* Copyright (c) 2000,2001 Fabrice Bellard.
* Copyright (c) 2002-2004 Michael Niedermayer
*
+ * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
*
- * This library is free software; you can redistribute it and/or
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg 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 of the License, or (at your option) any later version.
+ * version 2.1 of the License, or (at your option) any later version.
*
- * This library is distributed in the hope that it will be useful,
+ * FFmpeg 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 this library; if not, write to the Free Software
+ * License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
- *
- * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
*/
/**
+ (chroma ? FLAG_CHROMA : 0);
}
-static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
+/*! \brief compares a block (either a full macroblock or a partition thereof)
+ against a proposed motion-compensated prediction of that block
+ */
+static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
const int size, const int h, int ref_index, int src_index,
me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
MotionEstContext * const c= &s->me;
int d;
//FIXME check chroma 4mv, (no crashes ...)
if(flags&FLAG_DIRECT){
+ assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
const int time_pp= s->pp_time;
const int time_pb= s->pb_time;
static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
}
-void ff_init_me(MpegEncContext *s){
+int ff_init_me(MpegEncContext *s){
MotionEstContext * const c= &s->me;
+ int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
+ int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
+
+ if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){
+ av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
+ return -1;
+ }
+ //special case of snow is needed because snow uses its own iterative ME code
+ if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != CODEC_ID_SNOW){
+ av_log(s->avctx, AV_LOG_ERROR, "me_method is only allowed to be set to zero and epzs; for hex,umh,full and others see dia_size\n");
+ return -1;
+ }
+
c->avctx= s->avctx;
+ if(cache_size < 2*dia_size && !c->stride){
+ av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
+ }
+
ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
c->uvstride= 8*s->mb_width + 16;
}
- // 8x8 fullpel search would need a 4x4 chroma compare, which we dont have yet, and even if we had the motion estimation code doesnt expect it
+ /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
+ * not have yet, and even if we had, the motion estimation code
+ * does not expect it. */
if(s->codec_id != CODEC_ID_SNOW){
if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
s->dsp.me_cmp[2]= zero_cmp;
}
c->temp= c->scratchpad;
+
+ return 0;
}
#if 0
s = 0;
for (i = 0; i < 16; i++) {
for (j = 0; j < 16; j += 8) {
- s += ABS(pix[0]-mean);
- s += ABS(pix[1]-mean);
- s += ABS(pix[2]-mean);
- s += ABS(pix[3]-mean);
- s += ABS(pix[4]-mean);
- s += ABS(pix[5]-mean);
- s += ABS(pix[6]-mean);
- s += ABS(pix[7]-mean);
+ s += FFABS(pix[0]-mean);
+ s += FFABS(pix[1]-mean);
+ s += FFABS(pix[2]-mean);
+ s += FFABS(pix[3]-mean);
+ s += FFABS(pix[4]-mean);
+ s += FFABS(pix[5]-mean);
+ s += FFABS(pix[6]-mean);
+ s += FFABS(pix[7]-mean);
pix += 8;
}
pix += line_size - 16;
*my_ptr = 16 * s->mb_y;
}
-#if 0 /* the use of these functions is inside #if 0 */
-static int full_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr, int range,
- int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
-{
- int x1, y1, x2, y2, xx, yy, x, y;
- int mx, my, dmin, d;
- uint8_t *pix;
-
- xx = 16 * s->mb_x;
- yy = 16 * s->mb_y;
- x1 = xx - range + 1; /* we loose one pixel to avoid boundary pb with half pixel pred */
- if (x1 < xmin)
- x1 = xmin;
- x2 = xx + range - 1;
- if (x2 > xmax)
- x2 = xmax;
- y1 = yy - range + 1;
- if (y1 < ymin)
- y1 = ymin;
- y2 = yy + range - 1;
- if (y2 > ymax)
- y2 = ymax;
- pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
- dmin = 0x7fffffff;
- mx = 0;
- my = 0;
- for (y = y1; y <= y2; y++) {
- for (x = x1; x <= x2; x++) {
- d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
- s->linesize, 16);
- if (d < dmin ||
- (d == dmin &&
- (abs(x - xx) + abs(y - yy)) <
- (abs(mx - xx) + abs(my - yy)))) {
- dmin = d;
- mx = x;
- my = y;
- }
- }
- }
-
- *mx_ptr = mx;
- *my_ptr = my;
-
-#if 0
- if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
- *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
- av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
- }
-#endif
- return dmin;
-}
-
-
-static int log_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr, int range,
- int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
-{
- int x1, y1, x2, y2, xx, yy, x, y;
- int mx, my, dmin, d;
- uint8_t *pix;
-
- xx = s->mb_x << 4;
- yy = s->mb_y << 4;
-
- /* Left limit */
- x1 = xx - range;
- if (x1 < xmin)
- x1 = xmin;
-
- /* Right limit */
- x2 = xx + range;
- if (x2 > xmax)
- x2 = xmax;
-
- /* Upper limit */
- y1 = yy - range;
- if (y1 < ymin)
- y1 = ymin;
-
- /* Lower limit */
- y2 = yy + range;
- if (y2 > ymax)
- y2 = ymax;
-
- pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
- dmin = 0x7fffffff;
- mx = 0;
- my = 0;
-
- do {
- for (y = y1; y <= y2; y += range) {
- for (x = x1; x <= x2; x += range) {
- d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
- if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
- dmin = d;
- mx = x;
- my = y;
- }
- }
- }
-
- range = range >> 1;
-
- x1 = mx - range;
- if (x1 < xmin)
- x1 = xmin;
-
- x2 = mx + range;
- if (x2 > xmax)
- x2 = xmax;
-
- y1 = my - range;
- if (y1 < ymin)
- y1 = ymin;
-
- y2 = my + range;
- if (y2 > ymax)
- y2 = ymax;
-
- } while (range >= 1);
-
-#ifdef DEBUG
- av_log(s->avctx, AV_LOG_DEBUG, "log - MX: %d\tMY: %d\n", mx, my);
-#endif
- *mx_ptr = mx;
- *my_ptr = my;
- return dmin;
-}
-
-static int phods_motion_search(MpegEncContext * s,
- int *mx_ptr, int *my_ptr, int range,
- int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
-{
- int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
- int mx, my, dminx, dminy;
- uint8_t *pix;
-
- xx = s->mb_x << 4;
- yy = s->mb_y << 4;
-
- /* Left limit */
- x1 = xx - range;
- if (x1 < xmin)
- x1 = xmin;
-
- /* Right limit */
- x2 = xx + range;
- if (x2 > xmax)
- x2 = xmax;
-
- /* Upper limit */
- y1 = yy - range;
- if (y1 < ymin)
- y1 = ymin;
-
- /* Lower limit */
- y2 = yy + range;
- if (y2 > ymax)
- y2 = ymax;
-
- pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
- mx = 0;
- my = 0;
-
- x = xx;
- y = yy;
- do {
- dminx = 0x7fffffff;
- dminy = 0x7fffffff;
-
- lastx = x;
- for (x = x1; x <= x2; x += range) {
- d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
- if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
- dminx = d;
- mx = x;
- }
- }
-
- x = lastx;
- for (y = y1; y <= y2; y += range) {
- d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
- if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
- dminy = d;
- my = y;
- }
- }
-
- range = range >> 1;
-
- x = mx;
- y = my;
- x1 = mx - range;
- if (x1 < xmin)
- x1 = xmin;
-
- x2 = mx + range;
- if (x2 > xmax)
- x2 = xmax;
-
- y1 = my - range;
- if (y1 < ymin)
- y1 = ymin;
-
- y2 = my + range;
- if (y2 > ymax)
- y2 = ymax;
-
- } while (range >= 1);
-
-#ifdef DEBUG
- av_log(s->avctx, AV_LOG_DEBUG, "phods - MX: %d\tMY: %d\n", mx, my);
-#endif
-
- /* half pixel search */
- *mx_ptr = mx;
- *my_ptr = my;
- return dminy;
-}
-#endif /* 0 */
-
#define Z_THRESHOLD 256
#define CHECK_SAD_HALF_MV(suffix, x, y) \
static inline void get_limits(MpegEncContext *s, int x, int y)
{
MotionEstContext * const c= &s->me;
+ int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
/*
if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
else c->range= 16;
c->xmax = - x + s->mb_width *16 - 16;
c->ymax = - y + s->mb_height*16 - 16;
}
+ if(range){
+ c->xmin = FFMAX(c->xmin,-range);
+ c->xmax = FFMIN(c->xmax, range);
+ c->ymin = FFMAX(c->ymin,-range);
+ c->ymax = FFMIN(c->ymax, range);
+ }
}
static inline void init_mv4_ref(MotionEstContext *c){
{
MotionEstContext * const c= &s->me;
uint8_t *pix, *ppix;
- int sum, varc, vard, mx, my, dmin;
+ int sum, mx, my, dmin;
+ int varc; ///< the variance of the block (sum of squared (p[y][x]-average))
+ int vard; ///< sum of squared differences with the estimated motion vector
int P[10][2];
const int shift= 1+s->quarter_sample;
int mb_type=0;
vard= check_input_motion(s, mb_x, mb_y, 1);
if((vard+128)>>8 < c->avctx->me_threshold){
+ int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+ int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
c->mc_mb_var_sum_temp += (vard+128)>>8;
- if (vard <= 64<<8 || vard < varc) { //FIXME
- c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
- }else{
- c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
- }
+ c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
return;
}
if((vard+128)>>8 < c->avctx->mb_threshold)
my-= mb_y*16;
dmin = 0;
break;
-#if 0
- case ME_FULL:
- dmin = full_motion_search(s, &mx, &my, range, ref_picture);
- mx-= mb_x*16;
- my-= mb_y*16;
- break;
- case ME_LOG:
- dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
- mx-= mb_x*16;
- my-= mb_y*16;
- break;
- case ME_PHODS:
- dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
- mx-= mb_x*16;
- my-= mb_y*16;
- break;
-#endif
case ME_X1:
case ME_EPZS:
{
varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
#endif
if(mb_type){
- if (vard <= 64<<8 || vard < varc)
- c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
- else
- c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
+ int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+ int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+ c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
if(mb_type == CANDIDATE_MB_TYPE_INTER){
c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
}
}else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
- if (vard <= 64<<8 || vard < varc)
- c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
- else
- c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
+ int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+ int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+ c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
if (vard*2 + 200*256 > varc)
mb_type|= CANDIDATE_MB_TYPE_INTRA;
- if (varc*2 + 200*256 > vard){
+ if (varc*2 + 200*256 > vard || s->qscale > 24){
+// if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
mb_type|= CANDIDATE_MB_TYPE_INTER;
c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
if(s->flags&CODEC_FLAG_MV0)
}else
s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
- if (vard <= 64<<8 || vard < varc) { //FIXME
- c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
- }else{
- c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
+ {
+ int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
+ int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
+ c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
}
}
mx-= mb_x*16;
my-= mb_y*16;
break;
-#if 0
- case ME_FULL:
- dmin = full_motion_search(s, &mx, &my, range, ref_picture);
- mx-= mb_x*16;
- my-= mb_y*16;
- break;
- case ME_LOG:
- dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
- mx-= mb_x*16;
- my-= mb_y*16;
- break;
- case ME_PHODS:
- dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
- mx-= mb_x*16;
- my-= mb_y*16;
- break;
-#endif
case ME_X1:
case ME_EPZS:
{
}
#define CHECK_BIDIR2(a,b,c,d)\
CHECK_BIDIR(a,b,c,d)\
-CHECK_BIDIR(-a,-b,-c,-d)
+CHECK_BIDIR(-(a),-(b),-(c),-(d))
#define CHECK_BIDIRR(a,b,c,d)\
CHECK_BIDIR2(a,b,c,d)\
c->pred_x=0;
c->pred_y=0;
- P_LEFT[0] = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
- P_LEFT[1] = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
+ P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
+ P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
/* special case for first line */
- if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its cliped
- P_TOP[0] = clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
- P_TOP[1] = clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
- P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
- P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
+ if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
+ P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift);
+ P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift);
+ P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift);
+ P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift);
P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
- s->b_direct_mv_table[mot_xy][0]= mx;
- s->b_direct_mv_table[mot_xy][1]= my;
+ mv_table[mot_xy][0]= mx;
+ mv_table[mot_xy][1]= my;
c->flags &= ~FLAG_DIRECT;
c->sub_flags &= ~FLAG_DIRECT;
get_limits(s, 16*mb_x, 16*mb_y);
c->skip=0;
+
+ if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
+ int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
+
+ score= ((unsigned)(score*score + 128*256))>>16;
+ c->mc_mb_var_sum_temp += score;
+ s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
+ s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
+
+ return;
+ }
+
if(c->avctx->me_threshold){
int vard= check_input_motion(s, mb_x, mb_y, 0);
type |= CANDIDATE_MB_TYPE_BIDIR_I;
}
//FIXME something smarter
- if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
+ if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
+ if(s->codec_id == CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
+ type |= CANDIDATE_MB_TYPE_DIRECT0;
#if 0
if(s->out_format == FMT_MPEG1)
type |= CANDIDATE_MB_TYPE_INTRA;
continue;
for(j=0; j<fcode && j<8; j++){
- if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
+ if(s->pict_type==FF_B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
score[j]-= 170;
}
}
MotionEstContext * const c= &s->me;
const int f_code= s->f_code;
int y, range;
- assert(s->pict_type==P_TYPE);
+ assert(s->pict_type==FF_P_TYPE);
range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);