2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
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.
11 #include "./vpx_config.h"
12 #include "vpx_dsp/vpx_dsp_common.h"
13 #include "vpx_mem/vpx_mem.h"
14 #include "vp10/common/entropymode.h"
15 #include "vp10/common/thread_common.h"
16 #include "vp10/common/reconinter.h"
17 #include "vp10/common/loopfilter.h"
19 #if CONFIG_MULTITHREAD
20 static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
21 const int kMaxTryLocks = 4000;
25 for (i = 0; i < kMaxTryLocks; ++i) {
26 if (!pthread_mutex_trylock(mutex)) {
33 pthread_mutex_lock(mutex);
35 #endif // CONFIG_MULTITHREAD
37 static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
38 #if CONFIG_MULTITHREAD
39 const int nsync = lf_sync->sync_range;
41 if (r && !(c & (nsync - 1))) {
42 pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
45 while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
46 pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
48 pthread_mutex_unlock(mutex);
54 #endif // CONFIG_MULTITHREAD
57 static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
59 #if CONFIG_MULTITHREAD
60 const int nsync = lf_sync->sync_range;
62 // Only signal when there are enough filtered SB for next row to run.
65 if (c < sb_cols - 1) {
70 cur = sb_cols + nsync;
74 mutex_lock(&lf_sync->mutex_[r]);
76 lf_sync->cur_sb_col[r] = cur;
78 pthread_cond_signal(&lf_sync->cond_[r]);
79 pthread_mutex_unlock(&lf_sync->mutex_[r]);
86 #endif // CONFIG_MULTITHREAD
89 // Implement row loopfiltering for each thread.
91 void thread_loop_filter_rows(const YV12_BUFFER_CONFIG *const frame_buffer,
92 VP10_COMMON *const cm,
93 struct macroblockd_plane planes[MAX_MB_PLANE],
94 int start, int stop, int y_only,
95 VP9LfSync *const lf_sync) {
96 const int num_planes = y_only ? 1 : MAX_MB_PLANE;
97 const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
102 else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
104 else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
109 for (mi_row = start; mi_row < stop;
110 mi_row += lf_sync->num_workers * MI_BLOCK_SIZE) {
111 MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
113 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
114 const int r = mi_row >> MI_BLOCK_SIZE_LOG2;
115 const int c = mi_col >> MI_BLOCK_SIZE_LOG2;
116 LOOP_FILTER_MASK lfm;
119 sync_read(lf_sync, r, c);
121 vp10_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
123 // TODO(JBB): Make setup_mask work for non 420.
124 vp10_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
127 vp10_filter_block_plane_ss00(cm, &planes[0], mi_row, &lfm);
128 for (plane = 1; plane < num_planes; ++plane) {
131 vp10_filter_block_plane_ss11(cm, &planes[plane], mi_row, &lfm);
134 vp10_filter_block_plane_ss00(cm, &planes[plane], mi_row, &lfm);
137 vp10_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
143 sync_write(lf_sync, r, c, sb_cols);
148 // Row-based multi-threaded loopfilter hook
149 static int loop_filter_row_worker(VP9LfSync *const lf_sync,
150 LFWorkerData *const lf_data) {
151 thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
152 lf_data->start, lf_data->stop, lf_data->y_only,
157 static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame,
159 struct macroblockd_plane planes[MAX_MB_PLANE],
160 int start, int stop, int y_only,
161 VPxWorker *workers, int nworkers,
162 VP9LfSync *lf_sync) {
163 const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
164 // Number of superblock rows and cols
165 const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
166 // Decoder may allocate more threads than number of tiles based on user's
168 const int tile_cols = 1 << cm->log2_tile_cols;
169 const int num_workers = VPXMIN(nworkers, tile_cols);
172 if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
173 num_workers > lf_sync->num_workers) {
174 vp10_loop_filter_dealloc(lf_sync);
175 vp10_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
178 // Initialize cur_sb_col to -1 for all SB rows.
179 memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
181 // Set up loopfilter thread data.
182 // The decoder is capping num_workers because it has been observed that using
183 // more threads on the loopfilter than there are cores will hurt performance
184 // on Android. This is because the system will only schedule the tile decode
185 // workers on cores equal to the number of tile columns. Then if the decoder
186 // tries to use more threads for the loopfilter, it will hurt performance
187 // because of contention. If the multithreading code changes in the future
188 // then the number of workers used by the loopfilter should be revisited.
