54 |
#define MV16_THRESHOLD 192 |
#define MV16_THRESHOLD 192 |
55 |
#define MV8_THRESHOLD 56 |
#define MV8_THRESHOLD 56 |
56 |
|
|
57 |
|
#define NEIGH_MOVE_THRESH 8 |
58 |
|
// how much a block's MV must differ from his neighbour |
59 |
|
// to be search for INTER4V. The more, the faster... |
60 |
|
|
61 |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ |
62 |
/* nb = vop pixels * 2^(bpp-8) */ |
/* nb = vop pixels * 2^(bpp-8) */ |
63 |
#define MV16_00_BIAS (128+1) |
#define MV16_00_BIAS (128+1) |
64 |
#define MV8_00_BIAS (0) |
#define MV8_00_BIAS (0) |
65 |
|
|
66 |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ |
67 |
#define INTER_BIAS 512 |
#define MV16_INTER_BIAS 512 |
68 |
|
|
69 |
/* Parameters which control inter/inter4v decision */ |
/* Parameters which control inter/inter4v decision */ |
70 |
#define IMV16X16 5 |
#define IMV16X16 5 |
73 |
#define NEIGH_TEND_16X16 2 |
#define NEIGH_TEND_16X16 2 |
74 |
#define NEIGH_TEND_8X8 2 |
#define NEIGH_TEND_8X8 2 |
75 |
|
|
|
|
|
76 |
// fast ((A)/2)*2 |
// fast ((A)/2)*2 |
77 |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) |
78 |
|
|
79 |
|
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
80 |
|
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
81 |
|
|
82 |
int32_t PMVfastSearch16( |
int32_t PMVfastSearch16( |
83 |
const uint8_t * const pRef, |
const uint8_t * const pRef, |
276 |
{ |
{ |
277 |
const uint32_t iWcount = pParam->mb_width; |
const uint32_t iWcount = pParam->mb_width; |
278 |
const uint32_t iHcount = pParam->mb_height; |
const uint32_t iHcount = pParam->mb_height; |
279 |
MACROBLOCK * pMBs = current->mbs; |
MACROBLOCK * const pMBs = current->mbs; |
280 |
IMAGE * pCurrent = ¤t->image; |
MACROBLOCK * const prevMBs = reference->mbs; // previous frame |
|
|
|
|
MACROBLOCK * prevMBs = reference->mbs; // previous frame |
|
|
IMAGE * pRef = &reference->image; |
|
281 |
|
|
282 |
|
const IMAGE * const pCurrent = ¤t->image; |
283 |
|
const IMAGE * const pRef = &reference->image; |
284 |
|
|
285 |
uint32_t i, j, iIntra = 0; |
const VECTOR zeroMV = {0,0}; |
286 |
|
|
287 |
VECTOR mv16; |
int32_t x, y; |
288 |
VECTOR pmv16; |
int32_t iIntra = 0; |
289 |
|
VECTOR pmv; |
|
int32_t sad8 = 0; |
|
|
int32_t sad16; |
|
|
int32_t deviation; |
|
290 |
|
|
291 |
if (sadInit) |
if (sadInit) |
292 |
(*sadInit)(); |
(*sadInit)(); |
293 |
|
|
294 |
// note: i==horizontal, j==vertical |
for (y = 0; y < iHcount; y++) |
295 |
for (i = 0; i < iHcount; i++) |
for (x = 0; x < iWcount; x++) |
|
for (j = 0; j < iWcount; j++) |
|
296 |
{ |
{ |
297 |
MACROBLOCK *pMB = &pMBs[j + i * iWcount]; |
MACROBLOCK* const pMB = &pMBs[x + y * iWcount]; |
|
MACROBLOCK *prevMB = &prevMBs[j + i * iWcount]; |
|
|
|
|
|
sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
|
|
j, i, current->motion_flags, current->quant, current->fcode, |
|
|
pParam, pMBs, prevMBs, &mv16, &pmv16); |
|
|
pMB->sad16=sad16; |
|
298 |
|
|
299 |
|
pMB->sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
300 |
|
x, y, current->motion_flags, current->quant, current->fcode, |
301 |
|
pParam, pMBs, prevMBs, &pMB->mv16, &pMB->pmvs[0]); |
302 |
|
} |
303 |
