X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fmotion_est_template.c;h=7228744756bd1f89ff59a26f54a7b9d8381faf68;hb=6a8475dc4f10d65edb7faa8df6e035af0ddcce19;hp=980727d5a8dbbfea37dcb0d9a5dafc203f759df7;hpb=2421a01b258c0a0ba4b6f77289973fff3eba3b29;p=ffmpeg diff --git a/libavcodec/motion_est_template.c b/libavcodec/motion_est_template.c index 980727d5a8d..7228744756b 100644 --- a/libavcodec/motion_est_template.c +++ b/libavcodec/motion_est_template.c @@ -2,36 +2,35 @@ * Motion estimation * Copyright (c) 2002-2004 Michael Niedermayer * - * This file is part of FFmpeg. + * This file is part of Libav. * - * FFmpeg is free software; you can redistribute it and/or + * Libav is free software; you can redistribute it and/or * modify it under the terms of the GNU Lesser General Public * License as published by the Free Software Foundation; either * version 2.1 of the License, or (at your option) any later version. * - * FFmpeg is distributed in the hope that it will be useful, + * Libav is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU * Lesser General Public License for more details. * * You should have received a copy of the GNU Lesser General Public - * License along with FFmpeg; if not, write to the Free Software + * License along with Libav; if not, write to the Free Software * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA - * */ /** - * @file motion_est_template.c + * @file * Motion estimation template. */ -//lets hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...) +//Let us hope gcc will remove the unused vars ...(gcc 3.2.2 seems to do it ...) #define LOAD_COMMON\ - uint32_t attribute_unused * const score_map= c->score_map;\ - const int attribute_unused xmin= c->xmin;\ - const int attribute_unused ymin= c->ymin;\ - const int attribute_unused xmax= c->xmax;\ - const int attribute_unused ymax= c->ymax;\ + uint32_t av_unused * const score_map= c->score_map;\ + const int av_unused xmin= c->xmin;\ + const int av_unused ymin= c->ymin;\ + const int av_unused xmax= c->xmax;\ + const int av_unused ymax= c->ymax;\ uint8_t *mv_penalty= c->current_mv_penalty;\ const int pred_x= c->pred_x;\ const int pred_y= c->pred_y;\ @@ -40,80 +39,11 @@ {\ const int hx= 2*(x)+(dx);\ const int hy= 2*(y)+(dy);\ - d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\ + d= cmp_hpel(s, x, y, dx, dy, size, h, ref_index, src_index, cmp_sub, chroma_cmp_sub, flags);\ d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\ COPY3_IF_LT(dmin, d, bx, hx, by, hy)\ } -#if 0 -static int hpel_motion_search)(MpegEncContext * s, - int *mx_ptr, int *my_ptr, int dmin, - uint8_t *ref_data[3], - int size) -{ - const int xx = 16 * s->mb_x + 8*(n&1); - const int yy = 16 * s->mb_y + 8*(n>>1); - const int mx = *mx_ptr; - const int my = *my_ptr; - const int penalty_factor= c->sub_penalty_factor; - - LOAD_COMMON - - // INIT; - //FIXME factorize - me_cmp_func cmp, chroma_cmp, cmp_sub, chroma_cmp_sub; - - if(s->no_rounding /*FIXME b_type*/){ - hpel_put= &s->dsp.put_no_rnd_pixels_tab[size]; - chroma_hpel_put= &s->dsp.put_no_rnd_pixels_tab[size+1]; - }else{ - hpel_put=& s->dsp.put_pixels_tab[size]; - chroma_hpel_put= &s->dsp.put_pixels_tab[size+1]; - } - cmpf= s->dsp.me_cmp[size]; - chroma_cmpf= s->dsp.me_cmp[size+1]; - cmp_sub= s->dsp.me_sub_cmp[size]; - chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; - - if(c->skip){ //FIXME somehow move up (benchmark) - *mx_ptr = 0; - *my_ptr = 0; - return dmin; - } - - if(c->avctx->me_cmp != c->avctx->me_sub_cmp){ - CMP_HPEL(dmin, 0, 0, mx, my, size); - if(mx || my) - dmin += (mv_penalty[2*mx - pred_x] + mv_penalty[2*my - pred_y])*penalty_factor; - } - - if (mx > xmin && mx < xmax && - my > ymin && my < ymax) { - int bx=2*mx, by=2*my; - int d= dmin; - - CHECK_HALF_MV(1, 1, mx-1, my-1) - CHECK_HALF_MV(0, 1, mx , my-1) - CHECK_HALF_MV(1, 1, mx , my-1) - CHECK_HALF_MV(1, 0, mx-1, my ) - CHECK_HALF_MV(1, 0, mx , my ) - CHECK_HALF_MV(1, 1, mx-1, my ) - CHECK_HALF_MV(0, 1, mx , my ) - CHECK_HALF_MV(1, 1, mx , my ) - - assert(bx >= xmin*2 || bx <= xmax*2 || by >= ymin*2 || by <= ymax*2); - - *mx_ptr = bx; - *my_ptr = by; - }else{ - *mx_ptr =2*mx; - *my_ptr =2*my; - } - - return dmin; -} - -#else static int hpel_motion_search(MpegEncContext * s, int *mx_ptr, int *my_ptr, int dmin, int src_index, int ref_index, @@ -159,9 +89,8 @@ static int hpel_motion_search(MpegEncContext * s, const int b= score_map[(index+(1<penalty_factor; -#if 1 - int key; - int map_generation= c->map_generation; + unsigned key; + unsigned map_generation= c->map_generation; #ifndef NDEBUG uint32_t *map= c->map; #endif @@ -173,7 +102,6 @@ static int hpel_motion_search(MpegEncContext * s, assert(map[(index+1)&(ME_MAP_SIZE-1)] == key); key= ((my)<dsp.me_sub_cmp != s->dsp.mb_cmp; MotionEstContext * const c= &s->me; @@ -263,11 +191,17 @@ inline int ff_get_mb_score(MpegEncContext * s, int mx, int my, int src_index, return d; } +int ff_get_mb_score(MpegEncContext *s, int mx, int my, int src_index, + int ref_index, int size, int h, int add_rate) +{ + return get_mb_score(s, mx, my, src_index, ref_index, size, h, add_rate); +} + #define CHECK_QUARTER_MV(dx, dy, x, y)\ {\ const int hx= 4*(x)+(dx);\ const int hy= 4*(y)+(dy);\ - d= cmp(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ + d= cmp_qpel(s, x, y, dx, dy, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ d += (mv_penalty[hx - pred_x] + mv_penalty[hy - pred_y])*penalty_factor;\ COPY3_IF_LT(dmin, d, bx, hx, by, hy)\ } @@ -281,7 +215,7 @@ static int qpel_motion_search(MpegEncContext * s, const int mx = *mx_ptr; const int my = *my_ptr; const int penalty_factor= c->sub_penalty_factor; - const int map_generation= c->map_generation; + const unsigned map_generation = c->map_generation; const int subpel_quality= c->avctx->me_subpel_quality; uint32_t *map= c->map; me_cmp_func cmpf, chroma_cmpf; @@ -324,7 +258,6 @@ static int qpel_motion_search(MpegEncContext * s, int best_pos[8][2]; memset(best, 64, sizeof(int)*8); -#if 1 if(s->me.dia_size>=2){ const int tl= score_map[(index-(1<>2, ny>>2) } -#if 0 - const int tl= score_map[(index-(1<>2, (ny + oy[i])>>2) - } -#endif -#if 0 - //outer ring - CHECK_QUARTER_MV(1, 3, mx-1, my-1) - CHECK_QUARTER_MV(1, 2, mx-1, my-1) - CHECK_QUARTER_MV(1, 1, mx-1, my-1) - CHECK_QUARTER_MV(2, 1, mx-1, my-1) - CHECK_QUARTER_MV(3, 1, mx-1, my-1) - CHECK_QUARTER_MV(0, 1, mx , my-1) - CHECK_QUARTER_MV(1, 1, mx , my-1) - CHECK_QUARTER_MV(2, 1, mx , my-1) - CHECK_QUARTER_MV(3, 1, mx , my-1) - CHECK_QUARTER_MV(3, 2, mx , my-1) - CHECK_QUARTER_MV(3, 3, mx , my-1) - CHECK_QUARTER_MV(3, 0, mx , my ) - CHECK_QUARTER_MV(3, 1, mx , my ) - CHECK_QUARTER_MV(3, 2, mx , my ) - CHECK_QUARTER_MV(3, 3, mx , my ) - CHECK_QUARTER_MV(2, 3, mx , my ) - CHECK_QUARTER_MV(1, 3, mx , my ) - CHECK_QUARTER_MV(0, 3, mx , my ) - CHECK_QUARTER_MV(3, 3, mx-1, my ) - CHECK_QUARTER_MV(2, 3, mx-1, my ) - CHECK_QUARTER_MV(1, 3, mx-1, my ) - CHECK_QUARTER_MV(1, 2, mx-1, my ) - CHECK_QUARTER_MV(1, 1, mx-1, my ) - CHECK_QUARTER_MV(1, 0, mx-1, my ) -#endif assert(bx >= xmin*4 && bx <= xmax*4 && by >= ymin*4 && by <= ymax*4); *mx_ptr = bx; @@ -498,19 +361,17 @@ static int qpel_motion_search(MpegEncContext * s, #define CHECK_MV(x,y)\ {\ - const int key= ((y)<= xmin);\ assert((x) <= xmax);\ assert((y) >= ymin);\ assert((y) <= ymax);\ -/*printf("check_mv %d %d\n", x, y);*/\ if(map[index]!=key){\ d= cmp(s, x, y, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\ map[index]= key;\ score_map[index]= d;\ d += (mv_penalty[((x)<(ymax<<(S)) ) printf("%d %d %d %d %d ymax" #v, ymax, (x), (y), s->mb_x, const int qpel= flags&FLAG_QPEL;\ const int shift= 1+qpel;\ -static always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin, +static av_always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin, int src_index, int ref_index, int const penalty_factor, int size, int h, int flags) { @@ -564,13 +423,13 @@ static always_inline int small_diamond_search(MpegEncContext * s, int *best, int int next_dir=-1; LOAD_COMMON LOAD_COMMON2 - int map_generation= c->map_generation; + unsigned map_generation = c->map_generation; cmpf= s->dsp.me_cmp[size]; chroma_cmpf= s->dsp.me_cmp[size+1]; { /* ensure that the best point is in the MAP as h/qpel refinement needs it */ - const int key= (best[1]<xmin) CHECK_MV_DIR(x-1, y , 0) if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1) if(dir!=0 && xmap_generation; + unsigned map_generation = c->map_generation; cmpf= s->dsp.me_cmp[size]; chroma_cmpf= s->dsp.me_cmp[size+1]; @@ -635,25 +493,6 @@ static int funny_diamond_search(MpegEncContext * s, int *best, int dmin, 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]); -if(dy>dx){ - dx^=dy; dy^=dx; dx^=dy; -} -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("%8d ", stats[i]); - } - printf("\n"); -} -} -#endif } return dmin; } @@ -666,7 +505,7 @@ static int hex_search(MpegEncContext * s, int *best, int dmin, me_cmp_func cmpf, chroma_cmpf; LOAD_COMMON LOAD_COMMON2 - int map_generation= c->map_generation; + unsigned map_generation = c->map_generation; int x,y,d; const int dec= dia_size & (dia_size-1); @@ -700,7 +539,7 @@ static int l2s_dia_search(MpegEncContext * s, int *best, int dmin, me_cmp_func cmpf, chroma_cmpf; LOAD_COMMON LOAD_COMMON2 - int map_generation= c->map_generation; + unsigned map_generation = c->map_generation; int x,y,i,d; int dia_size= c->dia_size&0xFF; const int dec= dia_size & (dia_size-1); @@ -738,7 +577,7 @@ static int umh_search(MpegEncContext * s, int *best, int dmin, me_cmp_func cmpf, chroma_cmpf; LOAD_COMMON LOAD_COMMON2 - int map_generation= c->map_generation; + unsigned map_generation = c->map_generation; int x,y,x2,y2, i, j, d; const int dia_size= c->dia_size&0xFE; static const int hex[16][2]={{-4,-2}, {-4,-1}, {-4, 0}, {-4, 1}, {-4, 2}, @@ -777,17 +616,50 @@ static int umh_search(MpegEncContext * s, int *best, int dmin, return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, 2); } +static int full_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; + LOAD_COMMON + LOAD_COMMON2 + unsigned map_generation = c->map_generation; + int x,y, d; + const int dia_size= c->dia_size&0xFF; + + cmpf= s->dsp.me_cmp[size]; + chroma_cmpf= s->dsp.me_cmp[size+1]; + + for(y=FFMAX(-dia_size, ymin); y<=FFMIN(dia_size,ymax); y++){ + for(x=FFMAX(-dia_size, xmin); x<=FFMIN(dia_size,xmax); x++){ + CHECK_MV(x, y); + } + } + + x= best[0]; + y= best[1]; + d= dmin; + CHECK_CLIPPED_MV(x , y); + CHECK_CLIPPED_MV(x+1, y); + CHECK_CLIPPED_MV(x, y+1); + CHECK_CLIPPED_MV(x-1, y); + CHECK_CLIPPED_MV(x, y-1); + best[0]= x; + best[1]= y; + + return d; +} + #define SAB_CHECK_MV(ax,ay)\ {\ - const int key= ((ay)<map_generation; + unsigned map_generation = c->map_generation; cmpf= s->dsp.me_cmp[size]; chroma_cmpf= s->dsp.me_cmp[size+1]; - for(j=i=0; i>=ME_MAP_MV_BITS; minima[j].y= key & ((1< 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)<map_generation; + unsigned map_generation = c->map_generation; cmpf= s->dsp.me_cmp[size]; chroma_cmpf= s->dsp.me_cmp[size+1]; @@ -944,27 +823,11 @@ static int var_diamond_search(MpegEncContext * s, int *best, int dmin, 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 always_inline int diamond_search(MpegEncContext * s, int *best, int 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; @@ -974,6 +837,8 @@ static always_inline int diamond_search(MpegEncContext * s, int *best, int dmin, 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>1024) + return full_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) @@ -984,14 +849,25 @@ static always_inline int diamond_search(MpegEncContext * s, int *best, int dmin, return var_diamond_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags); } -static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx_ptr, int *my_ptr, +/** + @param P 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}; - int d, dmin; - int map_generation; + 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[]. */ + unsigned 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 @@ -1017,6 +893,10 @@ static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx map[0]= map_generation; score_map[0]= dmin; + //FIXME precalc first term below? + if((s->pict_type == AV_PICTURE_TYPE_B && !(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) @@ -1087,14 +967,14 @@ static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx *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) +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 @@ -1115,7 +995,7 @@ static int epzs_motion_search4(MpegEncContext * s, MotionEstContext * const c= &s->me; int best[2]={0, 0}; int d, dmin; - int map_generation; + unsigned map_generation; const int penalty_factor= c->penalty_factor; const int size=1; const int h=8; @@ -1132,7 +1012,7 @@ static int epzs_motion_search4(MpegEncContext * s, 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) @@ -1162,7 +1042,6 @@ static int epzs_motion_search4(MpegEncContext * s, *mx_ptr= best[0]; *my_ptr= best[1]; -// printf("%d %d %d \n", best[0], best[1], dmin); return dmin; } @@ -1175,7 +1054,7 @@ static int epzs_motion_search2(MpegEncContext * s, MotionEstContext * const c= &s->me; int best[2]={0, 0}; int d, dmin; - int map_generation; + unsigned map_generation; const int penalty_factor= c->penalty_factor; const int size=0; //FIXME pass as arg const int h=8; @@ -1192,7 +1071,7 @@ static int epzs_motion_search2(MpegEncContext * s, 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) @@ -1222,6 +1101,5 @@ static int epzs_motion_search2(MpegEncContext * s, *mx_ptr= best[0]; *my_ptr= best[1]; -// printf("%d %d %d \n", best[0], best[1], dmin); return dmin; }