189 for (i = 0; i < num_workers; ++i) {
190 VPxWorker *const worker = &workers[i];
191 LFWorkerData *const lf_data = &lf_sync->lfdata[i];
193 worker->hook = (VPxWorkerHook)loop_filter_row_worker;
194 worker->data1 = lf_sync;
195 worker->data2 = lf_data;
198 vp10_loop_filter_data_reset(lf_data, frame, cm, planes);
199 lf_data->start = start + i * MI_BLOCK_SIZE;
200 lf_data->stop = stop;
201 lf_data->y_only = y_only;
203 // Start loopfiltering
204 if (i == num_workers - 1) {
205 winterface->execute(worker);
207 winterface->launch(worker);
211 // Wait till all rows are finished
212 for (i = 0; i < num_workers; ++i) {
213 winterface->sync(&workers[i]);
217 void vp10_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
219 struct macroblockd_plane planes[MAX_MB_PLANE],
220 int frame_filter_level,
221 int y_only, int partial_frame,
222 VPxWorker *workers, int num_workers,
223 VP9LfSync *lf_sync) {
224 int start_mi_row, end_mi_row, mi_rows_to_filter;
226 if (!frame_filter_level) return;
229 mi_rows_to_filter = cm->mi_rows;
230 if (partial_frame && cm->mi_rows > 8) {
231 start_mi_row = cm->mi_rows >> 1;
232 start_mi_row &= 0xfffffff8;
233 mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
235 end_mi_row = start_mi_row + mi_rows_to_filter;
236 vp10_loop_filter_frame_init(cm, frame_filter_level);
238 loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row,
239 y_only, workers, num_workers, lf_sync);
242 // Set up nsync by width.
243 static INLINE int get_sync_range(int width) {
244 // nsync numbers are picked by testing. For example, for 4k
245 // video, using 4 gives best performance.
248 else if (width <= 1280)
250 else if (width <= 4096)
256 // Allocate memory for lf row synchronization
257 void vp10_loop_filter_alloc(VP9LfSync *lf_sync, VP10_COMMON *cm, int rows,
258 int width, int num_workers) {
259 lf_sync->rows = rows;
260 #if CONFIG_MULTITHREAD
264 CHECK_MEM_ERROR(cm, lf_sync->mutex_,
265 vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
266 if (lf_sync->mutex_) {
267 for (i = 0; i < rows; ++i) {
268 pthread_mutex_init(&lf_sync->mutex_[i], NULL);
272 CHECK_MEM_ERROR(cm, lf_sync->cond_,
273 vpx_malloc(sizeof(*lf_sync->cond_) * rows));
274 if (lf_sync->cond_) {
275 for (i = 0; i < rows; ++i) {
276 pthread_cond_init(&lf_sync->cond_[i], NULL);
280 #endif // CONFIG_MULTITHREAD
282 CHECK_MEM_ERROR(cm, lf_sync->lfdata,
283 vpx_malloc(num_workers * sizeof(*lf_sync->lfdata)));
284 lf_sync->num_workers = num_workers;
286 CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
287 vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
290 lf_sync->sync_range = get_sync_range(width);
293 // Deallocate lf synchronization related mutex and data
294 void vp10_loop_filter_dealloc(VP9LfSync *lf_sync) {
295 if (lf_sync != NULL) {
296 #if CONFIG_MULTITHREAD
299 if (lf_sync->mutex_ != NULL) {
300 for (i = 0; i < lf_sync->rows; ++i) {
301 pthread_mutex_destroy(&lf_sync->mutex_[i]);
303 vpx_free(lf_sync->mutex_);
305 if (lf_sync->cond_ != NULL) {
306 for (i = 0; i < lf_sync->rows; ++i) {
307 pthread_cond_destroy(&lf_sync->cond_[i]);
309 vpx_free(lf_sync->cond_);
311 #endif // CONFIG_MULTITHREAD
312 vpx_free(lf_sync->lfdata);
313 vpx_free(lf_sync->cur_sb_col);
314 // clear the structure as the source of this call may be a resize in which
315 // case this call will be followed by an _alloc() which may fail.
320 // Accumulate frame counts.
