generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Amend and improve VP8 multithreading implementation

Browse Source 
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
This commit is contained in:
Yunqing Wang
2016-01-06 18:27:37 -08:00
parent b8c2a4eb0c
commit ff0107f60d
7 changed files with 163 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
vp8/decoder/threading.c
View File

@@ -52,9 +52,6 @@ static void setup_decoding_thread_data(VP8D_COMP *pbi, MACROBLOCKD *xd, MB_ROW_D
mbd->subpixel_predict8x8 = xd->subpixel_predict8x8;
mbd->subpixel_predict16x16 = xd->subpixel_predict16x16;
mbd->mode_info_context = pc->mi + pc->mode_info_stride * (i + 1);
mbd->mode_info_stride = pc->mode_info_stride;
mbd->frame_type = pc->frame_type;
mbd->pre = xd->pre;
mbd->dst = xd->dst;
@@ -298,8 +295,8 @@ static void mt_decode_macroblock(VP8D_COMP *pbi, MACROBLOCKD *xd,
static void mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd, int start_mb_row)
{
volatile const int *last_row_current_mb_col;
volatile int *current_mb_col;
const int *last_row_current_mb_col;
int *current_mb_col;
int mb_row;
VP8_COMMON *pc = &pbi->common;
const int nsync = pbi->sync_range;
@@ -337,6 +334,9 @@ static void mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd, int start_mb_row)
xd->up_available = (start_mb_row != 0);
xd->mode_info_context = pc->mi + pc->mode_info_stride * start_mb_row;
xd->mode_info_stride = pc->mode_info_stride;
for (mb_row = start_mb_row; mb_row < pc->mb_rows; mb_row += (pbi->decoding_thread_count + 1))
{
int recon_yoffset, recon_uvoffset;
@@ -405,17 +405,15 @@ static void mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd, int start_mb_row)
xd->dst.uv_stride);
}
for (mb_col = 0; mb_col < pc->mb_cols; mb_col++)
{
*current_mb_col = mb_col - 1;
for (mb_col = 0; mb_col < pc->mb_cols; mb_col++) {
if (((mb_col - 1) % nsync) == 0) {
pthread_mutex_t *mutex = &pbi->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 = &pbi->pmutex[mb_row-1];
sync_read(mutex, mb_col, last_row_current_mb_col, nsync);
}
/* Distance of MB to the various image edges.
@@ -604,7 +602,7 @@ static void mt_decode_mb_rows(VP8D_COMP *pbi, MACROBLOCKD *xd, int start_mb_row)
xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
/* last MB of row is ready just after extension is done */
*current_mb_col = mb_col + nsync;
protected_write(&pbi->pmutex[mb_row], current_mb_col, mb_col + nsync);
++xd->mode_info_context; /* skip prediction column */
xd->up_available = 1;
@@ -629,12 +627,12 @@ static THREAD_FUNCTION thread_decoding_proc(void *p_data)
while (1)
{
if (pbi->b_multithreaded_rd == 0)
if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd) == 0)
break;
if (sem_wait(&pbi->h_event_start_decoding[ithread]) == 0)
{
if (pbi->b_multithreaded_rd == 0)
if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd) == 0)
break;
else
{
@@ -657,6 +655,7 @@ void vp8_decoder_create_threads(VP8D_COMP *pbi)
pbi->b_multithreaded_rd = 0;
pbi->allocated_decoding_thread_count = 0;
pthread_mutex_init(&pbi->mt_mutex, NULL);
/* limit decoding threads to the max number of token partitions */
core_count = (pbi->max_threads > 8) ? 8 : pbi->max_threads;
@@ -699,8 +698,17 @@ void vp8mt_de_alloc_temp_buffers(VP8D_COMP *pbi, int mb_rows)
{
int i;
if (pbi->b_multithreaded_rd)
if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd))
{
/* De-allocate mutex */
if (pbi->pmutex != NULL) {
for (i = 0; i < mb_rows; i++) {
pthread_mutex_destroy(&pbi->pmutex[i]);
}
vpx_free(pbi->pmutex);
pbi->pmutex = NULL;
}
vpx_free(pbi->mt_current_mb_col);
pbi->mt_current_mb_col = NULL ;
@@ -781,7 +789,7 @@ void vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows)
int i;
int uv_width;
if (pbi->b_multithreaded_rd)
if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd))
{
vp8mt_de_alloc_temp_buffers(pbi, prev_mb_rows);
@@ -796,6 +804,15 @@ void vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows)
uv_width = width >>1;
/* Allocate mutex */
CHECK_MEM_ERROR(pbi->pmutex, vpx_malloc(sizeof(*pbi->pmutex) *
pc->mb_rows));
if (pbi->pmutex) {
for (i = 0; i < pc->mb_rows; i++) {
pthread_mutex_init(&pbi->pmutex[i], NULL);
}
}
/* Allocate an int for each mb row. */
CALLOC_ARRAY(pbi->mt_current_mb_col, pc->mb_rows);
@@ -831,11 +848,11 @@ void vp8mt_alloc_temp_buffers(VP8D_COMP *pbi, int width, int prev_mb_rows)
void vp8_decoder_remove_threads(VP8D_COMP *pbi)
{
/* shutdown MB Decoding thread; */
if (pbi->b_multithreaded_rd)
if (protected_read(&pbi->mt_mutex, &pbi->b_multithreaded_rd))
{
int i;
pbi->b_multithreaded_rd = 0;
protected_write(&pbi->mt_mutex, &pbi->b_multithreaded_rd, 0);
/* allow all threads to exit */
for (i = 0; i < pbi->allocated_decoding_thread_count; i++)
@@ -863,6 +880,7 @@ void vp8_decoder_remove_threads(VP8D_COMP *pbi)
vpx_free(pbi->de_thread_data);
pbi->de_thread_data = NULL;
}
pthread_mutex_destroy(&pbi->mt_mutex);
}
void vp8mt_decode_mb_rows( VP8D_COMP *pbi, MACROBLOCKD *xd)