--- motion_est.c 2002/12/13 11:56:31 1.44.2.34 +++ motion_est.c 2002/12/29 11:22:20 1.44.2.40 @@ -1,7 +1,7 @@ /************************************************************************** * * XVID MPEG-4 VIDEO CODEC - * motion estimation + * motion estimation * * This program is an implementation of a part of one or more MPEG-4 * Video tools as specified in ISO/IEC 14496-2 standard. Those intending @@ -31,6 +31,7 @@ #include #include #include +#include // memcpy #include "../encoder.h" #include "../utils/mbfunctions.h" @@ -52,74 +53,74 @@ (*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } static __inline int -d_mv_bits(int x, int y, const uint32_t iFcode, const int qpel, const int rrv) +d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) { int xb, yb; if (qpel) { x *= 2; y *= 2;} else if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } + x -= pred.x; + y -= pred.y; - if (x == 0) xb = 1; - else { + if (x) { if (x < 0) x = -x; x += (1 << (iFcode - 1)) - 1; x >>= (iFcode - 1); if (x > 32) x = 32; xb = mvtab[x] + iFcode; - } + } else xb = 1; - if (y == 0) yb = 1; - else { + if (y) { if (y < 0) y = -y; y += (1 << (iFcode - 1)) - 1; y >>= (iFcode - 1); if (y > 32) y = 32; yb = mvtab[y] + iFcode; - } + } else yb = 1; return xb + yb; } -static int32_t +static int32_t ChromaSAD(int dx, int dy, const SearchData * const data) { int sad; + const uint32_t stride = data->iEdgedWidth/2; dx = (dx >> 1) + roundtab_79[dx & 0x3]; dy = (dy >> 1) + roundtab_79[dy & 0x3]; if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently + data->temp[5] = dx; data->temp[6] = dy; // backup switch (((dx & 1) << 1) | (dy & 1)) { case 0: - sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); - sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); + dx = dx / 2; dy = dy / 2; + sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride); + sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride); break; case 1: dx = dx / 2; dy = (dy - 1) / 2; - sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); - sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); + sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride); + sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride); break; case 2: dx = (dx - 1) / 2; dy = dy / 2; - sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); - sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); + sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride); + sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride); break; default: dx = (dx - 1) / 2; dy = (dy - 1) / 2; - interpolate8x8_halfpel_hv(data->RefQ, - data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, - data->rounding); - sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); - interpolate8x8_halfpel_hv(data->RefQ, - data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, - data->rounding); - sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); + interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); + sad = sad8(data->CurU, data->RefQ, stride); + + interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); + sad += sad8(data->CurV, data->RefQ, stride); break; } - data->temp[5] = dx; data->temp[6] = dy; data->temp[7] = sad; //backup + data->temp[7] = sad; //backup, part 2 return sad; } static __inline const uint8_t * -GetReference(const int x, const int y, const int dir, const SearchData * const data) +GetReferenceB(const int x, const int y, const int dir, const SearchData * const data) { // dir : 0 = forward, 1 = backward switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { @@ -134,7 +135,19 @@ } } -static uint8_t * +// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate +static __inline const uint8_t * +GetReference(const int x, const int y, const SearchData * const data) +{ + switch ( ((x&1)<<1) | (y&1) ) { + case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); + case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); + case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); + default : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); + } +} + +static uint8_t * Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) { // create or find a qpel-precision reference picture; return pointer to it @@ -145,31 +158,30 @@ const int halfpel_y = y/2; const uint8_t *ref1, *ref2, *ref3, *ref4; - ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases + ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; switch( ((x&1)<<1) + (y&1) ) { case 0: // pure halfpel position - Reference = (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); - Reference += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + return (uint8_t *) ref1; break; case 1: // x halfpel, y qpel - top or bottom during qpel refinement - ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); + ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); break; case 2: // x qpel, y halfpel - left or right during qpel refinement - ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); + ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); break; default: // x and y in qpel resolution - the "corners" (top left/right and // bottom left/right) during qpel refinement - ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); - ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); - ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); + ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); + ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); + ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; @@ -179,7 +191,7 @@ return Reference; } -static uint8_t * +static uint8_t * Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) { // create or find a qpel-precision reference picture; return pointer to it @@ -190,12 +202,12 @@ const int halfpel_y = y/2; const uint8_t *ref1, *ref2, *ref3, *ref4; - ref1 = GetReference(halfpel_x, halfpel_y, dir, data); // this reference is used in all cases + ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); switch( ((x&1)<<1) + (y&1) ) { case 0: // pure halfpel position - return (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); + return (uint8_t *) ref1; case 1: // x halfpel, y qpel - top or bottom during qpel refinement - ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); + ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); @@ -203,7 +215,7 @@ break; case 2: // x qpel, y halfpel - left or right during qpel refinement - ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); + ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); @@ -212,9 +224,9 @@ default: // x and y in qpel resolution - the "corners" (top left/right and // bottom left/right) during qpel refinement - ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); - ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); - ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); + ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); + ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); + ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); @@ -226,7 +238,7 @@ /* CHECK_CANDIATE FUNCTIONS START */ -static void +static void CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { int t, xc, yc; @@ -241,11 +253,11 @@ xc = x/2; yc = y/2; //for chroma sad current = data->currentQMV; } else { - Reference = GetReference(x, y, 0, data); + Reference = GetReference(x, y, data); current = data->currentMV; xc = x; yc = y; } - t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode, data->qpel && !data->qpel_precision, 0); + t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); @@ -270,7 +282,7 @@ } -static void +static void CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { int t; @@ -280,8 +292,8 @@ ( x > data->max_dx) || ( x < data->min_dx) || ( y > data->max_dy) || (y < data->min_dy)) return; - Reference = GetReference(x, y, 0, data); - t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode, 0, 1); + Reference = GetReference(x, y, data); + t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); @@ -303,7 +315,7 @@ data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } } -static void +static void CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { int32_t sad; @@ -321,11 +333,11 @@ Reference = Interpolate16x16qpel(x, y, 0, data); current = data->currentQMV; } else { - Reference = GetReference(x, y, 0, data); + Reference = GetReference(x, y, data); current = data->currentMV; } - t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode, - data->qpel && !data->qpel_precision && !data->rrv, data->rrv); + t = d_mv_bits(x, y, data->predMV, data->iFcode, + data->qpel && !data->qpel_precision, data->rrv); sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); sad += (data->lambda16 * t * sad)/1000; @@ -336,26 +348,32 @@ *dir = Direction; } } -static void -CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +static void +CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { // maximum speed - for P/B/I decision - int32_t sad; if (( x > data->max_dx) || ( x < data->min_dx) || ( y > data->max_dy) || (y < data->min_dy)) return; - sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), - data->iEdgedWidth, 256*4096); - - if (sad < *(data->iMinSAD)) { - *(data->iMinSAD) = sad; + data->temp[0] = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), + data->iEdgedWidth, data->temp+1); + if (data->temp[0] < *(data->iMinSAD)) { + *(data->iMinSAD) = data->temp[0]; data->currentMV[0].x = x; data->currentMV[0].y = y; *dir = Direction; } -} + if (data->temp[1] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } + if (data->temp[2] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } + if (data->temp[3] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } + if (data->temp[4] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } +} -static void +static void CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) { int32_t sad; @@ -372,14 +390,14 @@ current = data->currentQMV; ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); } else { - ReferenceF = GetReference(xf, yf, 0, data); + ReferenceF = GetReference(xf, yf, data); xb = data->currentMV[1].x; yb = data->currentMV[1].y; - ReferenceB = GetReference(xb, yb, 1, data); + ReferenceB = GetReferenceB(xb, yb, 1, data); current = data->currentMV; } - t = d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode, data->qpel && !data->qpel_precision, 0) - + d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode, data->qpel && !data->qpel_precision, 0); + t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0) + + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel && !data->qpel_precision, 0); sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); sad += (data->lambda16 * t * sad)/1000; @@ -390,7 +408,7 @@ *dir = Direction; } } -static void +static void CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { int32_t sad = 0; @@ -398,6 +416,7 @@ const uint8_t *ReferenceF; const uint8_t *ReferenceB; VECTOR mvs, b_mvs; + const VECTOR zeroMV={0,0}; if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; @@ -417,8 +436,8 @@ || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; - if (!data->qpel) { - mvs.x *= 2; mvs.y *= 2; + if (!data->qpel) { + mvs.x *= 2; mvs.y *= 2; b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway } ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); @@ -430,7 +449,7 @@ if (sad > *(data->iMinSAD)) return; } - sad += (data->lambda16 * d_mv_bits(x, y, 1, 0, 0) * sad)/1000; + sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -438,13 +457,14 @@ *dir = Direction; } } -static void +static void CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { int32_t sad; const uint8_t *ReferenceF; const uint8_t *ReferenceB; VECTOR mvs, b_mvs; + const VECTOR zeroMV = {0,0}; if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; @@ -463,15 +483,15 @@ || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; - if (!data->qpel) { - mvs.x *= 2; mvs.y *= 2; + if (!data->qpel) { + mvs.x *= 2; mvs.y *= 2; b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway } ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); - sad += (data->lambda16 * d_mv_bits(x, y, 1, 0, 0) * sad)/1000; + sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -489,10 +509,10 @@ || ( y > data->max_dy) || (y < data->min_dy)) return; if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); - else Reference = GetReference(x, y, 0, data); + else Reference = GetReference(x, y, data); sad = sad8(data->Cur, Reference, data->iEdgedWidth); - t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode, data->qpel && !data->qpel_precision, 0); + t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; @@ -532,12 +552,12 @@ CHECK_CANDIDATE(x, y - iDiamondSize, 4); } else { // what remains here is up or down CHECK_CANDIDATE(x + iDiamondSize, y, 2); - CHECK_CANDIDATE(x - iDiamondSize, y, 1); + CHECK_CANDIDATE(x - iDiamondSize, y, 1); } if (iDirection) { bDirection += iDirection; - x = data->currentMV->x; y = data->currentMV->y; + x = data->currentMV->x; y = data->currentMV->y; } } else { //about to quit, eh? not so fast.... switch (bDirection) { @@ -638,7 +658,7 @@ CHECK_CANDIDATE(x, y - iDiamondSize, 4); } else { // what remains here is up or down CHECK_CANDIDATE(x + iDiamondSize, y, 2); - CHECK_CANDIDATE(x - iDiamondSize, y, 1); + CHECK_CANDIDATE(x - iDiamondSize, y, 1); } bDirection += iDirection; x = data->currentMV->x; y = data->currentMV->y; @@ -662,16 +682,14 @@ backupMV = *(data->currentQMV); else backupMV = *(data->currentMV); - CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); + CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); - CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); - CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); - - CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); - + CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); - CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); + CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); + CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); + CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); } static __inline int @@ -773,12 +791,12 @@ for (x = 0; x < mb_width; x++) { MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; - if (Data.rrv) pMB->sad16 = + if (Data.rrv) pMB->sad16 = sad32v_c(pCurrent->y + (x + y * pParam->edged_width) * 32, pRef->y + (x + y * pParam->edged_width) * 32, pParam->edged_width, pMB->sad8 ); - else pMB->sad16 = + else pMB->sad16 = sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, pRef->y + (x + y * pParam->edged_width) * 16, pParam->edged_width, pMB->sad8 ); @@ -795,7 +813,7 @@ if (!