static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame,
VP10_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
const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
// Decoder may allocate more threads than number of tiles based on user's
// input.
const int tile_cols = 1 << cm->log2_tile_cols;
const int num_workers = VPXMIN(nworkers, tile_cols);
int i;
if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
num_workers > lf_sync->num_workers) {
vp10_loop_filter_dealloc(lf_sync);
vp10_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
}
// 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];
worker->hook = (VPxWorkerHook)loop_filter_row_worker;
worker->data1 = lf_sync;
worker->data2 = lf_data;
// Loopfilter data
vp10_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 vp10_decoder_remove(VP10Decoder *pbi) {
int i;
vpx_get_worker_interface()->end(&pbi->lf_worker);
vpx_free(pbi->lf_worker.data1);
vpx_free(pbi->tile_data);
for (i = 0; i < pbi->num_tile_workers; ++i) {
VPxWorker *const worker = &pbi->tile_workers[i];
vpx_get_worker_interface()->end(worker);
}
vpx_free(pbi->tile_worker_data);
vpx_free(pbi->tile_worker_info);
vpx_free(pbi->tile_workers);
if (pbi->num_tile_workers > 0) {
vp10_loop_filter_dealloc(&pbi->lf_row_sync);
}
vpx_free(pbi);
}
