aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorMichael Niedermayer <michaelni@gmx.at>2002-04-19 03:25:20 +0000
committerMichael Niedermayer <michaelni@gmx.at>2002-04-19 03:25:20 +0000
commit91029be79005b17311fcf75f3f9f977f4d20cfd7 (patch)
tree672c42c26c923719256f5e308d1d077e23ff13bb
parent1f0cd30fd9b656122436ecd625656a04f6235fb3 (diff)
downloadffmpeg-91029be79005b17311fcf75f3f9f977f4d20cfd7.tar.gz
much better ME for b frames (a bit slow though)
fixed MC rounding for b frames fixed hq mode with b-frames Originally committed as revision 406 to svn://svn.ffmpeg.org/ffmpeg/trunk
-rw-r--r--libavcodec/common.h10
-rw-r--r--libavcodec/h263.c56
-rw-r--r--libavcodec/motion_est.c269
-rw-r--r--libavcodec/mpegvideo.c194
-rw-r--r--libavcodec/mpegvideo.h11
5 files changed, 446 insertions, 94 deletions
diff --git a/libavcodec/common.h b/libavcodec/common.h
index 96b545d0dc..21cdd541b4 100644
--- a/libavcodec/common.h
+++ b/libavcodec/common.h
@@ -881,6 +881,16 @@ static inline int mid_pred(int a, int b, int c)
return a + b + c - vmin - vmax;
}
+static inline int clip(int a, int amin, int amax)
+{
+ if (a < amin)
+ return amin;
+ else if (a > amax)
+ return amax;
+ else
+ return a;
+}
+
/* memory */
void *av_mallocz(int size);
diff --git a/libavcodec/h263.c b/libavcodec/h263.c
index bcb9ea2508..42ae076ce4 100644
--- a/libavcodec/h263.c
+++ b/libavcodec/h263.c
@@ -18,7 +18,7 @@
* along with this program; if not, write to the Free Software
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
*
- * ac prediction encoding by Michael Niedermayer <michaelni@gmx.at>
+ * ac prediction encoding & b-frame support by Michael Niedermayer <michaelni@gmx.at>
*/
#include "common.h"
#include "dsputil.h"
@@ -282,7 +282,7 @@ void mpeg4_encode_mb(MpegEncContext * s,
s->mv[0][0][1]=
s->mv[1][0][0]=
s->mv[1][0][1]= 0;
-// s->mv_dir= MV_DIR_FORWARD; //doesnt matter
+ s->mv_dir= MV_DIR_FORWARD; //doesnt matter
return;
}
@@ -334,7 +334,8 @@ void mpeg4_encode_mb(MpegEncContext * s,
s->last_mv[0][0][0]= motion_x;
s->last_mv[0][0][1]= motion_y;
break;
- default:
+ default:
+ printf("unknown mb type\n");
return;
}
bits= get_bit_count(&s->pb);
@@ -959,6 +960,31 @@ static void put_string(PutBitContext * pbc, char *s)
put_bits(pbc, 8, 0);
}
+/* must be called before writing the header */
+void ff_set_mpeg4_time(MpegEncContext * s, int picture_number){
+ int time_div, time_mod;
+
+ if(s->pict_type==I_TYPE){ //we will encode a vol header
+ s->time_increment_resolution= s->frame_rate/ff_gcd(s->frame_rate, FRAME_RATE_BASE);
+ if(s->time_increment_resolution>=256*256) s->time_increment_resolution= 256*128;
+
+ s->time_increment_bits = av_log2(s->time_increment_resolution - 1) + 1;
+ }
+
+ s->time= picture_number*(int64_t)FRAME_RATE_BASE*s->time_increment_resolution/s->frame_rate;
+ time_div= s->time/s->time_increment_resolution;
+ time_mod= s->time%s->time_increment_resolution;
+
+ if(s->pict_type==B_TYPE){
+ s->bp_time= s->last_non_b_time - s->time;
+ }else{
+ s->last_time_base= s->time_base;
+ s->time_base= time_div;
+ s->pp_time= s->time - s->last_non_b_time;
+ s->last_non_b_time= s->time;
+ }
+}
+
static void mpeg4_encode_vol_header(MpegEncContext * s)
{
int vo_ver_id=1; //must be 2 if we want GMC or q-pel
@@ -983,11 +1009,7 @@ static void mpeg4_encode_vol_header(MpegEncContext * s)
put_bits(&s->pb, 2, RECT_SHAPE); /* vol shape= rectangle */
put_bits(&s->pb, 1, 1); /* marker bit */
- s->time_increment_resolution= s->frame_rate/ff_gcd(s->frame_rate, FRAME_RATE_BASE);
- if(s->time_increment_resolution>=256*256) s->time_increment_resolution= 256*128;
-
put_bits(&s->pb, 16, s->time_increment_resolution);
- s->time_increment_bits = av_log2(s->time_increment_resolution - 1) + 1;
if (s->time_increment_bits < 1)
s->time_increment_bits = 1;
put_bits(&s->pb, 1, 1); /* marker bit */
@@ -1034,9 +1056,6 @@ void mpeg4_encode_picture_header(MpegEncContext * s, int picture_number)
if(s->pict_type==I_TYPE) mpeg4_encode_vol_header(s);
- s->time= s->picture_number*(int64_t)FRAME_RATE_BASE*s->time_increment_resolution/s->frame_rate;
- time_div= s->time/s->time_increment_resolution;
- time_mod= s->time%s->time_increment_resolution;
//printf("num:%d rate:%d base:%d\n", s->picture_number, s->frame_rate, FRAME_RATE_BASE);
if(get_bit_count(&s->pb)!=0) mpeg4_stuffing(&s->pb);
@@ -1044,15 +1063,8 @@ void mpeg4_encode_picture_header(MpegEncContext * s, int picture_number)
put_bits(&s->pb, 16, 0x1B6); /* vop header */
put_bits(&s->pb, 2, s->pict_type - 1); /* pict type: I = 0 , P = 1 */
- if(s->pict_type==B_TYPE){
- s->bp_time= s->last_non_b_time - s->time;
- }else{
- s->last_time_base= s->time_base;
- s->time_base= time_div;
- s->pp_time= s->time - s->last_non_b_time;
- s->last_non_b_time= s->time;
- }
-
+ time_div= s->time/s->time_increment_resolution;
+ time_mod= s->time%s->time_increment_resolution;
time_incr= time_div - s->last_time_base;
while(time_incr--)
put_bits(&s->pb, 1, 1);
@@ -1770,6 +1782,7 @@ int h263_decode_mb(MpegEncContext *s,
s->last_mv[0][0][1]=
s->last_mv[1][0][0]=
s->last_mv[1][0][1]= 0;
+// printf("\n");
}
/* if we skipped it in the future P Frame than skip it now too */
@@ -1789,6 +1802,7 @@ int h263_decode_mb(MpegEncContext *s,
//FIXME is this correct?
/* s->last_mv[0][0][0]=
s->last_mv[0][0][1]=0;*/
+// printf("S");
return 0;
}
@@ -1837,6 +1851,7 @@ int h263_decode_mb(MpegEncContext *s,
s->mv[0][0][1] =
s->mv[1][0][0] =
s->mv[1][0][1] = 1000;*/
+// printf("D");
break;
case 1:
s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
@@ -1849,6 +1864,7 @@ int h263_decode_mb(MpegEncContext *s,
my = h263_decode_motion(s, s->last_mv[1][0][1], s->b_code);
s->last_mv[1][0][0]= s->mv[1][0][0] = mx;
s->last_mv[1][0][1]= s->mv[1][0][1] = my;
+// printf("I");
break;
case 2:
s->mv_dir = MV_DIR_BACKWARD;
@@ -1856,6 +1872,7 @@ int h263_decode_mb(MpegEncContext *s,
my = h263_decode_motion(s, s->last_mv[1][0][1], s->b_code);
s->last_mv[1][0][0]= s->mv[1][0][0] = mx;
s->last_mv[1][0][1]= s->mv[1][0][1] = my;
+// printf("B");
break;
case 3:
s->mv_dir = MV_DIR_FORWARD;
@@ -1863,6 +1880,7 @@ int h263_decode_mb(MpegEncContext *s,
my = h263_decode_motion(s, s->last_mv[0][0][1], s->f_code);
s->last_mv[0][0][0]= s->mv[0][0][0] = mx;
s->last_mv[0][0][1]= s->mv[0][0][1] = my;
+// printf("F");
break;
default: return -1;
}
diff --git a/libavcodec/motion_est.c b/libavcodec/motion_est.c
index 7fc0cec18c..1ba246ae5c 100644
--- a/libavcodec/motion_est.c
+++ b/libavcodec/motion_est.c
@@ -29,7 +29,7 @@
#define MAX(a,b) ((a) > (b) ? (a) : (b))
#define INTER_BIAS 257
-static void halfpel_motion_search(MpegEncContext * s,
+static int halfpel_motion_search(MpegEncContext * s,
int *mx_ptr, int *my_ptr, int dmin,
int xmin, int ymin, int xmax, int ymax,
int pred_x, int pred_y, uint8_t *ref_picture);
@@ -673,7 +673,7 @@ static int epzs_motion_search4(MpegEncContext * s, int block,
/* The idea would be to make half pel ME after Inter/Intra decision to
save time. */
-static inline void halfpel_motion_search(MpegEncContext * s,
+static inline int halfpel_motion_search(MpegEncContext * s,
int *mx_ptr, int *my_ptr, int dmin,
int xmin, int ymin, int xmax, int ymax,
int pred_x, int pred_y, uint8_t *ref_picture)
@@ -702,7 +702,7 @@ static inline void halfpel_motion_search(MpegEncContext * s,
if(dmin < Z_THRESHOLD && mx==0 && my==0){
*mx_ptr = 0;
*my_ptr = 0;
- return;
+ return dmin;
}
pen_x= pred_x + mx;
@@ -727,6 +727,7 @@ static inline void halfpel_motion_search(MpegEncContext * s,
*mx_ptr = mx;
*my_ptr = my;
+ return dminh;
}
static inline void halfpel_motion_search4(MpegEncContext * s,
@@ -1044,17 +1045,15 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
set_p_mv_tables(s, mx, my);
}
-void ff_estimate_motion_b(MpegEncContext * s,
+int ff_estimate_motion_b(MpegEncContext * s,
int mb_x, int mb_y, int16_t (*mv_table)[2], uint8_t *ref_picture, int f_code)
{
- UINT8 *pix, *ppix;
- int sum, varc, vard, mx, my, range, dmin, xx, yy;
+ int mx, my, range, dmin;
int xmin, ymin, xmax, ymax;
int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
int pred_x=0, pred_y=0;
int P[6][2];
const int shift= 1+s->quarter_sample;
- int mb_type=0;
const int mot_stride = s->mb_width + 2;
const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
@@ -1124,18 +1123,210 @@ void ff_estimate_motion_b(MpegEncContext * s,
/* At this point (mx,my) are full-pell and the absolute displacement */
// ppix = ref_picture + (my * s->linesize) + mx;
- halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y, ref_picture);
+ dmin= halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y, ref_picture);
// s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
mv_table[mot_xy][0]= mx;
mv_table[mot_xy][1]= my;
+ return dmin;
}
-int ff_decide_type(MpegEncContext * s,
- int mb_x, int mb_y)
+static inline int check_bidir_mv(MpegEncContext * s,
+ int mb_x, int mb_y,
+ int motion_fx, int motion_fy,
+ int motion_bx, int motion_by,
+ int pred_fx, int pred_fy,
+ int pred_bx, int pred_by)
{
+ //FIXME optimize?
+ UINT16 *mv_penalty= s->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;
+ int src_x, src_y;
+ int fbmin;
+
+ fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->qscale;
+
+ dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
+ src_x = mb_x * 16 + (motion_fx >> 1);
+ src_y = mb_y * 16 + (motion_fy >> 1);
+
+ ptr = s->last_picture[0] + (src_y * s->linesize) + src_x;
+ put_pixels_tab[dxy](dest_y , ptr , s->linesize, 16);
+ put_pixels_tab[dxy](dest_y + 8, ptr + 8, s->linesize, 16);
+
+ fbmin += (mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->qscale;
+
+ dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
+ src_x = mb_x * 16 + (motion_bx >> 1);
+ src_y = mb_y * 16 + (motion_by >> 1);
+
+ ptr = s->next_picture[0] + (src_y * s->linesize) + src_x;
+ avg_pixels_tab[dxy](dest_y , ptr , s->linesize, 16);
+ avg_pixels_tab[dxy](dest_y + 8, ptr + 8, s->linesize, 16);
+
+ fbmin += pix_abs16x16(s->new_picture[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
+ return fbmin;
+}
+/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
+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;
+ 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 pred_bx= s->b_bidir_back_mv_table[xy-1][0];
+ int pred_by= s->b_bidir_back_mv_table[xy-1][1];
+ int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
+ int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
+ int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
+ int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
+
+ //FIXME do refinement and add flag
+
+ fbmin= check_bidir_mv(s, mb_x, mb_y,
+ motion_fx, motion_fy,
+ motion_bx, motion_by,
+ pred_fx, pred_fy,
+ pred_bx, pred_by);
+
+ return fbmin;
+}
+
+static inline int direct_search(MpegEncContext * s,
+ int mb_x, int mb_y)
+{
+ int P[6][2];
+ const int mot_stride = s->mb_width + 2;
+ const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
+ int dmin, dmin2;
+ int motion_fx, motion_fy, motion_bx, motion_by, motion_bx0, motion_by0;
+ int motion_dx, motion_dy;
+ const int motion_px= s->p_mv_table[mot_xy][0];
+ const int motion_py= s->p_mv_table[mot_xy][1];
+ const int time_pp= s->pp_time;
+ const int time_bp= s->bp_time;
+ const int time_pb= time_pp - time_bp;
+ int bx, by;
+ int mx, my, mx2, my2;
+ uint8_t *ref_picture= s->me_scratchpad - (mb_x + 1 + (mb_y + 1)*s->linesize)*16;
+ int16_t (*mv_table)[2]= s->b_direct_mv_table;
+ uint16_t *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
+
+ /* thanks to iso-mpeg the rounding is different for the zero vector, so we need to handle that ... */
+ motion_fx= (motion_px*time_pb)/time_pp;
+ motion_fy= (motion_py*time_pb)/time_pp;
+ motion_bx0= (-motion_px*time_bp)/time_pp;
+ motion_by0= (-motion_py*time_bp)/time_pp;
+ motion_dx= motion_dy=0;
+ dmin2= check_bidir_mv(s, mb_x, mb_y,
+ motion_fx, motion_fy,
+ motion_bx0, motion_by0,
+ motion_fx, motion_fy,
+ motion_bx0, motion_by0) - s->qscale;
+
+ motion_bx= motion_fx - motion_px;
+ motion_by= motion_fy - motion_py;
+ for(by=-1; by<2; by++){
+ for(bx=-1; bx<2; bx++){
+ uint8_t *dest_y = s->me_scratchpad + (by+1)*s->linesize*16 + (bx+1)*16;
+ uint8_t *ptr;
+ int dxy;
+ int src_x, src_y;
+ const int width= s->width;
+ const int height= s->height;
+
+ dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
+ src_x = (mb_x + bx) * 16 + (motion_fx >> 1);
+ src_y = (mb_y + by) * 16 + (motion_fy >> 1);
+ src_x = clip(src_x, -16, width);
+ if (src_x == width) dxy &= ~1;
+ src_y = clip(src_y, -16, height);
+ if (src_y == height) dxy &= ~2;
+
+ ptr = s->last_picture[0] + (src_y * s->linesize) + src_x;
+ put_pixels_tab[dxy](dest_y , ptr , s->linesize, 16);
+ put_pixels_tab[dxy](dest_y + 8, ptr + 8, s->linesize, 16);
+
+ dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
+ src_x = (mb_x + bx) * 16 + (motion_bx >> 1);
