2013-12-28 00:25:54 +01:00
|
|
|
/*
|
|
|
|
* Copyright (c) 2014 The WebM project authors. All Rights Reserved.
|
|
|
|
*
|
|
|
|
* Use of this source code is governed by a BSD-style license
|
|
|
|
* that can be found in the LICENSE file in the root of the source
|
|
|
|
* tree. An additional intellectual property rights grant can be found
|
|
|
|
* in the file PATENTS. All contributing project authors may
|
|
|
|
* be found in the AUTHORS file in the root of the source tree.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "./vpx_config.h"
|
2014-03-03 23:34:39 +01:00
|
|
|
|
|
|
|
#include "vpx_mem/vpx_mem.h"
|
|
|
|
|
2013-12-28 00:25:54 +01:00
|
|
|
#include "vp9/common/vp9_reconinter.h"
|
2014-03-03 23:34:39 +01:00
|
|
|
|
2013-12-28 00:25:54 +01:00
|
|
|
#include "vp9/decoder/vp9_dthread.h"
|
2014-03-06 21:08:20 +01:00
|
|
|
#include "vp9/decoder/vp9_decoder.h"
|
2013-12-28 00:25:54 +01:00
|
|
|
|
|
|
|
#if CONFIG_MULTITHREAD
|
|
|
|
static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
|
|
|
|
const int kMaxTryLocks = 4000;
|
|
|
|
int locked = 0;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < kMaxTryLocks; ++i) {
|
|
|
|
if (!pthread_mutex_trylock(mutex)) {
|
|
|
|
locked = 1;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!locked)
|
|
|
|
pthread_mutex_lock(mutex);
|
|
|
|
}
|
|
|
|
#endif // CONFIG_MULTITHREAD
|
|
|
|
|
|
|
|
static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
|
|
|
|
#if CONFIG_MULTITHREAD
|
|
|
|
const int nsync = lf_sync->sync_range;
|
|
|
|
|
|
|
|
if (r && !(c & (nsync - 1))) {
|
2014-04-11 01:17:49 +02:00
|
|
|
pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
|
|
|
|
mutex_lock(mutex);
|
2013-12-28 00:25:54 +01:00
|
|
|
|
|
|
|
while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
|
2014-04-11 01:17:49 +02:00
|
|
|
pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
|
2013-12-28 00:25:54 +01:00
|
|
|
}
|
2014-04-11 01:17:49 +02:00
|
|
|
pthread_mutex_unlock(mutex);
|
2013-12-28 00:25:54 +01:00
|
|
|
}
|
|
|
|
#else
|
|
|
|
(void)lf_sync;
|
|
|
|
(void)r;
|
|
|
|
(void)c;
|
|
|
|
#endif // CONFIG_MULTITHREAD
|
|
|
|
}
|
|
|
|
|
|
|
|
static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
|
|
|
|
const int sb_cols) {
|
|
|
|
#if CONFIG_MULTITHREAD
|
|
|
|
const int nsync = lf_sync->sync_range;
|
|
|
|
int cur;
|
|
|
|
// Only signal when there are enough filtered SB for next row to run.
|
|
|
|
int sig = 1;
|
|
|
|
|
|
|
|
if (c < sb_cols - 1) {
|
|
|
|
cur = c;
|
|
|
|
if (c % nsync)
|
|
|
|
sig = 0;
|
|
|
|
} else {
|
|
|
|
cur = sb_cols + nsync;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (sig) {
|
|
|
|
mutex_lock(&lf_sync->mutex_[r]);
|
|
|
|
|
|
|
|
lf_sync->cur_sb_col[r] = cur;
|
|
|
|
|
|
|
|
pthread_cond_signal(&lf_sync->cond_[r]);
|
|
|
|
pthread_mutex_unlock(&lf_sync->mutex_[r]);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
(void)lf_sync;
|
|
|
|
(void)r;
|
|
|
|
(void)c;
|
|
|
|
(void)sb_cols;
|
|
|
|
#endif // CONFIG_MULTITHREAD
|
|
|
|
}
|
|
|
|
|
|
|
|
// Implement row loopfiltering for each thread.
|
|
|
|
static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer,
|
2014-05-16 18:48:26 +02:00
|
|
|
VP9_COMMON *const cm,
|
|
|
|
struct macroblockd_plane planes[MAX_MB_PLANE],
|
2013-12-28 00:25:54 +01:00
|
|
|
int start, int stop, int y_only,
|
|
|
|
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
|
|
|
|
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;
|
2014-04-11 01:17:49 +02:00
|
|
|
MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
|
2013-12-28 00:25:54 +01:00
|
|
|
|
|
|
|
for (c = 0; c < sb_cols; ++c) {
|
|
|
|
const int mi_col = c << MI_BLOCK_SIZE_LOG2;
|
2014-04-11 01:17:49 +02:00
|
|
|
LOOP_FILTER_MASK lfm;
|
2013-12-28 00:25:54 +01:00
|
|
|
int plane;
|
|
|
|
|
|
|
|
sync_read(lf_sync, r, c);
|
|
|
|
|
2014-05-16 18:48:26 +02:00
|
|
|
vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
|
2014-04-11 01:17:49 +02:00
|
|
|
vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm);
|
2013-12-28 00:25:54 +01:00
|
|
|
|
|
|
|
for (plane = 0; plane < num_planes; ++plane) {
|
2014-05-16 18:48:26 +02:00
|
|
|
vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm);
|
2013-12-28 00:25:54 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
sync_write(lf_sync, r, c, sb_cols);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Row-based multi-threaded loopfilter hook
|
|
|
|
static int loop_filter_row_worker(void *arg1, void *arg2) {
|
|
|
|
TileWorkerData *const tile_data = (TileWorkerData*)arg1;
|
|
|
|
LFWorkerData *const lf_data = &tile_data->lfdata;
|
2014-06-29 18:30:59 +02:00
|
|
|
(void) arg2;
|
2014-05-16 18:48:26 +02:00
|
|
|
loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
|
2013-12-28 00:25:54 +01:00
|
|
|
lf_data->start, lf_data->stop, lf_data->y_only,
|
|
|
|
lf_data->lf_sync, lf_data->num_lf_workers);
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
// VP9 decoder: Implement multi-threaded loopfilter that uses the tile
|
|
|
|
// threads.
|
2014-05-13 01:19:19 +02:00
|
|
|
void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
|
|
|
|
VP9Decoder *pbi, VP9_COMMON *cm,
|
2013-12-28 00:25:54 +01:00
|
|
|
int frame_filter_level,
|
2014-05-13 19:28:47 +02:00
|
|
|
int y_only) {
|
2014-04-11 01:17:49 +02:00
|
|
|
VP9LfSync *const lf_sync = &pbi->lf_row_sync;
|
2014-06-20 06:14:51 +02:00
|
|
|
const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
|
2013-12-28 00:25:54 +01:00
|
|
|
// Number of superblock rows and cols
|
|
|
|
const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
|
2014-04-07 05:07:14 +02:00
|
|
|
const int tile_cols = 1 << cm->log2_tile_cols;
|
2014-05-09 01:24:36 +02:00
|
|
|
const int num_workers = MIN(pbi->max_threads & ~1, tile_cols);
|
2013-12-28 00:25:54 +01:00
|
|
|
int i;
|
|
|
|
|
|
|
|
// Allocate memory used in thread synchronization.
|
|
|
|
// This always needs to be done even if frame_filter_level is 0.
|
2014-08-27 21:36:23 +02:00
|
|
|
if (!lf_sync->sync_range || cm->last_height != cm->height) {
|
2014-08-29 02:11:31 +02:00
|
|
|
vp9_loop_filter_dealloc(lf_sync);
|
2014-08-31 04:01:25 +02:00
|
|
|
vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width);
|
2013-12-28 00:25:54 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!frame_filter_level) return;
|
|
|
|
|
|
|
|
vp9_loop_filter_frame_init(cm, frame_filter_level);
|
|
|
|
|
|
|
|
// Initialize cur_sb_col to -1 for all SB rows.
|
2014-04-11 01:17:49 +02:00
|
|
|
vpx_memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
|
2013-12-28 00:25:54 +01:00
|
|
|
|
|
|
|
// Set up loopfilter thread data.
|
2014-04-07 05:07:14 +02:00
|
|
|
// 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) {
|
2013-12-28 00:25:54 +01:00
|
|
|
VP9Worker *const worker = &pbi->tile_workers[i];
|
|
|
|
TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
|
|
|
|
LFWorkerData *const lf_data = &tile_data->lfdata;
|
|
|
|
|
|
|
|
worker->hook = (VP9WorkerHook)loop_filter_row_worker;
|
|
|
|
|
|
|
|
// Loopfilter data
|
2014-05-13 01:19:19 +02:00
|
|
|
lf_data->frame_buffer = frame;
|
2013-12-28 00:25:54 +01:00
|
|
|
lf_data->cm = cm;
|
2014-05-16 18:48:26 +02:00
|
|
|
vp9_copy(lf_data->planes, pbi->mb.plane);
|
2013-12-28 00:25:54 +01:00
|
|
|
lf_data->start = i;
|
|
|
|
lf_data->stop = sb_rows;
|
|
|
|
lf_data->y_only = y_only; // always do all planes in decoder
|
|
|
|
|
2014-04-11 01:17:49 +02:00
|
|
|
lf_data->lf_sync = lf_sync;
|
2014-04-07 05:07:14 +02:00
|
|
|
lf_data->num_lf_workers = num_workers;
|
2013-12-28 00:25:54 +01:00
|
|
|
|
|
|
|
// Start loopfiltering
|
2014-04-07 05:07:14 +02:00
|
|
|
if (i == num_workers - 1) {
|
2014-06-20 06:14:51 +02:00
|
|
|
winterface->execute(worker);
|
2013-12-28 00:25:54 +01:00
|
|
|
} else {
|
2014-06-20 06:14:51 +02:00
|
|
|
winterface->launch(worker);
|
2013-12-28 00:25:54 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Wait till all rows are finished
|
2014-04-07 05:07:14 +02:00
|
|
|
for (i = 0; i < num_workers; ++i) {
|
2014-06-20 06:14:51 +02:00
|
|
|
winterface->sync(&pbi->tile_workers[i]);
|
2013-12-28 00:25:54 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Set up nsync by width.
|
|
|
|
static int get_sync_range(int width) {
|
|
|
|
// nsync numbers are picked by testing. For example, for 4k
|
|
|
|
// video, using 4 gives best performance.
|
|
|
|
if (width < 640)
|
|
|
|
return 1;
|
|
|
|
else if (width <= 1280)
|
|
|
|
return 2;
|
|
|
|
else if (width <= 4096)
|
|
|
|
return 4;
|
|
|
|
else
|
|
|
|
return 8;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Allocate memory for lf row synchronization
|
2014-08-31 04:01:25 +02:00
|
|
|
void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
|
2013-12-28 00:25:54 +01:00
|
|
|
int width) {
|
2014-08-29 02:11:31 +02:00
|
|
|
lf_sync->rows = rows;
|
2013-12-28 00:25:54 +01:00
|
|
|
#if CONFIG_MULTITHREAD
|
2014-08-29 02:11:31 +02:00
|
|
|
{
|
|
|
|
int i;
|
2013-12-28 00:25:54 +01:00
|
|
|
|
2014-08-29 02:11:31 +02:00
|
|
|
CHECK_MEM_ERROR(cm, lf_sync->mutex_,
|
|
|
|
vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
|
|
|
|
for (i = 0; i < rows; ++i) {
|
|
|
|
pthread_mutex_init(&lf_sync->mutex_[i], NULL);
|
|
|
|
}
|
2014-02-06 08:19:26 +01:00
|
|
|
|
2014-08-29 02:11:31 +02:00
|
|
|
CHECK_MEM_ERROR(cm, lf_sync->cond_,
|
|
|
|
vpx_malloc(sizeof(*lf_sync->cond_) * rows));
|
|
|
|
for (i = 0; i < rows; ++i) {
|
|
|
|
pthread_cond_init(&lf_sync->cond_[i], NULL);
|
|
|
|
}
|
2013-12-28 00:25:54 +01:00
|
|
|
}
|
|
|
|
#endif // CONFIG_MULTITHREAD
|
|
|
|
|
|
|
|
CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
|
|
|
|
vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
|
|
|
|
|
|
|
|
// Set up nsync.
|
|
|
|
lf_sync->sync_range = get_sync_range(width);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Deallocate lf synchronization related mutex and data
|
2014-08-29 02:11:31 +02:00
|
|
|
void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) {
|
2013-12-28 00:25:54 +01:00
|
|
|
if (lf_sync != NULL) {
|
2014-04-11 01:17:49 +02:00
|
|
|
#if CONFIG_MULTITHREAD
|
2013-12-28 00:25:54 +01:00
|
|
|
int i;
|
|
|
|
|
2014-02-06 08:19:26 +01:00
|
|
|
if (lf_sync->mutex_ != NULL) {
|
2014-08-29 02:11:31 +02:00
|
|
|
for (i = 0; i < lf_sync->rows; ++i) {
|
2014-02-06 08:19:26 +01:00
|
|
|
pthread_mutex_destroy(&lf_sync->mutex_[i]);
|
|
|
|
}
|
|
|
|
vpx_free(lf_sync->mutex_);
|
|
|
|
}
|
|
|
|
if (lf_sync->cond_ != NULL) {
|
2014-08-29 02:11:31 +02:00
|
|
|
for (i = 0; i < lf_sync->rows; ++i) {
|
2014-02-06 08:19:26 +01:00
|
|
|
pthread_cond_destroy(&lf_sync->cond_[i]);
|
|
|
|
}
|
|
|
|
vpx_free(lf_sync->cond_);
|
2013-12-28 00:25:54 +01:00
|
|
|
}
|
2014-04-11 01:17:49 +02:00
|
|
|
#endif // CONFIG_MULTITHREAD
|
2013-12-28 00:25:54 +01:00
|
|
|
vpx_free(lf_sync->cur_sb_col);
|
2014-02-06 08:19:26 +01:00
|
|
|
// 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.
|
2014-04-11 01:17:49 +02:00
|
|
|
vp9_zero(*lf_sync);
|
2013-12-28 00:25:54 +01:00
|
|
|
}
|
|
|
|
}
|