--- branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/11/25 13:38:56 676 +++ branches/dev-api-3/xvidcore/src/motion/motion_est.c 2003/02/12 12:57:27 836 @@ -1,7 +1,7 @@ /************************************************************************** * * XVID MPEG-4 VIDEO CODEC - * motion estimation + * motion estimation * * This program is an implementation of a part of one or more MPEG-4 * Video tools as specified in ISO/IEC 14496-2 standard. Those intending @@ -31,6 +31,8 @@ #include #include #include +#include // memcpy +#include // lrint #include "../encoder.h" #include "../utils/mbfunctions.h" @@ -42,82 +44,147 @@ #include "motion.h" #include "sad.h" #include "../utils/emms.h" +#include "../dct/fdct.h" #define INITIAL_SKIP_THRESH (10) #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 ); } +CheckCandidate((X),(Y), (D), &iDirection, data ); } -#define iDiamondSize 2 - -static __inline int -d_mv_bits(int x, int y, const uint32_t iFcode) +static __inline uint32_t +d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) { int xb, yb; - - if (x == 0) xb = 1; - else { - if (x < 0) x = -x; + x = qpel ? x<<1 : x; + y = qpel ? y<<1 : y; + if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } + + x -= pred.x; + y -= pred.y; + + if (x) { + x = ABS(x); x += (1 << (iFcode - 1)) - 1; x >>= (iFcode - 1); if (x > 32) x = 32; xb = mvtab[x] + iFcode; - } + } else xb = 1; - if (y == 0) yb = 1; - else { - if (y < 0) y = -y; + if (y) { + y = ABS(y); y += (1 << (iFcode - 1)) - 1; y >>= (iFcode - 1); if (y > 32) y = 32; yb = mvtab[y] + iFcode; - } + } else yb = 1; return xb + yb; } -static int32_t +static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data) +{ + int sad; + const uint32_t stride = data->iEdgedWidth/2; + uint8_t * f_refu = data->RefQ, + * f_refv = data->RefQ + 8, + * b_refu = data->RefQ + 16, + * b_refv = data->RefQ + 24; + + switch (((fx & 1) << 1) | (fy & 1)) { + case 0: + fx = fx / 2; fy = fy / 2; + f_refu = (uint8_t*)data->RefCU + fy * stride + fx, stride; + f_refv = (uint8_t*)data->RefCV + fy * stride + fx, stride; + break; + case 1: + fx = fx / 2; fy = (fy - 1) / 2; + interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); + interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); + break; + case 2: + fx = (fx - 1) / 2; fy = fy / 2; + interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); + interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); + break; + default: + fx = (fx - 1) / 2; fy = (fy - 1) / 2; + interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); + interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); + break; + } + + switch (((bx & 1) << 1) | (by & 1)) { + case 0: + bx = bx / 2; by = by / 2; + b_refu = (uint8_t*)data->b_RefCU + by * stride + bx, stride; + b_refv = (uint8_t*)data->b_RefCV + by * stride + bx, stride; + break; + case 1: + bx = bx / 2; by = (by - 1) / 2; + interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); + interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); + break; + case 2: + bx = (bx - 1) / 2; by = by / 2; + interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); + interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); + break; + default: + bx = (bx - 1) / 2; by = (by - 1) / 2; + interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); + interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); + break; + } + + sad = sad8bi(data->CurU, b_refu, f_refu, stride); + sad += sad8bi(data->CurV, b_refv, f_refv, stride); + + return sad; +} + + +static int32_t ChromaSAD(int dx, int dy, const SearchData * const data) { int sad; - dx = (dx >> 1) + roundtab_79[dx & 0x3]; - dy = (dy >> 1) + roundtab_79[dy & 0x3]; + const uint32_t stride = data->iEdgedWidth/2; + + if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently + data->temp[5] = dx; data->temp[6] = dy; // backup - switch (((dx & 1) << 1) + (dy & 1)) { // ((dx%2)?2:0)+((dy%2)?1:0) + switch (((dx & 1) << 1) | (dy & 1)) { case 0: - sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); - sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); + dx = dx / 2; dy = dy / 2; + sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride); + sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride); break; case 1: dx = dx / 2; dy = (dy - 1) / 2; - sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); - sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); + sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride); + sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride); break; case 2: dx = (dx - 1) / 2; dy = dy / 2; - sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); - sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); + sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride); + sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride); break; default: dx = (dx - 1) / 2; dy = (dy - 1) / 2; - interpolate8x8_halfpel_hv(data->RefQ, - data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, - data->rounding); - sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); - interpolate8x8_halfpel_hv(data->RefQ, - data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, - data->rounding); - sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); + interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); + sad = sad8(data->CurU, data->RefQ, stride); + + interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); + sad += sad8(data->CurV, data->RefQ, stride); break; } + data->temp[7] = sad; //backup, part 2 return sad; } static __inline const uint8_t * -GetReference(const int x, const int y, const int dir, const SearchData * const data) +GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) { // dir : 0 = forward, 1 = backward switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { @@ -129,46 +196,56 @@ 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); + } +} +// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate +static __inline const uint8_t * +GetReference(const int x, const int y, const SearchData * const data) +{ + switch ( ((x&1)<<1) | (y&1) ) { + case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); + case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); + case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); + default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2 } } -static uint8_t * -Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) +static uint8_t * +Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t 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; + uint8_t * Reference = data->RefQ + 16*dir; + const uint32_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 = GetReferenceB(halfpel_x, halfpel_y, dir, data); ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; switch( ((x&1)<<1) + (y&1) ) { case 0: // pure halfpel position - Reference = (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); - Reference += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + return (uint8_t *) ref1; break; case 1: // x halfpel, y qpel - top or bottom during qpel refinement - ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); + ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); break; case 2: // x qpel, y halfpel - left or right during qpel refinement - ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); + ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); break; default: // x and y in qpel resolution - the "corners" (top left/right and // bottom left/right) during qpel refinement - ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); - ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); - ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); + ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); + ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); + ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; @@ -178,23 +255,32 @@ return Reference; } -static uint8_t * -Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) +static uint8_t * +Interpolate16x16qpel(const int x, const int y, const uint32_t 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; + uint8_t * Reference = data->RefQ + 16*dir; + const uint32_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 = GetReferenceB(halfpel_x, halfpel_y, dir, data); switch( ((x&1)<<1) + (y&1) ) { - case 0: // pure halfpel position - return (uint8_t *) GetReference(halfpel_x, halfpel_y, dir, data); + case 3: // x and y in qpel resolution - the "corners" (top left/right and + // bottom left/right) during qpel refinement + ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); + ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); + ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); + interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); + interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); + interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); + 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; + case 1: // x halfpel, y qpel - top or bottom during qpel refinement - ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); + ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); @@ -202,71 +288,59 @@ break; case 2: // x qpel, y halfpel - left or right during qpel refinement - ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); + ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); 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; + case 0: // pure halfpel position + return (uint8_t *) ref1; } return Reference; } /* CHECK_CANDIATE FUNCTIONS START */ -static void +static void CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - int t, xc, yc; + int xc, yc; const uint8_t * Reference; VECTOR * current; + int32_t sad; uint32_t t; - if (( x > data->max_dx) || ( x < data->min_dx) - || ( y > data->max_dy) || (y < data->min_dy)) return; + if ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; - if (data->qpel_precision) { // x and y are in 1/4 precision + if (!data->qpel_precision) { + Reference = GetReference(x, y, data); + current = data->currentMV; + xc = x; yc = y; + } else { // x and y are in 1/4 precision Reference = Interpolate16x16qpel(x, y, 0, data); - t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); xc = x/2; yc = y/2; //for chroma sad current = data->currentQMV; - } else { - 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) t = d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode); - else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); - current = data->currentMV; - xc = x; yc = y; } - - data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); - data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; - data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; + sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); + t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); - if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data); + sad += (data->lambda16 * t * sad)>>10; + data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; - if (data->temp[0] < data->iMinSAD[0]) { - data->iMinSAD[0] = data->temp[0]; + if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); + + if (sad < data->iMinSAD[0]) { + data->iMinSAD[0] = sad; current[0].x = x; current[0].y = y; - *dir = Direction; } + *dir = Direction; + } if (data->temp[1] < data->iMinSAD[1]) { - data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[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]; current[2].x = x; current[2].y = y; } if (data->temp[3] < data->iMinSAD[3]) { @@ -276,111 +350,185 @@ } -static void +static void +CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad; uint32_t t; + const uint8_t * Reference; + + if ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; + + if (!data->qpel_precision) Reference = GetReference(x, y, data); + else Reference = Interpolate8x8qpel(x, y, 0, 0, data); + + sad = sad8(data->Cur, Reference, data->iEdgedWidth); + t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); + + sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; + + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; + } +} + + +static void +CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + uint32_t t; + const uint8_t * Reference; + + 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; + + Reference = GetReference(x, y, data); + t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); + + data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); + + data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; + data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; + + if (data->temp[0] < data->iMinSAD[0]) { + data->iMinSAD[0] = data->temp[0]; + data->currentMV[0].x = x; data->currentMV[0].y = y; + *dir = Direction; } + + if (data->temp[1] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } + if (data->temp[2] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } + if (data->temp[3] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } + if (data->temp[4] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } +} + +static void CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - int32_t sad; + int32_t sad, xc, yc; const uint8_t * Reference; - int t; + uint32_t t; VECTOR * current; - if (( x > data->max_dx) || ( x < data->min_dx) - || ( y > data->max_dy) || (y < data->min_dy)) return; + if ( (x > data->max_dx) | ( x < data->min_dx) + | (y > data->max_dy) | (y < data->min_dy) ) return; + + if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value if (data->qpel_precision) { // x and y are in 1/4 precision Reference = Interpolate16x16qpel(x, y, 0, data); - t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); current = data->currentQMV; + xc = x/2; yc = y/2; } else { - 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) t = d_mv_bits(2*x - data->predMV.x, 2*y - data->predMV.y, data->iFcode); - else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + Reference = GetReference(x, y, data); current = data->currentMV; + xc = x; yc = y; } - + t = d_mv_bits(x, y, data->predMV, data->iFcode, + data->qpel^data->qpel_precision, data->rrv); + sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); - sad += (data->lambda16 * t * sad)/1000; + sad += (data->lambda16 * t * sad)>>10; + + if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; current->x = x; current->y = y; - *dir = Direction; } + *dir = Direction; + } } -static void -CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +static void +CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { // maximum speed - for P/B/I decision int32_t sad; - if (( x > data->max_dx) || ( x < data->min_dx) - || ( y > data->max_dy) || (y < data->min_dy)) return; + if ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; - sad = sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), - data->iEdgedWidth, 256*4096); + sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), + data->iEdgedWidth, data->temp+1); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; data->currentMV[0].x = x; data->currentMV[0].y = y; - *dir = Direction; } -} + *dir = Direction; + } + if (data->temp[1] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } + if (data->temp[2] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } + if (data->temp[3] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } + if (data->temp[4] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } +} -static void +static void CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) { - int32_t sad; - int xb, yb, t; + int32_t sad, xb, yb, xcf, ycf, xcb, ycb; + uint32_t 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; + if ( (xf > data->max_dx) | (xf < data->min_dx) + | (yf > data->max_dy) | (yf < data->min_dy) ) return; - if (data->qpel_precision) { + if (!data->qpel_precision) { + ReferenceF = GetReference(xf, yf, data); + xb = data->currentMV[1].x; yb = data->currentMV[1].y; + ReferenceB = GetReferenceB(xb, yb, 1, data); + current = data->currentMV; + xcf = xf; ycf = yf; + xcb = xb; ycb = yb; + } else { 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); - t = 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); - } else { - ReferenceF = Interpolate16x16qpel(2*xf, 2*yf, 0, data); - xb = data->currentMV[1].x; yb = data->currentMV[1].y; - ReferenceB = Interpolate16x16qpel(2*xb, 2*yb, 1, data); - current = data->currentMV; - if (data->qpel) - t = 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); - else - t = 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); + xcf = xf/2; ycf = yf/2; + xcb = xb/2; ycb = yb/2; } + t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0) + + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0); + sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); - sad += (data->lambda16 * t * sad)/1000; + sad += (data->lambda16 * t * sad)>>10; + + if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], + (ycf >> 1) + roundtab_79[ycf & 0x3], + (xcb >> 1) + roundtab_79[xcb & 0x3], + (ycb >> 1) + roundtab_79[ycb & 0x3], data); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; current->x = xf; current->y = yf; - *dir = Direction; } + *dir = Direction; + } } -static void +static void CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - int32_t sad = 0; - int k; + int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; + uint32_t k; const uint8_t *ReferenceF; const uint8_t *ReferenceB; VECTOR mvs, b_mvs; - if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; + if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; for (k = 0; k < 4; k++) { mvs.x = data->directmvF[k].x + x; @@ -392,42 +540,53 @@ 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; - - if (!data->qpel) { - mvs.x *= 2; mvs.y *= 2; - b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway + + 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; + + if (data->qpel) { + xcf += mvs.x/2; ycf += mvs.y/2; + xcb += b_mvs.x/2; ycb += b_mvs.y/2; + } else { + xcf += mvs.x; ycf += mvs.y; + xcb += b_mvs.x; ycb += b_mvs.y; + mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway + b_mvs.x *= 2; b_mvs.y *= 2; } + 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, ReferenceB, - data->iEdgedWidth); + ReferenceF, ReferenceB, data->iEdgedWidth); if (sad > *(data->iMinSAD)) return; } - sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; + sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; + + if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], + (ycf >> 3) + roundtab_76[ycf & 0xf], + (xcb >> 3) + roundtab_76[xcb & 0xf], + (ycb >> 3) + roundtab_76[ycb & 0xf], data); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; data->currentMV->x = x; data->currentMV->y = y; - *dir = Direction; } + *dir = Direction; + } } -static void +static void CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - int32_t sad; + int32_t sad, xcf, ycf, xcb, ycb; const uint8_t *ReferenceF; const uint8_t *ReferenceB; VECTOR mvs, b_mvs; - if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; + if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return; mvs.x = data->directmvF[0].x + x; b_mvs.x = ((x == 0) ? @@ -438,50 +597,163 @@ 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; - - 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); + + 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; + + if (data->qpel) { + xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); + xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); + ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); + ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); + } else { + xcf = 4*mvs.x; ycf = 4*mvs.y; + xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; + ReferenceF = GetReference(mvs.x, mvs.y, data); + ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); + } sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); - sad += (data->lambda16 * d_mv_bits(x, y, 1) * sad)/1000; + sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; + + if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], + (ycf >> 