+ src_y = (mb_y + by) * 16 + (motion_by >> 1);
+ src_x = clip(src_x, -16, width);
+ if (src_x == width) dxy &= ~1;
+ src_y = clip(src_y, -16, height);
+ if (src_y == height) dxy &= ~2;
+
+ avg_pixels_tab[dxy](dest_y , ptr , s->linesize, 16);
+ avg_pixels_tab[dxy](dest_y + 8, ptr + 8, s->linesize, 16);
+ }
+ }
+
+ P[0][0] = mv_table[mot_xy ][0];
+ P[0][1] = mv_table[mot_xy ][1];
+ P[1][0] = mv_table[mot_xy - 1][0];
+ P[1][1] = mv_table[mot_xy - 1][1];
+
+ /* special case for first line */
+ if ((mb_y == 0 || s->first_slice_line || s->first_gob_line)) {
+ P[4][0] = P[1][0];
+ P[4][1] = P[1][1];
+ } else {
+ P[2][0] = mv_table[mot_xy - mot_stride ][0];
+ P[2][1] = mv_table[mot_xy - mot_stride ][1];
+ P[3][0] = mv_table[mot_xy - mot_stride + 1 ][0];
+ P[3][1] = mv_table[mot_xy - mot_stride + 1 ][1];
+
+ P[4][0]= mid_pred(P[1][0], P[2][0], P[3][0]);
+ P[4][1]= mid_pred(P[1][1], P[2][1], P[3][1]);
+ }
+ dmin = epzs_motion_search(s, &mx, &my, P, 0, 0, -16, -16, 15, 15, ref_picture);
+ if(mx==0 && my==0) dmin=99999999; // not representable, due to rounding stuff
+ if(dmin2<dmin){
+ dmin= dmin2;
+ mx=0;
+ my=0;
+ }
+#if 1
+ mx2= mx= mx*2;
+ my2= my= my*2;
+ for(by=-1; by<2; by++){
+ if(my2+by < -32) continue;
+ for(bx=-1; bx<2; bx++){
+ if(bx==0 && by==0) continue;
+ if(mx2+bx < -32) continue;
+ dmin2= check_bidir_mv(s, mb_x, mb_y,
+ mx2+bx+motion_fx, my2+by+motion_fy,
+ mx2+bx+motion_bx, my2+by+motion_by,
+ mx2+bx+motion_fx, my2+by+motion_fy,
+ motion_bx, motion_by) - s->qscale;
+
+ if(dmin2<dmin){
+ dmin=dmin2;
+ mx= mx2 + bx;
+ my= my2 + by;
+ }
+ }
+ }
+#else
+ mx*=2; my*=2;
+#endif
+ if(mx==0 && my==0){
+ motion_bx= motion_bx0;
+ motion_by= motion_by0;
+ }
+
+ s->b_direct_mv_table[mot_xy][0]= mx;
+ s->b_direct_mv_table[mot_xy][1]= my;
+ s->b_direct_forw_mv_table[mot_xy][0]= motion_fx + mx;
+ s->b_direct_forw_mv_table[mot_xy][1]= motion_fy + my;
+ s->b_direct_back_mv_table[mot_xy][0]= motion_bx + mx;
+ s->b_direct_back_mv_table[mot_xy][1]= motion_by + my;
+ return dmin;
}
void ff_estimate_b_frame_motion(MpegEncContext * s,
@@ -1143,16 +1334,41 @@ void ff_estimate_b_frame_motion(MpegEncContext * s,
{
const int mot_stride = s->mb_width + 2;
const int xy = (mb_y + 1)*mot_stride + mb_x + 1;
+ const int quant= s->qscale;
+ int fmin, bmin, dmin, fbmin;
+ int type=0;
+ int motion_fx, motion_fy, motion_bx, motion_by;
+
+ dmin= direct_search(s, mb_x, mb_y);
- ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, s->last_picture[0], s->f_code);
- ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, s->next_picture[0], s->b_code);
+ fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, s->last_picture[0], s->f_code);
+ bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, s->next_picture[0], s->b_code) - quant;
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
- s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
- s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
- s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
- s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
-
- s->mb_type[mb_y*s->mb_width + mb_x]= MB_TYPE_FORWARD; //FIXME
+
+ fbmin= bidir_refine(s, mb_x, mb_y);
+
+ if(s->flags&CODEC_FLAG_HQ){
+ type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT;
+ }else{
+ int score= dmin;
+ type=MB_TYPE_DIRECT;
+
+ if(fmin<score){
+ score=fmin;
+ type= MB_TYPE_FORWARD;
+ }
+ if(bmin<score){
+ score=bmin;
+ type= MB_TYPE_BACKWARD;
+ }
+ if(fbmin<score){
+ score=fbmin;
+ type= MB_TYPE_BIDIR;
+ }
+ s->mc_mb_var += score;
+ }
+
+ s->mb_type[mb_y*s->mb_width + mb_x]= type;
}
/* find best f_code for ME which do unlimited searches */
@@ -1184,8 +1400,12 @@ int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
}
for(i=MAX_FCODE; i>1; i--){
+ int threshold;
loose+= mv_num[i];
- if(loose > s->mb_num/20) break; //FIXME this is pretty ineffective
