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"
13 #include "vpx_mem/vpx_mem.h"
15 #include "vp9/common/vp9_reconinter.h"
17 #include "vp9/decoder/vp9_dthread.h"
18 #include "vp9/decoder/vp9_decoder.h"
20 #if CONFIG_MULTITHREAD
21 static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
22 const int kMaxTryLocks = 4000;
26 for (i = 0; i < kMaxTryLocks; ++i) {
27 if (!pthread_mutex_trylock(mutex)) {
34 pthread_mutex_lock(mutex);
36 #endif // CONFIG_MULTITHREAD
38 static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
39 #if CONFIG_MULTITHREAD
40 const int nsync = lf_sync->sync_range;
42 if (r && !(c & (nsync - 1))) {
43 pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
46 while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
47 pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
49 pthread_mutex_unlock(mutex);
55 #endif // CONFIG_MULTITHREAD
58 static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
60 #if CONFIG_MULTITHREAD
61 const int nsync = lf_sync->sync_range;
63 // Only signal when there are enough filtered SB for next row to run.
66 if (c < sb_cols - 1) {
71 cur = sb_cols + nsync;
75 mutex_lock(&lf_sync->mutex_[r]);
77 lf_sync->cur_sb_col[r] = cur;
79 pthread_cond_signal(&lf_sync->cond_[r]);
80 pthread_mutex_unlock(&lf_sync->mutex_[r]);
87 #endif // CONFIG_MULTITHREAD
90 // Implement row loopfiltering for each thread.
91 static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer,
93 struct macroblockd_plane planes[MAX_MB_PLANE],
94 int start, int stop, int y_only,
95 VP9LfSync *const lf_sync, int num_lf_workers) {
96 const int num_planes = y_only ? 1 : MAX_MB_PLANE;
97 int r, c; // SB row and col
98 const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
100 for (r = start; r < stop; r += num_lf_workers) {
101 const int mi_row = r << MI_BLOCK_SIZE_LOG2;
102 MODE_INFO *const mi = cm->mi + mi_row * cm->mi_stride;
104 for (c = 0; c < sb_cols; ++c) {
105 const int mi_col = c << MI_BLOCK_SIZE_LOG2;
106 LOOP_FILTER_MASK lfm;
109 sync_read(lf_sync, r, c);
111 vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
112 vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm);
114 for (plane = 0; plane < num_planes; ++plane) {
115 vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
118 sync_write(lf_sync, r, c, sb_cols);
123 // Row-based multi-threaded loopfilter hook
124 static int loop_filter_row_worker(TileWorkerData *const tile_data,
126 LFWorkerData *const lf_data = &tile_data->lfdata;
128 loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
129 lf_data->start, lf_data->stop, lf_data->y_only,
130 lf_data->lf_sync, lf_data->num_lf_workers);
134 // VP9 decoder: Implement multi-threaded loopfilter that uses the tile
136 void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
137 VP9Decoder *pbi, VP9_COMMON *cm,
138 int frame_filter_level,
140 VP9LfSync *const lf_sync = &pbi->lf_row_sync;
141 const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
142 // Number of superblock rows and cols
143 const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
144 const int tile_cols = 1 << cm->log2_tile_cols;
145 const int num_workers = MIN(pbi->max_threads & ~1, tile_cols);
148 if (!frame_filter_level) return;
150 if (!lf_sync->sync_range || cm->last_height != cm->height) {
151 vp9_loop_filter_dealloc(lf_sync);
152 vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width);
155 vp9_loop_filter_frame_init(cm, frame_filter_level);
157 // Initialize cur_sb_col to -1 for all SB rows.
158 vpx_memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
160 // Set up loopfilter thread data.
161 // The decoder is using num_workers instead of pbi->num_tile_workers
162 // because it has been observed that using more threads on the
163 // loopfilter, than there are tile columns in the frame will hurt
164 // performance on Android. This is because the system will only
165 // schedule the tile decode workers on cores equal to the number
166 // of tile columns. Then if the decoder tries to use more threads for the
167 // loopfilter, it will hurt performance because of contention. If the
168 // multithreading code changes in the future then the number of workers
169 // used by the loopfilter should be revisited.
170 for (i = 0; i < num_workers; ++i) {
171 VP9Worker *const worker = &pbi->tile_workers[i];
172 TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
173 LFWorkerData *const lf_data = &tile_data->lfdata;
175 worker->hook = (VP9WorkerHook)loop_filter_row_worker;
178 lf_data->frame_buffer = frame;
180 vp9_copy(lf_data->planes, pbi->mb.plane);
182 lf_data->stop = sb_rows;
183 lf_data->y_only = y_only; // always do all planes in decoder
185 lf_data->lf_sync = lf_sync;
186 lf_data->num_lf_workers = num_workers;
188 // Start loopfiltering
189 if (i == num_workers - 1) {
190 winterface->execute(worker);
192 winterface->launch(worker);
196 // Wait till all rows are finished
197 for (i = 0; i < num_workers; ++i) {
198 winterface->sync(&pbi->tile_workers[i]);
202 // Set up nsync by width.
203 static int get_sync_range(int width) {
204 // nsync numbers are picked by testing. For example, for 4k
205 // video, using 4 gives best performance.
208 else if (width <= 1280)
210 else if (width <= 4096)
216 // Allocate memory for lf row synchronization
217 void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
219 lf_sync->rows = rows;
220 #if CONFIG_MULTITHREAD
224 CHECK_MEM_ERROR(cm, lf_sync->mutex_,
225 vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
226 if (lf_sync->mutex_) {
227 for (i = 0; i < rows; ++i) {
228 pthread_mutex_init(&lf_sync->mutex_[i], NULL);
232 CHECK_MEM_ERROR(cm, lf_sync->cond_,
233 vpx_malloc(sizeof(*lf_sync->cond_) * rows));
234 if (lf_sync->cond_) {
235 for (i = 0; i < rows; ++i) {
236 pthread_cond_init(&lf_sync->cond_[i], NULL);
240 #endif // CONFIG_MULTITHREAD
242 CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
243 vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
246 lf_sync->sync_range = get_sync_range(width);
249 // Deallocate lf synchronization related mutex and data
250 void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) {
251 if (lf_sync != NULL) {
252 #if CONFIG_MULTITHREAD
255 if (lf_sync->mutex_ != NULL) {
256 for (i = 0; i < lf_sync->rows; ++i) {
257 pthread_mutex_destroy(&lf_sync->mutex_[i]);
259 vpx_free(lf_sync->mutex_);
261 if (lf_sync->cond_ != NULL) {
262 for (i = 0; i < lf_sync->rows; ++i) {
263 pthread_cond_destroy(&lf_sync->cond_[i]);
265 vpx_free(lf_sync->cond_);
267 #endif // CONFIG_MULTITHREAD
268 vpx_free(lf_sync->cur_sb_col);
269 // clear the structure as the source of this call may be a resize in which
270 // case this call will be followed by an _alloc() which may fail.