321 void vp10_accumulate_frame_counts(VP10_COMMON *cm, FRAME_COUNTS *counts,
325 for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
326 for (j = 0; j < INTRA_MODES; j++)
327 cm->counts.y_mode[i][j] += counts->y_mode[i][j];
329 for (i = 0; i < INTRA_MODES; i++)
330 for (j = 0; j < INTRA_MODES; j++)
331 cm->counts.uv_mode[i][j] += counts->uv_mode[i][j];
333 for (i = 0; i < PARTITION_CONTEXTS; i++)
334 for (j = 0; j < PARTITION_TYPES; j++)
335 cm->counts.partition[i][j] += counts->partition[i][j];
339 for (i = 0; i < TX_SIZES; i++)
340 for (j = 0; j < PLANE_TYPES; j++)
341 for (k = 0; k < REF_TYPES; k++)
342 for (l = 0; l < COEF_BANDS; l++)
343 for (m = 0; m < COEFF_CONTEXTS; m++) {
344 cm->counts.eob_branch[i][j][k][l][m] +=
345 counts->eob_branch[i][j][k][l][m];
346 for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
347 cm->counts.coef[i][j][k][l][m][n] +=
348 counts->coef[i][j][k][l][m][n];
351 for (i = 0; i < TX_SIZES; i++)
352 for (j = 0; j < PLANE_TYPES; j++)
353 for (k = 0; k < REF_TYPES; k++)
354 for (l = 0; l < COEF_BANDS; l++)
355 for (m = 0; m < COEFF_CONTEXTS; m++)
356 cm->counts.eob_branch[i][j][k][l][m] +=
357 counts->eob_branch[i][j][k][l][m];
358 // In the encoder, cm->counts.coef is only updated at frame
359 // level, so not need to accumulate it here.
360 // for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
361 // cm->counts.coef[i][j][k][l][m][n] +=
362 // counts->coef[i][j][k][l][m][n];
365 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
366 for (j = 0; j < SWITCHABLE_FILTERS; j++)
367 cm->counts.switchable_interp[i][j] += counts->switchable_interp[i][j];
369 for (i = 0; i < INTER_MODE_CONTEXTS; i++)
370 for (j = 0; j < INTER_MODES; j++)
371 cm->counts.inter_mode[i][j] += counts->inter_mode[i][j];
373 for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
374 for (j = 0; j < 2; j++)
375 cm->counts.intra_inter[i][j] += counts->intra_inter[i][j];
377 for (i = 0; i < COMP_INTER_CONTEXTS; i++)
378 for (j = 0; j < 2; j++)
379 cm->counts.comp_inter[i][j] += counts->comp_inter[i][j];
381 for (i = 0; i < REF_CONTEXTS; i++)
382 for (j = 0; j < 2; j++)
383 for (k = 0; k < 2; k++)
384 cm->counts.single_ref[i][j][k] += counts->single_ref[i][j][k];
386 for (i = 0; i < REF_CONTEXTS; i++)
387 for (j = 0; j < 2; j++)
388 cm->counts.comp_ref[i][j] += counts->comp_ref[i][j];
390 for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
391 for (j = 0; j < TX_SIZES; j++)
392 cm->counts.tx.p32x32[i][j] += counts->tx.p32x32[i][j];
394 for (j = 0; j < TX_SIZES - 1; j++)
395 cm->counts.tx.p16x16[i][j] += counts->tx.p16x16[i][j];
397 for (j = 0; j < TX_SIZES - 2; j++)
398 cm->counts.tx.p8x8[i][j] += counts->tx.p8x8[i][j];
401 for (i = 0; i < TX_SIZES; i++)
402 cm->counts.tx.tx_totals[i] += counts->tx.tx_totals[i];
404 for (i = 0; i < SKIP_CONTEXTS; i++)
405 for (j = 0; j < 2; j++)
406 cm->counts.skip[i][j] += counts->skip[i][j];
408 for (i = 0; i < MV_JOINTS; i++)
409 cm->counts.mv.joints[i] += counts->mv.joints[i];
411 for (k = 0; k < 2; k++) {
412 nmv_component_counts *comps = &cm->counts.mv.comps[k];
413 nmv_component_counts *comps_t = &counts->mv.comps[k];
415 for (i = 0; i < 2; i++) {
416 comps->sign[i] += comps_t->sign[i];
417 comps->class0_hp[i] += comps_t->class0_hp[i];
418 comps->hp[i] += comps_t->hp[i];
421 for (i = 0; i < MV_CLASSES; i++)
422 comps->classes[i] += comps_t->classes[i];
424 for (i = 0; i < CLASS0_SIZE; i++) {
425 comps->class0[i] += comps_t->class0[i];
426 for (j = 0; j < MV_FP_SIZE; j++)
427 comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
430 for (i = 0; i < MV_OFFSET_BITS; i++)
431 for (j = 0; j < 2; j++)
432 comps->bits[i][j] += comps_t->bits[i][j];
434 for (i = 0; i < MV_FP_SIZE; i++)
435 comps->fp[i] += comps_t->fp[i];