+
+ if(s->pict_type==B_TYPE) threshold= 0;
+ else threshold= s->mb_num/20; //FIXME
+ if(loose > threshold) break;
}
// printf("fcode: %d type: %d\n", i, s->pict_type);
return i;
@@ -1275,11 +1495,12 @@ void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, i
|| fcode_tab[mv_table[xy][0] + MAX_MV] == 0
|| fcode_tab[mv_table[xy][1] + MAX_MV] > f_code
|| fcode_tab[mv_table[xy][1] + MAX_MV] == 0 ){
- s->mb_type[i] &= ~type;
- if(s->mb_type[i]==0) s->mb_type[i]= MB_TYPE_FORWARD; //FIXME
- mv_table[xy][0] = 0;
- mv_table[xy][1] = 0;
- //this is certainly bad FIXME
+ if(s->mb_type[i]&(~type)) s->mb_type[i] &= ~type;
+ else{
+ mv_table[xy][0] = 0;
+ mv_table[xy][1] = 0;
+ //this is certainly bad FIXME
+ }
}
}
xy++;
diff --git a/libavcodec/mpegvideo.c b/libavcodec/mpegvideo.c
index af1d1c3396..d1c7e16372 100644
--- a/libavcodec/mpegvideo.c
+++ b/libavcodec/mpegvideo.c
@@ -225,6 +225,12 @@ int MPV_common_init(MpegEncContext *s)
goto fail;
}
+ s->me_scratchpad = av_mallocz( s->linesize*16*3*sizeof(uint8_t));
+ if (s->me_scratchpad == NULL) {
+ perror("malloc");
+ goto fail;
+ }
+
if(s->max_b_frames){
for(j=0; j<REORDER_BUFFER_SIZE; j++){
int i;
@@ -297,7 +303,7 @@ int MPV_common_init(MpegEncContext *s)
if (!s->mbskip_table)
goto fail;
- s->block= s->intra_block;
+ s->block= s->blocks[0];
s->context_initialized = 1;
return 0;
@@ -333,6 +339,7 @@ void MPV_common_end(MpegEncContext *s)
CHECK_FREE(s->ac_val[0]);
CHECK_FREE(s->coded_block);
CHECK_FREE(s->mbintra_table);
+ CHECK_FREE(s->me_scratchpad);
CHECK_FREE(s->mbskip_table);
for(i=0;i<3;i++) {
@@ -761,16 +768,6 @@ int MPV_encode_picture(AVCodecContext *avctx,
return pbBufPtr(&s->pb) - s->pb.buf;
}
-static inline int clip(int a, int amin, int amax)
-{
- if (a < amin)
- return amin;
- else if (a > amax)
- return amax;
- else
- return a;
-}
-
static inline void gmc1_motion(MpegEncContext *s,
UINT8 *dest_y, UINT8 *dest_cb, UINT8 *dest_cr,
int dest_offset,
@@ -1225,7 +1222,7 @@ void MPV_decode_mb(MpegEncContext *s, DCTELEM block[6][64])
if (!s->mb_intra) {
/* motion handling */
if((s->flags&CODEC_FLAG_HQ) || (!s->encoding)){
- if (!s->no_rounding){
+ if ((!s->no_rounding) || s->pict_type==B_TYPE){
op_pix = put_pixels_tab;
op_qpix= qpel_mc_rnd_tab;
}else{
@@ -1235,7 +1232,7 @@ void MPV_decode_mb(MpegEncContext *s, DCTELEM block[6][64])
if (s->mv_dir & MV_DIR_FORWARD) {
MPV_motion(s, dest_y, dest_cb, dest_cr, 0, s->last_picture, op_pix, op_qpix);
- if (!s->no_rounding)
+ if ((!s->no_rounding) || s->pict_type==B_TYPE)
op_pix = avg_pixels_tab;
else
op_pix = avg_no_rnd_pixels_tab;
@@ -1312,7 +1309,7 @@ static void encode_mb(MpegEncContext *s, int motion_x, int motion_y)
dest_cb = s->current_picture[1] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8;
dest_cr = s->current_picture[2] + (mb_y * 8 * (s->linesize >> 1)) + mb_x * 8;
- if (!s->no_rounding){
+ if ((!s->no_rounding) || s->pict_type==B_TYPE){
op_pix = put_pixels_tab;
op_qpix= qpel_mc_rnd_tab;
}else{
@@ -1322,7 +1319,7 @@ static void encode_mb(MpegEncContext *s, int motion_x, int motion_y)
if (s->mv_dir & MV_DIR_FORWARD) {
MPV_motion(s, dest_y, dest_cb, dest_cr, 0, s->last_picture, op_pix, op_qpix);
- if (!s->no_rounding)
+ if ((!s->no_rounding) || s->pict_type==B_TYPE)
op_pix = avg_pixels_tab;
else
op_pix = avg_no_rnd_pixels_tab;
@@ -1429,6 +1426,8 @@ static void copy_context_before_encode(MpegEncContext *d, MpegEncContext *s, int
d->skip_count= s->skip_count;
d->misc_bits= s->misc_bits;
d->last_bits= s->last_bits;
+
+ d->mb_skiped= s->mb_skiped;
}
static void copy_context_after_encode(MpegEncContext *d, MpegEncContext *s, int type){
@@ -1453,6 +1452,7 @@ static void copy_context_after_encode(MpegEncContext *d, MpegEncContext *s, int
d->last_bits= s->last_bits;
d->mb_intra= s->mb_intra;
+ d->mb_skiped= s->mb_skiped;
d->mv_type= s->mv_type;
