Chiebot-Mirror
/
FFmpeg
mirror of https://github.com/FFmpeg/FFmpeg.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

Restructure spatial direct MV generation so that the zero predictior shortcut

Browse Source 
is closer at the top.
50-130 cpu cycles faster depending on which kind of direct MBs are meassured.

Originally committed as revision 21788 to svn://svn.ffmpeg.org/ffmpeg/trunk
release/0.6
Michael Niedermayer 15 years ago
parent edb4df436f
commit af0d6581e1
1 changed files with 70 additions and 65 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 131
      libavcodec/h264_direct.c

131
libavcodec/h264_direct.c
Unescape View File

@ -151,73 +151,16 @@ static void pred_spatial_direct_motion(H264Context * const h, int *mb_type){
const int is_b8x8 = IS_8X8(*mb_type);
unsigned int sub_mb_type;
int i8, i4;
int ref[2];
int mv[2];
int list;
assert(h->ref_list[1][0].reference&3);
#define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
if(!IS_INTERLACED(*mb_type)){ // AFR/FR -> AFL/FL
mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride;
b8_stride = 0;
}else{
mb_xy += h->col_fieldoff; // non zero for FL -> FL & differ parity
}
goto single_col;
}else{ // AFL/AFR/FR/FL -> AFR/FR
if(IS_INTERLACED(*mb_type)){ // AFL /FL -> AFR/FR
mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
b8_stride *= 3;
b4_stride *= 6;
sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
if( (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
&& (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
&& !is_b8x8){
*mb_type |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
}else{
*mb_type |= MB_TYPE_8x8|MB_TYPE_L0L1;
}
}else{ // AFR/FR -> AFR/FR
single_col:
mb_type_col[0] =
mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
*mb_type |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
}else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
*mb_type |= MB_TYPE_L0L1|MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
}else{
if(!h->sps.direct_8x8_inference_flag){
/* FIXME save sub mb types from previous frames (or derive from MVs)
* so we know exactly what block size to use */
sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
}
*mb_type |= MB_TYPE_8x8|MB_TYPE_L0L1;
}
}
}
l1mv0 = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
l1mv1 = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
if(!b8_stride){
if(s->mb_y&1){
l1ref0 += h->b8_stride;
l1ref1 += h->b8_stride;
l1mv0 += 2*b4_stride;
l1mv1 += 2*b4_stride;
}
}
{
int ref[2];
int mv[2];
int list;
*mb_type |= MB_TYPE_L0L1;
sub_mb_type |= MB_TYPE_L0L1;
/* ref = min(neighbors) */
for(list=0; list<2; list++){
@ -271,7 +214,69 @@ single_col:
fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
*mb_type= (*mb_type & ~(MB_TYPE_8x8|MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_P1L0|MB_TYPE_P1L1))|MB_TYPE_16x16|MB_TYPE_DIRECT2;
}else if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
return;
}
if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
if(!IS_INTERLACED(*mb_type)){ // AFR/FR -> AFL/FL
mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride;
b8_stride = 0;
}else{
mb_xy += h->col_fieldoff; // non zero for FL -> FL & differ parity
}
goto single_col;
}else{ // AFL/AFR/FR/FL -> AFR/FR
if(IS_INTERLACED(*mb_type)){ // AFL /FL -> AFR/FR
mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
b8_stride *= 3;
b4_stride *= 6;
sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
if( (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
&& (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
&& !is_b8x8){
*mb_type |= MB_TYPE_16x8 |MB_TYPE_DIRECT2; /* B_16x8 */
}else{
*mb_type |= MB_TYPE_8x8;
}
}else{ // AFR/FR -> AFR/FR
single_col:
mb_type_col[0] =
mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
*mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_16x16 */
}else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
*mb_type |= MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
}else{
if(!h->sps.direct_8x8_inference_flag){
/* FIXME save sub mb types from previous frames (or derive from MVs)
* so we know exactly what block size to use */
sub_mb_type += (MB_TYPE_8x8-MB_TYPE_16x16); /* B_SUB_4x4 */
}
*mb_type |= MB_TYPE_8x8;
}
}
}
l1mv0 = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
l1mv1 = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
if(!b8_stride){
if(s->mb_y&1){
l1ref0 += h->b8_stride;
l1ref1 += h->b8_stride;
l1mv0 += 2*b4_stride;
l1mv1 += 2*b4_stride;
}
}
if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
int n=0;
for(i8=0; i8<4; i8++){
int x8 = i8&1;

Write Preview
Loading…
Cancel
Save