 2; k++)
373 0 : accum->single_ref[i][j][k] += counts->single_ref[i][j][k];
374 :
375 0 : for (i = 0; i < REF_CONTEXTS; i++)
376 0 : for (j = 0; j < 2; j++) accum->comp_ref[i][j] += counts->comp_ref[i][j];
377 :
378 0 : for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
379 0 : for (j = 0; j < TX_SIZES; j++)
380 0 : accum->tx.p32x32[i][j] += counts->tx.p32x32[i][j];
381 :
382 0 : for (j = 0; j < TX_SIZES - 1; j++)
383 0 : accum->tx.p16x16[i][j] += counts->tx.p16x16[i][j];
384 :
385 0 : for (j = 0; j < TX_SIZES - 2; j++)
386 0 : accum->tx.p8x8[i][j] += counts->tx.p8x8[i][j];
387 : }
388 :
389 0 : for (i = 0; i < TX_SIZES; i++)
390 0 : accum->tx.tx_totals[i] += counts->tx.tx_totals[i];
391 :
392 0 : for (i = 0; i < SKIP_CONTEXTS; i++)
393 0 : for (j = 0; j < 2; j++) accum->skip[i][j] += counts->skip[i][j];
394 :
395 0 : for (i = 0; i < MV_JOINTS; i++) accum->mv.joints[i] += counts->mv.joints[i];
396 :
397 0 : for (k = 0; k < 2; k++) {
398 0 : nmv_component_counts *const comps = &accum->mv.comps[k];
399 0 : const nmv_component_counts *const comps_t = &counts->mv.comps[k];
400 :
401 0 : for (i = 0; i < 2; i++) {
402 0 : comps->sign[i] += comps_t->sign[i];
403 0 : comps->class0_hp[i] += comps_t->class0_hp[i];
404 0 : comps->hp[i] += comps_t->hp[i];
405 : }
406 :
407 0 : for (i = 0; i < MV_CLASSES; i++) comps->classes[i] += comps_t->classes[i];
408 :
409 0 : for (i = 0; i < CLASS0_SIZE; i++) {
410 0 : comps->class0[i] += comps_t->class0[i];
411 0 : for (j = 0; j < MV_FP_SIZE; j++)
412 0 : comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
413 : }
414 :
415 0 : for (i = 0; i < MV_OFFSET_BITS; i++)
416 0 : for (j = 0; j < 2; j++) comps->bits[i][j] += comps_t->bits[i][j];
417 :
418 0 : for (i = 0; i < MV_FP_SIZE; i++) comps->fp[i] += comps_t->fp[i];
419 : }
420 0 : }
|