(current->global_flags & XVID_LUMIMASKING)) { pMB->dquant = NO_CHANGE; - pMB->quant = current->quant; + pMB->quant = current->quant; } else { if (pMB->dquant != NO_CHANGE) { quant += DQtab[pMB->dquant]; @@ -821,7 +839,7 @@ current->global_flags & XVID_INTER4V, pMB); /* final skip decision, a.k.a. "the vector you found, really that good?" */ - if (current->coding_type == P_VOP) { + if (current->coding_type == P_VOP) { if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) && ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) ) if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { @@ -892,7 +910,7 @@ if (pmv[i].x == pmv[j].x) { if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } - } else + } else if (pmv[i].y == pmv[j].y) { if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } @@ -901,7 +919,7 @@ return mask; } -static __inline void +static __inline void PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, int iHcount, const MACROBLOCK * const prevMB, int rrv) { @@ -911,7 +929,7 @@ if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour pmv[5].x = EVEN(pmv[3].x); - pmv[5].y = EVEN(pmv[3].y); + pmv[5].y = EVEN(pmv[3].y); } else pmv[5].x = pmv[5].y = 0; if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour @@ -933,7 +951,7 @@ if ((x < iWcount-1) && (y < iHcount-1)) { pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame - pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); + pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); } else pmv[6].x = pmv[6].y = 0; if (rrv) { @@ -996,7 +1014,7 @@ if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); else Data->predMV = pmv[0]; - i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode, 0, 0); + i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; Data->iMinSAD[2] = pMB->sad8[1]; @@ -1030,7 +1048,7 @@ inter4v = 0; } else { - MainSearchFunc * MainSearchPtr; + MainSearchFunc * MainSearchPtr; if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; @@ -1047,7 +1065,7 @@ if (Data->rrv) { startMV.x = RRV_MV_SCALEUP(startMV.x); startMV.y = RRV_MV_SCALEUP(startMV.y); - } else + } else if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); if (!(MVequal(startMV, backupMV))) { @@ -1082,7 +1100,7 @@ Data->currentQMV[i].y = 2 * Data->currentMV[i].y; } - if((!Data->rrv) && (pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { + if((Data->qpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { Data->qpel_precision = 1; get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, @@ -1126,8 +1144,9 @@ Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); } + if (!(inter4v) || - (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + + (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { // INTER MODE pMB->mode = MODE_INTER; @@ -1147,9 +1166,9 @@ pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; } } else { -// INTER4V MODE; all other things are already set in Search8 +// INTER4V MODE; all other things are already set in Search8 pMB->mode = MODE_INTER4V; - pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + + pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; } } @@ -1171,14 +1190,14 @@ if(pParam->m_quarterpel) { Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); - if (block != 0) i = d_mv_bits( Data->currentQMV->x - Data->predMV.x, - Data->currentQMV->y - Data->predMV.y, Data->iFcode, 0, 0); + if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, + Data->predMV, Data->iFcode, 0, 0); } else { Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); if (block != 0) { - if (block != 0) i = d_mv_bits( Data->currentMV->x - Data->predMV.x, - Data->currentMV->y - Data->predMV.y, Data->iFcode, 0, Data->rrv); + if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, + Data->predMV, Data->iFcode, 0, Data->rrv); } } @@ -1228,7 +1247,7 @@ } } - if(!Data->rrv && Data->qpel) { + if(Data->qpel) { if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && (MotionFlags & PMV_QUARTERPELREFINE8)) { Data->qpel_precision = 1; @@ -1284,7 +1303,7 @@ if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); - pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); + pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); } else pmv[3].x = pmv[3].y = 0; if (y != 0) { @@ -1326,7 +1345,7 @@ const int32_t iEdgedWidth = pParam->edged_width; - int i, iDirection, mask; + int i, iDirection = 255, mask; VECTOR pmv[7]; MainSearchFunc *MainSearchPtr; *Data->iMinSAD = MV_MAX_ERROR; @@ -1351,7 +1370,7 @@ CheckCandidate = CheckCandidate16no4v; // main loop. checking all predictions - for (i = 0; i < 8; i++) { + for (i = 0; i < 7; i++) { if (!