X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fmotion_est_template.c;h=c655e199b764ddebfcf124edf1dc957fe0d821f4;hb=e328178da90f44690e0076f4dbfd16da9175f441;hp=837d49a7fc8595b0b85c2fb444e2fab4660fbfce;hpb=f7f8af46f7a0ae7754c5b20e55550a3554384162;p=ffmpeg diff --git a/libavcodec/motion_est_template.c b/libavcodec/motion_est_template.c index 837d49a7fc8..c655e199b76 100644 --- a/libavcodec/motion_est_template.c +++ b/libavcodec/motion_est_template.c @@ -2,36 +2,37 @@ * 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 ...) +#include "mpegvideo.h" + +//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 +41,11 @@ {\ 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, @@ -131,8 +63,8 @@ static int hpel_motion_search(MpegEncContext * s, //FIXME factorize - cmp_sub= s->dsp.me_sub_cmp[size]; - chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; + cmp_sub = s->mecc.me_sub_cmp[size]; + chroma_cmp_sub = s->mecc.me_sub_cmp[size + 1]; if(c->skip){ //FIXME move out of hpel? *mx_ptr = 0; @@ -159,9 +91,8 @@ 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 +104,6 @@ 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; const int penalty_factor= c->mb_penalty_factor; const int flags= c->mb_flags; @@ -249,8 +178,8 @@ inline int ff_get_mb_score(MpegEncContext * s, int mx, int my, int src_index, //FIXME factorize - cmp_sub= s->dsp.mb_cmp[size]; - chroma_cmp_sub= s->dsp.mb_cmp[size+1]; + cmp_sub = s->mecc.mb_cmp[size]; + chroma_cmp_sub = s->mecc.mb_cmp[size + 1]; // assert(!c->skip); // assert(c->avctx->me_sub_cmp != c->avctx->mb_cmp); @@ -263,11 +192,17 @@ 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 +216,7 @@ 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; @@ -290,12 +225,12 @@ static int qpel_motion_search(MpegEncContext * s, LOAD_COMMON int flags= c->sub_flags; - cmpf= s->dsp.me_cmp[size]; - chroma_cmpf= s->dsp.me_cmp[size+1]; //factorize FIXME + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.me_cmp[size + 1]; // FIXME: factorize //FIXME factorize - cmp_sub= s->dsp.me_sub_cmp[size]; - chroma_cmp_sub= s->dsp.me_sub_cmp[size+1]; + cmp_sub = s->mecc.me_sub_cmp[size]; + chroma_cmp_sub = s->mecc.me_sub_cmp[size + 1]; if(c->skip){ //FIXME somehow move up (benchmark) *mx_ptr = 0; @@ -324,7 +259,6 @@ 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; @@ -498,43 +362,39 @@ 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, - int src_index, int ref_index, int const penalty_factor, +static av_always_inline int small_diamond_search(MpegEncContext * s, int *best, int dmin, + int src_index, int ref_index, const int penalty_factor, int size, int h, int flags) { MotionEstContext * const c= &s->me; @@ -564,15 +424,15 @@ 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]; + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.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 && xme; @@ -606,10 +465,10 @@ static int funny_diamond_search(MpegEncContext * s, int *best, int dmin, int dia_size; 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]; + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.me_cmp[size + 1]; for(dia_size=1; dia_size<=4; dia_size++){ int dir; @@ -635,40 +494,173 @@ static int funny_diamond_search(MpegEncContext * s, int *best, int dmin, if(x!=best[0] || y!=best[1]) dia_size=0; -#if 0 + } + return dmin; +} + +static int hex_search(MpegEncContext * s, int *best, int dmin, + int src_index, int ref_index, const int penalty_factor, + int size, int h, int flags, int dia_size) { -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; + 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 dec= dia_size & (dia_size-1); + + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.me_cmp[size + 1]; + + for(;dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){ + do{ + x= best[0]; + y= best[1]; + + CHECK_CLIPPED_MV(x -dia_size , y); + CHECK_CLIPPED_MV(x+ dia_size , y); + CHECK_CLIPPED_MV(x+( dia_size>>1), y+dia_size); + CHECK_CLIPPED_MV(x+( dia_size>>1), y-dia_size); + if(dia_size>1){ + CHECK_CLIPPED_MV(x+(-dia_size>>1), y+dia_size); + CHECK_CLIPPED_MV(x+(-dia_size>>1), y-dia_size); + } + }while(best[0] != x || best[1] != y); + } + + return dmin; } -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]); + +static int l2s_dia_search(MpegEncContext * s, int *best, int dmin, + int src_index, int ref_index, const int 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,i,d; + int dia_size= c->dia_size&0xFF; + const int dec= dia_size & (dia_size-1); + static const int hex[8][2]={{-2, 0}, {-1,-1}, { 0,-2}, { 1,-1}, + { 2, 0}, { 1, 1}, { 0, 2}, {-1, 1}}; + + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.me_cmp[size + 1]; + + for(; dia_size; dia_size= dec ? dia_size-1 : dia_size>>1){ + do{ + x= best[0]; + y= best[1]; + for(i=0; i<8; i++){ + CHECK_CLIPPED_MV(x+hex[i][0]*dia_size, y+hex[i][1]*dia_size); + } + }while(best[0] != x || best[1] != y); } - printf("\n"); + + x= best[0]; + y= best[1]; + CHECK_CLIPPED_MV(x+1, y); + CHECK_CLIPPED_MV(x, y+1); + CHECK_CLIPPED_MV(x-1, y); + CHECK_CLIPPED_MV(x, y-1); + + return dmin; } + +static int umh_search(MpegEncContext * s, int *best, int dmin, + int src_index, int ref_index, const int 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,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}, + { 4,-2}, { 4,-1}, { 4, 0}, { 4, 1}, { 4, 2}, + {-2, 3}, { 0, 4}, { 2, 3}, + {-2,-3}, { 0,-4}, { 2,-3},}; + + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.me_cmp[size + 1]; + + x= best[0]; + y= best[1]; + for(x2=FFMAX(x-dia_size+1, xmin); x2<=FFMIN(x+dia_size-1,xmax); x2+=2){ + CHECK_MV(x2, y); + } + for(y2=FFMAX(y-dia_size/2+1, ymin); y2<=FFMIN(y+dia_size/2-1,ymax); y2+=2){ + CHECK_MV(x, y2); + } + + x= best[0]; + y= best[1]; + for(y2=FFMAX(y-2, ymin); y2<=FFMIN(y+2,ymax); y2++){ + for(x2=FFMAX(x-2, xmin); x2<=FFMIN(x+2,xmax); x2++){ + CHECK_MV(x2, y2); + } + } + +//FIXME prevent the CLIP stuff + + for(j=1; j<=dia_size/4; j++){ + for(i=0; i<16; i++){ + CHECK_CLIPPED_MV(x+hex[i][0]*j, y+hex[i][1]*j); + } + } + + return hex_search(s, best, dmin, src_index, ref_index, penalty_factor, size, h, flags, 2); } -#endif + +static int full_search(MpegEncContext * s, int *best, int dmin, + int src_index, int ref_index, const int 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->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.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); + } } - return dmin; + + 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)<me; @@ -699,25 +691,33 @@ static int sab_diamond_search(MpegEncContext * s, int *best, int dmin, int i, j; 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]; + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.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)< xmin && best[1] < ymax && best[1] > ymin){ int d; - //ensure that the refernece samples for hpel refinement are in the map + // ensure that the reference 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) @@ -769,7 +769,7 @@ static int sab_diamond_search(MpegEncContext * s, int *best, int dmin, } static int var_diamond_search(MpegEncContext * s, int *best, int dmin, - int src_index, int ref_index, int const penalty_factor, + int src_index, int ref_index, const int penalty_factor, int size, int h, int flags) { MotionEstContext * const c= &s->me; @@ -777,10 +777,10 @@ static int var_diamond_search(MpegEncContext * s, int *best, int dmin, int dia_size; 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]; + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.me_cmp[size + 1]; for(dia_size=1; dia_size<=c->dia_size; dia_size++){ int dir, start, end; @@ -825,28 +825,12 @@ 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, - int src_index, int ref_index, int const penalty_factor, +static av_always_inline int diamond_search(MpegEncContext * s, int *best, int dmin, + int src_index, int ref_index, const int penalty_factor, int size, int h, int flags){ MotionEstContext * const c= &s->me; if(c->dia_size==-1) @@ -855,21 +839,40 @@ 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) + 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); } -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 + const int ref_mv_xy = s->mb_x + s->mb_y * ref_mv_stride; // add to last_mv before passing FIXME me_cmp_func cmpf, chroma_cmpf; LOAD_COMMON @@ -877,12 +880,12 @@ static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx 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]; + cmpf = s->mecc.me_pre_cmp[size]; + chroma_cmpf = s->mecc.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]; + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.me_cmp[size + 1]; } map_generation= update_map_generation(c); @@ -892,10 +895,15 @@ 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->mpv_flags & FF_MPV_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_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, + 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) @@ -908,30 +916,28 @@ static always_inline int epzs_motion_search_internal(MpegEncContext * s, int *mx return dmin; } CHECK_MV( P_MEDIAN[0] >>shift , P_MEDIAN[1] >>shift) - CHECK_CLIPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)-1) - CHECK_CLIPED_MV((P_MEDIAN[0]>>shift) , (P_MEDIAN[1]>>shift)+1) - CHECK_CLIPED_MV((P_MEDIAN[0]>>shift)-1, (P_MEDIAN[1]>>shift) ) - CHECK_CLIPED_MV((P_MEDIAN[0]>>shift)+1, (P_MEDIAN[1]>>shift) ) - if(dmin>h*h*2){ - CHECK_CLIPED_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) - } + 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_CLIPED_MV((last_mv[ref_mv_xy-1][0]*ref_mv_scale + (1<<15))>>16, + 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_CLIPED_MV((last_mv[ref_mv_xy-ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, + 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_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, + 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+1end_mb_y) //FIXME replace at least with last_slice_line - CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, + 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) } } @@ -964,14 +970,14 @@ 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) +//this function is dedicated to the brain damaged gcc +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 @@ -992,7 +998,7 @@ 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; @@ -1003,36 +1009,34 @@ static int epzs_motion_search4(MpegEncContext * s, int flags= c->flags; LOAD_COMMON2 - cmpf= s->dsp.me_cmp[size]; - chroma_cmpf= s->dsp.me_cmp[size+1]; + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.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_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, + 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 - if(dmin>64*2){ - 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_CLIPED_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_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_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, + 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+1end_mb_y) //FIXME replace at least with last_slice_line - CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, + 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) } @@ -1041,7 +1045,6 @@ 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; } @@ -1054,7 +1057,7 @@ 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; @@ -1065,36 +1068,34 @@ static int epzs_motion_search2(MpegEncContext * s, int flags= c->flags; LOAD_COMMON2 - cmpf= s->dsp.me_cmp[size]; - chroma_cmpf= s->dsp.me_cmp[size+1]; + cmpf = s->mecc.me_cmp[size]; + chroma_cmpf = s->mecc.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_CLIPED_MV((last_mv[ref_mv_xy][0]*ref_mv_scale + (1<<15))>>16, + 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 - if(dmin>64*2){ - 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_CLIPED_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_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_CLIPED_MV((last_mv[ref_mv_xy+1][0]*ref_mv_scale + (1<<15))>>16, + 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+1end_mb_y) //FIXME replace at least with last_slice_line - CHECK_CLIPED_MV((last_mv[ref_mv_xy+ref_mv_stride][0]*ref_mv_scale + (1<<15))>>16, + 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) } @@ -1103,6 +1104,5 @@ 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; }