d->mv_dir= s->mv_dir;
d->pb= s->pb;
@@ -1468,7 +1468,7 @@ static void encode_picture(MpegEncContext *s, int picture_number)
int i;
int bits;
MpegEncContext best_s, backup_s;
- UINT8 bit_buf[4][3000]; //FIXME check that this is ALLWAYS large enogh for a MB
+ UINT8 bit_buf[7][3000]; //FIXME check that this is ALLWAYS large enogh for a MB
s->picture_number = picture_number;
@@ -1483,7 +1483,11 @@ static void encode_picture(MpegEncContext *s, int picture_number)
/* Reset the average MB variance */
s->avg_mb_var = 0;
s->mc_mb_var = 0;
-
+
+ /* we need to initialize some time vars before we can encode b-frames */
+ if (s->h263_pred && !s->h263_msmpeg4)
+ ff_set_mpeg4_time(s, s->picture_number);
+
/* Estimate motion for every MB */
if(s->pict_type != I_TYPE){
// int16_t (*tmp)[2]= s->p_mv_table;
@@ -1535,9 +1539,11 @@ static void encode_picture(MpegEncContext *s, int picture_number)
if(s->pict_type==B_TYPE){
s->f_code= ff_get_best_fcode(s, s->b_forw_mv_table, MB_TYPE_FORWARD);
s->b_code= ff_get_best_fcode(s, s->b_back_mv_table, MB_TYPE_BACKWARD);
- //FIXME if BIDIR != for&back
- ff_fix_long_b_mvs(s, s->b_forw_mv_table, s->f_code, MB_TYPE_FORWARD |MB_TYPE_BIDIR);
- ff_fix_long_b_mvs(s, s->b_back_mv_table, s->b_code, MB_TYPE_BACKWARD|MB_TYPE_BIDIR);
+
+ ff_fix_long_b_mvs(s, s->b_forw_mv_table, s->f_code, MB_TYPE_FORWARD);
+ ff_fix_long_b_mvs(s, s->b_back_mv_table, s->b_code, MB_TYPE_BACKWARD);
+ ff_fix_long_b_mvs(s, s->b_bidir_forw_mv_table, s->f_code, MB_TYPE_BIDIR);
+ ff_fix_long_b_mvs(s, s->b_bidir_back_mv_table, s->b_code, MB_TYPE_BIDIR);
}
//printf("f_code %d ///\n", s->f_code);
@@ -1632,7 +1638,8 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->block_index[4]= s->block_wrap[4]*(mb_y + 1) + s->block_wrap[0]*(s->mb_height*2 + 2);
s->block_index[5]= s->block_wrap[4]*(mb_y + 1 + s->mb_height + 2) + s->block_wrap[0]*(s->mb_height*2 + 2);
for(mb_x=0; mb_x < s->mb_width; mb_x++) {
- /*const */int mb_type= s->mb_type[mb_y * s->mb_width + mb_x];
+ const int mb_type= s->mb_type[mb_y * s->mb_width + mb_x];
+ const int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
PutBitContext pb;
int d;
int dmin=10000000;
@@ -1647,19 +1654,19 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->block_index[4]++;
s->block_index[5]++;
if(mb_type & (mb_type-1)){ // more than 1 MB type possible
+ int next_block=0;
pb= s->pb;
- s->mv_dir = MV_DIR_FORWARD;
copy_context_before_encode(&backup_s, s, -1);
if(mb_type&MB_TYPE_INTER){
- int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
+ s->mv_dir = MV_DIR_FORWARD;
s->mv_type = MV_TYPE_16X16;
s->mb_intra= 0;
s->mv[0][0][0] = s->p_mv_table[xy][0];
s->mv[0][0][1] = s->p_mv_table[xy][1];
init_put_bits(&s->pb, bit_buf[1], 3000, NULL, NULL);
- s->block= s->inter_block;
+ s->block= s->blocks[next_block];
encode_mb(s, s->mv[0][0][0], s->mv[0][0][1]);
d= get_bit_count(&s->pb);
@@ -1668,10 +1675,12 @@ static void encode_picture(MpegEncContext *s, int picture_number)
dmin=d;
copy_context_after_encode(&best_s, s, MB_TYPE_INTER);
best=1;
+ next_block^=1;
}
}
if(mb_type&MB_TYPE_INTER4V){
copy_context_before_encode(s, &backup_s, MB_TYPE_INTER4V);
+ s->mv_dir = MV_DIR_FORWARD;
s->mv_type = MV_TYPE_8X8;
s->mb_intra= 0;
for(i=0; i<4; i++){
@@ -1679,25 +1688,111 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->mv[0][i][1] = s->motion_val[s->block_index[i]][1];
}
init_put_bits(&s->pb, bit_buf[2], 3000, NULL, NULL);
- s->block= s->inter4v_block;
+ s->block= s->blocks[next_block];
encode_mb(s, 0, 0);
d= get_bit_count(&s->pb);
- if(d<dmin && 0){
+ if(d<dmin){
flush_put_bits(&s->pb);
dmin=d;
copy_context_after_encode(&best_s, s, MB_TYPE_INTER4V);
best=2;
+ next_block^=1;
+ }
+ }
+ if(mb_type&MB_TYPE_FORWARD){
+ copy_context_before_encode(s, &backup_s, MB_TYPE_FORWARD);
+ s->mv_dir = MV_DIR_FORWARD;
+ s->mv_type = MV_TYPE_16X16;
+ s->mb_intra= 0;
+ s->mv[0][0][0] = s->b_forw_mv_table[xy][0];
+ s->mv[0][0][1] = s->b_forw_mv_table[xy][1];
+ init_put_bits(&s->pb, bit_buf[3], 3000, NULL, NULL);
+ s->block= s->blocks[next_block];
+
+ encode_mb(s, s->mv[0][0][0], s->mv[0][0][1]);
+ d= get_bit_count(&s->pb);
+ if(d<dmin){
+ flush_put_bits(&s->pb);
+ dmin=d;
+ copy_context_after_encode(&best_s, s, MB_TYPE_FORWARD);
+ best=3;
+ next_block^=1;
+ }
+ }
+ if(mb_type&MB_TYPE_BACKWARD){
+ copy_context_before_encode(s, &backup_s, MB_TYPE_BACKWARD);
+ s->mv_dir = MV_DIR_BACKWARD;
+ s->mv_type = MV_TYPE_16X16;
+ s->mb_intra= 0;
+ s->mv[1][0][0] = s->b_back_mv_table[xy][0];
+ s->mv[1][0][1] = s->b_back_mv_table[xy][1];
+ init_put_bits(&s->pb, bit_buf[4], 3000, NULL, NULL);
+ s->block= s->blocks[next_block];
+
+ encode_mb(s, s->mv[1][0][0], s->mv[1][0][1]);
+ d= get_bit_count(&s->pb);
+ if(d<dmin){
+ flush_put_bits(&s->pb);
+ dmin=d;
+ copy_context_after_encode(&best_s, s, MB_TYPE_BACKWARD);
+ best=4;
+ next_block^=1;
+ }
+ }
+ if(mb_type&MB_TYPE_BIDIR){
+ copy_context_before_encode(s, &backup_s, MB_TYPE_BIDIR);
+ s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
+ s->mv_type = MV_TYPE_16X16;
+ s->mb_intra= 0;
+ s->mv[0][0][0] = s->b_bidir_forw_mv_table[xy][0];
+ s->mv[0][0][1] = s->b_bidir_forw_mv_table[xy][1];
+ s->mv[1][0][0] = s->b_bidir_back_mv_table[xy][0];
+ s->mv[1][0][1] = s->b_bidir_back_mv_table[xy][1];
+ init_put_bits(&s->pb, bit_buf[5], 3000, NULL, NULL);
+ s->block= s->blocks[next_block];
+
+ encode_mb(s, 0, 0);
+ d= get_bit_count(&s->pb);
+ if(d<dmin){
+ flush_put_bits(&s->pb);
+ dmin=d;
+ copy_context_after_encode(&best_s, s, MB_TYPE_BIDIR);
+ best=5;
+ next_block^=1;
+ }
+ }
+ if(mb_type&MB_TYPE_DIRECT){
+ copy_context_before_encode(s, &backup_s, MB_TYPE_DIRECT);
+ s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
+ s->mv_type = MV_TYPE_16X16; //FIXME
+ s->mb_intra= 0;
+ s->mv[0][0][0] = s->b_direct_forw_mv_table[xy][0];
+ s->mv[0][0][1] = s->b_direct_forw_mv_table[xy][1];
+ s->mv[1][0][0] = s->b_direct_back_mv_table[xy][0];
+ s->mv[1][0][1] = s->b_direct_back_mv_table[xy][1];
+ init_put_bits(&s->pb, bit_buf[6], 3000, NULL, NULL);
+ s->block= s->blocks[next_block];
+
+ encode_mb(s, s->b_direct_mv_table[xy][0], s->b_direct_mv_table[xy][1]);
+ d= get_bit_count(&s->pb);
+ if(d<dmin){
+ flush_put_bits(&s->pb);
+ dmin=d;
+ copy_context_after_encode(&best_s, s, MB_TYPE_DIRECT);
+ best=6;
+ next_block^=1;
}
}
if(mb_type&MB_TYPE_INTRA){
copy_context_before_encode(s, &backup_s, MB_TYPE_INTRA);
+ s->mv_dir = MV_DIR_FORWARD;
s->mv_type = MV_TYPE_16X16;
s->mb_intra= 1;
s->mv[0][0][0] = 0;
s->mv[0][0][1] = 0;
init_put_bits(&s->pb, bit_buf[0], 3000, NULL, NULL);
- s->block= s->intra_block;
+ s->block= s->blocks[next_block];
encode_mb(s, 0, 0);
d= get_bit_count(&s->pb);
@@ -1706,6 +1801,7 @@ static void encode_picture(MpegEncContext *s, int picture_number)
dmin=d;
copy_context_after_encode(&best_s, s, MB_TYPE_INTRA);
best=0;
+ next_block^=1;
}
/* force cleaning of ac/dc pred stuff if needed ... */
if(s->h263_pred || s->h263_aic)
@@ -1718,30 +1814,30 @@ static void encode_picture(MpegEncContext *s, int picture_number)
int motion_x, motion_y;
s->mv_type=MV_TYPE_16X16;
// only one MB-Type possible
- //FIXME convert to swicth()
- if(mb_type&MB_TYPE_INTRA){
+ switch(mb_type){
+ case MB_TYPE_INTRA:
s->mv_dir = MV_DIR_FORWARD;
s->mb_intra= 1;
motion_x= s->mv[0][0][0] = 0;
motion_y= s->mv[0][0][1] = 0;
- }else if(mb_type&MB_TYPE_INTER){
- int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
+ break;
+ case MB_TYPE_INTER:
s->mv_dir = MV_DIR_FORWARD;
s->mb_intra= 0;
motion_x= s->mv[0][0][0] = s->p_mv_table[xy][0];
motion_y= s->mv[0][0][1] = s->p_mv_table[xy][1];
- }else if(mb_type&MB_TYPE_DIRECT){
- int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