(mask = make_mask(pmv, i)) ) continue; CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); } @@ -1362,11 +1381,11 @@ MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; - (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); + (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); SubpelRefine(Data); - if (Data->qpel) { + if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { Data->currentQMV->x = 2*Data->currentMV->x; Data->currentQMV->y = 2*Data->currentMV->y; Data->qpel_precision = 1; @@ -1376,7 +1395,7 @@ } // three bits are needed to code backward mode. four for forward -// we treat the bits just like they were vector's + if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; else *Data->iMinSAD += 3 * Data->lambda16; @@ -1386,27 +1405,26 @@ if (Data->qpel) { pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; - if (mode_current == MODE_FORWARD) + if (mode_current == MODE_FORWARD) pMB->qmvs[0] = *Data->currentQMV; - else + else pMB->b_qmvs[0] = *Data->currentQMV; } else { pMB->pmvs[0].x = Data->currentMV->x - predMV->x; pMB->pmvs[0].y = Data->currentMV->y - predMV->y; } - if (mode_current == MODE_FORWARD) - pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; - else - pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search - + if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; + else pMB->b_mvs[0] = *Data->currentMV; } - + if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; + else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search + } -static void +static void SkipDecisionB(const IMAGE * const pCur, const IMAGE * const f_Ref, - const IMAGE * const b_Ref, + const IMAGE * const b_Ref, MACROBLOCK * const pMB, const uint32_t quant, const uint32_t x, const uint32_t y, @@ -1534,7 +1552,6 @@ } } - if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; else CheckCandidate = CheckCandidateDirectno4v; @@ -1561,8 +1578,7 @@ *best_sad = *Data->iMinSAD; - if (b_mb->mode == MODE_INTER4V) - pMB->mode = MODE_DIRECT; + if (b_mb->mode == MODE_INTER4V || Data->qpel) pMB->mode = MODE_DIRECT; else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation pMB->pmvs[3] = *Data->currentMV; @@ -1624,7 +1640,7 @@ fData->qpel_precision = 0; memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data *fData->iMinSAD = 4096*256; - bData.currentMV ++; bData.currentQMV ++; + bData.currentMV++; bData.currentQMV++; fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; i = (x + y * fData->iEdgedWidth) * 16; @@ -1679,6 +1695,7 @@ } while (!(iDirection)); if (fData->qpel) { + if (*fData->iMinSAD > *best_sad + 500) return; CheckCandidate = CheckCandidateInt; fData->qpel_precision = bData.qpel_precision = 1; get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); @@ -1688,11 +1705,12 @@ fData->currentQMV[1].x = 2 * fData->currentMV[1].x; fData->currentQMV[1].y = 2 * fData->currentMV[1].y; SubpelRefine(fData); + if (*fData->iMinSAD > *best_sad + 300) return; fData->currentQMV[2] = fData->currentQMV[0]; SubpelRefine(&bData); } - *fData->iMinSAD += (2+2) * fData->lambda16; // two bits are needed to code interpolate mode. + *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. if (*fData->iMinSAD < *best_sad) { *best_sad = *fData->iMinSAD; @@ -1800,7 +1818,7 @@ &Data); if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } - + // forward search SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, &frame->image, i, j, @@ -1818,7 +1836,6 @@ MODE_BACKWARD, &Data); // interpolate search comes last, because it uses data from forward and backward as prediction - SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, b_ref->y, b_refH->y, b_refV->y, b_refHV->y, &frame->image, @@ -1851,7 +1868,7 @@ if (Data.qpel) { f_predMV = pMB->qmvs[0]; b_predMV = pMB->b_qmvs[0]; - } else { + } else { f_predMV = pMB->mvs[0]; b_predMV = pMB->b_mvs[0]; } @@ -1867,66 +1884,67 @@ free(qimage); } -static __inline int +static __inline void MEanalyzeMB ( const uint8_t * const pRef, const uint8_t * const pCur, const int x, const int y, const MBParam * const pParam, - const MACROBLOCK * const pMBs, - MACROBLOCK * const pMB, + MACROBLOCK * const pMBs, SearchData * const Data) { - int i = 255, mask; + int i, mask; VECTOR pmv[3]; - *(Data->iMinSAD) = MV_MAX_ERROR; + MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; + + for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; //median is only used as prediction. it doesn't have to be real if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; else if (x == 1) //left macroblock does not have any vector now Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median - else if (y == 1) // top macroblock don't have it's vector + else if (y == 1) // top macroblock doesn't have it's vector Data->predMV = (pMB - 1)->mvs[0]; // left instead of median else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, - pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0); + pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); Data->Cur = pCur + (x + y * pParam->edged_width) * 16; Data->Ref = pRef + (x + y * pParam->edged_width) * 16; - + pmv[1].x = EVEN(pMB->mvs[0].x); pmv[1].y = EVEN(pMB->mvs[0].y); pmv[2].x = EVEN(Data->predMV.x); pmv[2].y = EVEN(Data->predMV.y); pmv[0].x = pmv[0].y = 0; - CheckCandidate16no4vI(0, 0, 255, &i, Data); + CheckCandidate32I(0, 0, 255, &i, Data); -//early skip for 0,0 - if (*Data->iMinSAD < MAX_SAD00_FOR_SKIP * 4) { - pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; - pMB->mode = MODE_NOT_CODED; - return 0; - } + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { - if (!(mask = make_mask(pmv, 1))) - CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); - if (!(mask = make_mask(pmv, 2))) - CheckCandidate16no4vI(pmv[2].x, pmv[2].y, mask, &i, Data); + if (!(mask = make_mask(pmv, 1))) + CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); + if (!(mask = make_mask(pmv, 2))) + CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); - if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 6) // diamond only if needed - DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed + DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); - pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; - pMB->mode = MODE_INTER; - return *(Data->iMinSAD); + for (i = 0; i < 4; i++) { + MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1) * pParam->mb_width)]; + MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; + MB->mode = MODE_INTER; + MB->sad16 = Data->iMinSAD[i+1]; + } + } } -#define INTRA_THRESH 1350 -#define INTER_THRESH 1200 +#define INTRA_BIAS 2500 +#define INTRA_THRESH 1500 +#define INTER_THRESH 1400 int @@ -1942,15 +1960,18 @@ MACROBLOCK * const pMBs = Current->mbs; const IMAGE * const pCurrent = &Current->image; int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; + const VECTOR zeroMV = {0,0}; - VECTOR currentMV; - int32_t iMinSAD; + int32_t iMinSAD[5], temp[5]; + VECTOR currentMV[5]; SearchData Data; Data.iEdgedWidth = pParam->edged_width; - Data.currentMV = ¤tMV; - Data.iMinSAD = &iMinSAD; + Data.currentMV = currentMV; + Data.iMinSAD = iMinSAD; Data.iFcode = Current->fcode; - CheckCandidate = CheckCandidate16no4vI; + Data.rrv = Current->global_flags & XVID_REDUCED; + Data.temp = temp; + CheckCandidate = CheckCandidate32I; if (intraCount < 10) // we're right after an I frame IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); @@ -1958,32 +1979,36 @@ if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; - InterThresh += 400 * (1 - bCount); - if (InterThresh < 200) InterThresh = 200; + if (InterThresh < 300) InterThresh = 300; if (sadInit) (*sadInit) (); - for (y = 1; y < pParam->mb_height-1; y++) { - for (x = 1; x < pParam->mb_width-1; x++) { - int sad, dev; - MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; + for (y = 1; y < pParam->mb_height-1; y+=2) { + for (x = 1; x < pParam->mb_width-1; x+=2) { + int i; - sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, - pParam, pMBs, pMB, &Data); + if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; + + MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); - if (sad > IntraThresh) { - dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, - pParam->edged_width); - if (dev + IntraThresh < sad) { - pMB->mode = MODE_INTRA; - if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; + for (i = 0; i < 4; i++) { + int dev; + MACROBLOCK *pMB = &pMBs[x+(i&1) + y+(i>>1) * pParam->mb_width]; + if (pMB->sad16 > IntraThresh) { + dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1))* pParam->edged_width) * 16, + pParam->edged_width); + if (dev + IntraThresh < pMB->sad16) { + pMB->mode = MODE_INTRA; + if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; + } } + sSAD += pMB->sad16; } - sSAD += sad; } } sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); + if (sSAD > IntraThresh + INTRA_BIAS ) return I_VOP; if (sSAD > InterThresh ) return P_VOP; emms(); return B_VOP; @@ -1991,7 +2016,7 @@ } static void -CheckGMC(int x, int y, const int dir, int * iDirection, +CheckGMC(int x, int y, const int dir, int * iDirection, const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, const MBParam * const pParam) { @@ -2010,12 +2035,12 @@ if (count > *bestcount) { *bestcount = count; *iDirection = dir; - GMC->x = x; GMC->y = y; + GMC->x = x; GMC->y = y; } } -static __inline VECTOR +static VECTOR GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) { @@ -2041,7 +2066,7 @@ const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; VECTOR mv; - if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) + if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; mv = pMB->mvs[0]; @@ -2054,13 +2079,12 @@ max_x = gmc.x + step; min_y = gmc.y - step; max_y = gmc.y + step; - } - if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) + if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) gmc.x = gmc.y = 0; //no camara pan, no GMC -// step2: let's refine camera panning using gradiend-descent approach. +// step2: let's refine camera panning using gradiend-descent approach // TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) bestcount = 0; CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam);