*/
#include "./vpx_config.h"
+
+#include "vpx_mem/vpx_mem.h"
+
#include "vp9/common/vp9_reconinter.h"
+
#include "vp9/decoder/vp9_dthread.h"
-#include "vp9/decoder/vp9_onyxd_int.h"
-#include "vpx_mem/vpx_mem.h"
+#include "vp9/decoder/vp9_decoder.h"
#if CONFIG_MULTITHREAD
static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
const int nsync = lf_sync->sync_range;
if (r && !(c & (nsync - 1))) {
- mutex_lock(&lf_sync->mutex_[r - 1]);
+ pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
+ mutex_lock(mutex);
while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
- pthread_cond_wait(&lf_sync->cond_[r - 1],
- &lf_sync->mutex_[r - 1]);
+ pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
}
- pthread_mutex_unlock(&lf_sync->mutex_[r - 1]);
+ pthread_mutex_unlock(mutex);
}
#else
(void)lf_sync;
VP9LfSync *const lf_sync, int num_lf_workers) {
const int num_planes = y_only ? 1 : MAX_MB_PLANE;
int r, c; // SB row and col
- LOOP_FILTER_MASK lfm;
const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
for (r = start; r < stop; r += num_lf_workers) {
const int mi_row = r << MI_BLOCK_SIZE_LOG2;
- MODE_INFO **mi_8x8 = cm->mi_grid_visible + mi_row * cm->mode_info_stride;
+ MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
for (c = 0; c < sb_cols; ++c) {
const int mi_col = c << MI_BLOCK_SIZE_LOG2;
+ LOOP_FILTER_MASK lfm;
int plane;
sync_read(lf_sync, r, c);
- setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
- vp9_setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mode_info_stride,
- &lfm);
+ vp9_setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
+ vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm);
for (plane = 0; plane < num_planes; ++plane) {
vp9_filter_block_plane(cm, &xd->plane[plane], mi_row, &lfm);
// VP9 decoder: Implement multi-threaded loopfilter that uses the tile
// threads.
-void vp9_loop_filter_frame_mt(VP9D_COMP *pbi,
+void vp9_loop_filter_frame_mt(VP9Decoder *pbi,
VP9_COMMON *cm,
- MACROBLOCKD *xd,
int frame_filter_level,
int y_only, int partial_frame) {
+ VP9LfSync *const lf_sync = &pbi->lf_row_sync;
// Number of superblock rows and cols
const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
+ const int tile_cols = 1 << cm->log2_tile_cols;
+ const int num_workers = MIN(pbi->oxcf.max_threads & ~1, tile_cols);
int i;
// Allocate memory used in thread synchronization.
// This always needs to be done even if frame_filter_level is 0.
if (!cm->current_video_frame || cm->last_height != cm->height) {
- VP9LfSync *const lf_sync = &pbi->lf_row_sync;
-
if (cm->last_height != cm->height) {
const int aligned_last_height =
ALIGN_POWER_OF_TWO(cm->last_height, MI_SIZE_LOG2);
vp9_loop_filter_frame_init(cm, frame_filter_level);
// Initialize cur_sb_col to -1 for all SB rows.
- vpx_memset(pbi->lf_row_sync.cur_sb_col, -1,
- sizeof(*pbi->lf_row_sync.cur_sb_col) * sb_rows);
+ vpx_memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
// Set up loopfilter thread data.
- for (i = 0; i < pbi->num_tile_workers; ++i) {
+ // The decoder is using num_workers instead of pbi->num_tile_workers
+ // because it has been observed that using more threads on the
+ // loopfilter, than there are tile columns in the frame will hurt
+ // performance on Android. This is because the system will only
+ // schedule the tile decode workers on cores equal to the number
+ // of tile columns. Then if the decoder tries to use more threads for the
+ // loopfilter, it will hurt performance because of contention. If the
+ // multithreading code changes in the future then the number of workers
+ // used by the loopfilter should be revisited.
+ for (i = 0; i < num_workers; ++i) {
VP9Worker *const worker = &pbi->tile_workers[i];
TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
LFWorkerData *const lf_data = &tile_data->lfdata;
lf_data->stop = sb_rows;
lf_data->y_only = y_only; // always do all planes in decoder
- lf_data->lf_sync = &pbi->lf_row_sync;
- lf_data->num_lf_workers = pbi->num_tile_workers;
+ lf_data->lf_sync = lf_sync;
+ lf_data->num_lf_workers = num_workers;
// Start loopfiltering
- if (i == pbi->num_tile_workers - 1) {
+ if (i == num_workers - 1) {
vp9_worker_execute(worker);
} else {
vp9_worker_launch(worker);
}
// Wait till all rows are finished
- for (i = 0; i < pbi->num_tile_workers; ++i) {
+ for (i = 0; i < num_workers; ++i) {
vp9_worker_sync(&pbi->tile_workers[i]);
}
}
// Deallocate lf synchronization related mutex and data
void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows) {
-#if CONFIG_MULTITHREAD
+#if !CONFIG_MULTITHREAD
+ (void)rows;
+#endif // !CONFIG_MULTITHREAD
+
if (lf_sync != NULL) {
+#if CONFIG_MULTITHREAD
int i;
if (lf_sync->mutex_ != NULL) {
}
vpx_free(lf_sync->cond_);
}
-
+#endif // CONFIG_MULTITHREAD
vpx_free(lf_sync->cur_sb_col);
// clear the structure as the source of this call may be a resize in which
// case this call will be followed by an _alloc() which may fail.
- vpx_memset(lf_sync, 0, sizeof(*lf_sync));
+ vp9_zero(*lf_sync);
}
-#else
- (void)rows;
- if (lf_sync != NULL) {
- vpx_free(lf_sync->cur_sb_col);
- vpx_memset(lf_sync, 0, sizeof(*lf_sync));
- }
-#endif // CONFIG_MULTITHREAD
}
projects / platform / framework / web / crosswalk.git / blobdiff