--- branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/11/19 13:48:42 662 +++ branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/12/13 11:56:31 712 @@ -47,38 +47,16 @@ #define FINAL_SKIP_THRESH (50) #define MAX_SAD00_FOR_SKIP (20) #define MAX_CHROMA_SAD_FOR_SKIP (22) -#define SKIP_THRESH_B (25) #define CHECK_CANDIDATE(X,Y,D) { \ (*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } -#define GET_REFERENCE(X, Y, REF) { \ - switch ( (((X)&1)<<1) + ((Y)&1) ) \ - { \ - case 0 : REF = (uint8_t *)data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ - case 1 : REF = (uint8_t *)data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ - case 2 : REF = (uint8_t *)data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ - default : REF = (uint8_t *)data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ - } \ -} -// I hate those macros :/ -#define GET_REFERENCE2(X, Y, REF) { \ - switch ( (((X)&1)<<1) + ((Y)&1) ) \ - { \ - case 0 : REF = (uint8_t *)data->bRef + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ - case 1 : REF = (uint8_t *)data->bRefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ - case 2 : REF = (uint8_t *)data->bRefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ - default : REF = (uint8_t *)data->bRefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ - } \ -} - - -#define iDiamondSize 2 - static __inline int -d_mv_bits(int x, int y, const uint32_t iFcode) +d_mv_bits(int x, int y, 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); } if (x == 0) xb = 1; else { @@ -107,7 +85,9 @@ dx = (dx >> 1) + roundtab_79[dx & 0x3]; dy = (dy >> 1) + roundtab_79[dy & 0x3]; - switch (((dx & 1) << 1) + (dy & 1)) { // ((dx%2)?2:0)+((dy%2)?1:0) + if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently + + 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); @@ -134,223 +114,226 @@ sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); break; } + data->temp[5] = dx; data->temp[6] = dy; data->temp[7] = sad; //backup return sad; } +static __inline const uint8_t * +GetReference(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) ) { + 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); + case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); + case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); + case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); + case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); + default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); + } +} -/* CHECK_CANDIATE FUNCTIONS START */ +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 + uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; + const int32_t iEdgedWidth = data->iEdgedWidth; + const uint32_t rounding = data->rounding; + const int halfpel_x = x/2; + 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 += 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; + break; + case 1: // x halfpel, y qpel - top or bottom during qpel refinement + ref2 = GetReference(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 += 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 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); + break; + } + return Reference; +} + +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 + uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; + const int32_t iEdgedWidth = data->iEdgedWidth; + const uint32_t rounding = data->rounding; + const int halfpel_x = x/2; + 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 + switch( ((x&1)<<1) + (y&1) ) { + case 0: // pure halfpel position + return (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); + case 1: // x halfpel, y qpel - top or bottom during qpel refinement + ref2 = GetReference(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); + interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, 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); + 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); + interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, 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); + 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); + interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); + break; + } + return Reference; +} + +/* CHECK_CANDIATE FUNCTIONS START */ static void CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - int t; + int t, xc, yc; const uint8_t * Reference; + VECTOR * current; if (( x > data->max_dx) || ( x < data->min_dx) || ( y > data->max_dy) || (y < data->min_dy)) return; - switch ( ((x&1)<<1) + (y&1) ) { - case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; - case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; - case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; - default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; + if (data->qpel_precision) { // x and y are in 1/4 precision + Reference = Interpolate16x16qpel(x, y, 0, data); + xc = x/2; yc = y/2; //for chroma sad + current = data->currentQMV; + } else { + Reference = GetReference(x, y, 0, 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); data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); - if (data->qpel) t = d_mv_bits(2*x - data->predQMV.x, 2*y - data->predQMV.y, data->iFcode); - else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); - data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; - if (data->chroma) data->temp[0] += ChromaSAD(x, y, data); + if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data); if (data->temp[0] < data->iMinSAD[0]) { data->iMinSAD[0] = data->temp[0]; - data->currentMV[0].x = x; data->currentMV[0].y = y; + current[0].x = x; current[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; } + data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[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; } + data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[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; } + data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[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 -CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) -{ - int32_t sad; - const uint8_t * Reference; - - if (( x > data->max_dx) || ( x < data->min_dx) - || ( y > data->max_dy) || (y < data->min_dy)) return; + data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } - switch ( ((x&1)<<1) + (y&1) ) - { - case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; - case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; - case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; - default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; - } - - sad = sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); - if (data->qpel) //only to be used in b-frames' ME - sad += (data->lambda16 * d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode) * sad)/1000; - else - sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000; - - if (sad < *(data->iMinSAD)) { - *(data->iMinSAD) = sad; - data->currentMV[0].x = x; data->currentMV[0].y = y; - *dir = Direction; } } static void -CheckCandidate16_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) - -// CheckCandidate16 variant which expects x and y in quarter pixel resolution -// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) -// around currentMV! +CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { int t; - uint8_t * Reference = (uint8_t *)data->RefQ; - const uint8_t *ref1, *ref2, *ref3, *ref4; - VECTOR halfpelMV = *(data->currentMV); - - int32_t iEdgedWidth = data->iEdgedWidth; - uint32_t rounding = data->rounding; + const uint8_t * Reference; - if (( x > data->max_dx) || ( x < data->min_dx) + if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value + ( x > data->max_dx) || ( x < data->min_dx) || ( y > data->max_dy) || (y < data->min_dy)) return; - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases - switch( ((x&1)<<1) + (y&1) ) - { - case 0: // pure halfpel position - shouldn't happen during a refinement step - GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); - break; - - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); - interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); - break; + Reference = GetReference(x, y, 0, data); + t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode, 0, 1); - default: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement - GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); - - 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); - interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); - break; - } - - data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp+1); + data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); - t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; - if (data->chroma) - data->temp[0] += ChromaSAD(x/2, y/2, data); - if (data->temp[0] < data->iMinSAD[0]) { data->iMinSAD[0] = data->temp[0]; - data->currentQMV[0].x = x; data->currentQMV[0].y = y; - /* *dir = Direction;*/ } + 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->currentQMV[1].x = x; data->currentQMV[1].y = y; } + 