1 |
/************************************************************************** |
/***************************************************************************** |
2 |
* |
* |
3 |
* XVID MPEG-4 VIDEO CODEC |
* XVID MPEG-4 VIDEO CODEC |
4 |
* motion estimation |
* - Motion Estimation module - |
5 |
|
* |
6 |
|
* Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
7 |
|
* Copyright(C) 2002 Michael Militzer <michael@xvid.org> |
8 |
|
* Copyright(C) 2002 Edouard Gomez <ed.gomez@wanadoo.fr> |
9 |
* |
* |
10 |
* This program is an implementation of a part of one or more MPEG-4 |
* This program is an implementation of a part of one or more MPEG-4 |
11 |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
28 |
* |
* |
29 |
* You should have received a copy of the GNU General Public License |
* You should have received a copy of the GNU General Public License |
30 |
* along with this program; if not, write to the Free Software |
* along with this program; if not, write to the Free Software |
31 |
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
32 |
* |
* |
33 |
*************************************************************************/ |
*************************************************************************/ |
34 |
|
|
|
/************************************************************************** |
|
|
* |
|
|
* Modifications: |
|
|
* |
|
|
* 01.05.2002 updated MotionEstimationBVOP |
|
|
* 25.04.2002 partial prevMB conversion |
|
|
* 22.04.2002 remove some compile warning by chenm001 <chenm001@163.com> |
|
|
* 14.04.2002 added MotionEstimationBVOP() |
|
|
* 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between |
|
|
* EPZS and EPZS^2 |
|
|
* 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop |
|
|
* PMVFast_Refine to support multiple searches with different start points |
|
|
* 07.01.2002 uv-block-based interpolation |
|
|
* 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) |
|
|
* changed INTER_BIAS to 150 (as suggested by suxen_drol) |
|
|
* removed halfpel refinement step in PMVfastSearch8 + quality=5 |
|
|
* added new quality mode = 6 which performs halfpel refinement |
|
|
* filesize difference between quality 5 and 6 is smaller than 1% |
|
|
* (Isibaar) |
|
|
* 31.12.2001 PMVfastSearch16 and PMVfastSearch8 (gruel) |
|
|
* 30.12.2001 get_range/MotionSearchX simplified; blue/green bug fix |
|
|
* 22.12.2001 commented best_point==99 check |
|
|
* 19.12.2001 modified get_range (purple bug fix) |
|
|
* 15.12.2001 moved pmv displacement from mbprediction |
|
|
* 02.12.2001 motion estimation/compensation split (Isibaar) |
|
|
* 16.11.2001 rewrote/tweaked search algorithms; pross@cs.rmit.edu.au |
|
|
* 10.11.2001 support for sad16/sad8 functions |
|
|
* 28.08.2001 reactivated MODE_INTER4V for EXT_MODE |
|
|
* 24.08.2001 removed MODE_INTER4V_Q, disabled MODE_INTER4V for EXT_MODE |
|
|
* 22.08.2001 added MODE_INTER4V_Q |
|
|
* 20.08.2001 added pragma to get rid of internal compiler error with VC6 |
|
|
* idea by Cyril. Thanks. |
|
|
* |
|
|
* Michael Militzer <isibaar@videocoding.de> |
|
|
* |
|
|
**************************************************************************/ |
|
|
|
|
35 |
#include <assert.h> |
#include <assert.h> |
36 |
#include <stdio.h> |
#include <stdio.h> |
37 |
#include <stdlib.h> |
#include <stdlib.h> |
153 |
|
|
154 |
MACROBLOCK *const pMB = &pMBs[x + y * iWcount]; |
MACROBLOCK *const pMB = &pMBs[x + y * iWcount]; |
155 |
|
|
156 |
|
if (pMB->mode == MODE_NOT_CODED) |
157 |
|
continue; |
158 |
|
|
159 |
predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
160 |
|
|
161 |
pMB->sad16 = |
pMB->sad16 = |
2858 |
*currMV = pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV; |
*currMV = pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV; |
2859 |
|
|
2860 |
} else { |
} else { |
2861 |
|
|
2862 |
|
bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
2863 |
|
|
2864 |
threshA = psad[0]; |
threshA = psad[0]; |
2865 |
threshB = threshA + 256; |
threshB = threshA + 256; |
2866 |
if (threshA < 512) |
if (threshA < 512) |
2870 |
if (threshB > 1792) |
if (threshB > 1792) |
2871 |
threshB = 1792; |
threshB = 1792; |
2872 |
|
|
|
bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); |
|
2873 |
*currMV = pmv[0]; /* current best := prediction */ |
*currMV = pmv[0]; /* current best := prediction */ |
2874 |
} |
} |
2875 |
|
|
3081 |
return iMinSAD; |
return iMinSAD; |
3082 |
} |
} |
3083 |
|
|
|
|
|
|
|
|
|
/* *********************************************************** |
|
|
bvop motion estimation |
|
|
***************************************************************/ |
|
|
|
|
|
void |
|
|
MotionEstimationBVOP(MBParam * const pParam, |
|
|
FRAMEINFO * const frame, |
|
|
const int32_t time_bp, |
|
|
const int32_t time_pp, |
|
|
// forward (past) reference |
|
|
const MACROBLOCK * const f_mbs, |
|
|
const IMAGE * const f_ref, |
|
|
const IMAGE * const f_refH, |
|
|
const IMAGE * const f_refV, |
|
|
const IMAGE * const f_refHV, |
|
|
// backward (future) reference |
|
|
const MACROBLOCK * const b_mbs, |
|
|
const IMAGE * const b_ref, |
|
|
const IMAGE * const b_refH, |
|
|
const IMAGE * const b_refV, |
|
|
const IMAGE * const b_refHV) |
|
|
{ |
|
|
const int mb_width = pParam->mb_width; |
|
|
const int mb_height = pParam->mb_height; |
|
|
const int edged_width = pParam->edged_width; |
|
|
|
|
|
const int32_t iWidth = pParam->width; |
|
|
const int32_t iHeight = pParam->height; |
|
|
|
|
|
int i, j, k; |
|
|
|
|
|
static const VECTOR zeroMV={0,0}; |
|
|
|
|
|
int f_sad16; /* forward (as usual) search */ |
|
|
int b_sad16; /* backward (only in b-frames) search */ |
|
|
int i_sad16; /* interpolated (both direction, b-frames only) */ |
|
|
int d_sad16; /* direct mode (assume almost linear motion) */ |
|
|
|
|
|
int best_sad; |
|
|
|
|
|
VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ |
|
|
VECTOR f_interpolMV, b_interpolMV; |
|
|
VECTOR pmv_dontcare; |
|
|
|
|
|
int min_dx, max_dx, min_dy, max_dy; |
|
|
int f_min_dx, f_max_dx, f_min_dy, f_max_dy; |
|
|
int b_min_dx, b_max_dx, b_min_dy, b_max_dy; |
|
|
|
|
|
int f_count=0; |
|
|
int b_count=0; |
|
|
int i_count=0; |
|
|
int d_count=0; |
|
|
|
|
|
const int64_t TRB = (int32_t)time_pp - (int32_t)time_bp; |
|
|
const int64_t TRD = (int32_t)time_pp; |
|
|
|
|
|
// fprintf(stderr,"TRB = %lld TRD = %lld time_bp =%d time_pp =%d\n\n",TRB,TRD,time_bp,time_pp); |
|
|
// note: i==horizontal, j==vertical |
|
|
for (j = 0; j < mb_height; j++) { |
|
|
|
|
|
f_predMV = zeroMV; /* prediction is reset at left boundary */ |
|
|
b_predMV = zeroMV; |
|
|
|
|
|
for (i = 0; i < mb_width; i++) { |
|
|
MACROBLOCK *mb = &frame->mbs[i + j * mb_width]; |
|
|
const MACROBLOCK *f_mb = &f_mbs[i + j * mb_width]; |
|
|
const MACROBLOCK *b_mb = &b_mbs[i + j * mb_width]; |
|
|
|
|
|
mb->deltamv=zeroMV; |
|
|
|
|
|
/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ |
|
|
|
|
|
if (b_mb->mode == MODE_INTER && b_mb->cbp == 0 && |
|
|
b_mb->mvs[0].x == 0 && b_mb->mvs[0].y == 0) { |
|
|
mb->mode = MODE_NOT_CODED; |
|
|
mb->b_mvs[0] = mb->mvs[0] = zeroMV; |
|
|
continue; |
|
|
} |
|
|
|
|
|
if (b_mb->mode == MODE_INTER4V) |
|
|
{ |
|
|
d_sad16 = 0; |
|
|
/* same method of scaling as in decoder.c, so we copy from there */ |
|
|
for (k = 0; k < 4; k++) { |
|
|
|
|
|
mb->directmv[k] = b_mb->mvs[k]; |
|
|
|
|
|
mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); |
|
|
mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].x) / TRD |
|
|
: mb->mvs[k].x - mb->directmv[k].x); |
|
|
|
|
|
mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); |
|
|
mb->b_mvs[k].y = (int32_t) ((mb->deltamv.y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].y) / TRD |
|
|
: mb->mvs[k].y - mb->directmv[k].y); |
|
|
|
|
|
d_sad16 += |
|
|
sad8bi(frame->image.y + 2*(i+(k&1))*8 + 2*(j+(k>>1))*8*edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
2*(i+(k&1)), 2*(j+(k>>1)), 8, &mb->mvs[k], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
2*(i+(k&1)), 2*(j+(k>>1)), 8, &mb->b_mvs[k], edged_width), |
|
|
edged_width); |
|
|
} |
|
|
} |
|
|
else |
|
|
{ |
|
|
mb->directmv[3] = mb->directmv[2] = mb->directmv[1] = |
|
|
mb->directmv[0] = b_mb->mvs[0]; |
|
|
|
|
|
mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); |
|
|
mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].x) / TRD |
|
|
: mb->mvs[0].x - mb->directmv[0].x); |
|
|
|
|
|
mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); |
|
|
mb->b_mvs[0].y = (int32_t) ((mb->directmv[0].y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].y) / TRD |
|
|
: mb->mvs[0].y - mb->directmv[0].y); |
|
|
|
|
|
d_sad16 = sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, 16, &mb->mvs[0], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, &mb->b_mvs[0], edged_width), |
|
|
edged_width); |
|
|
|
|
|
} |
|
|
d_sad16 += calc_delta_16(mb->deltamv.x, mb->deltamv.y, 1, frame->quant); |
|
|
|
|
|
// forward search |
|
|
f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
&frame->image, i, j, |
|
|
mb->mvs[0].x, mb->mvs[0].y, /* start point f_directMV */ |
|
|
f_predMV.x, f_predMV.y, /* center is f-prediction */ |
|
|
frame->motion_flags, |
|
|
frame->quant, frame->fcode, pParam, |
|
|
f_mbs, f_mbs, |
|
|
&mb->mvs[0], &pmv_dontcare); |
|
|
|
|
|
|
|
|
// backward search |
|
|
b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
&frame->image, i, j, |
|
|
mb->b_mvs[0].x, mb->b_mvs[0].y, /* start point b_directMV */ |
|
|
b_predMV.x, b_predMV.y, /* center is b-prediction */ |
|
|
frame->motion_flags, |
|
|
frame->quant, frame->bcode, pParam, |
|
|
b_mbs, b_mbs, |
|
|
&mb->b_mvs[0], &pmv_dontcare); |
|
|
|
|
|
i_sad16 = |
|
|
sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
i, j, 16, &mb->mvs[0], edged_width), |
|
|
get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, 16, &mb->b_mvs[0], edged_width), |
|
|
edged_width); |
|
|
i_sad16 += calc_delta_16(mb->mvs[0].x-f_predMV.x, mb->mvs[0].y-f_predMV.y, |
|
|
frame->fcode, frame->quant); |
|
|
i_sad16 += calc_delta_16(mb->b_mvs[0].x-b_predMV.x, mb->b_mvs[0].y-b_predMV.y, |
|
|
frame->bcode, frame->quant); |
|
|
|
|
|
get_range(&f_min_dx, &f_max_dx, &f_min_dy, &f_max_dy, i, j, 16, iWidth, iHeight, |
|
|
frame->fcode); |
|
|
get_range(&b_min_dx, &b_max_dx, &b_min_dy, &b_max_dy, i, j, 16, iWidth, iHeight, |
|
|
frame->bcode); |
|
|
|
|
|
/* Interpolated MC motion vector search, this is tedious and more complicated because there are |
|
|
two values for everything, always one for backward and one for forward ME. Still, we don't gain |
|
|
much from this search, maybe it should simply be skipped and simply current i_sad16 value used |
|
|
as "optimal". */ |
|
|
|
|
|
i_sad16 = Diamond16_InterpolMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
mb->mvs[0].x, mb->mvs[0].y, |
|
|
mb->b_mvs[0].x, mb->b_mvs[0].y, |
|
|
i_sad16, |
|
|
&f_interpolMV, &b_interpolMV, |
|
|
f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, |
|
|
f_min_dx, f_max_dx, f_min_dy, f_max_dy, |
|
|
b_min_dx, b_max_dx, b_min_dy, b_max_dy, |
|
|
edged_width, 2, |
|
|
frame->fcode, frame->bcode,frame->quant,0); |
|
|
|
|
|
i_sad16 = Diamond16_InterpolMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i * 16 + j * 16 * edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
f_interpolMV.x, f_interpolMV.y, |
|
|
b_interpolMV.x, b_interpolMV.y, |
|
|
i_sad16, |
|
|
&f_interpolMV, &b_interpolMV, |
|
|
f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, |
|
|
f_min_dx, f_max_dx, f_min_dy, f_max_dy, |
|
|
b_min_dx, b_max_dx, b_min_dy, b_max_dy, |
|
|
edged_width, 1, |
|
|
frame->fcode, frame->bcode,frame->quant,0); // equiv to halfpel refine |
|
|
|
|
|
|
|
|
/* DIRECT MODE DELTA VECTOR SEARCH. |
|
|
This has to be made more effective, but at the moment I'm happy it's running at all */ |
|
|
|
|
|
/* There are two range restrictions for direct mode: deltaMV is limited to [-32,31] in halfpel units, and |
|
|
absolute vector must not lie outside of image dimensions. Constraint one is dealt with by CHECK_MV16_DIRECT |
|
|
and for constraint two we need distance to boundary. This is done by get_range very large fcode (hack!) */ |
|
|
|
|
|
get_range(&min_dx, &max_dx, &min_dy, &max_dy, i, j, 16, iWidth, iHeight, 19); |
|
|
|
|
|
d_sad16 = Diamond16_DirectMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i*16 + j*16*edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
TRB,TRD, |
|
|
0,0, |
|
|
d_sad16, |
|
|
&mb->deltamv, |
|
|
mb->directmv, // this has to be pre-initialized with b_mb->mvs[] |
|
|
min_dx, max_dx, min_dy, max_dy, |
|
|
edged_width, 2, frame->quant, 0); |
|
|
|
|
|
d_sad16 = Diamond16_DirectMainSearch( |
|
|
f_ref->y, f_refH->y, f_refV->y, f_refHV->y, |
|
|
frame->image.y + i*16 + j*16*edged_width, |
|
|
b_ref->y, b_refH->y, b_refV->y, b_refHV->y, |
|
|
i, j, |
|
|
TRB,TRD, |
|
|
mb->deltamv.x, mb->deltamv.y, |
|
|
d_sad16, |
|
|
&mb->deltamv, |
|
|
mb->directmv, // this has to be pre-initialized with b_mb->mvs[] |
|
|
min_dx, max_dx, min_dy, max_dy, |
|
|
edged_width, 1, frame->quant, 0); // equiv to halfpel refine |
|
|
|
|
|
|
|
|
// i_sad16 = 65535; /* remove the comment to disable any of the MODEs */ |
|
|
// f_sad16 = 65535; |
|
|
// b_sad16 = 65535; |
|
|
// d_sad16 = 65535; |
|
|
|
|
|
if (f_sad16 < b_sad16) { |
|
|
best_sad = f_sad16; |
|
|
mb->mode = MODE_FORWARD; |
|
|
} else { |
|
|
best_sad = b_sad16; |
|
|
mb->mode = MODE_BACKWARD; |
|
|
} |
|
|
|
|
|
if (i_sad16 < best_sad) { |
|
|
best_sad = i_sad16; |
|
|
mb->mode = MODE_INTERPOLATE; |
|
|
} |
|
|
|
|
|
if (d_sad16 < best_sad) { |
|
|
|
|
|
if (b_mb->mode == MODE_INTER4V) |
|
|
{ |
|
|
|
|
|
/* how to calc vectors is defined in standard. mvs[] and b_mvs[] are only for motion compensation */ |
|
|
/* for the bitstream, the value mb->deltamv is read directly */ |
|
|
|
|
|
for (k = 0; k < 4; k++) { |
|
|
|
|
|
mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); |
|
|
mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].x) / TRD |
|
|
: mb->mvs[k].x - mb->directmv[k].x); |
|
|
|
|
|
mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); |
|
|
mb->b_mvs[k].y = (int32_t) ((mb->deltamv.y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[k].y) / TRD |
|
|
: mb->mvs[k].y - mb->directmv[k].y); |
|
|
} |
|
|
} |
|
|
else |
|
|
{ |
|
|
mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); |
|
|
|
|
|
mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].x) / TRD |
|
|
: mb->mvs[0].x - mb->directmv[0].x); |
|
|
|
|
|
mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); |
|
|
|
|
|
mb->b_mvs[0].y = (int32_t) ((mb->deltamv.y == 0) |
|
|
? ((TRB - TRD) * mb->directmv[0].y) / TRD |
|
|
: mb->mvs[0].y - mb->directmv[0].y); |
|
|
|
|
|
mb->mvs[3] = mb->mvs[2] = mb->mvs[1] = mb->mvs[0]; |
|
|
mb->b_mvs[3] = mb->b_mvs[2] = mb->b_mvs[1] = mb->b_mvs[0]; |
|
|
} |
|
|
|
|
|
best_sad = d_sad16; |
|
|
mb->mode = MODE_DIRECT; |
|
|
} |
|
|
|
|
|
switch (mb->mode) |
|
|
{ |
|
|
case MODE_FORWARD: |
|
|
f_count++; |
|
|
f_predMV = mb->mvs[0]; |
|
|
break; |
|
|
case MODE_BACKWARD: |
|
|
b_count++; |
|
|
b_predMV = mb->b_mvs[0]; |
|
|
|
|
|
break; |
|
|
case MODE_INTERPOLATE: |
|
|
i_count++; |
|
|
mb->mvs[0] = f_interpolMV; |
|
|
mb->b_mvs[0] = b_interpolMV; |
|
|
f_predMV = mb->mvs[0]; |
|
|
b_predMV = mb->b_mvs[0]; |
|
|
break; |
|
|
case MODE_DIRECT: |
|
|
d_count++; |
|
|
break; |
|
|
default: |
|
|
break; |
|
|
} |
|
|
|
|
|
} |
|
|
} |
|
|
|
|
|
#ifdef _DEBUG_BFRAME_STAT |
|
|
fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D: %04d\n", |
|
|
f_count,b_count,i_count,d_count); |
|
|
#endif |
|
|
|
|
|
} |
|