Limit row mt threads based on tile size for decoder
This patch limits the number of threads assigned to
decode a tile based on its sb rows and columns.
Theoretically thread count of tile decoding can not
exceed two.
This change does not limit number of threads used in
post-processing stages.
BUG=aomedia:2215
Change-Id: I0d45f1f2059eb109e1a8b22228b1c947f60b4527
diff --git a/av1/decoder/decodeframe.c b/av1/decoder/decodeframe.c
index d87cfeb..f01b890 100644
--- a/av1/decoder/decodeframe.c
+++ b/av1/decoder/decodeframe.c
@@ -64,6 +64,9 @@
#define ACCT_STR __func__
+#define AOM_MIN_THREADS_PER_TILE 1
+#define AOM_MAX_THREADS_PER_TILE 2
+
// This is needed by ext_tile related unit tests.
#define EXT_TILE_DEBUG 1
#define MC_TEMP_BUF_PELS \
@@ -3361,6 +3364,20 @@
return !td->xd.corrupted;
}
+static INLINE int get_max_row_mt_workers_per_tile(AV1_COMMON *cm,
+ TileInfo tile) {
+ // NOTE: Currently value of max workers is calculated based
+ // on the parse and decode time. As per the theoretical estimate
+ // when percentage of parse time is equal to percentage of decode
+ // time, number of workers needed to parse + decode a tile can not
+ // exceed more than 2.
+ // TODO(any): Modify this value if parsing is optimized in future.
+ int sb_rows = av1_get_sb_rows_in_tile(cm, tile);
+ int max_workers =
+ sb_rows == 1 ? AOM_MIN_THREADS_PER_TILE : AOM_MAX_THREADS_PER_TILE;
+ return max_workers;
+}
+
// The caller must hold pbi->row_mt_mutex_ when calling this function.
// Returns 1 if either the next job is stored in *next_job_info or 1 is stored
// in *end_of_frame.
@@ -3391,8 +3408,8 @@
int min_threads_working = INT_MAX;
int max_mis_to_decode = 0;
int tile_row_idx, tile_col_idx;
- int tile_row = 0;
- int tile_col = 0;
+ int tile_row = -1;
+ int tile_col = -1;
memset(next_job_info, 0, sizeof(*next_job_info));
@@ -3440,7 +3457,9 @@
max_mis_to_decode = 0;
}
if (num_threads_working == min_threads_working &&
- num_mis_to_decode > max_mis_to_decode) {
+ num_mis_to_decode > max_mis_to_decode &&
+ num_threads_working <
+ get_max_row_mt_workers_per_tile(cm, tile_data->tile_info)) {
max_mis_to_decode = num_mis_to_decode;
tile_row = tile_row_idx;
tile_col = tile_col_idx;
@@ -3448,6 +3467,8 @@
}
}
}
+ // No job found to process
+ if (tile_row == -1 || tile_col == -1) return 0;
tile_data = pbi->tile_data + tile_row * cm->tile_cols + tile_col;
tile_info = tile_data->tile_info;
@@ -3576,9 +3597,22 @@
TileDataDec *const tile_data = cur_job_info->tile_data;
tile_worker_hook_init(pbi, thread_data, tile_buffer, tile_data,
allow_update_cdf);
-
+#if CONFIG_MULTITHREAD
+ pthread_mutex_lock(pbi->row_mt_mutex_);
+#endif
+ tile_data->dec_row_mt_sync.num_threads_working++;
+#if CONFIG_MULTITHREAD
+ pthread_mutex_unlock(pbi->row_mt_mutex_);
+#endif
// decode tile
parse_tile_row_mt(pbi, td, tile_data);
+#if CONFIG_MULTITHREAD
+ pthread_mutex_lock(pbi->row_mt_mutex_);
+#endif
+ tile_data->dec_row_mt_sync.num_threads_working--;
+#if CONFIG_MULTITHREAD
+ pthread_mutex_unlock(pbi->row_mt_mutex_);
+#endif
} else {
break;
}
@@ -4055,7 +4089,8 @@
int tile_cols_start;
int tile_cols_end;
int tile_count_tg;
- int num_workers;
+ int num_workers = 0;
+ int max_threads;
const uint8_t *raw_data_end = NULL;
int max_sb_rows = 0;
@@ -4071,7 +4106,7 @@
tile_cols_end = tile_cols;
}
tile_count_tg = end_tile - start_tile + 1;
- num_workers = pbi->max_threads;
+ max_threads = pbi->max_threads;
// No tiles to decode.
if (tile_rows_end <= tile_rows_start || tile_cols_end <= tile_cols_start ||
@@ -4084,7 +4119,7 @@
assert(tile_rows <= MAX_TILE_ROWS);
assert(tile_cols <= MAX_TILE_COLS);
assert(tile_count_tg > 0);
- assert(num_workers > 0);
+ assert(max_threads > 0);
assert(start_tile <= end_tile);
assert(start_tile >= 0 && end_tile < n_tiles);
@@ -4116,8 +4151,10 @@
max_sb_rows = AOMMAX(max_sb_rows,
av1_get_sb_rows_in_tile(cm, tile_data->tile_info));
+ num_workers += get_max_row_mt_workers_per_tile(cm, tile_data->tile_info);
}
}
+ num_workers = AOMMIN(num_workers, max_threads);
if (pbi->allocated_row_mt_sync_rows != max_sb_rows) {
for (int i = 0; i < n_tiles; ++i) {