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->currentQMV[2].x = x; data->currentQMV[2].y = y; } + 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->currentQMV[3].x = x; data->currentQMV[3].y = y; } + 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->currentQMV[4].x = x; data->currentQMV[4].y = y; } + data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } } static void -CheckCandidate16no4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) - -// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution -// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) -// around currentMV! -// this function is for B-frames' search only +CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - uint8_t * Reference = (uint8_t *)data->RefQ; - const uint8_t *ref1, *ref2, *ref3, *ref4; - VECTOR halfpelMV = *(data->currentMV); - - int32_t iEdgedWidth = data->iEdgedWidth; int32_t sad; + const uint8_t * Reference; + int t; + VECTOR * current; if (( x > data->max_dx) || ( x < data->min_dx) || ( y > data->max_dy) || (y < data->min_dy)) return; - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this refenrence is used in all cases - switch( ((x&1)<<1) + (y&1) ) - { - case 0: // pure halfpel position - shouldn't happen during a refinement step - GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); - break; + if (data->rrv) + if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) ) return; //non-zero integer value - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - default: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement - GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); - - interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, 0); - interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); - interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); - break; + if (data->qpel_precision) { // x and y are in 1/4 precision + Reference = Interpolate16x16qpel(x, y, 0, data); + current = data->currentQMV; + } else { + Reference = GetReference(x, y, 0, 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); sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); - sad += (data->lambda16 * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode) * sad)/1000; + sad += (data->lambda16 * t * sad)/1000; - if (sad < data->iMinSAD[0]) { - data->iMinSAD[0] = sad; - data->currentQMV[0].x = x; data->currentQMV[0].y = y; - /* *dir = Direction;*/ } + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + current->x = x; current->y = y; + *dir = Direction; } } static void @@ -376,147 +359,34 @@ CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) { int32_t sad; - const int xb = data->currentMV[1].x; - const int yb = data->currentMV[1].y; + int xb, yb, t; const uint8_t *ReferenceF, *ReferenceB; + VECTOR *current; if (( xf > data->max_dx) || ( xf < data->min_dx) || ( yf > data->max_dy) || (yf < data->min_dy)) return; - switch ( ((xf&1)<<1) + (yf&1) ) { - case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; - case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; - case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; - default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; - } - - switch ( ((xb&1)<<1) + (yb&1) ) { - case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; - case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; - case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; - default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; - } - - sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); - - if (data->qpel) - sad += (data->lambda16 * - ( d_mv_bits(2*xf - data->predMV.x, 2*yf - data->predMV.y, data->iFcode) + - d_mv_bits(2*xb - data->bpredMV.x, 2*yb - data->bpredMV.y, data->iFcode)) * sad)/1000; - else - sad += (data->lambda16 * - ( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + - d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000; - - if (sad < *(data->iMinSAD)) { - *(data->iMinSAD) = sad; - data->currentMV->x = xf; data->currentMV->y = yf; - *dir = Direction; } -} - - -static void -CheckCandidateInt_qpel(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) -{ -// CheckCandidateInt variant which expects x and y in quarter pixel resolution - - int32_t sad; - const int xb = data->currentQMV[1].x; - const int yb = data->currentQMV[1].y; - uint8_t * ReferenceF = (uint8_t *)data->RefQ; - uint8_t * ReferenceB = (uint8_t *)data->RefQ + 16; - const uint8_t *ref1, *ref2, *ref3, *ref4; - VECTOR halfpelMV; - const int32_t iEdgedWidth = data->iEdgedWidth; - - if (( xf > data->max_dx) || ( xf < data->min_dx) - || ( yf > data->max_dy) || (yf < data->min_dy)) return; - - halfpelMV.x = xf/2; //forward first - halfpelMV.y = yf/2; - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases - switch( ((xf&1)<<1) + (yf&1) ) - { - case 0: // pure halfpel position - shouldn't happen during a refinement step - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ReferenceF); - break; - - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE(halfpelMV.x, yf - halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE(xf - halfpelMV.x, halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - default: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement - GET_REFERENCE(halfpelMV.x, yf - halfpelMV.y, ref2); - GET_REFERENCE(xf - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE(xf - halfpelMV.x, yf - halfpelMV.y, ref4); - - interpolate8x8_avg4(ReferenceF, ref1, ref2, ref3, ref4, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceF+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - } - - halfpelMV.x = xb/2; //backward - halfpelMV.y = yb/2; - GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases - switch( ((xb&1)<<1) + (yb&1) ) - { - case 0: // pure halfpel position - shouldn't happen during a refinement step - GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ReferenceB); - break; - - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE2(halfpelMV.x, yb - halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE2(xb - halfpelMV.x, halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - default: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement - GET_REFERENCE2(halfpelMV.x, yb - halfpelMV.y, ref2); - GET_REFERENCE2(xb - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE2(xb - halfpelMV.x, yb - halfpelMV.y, ref4); - - interpolate8x8_avg4(ReferenceB, ref1, ref2, ref3, ref4, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceB+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); - break; + if (data->qpel_precision) { + ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); + xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; + current = data->currentQMV; + ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); + } else { + ReferenceF = GetReference(xf, yf, 0, data); + xb = data->currentMV[1].x; yb = data->currentMV[1].y; + ReferenceB = GetReference(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); + sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); - - sad += (data->lambda16 * - ( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + - d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode)) * sad)/1000; + sad += (data->lambda16 * t * sad)/1000; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; - data->currentQMV->x = xf; data->currentQMV->y = yf; + current->x = xf; current->y = yf; *dir = Direction; } } @@ -547,249 +417,20 @@ || ( 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; - switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { - case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; - case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; - case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; - default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; - } - - switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { - case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; - case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; - case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; - default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; + 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); + ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), + ReferenceF, ReferenceB, data->iEdgedWidth); if (sad > *(data->iMinSAD)) return; } - sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; - - if (sad < *(data->iMinSAD)) { - *(data->iMinSAD) = sad; - data->currentMV->x = x; data->currentMV->y = y; - *dir = Direction; } -} - - -static void -CheckCandidateDirect_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) -{ - int32_t sad = 0; - int k; - VECTOR mvs, b_mvs, halfpelMV; - const uint8_t *ref1, *ref2, *ref3, *ref4; - uint8_t *ReferenceF, *ReferenceB; - const uint32_t iEdgedWidth = data->iEdgedWidth; - - if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; - - for (k = 0; k < 4; k++) { - ReferenceF = (uint8_t *)data->RefQ; - ReferenceB = (uint8_t *)data->RefQ + 64; - - mvs.x = data->directmvF[k].x + x; - b_mvs.x = ((x == 0) ? - data->directmvB[k].x - : mvs.x - data->referencemv[k].x); - - mvs.y = data->directmvF[k].y + y; - b_mvs.y = ((y == 0) ? - data->directmvB[k].y - : mvs.y - data->referencemv[k].y); - - if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) - || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) - || ( 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; - - halfpelMV.x = mvs.x/2; //forward first - halfpelMV.y = mvs.y/2; - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases - switch( ((mvs.x&1)<<1) + (mvs.y&1) ) { - case 0: // pure halfpel position - GET_REFERENCE(halfpelMV.x + 16*(k&1), halfpelMV.y + 16*(k>>1), ReferenceF); - break; - - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceF, ref1+8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref2+ 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceF, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); - break; - - default: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement - GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); - GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE(mvs.x - halfpelMV.x, mvs.y - halfpelMV.y, ref4); - interpolate8x8_avg4(ReferenceF, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref3 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref4 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); - break; - } - - halfpelMV.x = b_mvs.x/2; - halfpelMV.y = b_mvs.y/2; - GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in most cases - switch( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { - case 0: // pure halfpel position - GET_REFERENCE2(halfpelMV.x + 16*(k&1), halfpelMV.y + 16*(k>>1), ReferenceB); - break; - - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceB, ref1+8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref2+ 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceB, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); - break; - - default: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement - GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); - GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE2(b_mvs.x - halfpelMV.x, b_mvs.y - halfpelMV.y, ref4); - interpolate8x8_avg4(ReferenceB, ref1 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref2 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref3 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ref4 + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), iEdgedWidth, 0); - break; - } - - sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), - ReferenceF, - ReferenceB, - data->iEdgedWidth); - if (sad > *(data->iMinSAD)) return; - } - - sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; - - if (sad < *(data->iMinSAD)) { - *(data->iMinSAD) = sad; - data->currentMV->x = x; data->currentMV->y = y; - *dir = Direction; } -} - -static void -CheckCandidateDirectno4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) -{ - int32_t sad = 0; - VECTOR mvs, b_mvs, halfpelMV; - const uint8_t *ref1, *ref2, *ref3, *ref4; - const uint32_t iEdgedWidth = data->iEdgedWidth; - uint8_t * ReferenceF = (uint8_t *)data->RefQ; - uint8_t * ReferenceB = (uint8_t *)data->RefQ + 64; - - if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; - - mvs.x = data->directmvF[0].x + x; - b_mvs.x = ((x == 0) ? - data->directmvB[0].x - : mvs.x - data->referencemv[0].x); - - mvs.y = data->directmvF[0].y + y; - b_mvs.y = ((y == 0) ? - data->directmvB[0].y - : mvs.y - data->referencemv[0].y); - - if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) - || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) - || ( 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; - - halfpelMV.x = mvs.x/2; //forward first - halfpelMV.y = mvs.y/2; - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); // this reference is used in all cases - switch( ((mvs.x&1)<<1) + (mvs.y&1) ) { - case 0: // pure halfpel position - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ReferenceF); - break; - - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceF, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - default: // x and y in qpel resolution - GET_REFERENCE(halfpelMV.x, mvs.y - halfpelMV.y, ref2); - GET_REFERENCE(mvs.x - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE(mvs.x - halfpelMV.x, mvs.y - halfpelMV.y, ref4); - - interpolate8x8_avg4(ReferenceF, ref1, ref2, ref3, ref4, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceF+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceF+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceF+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - } - - halfpelMV.x = b_mvs.x/2; //backward - halfpelMV.y = b_mvs.y/2; - GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ref1); - switch( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) - { - case 0: // pure halfpel position - GET_REFERENCE2(halfpelMV.x, halfpelMV.y, ReferenceB); - break; - - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref2); - interpolate8x8_avg2(ReferenceB, ref1, ref2, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8, ref1+8, ref2+8, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg2(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - - default: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement - GET_REFERENCE2(halfpelMV.x, b_mvs.y - halfpelMV.y, ref2); - GET_REFERENCE2(b_mvs.x - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE2(b_mvs.x - halfpelMV.x, b_mvs.y - halfpelMV.y, ref4); - - interpolate8x8_avg4(ReferenceB, ref1, ref2, ref3, ref4, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceB+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceB+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, 0); - interpolate8x8_avg4(ReferenceB+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, 0); - break; - } - - sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); - sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; + sad += (data->lambda16 * d_mv_bits(x, y, 1, 0, 0) * sad)/1000; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -797,7 +438,6 @@ *dir = Direction; } } - static void CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { @@ -823,22 +463,15 @@ || ( 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; - switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { - case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; - case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; - case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; - default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; + 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); - switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { - case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; - case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; - case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; - default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; - } - sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); - sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; + sad += (data->lambda16 * d_mv_bits(x, y, 1, 0, 0) * sad)/1000; if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -855,17 +488,11 @@ if (( x > data->max_dx) || ( x < data->min_dx) || ( y > data->max_dy) || (y < data->min_dy)) return; - switch ( ((x&1)<<1) + (y&1) ) - { - case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; - case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; - case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; - default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; - } + if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); + else Reference = GetReference(x, y, 0, data); sad = sad8(data->Cur, Reference, data->iEdgedWidth); - if (data->qpel) t = d_mv_bits(2 * x - data->predQMV.x, 2 * y - data->predQMV.y, data->iFcode); - else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode, data->qpel && !data->qpel_precision, 0); sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; @@ -875,62 +502,6 @@ *dir = Direction; } } -static void -CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) -// CheckCandidate8 variant which expects x and y in quarter pixel resolution -// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) -// around currentMV! - -{ - int32_t sad; - uint8_t *Reference = (uint8_t *) data->RefQ; - const uint8_t *ref1, *ref2, *ref3, *ref4; - VECTOR halfpelMV = *(data->currentMV); - - int32_t iEdgedWidth = data->iEdgedWidth; - uint32_t rounding = data->rounding; - - if (( x > data->max_dx) || ( x < data->min_dx) - || ( y > data->max_dy) || (y < data->min_dy)) return; - - GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); - switch( ((x&1)<<1) + (y&1) ) - { - case 0: // pure halfpel position - shouldn't happen during a refinement step - GET_REFERENCE(halfpelMV.x, halfpelMV.y, Reference); - break; - - case 1: // x halfpel, y qpel - top or bottom during qpel refinement - GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); - break; - - case 2: // x qpel, y halfpel - left or right during qpel refinement - GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); - - interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); - break; - - default: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement - GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); - GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); - - interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); - break; - } - - sad = sad8(data->Cur, Reference, data->iEdgedWidth); - sad += (data->lambda8 * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode) * (sad+NEIGH_8X8_BIAS))/100; - - if (sad < *(data->iMinSAD)) { - *(data->iMinSAD) = sad; - data->currentQMV->x = x; data->currentQMV->y = y; - *dir = Direction; } -} - /* CHECK_CANDIATE FUNCTIONS END */ /* MAINSEARCH FUNCTIONS START */ @@ -941,82 +512,83 @@ /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - int iDirection; + int iDirection; - do { + for(;;) { //forever + iDirection = 0; + if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); + if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); + if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); + if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); + + /* now we're doing diagonal checks near our candidate */ + + if (iDirection) { //checking if anything found + bDirection = iDirection; iDirection = 0; - if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); - if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); - if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); - if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); - - /* now we're doing diagonal checks near our candidate */ - - if (iDirection) { //checking if anything found - bDirection = iDirection; - iDirection = 0; - x = data->currentMV->x; y = data->currentMV->y; - if (bDirection & 3) { //our candidate is left or right - CHECK_CANDIDATE(x, y + iDiamondSize, 8); - 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); } - - if (iDirection) { - bDirection += iDirection; - x = data->currentMV->x; y = data->currentMV->y; } - } else { //about to quit, eh? not so fast.... - switch (bDirection) { - case 2: - CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); - CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); - break; - case 1: - CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); - CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); - break; - case 2 + 4: - CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); - CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); - CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); - break; - case 4: - CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); - CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); - break; - case 8: - CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); - CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); - break; - case 1 + 4: - CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); - CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); - CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); - break; - case 2 + 8: - CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); - CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); - CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); - break; - case 1 + 8: - CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); - CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); - CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); - break; - default: //1+2+4+8 == we didn't find anything at all - CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); - CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); - CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); - CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); - break; - } - if (!iDirection) break; //ok, the end. really - bDirection = iDirection; - x = data->currentMV->x; y = data->currentMV->y; + x = data->currentMV->x; y = data->currentMV->y; + if (bDirection & 3) { //our candidate is left or right + CHECK_CANDIDATE(x, y + iDiamondSize, 8); + 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); + } + + if (iDirection) { + bDirection += iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } + } else { //about to quit, eh? not so fast.... + switch (bDirection) { + case 2: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 1: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + case 2 + 4: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 4: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + break; + case 8: + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + case 1 + 4: + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + break; + case 2 + 8: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 1 + 8: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + default: //1+2+4+8 == we didn't find anything at all + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; } + if (!iDirection) break; //ok, the end. really + bDirection = iDirection; + x = data->currentMV->x; y = data->currentMV->y; } - while (1); //forever + } } static void @@ -1046,33 +618,33 @@ /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - int iDirection; + int iDirection; - do { - iDirection = 0; - if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); - if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); - if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); - if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); - - /* now we're doing diagonal checks near our candidate */ - - if (iDirection) { //checking if anything found - bDirection = iDirection; - iDirection = 0; - x = data->currentMV->x; y = data->currentMV->y; - if (bDirection & 3) { //our candidate is left or right - CHECK_CANDIDATE(x, y + iDiamondSize, 8); - 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); } + do { + iDirection = 0; + if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); + if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); + if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); + if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); - bDirection += iDirection; - x = data->currentMV->x; y = data->currentMV->y; + /* now we're doing diagonal checks near our candidate */ + + if (iDirection) { //checking if anything found + bDirection = iDirection; + iDirection = 0; + x = data->currentMV->x; y = data->currentMV->y; + if (bDirection & 3) { //our candidate is left or right + CHECK_CANDIDATE(x, y + iDiamondSize, 8); + 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); } + bDirection += iDirection; + x = data->currentMV->x; y = data->currentMV->y; } - while (iDirection); + } + while (iDirection); } /* MAINSEARCH FUNCTIONS END */ @@ -1080,33 +652,15 @@ /* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ static void -HalfpelRefine(const SearchData * const data) +SubpelRefine(const SearchData * const data) { -/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ - - VECTOR backupMV = *(data->currentMV); +/* Do a half-pel or q-pel refinement */ + VECTOR backupMV; int iDirection; //not needed - 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 + 1, 0); - - CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); - CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); - - CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); - CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); -} - - -static void -QuarterpelRefine(const SearchData * const data) -{ -/* Perform quarter pixel refinement*/ - - VECTOR backupMV = *(data->currentQMV); - int iDirection; //not needed + if (data->qpel_precision) + backupMV = *(data->currentQMV); + else backupMV = *(data->currentMV); CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); @@ -1118,27 +672,34 @@ CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); - } static __inline int SkipDecisionP(const IMAGE * current, const IMAGE * reference, const int x, const int y, - const uint32_t iEdgedWidth, const uint32_t iQuant) + const uint32_t iEdgedWidth, const uint32_t iQuant, int rrv) { /* keep repeating checks for all b-frames before this P frame, to make sure that SKIP is possible (todo) how: if skip is not possible set sad00 to a very high value */ - - uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, - reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); - if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; - sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, - reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); - if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; - - return 1; + if(rrv) { + uint32_t sadC = sad16(current->u + x*16 + y*(iEdgedWidth/2)*16, + reference->u + x*16 + y*(iEdgedWidth/2)*16, iEdgedWidth/2, 256*4096); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; + sadC += sad16(current->v + (x + y*(iEdgedWidth/2))*16, + reference->v + (x + y*(iEdgedWidth/2))*16, iEdgedWidth/2, 256*4096); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; + return 1; + } else { + uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, + reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; + sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8, + reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; + return 