📄 motion_est_template.c
字号:
minima[j].height= score_map[i];
minima[j].x= key & ((1<<ME_MAP_MV_BITS)-1); key>>=ME_MAP_MV_BITS;
minima[j].y= key & ((1<<ME_MAP_MV_BITS)-1);
minima[j].x-= (1<<(ME_MAP_MV_BITS-1));
minima[j].y-= (1<<(ME_MAP_MV_BITS-1));
// all entries in map should be in range except if the mv overflows their ME_MAP_MV_BITS bits space
if( minima[j].x > xmax || minima[j].x < xmin
|| minima[j].y > ymax || minima[j].y < ymin)
continue;
minima[j].checked=0;
if(minima[j].x || minima[j].y)
minima[j].height+= (mv_penalty[((minima[j].x)<<shift)-pred_x] + mv_penalty[((minima[j].y)<<shift)-pred_y])*penalty_factor;
j++;
}
qsort(minima, j, sizeof(Minima), minima_cmp);
for(; j<minima_count; j++){
minima[j].height=256*256*256*64;
minima[j].checked=0;
minima[j].x= minima[j].y=0;
}
for(i=0; i<minima_count; i++){
const int x= minima[i].x;
const int y= minima[i].y;
int d;
if(minima[i].checked) continue;
if( x >= xmax || x <= xmin
|| y >= ymax || y <= ymin)
continue;
SAB_CHECK_MV(x-1, y)
SAB_CHECK_MV(x+1, y)
SAB_CHECK_MV(x , y-1)
SAB_CHECK_MV(x , y+1)
minima[i].checked= 1;
}
best[0]= minima[0].x;
best[1]= minima[0].y;
dmin= minima[0].height;
if( best[0] < xmax && best[0] > xmin
&& best[1] < ymax && best[1] > ymin){
int d;
//ensure that the refernece samples for hpel refinement are in the map
CHECK_MV(best[0]-1, best[1])
CHECK_MV(best[0]+1, best[1])
CHECK_MV(best[0], best[1]-1)
CHECK_MV(best[0], best[1]+1)
}
return dmin;
}
static int var_diamond_search(MpegEncContext * s, int *best, int dmin,
int src_index, int ref_index, int const penalty_factor,
int size, int h, int flags)
{
MotionEstContext * const c= &s->me;
me_cmp_func cmpf, chroma_cmpf;
int dia_size;
LOAD_COMMON
LOAD_COMMON2
int map_generation= c->map_generation;
cmpf= s->dsp.me_cmp[size];
chroma_cmpf= s->dsp.me_cmp[size+1];
for(dia_size=1; dia_size<=c->dia_size; dia_size++){
int dir, start, end;
const int x= best[0];
const int y= best[1];
start= FFMAX(0, y + dia_size - ymax);
end = FFMIN(dia_size, xmax - x + 1);
for(dir= start; dir<end; dir++){
int d;
//check(x + dir,y + dia_size - dir,0, a0)
CHECK_MV(x + dir , y + dia_size - dir);
}
start= FFMAX(0, x + dia_size - xmax);
end = FFMIN(dia_size, y - ymin + 1);
for(dir= start; dir<end; dir++){
int d;
//check(x + dia_size - dir, y - dir,0, a1)
CHECK_MV(x + dia_size - dir, y - dir );
}
start= FFMAX(0, -y + dia_size + ymin );
end = FFMIN(dia_size, x - xmin + 1);
for(dir= start; dir<end; dir++){
int d;
//check(x - dir,y - dia_size + dir,0, a2)
CHECK_MV(x - dir , y - dia_size + dir);
}
start= FFMAX(0, -x + dia_size + xmin );
end = FFMIN(dia_size, ymax - y + 1);
for(dir= start; dir<end; dir++){
int d;
//check(x - dia_size + dir, y + dir,0, a3)
CHECK_MV(x - dia_size + dir, y + dir );
}
if(x!=best[0] || y!=best[1])
dia_size=0;
#if 0
{
int dx, dy, i;
static int stats[8*8];
dx= FFABS(x-best[0]);
dy= FFABS(y-best[1]);
stats[dy*8 + dx] ++;
if(256*256*256*64 % (stats[0]+1)==0){
for(i=0; i<64; i++){
if((i&7)==0) printf("\n");
printf("%6d ", stats[i]);
}
printf("\n");
}
}
#endif
}
return dmin;
}
static av_always_inline int diamond_search(MpegEncContext * s, int *best, int dmin,
int src_index, int ref_index, int const penalty_factor,
int size, int h, int flags){
MotionEstContext * const c= &s->me;
if(c->dia_size==-1)
return funny_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
else if(c->dia_size<-1)
return sab_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
else if(c->dia_size<2)
return small_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
else if(c->dia_size>768)
return umh_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
else if(c->dia_size>512)
return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, c->dia_size&0xFF);
else if(c->dia_size>256)
return l2s_dia_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
else
return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
}
/*!
\param P[10][2] a list of candidate mvs to check before starting the
iterative search. If one of the candidates is close to the optimal mv, then
it takes fewer iterations. And it increases the chance that we find the
optimal mv.
*/
static av_always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr,
int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
int ref_mv_scale, int flags, int size, int h)
{
MotionEstContext * const c= &s->me;
int best[2]={0, 0}; /*!< x and y coordinates of the best motion vector.
i.e. the difference between the position of the
block currently being encoded and the position of
the block chosen to predict it from. */
int d; ///< the score (cmp + penalty) of any given mv
int dmin; /*!< the best value of d, i.e. the score
corresponding to the mv stored in best[]. */
int map_generation;
int penalty_factor;
const int ref_mv_stride= s->mb_stride; //pass as arg FIXME
const int ref_mv_xy= s->mb_x + s->mb_y*ref_mv_stride; //add to last_mv beforepassing FIXME
me_cmp_func cmpf, chroma_cmpf;
LOAD_COMMON
LOAD_COMMON2
if(c->pre_pass){
penalty_factor= c->pre_penalty_factor;
cmpf= s->dsp.me_pre_cmp[size];
chroma_cmpf= s->dsp.me_pre_cmp[size+1];
}else{
penalty_factor= c->penalty_factor;
cmpf= s->dsp.me_cmp[size];
chroma_cmpf= s->dsp.me_cmp[size+1];
}
map_generation= update_map_generation(c);
assert(cmpf);
dmin= cmp(s, 0, 0, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
map[0]= map_generation;
score_map[0]= dmin;
//FIXME precalc first term below?
if((s->pict_type == B_TYPE && !(c->flags & FLAG_DIRECT)) || s->flags&CODEC_FLAG_MV0)
dmin += (mv_penalty[pred_x] + mv_penalty[pred_y])*penalty_factor;
/* first line */
if (s->first_slice_line) {
CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
}else{
if(dmin<((h*h*s->avctx->mv0_threshold)>>8)
&& ( P_LEFT[0] |P_LEFT[1]
|P_TOP[0] |P_TOP[1]
|P_TOPRIGHT[0]|P_TOPRIGHT[1])==0){
*mx_ptr= 0;
*my_ptr= 0;
c->skip=1;
return dmin;
}
CHECK_MV( P_MEDIAN[0] >>shift , P_MEDIAN[1] >>shift)
CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)-1)
CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)+1)
CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)-1, (P_MEDIAN[1]>>shift) )
CHECK_CLIPPED_MV((P_MEDIAN[0]>>shift)+1, (P_MEDIAN[1]>>shift) )
CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
CHECK_MV(P_LEFT[0] >>shift, P_LEFT[1] >>shift)
CHECK_MV(P_TOP[0] >>shift, P_TOP[1] >>shift)
CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
}
if(dmin>h*h*4){
if(c->pre_pass){
CHECK_CLIPPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy-1][1]*ref_mv_scale + (1<<15))>>16)
if(!s->first_slice_line)
CHECK_CLIPPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy-ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
}else{
CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
}
}
if(c->avctx->last_predictor_count){
const int count= c->avctx->last_predictor_count;
const int xstart= FFMAX(0, s->mb_x - count);
const int ystart= FFMAX(0, s->mb_y - count);
const int xend= FFMIN(s->mb_width , s->mb_x + count + 1);
const int yend= FFMIN(s->mb_height, s->mb_y + count + 1);
int mb_y;
for(mb_y=ystart; mb_y<yend; mb_y++){
int mb_x;
for(mb_x=xstart; mb_x<xend; mb_x++){
const int xy= mb_x + 1 + (mb_y + 1)*ref_mv_stride;
int mx= (last_mv[xy][0]*ref_mv_scale + (1<<15))>>16;
int my= (last_mv[xy][1]*ref_mv_scale + (1<<15))>>16;
if(mx>xmax || mx<xmin || my>ymax || my<ymin) continue;
CHECK_MV(mx,my)
}
}
}
//check(best[0],best[1],0, b0)
dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
//check(best[0],best[1],0, b1)
*mx_ptr= best[0];
*my_ptr= best[1];
// printf("%d %d %d \n", best[0], best[1], dmin);
return dmin;
}
//this function is dedicated to the braindamaged gcc
inline int ff_epzs_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr,
int P[10][2], int src_index, int ref_index, int16_t (*last_mv)[2],
int ref_mv_scale, int size, int h)
{
MotionEstContext * const c= &s->me;
//FIXME convert other functions in the same way if faster
if(c->flags==0 && h==16 && size==0){
return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, 0, 0, 16);
// case FLAG_QPEL:
// return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, FLAG_QPEL);
}else{
return epzs_motion_search_internal(s, mx_ptr, my_ptr, P, src_index, ref_index, last_mv, ref_mv_scale, c->flags, size, h);
}
}
static int epzs_motion_search4(MpegEncContext * s,
int *mx_ptr, int *my_ptr, int P[10][2],
int src_index, int ref_index, int16_t (*last_mv)[2],
int ref_mv_scale)
{
MotionEstContext * const c= &s->me;
int best[2]={0, 0};
int d, dmin;
int map_generation;
const int penalty_factor= c->penalty_factor;
const int size=1;
const int h=8;
const int ref_mv_stride= s->mb_stride;
const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
me_cmp_func cmpf, chroma_cmpf;
LOAD_COMMON
int flags= c->flags;
LOAD_COMMON2
cmpf= s->dsp.me_cmp[size];
chroma_cmpf= s->dsp.me_cmp[size+1];
map_generation= update_map_generation(c);
dmin = 1000000;
//printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
/* first line */
if (s->first_slice_line) {
CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
}else{
CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
//FIXME try some early stop
CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
}
if(dmin>64*4){
CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
}
dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
*mx_ptr= best[0];
*my_ptr= best[1];
// printf("%d %d %d \n", best[0], best[1], dmin);
return dmin;
}
//try to merge with above FIXME (needs PSNR test)
static int epzs_motion_search2(MpegEncContext * s,
int *mx_ptr, int *my_ptr, int P[10][2],
int src_index, int ref_index, int16_t (*last_mv)[2],
int ref_mv_scale)
{
MotionEstContext * const c= &s->me;
int best[2]={0, 0};
int d, dmin;
int map_generation;
const int penalty_factor= c->penalty_factor;
const int size=0; //FIXME pass as arg
const int h=8;
const int ref_mv_stride= s->mb_stride;
const int ref_mv_xy= s->mb_x + s->mb_y *ref_mv_stride;
me_cmp_func cmpf, chroma_cmpf;
LOAD_COMMON
int flags= c->flags;
LOAD_COMMON2
cmpf= s->dsp.me_cmp[size];
chroma_cmpf= s->dsp.me_cmp[size+1];
map_generation= update_map_generation(c);
dmin = 1000000;
//printf("%d %d %d %d //",xmin, ymin, xmax, ymax);
/* first line */
if (s->first_slice_line) {
CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
}else{
CHECK_MV(P_MV1[0]>>shift, P_MV1[1]>>shift)
//FIXME try some early stop
CHECK_MV(P_MEDIAN[0]>>shift, P_MEDIAN[1]>>shift)
CHECK_MV(P_LEFT[0]>>shift, P_LEFT[1]>>shift)
CHECK_MV(P_TOP[0]>>shift, P_TOP[1]>>shift)
CHECK_MV(P_TOPRIGHT[0]>>shift, P_TOPRIGHT[1]>>shift)
CHECK_CLIPPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy][1]*ref_mv_scale + (1<<15))>>16)
}
if(dmin>64*4){
CHECK_CLIPPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy+1][1]*ref_mv_scale + (1<<15))>>16)
if(s->mb_y+1<s->end_mb_y) //FIXME replace at least with last_slice_line
CHECK_CLIPPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16,
(last_mv[ref_mv_xy+ref_mv_stride][1]*ref_mv_scale + (1<<15))>>16)
}
dmin= diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags);
*mx_ptr= best[0];
*my_ptr= best[1];
// printf("%d %d %d \n", best[0], best[1], dmin);
return dmin;
}
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -