Skip to content

GitLab

Commit a8703156 authored by Yunqing Wang's avatar Yunqing Wang Committed by Code Review
Browse files
Options

Merge "Improved encoder threading"

parents a5ecaca6 385c2a76
Show whitespace changes
Inline Side-by-side
Showing with 244 additions and 223 deletions
vp8/encoder/encodeframe.c
View file @ a8703156
......@@ -487,6 +487,16 @@ void encode_mb_row(VP8_COMP *cpi,
int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
int seg_map_index = (mb_row * cpi->common.mb_cols);
#if CONFIG_MULTITHREAD
const int nsync = cpi->mt_sync_range;
const int rightmost_col = cm->mb_cols - 1;
volatile const int *last_row_current_mb_col;
if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
else
last_row_current_mb_col = &rightmost_col;
#endif
// reset above block coeffs
xd->above_context = cm->above_context;
......@@ -532,6 +542,21 @@ void encode_mb_row(VP8_COMP *cpi,
x->rddiv = cpi->RDDIV;
x->rdmult = cpi->RDMULT;
#if CONFIG_MULTITHREAD
if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
{
if ((mb_col & (nsync - 1)) == 0)
{
while (mb_col > (*last_row_current_mb_col - nsync)
&& (*last_row_current_mb_col) != (cm->mb_cols - 1))
{
x86_pause_hint();
thread_sleep(0);
}
}
}
#endif
if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
activity_sum += vp8_activity_masking(cpi, x);
......@@ -628,7 +653,12 @@ void encode_mb_row(VP8_COMP *cpi,
x->partition_info++;
xd->above_context++;
cpi->current_mb_col_main = mb_col;
#if CONFIG_MULTITHREAD
if (cpi->b_multi_threaded != 0)
{
cpi->mt_current_mb_col[mb_row] = mb_col;
}
#endif
}
//extend the recon for intra prediction
......@@ -642,11 +672,14 @@ void encode_mb_row(VP8_COMP *cpi,
xd->mode_info_context++;
x->partition_info++;
x->activity_sum += activity_sum;
}
#if CONFIG_MULTITHREAD
if ((cpi->b_multi_threaded != 0) && (mb_row == cm->mb_rows - 1))
{
sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
}
#endif
}
void vp8_encode_frame(VP8_COMP *cpi)
{
......@@ -800,22 +833,16 @@ void vp8_encode_frame(VP8_COMP *cpi)
vp8cx_init_mbrthread_data(cpi, x, cpi->mb_row_ei, 1, cpi->encoding_thread_count);
for (mb_row = 0; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
{
cpi->current_mb_col_main = -1;
for (i = 0; i < cm->mb_rows; i++)
cpi->mt_current_mb_col[i] = 0;
for (i = 0; i < cpi->encoding_thread_count; i++)
{
if ((mb_row + i + 1) >= cm->mb_rows)
break;
cpi->mb_row_ei[i].mb_row = mb_row + i + 1;
cpi->mb_row_ei[i].tp = cpi->tok + (mb_row + i + 1) * (cm->mb_cols * 16 * 24);
cpi->mb_row_ei[i].current_mb_col = -1;
//SetEvent(cpi->h_event_mbrencoding[i]);
sem_post(&cpi->h_event_mbrencoding[i]);
sem_post(&cpi->h_event_start_encoding[i]);
}
for (mb_row = 0; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
{
vp8_zero(cm->left_context)
tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
......@@ -830,26 +857,10 @@ void vp8_encode_frame(VP8_COMP *cpi)
xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
if (mb_row < cm->mb_rows - 1)
//WaitForSingleObject(cpi->h_event_main, INFINITE);
sem_wait(&cpi->h_event_main);
}
/*
for( ;mb_row<cm->mb_rows; mb_row ++)
{
vp8_zero( cm->left_context)
tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
sem_wait(&cpi->h_event_end_encoding); /* wait for other threads to finish */
encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
// adjust to the next row of mbs
x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
}
*/
cpi->tok_count = 0;
for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
......@@ -859,7 +870,6 @@ void vp8_encode_frame(VP8_COMP *cpi)
if (xd->segmentation_enabled)
{
int i, j;
if (xd->segmentation_enabled)
......@@ -871,7 +881,6 @@ void vp8_encode_frame(VP8_COMP *cpi)
segment_counts[j] += cpi->mb_row_ei[i].segment_counts[j];
}
}
}
for (i = 0; i < cpi->encoding_thread_count; i++)
......
vp8/encoder/ethreading.c
View file @ a8703156
......@@ -8,15 +8,16 @@
* be found in the AUTHORS file in the root of the source tree.
*/
#include "onyx_int.h"
#include "threading.h"
#include "common.h"
#include "extend.h"
extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t, int recon_yoffset, int recon_uvoffset);
extern int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t);
extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
TOKENEXTRA **t, int recon_yoffset,
int recon_uvoffset);
extern int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x,
TOKENEXTRA **t);
extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x);
extern void vp8_build_block_offsets(MACROBLOCK *x);
extern void vp8_setup_block_ptrs(MACROBLOCK *x);
......@@ -30,6 +31,7 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
MB_ROW_COMP *mbri = (MB_ROW_COMP *)(((ENCODETHREAD_DATA *)p_data)->ptr2);
ENTROPY_CONTEXT_PLANES mb_row_left_context;
const int nsync = cpi->mt_sync_range;
//printf("Started thread %d\n", ithread);
while (1)
......@@ -38,21 +40,23 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
break;
//if(WaitForSingleObject(cpi->h_event_mbrencoding[ithread], INFINITE) == WAIT_OBJECT_0)
if (sem_wait(&cpi->h_event_mbrencoding[ithread]) == 0)
{
if (cpi->b_multi_threaded == FALSE) // we're shutting down
break;
else
if (sem_wait(&cpi->h_event_start_encoding[ithread]) == 0)
{
VP8_COMMON *cm = &cpi->common;
int mb_row = mbri->mb_row;
int mb_row;
MACROBLOCK *x = &mbri->mb;
MACROBLOCKD *xd = &x->e_mbd;
TOKENEXTRA **tp = &mbri->tp;
TOKENEXTRA *tp ;
int *segment_counts = mbri->segment_counts;
int *totalrate = &mbri->totalrate;
if (cpi->b_multi_threaded == FALSE) // we're shutting down
break;
for (mb_row = ithread + 1; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
{
int i;
int recon_yoffset, recon_uvoffset;
int mb_col;
......@@ -63,10 +67,9 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
volatile int *last_row_current_mb_col;
INT64 activity_sum = 0;
if (ithread > 0)
last_row_current_mb_col = &cpi->mb_row_ei[ithread-1].current_mb_col;
else
last_row_current_mb_col = &cpi->current_mb_col_main;
tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
// reset above block coeffs
xd->above_context = cm->above_context;
......@@ -78,8 +81,7 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
recon_yoffset = (mb_row * recon_y_stride * 16);
recon_uvoffset = (mb_row * recon_uv_stride * 8);
cpi->tplist[mb_row].start = *tp;
cpi->tplist[mb_row].start = tp;
//printf("Thread mb_row = %d\n", mb_row);
......@@ -88,11 +90,14 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
{
int seg_map_index = (mb_row * cm->mb_cols);
while (mb_col > (*last_row_current_mb_col - 1) && *last_row_current_mb_col != cm->mb_cols - 1)
if ((mb_col & (nsync - 1)) == 0)
{
while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != cm->mb_cols - 1)
{
x86_pause_hint();
thread_sleep(0);
}
}
// Distance of Mb to the various image edges.
// These specified to 8th pel as they are always compared to values that are in 1/8th pel units
......@@ -115,7 +120,7 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
x->rddiv = cpi->RDDIV;
x->rdmult = cpi->RDMULT;
if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
activity_sum += vp8_activity_masking(cpi, x);
// Is segmentation enabled
......@@ -123,8 +128,8 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
if (xd->segmentation_enabled)
{
// Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
if (cpi->segmentation_map[seg_map_index+mb_col] <= 3)
xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[seg_map_index+mb_col];
if (cpi->segmentation_map[seg_map_index + mb_col] <= 3)
xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[seg_map_index + mb_col];
else
xd->mode_info_context->mbmi.segment_id = 0;
......@@ -137,14 +142,14 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
if (cm->frame_type == KEY_FRAME)
{
*totalrate += vp8cx_encode_intra_macro_block(cpi, x, tp);
*totalrate += vp8cx_encode_intra_macro_block(cpi, x, &tp);
#ifdef MODE_STATS
y_modes[xd->mbmi.mode] ++;
#endif
}
else
{
*totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset);
*totalrate += vp8cx_encode_inter_macroblock(cpi, x, &tp, recon_yoffset, recon_uvoffset);
#ifdef MODE_STATS
inter_y_modes[xd->mbmi.mode] ++;
......@@ -163,31 +168,32 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
// Count of last ref frame 0,0 useage
if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
cpi->inter_zz_count ++;
cpi->inter_zz_count++;
// Special case code for cyclic refresh
// If cyclic update enabled then copy xd->mbmi.segment_id; (which may have been updated based on mode
// during vp8cx_encode_inter_macroblock()) back into the global sgmentation map
if (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)
{
cpi->segmentation_map[seg_map_index+mb_col] = xd->mode_info_context->mbmi.segment_id;
const MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
cpi->segmentation_map[seg_map_index + mb_col] = mbmi->segment_id;
// If the block has been refreshed mark it as clean (the magnitude of the -ve influences how long it will be before we consider another refresh):
// Else if it was coded (last frame 0,0) and has not already been refreshed then mark it as a candidate for cleanup next time (marked 0)
// else mark it as dirty (1).
if (xd->mode_info_context->mbmi.segment_id)
cpi->cyclic_refresh_map[seg_map_index+mb_col] = -1;
else if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
if (mbmi->segment_id)
cpi->cyclic_refresh_map[seg_map_index + mb_col] = -1;
else if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
{
if (cpi->cyclic_refresh_map[seg_map_index+mb_col] == 1)
cpi->cyclic_refresh_map[seg_map_index+mb_col] = 0;
if (cpi->cyclic_refresh_map[seg_map_index + mb_col] == 1)
cpi->cyclic_refresh_map[seg_map_index + mb_col] = 0;
}
else
cpi->cyclic_refresh_map[seg_map_index+mb_col] = 1;
cpi->cyclic_refresh_map[seg_map_index + mb_col] = 1;
}
}
cpi->tplist[mb_row].stop = *tp;
cpi->tplist[mb_row].stop = tp;
x->gf_active_ptr++; // Increment pointer into gf useage flags structure for next mb
......@@ -203,16 +209,14 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
recon_uvoffset += 8;
// Keep track of segment useage
segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
segment_counts[xd->mode_info_context->mbmi.segment_id]++;
// skip to next mb
xd->mode_info_context++;
x->partition_info++;
xd->above_context++;
cpi->mb_row_ei[ithread].current_mb_col = mb_col;
cpi->mt_current_mb_col[mb_row] = mb_col;
}
//extend the recon for intra prediction
......@@ -234,14 +238,11 @@ THREAD_FUNCTION thread_encoding_proc(void *p_data)
xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
if (ithread == (cpi->encoding_thread_count - 1) || mb_row == cm->mb_rows - 1)
if (mb_row == cm->mb_rows - 1)
{
//SetEvent(cpi->h_event_main);
sem_post(&cpi->h_event_main);
}
sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
}
}
}
}
......@@ -363,7 +364,6 @@ static void setup_mbby_copy(MACROBLOCK *mbdst, MACROBLOCK *mbsrc)
}
}
void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
MACROBLOCK *x,
MB_ROW_COMP *mbr_ei,
......@@ -414,7 +414,6 @@ void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
mb->src.u_buffer += 8 * x->src.uv_stride * (i + 1);
mb->src.v_buffer += 8 * x->src.uv_stride * (i + 1);
vp8_build_block_offsets(mb);
vp8_setup_block_dptrs(mbd);
......@@ -431,7 +430,6 @@ void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
}
}
void vp8cx_create_encoder_threads(VP8_COMP *cpi)
{
cpi->b_multi_threaded = 0;
......@@ -451,14 +449,15 @@ void vp8cx_create_encoder_threads(VP8_COMP *cpi)
else
cpi->encoding_thread_count = cpi->oxcf.multi_threaded - 1;
CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * cpi->encoding_thread_count));
CHECK_MEM_ERROR(cpi->h_event_mbrencoding, vpx_malloc(sizeof(sem_t) * cpi->encoding_thread_count));
CHECK_MEM_ERROR(cpi->h_event_start_encoding, vpx_malloc(sizeof(sem_t) * cpi->encoding_thread_count));
CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count));
vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count);
CHECK_MEM_ERROR(cpi->en_thread_data, vpx_malloc(sizeof(ENCODETHREAD_DATA) * cpi->encoding_thread_count));
CHECK_MEM_ERROR(cpi->mt_current_mb_col, vpx_malloc(sizeof(*cpi->mt_current_mb_col) * cpi->common.mb_rows));
//cpi->h_event_main = CreateEvent(NULL, FALSE, FALSE, NULL);
sem_init(&cpi->h_event_main, 0, 0);
sem_init(&cpi->h_event_end_encoding, 0, 0);
cpi->b_multi_threaded = 1;
......@@ -466,11 +465,13 @@ void vp8cx_create_encoder_threads(VP8_COMP *cpi)
