Skip to content

F
ffmpeg-mt
Commit ae3c9558 authored by michael's avatar michael
Browse files
Options

Restructure spatial direct MV generation so that the zero predictior shortcut 

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


git-svn-id: file:///var/local/repositories/ffmpeg/trunk@21788 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b
parent 5a76e75c
Hide whitespace changes
Inline Side-by-side
Showing with 70 additions and 65 deletions
libavcodec/h264_direct.c
View file @ ae3c9558
......@@ -151,11 +151,72 @@ 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)
*mb_type |= MB_TYPE_L0L1;
sub_mb_type |= MB_TYPE_L0L1;
/* ref = min(neighbors) */
for(list=0; list<2; list++){
int left_ref = h->ref_cache[list][scan8[0] - 1];
int top_ref = h->ref_cache[list][scan8[0] - 8];
int refc = h->ref_cache[list][scan8[0] - 8 + 4];
const int16_t *C= h->mv_cache[list][ scan8[0] - 8 + 4];
if(refc == PART_NOT_AVAILABLE){
refc = h->ref_cache[list][scan8[0] - 8 - 1];
C = h-> mv_cache[list][scan8[0] - 8 - 1];
}
ref[list] = FFMIN3((unsigned)left_ref, (unsigned)top_ref, (unsigned)refc);
if(ref[list] >= 0){
//this is just pred_motion() but with the cases removed that cannot happen for direct blocks
const int16_t * const A= h->mv_cache[list][ scan8[0] - 1 ];
const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
int match_count= (left_ref==ref[list]) + (top_ref==ref[list]) + (refc==ref[list]);
if(match_count > 1){ //most common
mv[list]= (mid_pred(A[0], B[0], C[0])&0xFFFF)
+(mid_pred(A[1], B[1], C[1])<<16);
}else {
assert(match_count==1);
if(left_ref==ref[list]){
mv[list]= *(uint32_t*)A;
}else if(top_ref==ref[list]){
mv[list]= *(uint32_t*)B;
}else{
mv[list]= *(uint32_t*)C;
}
}
}else{
int mask= ~(MB_TYPE_L0 << (2*list));
mv[list] = 0;
ref[list] = -1;
if(!is_b8x8)
*mb_type &= mask;
sub_mb_type &= mask;
}
}
if(ref[0] < 0 && ref[1] < 0){
ref[0] = ref[1] = 0;
if(!is_b8x8)
*mb_type |= MB_TYPE_L0L1;
sub_mb_type |= MB_TYPE_L0L1;
}
if(!is_b8x8 && (mv[0]|mv[1]) == 0){
fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
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;
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;
......@@ -172,31 +233,31 @@ static void pred_spatial_direct_motion(H264Context * const h, int *mb_type){
b8_stride *= 3;
b4_stride *= 6;
sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
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_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
*mb_type |= MB_TYPE_16x8 |MB_TYPE_DIRECT2; /* B_16x8 */
}else{
*mb_type |= MB_TYPE_8x8|MB_TYPE_L0L1;
*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_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
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_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
*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_L0L1|MB_TYPE_DIRECT2 | (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_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
sub_mb_type += (MB_TYPE_8x8-MB_TYPE_16x16); /* B_SUB_4x4 */
}
*mb_type |= MB_TYPE_8x8|MB_TYPE_L0L1;
*mb_type |= MB_TYPE_8x8;
}
}
}
......@@ -214,64 +275,8 @@ single_col:
}
}
{
int ref[2];
int mv[2];
int list;
/* ref = min(neighbors) */
for(list=0; list<2; list++){
int left_ref = h->ref_cache[list][scan8[0] - 1];
int top_ref = h->ref_cache[list][scan8[0] - 8];
int refc = h->ref_cache[list][scan8[0] - 8 + 4];
const int16_t *C= h->mv_cache[list][ scan8[0] - 8 + 4];
if(refc == PART_NOT_AVAILABLE){
refc = h->ref_cache[list][scan8[0] - 8 - 1];
C = h-> mv_cache[list][scan8[0] - 8 - 1];
}
ref[list] = FFMIN3((unsigned)left_ref, (unsigned)top_ref, (unsigned)refc);
if(ref[list] >= 0){
//this is just pred_motion() but with the cases removed that cannot happen for direct blocks
const int16_t * const A= h->mv_cache[list][ scan8[0] - 1 ];
const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
int match_count= (left_ref==ref[list]) + (top_ref==ref[list]) + (refc==ref[list]);
if(match_count > 1){ //most common
mv[list]= (mid_pred(A[0], B[0], C[0])&0xFFFF)
+(mid_pred(A[1], B[1], C[1])<<16);
}else {
assert(match_count==1);
if(left_ref==ref[list]){
mv[list]= *(uint32_t*)A;
}else if(top_ref==ref[list]){
mv[list]= *(uint32_t*)B;
}else{
mv[list]= *(uint32_t*)C;
}
}
}else{
int mask= ~(MB_TYPE_L0 << (2*list));
mv[list] = 0;
ref[list] = -1;
if(!is_b8x8)
*mb_type &= mask;
sub_mb_type &= mask;
}
}
if(ref[0] < 0 && ref[1] < 0){
ref[0] = ref[1] = 0;
if(!is_b8x8)
*mb_type |= MB_TYPE_L0L1;
sub_mb_type |= MB_TYPE_L0L1;
}
if(!is_b8x8 && (mv[0]|mv[1]) == 0){
fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
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])){
if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
int n=0;
for(i8=0; i8<4; i8++){
int x8 = i8&1;
......
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