1; + } } static __inline void @@ -1169,17 +730,21 @@ const VECTOR zeroMV = { 0, 0 }; + uint32_t mb_width = pParam->mb_width; + uint32_t mb_height = pParam->mb_height; + uint32_t x, y; uint32_t iIntra = 0; int32_t InterBias, quant = current->quant, sad00; uint8_t *qimage; // some pre-initialized thingies for SearchP - int32_t temp[5]; + int32_t temp[8]; VECTOR currentMV[5]; VECTOR currentQMV[5]; int32_t iMinSAD[5]; SearchData Data; + memset(&Data, 0, sizeof(SearchData)); Data.iEdgedWidth = pParam->edged_width; Data.currentMV = currentMV; Data.currentQMV = currentQMV; @@ -1189,6 +754,13 @@ Data.rounding = pParam->m_rounding_type; Data.qpel = pParam->m_quarterpel; Data.chroma = current->global_flags & XVID_ME_COLOUR; + Data.rrv = current->global_flags & XVID_REDUCED; + + if ((current->global_flags & XVID_REDUCED)) { + mb_width = (pParam->width + 31) / 32; + mb_height = (pParam->height + 31) / 32; + Data.qpel = Data.chroma = 0; + } if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) return 1; // allocate some mem for qpel interpolated blocks @@ -1197,12 +769,17 @@ Data.RefQ = qimage; if (sadInit) (*sadInit) (); - for (y = 0; y < pParam->mb_height; y++) { - for (x = 0; x < pParam->mb_width; x++) { + for (y = 0; y < mb_height; y++) { + for (x = 0; x < mb_width; x++) { MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; - pMB->sad16 - = sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, + 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 = + sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, pRef->y + (x + y * pParam->edged_width) * 16, pParam->edged_width, pMB->sad8 ); @@ -1231,8 +808,8 @@ //initial skip decision /* no early skip for GMC (global vector = skip vector is unknown!) */ if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ - if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH) - if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { + if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) + if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { SkipMacroblockP(pMB, sad00); continue; } @@ -1246,8 +823,8 @@ /* final skip decision, a.k.a. "the vector you found, really that good?" */ 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) ) - if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) { + && ((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)) { SkipMacroblockP(pMB, sad00); continue; } @@ -1263,13 +840,24 @@ if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; if (Data.chroma) InterBias += 50; // to compensate bigger SAD + if (Data.rrv) InterBias *= 4; //?? if (InterBias < pMB->sad16) { - const int32_t deviation = - dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, + int32_t deviation; + if (Data.rrv) { + deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 32, + pParam->edged_width) + + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16, + pParam->edged_width) + + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * pParam->edged_width, + pParam->edged_width) + + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * (pParam->edged_width+1), + pParam->edged_width); + } else + deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, pParam->edged_width); - if (deviation < (pMB->sad16 - InterBias)) { + if (deviation < (pMB->sad16 - InterBias)) { if (++iIntra >= iLimit) { free(qimage); return 1; } pMB->mode = MODE_INTRA; pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = @@ -1314,13 +902,14 @@ } static __inline void -PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, - const int iHcount, const MACROBLOCK * const prevMB) +PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, + int iHcount, const MACROBLOCK * const prevMB, int rrv) { //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself + if (rrv) { iWcount /= 2; iHcount /= 2; } - if ( (y != 0) && (x != (iWcount-1)) ) { // [5] top-right neighbour + if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour pmv[5].x = EVEN(pmv[3].x); pmv[5].y = EVEN(pmv[3].y); } else pmv[5].x = pmv[5].y = 0; @@ -1332,17 +921,28 @@ else pmv[4].x = pmv[4].y = 0; // [1] median prediction - pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); + if (rrv) { //median is in halfzero-precision + pmv[1].x = RRV_MV_SCALEUP(pmv[0].x); + pmv[1].y = RRV_MV_SCALEUP(pmv[0].y); + } else { pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); } pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame pmv[2].y = EVEN(prevMB->mvs[0].y); - if ((x != iWcount-1) && (y != iHcount-1)) { + 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); } else pmv[6].x = pmv[6].y = 0; + + if (rrv) { + int i; + for (i = 0; i < 7; i++) { + pmv[i].x = RRV_MV_SCALEDOWN(pmv[i].x); + pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // a trick + } + } } static void @@ -1366,42 +966,37 @@ int i, iDirection = 255, mask, threshA; VECTOR pmv[7]; - get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() 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); + pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); - Data->predMV = pmv[0]; + get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() - Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; - Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; - Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; - - Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; - Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; - Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; - Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; - Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; - Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; + Data->temp[5] = Data->temp[7] = 256*4096; // to reset chroma-sad cache + if (Data->rrv) i = 2; else i = 1; + Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; + Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; + Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; + + Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; + Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; Data->lambda16 = lambda_vec16[iQuant]; Data->lambda8 = lambda_vec8[iQuant]; + Data->qpel_precision = 0; - if (!(MotionFlags & PMV_HALFPEL16)) { - Data->min_dx = EVEN(Data->min_dx); - Data->max_dx = EVEN(Data->max_dx); - Data->min_dy = EVEN(Data->min_dy); - Data->max_dy = EVEN(Data->max_dy); } - if (pMB->dquant != NO_CHANGE) inter4v = 0; for(i = 0; i < 5; i++) Data->currentMV[i].x = Data->currentMV[i].y = 0; - if (pParam->m_quarterpel) { - Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); - i = d_mv_bits(Data->predQMV.x, Data->predQMV.y, Data->iFcode); - } else i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); + 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); 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]; @@ -1415,10 +1010,11 @@ if (threshA > 1024) threshA = 1024; } PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, - prevMBs + x + y * pParam->mb_width); + prevMBs + x + y * pParam->mb_width, Data->rrv); - if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16; - else CheckCandidate = CheckCandidate16no4v; + if (Data->rrv) CheckCandidate = CheckCandidate32; + else if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; //for extra speed /* main loop. checking all predictions */ @@ -1434,7 +1030,7 @@ inter4v = 0; } else { - MainSearchFunc * MainSearchPtr; + MainSearchFunc * MainSearchPtr; if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; @@ -1448,8 +1044,12 @@ if (MotionFlags & PMV_EXTSEARCH16) { int32_t bSAD; VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; - if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? - startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); + if (Data->rrv) { + startMV.x = RRV_MV_SCALEUP(startMV.x); + startMV.y = RRV_MV_SCALEUP(startMV.y); + } 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))) { bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; @@ -1461,8 +1061,8 @@ } backupMV = Data->currentMV[0]; - if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; - else startMV.x = startMV.y = 0; + if (!MotionFlags & PMV_HALFPELREFINE16 || Data->rrv) startMV.x = startMV.y = 0; + else startMV.x = startMV.y = 1; if (!(MVequal(startMV, backupMV))) { bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; @@ -1475,20 +1075,20 @@ } } - if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); + if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); for(i = 0; i < 5; i++) { Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors Data->currentQMV[i].y = 2 * Data->currentMV[i].y; } - if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { + if((!Data->rrv) && (pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { - CheckCandidate = CheckCandidate16_qpel; + Data->qpel_precision = 1; get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, - pParam->width, pParam->height, Data->iFcode, 0); + pParam->width, pParam->height, Data->iFcode, 1, 0); - QuarterpelRefine(Data); + SubpelRefine(Data); } if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; @@ -1499,28 +1099,33 @@ Data8.iEdgedWidth = Data->iEdgedWidth; Data8.RefQ = Data->RefQ; Data8.qpel = Data->qpel; + Data8.rrv = Data->rrv; Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); if (Data->chroma) { - int sum, dx, dy; - - if(pParam->m_quarterpel) { - sum = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; - } else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; - dy = (sum >> 3) + roundtab_76[sum & 0xf]; + int sumx, sumy, dx, dy; if(pParam->m_quarterpel) { - sum = pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; - } else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; - dx = (sum >> 3) + roundtab_76[sum & 0xf]; + sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; + sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; + } else { + sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; + sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; + } + dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; + dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); } } + if (Data->rrv) { + 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[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { @@ -1529,15 +1134,14 @@ pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; - pMB->qmvs[0] = pMB->qmvs[1] - = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; - pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; if(pParam->m_quarterpel) { - pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; - pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; + pMB->qmvs[0] = pMB->qmvs[1] + = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; + pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; + pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; } else { pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; @@ -1560,36 +1164,42 @@ const int block, SearchData * const Data) { + int i = 0; Data->iMinSAD = OldData->iMinSAD + 1 + block; Data->currentMV = OldData->currentMV + 1 + block; Data->currentQMV = OldData->currentQMV + 1 + block; if(pParam->m_quarterpel) { - Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); - if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * - d_mv_bits( Data->currentQMV->x - Data->predQMV.x, - Data->currentQMV->y - Data->predQMV.y, - Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; + 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); + } else { - Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); - if (block != 0) *(Data->iMinSAD) += (Data->lambda8 * - d_mv_bits( Data->currentMV->x - Data->predMV.x, - Data->currentMV->y - Data->predMV.y, - Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; + 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); + } } + *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; + if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { + if (Data->rrv) i = 2; else i = 1; - Data->Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); - Data->RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); - Data->RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); - Data->RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->Ref = OldData->Ref + i*8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->RefH = OldData->RefH + i*8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->RefV = OldData->RefV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->RefHV = OldData->RefHV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); + + Data->Cur = OldData->Cur + i*8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->qpel_precision = 0; - Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); - get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, - pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); - CheckCandidate = CheckCandidate8; + pParam->width, pParam->height, OldData->iFcode - Data->qpel, 0, Data->rrv); + + if (Data->rrv) CheckCandidate = CheckCandidate16no4v; + else CheckCandidate = CheckCandidate8; if (MotionFlags & PMV_EXTSEARCH8) { int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD @@ -1610,7 +1220,7 @@ if (MotionFlags & PMV_HALFPELREFINE8) { int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD - HalfpelRefine(Data); // perform halfpel refine of current best vector + SubpelRefine(Data); // perform halfpel refine of current best vector if(*(Data->iMinSAD) < temp_sad) { // we have found a better match Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector @@ -1618,29 +1228,32 @@ } } - if(pParam->m_quarterpel) { + if(!Data->rrv && Data->qpel) { if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && (MotionFlags & PMV_QUARTERPELREFINE8)) { - CheckCandidate = CheckCandidate8_qpel; + Data->qpel_precision = 1; get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, - pParam->width, pParam->height, OldData->iFcode, pParam->m_quarterpel); - QuarterpelRefine(Data); + pParam->width, pParam->height, OldData->iFcode, 1, 0); + SubpelRefine(Data); } } } - if(pParam->m_quarterpel) { - pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; - pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; + if (Data->rrv) { + Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); + Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); } - else { + + if(Data->qpel) { + pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; + pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; + pMB->qmvs[block] = *(Data->currentQMV); + } else { pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; } pMB->mvs[block] = *(Data->currentMV); - pMB->qmvs[block] = *(Data->currentQMV); - pMB->sad8[block] = 4 * (*Data->iMinSAD); } @@ -1718,6 +1331,7 @@ MainSearchFunc *MainSearchPtr; *Data->iMinSAD = MV_MAX_ERROR; Data->iFcode = iFcode; + Data->qpel_precision = 0; Data->Ref = pRef + (x + y * iEdgedWidth) * 16; Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; @@ -1727,14 +1341,13 @@ Data->predMV = *predMV; get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, - pParam->width, pParam->height, iFcode, pParam->m_quarterpel); + pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); pmv[0] = Data->predMV; if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); Data->currentMV->x = Data->currentMV->y = 0; - CheckCandidate = CheckCandidate16no4v; // main loop. checking all predictions @@ -1751,15 +1364,15 @@ (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); - HalfpelRefine(Data); + SubpelRefine(Data); if (Data->qpel) { Data->currentQMV->x = 2*Data->currentMV->x; Data->currentQMV->y = 2*Data->currentMV->y; - CheckCandidate = CheckCandidate16no4v_qpel; + Data->qpel_precision = 1; get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, - pParam->width, pParam->height, iFcode, pParam->m_quarterpel); - QuarterpelRefine(Data); + pParam->width, pParam->height, iFcode, 1, 0); + SubpelRefine(Data); } // three bits are needed to code backward mode. four for forward @@ -1790,7 +1403,66 @@ } -static int32_t +static void +SkipDecisionB(const IMAGE * const pCur, + const IMAGE * const f_Ref, + const IMAGE * const b_Ref, + MACROBLOCK * const pMB, + const uint32_t quant, + const uint32_t x, const uint32_t y, + const SearchData * const Data) +{ + int dx, dy, b_dx, b_dy; + uint32_t sum; +//this is not full chroma compensation, only it's fullpel approximation. should work though + if (Data->qpel) { + dy = Data->directmvF[0].y/2 + Data->directmvF[1].y/2 + + Data->directmvF[2].y/2 + Data->directmvF[3].y/2; + + dx = Data->directmvF[0].x/2 + Data->directmvF[1].x/2 + + Data->directmvF[2].x/2 + Data->directmvF[3].x/2; + + b_dy = Data->directmvB[0].y/2 + Data->directmvB[1].y/2 + + Data->directmvB[2].y/2 + Data->directmvB[3].y/2; + + b_dx = Data->directmvB[0].x/2 + Data->directmvB[1].x/2 + + Data->directmvB[2].x/2 + Data->directmvB[3].x/2; + + } else { + dy = Data->directmvF[0].y + Data->directmvF[1].y + + Data->directmvF[2].y + Data->directmvF[3].y; + + dx = Data->directmvF[0].x + Data->directmvF[1].x + + Data->directmvF[2].x + Data->directmvF[3].x; + + b_dy = Data->directmvB[0].y + Data->directmvB[1].y + + Data->directmvB[2].y + Data->directmvB[3].y; + + b_dx = Data->directmvB[0].x + Data->directmvB[1].x + + Data->directmvB[2].x + Data->directmvB[3].x; + } + + + dy = (dy >> 3) + roundtab_76[dy & 0xf]; + dx = (dx >> 3) + roundtab_76[dx & 0xf]; + b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; + b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; + + sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), + f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, + b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, + Data->iEdgedWidth/2); + sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), + f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, + b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, + Data->iEdgedWidth/2); + + if (sum < 2*MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped +} + + + +static __inline uint32_t SearchDirect(const IMAGE * const f_Ref, const uint8_t * const f_RefH, const uint8_t * const f_RefV, @@ -1837,6 +1509,7 @@ Data->min_dy *= 2; Data->referencemv = b_mb->qmvs; } else Data->referencemv = b_mb->mvs; + Data->qpel_precision = 0; // it's a trick. it's 1 not 0, but we need 0 here for (k = 0; k < 4; k++) { pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); @@ -1850,7 +1523,7 @@ *best_sad = 256*4096; // in that case, we won't use direct mode pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; - return 0; + return 256*4096; } if (b_mb->mode != MODE_INTER4V) { pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; @@ -1861,58 +1534,16 @@ } } - if (Data->qpel) { - if (b_mb->mode == MODE_INTER4V) - CheckCandidate = CheckCandidateDirect_qpel; - else CheckCandidate = CheckCandidateDirectno4v_qpel; - } else { - if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; - else CheckCandidate = CheckCandidateDirectno4v; - } + + if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; + else CheckCandidate = CheckCandidateDirectno4v; (*CheckCandidate)(0, 0, 255, &k, Data); -// skip decision - if (*Data->iMinSAD < pMB->quant * SKIP_THRESH_B) { - //possible skip - checking chroma. everything copied from MC - //this is not full chroma compensation, only it's fullpel approximation. should work though - int sum, dx, dy, b_dx, b_dy; - - if (Data->qpel) { - sum = pMB->mvs[0].y/2 + pMB->mvs[1].y/2 + pMB->mvs[2].y/2 + pMB->mvs[3].y/2; - dy = (sum >> 3) + roundtab_76[sum & 0xf]; - sum = pMB->mvs[0].x/2 + pMB->mvs[1].x/2 + pMB->mvs[2].x/2 + pMB->mvs[3].x/2; - dx = (sum >> 3) + roundtab_76[sum & 0xf]; - - sum = pMB->b_mvs[0].y/2 + pMB->b_mvs[1].y/2 + pMB->b_mvs[2].y/2 + pMB->b_mvs[3].y/2; - b_dy = (sum >> 3) + roundtab_76[sum & 0xf]; - sum = pMB->b_mvs[0].x/2 + pMB->b_mvs[1].x/2 + pMB->b_mvs[2].x/2 + pMB->b_mvs[3].x/2; - b_dx = (sum >> 3) + roundtab_76[sum & 0xf]; - - } else { - sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; - dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); - sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; - dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); - - sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; - b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); - sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; - b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); - } - sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), - f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, - b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, - Data->iEdgedWidth/2); - sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), - f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, - b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, - Data->iEdgedWidth/2); - - if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { - pMB->mode = MODE_DIRECT_NONE_MV; - return *Data->iMinSAD; - } +// initial (fast) skip decision + if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH*2) { + SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data->chroma, Data); //possible skip - checking chroma + if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. } skip_sad = *Data->iMinSAD; @@ -1926,9 +1557,8 @@ (*MainSearchPtr)(0, 0, Data, 255); - HalfpelRefine(Data); //or qpel refine, if we're in qpel mode + SubpelRefine(Data); - *Data->iMinSAD += 1 * Data->lambda16; // one bit is needed to code direct mode *best_sad = *Data->iMinSAD; if (b_mb->mode == MODE_INTER4V) @@ -1965,7 +1595,7 @@ } -static __inline void +static void SearchInterpolate(const uint8_t * const f_Ref, const uint8_t * const f_RefH, const uint8_t * const f_RefV, @@ -1988,43 +1618,41 @@ { - const int32_t iEdgedWidth = pParam->edged_width; int iDirection, i, j; SearchData bData; - *(bData.iMinSAD = fData->iMinSAD) = 4096*256; - bData.Cur = fData->Cur; - fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; - bData.currentMV = fData->currentMV + 1; bData.currentQMV = fData->currentQMV + 1; - bData.lambda16 = fData->lambda16; + fData->qpel_precision = 0; + memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data + *fData->iMinSAD = 4096*256; + bData.currentMV ++; bData.currentQMV ++; fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; - bData.bRef = fData->Ref = f_Ref + (x + y * iEdgedWidth) * 16; - bData.bRefH = fData->RefH = f_RefH + (x + y * iEdgedWidth) * 16; - bData.bRefV = fData->RefV = f_RefV + (x + y * iEdgedWidth) * 16; - bData.bRefHV = fData->RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; - bData.Ref = fData->bRef = b_Ref + (x + y * iEdgedWidth) * 16; - bData.RefH = fData->bRefH = b_RefH + (x + y * iEdgedWidth) * 16; - bData.RefV = fData->bRefV = b_RefV + (x + y * iEdgedWidth) * 16; - bData.RefHV = fData->bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; - bData.RefQ = fData->RefQ; + i = (x + y * fData->iEdgedWidth) * 16; + bData.bRef = fData->Ref = f_Ref + i; + bData.bRefH = fData->RefH = f_RefH + i; + bData.bRefV = fData->RefV = f_RefV + i; + bData.bRefHV = fData->RefHV = f_RefHV + i; + bData.Ref = fData->bRef = b_Ref + i; + bData.RefH = fData->bRefH = b_RefH + i; + bData.RefV = fData->bRefV = b_RefV + i; + bData.RefHV = fData->bRefHV = b_RefHV + i; bData.bpredMV = fData->predMV = *f_predMV; fData->bpredMV = bData.predMV = *b_predMV; - fData->currentMV[0] = fData->currentMV[2]; - get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, pParam->m_quarterpel); - get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel); + + get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); + get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; - if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy; - if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx; - if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy; + if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx; + if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; + if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy; if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; - if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy; - if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx; - if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy; + if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx; + if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; + if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy; CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); @@ -2050,20 +1678,21 @@ } while (!(iDirection)); - *fData->iMinSAD += 2 * fData->lambda16; // two bits are needed to code interpolate mode. - if (fData->qpel) { - CheckCandidate = CheckCandidateInt_qpel; - get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 0); - get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 0); + 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); + get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; fData->currentQMV[1].x = 2 * fData->currentMV[1].x; fData->currentQMV[1].y = 2 * fData->currentMV[1].y; -// QuarterpelRefine(fData); + SubpelRefine(fData); fData->currentQMV[2] = fData->currentQMV[0]; -// QuarterpelRefine(&bData); + SubpelRefine(&bData); } + + *fData->iMinSAD += (2+2) * fData->lambda16; // two bits are needed to code interpolate mode. if (*fData->iMinSAD < *best_sad) { *best_sad = *fData->iMinSAD; @@ -2105,7 +1734,8 @@ const IMAGE * const b_refHV) { uint32_t i, j; - int32_t best_sad, skip_sad; + int32_t best_sad; + uint32_t skip_sad; int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; static const VECTOR zeroMV={0,0}; const MACROBLOCK * const b_mbs = b_reference->mbs; @@ -2122,11 +1752,14 @@ int32_t iMinSAD; VECTOR currentMV[3]; VECTOR currentQMV[3]; + memset(&Data, 0, sizeof(SearchData)); Data.iEdgedWidth = pParam->edged_width; Data.currentMV = currentMV; Data.currentQMV = currentQMV; Data.iMinSAD = &iMinSAD; Data.lambda16 = lambda_vec16[frame->quant]; + Data.chroma = frame->quant; Data.qpel = pParam->m_quarterpel; + Data.rounding = 0; if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) return; // allocate some mem for qpel interpolated blocks @@ -2197,20 +1830,25 @@ pMB, &best_sad, &Data); +// final skip decision + if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2) + && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) + SkipDecisionB(&frame->image, f_ref, b_ref, pMB,frame->quant, i, j, &Data); + switch (pMB->mode) { case MODE_FORWARD: f_count++; - if (pParam->m_quarterpel) f_predMV = pMB->qmvs[0]; + if (Data.qpel) f_predMV = pMB->qmvs[0]; else f_predMV = pMB->mvs[0]; break; case MODE_BACKWARD: b_count++; - if (pParam->m_quarterpel) b_predMV = pMB->b_qmvs[0]; + if (Data.qpel) b_predMV = pMB->b_qmvs[0]; else b_predMV = pMB->b_mvs[0]; break; case MODE_INTERPOLATE: i_count++; - if (pParam->m_quarterpel) { + if (Data.qpel) { f_predMV = pMB->qmvs[0]; b_predMV = pMB->b_qmvs[0]; } else { @@ -2221,7 +1859,6 @@ case MODE_DIRECT: case MODE_DIRECT_NO4V: d_count++; - break; default: break; } @@ -2230,224 +1867,6 @@ free(qimage); } -/* Hinted ME starts here */ - -static void -SearchPhinted ( const IMAGE * 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 MBParam * const pParam, - const MACROBLOCK * const pMBs, - int inter4v, - MACROBLOCK * const pMB, - SearchData * const Data) -{ - - int i, t; - MainSearchFunc * MainSearchPtr; - - Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); - Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); - 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); - - Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; - Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8; - Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8; - - Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16; - Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; - Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; - Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; - Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; - Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; - - if (!(MotionFlags & PMV_HALFPEL16)) { - Data->min_dx = EVEN(Data->min_dx); - Data->max_dx = EVEN(Data->max_dx); - Data->min_dy = EVEN(Data->min_dy); - Data->max_dy = EVEN(Data->max_dy); - } - - for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; - - if (pMB->dquant != NO_CHANGE) inter4v = 0; - - if (inter4v || pParam->m_quarterpel || Data->chroma) CheckCandidate = CheckCandidate16; - else CheckCandidate = CheckCandidate16no4v; - - pMB->mvs[0].x = EVEN(pMB->mvs[0].x); - pMB->mvs[0].y = EVEN(pMB->mvs[0].y); - if (pMB->mvs[0].x > Data->max_dx) pMB->mvs[0].x = Data->max_dx; // this is in case iFcode changed - if (pMB->mvs[0].x < Data->min_dx) pMB->mvs[0].x = Data->min_dx; - if (pMB->mvs[0].y > Data->max_dy) pMB->mvs[0].y = Data->max_dy; - if (pMB->mvs[0].y < Data->min_dy) pMB->mvs[0].y = Data->min_dy; - - (*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); - - if (pMB->mode == MODE_INTER4V) - for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search - pMB->mvs[i].x = EVEN(pMB->mvs[i].x); - pMB->mvs[i].y = EVEN(pMB->mvs[i].y); - if (!(make_mask(pMB->mvs, i))) - (*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); - } - - if (MotionFlags & PMV_USESQUARES16) - MainSearchPtr = SquareSearch; - else if (MotionFlags & PMV_ADVANCEDDIAMOND16) - MainSearchPtr = AdvDiamondSearch; - else MainSearchPtr = DiamondSearch; - - (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); - - if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); - - for(i = 0; i < 5; i++) { - Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors - Data->currentQMV[i].y = 2 * Data->currentMV[i].y; - } - - if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { - get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, - pParam->width, pParam->height, Data->iFcode, 0); - CheckCandidate = CheckCandidate16_qpel; - QuarterpelRefine(Data); - } - - if (inter4v) { - SearchData Data8; - Data8.iFcode = Data->iFcode; - Data8.lambda8 = Data->lambda8; - Data8.iEdgedWidth = Data->iEdgedWidth; - Data8.RefQ = Data->RefQ; - Data8.qpel = Data->qpel; - Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); - Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); - Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); - Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); - - if (Data->chroma) { - int sum, dx, dy; - - if(pParam->m_quarterpel) - sum = (pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2); - else sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; - dy = (sum ? SIGN(sum) * - (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0); - - if(pParam->m_quarterpel) - sum = (pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2); - else sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; - dx = (sum ? SIGN(sum) * - (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2) : 0); - Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); - } - } - - if (!(inter4v) || - (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; - pMB->mvs[0] = pMB->mvs[1] - = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; - - pMB->qmvs[0] = pMB->qmvs[1] - = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; - - pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = - pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; - - if(pParam->m_quarterpel) { - pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; - pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; - } else { - pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; - pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; - } - } else { -// INTER4V MODE; all other things are already set in Search8 - pMB->mode = MODE_INTER4V; - pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] - + Data->iMinSAD[4] + IMV16X16 * iQuant; - } - -} - -void -MotionEstimationHinted( MBParam * const pParam, - FRAMEINFO * const current, - FRAMEINFO * const reference, - const IMAGE * const pRefH, - const IMAGE * const pRefV, - const IMAGE * const pRefHV) -{ - MACROBLOCK *const pMBs = current->mbs; - const IMAGE *const pCurrent = ¤t->image; - const IMAGE *const pRef = &reference->image; - - uint32_t x, y; - uint8_t * qimage; - int32_t temp[5], quant = current->quant; - int32_t iMinSAD[5]; - VECTOR currentMV[5], currentQMV[5]; - SearchData Data; - Data.iEdgedWidth = pParam->edged_width; - Data.currentMV = currentMV; - Data.currentQMV = currentQMV; - Data.iMinSAD = iMinSAD; - Data.temp = temp; - Data.iFcode = current->fcode; - Data.rounding = pParam->m_rounding_type; - Data.qpel = pParam->m_quarterpel; - Data.chroma = current->global_flags & XVID_ME_COLOUR; - - if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) - return; // allocate some mem for qpel interpolated blocks - // somehow this is dirty since I think we shouldn't use malloc outside - // encoder_create() - so please fix me! - - Data.RefQ = qimage; - - if (sadInit) (*sadInit) (); - - for (y = 0; y < pParam->mb_height; y++) { - for (x = 0; x < pParam->mb_width; x++) { - - MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; - -//intra mode is copied from the first pass. At least for the time being - if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; - - if (!(current->global_flags & XVID_LUMIMASKING)) { - pMB->dquant = NO_CHANGE; - pMB->quant = current->quant; } - else { - if (pMB->dquant != NO_CHANGE) { - quant += DQtab[pMB->dquant]; - if (quant > 31) quant = 31; - else if (quant < 1) quant = 1; - } - pMB->quant = quant; - } - - SearchPhinted(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, - y, current->motion_flags, pMB->quant, - pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, - &Data); - - } - } - free(qimage); -} - static __inline int MEanalyzeMB ( const uint8_t * const pRef, const uint8_t * const pCur, @@ -2473,7 +1892,7 @@ 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); + pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0); Data->Cur = pCur + (x + y * pParam->edged_width) * 16; Data->Ref = pRef + (x + y * pParam->edged_width) * 16; @@ -2484,7 +1903,7 @@ pmv[2].y = EVEN(Data->predMV.y); pmv[0].x = pmv[0].y = 0; - (*CheckCandidate)(0, 0, 255, &i, Data); + CheckCandidate16no4vI(0, 0, 255, &i, Data); //early skip for 0,0 if (*Data->iMinSAD < MAX_SAD00_FOR_SKIP * 4) { @@ -2494,11 +1913,11 @@ } if (!(mask = make_mask(pmv, 1))) - (*CheckCandidate)(pmv[1].x, pmv[1].y, mask, &i, Data); + CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); if (!(mask = make_mask(pmv, 2))) - (*CheckCandidate)(pmv[2].x, pmv[2].y, mask, &i, Data); + CheckCandidate16no4vI(pmv[2].x, pmv[2].y, mask, &i, Data); - if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 4) // diamond only if needed + if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 6) // 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]; @@ -2507,7 +1926,7 @@ } #define INTRA_THRESH 1350 -#define INTER_THRESH 900 +#define INTER_THRESH 1200 int @@ -2540,7 +1959,7 @@ IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; - InterThresh += 300 * (1 - bCount); + InterThresh += 400 * (1 - bCount); if (InterThresh < 200) InterThresh = 200; if (sadInit) (*sadInit) (); @@ -2558,47 +1977,17 @@ pParam->edged_width); if (dev + IntraThresh < sad) { pMB->mode = MODE_INTRA; - if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame + if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; } } sSAD += sad; } } sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); - if (sSAD > InterThresh ) return 1; //P frame + if (sSAD > InterThresh ) return P_VOP; emms(); - return 0; // B frame - -} - -int -FindFcode( const MBParam * const pParam, - const FRAMEINFO * const current) -{ - uint32_t x, y; - int max = 0, min = 0, i; + return B_VOP; - for (y = 0; y < pParam->mb_height; y++) { - for (x = 0; x < pParam->mb_width; x++) { - - MACROBLOCK *pMB = ¤t->mbs[x + y * pParam->mb_width]; - for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) { - if (pMB->mvs[i].x > max) max = pMB->mvs[i].x; - if (pMB->mvs[i].y > max) max = pMB->mvs[i].y; - - if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; - if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; - } - } - } - - min = -min; - max += 1; - if (min > max) max = min; - if (pParam->m_quarterpel) max *= 2; - - for (i = 1; (max > 32 << (i - 1)); i++); - return i; } static void @@ -2626,7 +2015,7 @@ } -static VECTOR +static __inline VECTOR GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) {