*
* new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
*/
+
+/**
+ * @file motion_est.c
+ * Motion estimation.
+ */
+
#include <stdlib.h>
#include <stdio.h>
#include "avcodec.h"
int *mx_ptr, int *my_ptr, int dmin,
int xmin, int ymin, int xmax, int ymax,
int pred_x, int pred_y, Picture *picture,
- int n, int size, uint16_t * const mv_penalty);
+ int n, int size, uint8_t * const mv_penalty);
static inline int update_map_generation(MpegEncContext * s)
{
}Minima;
static int minima_cmp(const void *a, const void *b){
- Minima *da = (Minima *) a;
- Minima *db = (Minima *) b;
+ const Minima *da = (const Minima *) a;
+ const Minima *db = (const Minima *) b;
return da->height - db->height;
}
}
}
-static int pix_dev(UINT8 * pix, int line_size, int mean)
+static int pix_dev(uint8_t * pix, int line_size, int mean)
{
int s, i, j;
{
int x1, y1, x2, y2, xx, yy, x, y;
int mx, my, dmin, d;
- UINT8 *pix;
+ uint8_t *pix;
xx = 16 * s->mb_x;
yy = 16 * s->mb_y;
{
int x1, y1, x2, y2, xx, yy, x, y;
int mx, my, dmin, d;
- UINT8 *pix;
+ uint8_t *pix;
xx = s->mb_x << 4;
yy = s->mb_y << 4;
{
int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
int mx, my, dminx, dminy;
- UINT8 *pix;
+ uint8_t *pix;
xx = s->mb_x << 4;
yy = s->mb_y << 4;
int *mx_ptr, int *my_ptr, int dmin,
int xmin, int ymin, int xmax, int ymax,
int pred_x, int pred_y, Picture *picture,
- int n, int size, uint16_t * const mv_penalty)
+ int n, int size, uint8_t * const mv_penalty)
{
uint8_t *ref_picture= picture->data[0];
uint32_t *score_map= s->me.score_map;
const int penalty_factor= s->me.sub_penalty_factor;
int mx, my, xx, yy, dminh;
- UINT8 *pix, *ptr;
+ uint8_t *pix, *ptr;
op_pixels_abs_func pix_abs_x2;
op_pixels_abs_func pix_abs_y2;
op_pixels_abs_func pix_abs_xy2;
static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
{
- const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2);
+ const int xy= s->mb_x + s->mb_y*s->mb_stride;
s->p_mv_table[xy][0] = mx;
s->p_mv_table[xy][1] = my;
}
}
-static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code)
+/**
+ * get fullpel ME search limits.
+ * @param range the approximate search range for the old ME code, unused for EPZS and newer
+ */
+static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax)
{
- *range = 8 * (1 << (f_code - 1));
- /* XXX: temporary kludge to avoid overflow for msmpeg4 */
- if (s->out_format == FMT_H263 && !s->h263_msmpeg4)
- *range *= 2;
+ if(s->avctx->me_range) *range= s->avctx->me_range >> 1;
+ else *range= 16;
if (s->unrestricted_mv) {
*xmin = -16;
*ymin = -16;
- if (s->h263_plus)
- *range *= 2;
- if(s->avctx->codec->id!=CODEC_ID_MPEG4){
- *xmax = s->mb_width*16;
- *ymax = s->mb_height*16;
- }else {
- *xmax = s->width;
- *ymax = s->height;
- }
+ *xmax = s->mb_width*16;
+ *ymax = s->mb_height*16;
} else {
*xmin = 0;
*ymin = 0;
*xmax = s->mb_width*16 - 16;
*ymax = s->mb_height*16 - 16;
}
+
+ //FIXME try to limit x/y min/max if me_range is set
}
static inline int h263_mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
int block;
int P[10][2];
int dmin_sum=0, mx4_sum=0, my4_sum=0;
- uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
+ uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
for(block=0; block<4; block++){
int mx4, my4;
void ff_estimate_p_frame_motion(MpegEncContext * s,
int mb_x, int mb_y)
{
- UINT8 *pix, *ppix;
+ uint8_t *pix, *ppix;
int sum, varc, vard, mx, my, range, dmin, xx, yy;
int xmin, ymin, xmax, ymax;
int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
int mb_type=0;
uint8_t *ref_picture= s->last_picture.data[0];
Picture * const pic= &s->current_picture;
- uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
+ uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
assert(s->quarter_sample==0 || s->quarter_sample==1);
s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
- get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
+ get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
rel_xmin= xmin - mb_x*16;
rel_xmax= xmax - mb_x*16;
rel_ymin= ymin - mb_y*16;
vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
//printf("%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);
- pic->mb_var [s->mb_width * mb_y + mb_x] = varc;
- pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard;
- pic->mb_mean [s->mb_width * mb_y + mb_x] = (sum+128)>>8;
-// pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin;
+ pic->mb_var [s->mb_stride * mb_y + mb_x] = varc;
+ pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
+ pic->mb_mean [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
+// pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
pic->mb_var_sum += varc;
pic->mc_mb_var_sum += vard;
//printf("E%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);
printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
#endif
- if(s->flags&CODEC_FLAG_HQ){
+ if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
if (vard <= 64 || vard < varc)
s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
else
dmin=dmin4;
}
}
- pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin;
+ pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
if (vard <= 64 || vard < varc) {
}
}
- s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
+ s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
}
int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
int pred_x=0, pred_y=0;
int P[10][2];
const int shift= 1+s->quarter_sample;
- uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
- const int mv_stride= s->mb_width + 2;
- const int xy= mb_x + 1 + (mb_y + 1)*mv_stride;
+ uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
+ const int xy= mb_x + mb_y*s->mb_stride;
assert(s->quarter_sample==0 || s->quarter_sample==1);
s->me.pre_penalty_factor = get_penalty_factor(s, s->avctx->me_pre_cmp);
- get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
+ get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
rel_xmin= xmin - mb_x*16;
rel_xmax= xmax - mb_x*16;
rel_ymin= ymin - mb_y*16;
P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
} else {
- P_TOP[0] = s->p_mv_table[xy + mv_stride ][0];
- P_TOP[1] = s->p_mv_table[xy + mv_stride ][1];
- P_TOPRIGHT[0] = s->p_mv_table[xy + mv_stride - 1][0];
- P_TOPRIGHT[1] = s->p_mv_table[xy + mv_stride - 1][1];
+ P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0];
+ P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1];
+ P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
+ P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
if(P_TOP[1] < (rel_ymin<<shift)) P_TOP[1] = (rel_ymin<<shift);
if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
return dmin;
}
-int ff_estimate_motion_b(MpegEncContext * s,
+static int ff_estimate_motion_b(MpegEncContext * s,
int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
{
int mx, my, range, dmin;
int pred_x=0, pred_y=0;
int P[10][2];
const int shift= 1+s->quarter_sample;
- const int mot_stride = s->mb_width + 2;
- const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
+ const int mot_stride = s->mb_stride;
+ const int mot_xy = mb_y*mot_stride + mb_x;
uint8_t * const ref_picture= picture->data[0];
- uint16_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
+ uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
int mv_scale;
s->me.penalty_factor = get_penalty_factor(s, s->avctx->me_cmp);
s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
- get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code);
+ get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
rel_xmin= xmin - mb_x*16;
rel_xmax= xmax - mb_x*16;
rel_ymin= ymin - mb_y*16;
//FIXME optimize?
//FIXME move into template?
//FIXME better f_code prediction (max mv & distance)
- UINT16 *mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
+ uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
uint8_t *dest_y = s->me.scratchpad;
uint8_t *ptr;
int dxy;
dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
src_x = mb_x * 16 + (motion_fx >> 2);
src_y = mb_y * 16 + (motion_fy >> 2);
- assert(src_x >=-16 && src_x<=s->width);
- assert(src_y >=-16 && src_y<=s->height);
+ assert(src_x >=-16 && src_x<=s->h_edge_pos);
+ assert(src_y >=-16 && src_y<=s->v_edge_pos);
ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
s->dsp.put_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
src_x = mb_x * 16 + (motion_bx >> 2);
src_y = mb_y * 16 + (motion_by >> 2);
- assert(src_x >=-16 && src_x<=s->width);
- assert(src_y >=-16 && src_y<=s->height);
+ assert(src_x >=-16 && src_x<=s->h_edge_pos);
+ assert(src_y >=-16 && src_y<=s->v_edge_pos);
ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y , ptr , s->linesize);
dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
src_x = mb_x * 16 + (motion_fx >> 1);
src_y = mb_y * 16 + (motion_fy >> 1);
- assert(src_x >=-16 && src_x<=s->width);
- assert(src_y >=-16 && src_y<=s->height);
+ assert(src_x >=-16 && src_x<=s->h_edge_pos);
+ assert(src_y >=-16 && src_y<=s->v_edge_pos);
ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
s->dsp.put_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
src_x = mb_x * 16 + (motion_bx >> 1);
src_y = mb_y * 16 + (motion_by >> 1);
- assert(src_x >=-16 && src_x<=s->width);
- assert(src_y >=-16 && src_y<=s->height);
+ assert(src_x >=-16 && src_x<=s->h_edge_pos);
+ assert(src_y >=-16 && src_y<=s->v_edge_pos);
ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
s->dsp.avg_pixels_tab[0][dxy](dest_y , ptr , s->linesize, 16);
static inline int bidir_refine(MpegEncContext * s,
int mb_x, int mb_y)
{
- const int mot_stride = s->mb_width + 2;
- const int xy = (mb_y + 1)*mot_stride + mb_x + 1;
+ const int mot_stride = s->mb_stride;
+ const int xy = mb_y *mot_stride + mb_x;
int fbmin;
int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
int mb_x, int mb_y)
{
int P[10][2];
- const int mot_stride = s->mb_width + 2;
- const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
+ const int mot_stride = s->mb_stride;
+ const int mot_xy = mb_y*mot_stride + mb_x;
const int shift= 1+s->quarter_sample;
int dmin, i;
const int time_pp= s->pp_time;
const int time_pb= s->pb_time;
int mx, my, xmin, xmax, ymin, ymax;
int16_t (*mv_table)[2]= s->b_direct_mv_table;
- uint16_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
+ uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
ymin= xmin=(-32)>>shift;
ymax= xmax= 31>>shift;
- if(s->co_located_type_table[mb_x + mb_y*s->mb_width]==CO_LOCATED_TYPE_4MV){
+ if(IS_8X8(s->next_picture.mb_type[mot_xy])){
s->mv_type= MV_TYPE_8X8;
}else{
s->mv_type= MV_TYPE_16X16;
score= ((unsigned)(score*score + 128*256))>>16;
s->current_picture.mc_mb_var_sum += score;
- s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSE
+ s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
}
- if(s->flags&CODEC_FLAG_HQ){
+ if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
}
- s->mb_type[mb_y*s->mb_width + mb_x]= type;
+ s->mb_type[mb_y*s->mb_stride + mb_x]= type;
}
/* find best f_code for ME which do unlimited searches */
if(s->me_method>=ME_EPZS){
int score[8];
int i, y;
- UINT8 * fcode_tab= s->fcode_tab;
+ uint8_t * fcode_tab= s->fcode_tab;
int best_fcode=-1;
int best_score=-10000000;
for(y=0; y<s->mb_height; y++){
int x;
- int xy= (y+1)* (s->mb_width+2) + 1;
- i= y*s->mb_width;
+ int xy= y*s->mb_stride;
for(x=0; x<s->mb_width; x++){
- if(s->mb_type[i] & type){
+ if(s->mb_type[xy] & type){
int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
fcode_tab[mv_table[xy][1] + MAX_MV]);
int j;
for(j=0; j<fcode && j<8; j++){
- if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[i] < s->current_picture.mb_var[i])
+ if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
score[j]-= 170;
}
}
- i++;
xy++;
}
}
void ff_fix_long_p_mvs(MpegEncContext * s)
{
const int f_code= s->f_code;
- int y;
- UINT8 * fcode_tab= s->fcode_tab;
-//int clip=0;
-//int noclip=0;
+ int y, range;
+
+ range = (((s->codec_id == CODEC_ID_MPEG1VIDEO) ? 8 : 16) << f_code);
+
+ if(s->msmpeg4_version) range= 16;
+
+ if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
+
/* clip / convert to intra 16x16 type MVs */
for(y=0; y<s->mb_height; y++){
int x;
- int xy= (y+1)* (s->mb_width+2)+1;
- int i= y*s->mb_width;
+ int xy= y*s->mb_stride;
for(x=0; x<s->mb_width; x++){
- if(s->mb_type[i]&MB_TYPE_INTER){
- if( fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code
- || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0
- || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code
- || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){
- s->mb_type[i] &= ~MB_TYPE_INTER;
- s->mb_type[i] |= MB_TYPE_INTRA;
+ if(s->mb_type[xy]&MB_TYPE_INTER){
+ if( s->p_mv_table[xy][0] >=range || s->p_mv_table[xy][0] <-range
+ || s->p_mv_table[xy][1] >=range || s->p_mv_table[xy][1] <-range){
+ s->mb_type[xy] &= ~MB_TYPE_INTER;
+ s->mb_type[xy] |= MB_TYPE_INTRA;
s->p_mv_table[xy][0] = 0;
s->p_mv_table[xy][1] = 0;
-//clip++;
}
-//else
-// noclip++;
}
xy++;
- i++;
}
}
//printf("%d no:%d %d//\n", clip, noclip, f_code);
/* clip / convert to intra 8x8 type MVs */
for(y=0; y<s->mb_height; y++){
int xy= (y*2 + 1)*wrap + 1;
- int i= y*s->mb_width;
+ int i= y*s->mb_stride;
int x;
for(x=0; x<s->mb_width; x++){
int mx= s->motion_val[ xy + off ][0];
int my= s->motion_val[ xy + off ][1];
- if( fcode_tab[mx + MAX_MV] > f_code
- || fcode_tab[mx + MAX_MV] == 0
- || fcode_tab[my + MAX_MV] > f_code
- || fcode_tab[my + MAX_MV] == 0 ){
+ if( mx >=range || mx <-range
+ || my >=range || my <-range){
s->mb_type[i] &= ~MB_TYPE_INTER4V;
s->mb_type[i] |= MB_TYPE_INTRA;
}
void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
{
int y;
- UINT8 * fcode_tab= s->fcode_tab;
// RAL: 8 in MPEG-1, 16 in MPEG-4
int range = (((s->codec_id == CODEC_ID_MPEG1VIDEO) ? 8 : 16) << f_code);
+
+ if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
/* clip / convert to intra 16x16 type MVs */
for(y=0; y<s->mb_height; y++){
int x;
- int xy= (y+1)* (s->mb_width+2)+1;
- int i= y*s->mb_width;
- for(x=0; x<s->mb_width; x++)
- {
- if (s->mb_type[i] & type) // RAL: "type" test added...
- {
- if (fcode_tab[mv_table[xy][0] + MAX_MV] > f_code || fcode_tab[mv_table[xy][0] + MAX_MV] == 0)
- {
- if(mv_table[xy][0]>0)
- mv_table[xy][0]= range-1;
- else
- mv_table[xy][0]= -range;
+ int xy= y*s->mb_stride;
+ for(x=0; x<s->mb_width; x++){
+ if (s->mb_type[xy] & type){ // RAL: "type" test added...
+ if( mv_table[xy][0] >=range || mv_table[xy][0] <-range
+ || mv_table[xy][1] >=range || mv_table[xy][1] <-range){
+
+ if(s->codec_id == CODEC_ID_MPEG1VIDEO && 0){
+ }else{
+ if (mv_table[xy][0] > range-1) mv_table[xy][0]= range-1;
+ else if(mv_table[xy][0] < -range ) mv_table[xy][0]= -range;
+ if (mv_table[xy][1] > range-1) mv_table[xy][1]= range-1;
+ else if(mv_table[xy][1] < -range ) mv_table[xy][1]= -range;
}
- if (fcode_tab[mv_table[xy][1] + MAX_MV] > f_code || fcode_tab[mv_table[xy][1] + MAX_MV] == 0)
- {
- if(mv_table[xy][1]>0)
- mv_table[xy][1]= range-1;
- else
- mv_table[xy][1]= -range;
- }
+ }
}
xy++;
- i++;
}
}
}
projects / ffmpeg / blobdiff