+ break;
+ case MB_TYPE_DIRECT:
s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD | MV_DIRECT;
s->mb_intra= 0;
- motion_x=0;
- motion_y=0;
- s->mv[0][0][0] = 0;
- s->mv[0][0][1] = 0;
- s->mv[1][0][0] = 0;
- s->mv[1][0][1] = 0;
- }else if(mb_type&MB_TYPE_BIDIR){
- int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
+ motion_x=s->b_direct_mv_table[xy][0];
+ motion_y=s->b_direct_mv_table[xy][1];
+ s->mv[0][0][0] = s->b_direct_forw_mv_table[xy][0];
+ s->mv[0][0][1] = s->b_direct_forw_mv_table[xy][1];
+ s->mv[1][0][0] = s->b_direct_back_mv_table[xy][0];
+ s->mv[1][0][1] = s->b_direct_back_mv_table[xy][1];
+ break;
+ case MB_TYPE_BIDIR:
s->mv_dir = MV_DIR_FORWARD | MV_DIR_BACKWARD;
s->mb_intra= 0;
motion_x=0;
@@ -1750,25 +1846,31 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->mv[0][0][1] = s->b_bidir_forw_mv_table[xy][1];
s->mv[1][0][0] = s->b_bidir_back_mv_table[xy][0];
s->mv[1][0][1] = s->b_bidir_back_mv_table[xy][1];
- }else if(mb_type&MB_TYPE_BACKWARD){
- int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
+ break;
+ case MB_TYPE_BACKWARD:
s->mv_dir = MV_DIR_BACKWARD;
s->mb_intra= 0;
motion_x= s->mv[1][0][0] = s->b_back_mv_table[xy][0];
motion_y= s->mv[1][0][1] = s->b_back_mv_table[xy][1];
- }else if(mb_type&MB_TYPE_FORWARD){
- int xy= (mb_y+1) * (s->mb_width+2) + mb_x + 1;
+ break;
+ case MB_TYPE_FORWARD:
s->mv_dir = MV_DIR_FORWARD;
s->mb_intra= 0;
motion_x= s->mv[0][0][0] = s->b_forw_mv_table[xy][0];
motion_y= s->mv[0][0][1] = s->b_forw_mv_table[xy][1];
// printf(" %d %d ", motion_x, motion_y);
- }else{
+ break;
+ default:
motion_x=motion_y=0; //gcc warning fix
printf("illegal MB type\n");
}
encode_mb(s, motion_x, motion_y);
}
+ /* clean the MV table in IPS frames for direct mode in B frames */
+ if(s->mb_intra /* && I,P,S_TYPE */){
+ s->p_mv_table[xy][0]=0;
+ s->p_mv_table[xy][1]=0;
+ }
MPV_decode_mb(s, s->block);
}
diff --git a/libavcodec/mpegvideo.h b/libavcodec/mpegvideo.h
index 841d40fb3c..24943e5b19 100644
--- a/libavcodec/mpegvideo.h
+++ b/libavcodec/mpegvideo.h
@@ -141,7 +141,8 @@ typedef struct MpegEncContext {
INT16 (*b_direct_forw_mv_table)[2];/* MV table (1MV per MB) direct mode b-frame encoding */
INT16 (*b_direct_back_mv_table)[2];/* MV table (1MV per MB) direct mode b-frame encoding */
INT16 (*b_direct_mv_table)[2]; /* MV table (1MV per MB) direct mode b-frame encoding */
- int me_method; /* ME algorithm */
+ int me_method; /* ME algorithm */
+ uint8_t *me_scratchpad; /* data area for the me algo, so that the ME doesnt need to malloc/free */
int mv_dir;
#define MV_DIR_BACKWARD 1
#define MV_DIR_FORWARD 2
@@ -164,7 +165,8 @@ typedef struct MpegEncContext {
UINT8 *fcode_tab; /* smallest fcode needed for each MV */
int has_b_frames;
- int no_rounding; /* apply no rounding to motion compensation (MPEG4, msmpeg4, ...) */
+ int no_rounding; /* apply no rounding to motion compensation (MPEG4, msmpeg4, ...)
+ for b-frames rounding mode is allways 0 */
/* macroblock layer */
int mb_x, mb_y;
@@ -335,9 +337,7 @@ typedef struct MpegEncContext {
UINT32 mb_line_avgsize;
DCTELEM (*block)[64]; /* points to one of the following blocks */
- DCTELEM intra_block[6][64] __align8;
- DCTELEM inter_block[6][64] __align8;
- DCTELEM inter4v_block[6][64] __align8;
+ DCTELEM blocks[2][6][64] __align8; // for HQ mode we need to keep the best block
void (*dct_unquantize_mpeg1)(struct MpegEncContext *s,
DCTELEM *block, int n, int qscale);
void (*dct_unquantize_mpeg2)(struct MpegEncContext *s,
@@ -421,6 +421,7 @@ INT16 *h263_pred_motion(MpegEncContext * s, int block,
int *px, int *py);
void mpeg4_pred_ac(MpegEncContext * s, INT16 *block, int n,
int dir);
+void ff_set_mpeg4_time(MpegEncContext * s, int picture_number);
void mpeg4_encode_picture_header(MpegEncContext *s, int picture_number);
void h263_encode_init(MpegEncContext *s);