|
|
304 |
/* decide: MODE_INTER or MODE_INTRA |
for (y = 0; y < iHcount; y++) |
305 |
if (dev_intra < sad_inter - 2 * nb) use_intra |
for (x = 0; x < iWcount; x++) |
306 |
*/ |
{ |
307 |
|
MACROBLOCK* const pMB = &pMBs[x + y * iWcount]; |
308 |
|
|
309 |
deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); |
if (0 < (pMB->sad16 - MV16_INTER_BIAS)) |
310 |
|
{ |
311 |
|
int32_t deviation; |
312 |
|
deviation = dev16(pCurrent->y + x*16 + y*16*pParam->edged_width, |
313 |
|
pParam->edged_width); |
314 |
|
|
315 |
if (deviation < (sad16 - INTER_BIAS)) |
if (deviation < (pMB->sad16 - MV16_INTER_BIAS)) |
316 |
{ |
{ |
317 |
pMB->mode = MODE_INTRA; |
pMB->mode = MODE_INTRA; |
318 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; |
pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] |
319 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; |
= pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
320 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] |
321 |
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; |
= pMB->sad8[2] = pMB->sad8[3] = 0; |
322 |
|
|
323 |
iIntra++; |
iIntra++; |
324 |
if(iIntra >= iLimit) |
if(iIntra >= iLimit) |
326 |
|
|
327 |
continue; |
continue; |
328 |
} |
} |
329 |
|
} |
330 |
|
|
331 |
if (current->global_flags & XVID_INTER4V) |
if ( (current->global_flags & XVID_INTER4V) |
332 |
|
&& (!(current->global_flags & XVID_LUMIMASKING) |
333 |
|
|| pMB->dquant == NO_CHANGE) ) |
334 |
{ |
{ |
335 |
pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
int32_t neigh=0; |
336 |
2 * j, 2 * i, mv16.x, mv16.y, |
|
337 |
|
if (x>0) |
338 |
|
{ neigh += abs((pMB->mv16.x)-((pMB-1)->mv16.x)); |
339 |
|
neigh += abs((pMB->mv16.y)-((pMB-1)->mv16.y)); |
340 |
|
} |
341 |
|
if (y>0) |
342 |
|
{ neigh += abs((pMB->mv16.x)-((pMB-iWcount)->mv16.x)); |
343 |
|
neigh += abs((pMB->mv16.y)-((pMB-iWcount)->mv16.y)); |
344 |
|
} |
345 |
|
if (x<(iWcount-1)) |
346 |
|
{ neigh += abs((pMB->mv16.x)-((pMB+1)->mv16.x)); |
347 |
|
neigh += abs((pMB->mv16.y)-((pMB+1)->mv16.y)); |
348 |
|
} |
349 |
|
if (y<(iHcount-1)) |
350 |
|
{ neigh += abs((pMB->mv16.x)-((pMB+iHcount)->mv16.x)); |
351 |
|
neigh += abs((pMB->mv16.y)-((pMB+iHcount)->mv16.y)); |
352 |
|
} |
353 |
|
|
354 |
|
if (neigh > NEIGH_MOVE_THRESH) |
355 |
|
{ |
356 |
|
int32_t sad8 = 129; //IMV16X16 * current->quant; |
357 |
|
|
358 |
|
if (sad8 < pMB->sad16) |
359 |
|
sad8 += pMB->sad8[0] |
360 |
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
361 |
|
2*x, 2*y, pMB->mv16.x, pMB->mv16.y, |
362 |
current->motion_flags, current->quant, current->fcode, |
current->motion_flags, current->quant, current->fcode, |
363 |
pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
pParam, pMBs, prevMBs, &pMB->mvs[0], &pMB->pmvs[0]); |
364 |
|
|
365 |
pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (sad8 < pMB->sad16) |
366 |
2 * j + 1, 2 * i, mv16.x, mv16.y, |
sad8 += pMB->sad8[1] |
367 |
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
368 |
|
2*x+1, 2*y, pMB->mv16.x, pMB->mv16.y, |
369 |
current->motion_flags, current->quant, current->fcode, |
current->motion_flags, current->quant, current->fcode, |
370 |
pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
pParam, pMBs, prevMBs, &pMB->mvs[1], &pMB->pmvs[1]); |
371 |
|
|
372 |
pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (sad8 < pMB->sad16) |
373 |
2 * j, 2 * i + 1, mv16.x, mv16.y, |
sad8 += pMB->sad8[2] |
374 |
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
375 |
|
2*x, 2*y+1, pMB->mv16.x, pMB->mv16.y, |
376 |
current->motion_flags, current->quant, current->fcode, |
current->motion_flags, current->quant, current->fcode, |
377 |
pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
pParam, pMBs, prevMBs, &pMB->mvs[2], &pMB->pmvs[2]); |
378 |
|
|
379 |
pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
if (sad8 < pMB->sad16) |
380 |
2 * j + 1, 2 * i + 1, mv16.x, mv16.y, |
sad8 += pMB->sad8[3] |
381 |
|
= SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, |
382 |
|
2*x+1, 2*y+1, pMB->mv16.x, pMB->mv16.y, |
383 |
current->motion_flags, current->quant, current->fcode, |
current->motion_flags, current->quant, current->fcode, |
384 |
pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
pParam, pMBs, prevMBs, &pMB->mvs[3], &pMB->pmvs[3]); |
385 |
|
|
|
sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; |
|
|
} |
|
|
|
|
|
|
|
386 |
/* decide: MODE_INTER or MODE_INTER4V |
/* decide: MODE_INTER or MODE_INTER4V |
387 |
mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v |
mpeg4: if (sad8 < pMB->sad16 - nb/2+1) use_inter4v |
388 |
*/ |
*/ |
389 |
|
|
390 |
if (!(current->global_flags & XVID_LUMIMASKING) || pMB->dquant == NO_CHANGE) |
if (sad8 < pMB->sad16) |
|
{ |
|
|
if (((current->global_flags & XVID_INTER4V)==0) || |
|
|
(sad16 < (sad8 + (int32_t)(IMV16X16 * current->quant)))) |
|
|
{ |
|
|
|
|
|
sad8 = sad16; |
|
|
pMB->mode = MODE_INTER; |
|
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
|
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
|
|
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad16; |
|
|
pMB->pmvs[0].x = pmv16.x; |
|
|
pMB->pmvs[0].y = pmv16.y; |
|
|
} |
|
|
else |
|
391 |
{ |
{ |
392 |
pMB->mode = MODE_INTER4V; |
pMB->mode = MODE_INTER4V; |
393 |
pMB->sad8[0] *= 4; |
pMB->sad8[0] *= 4; |
394 |
pMB->sad8[1] *= 4; |
pMB->sad8[1] *= 4; |
395 |
pMB->sad8[2] *= 4; |
pMB->sad8[2] *= 4; |
396 |
pMB->sad8[3] *= 4; |
pMB->sad8[3] *= 4; |
397 |
|
continue; |
398 |
} |
} |
399 |
} |
} |
400 |
else |
} |
401 |
{ |
|
|
sad8 = sad16; |
|
402 |
pMB->mode = MODE_INTER; |
pMB->mode = MODE_INTER; |
403 |
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; |
404 |
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; |
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = pMB->sad16; |
|
pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad16; |
|
405 |
|
|
406 |
pMB->pmvs[0].x = pmv16.x; |
pmv = get_pmv(pMBs, x, y, pParam->mb_width, 0); |
407 |
pMB->pmvs[0].y = pmv16.y; |
// get_pmv has to be called again. |
408 |
} |
// intra-decision and inter4v change predictors |
409 |
|
|
410 |
|
pMB->pmvs[0].x = pMB->mv16.x - pmv.x; |
411 |
|
pMB->pmvs[0].y = pMB->mv16.y - pmv.y; |
412 |
} |
} |
413 |
|
|
414 |
return 0; |
return 0; |
415 |
} |
} |
416 |
|
|
|
#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) |
|
|
|
|
|
#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) |
|
|
|
|
|
|
|
417 |
#define CHECK_MV16_ZERO {\ |
#define CHECK_MV16_ZERO {\ |
418 |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
if ( (0 <= max_dx) && (0 >= min_dx) \ |
419 |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
&& (0 <= max_dy) && (0 >= min_dy) ) \ |
877 |
VECTOR pmv[4]; |
VECTOR pmv[4]; |
878 |
int32_t psad[4]; |
int32_t psad[4]; |
879 |
|
|
880 |
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
881 |
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
882 |
|
|
883 |
static int32_t threshA,threshB; |
static int32_t threshA,threshB; |
917 |
|
|
918 |
iFound=0; |
iFound=0; |
919 |
|
|
|
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
|
|
iFound=2; |
|
|
|
|
|
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) |
|
|
iDiamondSize=1; // halfpel! |
|
|
else |
|
|
iDiamondSize=2; // halfpel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
|
|
iDiamondSize*=2; |
|
|
|
|
920 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
921 |
MinSAD=SAD |
MinSAD=SAD |
922 |
If Motion Vector equal to Previous frame motion vector |
If Motion Vector equal to Previous frame motion vector |
924 |
If SAD<=256 goto Step 10. |
If SAD<=256 goto Step 10. |
925 |
*/ |
*/ |
926 |
|
|
|
|
|
|
// Prepare for main loop |
|
|
|
|
927 |
*currMV=pmv[0]; /* current best := prediction */ |
*currMV=pmv[0]; /* current best := prediction */ |
928 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
929 |
{ /* This should NOT be necessary! */ |
{ /* This should NOT be necessary! */ |
955 |
|
|
956 |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,prevMB->mvs[0])) && ((uint32_t)iMinSAD < prevMB->sad16) ) ) |
957 |
{ |
{ |
958 |
|
if (iMinSAD < 2*iQuant) // high chances for SKIP-mode |
959 |
|
{ |
960 |
|
if (!MVzero(*currMV)) |
961 |
|
{ |
962 |
|
iMinSAD += MV16_00_BIAS; |
963 |
|
CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures |
964 |
|
iMinSAD -= MV16_00_BIAS; |
965 |
|
} |
966 |
|
} |
967 |
|
|
968 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
969 |
goto PMVfast16_Terminate_without_Refine; |
goto PMVfast16_Terminate_without_Refine; |
971 |
goto PMVfast16_Terminate_with_Refine; |
goto PMVfast16_Terminate_with_Refine; |
972 |
} |
} |
973 |
|
|
974 |
|
|
975 |
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
976 |
|
vector of the median. |
977 |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
978 |
|
*/ |
979 |
|
|
980 |
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[0]) ) ) |
981 |
|
iFound=2; |
982 |
|
|
983 |
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
984 |
|
Otherwise select large Diamond Search. |
985 |
|
*/ |
986 |
|
|
987 |
|
if ( (!MVzero(pmv[0])) || (threshB<1536) || (bPredEq) ) |
988 |
|
iDiamondSize=1; // halfpel! |
989 |
|
else |
990 |
|
iDiamondSize=2; // halfpel! |
991 |
|
|
992 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) |
993 |
|
iDiamondSize*=2; |
994 |
|
|
995 |
/* |
/* |
996 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
997 |
Also calculate (0,0) but do not subtract offset. |
