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, 32 insertions, 22 deletions

diff --git a/vp8/encoder/ethreading.c b/vp8/encoder/ethreading.c
index 4e234ccd5..4f689c4bc 100644
--- a/vp8/encoder/ethreading.c
+++ b/vp8/encoder/ethreading.c
@@ -26,12 +26,13 @@ static THREAD_FUNCTION thread_loopfilter(void *p_data)
 
     while (1)
     {
-        if (cpi->b_multi_threaded == 0)
+        if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0)
             break;
 
         if (sem_wait(&cpi->h_event_start_lpf) == 0)
         {
-            if (cpi->b_multi_threaded == 0) /* we're shutting down */
+            /* we're shutting down */
+            if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0)
                 break;
 
             vp8_loopfilter_frame(cpi, cm);
@@ -53,7 +54,7 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
 
     while (1)
     {
-        if (cpi->b_multi_threaded == 0)
+        if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0)
             break;
 
         if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0)
@@ -72,9 +73,14 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
             int *segment_counts = mbri->segment_counts;
             int *totalrate = &mbri->totalrate;
 
-            if (cpi->b_multi_threaded == 0) /* we're shutting down */
+            /* we're shutting down */
+            if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded) == 0)
                 break;
 
+            xd->mode_info_context = cm->mi + cm->mode_info_stride *
+                (ithread + 1);
+            xd->mode_info_stride = cm->mode_info_stride;
+
             for (mb_row = ithread + 1; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
             {
 
@@ -85,8 +91,8 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
                 int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
                 int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
                 int map_index = (mb_row * cm->mb_cols);
-                volatile const int *last_row_current_mb_col;
-                volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
+                const int *last_row_current_mb_col;
+                int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
 
 #if  (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
                 vp8_writer *w = &cpi->bc[1 + (mb_row % num_part)];
@@ -113,15 +119,14 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
                 /* for each macroblock col in image */
                 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
                 {
-                    *current_mb_col = mb_col - 1;
+                    if (((mb_col - 1) % nsync) == 0) {
+                        pthread_mutex_t *mutex = &cpi->pmutex[mb_row];
+                        protected_write(mutex, current_mb_col, mb_col - 1);
+                    }
 
-                    if ((mb_col & (nsync - 1)) == 0)
-                    {
-                        while (mb_col > (*last_row_current_mb_col - nsync))
-                        {
-                            x86_pause_hint();
-                            thread_sleep(0);
-                        }
+                    if (mb_row && !(mb_col & (nsync - 1))) {
+                      pthread_mutex_t *mutex = &cpi->pmutex[mb_row-1];
+                      sync_read(mutex, mb_col, last_row_current_mb_col, nsync);
                     }
 
 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
@@ -296,7 +301,8 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
                                     xd->dst.u_buffer + 8,
                                     xd->dst.v_buffer + 8);
 
-                *current_mb_col = mb_col + nsync;
+                protected_write(&cpi->pmutex[mb_row], current_mb_col,
+                                mb_col + nsync);
 
                 /* this is to account for the border */
                 xd->mode_info_context++;
@@ -473,9 +479,6 @@ void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
 
         mb->partition_info = x->pi + x->e_mbd.mode_info_stride * (i + 1);
 
-        mbd->mode_info_context = cm->mi   + x->e_mbd.mode_info_stride * (i + 1);
-        mbd->mode_info_stride  = cm->mode_info_stride;
-
         mbd->frame_type = cm->frame_type;
 
         mb->src = * cpi->Source;
@@ -517,6 +520,8 @@ int vp8cx_create_encoder_threads(VP8_COMP *cpi)
     cpi->encoding_thread_count = 0;
     cpi->b_lpf_running = 0;
 
+    pthread_mutex_init(&cpi->mt_mutex, NULL);
+
     if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
     {
         int ithread;
@@ -580,7 +585,7 @@ int vp8cx_create_encoder_threads(VP8_COMP *cpi)
         if(rc)
         {
             /* shutdown other threads */
-            cpi->b_multi_threaded = 0;
+            protected_write(&cpi->mt_mutex, &cpi->b_multi_threaded, 0);
             for(--ithread; ithread >= 0; ithread--)
             {
                 pthread_join(cpi->h_encoding_thread[ithread], 0);
@@ -594,6 +599,8 @@ int vp8cx_create_encoder_threads(VP8_COMP *cpi)
             vpx_free(cpi->mb_row_ei);
             vpx_free(cpi->en_thread_data);
 
+            pthread_mutex_destroy(&cpi->mt_mutex);
+
             return -1;
         }
 
@@ -611,7 +618,7 @@ int vp8cx_create_encoder_threads(VP8_COMP *cpi)
             if(rc)
             {
                 /* shutdown other threads */
-                cpi->b_multi_threaded = 0;
+                protected_write(&cpi->mt_mutex, &cpi->b_multi_threaded, 0);
                 for(--ithread; ithread >= 0; ithread--)
                 {
                     sem_post(&cpi->h_event_start_encoding[ithread]);
@@ -628,6 +635,8 @@ int vp8cx_create_encoder_threads(VP8_COMP *cpi)
                 vpx_free(cpi->mb_row_ei);
                 vpx_free(cpi->en_thread_data);
 
+                pthread_mutex_destroy(&cpi->mt_mutex);
+
                 return -2;
             }
         }
@@ -637,10 +646,10 @@ int vp8cx_create_encoder_threads(VP8_COMP *cpi)
 
 void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
 {
-    if (cpi->b_multi_threaded)
+    if (protected_read(&cpi->mt_mutex, &cpi->b_multi_threaded))
     {
         /* shutdown other threads */
-        cpi->b_multi_threaded = 0;
+        protected_write(&cpi->mt_mutex, &cpi->b_multi_threaded, 0);
         {
             int i;
 
@@ -666,5 +675,6 @@ void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
         vpx_free(cpi->mb_row_ei);
         vpx_free(cpi->en_thread_data);
     }
+    pthread_mutex_destroy(&cpi->mt_mutex);
 }
 #endif