--- trunk/xvidcore/src/motion/motion_est.c 2002/07/10 19:16:56 285 +++ trunk/xvidcore/src/motion/motion_est.c 2002/07/19 14:56:00 317 @@ -171,9 +171,8 @@ const IMAGE *const pCurrent = ¤t->image; const IMAGE *const pRef = &reference->image; - const VECTOR zeroMV = { 0, 0 }; + static const VECTOR zeroMV = { 0, 0 }; - int64_t time; int32_t x, y; int32_t iIntra = 0; VECTOR pmv; @@ -1047,7 +1046,7 @@ return iMinSAD; } - +Halfpel8_RefineFuncPtr Halfpel8_Refine; int32_t Halfpel16_Refine(const uint8_t * const pRef, @@ -1209,7 +1208,7 @@ if ((iMinSAD < 256) || ((MVequal(*currMV, prevMB->mvs[0])) && - ((uint32_t) iMinSAD < prevMB->sad16))) { + ((int32_t) iMinSAD < prevMB->sad16))) { if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode { if (!MVzero(*currMV)) { @@ -1313,7 +1312,7 @@ if ((iMinSAD <= threshA) || (MVequal(*currMV, prevMB->mvs[0]) && - ((uint32_t) iMinSAD < prevMB->sad16))) { + ((int32_t) iMinSAD < prevMB->sad16))) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast16_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -1473,7 +1472,7 @@ } int32_t -Halfpel8_Refine(const uint8_t * const pRef, +Halfpel8_Refine_c(const uint8_t * const pRef, const uint8_t * const pRefH, const uint8_t * const pRefV, const uint8_t * const pRefHV, @@ -1628,7 +1627,7 @@ (uint8_t) iFcode, iQuant); if ((iMinSAD < 256 / 4) || ((MVequal(*currMV, prevMB->mvs[iSubBlock])) - && ((uint32_t) iMinSAD < + && ((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; @@ -1683,7 +1682,7 @@ if ((iMinSAD <= threshA) || (MVequal(*currMV, prevMB->mvs[iSubBlock]) && - ((uint32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { + ((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -1742,7 +1741,7 @@ if ((iMinSAD <= threshA) || (MVequal(*currMV, prevMB->mvs[iSubBlock]) && - ((uint32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { + ((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast8_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -1928,7 +1927,7 @@ // thresh1 is fixed to 256 if ((iMinSAD < 256) || ((MVequal(*currMV, prevMB->mvs[0])) && - ((uint32_t) iMinSAD < prevMB->sad16))) { + ((int32_t) iMinSAD < prevMB->sad16))) { if (MotionFlags & PMV_QUICKSTOP16) goto EPZS16_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -1988,7 +1987,7 @@ if ((iMinSAD <= thresh2) || (MVequal(*currMV, prevMB->mvs[0]) && - ((uint32_t) iMinSAD <= prevMB->sad16))) { + ((int32_t) iMinSAD <= prevMB->sad16))) { if (MotionFlags & PMV_QUICKSTOP16) goto EPZS16_Terminate_without_Refine; if (MotionFlags & PMV_EARLYSTOP16) @@ -2348,6 +2347,291 @@ +int32_t +PMVfastIntSearch16(const uint8_t * const pRef, + const uint8_t * const pRefH, + const uint8_t * const pRefV, + const uint8_t * const pRefHV, + const IMAGE * const pCur, + const int x, + const int y, + const uint32_t MotionFlags, + const uint32_t iQuant, + const uint32_t iFcode, + const MBParam * const pParam, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, + VECTOR * const currMV, + VECTOR * const currPMV) +{ + const uint32_t iWcount = pParam->mb_width; + const int32_t iWidth = pParam->width; + const int32_t iHeight = pParam->height; + const int32_t iEdgedWidth = pParam->edged_width; + + const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; + const VECTOR zeroMV = { 0, 0 }; + + int32_t iDiamondSize; + + int32_t min_dx; + int32_t max_dx; + int32_t min_dy; + int32_t max_dy; + + int32_t iFound; + + VECTOR newMV; + VECTOR backupMV; /* just for PMVFAST */ + + VECTOR pmv[4]; + int32_t psad[4]; + + MainSearch16FuncPtr MainSearchPtr; + + const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; + MACROBLOCK *const pMB = pMBs + x + y * iWcount; + + int32_t threshA, threshB; + int32_t bPredEq; + int32_t iMinSAD, iSAD; + +/* Get maximum range */ + get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, + iFcode); + +/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ + + if ((x == 0) && (y == 0)) { + threshA = 512; + threshB = 1024; + + bPredEq = 0; + psad[0] = psad[1] = psad[2] = psad[3] = 0; + *currMV = pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV; + + } else { + threshA = psad[0]; + threshB = threshA + 256; + if (threshA < 512) + threshA = 512; + if (threshA > 1024) + threshA = 1024; + if (threshB > 1792) + threshB = 1792; + + bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); + *currMV = pmv[0]; /* current best := prediction */ + } + + iFound = 0; + +/* Step 4: Calculate SAD around the Median prediction. + MinSAD=SAD + If Motion Vector equal to Previous frame motion vector + and MinSADx > max_dx) { + currMV->x = EVEN(max_dx); + } + if (currMV->x < min_dx) { + currMV->x = EVEN(min_dx); + } + if (currMV->y > max_dy) { + currMV->y = EVEN(max_dy); + } + if (currMV->y < min_dy) { + currMV->y = EVEN(min_dy); + } + + iMinSAD = + sad16(cur, + get_iref_mv(pRef, x, y, 16, currMV, + iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); + iMinSAD += + calc_delta_16(currMV->x - pmv[0].x, currMV->y - pmv[0].y, + (uint8_t) iFcode, iQuant); + + if ((iMinSAD < 256) || + ((MVequal(*currMV, prevMB->i_mvs[0])) && + ((int32_t) iMinSAD < prevMB->i_sad16))) { + if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode + { + if (!MVzero(*currMV)) { + iMinSAD += MV16_00_BIAS; + CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures + iMinSAD -= MV16_00_BIAS; + } + } + + if (MotionFlags & PMV_EARLYSTOP16) + goto PMVfastInt16_Terminate_with_Refine; + } + + +/* Step 2 (lazy eval): 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->i_mvs[0]))) + iFound = 2; + +/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. + Otherwise select large Diamond Search. +*/ + + if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) + iDiamondSize = 2; // halfpel units! + else + iDiamondSize = 4; // halfpel units! + +/* + Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. + Also calculate (0,0) but do not subtract offset. + Let MinSAD be the smallest SAD up to this point. + If MV is (0,0) subtract offset. +*/ + +// (0,0) is often a good choice + + if (!MVzero(pmv[0])) + CHECK_MV16_ZERO; + +// previous frame MV is always possible + + if (!MVzero(prevMB->i_mvs[0])) + if (!MVequal(prevMB->i_mvs[0], pmv[0])) + CHECK_MV16_CANDIDATE(prevMB->i_mvs[0].x, prevMB->i_mvs[0].y); + +// left neighbour, if allowed + + if (!MVzero(pmv[1])) + if (!MVequal(pmv[1], prevMB->i_mvs[0])) + if (!MVequal(pmv[1], pmv[0])) + CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); + +// top neighbour, if allowed + if (!MVzero(pmv[2])) + if (!MVequal(pmv[2], prevMB->i_mvs[0])) + if (!MVequal(pmv[2], pmv[0])) + if (!MVequal(pmv[2], pmv[1])) + CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); + +// top right neighbour, if allowed + if (!MVzero(pmv[3])) + if (!MVequal(pmv[3], prevMB->i_mvs[0])) + if (!MVequal(pmv[3], pmv[0])) + if (!MVequal(pmv[3], pmv[1])) + if (!MVequal(pmv[3], pmv[2])) + CHECK_MV16_CANDIDATE(pmv[3].x, + pmv[3].y); + + if ((MVzero(*currMV)) && + (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) + iMinSAD -= MV16_00_BIAS; + + +/* Step 6: If MinSAD <= thresa goto Step 10. + If Motion Vector equal to Previous frame motion vector and MinSADi_mvs[0]) && + ((int32_t) iMinSAD < prevMB->i_sad16))) { + + if (MotionFlags & PMV_EARLYSTOP16) + goto PMVfastInt16_Terminate_with_Refine; + } + + +/************ (Diamond Search) **************/ +/* + Step 7: Perform Diamond search, with either the small or large diamond. + If Found=2 only examine one Diamond pattern, and afterwards goto step 10 + Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. + If center then goto step 10. + Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. + Refine by using small diamond and goto step 10. +*/ + + if (MotionFlags & PMV_USESQUARES16) + MainSearchPtr = Square16_MainSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) + MainSearchPtr = AdvDiamond16_MainSearch; + else + MainSearchPtr = Diamond16_MainSearch; + + backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + + +/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ + iSAD = + (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, + currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, + min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, + iQuant, iFound); + + if (iSAD < iMinSAD) { + *currMV = newMV; + iMinSAD = iSAD; + } + + if (MotionFlags & PMV_EXTSEARCH16) { +/* extended: search (up to) two more times: orignal prediction and (0,0) */ + + if (!(MVequal(pmv[0], backupMV))) { + iSAD = + (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, + pmv[0].x, pmv[0].y, iMinSAD, &newMV, pmv, + min_dx, max_dx, min_dy, max_dy, iEdgedWidth, + iDiamondSize, iFcode, iQuant, iFound); + + if (iSAD < iMinSAD) { + *currMV = newMV; + iMinSAD = iSAD; + } + } + + if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { + iSAD = + (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, + iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, + max_dy, iEdgedWidth, iDiamondSize, iFcode, + iQuant, iFound); + + if (iSAD < iMinSAD) { + *currMV = newMV; + iMinSAD = iSAD; + } + } + } + +/* + Step 10: The motion vector is chosen according to the block corresponding to MinSAD. +*/ + +PMVfastInt16_Terminate_with_Refine: + + pMB->i_mvs[0] = pMB->i_mvs[1] = pMB->i_mvs[2] = pMB->i_mvs[3] = pMB->i_mv16 = *currMV; + pMB->i_sad8[0] = pMB->i_sad8[1] = pMB->i_sad8[2] = pMB->i_sad8[3] = pMB->i_sad16 = iMinSAD; + + if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step + iMinSAD = + Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, + iMinSAD, pmv, min_dx, max_dx, min_dy, max_dy, + iFcode, iQuant, iEdgedWidth); + + pmv[0] = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); // get _REAL_ prediction (halfpel possible) + +PMVfastInt16_Terminate_without_Refine: + currPMV->x = currMV->x - pmv[0].x; + currPMV->y = currMV->y - pmv[0].y; + return iMinSAD; +} + /* *********************************************************** @@ -2359,6 +2643,8 @@ 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, @@ -2372,20 +2658,33 @@ const IMAGE * const b_refV, const IMAGE * const b_refHV) { - const uint32_t mb_width = pParam->mb_width; - const uint32_t mb_height = pParam->mb_height; - const int32_t edged_width = pParam->edged_width; - - uint32_t i, j; - - int32_t f_sad16; - int32_t b_sad16; - int32_t i_sad16; - int32_t d_sad16; - int32_t best_sad; + const int mb_width = pParam->mb_width; + const int mb_height = pParam->mb_height; + const int edged_width = pParam->edged_width; + + 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 linear motion) */ + + int best_sad; VECTOR pmv_dontcare; + int f_count=0; + int b_count=0; + int i_count=0; + int d_count=0; + int dnv_count=0; + int s_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++) { for (i = 0; i < mb_width; i++) { @@ -2393,8 +2692,13 @@ const MACROBLOCK *f_mb = &f_mbs[i + j * mb_width]; const MACROBLOCK *b_mb = &b_mbs[i + j * mb_width]; + VECTOR directMV; + VECTOR deltaMV=zeroMV; + +/* special case, if collocated block is SKIPed: encoding is forward(0,0) */ + if (b_mb->mode == MODE_INTER && b_mb->cbp == 0 && - b_mb->mvs[0].x == 0 && b_mb->mvs[0].y == 0) { + b_mb->mvs[0].x == 0 && b_mb->mvs[0].y == 0) { mb->mode = MODE_NOT_CODED; mb->mvs[0].x = 0; mb->mvs[0].y = 0; @@ -2402,26 +2706,37 @@ mb->b_mvs[0].y = 0; continue; } - /* force F_SAD16 - f_sad16 = 100; - b_sad16 = 65535; - - mb->mode = MODE_FORWARD; - mb->mvs[0].x = 1; - mb->mvs[0].y = 1; - mb->b_mvs[0].x = 1; - mb->b_mvs[0].y = 1; - continue; - ^^ force F_SAD16 */ + /* same method of scaling as in decoder.c, so we copy 1:1 from there */ + for (k = 0; k < 4; k++) { + directMV = b_mb->mvs[k]; + + mb->mvs[k].x = (int32_t) ((TRB * directMV.x) / TRD + deltaMV.x); + mb->b_mvs[k].x = (int32_t) ((deltaMV.x == 0) + ? ((TRB - TRD) * directMV.x) / TRD + : mb->mvs[k].x - directMV.x); + mb->mvs[k].y = (int32_t) ((TRB * directMV.y) / TRD + deltaMV.y); + mb->b_mvs[k].y = (int32_t) ((deltaMV.y == 0) + ? ((TRB - TRD) * directMV.y) / TRD + : mb->mvs[k].y - directMV.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); + // forward search - f_sad16 = - SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - &frame->image, i, j, frame->motion_flags, - frame->quant, frame->fcode, pParam, - f_mbs, f_mbs, /* todo */ - &mb->mvs[0], &pmv_dontcare); // ignore pmv + f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + &frame->image, i, j, frame->motion_flags, + frame->quant, frame->fcode, pParam, + f_mbs, f_mbs, /* todo */ + &mb->mvs[0], &pmv_dontcare); // ignore pmv (why?) + // backward search b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, @@ -2429,29 +2744,29 @@ frame->quant, frame->bcode, pParam, b_mbs, b_mbs, /* todo */ &mb->b_mvs[0], &pmv_dontcare); // ignore pmv - - // interpolate search (simple, but effective) - i_sad16 = 65535; - - /* - x/y range somewhat buggy + i_sad16 = - sad16bi_c(frame->image.y + i * 16 + j * 16 * edged_width, - get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - i, j, 16, mb->mvs[0].x, mb->mvs[0].y, - edged_width), get_ref(b_ref->y, b_refH->y, - b_refV->y, b_refHV->y, - i, j, 16, - mb->b_mvs[0].x, - mb->b_mvs[0].x, - edged_width), + 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(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + i, j, 16, -mb->b_mvs[0].x, -mb->b_mvs[0].y, edged_width), edged_width); - */ // TODO: direct search // predictor + range of [-32,32] - d_sad16 = 65535; + DEBUG2("f_MV: ",mb->mvs[0].x,mb->mvs[0].y); + DEBUG2("b_MV: ",mb->b_mvs[0].x,mb->b_mvs[0].y); + +/* fprintf(stderr,"f_sad16 = %d, b_sad16 = %d, i_sad16 = %d, d_sad16 = %d\n", + f_sad16,b_sad16,i_sad16,d_sad16); +*/ + +// d_sad16 -= 50; +// d_sad16 = 65535; +// i_sad16 = 65535; +// b_sad16 = 65535; if (f_sad16 < b_sad16) { best_sad = f_sad16; @@ -2468,9 +2783,31 @@ if (d_sad16 < best_sad) { best_sad = d_sad16; - mb->mode = MODE_DIRECT; + mb->mode = MODE_DIRECT_NONE_MV; } + switch (mb->mode) + { + case MODE_FORWARD: + f_count++; break; + case MODE_BACKWARD: + b_count++; break; + case MODE_INTERPOLATE: + i_count++; break; + case MODE_DIRECT: + d_count++; break; + case MODE_DIRECT_NONE_MV: + dnv_count++; break; + default: + s_count++; break; + } + } } + +#ifdef _DEBUG_BFRAME_STAT + fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D0: %04d D: %04d S: %04d\n", + f_count,b_count,i_count,dnv_count,d_count,s_count); +#endif + }