for (ithread = 0; ithread < cpi->encoding_thread_count; ithread++)
{
ENCODETHREAD_DATA * ethd = &cpi->en_thread_data[ithread];
//cpi->h_event_mbrencoding[ithread] = CreateEvent(NULL, FALSE, FALSE, NULL);
sem_init(&cpi->h_event_mbrencoding[ithread], 0, 0);
cpi->en_thread_data[ithread].ithread = ithread;
cpi->en_thread_data[ithread].ptr1 = (void *)cpi;
cpi->en_thread_data[ithread].ptr2 = (void *)&cpi->mb_row_ei[ithread];
sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
ethd->ithread = ithread;
ethd->ptr1 = (void *)cpi;
ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
//printf(" call begin thread %d \n", ithread);
......@@ -482,8 +483,7 @@ void vp8cx_create_encoder_threads(VP8_COMP *cpi)
// 0,
// NULL);
pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, (&cpi->en_thread_data[ithread]));
pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, ethd);
}
}
......@@ -505,18 +505,21 @@ void vp8cx_remove_encoder_threads(VP8_COMP *cpi)
for (i = 0; i < cpi->encoding_thread_count; i++)
{
//SetEvent(cpi->h_event_mbrencoding[i]);
sem_post(&cpi->h_event_mbrencoding[i]);
sem_post(&cpi->h_event_start_encoding[i]);
pthread_join(cpi->h_encoding_thread[i], 0);
}
for (i = 0; i < cpi->encoding_thread_count; i++)
sem_destroy(&cpi->h_event_mbrencoding[i]);
sem_destroy(&cpi->h_event_start_encoding[i]);
}
}
sem_destroy(&cpi->h_event_end_encoding);
//free thread related resources
vpx_free(cpi->h_event_mbrencoding);
vpx_free(cpi->h_event_start_encoding);
vpx_free(cpi->h_encoding_thread);
vpx_free(cpi->mb_row_ei);
vpx_free(cpi->en_thread_data);
vpx_free(cpi->mt_current_mb_col);
}
#endif
......
vp8/encoder/onyx_if.c
View file @ a8703156
......@@ -1465,6 +1465,17 @@ void vp8_alloc_compressor_data(VP8_COMP *cpi)
vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
"Failed to allocate firstpass stats");
#endif
#if CONFIG_MULTITHREAD
if (width < 640)
cpi->mt_sync_range = 1;
else if (width <= 1280)
cpi->mt_sync_range = 4;
else if (width <= 2560)
cpi->mt_sync_range = 8;
else
cpi->mt_sync_range = 16;
#endif
}
......@@ -4413,7 +4424,7 @@ static void encode_frame_to_data_rate
}
// Update the GF useage maps.
// Update the GF useage maps.
// This is done after completing the compression of a frame when all modes etc. are finalized but before loop filter
// This is done after completing the compression of a frame when all modes etc. are finalized but before loop filter
vp8_update_gf_useage_maps(cpi, cm, &cpi->mb);
......
vp8/encoder/onyx_int.h
View file @ a8703156
......@@ -191,11 +191,8 @@ typedef struct
typedef struct
{
MACROBLOCK mb;
int mb_row;
TOKENEXTRA *tp;
int segment_counts[MAX_MB_SEGMENTS];
int totalrate;
int current_mb_col;
} MB_ROW_COMP;
typedef struct
......@@ -593,7 +590,8 @@ typedef struct
signed char *cyclic_refresh_map;
// multithread data
int current_mb_col_main;
int * mt_current_mb_col;
int mt_sync_range;
int processor_core_count;
int b_multi_threaded;
int encoding_thread_count;
......@@ -606,8 +604,8 @@ typedef struct
#if CONFIG_MULTITHREAD
//events
sem_t *h_event_mbrencoding;
sem_t h_event_main;
sem_t *h_event_start_encoding;
sem_t h_event_end_encoding;
#endif
TOKENLIST *tplist;
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment