* 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;\
{\
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,
//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;
const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]
+ (mv_penalty[bx - pred_x] + mv_penalty[by+2 - pred_y])*c->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
assert(map[(index+1)&(ME_MAP_SIZE-1)] == key);
key= ((my)<<ME_MAP_MV_BITS) + (mx-1) + map_generation;
assert(map[(index-1)&(ME_MAP_SIZE-1)] == key);
-#endif
if(t<=b){
CHECK_HALF_MV(0, 1, mx ,my-1)
if(l<=r){
return dmin;
}
-#endif
static int no_sub_motion_search(MpegEncContext * s,
int *mx_ptr, int *my_ptr, int dmin,
return dmin;
}
-inline int ff_get_mb_score(MpegEncContext * s, int mx, int my, int src_index,
- int ref_index, int size, int h, int add_rate)
+static inline int get_mb_score(MpegEncContext *s, int mx, int my,
+ int src_index, int ref_index, int size,
+ int h, int add_rate)
{
-// const int check_luma= s->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;
//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);
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)\
}
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;
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;
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<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
CHECK_QUARTER_MV(nx&3, ny&3, nx>>2, ny>>2)
}
-#if 0
- const int tl= score_map[(index-(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
- const int bl= score_map[(index+(1<<ME_MAP_SHIFT)-1)&(ME_MAP_SIZE-1)];
- const int tr= score_map[(index-(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
- const int br= score_map[(index+(1<<ME_MAP_SHIFT)+1)&(ME_MAP_SIZE-1)];
-// if(l < r && l < t && l < b && l < tl && l < bl && l < tr && l < br && bl < tl){
- if(tl<br){
-
-// nx= FFMAX(4*mx - bx, bx - 4*mx);
-// ny= FFMAX(4*my - by, by - 4*my);
-
- static int stats[7][7], count;
- count++;
- stats[4*mx - bx + 3][4*my - by + 3]++;
- if(256*256*256*64 % count ==0){
- for(i=0; i<49; i++){
- if((i%7)==0) printf("\n");
- printf("%6d ", stats[0][i]);
- }
- printf("\n");
- }
- }
-#endif
-#else
-
- CHECK_QUARTER_MV(2, 2, mx-1, my-1)
- CHECK_QUARTER_MV(0, 2, mx , my-1)
- CHECK_QUARTER_MV(2, 2, mx , my-1)
- CHECK_QUARTER_MV(2, 0, mx , my )
- CHECK_QUARTER_MV(2, 2, mx , my )
- CHECK_QUARTER_MV(0, 2, mx , my )
- CHECK_QUARTER_MV(2, 2, mx-1, my )
- CHECK_QUARTER_MV(2, 0, mx-1, my )
-
- nx= bx;
- ny= by;
-
- for(i=0; i<8; i++){
- int ox[8]= {0, 1, 1, 1, 0,-1,-1,-1};
- int oy[8]= {1, 1, 0,-1,-1,-1, 0, 1};
- CHECK_QUARTER_MV((nx + ox[i])&3, (ny + oy[i])&3, (nx + ox[i])>>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;
#define CHECK_MV(x,y)\
{\
- const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
+ const unsigned key = ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
assert((x) >= 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)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
-/*printf("score:%d\n", d);*/\
COPY3_IF_LT(dmin, d, best[0], x, best[1], y)\
}\
}
-#define CHECK_CLIPED_MV(ax,ay)\
+#define CHECK_CLIPPED_MV(ax,ay)\
{\
- const int x= ax;\
- const int y= ay;\
- const int x2= FFMAX(xmin, FFMIN(x, xmax));\
- const int y2= FFMAX(ymin, FFMIN(y, ymax));\
- CHECK_MV(x2, y2)\
+ const int Lx= ax;\
+ const int Ly= ay;\
+ const int Lx2= FFMAX(xmin, FFMIN(Lx, xmax));\
+ const int Ly2= FFMAX(ymin, FFMIN(Ly, ymax));\
+ CHECK_MV(Lx2, Ly2)\
}
#define CHECK_MV_DIR(x,y,new_dir)\
{\
- const int key= ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
+ const unsigned key = ((y)<<ME_MAP_MV_BITS) + (x) + map_generation;\
const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);\
-/*printf("check_mv_dir %d %d %d\n", x, y, new_dir);*/\
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)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*penalty_factor;\
-/*printf("score:%d\n", d);*/\
if(d<dmin){\
best[0]=x;\
best[1]=y;\
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)
{
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]<<ME_MAP_MV_BITS) + best[0] + map_generation;
+ const unsigned key = (best[1]<<ME_MAP_MV_BITS) + best[0] + map_generation;
const int index= ((best[1]<<ME_MAP_SHIFT) + best[0])&(ME_MAP_SIZE-1);
if(map[index]!=key){ //this will be executed only very rarey
score_map[index]= cmp(s, best[0], best[1], 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);
const int y= best[1];
next_dir=-1;
-//printf("%d", dir);
if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y , 0)
if(dir!=3 && y>ymin) CHECK_MV_DIR(x , y-1, 1)
if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y , 2)
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;
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, int const 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, 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,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, 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,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, 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->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_MAP_MV_BITS) + (ax) + map_generation;\
+ const unsigned key = ((ay)<<ME_MAP_MV_BITS) + (ax) + map_generation;\
const int index= (((ay)<<ME_MAP_SHIFT) + (ax))&(ME_MAP_SIZE-1);\
-/*printf("sab check %d %d\n", ax, ay);*/\
if(map[index]!=key){\
d= cmp(s, ax, ay, 0, 0, size, h, ref_index, src_index, cmpf, chroma_cmpf, flags);\
map[index]= key;\
score_map[index]= d;\
d += (mv_penalty[((ax)<<shift)-pred_x] + mv_penalty[((ay)<<shift)-pred_y])*penalty_factor;\
-/*printf("score: %d\n", d);*/\
if(d < minima[minima_count-1].height){\
int j=0;\
\
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_SIZE; i++){
+ /*Note j<MAX_SAB_SIZE is needed if MAX_SAB_SIZE < ME_MAP_SIZE as j can
+ become larger due to MVs overflowing their ME_MAP_MV_BITS bits space in map
+ */
+ for(j=i=0; i<ME_MAP_SIZE && j<MAX_SAB_SIZE; i++){
uint32_t key= map[i];
key += (1<<(ME_MAP_MV_BITS-1)) + (1<<(2*ME_MAP_MV_BITS-1));
if((key&((-1)<<(2*ME_MAP_MV_BITS))) != map_generation) continue;
- assert(j<MAX_SAB_SIZE); //max j = number of predictors
-
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;
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;
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;
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
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);
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)
c->skip=1;
return dmin;
}
- CHECK_MV(P_MEDIAN[0]>>shift, 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_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_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+1<s->end_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)
}
}
*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
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;
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+1<s->end_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)
}
*mx_ptr= best[0];
*my_ptr= best[1];
-// printf("%d %d %d \n", best[0], best[1], dmin);
return dmin;
}
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;
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+1<s->end_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)
}
*mx_ptr= best[0];
*my_ptr= best[1];
-// printf("%d %d %d \n", best[0], best[1], dmin);
return dmin;
}