/* * 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.
*/
staticINLINEvoid sync_write(VP9LfSync *const lf_sync, int r, int c, constint sb_cols) { #if CONFIG_MULTITHREAD constint 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;
}
staticvoid loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame, VP9_COMMON *cm, struct macroblockd_plane planes[MAX_MB_PLANE], int start, int stop, int y_only,
VPxWorker *workers, int nworkers,
VP9LfSync *lf_sync) { const VPxWorkerInterface *const winterface = vpx_get_worker_interface(); // Number of superblock rows and cols constint sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2; constint num_tile_cols = 1 << cm->log2_tile_cols; // Limit the number of workers to prevent changes in frame dimensions from // causing incorrect sync calculations when sb_rows < threads/tile_cols. // Further restrict them by the number of tile columns should the user // request more as this implementation doesn't scale well beyond that. constint num_workers = VPXMIN(nworkers, VPXMIN(num_tile_cols, sb_rows)); int i;
// Initialize cur_sb_col to -1 for all SB rows.
memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
// Set up loopfilter thread data. // The decoder is capping num_workers because it has been observed that using // more threads on the loopfilter than there are cores 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) {
VPxWorker *const worker = &workers[i];
LFWorkerData *const lf_data = &lf_sync->lfdata[i];
// Loopfilter data
vp9_loop_filter_data_reset(lf_data, frame, cm, planes);
lf_data->start = start + i * MI_BLOCK_SIZE;
lf_data->stop = stop;
lf_data->y_only = y_only;
// Start loopfiltering if (i == num_workers - 1) {
winterface->execute(worker);
} else {
winterface->launch(worker);
}
}
// Wait till all rows are finished for (i = 0; i < num_workers; ++i) {
winterface->sync(&workers[i]);
}
}
void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame, VP9_COMMON *cm, struct macroblockd_plane planes[MAX_MB_PLANE], int frame_filter_level, int y_only, int partial_frame, VPxWorker *workers, int num_workers, VP9LfSync *lf_sync) { int start_mi_row, end_mi_row, mi_rows_to_filter;
// Set up nsync by width. staticINLINEint 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; elseif (width <= 1280) return 2; elseif (width <= 4096) return 4; else return 8;
}
// Allocate memory for lf row synchronization void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows, int width, int num_workers) {
lf_sync->rows = rows; #if CONFIG_MULTITHREAD
{ int i;
CHECK_MEM_ERROR(&cm->error, lf_sync->mutex,
vpx_malloc(sizeof(*lf_sync->mutex) * rows)); if (lf_sync->mutex) { for (i = 0; i < rows; ++i) {
pthread_mutex_init(&lf_sync->mutex[i], NULL);
}
}
CHECK_MEM_ERROR(&cm->error, lf_sync->cond,
vpx_malloc(sizeof(*lf_sync->cond) * rows)); if (lf_sync->cond) { for (i = 0; i < rows; ++i) {
pthread_cond_init(&lf_sync->cond[i], NULL);
}
}
// Set up nsync.
lf_sync->sync_range = get_sync_range(width);
}
// Deallocate lf synchronization related mutex and data void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) {
assert(lf_sync != NULL);
#if CONFIG_MULTITHREAD if (lf_sync->mutex != NULL) { int i; for (i = 0; i < lf_sync->rows; ++i) {
pthread_mutex_destroy(&lf_sync->mutex[i]);
}
vpx_free(lf_sync->mutex);
} if (lf_sync->cond != NULL) { int i; for (i = 0; i < lf_sync->rows; ++i) {
pthread_cond_destroy(&lf_sync->cond[i]);
}
vpx_free(lf_sync->cond);
} if (lf_sync->recon_done_mutex != NULL) { int i; for (i = 0; i < lf_sync->rows; ++i) {
pthread_mutex_destroy(&lf_sync->recon_done_mutex[i]);
}
vpx_free(lf_sync->recon_done_mutex);
}
if (lf_sync->lf_mutex != NULL) {
pthread_mutex_destroy(lf_sync->lf_mutex);
vpx_free(lf_sync->lf_mutex);
} if (lf_sync->recon_done_cond != NULL) { int i; for (i = 0; i < lf_sync->rows; ++i) {
pthread_cond_destroy(&lf_sync->recon_done_cond[i]);
}
vpx_free(lf_sync->recon_done_cond);
} #endif// CONFIG_MULTITHREAD
vpx_free(lf_sync->lfdata);
vpx_free(lf_sync->cur_sb_col);
vpx_free(lf_sync->num_tiles_done); // 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.
vp9_zero(*lf_sync);
}
#if CONFIG_MULTITHREAD int cur_row; constint tile_cols = 1 << cm->log2_tile_cols;
pthread_mutex_lock(lf_sync->lf_mutex); if (cm->lf_row < max_rows) {
cur_row = cm->lf_row >> MI_BLOCK_SIZE_LOG2;
return_val = cm->lf_row;
cm->lf_row += MI_BLOCK_SIZE; if (cm->lf_row < max_rows) { /* If this is not the last row, make sure the next row is also decoded.
* This is because the intra predict has to happen before loop filter */
cur_row += 1;
}
}
pthread_mutex_unlock(lf_sync->lf_mutex);
void vp9_set_row(VP9LfSync *lf_sync, int num_tiles, int row, int is_last_row, int corrupted) { #if CONFIG_MULTITHREAD
pthread_mutex_lock(lf_sync->lf_mutex);
lf_sync->corrupted |= corrupted;
pthread_mutex_unlock(lf_sync->lf_mutex);
pthread_mutex_lock(&lf_sync->recon_done_mutex[row]);
lf_sync->num_tiles_done[row] += 1; if (num_tiles == lf_sync->num_tiles_done[row]) { if (is_last_row) { /* The last 2 rows wait on the last row to be done. * So, we have to broadcast the signal in this case.
*/
pthread_cond_broadcast(&lf_sync->recon_done_cond[row]);
} else {
pthread_cond_signal(&lf_sync->recon_done_cond[row]);
}
}
pthread_mutex_unlock(&lf_sync->recon_done_mutex[row]); #else
(void)lf_sync;
(void)num_tiles;
(void)row;
(void)is_last_row;
(void)corrupted; #endif// CONFIG_MULTITHREAD
}
// Accumulate frame counts. void vp9_accumulate_frame_counts(FRAME_COUNTS *accum, const FRAME_COUNTS *counts, int is_dec) { int i, j, k, l, m;
for (i = 0; i < BLOCK_SIZE_GROUPS; i++) for (j = 0; j < INTRA_MODES; j++)
accum->y_mode[i][j] += counts->y_mode[i][j];
for (i = 0; i < INTRA_MODES; i++) for (j = 0; j < INTRA_MODES; j++)
accum->uv_mode[i][j] += counts->uv_mode[i][j];
for (i = 0; i < PARTITION_CONTEXTS; i++) for (j = 0; j < PARTITION_TYPES; j++)
accum->partition[i][j] += counts->partition[i][j];
if (is_dec) { int n; for (i = 0; i < TX_SIZES; i++) for (j = 0; j < PLANE_TYPES; j++) for (k = 0; k < REF_TYPES; k++) for (l = 0; l < COEF_BANDS; l++) for (m = 0; m < COEFF_CONTEXTS; m++) {
accum->eob_branch[i][j][k][l][m] +=
counts->eob_branch[i][j][k][l][m]; for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
accum->coef[i][j][k][l][m][n] += counts->coef[i][j][k][l][m][n];
}
} else { for (i = 0; i < TX_SIZES; i++) for (j = 0; j < PLANE_TYPES; j++) for (k = 0; k < REF_TYPES; k++) for (l = 0; l < COEF_BANDS; l++) for (m = 0; m < COEFF_CONTEXTS; m++)
accum->eob_branch[i][j][k][l][m] +=
counts->eob_branch[i][j][k][l][m]; // In the encoder, coef is only updated at frame // level, so not need to accumulate it here. // for (n = 0; n < UNCONSTRAINED_NODES + 1; n++) // accum->coef[i][j][k][l][m][n] += // counts->coef[i][j][k][l][m][n];
}
for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) for (j = 0; j < SWITCHABLE_FILTERS; j++)
accum->switchable_interp[i][j] += counts->switchable_interp[i][j];
for (i = 0; i < INTER_MODE_CONTEXTS; i++) for (j = 0; j < INTER_MODES; j++)
accum->inter_mode[i][j] += counts->inter_mode[i][j];
for (i = 0; i < INTRA_INTER_CONTEXTS; i++) for (j = 0; j < 2; j++)
accum->intra_inter[i][j] += counts->intra_inter[i][j];
for (i = 0; i < COMP_INTER_CONTEXTS; i++) for (j = 0; j < 2; j++) accum->comp_inter[i][j] += counts->comp_inter[i][j];
for (i = 0; i < REF_CONTEXTS; i++) for (j = 0; j < 2; j++) for (k = 0; k < 2; k++)
accum->single_ref[i][j][k] += counts->single_ref[i][j][k];
for (i = 0; i < REF_CONTEXTS; i++) for (j = 0; j < 2; j++) accum->comp_ref[i][j] += counts->comp_ref[i][j];
for (i = 0; i < TX_SIZE_CONTEXTS; i++) { for (j = 0; j < TX_SIZES; j++)
accum->tx.p32x32[i][j] += counts->tx.p32x32[i][j];
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.