Also calculate (0,0) but do not subtract offset. |
1001 |
|
|
1002 |
// (0,0) is always possible |
// (0,0) is always possible |
1003 |
|
|
1004 |
|
if (!MVzero(pmv[0])) |
1005 |
CHECK_MV16_ZERO; |
CHECK_MV16_ZERO; |
1006 |
|
|
1007 |
// previous frame MV is always possible |
// previous frame MV is always possible |
1008 |
|
|
1009 |
|
if (!MVzero(prevMB->mvs[0])) |
1010 |
|
if (!MVequal(prevMB->mvs[0],pmv[0])) |
1011 |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
CHECK_MV16_CANDIDATE(prevMB->mvs[0].x,prevMB->mvs[0].y); |
1012 |
|
|
1013 |
// left neighbour, if allowed |
// left neighbour, if allowed |
1014 |
if (x != 0) |
|
1015 |
|
if (!MVzero(pmv[1])) |
1016 |
|
if (!MVequal(pmv[1],prevMB->mvs[0])) |
1017 |
|
if (!MVequal(pmv[1],pmv[0])) |
1018 |
{ |
{ |
1019 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1020 |
{ pmv[1].x = EVEN(pmv[1].x); |
{ pmv[1].x = EVEN(pmv[1].x); |
1021 |
pmv[1].y = EVEN(pmv[1].y); |
pmv[1].y = EVEN(pmv[1].y); |
1022 |
} |
} |
1023 |
|
|
1024 |
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); |
1025 |
} |
} |
1026 |
|
|
1027 |
// top neighbour, if allowed |
// top neighbour, if allowed |
1028 |
if (y != 0) |
if (!MVzero(pmv[2])) |
1029 |
|
if (!MVequal(pmv[2],prevMB->mvs[0])) |
1030 |
|
if (!MVequal(pmv[2],pmv[0])) |
1031 |
|
if (!MVequal(pmv[2],pmv[1])) |
1032 |
{ |
{ |
1033 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1034 |
{ pmv[2].x = EVEN(pmv[2].x); |
{ pmv[2].x = EVEN(pmv[2].x); |
1037 |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); |
1038 |
|
|
1039 |
// top right neighbour, if allowed |
// top right neighbour, if allowed |
1040 |
if ((uint32_t)x != (iWcount-1)) |
if (!MVzero(pmv[3])) |
1041 |
|
if (!MVequal(pmv[3],prevMB->mvs[0])) |
1042 |
|
if (!MVequal(pmv[3],pmv[0])) |
1043 |
|
if (!MVequal(pmv[3],pmv[1])) |
1044 |
|
if (!MVequal(pmv[3],pmv[2])) |
1045 |
{ |
{ |
1046 |
if (!(MotionFlags & PMV_HALFPEL16 )) |
if (!(MotionFlags & PMV_HALFPEL16 )) |
1047 |
{ pmv[3].x = EVEN(pmv[3].x); |
{ pmv[3].x = EVEN(pmv[3].x); |
1273 |
int32_t psad[4]; |
int32_t psad[4]; |
1274 |
VECTOR newMV; |
VECTOR newMV; |
1275 |
VECTOR backupMV; |
VECTOR backupMV; |
1276 |
|
VECTOR startMV; |
1277 |
|
|
1278 |
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1279 |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
1280 |
|
|
1281 |
static int32_t threshA,threshB; |
static int32_t threshA,threshB; |
1282 |
int32_t iFound,bPredEq; |
int32_t iFound,bPredEq; |
1283 |
int32_t iMinSAD,iSAD; |
int32_t iMinSAD,iSAD; |
1284 |
|
|
1285 |
int32_t iSubBlock = ((y&1)<<1) + (x&1); |
int32_t iSubBlock = (y&1)+(y&1) + (x&1); |
1286 |
|
|
1287 |
|
/* Init variables */ |
1288 |
|
startMV.x = start_x; |
1289 |
|
startMV.y = start_y; |
1290 |
|
|
1291 |
/* Get maximum range */ |
/* Get maximum range */ |
1292 |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
get_range(&min_dx, &max_dx, &min_dy, &max_dy, |
1319 |
|
|
1320 |
iFound=0; |
iFound=0; |
1321 |
|
|
|
/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
|
|
vector of the median. |
|
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
|
|
*/ |
|
|
|
|
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[iSubBlock]) ) ) |
|
|
iFound=2; |
|
|
|
|
|
/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
|
|
Otherwise select large Diamond Search. |
|
|
*/ |
|
|
|
|
|
if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) |
|
|
iDiamondSize=1; // 1 halfpel! |
|
|
else |
|
|
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
|
|
|
|
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
|
|
iDiamondSize*=2; |
|
|
|
|
1322 |
/* Step 4: Calculate SAD around the Median prediction. |
/* Step 4: Calculate SAD around the Median prediction. |
1323 |
MinSAD=SAD |
MinSAD=SAD |
1324 |
If Motion Vector equal to Previous frame motion vector |
If Motion Vector equal to Previous frame motion vector |
1329 |
|
|
1330 |
// Prepare for main loop |
// Prepare for main loop |
1331 |
|
|
1332 |
currMV->x=start_x; /* start with mv16 */ |
*currMV = startMV; |
|
currMV->y=start_y; |
|
1333 |
|
|
1334 |
iMinSAD = sad8( cur, |
iMinSAD = sad8( cur, |
1335 |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), |
1336 |
iEdgedWidth); |
iEdgedWidth); |
1337 |
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode, iQuant); |
iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode, iQuant); |
1338 |
|
|
1339 |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,prevMB->mvs[iSubBlock])) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,prevMB->mvs[iSubBlock])) |
1340 |
|
&& ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1341 |
{ |
{ |
1342 |
if (MotionFlags & PMV_QUICKSTOP16) |
if (MotionFlags & PMV_QUICKSTOP16) |
1343 |
goto PMVfast8_Terminate_without_Refine; |
goto PMVfast8_Terminate_without_Refine; |
1345 |
goto PMVfast8_Terminate_with_Refine; |
goto PMVfast8_Terminate_with_Refine; |
1346 |
} |
} |
1347 |
|
|
1348 |
|
/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion |
1349 |
|
vector of the median. |
1350 |
|
If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 |
1351 |
|
*/ |
1352 |
|
|
1353 |
|
if ((bPredEq) && (MVequal(pmv[0],prevMB->mvs[iSubBlock]) ) ) |
1354 |
|
iFound=2; |
1355 |
|
|
1356 |
|
/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. |
1357 |
|
Otherwise select large Diamond Search. |
1358 |
|
*/ |
1359 |
|
|
1360 |
|
if ( (!MVzero(pmv[0])) || (threshB<1536/4) || (bPredEq) ) |
1361 |
|
iDiamondSize=1; // 1 halfpel! |
1362 |
|
else |
1363 |
|
iDiamondSize=2; // 2 halfpel = 1 full pixel! |
1364 |
|
|
1365 |
|
if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) |
1366 |
|
iDiamondSize*=2; |
1367 |
|
|
1368 |
|
|
1369 |
/* |
/* |
1370 |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. |
1373 |
If MV is (0,0) subtract offset. |
If MV is (0,0) subtract offset. |
1374 |
*/ |
*/ |
1375 |
|
|
1376 |
// the prediction might be even better than mv16 |
// the median prediction might be even better than mv16 |
1377 |
|
|
1378 |
|
if (!MVequal(pmv[0],startMV)) |
1379 |
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); |
1380 |
|
|
1381 |
// (0,0) is always possible |
// (0,0) if needed |
1382 |
|
if (!MVzero(pmv[0])) |
1383 |
|
if (!MVzero(startMV)) |
1384 |
CHECK_MV8_ZERO; |
CHECK_MV8_ZERO; |
1385 |
|
|
1386 |
// previous frame MV is always possible |
// previous frame MV if needed |
1387 |
|
if (!MVzero(prevMB->mvs[iSubBlock])) |
1388 |
|
if (!MVequal(prevMB->mvs[iSubBlock],startMV)) |
1389 |
|
if (!MVequal(prevMB->mvs[iSubBlock],pmv[0])) |
1390 |
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x,prevMB->mvs[iSubBlock].y); |
1391 |
|
|
1392 |
// left neighbour, if allowed |
if ( (iMinSAD <= threshA) || ( MVequal(*currMV,prevMB->mvs[iSubBlock]) && ((uint32_t)iMinSAD < prevMB->sad8[iSubBlock]) ) ) |
1393 |
if (psad[1] != MV_MAX_ERROR) |
{ |
1394 |
|
if (MotionFlags & PMV_QUICKSTOP16) |
1395 |
|
goto PMVfast8_Terminate_without_Refine; |
1396 |
|
if (MotionFlags & PMV_EARLYSTOP16) |
1397 |
|
goto PMVfast8_Terminate_with_Refine; |
1398 |
|
} |
1399 |
|
|
1400 |
|
|
1401 |
|
// left neighbour, if allowed and needed |
1402 |
|
if (!MVzero(pmv[1])) |
1403 |
|
if (!MVequal(pmv[1],startMV)) |
1404 |
|
if (!MVequal(pmv[1],prevMB->mvs[iSubBlock])) |
1405 |
|
if (!MVequal(pmv[1],pmv[0])) |
1406 |
{ |
{ |
1407 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
1408 |
{ pmv[1].x = EVEN(pmv[1].x); |
{ pmv[1].x = EVEN(pmv[1].x); |
1411 |
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); |
1412 |
} |
} |
1413 |
|
|
1414 |
// top neighbour, if allowed |
// top neighbour, if allowed and needed |
1415 |
if (psad[2] != MV_MAX_ERROR) |
if (!MVzero(pmv[2])) |
1416 |
|
if (!MVequal(pmv[2],startMV)) |
1417 |
|
if (!MVequal(pmv[2],prevMB->mvs[iSubBlock])) |
1418 |
|
if (!MVequal(pmv[2],pmv[0])) |
1419 |
|
if (!MVequal(pmv[2],pmv[1])) |
1420 |
{ |
{ |
1421 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
1422 |
{ pmv[2].x = EVEN(pmv[2].x); |
{ pmv[2].x = EVEN(pmv[2].x); |
1424 |
} |
} |
1425 |
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); |
1426 |
|
|
1427 |
// top right neighbour, if allowed |
// top right neighbour, if allowed and needed |
1428 |
if (psad[3] != MV_MAX_ERROR) |
if (!MVzero(pmv[3])) |
1429 |
|
if (!MVequal(pmv[3],startMV)) |
1430 |
|
if (!MVequal(pmv[3],prevMB->mvs[iSubBlock])) |
1431 |
|
if (!MVequal(pmv[3],pmv[0])) |
1432 |
|
if (!MVequal(pmv[3],pmv[1])) |
1433 |
|
if (!MVequal(pmv[3],pmv[2])) |
1434 |
{ |
{ |
1435 |
if (!(MotionFlags & PMV_HALFPEL8 )) |
if (!(MotionFlags & PMV_HALFPEL8 )) |
1436 |
{ pmv[3].x = EVEN(pmv[3].x); |
{ pmv[3].x = EVEN(pmv[3].x); |
1567 |
int32_t psad[8]; |
int32_t psad[8]; |
1568 |
|
|
1569 |
static MACROBLOCK * oldMBs = NULL; |
static MACROBLOCK * oldMBs = NULL; |
1570 |
const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; |
1571 |
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
const MACROBLOCK * const prevMB = prevMBs + x + y * iWcount; |
1572 |
MACROBLOCK * oldMB = NULL; |
MACROBLOCK * oldMB = NULL; |
1573 |
|
|
1854 |
|
|
1855 |
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
const int32_t iSubBlock = ((y&1)<<1) + (x&1); |
1856 |
|
|
1857 |
const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; |
1858 |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
const MACROBLOCK * const prevMB = prevMBs + (x>>1) + (y>>1) * iWcount; |
1859 |
|
|
1860 |
int32_t bPredEq; |
int32_t bPredEq; |