Amend and improve VP8 multithreading implementation

There are flaws in current implementation of VP8 multithreading encoder
and decoder as reported in the following issue:
https://code.google.com/p/chromium/issues/detail?id=158922

Although the data race warnings are harmless, and wouldn't cause real
problems while encoding and decoding videos, it is better to fix the
warnings so that VP8 code could pass the TSan test.

To synchronize the thread-shared data access and maintain the speed
(i.e. decoding speed), use multiple mutexes based on mb_rows to reduce
the number of synchronizations needed, make the reads and writes of
the shared data protected, and reduce the number of mb_col writes by
nsync times.

The decoder speed tests showed < 3% speed loss while using 2 ~ 4
threads.

Change-Id: Ie296defffcd86a693188b668270d811964227882

1 files changed, 2 insertions, 0 deletions

diff --git a/vp8/encoder/onyx_int.h b/vp8/encoder/onyx_int.h
index 317e4b9e4..2b2f7a0a9 100644
--- a/vp8/encoder/onyx_int.h
+++ b/vp8/encoder/onyx_int.h
@@ -530,6 +530,8 @@ typedef struct VP8_COMP
 
 #if CONFIG_MULTITHREAD
     /* multithread data */
+    pthread_mutex_t *pmutex;
+    pthread_mutex_t mt_mutex;           /* mutex for b_multi_threaded */
     int * mt_current_mb_col;
     int mt_sync_range;
     int b_multi_threaded;