3) + roundtab_76[ycf & 0xf], + (xcb >> 3) + roundtab_76[xcb & 0xf], + (ycb >> 3) + roundtab_76[ycb & 0xf], data); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; data->currentMV->x = x; data->currentMV->y = y; - *dir = Direction; } + *dir = Direction; + } } + static void -CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - int32_t sad; int t; - const uint8_t * Reference; - if (( x > data->max_dx) || ( x < data->min_dx) - || ( y > data->max_dy) || (y < data->min_dy)) return; + static int16_t in[64], coeff[64]; + int32_t bits = 0, sum; + VECTOR * current; + const uint8_t * ptr; + int i, cbp = 0, t, xc, yc; - if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); - else Reference = Interpolate16x16qpel(2*x, 2*y, 0, data); + if ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; - sad = sad8(data->Cur, Reference, data->iEdgedWidth); - if (data->qpel) t = d_mv_bits(2 * x - data->predMV.x, 2 * y - data->predMV.y, data->iFcode); - else t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + if (!data->qpel_precision) { + ptr = GetReference(x, y, data); + current = data->currentMV; + xc = x; yc = y; + } else { // x and y are in 1/4 precision + ptr = Interpolate16x16qpel(x, y, 0, data); + current = data->currentQMV; + xc = x/2; yc = y/2; + } + + for(i = 0; i < 4; i++) { + int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); + transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); + fdct(in); + if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); + else sum = quant4_inter(coeff, in, data->lambda16); + if (sum > 0) { + cbp |= 1 << (5 - i); + bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + } else data->temp[i] = 0; + } + + bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); + + if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma + xc = (xc >> 1) + roundtab_79[xc & 0x3]; + yc = (yc >> 1) + roundtab_79[yc & 0x3]; + + //chroma U + ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCU, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); + transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); + fdct(in); + if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); + else sum = quant4_inter(coeff, in, data->lambda16); + if (sum > 0) { + cbp |= 1 << (5 - 4); + bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + } + + if (bits < data->iMinSAD[0]) { + //chroma V + ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCV, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); + transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2); + fdct(in); + if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); + else sum = quant4_inter(coeff, in, data->lambda16); + if (sum > 0) { + cbp |= 1 << (5 - 5); + bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + } + } + } - sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; + bits += cbpy_tab[15-(cbp>>2)].len; + bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; - if (sad < *(data->iMinSAD)) { - *(data->iMinSAD) = sad; - data->currentMV->x = x; data->currentMV->y = y; - *dir = Direction; } + if (bits < data->iMinSAD[0]) { + data->iMinSAD[0] = bits; + current[0].x = x; current[0].y = y; + *dir = Direction; + } + + if (data->temp[0] + t < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; } + if (data->temp[1] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } + if (data->temp[2] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } + if (data->temp[3] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } + +} +static void +CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + + static int16_t in[64], coeff[64]; + int32_t sum, bits; + VECTOR * current; + const uint8_t * ptr; + int cbp; + + if ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; + + if (!data->qpel_precision) { + ptr = GetReference(x, y, data); + current = data->currentMV; + } else { // x and y are in 1/4 precision + ptr = Interpolate8x8qpel(x, y, 0, 0, data); + current = data->currentQMV; + } + + transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); + fdct(in); + if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); + else sum = quant4_inter(coeff, in, data->lambda16); + if (sum > 0) { + bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + cbp = 1; + } else cbp = bits = 0; + + bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); + + if (bits < data->iMinSAD[0]) { + data->temp[0] = cbp; + data->iMinSAD[0] = bits; + current[0].x = x; current[0].y = y; + *dir = Direction; + } } /* CHECK_CANDIATE FUNCTIONS END */ @@ -494,82 +766,83 @@ /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - int iDirection; + int iDirection; + + 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 */ - do { + if (iDirection) { //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; + 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 @@ -599,92 +872,86 @@ /* 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 */ -/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ - static void SubpelRefine(const SearchData * const data) { /* Do a half-pel or q-pel refinement */ - VECTOR backupMV; - int iDirection; //not needed + const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; + int iDirection; //only needed because macro expects it - 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); - 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); + CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); + CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); + CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); + CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); + CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); + CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); + CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); + CHECK_CANDIDATE(centerMV.x - 1, centerMV.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 stride, 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; + if(!rrv) { + uint32_t sadC = sad8(current->u + x*8 + y*stride*8, + reference->u + x*8 + y*stride*8, stride); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; + sadC += sad8(current->v + (x + y*stride)*8, + reference->v + (x + y*stride)*8, stride); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; + return 1; - return 1; + } else { + uint32_t sadC = sad16(current->u + x*16 + y*stride*16, + reference->u + x*16 + y*stride*16, stride, 256*4096); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; + sadC += sad16(current->v + (x + y*stride)*16, + reference->v + (x + y*stride)*16, stride, 256*4096); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; + return 1; + } } static __inline void SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) { pMB->mode = MODE_NOT_CODED; - pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; - pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; - - pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; - pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; - + pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; + pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; } @@ -701,20 +968,23 @@ const IMAGE *const pCurrent = ¤t->image; const IMAGE *const pRef = &reference->image; - const VECTOR zeroMV = { 0, 0 }; + uint32_t mb_width = pParam->mb_width; + uint32_t mb_height = pParam->mb_height; + const uint32_t iEdgedWidth = pParam->edged_width; + const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags); uint32_t x, y; uint32_t iIntra = 0; - int32_t InterBias, quant = current->quant, sad00; - uint8_t *qimage; + int32_t quant = current->quant, sad00; // 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; - Data.iEdgedWidth = pParam->edged_width; + memset(&Data, 0, sizeof(SearchData)); + Data.iEdgedWidth = iEdgedWidth; Data.currentMV = currentMV; Data.currentQMV = currentQMV; Data.iMinSAD = iMinSAD; @@ -722,113 +992,95 @@ Data.iFcode = current->fcode; Data.rounding = pParam->m_rounding_type; Data.qpel = pParam->m_quarterpel; - Data.chroma = current->global_flags & XVID_ME_COLOUR; + Data.chroma = MotionFlags & PMV_CHROMA16; + 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 = 0; + } - if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) - return 1; // 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; + Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) 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, - pRef->y + (x + y * pParam->edged_width) * 16, + + if (!Data.rrv) pMB->sad16 = + sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, + pRef->y + (x + y * iEdgedWidth) * 16, pParam->edged_width, pMB->sad8 ); - if (Data.chroma) { - pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, - pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); + else pMB->sad16 = + sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, + pRef->y + (x + y * iEdgedWidth) * 32, + pParam->edged_width, pMB->sad8 ); - pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, - pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); + if (Data.chroma) { + Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, + pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) + + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, + pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); + pMB->sad16 += Data.temp[7]; } - sad00 = pMB->sad16; //if no gmc; else sad00 = (..) + sad00 = pMB->sad16; 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; } + pMB->quant = current->quant; //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 (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ + if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) + if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { SkipMacroblockP(pMB, sad00); continue; } } - + SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, - y, current->motion_flags, pMB->quant, + y, MotionFlags, current->global_flags, pMB->quant, &Data, pParam, pMBs, reference->mbs, current->global_flags & XVID_INTER4V, pMB); /* final skip decision, a.k.a. "the vector you found, really that good?" */ - if (current->coding_type == P_VOP) { - if ( (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)) { - SkipMacroblockP(pMB, sad00); - continue; - } - } - -/* finally, intra decision */ + if (!(current->global_flags & XVID_GMC)) { + if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) { + if (!(current->global_flags & XVID_MODEDECISION_BITS)) { + if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) ) + if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) + SkipMacroblockP(pMB, sad00); + } else { // BITS mode decision + if (pMB->sad16 > 10) + SkipMacroblockP(pMB, sad00); // more than 10 bits would be used for this MB - skip - InterBias = MV16_INTER_BIAS; - if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work - if (y != 0) - if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; - if (x != 0) - if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; - - if (Data.chroma) InterBias += 50; // to compensate bigger SAD - - if (InterBias < pMB->sad16) { - const int32_t deviation = - dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, - pParam->edged_width); - - 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] = - pMB->mvs[3] = zeroMV; - pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = - pMB->qmvs[3] = zeroMV; - pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = - pMB->sad8[3] = 0; + } } } + if (pMB->mode == MODE_INTRA) + if (++iIntra > iLimit) return 1; } } - free(qimage); - - if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ - current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); - else - current->GMC_MV = zeroMV; + if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */ + { + current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); + } return 0; } -#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) - static __inline int make_mask(const VECTOR * const pmv, const int i) { @@ -836,34 +1088,35 @@ for (j = 0; j < i; j++) { if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already if (pmv[i].x == pmv[j].x) { - if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } - if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } - } else + if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; + else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; + } else if (pmv[i].y == pmv[j].y) { - if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } - if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } + if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; + else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; } } return mask; } -static __inline void -PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, - const int iHcount, const MACROBLOCK * const prevMB) +static __inline void +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); + pmv[5].y = EVEN(pmv[3].y); } else pmv[5].x = pmv[5].y = 0; if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour else pmv[3].x = pmv[3].y = 0; if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour - else pmv[4].x = pmv[4].y = 0; + 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); @@ -873,10 +1126,99 @@ 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); + 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_SCALEUP(pmv[i].x); + pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); + } + } +} + +static int +ModeDecision(const uint32_t iQuant, SearchData * const Data, + int inter4v, + MACROBLOCK * const pMB, + const MACROBLOCK * const pMBs, + const int x, const int y, + const MBParam * const pParam, + const uint32_t MotionFlags, + const uint32_t GlobalFlags) +{ + + int mode = MODE_INTER; + + if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision + int intra = 0; + int sad; + int InterBias = MV16_INTER_BIAS; + if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { + mode = 0; //inter + sad = Data->iMinSAD[0]; + } else { + mode = MODE_INTER4V; + sad = Data->iMinSAD[1] + Data->iMinSAD[2] + + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; + Data->iMinSAD[0] = sad; + } + + /* intra decision */ + + if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work + if (y != 0) + if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; + if (x != 0) + 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) { + int32_t deviation; + if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth); + else deviation = dev16(Data->Cur, Data->iEdgedWidth) + + dev16(Data->Cur+8, Data->iEdgedWidth) + + dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) + + dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth); + + if (deviation < (sad - InterBias)) return MODE_INTRA;// intra + } + return mode; + + } else { + + int bits, intra, i; + VECTOR backup[5], *v; + Data->lambda16 = iQuant; + Data->lambda8 = pParam->m_quant_type; + + v = Data->qpel ? Data->currentQMV : Data->currentMV; + for (i = 0; i < 5; i++) { + Data->iMinSAD[i] = 256*4096; + backup[i] = v[i]; + } + + bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags); + if (bits == 0) return MODE_INTER; // quick stop + + if (inter4v) { + int inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); + if (inter4v < bits) { Data->iMinSAD[0] = bits = inter4v; mode = MODE_INTER4V; } + } + + + intra = CountMBBitsIntra(Data); + + if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; } + + return mode; + } } static void @@ -888,6 +1230,7 @@ const int x, const int y, const uint32_t MotionFlags, + const uint32_t GlobalFlags, const uint32_t iQuant, SearchData * const Data, const MBParam * const pParam, @@ -900,78 +1243,78 @@ 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->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; + get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); + + Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache + i = Data->rrv ? 2 : 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++) + for(i = 0; i < 5; i++) Data->currentMV[i].x = Data->currentMV[i].y = 0; - if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); + if (Data->qpel) 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); - 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; + i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); + Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); + Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); Data->iMinSAD[2] = pMB->sad8[1]; Data->iMinSAD[3] = pMB->sad8[2]; Data->iMinSAD[4] = pMB->sad8[3]; - if ((x == 0) && (y == 0)) threshA = 512; - else { - threshA = Data->temp[0]; // that's when we keep this SAD atm + if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) { + threshA = Data->temp[0]; // that's where we keep this SAD atm if (threshA < 512) threshA = 512; - if (threshA > 1024) threshA = 1024; } + else if (threshA > 1024) threshA = 1024; + } else + threshA = 512; 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 || Data->chroma) CheckCandidate = CheckCandidate16; - else CheckCandidate = CheckCandidate16no4v; //for extra speed + if (!Data->rrv) { + if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; //for extra speed + } else CheckCandidate = CheckCandidate32; -/* main loop. checking all predictions */ +/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ for (i = 1; i < 7; i++) { if (!(mask = make_mask(pmv, i)) ) continue; - (*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); + CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data); if (Data->iMinSAD[0] <= threshA) break; } if ((Data->iMinSAD[0] <= threshA) || (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { - inter4v = 0; - } else { + if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; } + else { MainSearchFunc * MainSearchPtr; if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; - (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); + MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); /* extended search, diamond starting in 0,0 and in prediction. note that this search is/might be done in halfpel positions, @@ -980,26 +1323,27 @@ 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); + } if (!(MVequal(startMV, backupMV))) { bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; - (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); - (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); + CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); + MainSearchPtr(startMV.x, startMV.y, Data, 255); if (bSAD < Data->iMinSAD[0]) { Data->currentMV[0] = backupMV; Data->iMinSAD[0] = bSAD; } } backupMV = Data->currentMV[0]; - if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; - else startMV.x = startMV.y = 0; + startMV.x = startMV.y = 1; if (!(MVequal(startMV, backupMV))) { bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; - (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data); - (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); + CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); + MainSearchPtr(startMV.x, startMV.y, Data, 255); if (bSAD < Data->iMinSAD[0]) { Data->currentMV[0] = backupMV; Data->iMinSAD[0] = bSAD; } @@ -1007,64 +1351,68 @@ } } - if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data); + if (MotionFlags & PMV_HALFPELREFINE16) + if ((!(MotionFlags & HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant) + 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 (MotionFlags & PMV_QUARTERPELREFINE16) + if ((!(MotionFlags & QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) { + 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, 1, 0); - 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); + SubpelRefine(Data); + } - SubpelRefine(Data); - } + if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0; + + if (inter4v && (!(GlobalFlags & XVID_MODEDECISION_BITS) || + (!(MotionFlags & QUARTERPELREFINE8_BITS)) || (!(MotionFlags & HALFPELREFINE8_BITS)) || + ((!(MotionFlags & EXTSEARCH_BITS)) && (!(MotionFlags&PMV_EXTSEARCH8)) ))) { + // if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop - if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; - if (inter4v) { SearchData Data8; - Data8.iFcode = Data->iFcode; - Data8.lambda8 = Data->lambda8; - Data8.iEdgedWidth = Data->iEdgedWidth; - Data8.RefQ = Data->RefQ; - Data8.qpel = Data->qpel; + memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data + 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 sumx, sumy, dx, dy; - if(pParam->m_quarterpel) { - 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; + if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) { + // chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it + int sumx = 0, sumy = 0; + const int div = 1 + Data->qpel; + const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; + + for (i = 0; i < 4; i++) { + sumx += mv[i].x / div; + sumy += mv[i].y / div; } - dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; - dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; - - Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); + + Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], + (sumy >> 3) + roundtab_76[sumy & 0xf], 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]; + inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags); - pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = - pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; + 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(pParam->m_quarterpel) { + if (inter4v == MODE_INTER) { + pMB->mode = MODE_INTER; + pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; + + if(Data->qpel) { 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; @@ -1073,12 +1421,15 @@ 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 + + } else if (inter4v == MODE_INTER4V) { pMB->mode = MODE_INTER4V; - pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + - Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; + pMB->sad16 = Data->iMinSAD[0]; + } else { // INTRA mode + SkipMacroblockP(pMB, 0); // not skip, but similar enough + pMB->mode = MODE_INTRA; } + } static void @@ -1091,47 +1442,49 @@ 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->predMV = 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->predMV.x, - Data->currentQMV->y - Data->predMV.y, - Data->iFcode) * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; + if(Data->qpel) { + 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->currentQMV->y, + Data->predMV, Data->iFcode, 0, 0); } else { - Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); - if (block != 0) *(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) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, + Data->predMV, Data->iFcode, 0, Data->rrv); } - if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { + *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; + + if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { + 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) + Data->iEdgedWidth*(block>>1)); + Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); - Data->Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); Data->qpel_precision = 0; - + 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, Data->iFcode - Data->qpel, 0, Data->rrv); - if (MotionFlags & PMV_EXTSEARCH8) { + if (!Data->rrv) CheckCandidate = CheckCandidate8; + else CheckCandidate = CheckCandidate16no4v; + + if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) { int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD - + MainSearchFunc *MainSearchPtr; if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; - (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); + MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); if(*(Data->iMinSAD) < temp_sad) { Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector @@ -1150,39 +1503,39 @@ } } - if(pParam->m_quarterpel) { - if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && - (MotionFlags & PMV_QUARTERPELREFINE8)) { - 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, 0); - SubpelRefine(Data); - } + if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) { + 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, Data->iFcode, 1, 0); + SubpelRefine(Data); } } - if(pParam->m_quarterpel) { + if (Data->rrv) { + Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); + Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); + } + + 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->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->sad8[block] = 4 * (*Data->iMinSAD); + pMB->mvs[block] = *Data->currentMV; + pMB->sad8[block] = 4 * *Data->iMinSAD; } -/* B-frames code starts here */ +/* motion estimation for B-frames */ static __inline VECTOR ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) { /* the stupidiest function ever */ - if (mode == MODE_FORWARD) return pMB->mvs[0]; - else return pMB->b_mvs[0]; + return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); } static void __inline @@ -1202,7 +1555,7 @@ if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); - pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); + pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); } else pmv[3].x = pmv[3].y = 0; if (y != 0) { @@ -1215,18 +1568,16 @@ pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); } else pmv[5].x = pmv[5].y = 0; - if ((x != 0)&&(y != 0)) { + if (x != 0 && y != 0) { pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); - pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); + pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); } else pmv[6].x = pmv[6].y = 0; - -// more? } -/* search backward or forward, for b-frames */ +/* search backward or forward */ static void -SearchBF( const uint8_t * const pRef, +SearchBF( const IMAGE * const pRef, const uint8_t * const pRefH, const uint8_t * const pRefV, const uint8_t * const pRefHV, @@ -1242,61 +1593,61 @@ SearchData * const Data) { - const int32_t iEdgedWidth = pParam->edged_width; - - int i, iDirection, mask; + int i, iDirection = 255, mask; VECTOR pmv[7]; MainSearchFunc *MainSearchPtr; *Data->iMinSAD = MV_MAX_ERROR; Data->iFcode = iFcode; Data->qpel_precision = 0; + Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache - Data->Ref = pRef + (x + y * iEdgedWidth) * 16; - Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; - Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; - Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; + Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16; + Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16; + Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16; + Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16; + Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8; + Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8; 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 - for (i = 0; i < 8; i++) { + for (i = 0; i < 7; i++) { if (!(mask = make_mask(pmv, i)) ) continue; CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); } - if (MotionFlags & PMV_USESQUARES16) - MainSearchPtr = SquareSearch; - else if (MotionFlags & PMV_ADVANCEDDIAMOND16) - MainSearchPtr = AdvDiamondSearch; + 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); + MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); SubpelRefine(Data); - - if (Data->qpel) { + + if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { Data->currentQMV->x = 2*Data->currentMV->x; Data->currentQMV->y = 2*Data->currentMV->y; Data->qpel_precision = 1; get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, - pParam->width, pParam->height, iFcode, 0); + pParam->width, pParam->height, iFcode, 1, 0); SubpelRefine(Data); } // three bits are needed to code backward mode. four for forward -// we treat the bits just like they were vector's - if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; - else *Data->iMinSAD += 3 * Data->lambda16; + + if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; + else *Data->iMinSAD += 3 * Data->lambda16; if (*Data->iMinSAD < *best_sad) { *best_sad = *Data->iMinSAD; @@ -1304,24 +1655,65 @@ if (Data->qpel) { pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; - if (mode_current == MODE_FORWARD) + if (mode_current == MODE_FORWARD) pMB->qmvs[0] = *Data->currentQMV; - else + else pMB->b_qmvs[0] = *Data->currentQMV; } else { pMB->pmvs[0].x = Data->currentMV->x - predMV->x; pMB->pmvs[0].y = Data->currentMV->y - predMV->y; } - if (mode_current == MODE_FORWARD) - pMB->mvs[0] = *(Data->currentMV+2) = *Data->currentMV; - else - pMB->b_mvs[0] = *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search + if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; + else pMB->b_mvs[0] = *Data->currentMV; + } + + if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; + else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search +} +static void +SkipDecisionB(const IMAGE * const pCur, + const IMAGE * const f_Ref, + const IMAGE * const b_Ref, + MACROBLOCK * const pMB, + const uint32_t x, const uint32_t y, + const SearchData * const Data) +{ + int dx = 0, dy = 0, b_dx = 0, b_dy = 0; + int32_t sum; + const int div = 1 + Data->qpel; + int k; + const uint32_t stride = Data->iEdgedWidth/2; +//this is not full chroma compensation, only it's fullpel approximation. should work though + + for (k = 0; k < 4; k++) { + dy += Data->directmvF[k].y / div; + dx += Data->directmvF[0].x / div; + b_dy += Data->directmvB[0].y / div; + b_dx += Data->directmvB[0].x / div; } - + + 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 * stride, + f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, + b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, + stride); + + if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip + + sum += sad8bi(pCur->v + 8*x + 8 * y * stride, + f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, + b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, + stride); + + if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped } -static int32_t +static __inline uint32_t SearchDirect(const IMAGE * const f_Ref, const uint8_t * const f_RefH, const uint8_t * const f_RefV, @@ -1342,33 +1734,31 @@ { int32_t skip_sad; - int k; - + int k = (x + Data->iEdgedWidth*y) * 16; MainSearchFunc *MainSearchPtr; *Data->iMinSAD = 256*4096; + Data->Ref = f_Ref->y + k; + Data->RefH = f_RefH + k; + Data->RefV = f_RefV + k; + Data->RefHV = f_RefHV + k; + Data->bRef = b_Ref->y + k; + Data->bRefH = b_RefH + k; + Data->bRefV = b_RefV + k; + Data->bRefHV = b_RefHV + k; + Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; + Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; + Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; + Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; + + k = Data->qpel ? 4 : 2; + Data->max_dx = k * (pParam->width - x * 16); + Data->max_dy = k * (pParam->height - y * 16); + Data->min_dx = -k * (16 + x * 16); + Data->min_dy = -k * (16 + y * 16); - Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; - Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; - Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; - Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; - Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; - Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; - Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; - Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; - - Data->max_dx = 2 * pParam->width - 2 * (x) * 16; - Data->max_dy = 2 * pParam->height - 2 * (y) * 16; - Data->min_dx = -(2 * 16 + 2 * (x) * 16); - Data->min_dy = -(2 * 16 + 2 * (y) * 16); - if (Data->qpel) { //we measure in qpixels - Data->max_dx *= 2; - Data->max_dy *= 2; - Data->min_dx *= 2; - 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 + Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; + Data->qpel_precision = 0; for (k = 0; k < 4; k++) { pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); @@ -1376,13 +1766,13 @@ pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; - if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx ) - || ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) { + if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) + | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { *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]; @@ -1393,73 +1783,38 @@ } } - - 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; + CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; - 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); + CheckCandidate(0, 0, 255, &k, Data); - if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { +// initial (fast) skip decision + if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { + //possible skip + if (Data->chroma) { pMB->mode = MODE_DIRECT_NONE_MV; - return *Data->iMinSAD; + return *Data->iMinSAD; // skip. + } else { + SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); + if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. } } skip_sad = *Data->iMinSAD; -// DIRECT MODE DELTA VECTOR SEARCH. +// DIRECT MODE DELTA VECTOR SEARCH. // This has to be made more effective, but at the moment I'm happy it's running at all if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; - (*MainSearchPtr)(0, 0, Data, 255); + MainSearchPtr(0, 0, Data, 255); 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) - pMB->mode = MODE_DIRECT; + if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation pMB->pmvs[3] = *Data->currentMV; @@ -1491,13 +1846,12 @@ return skip_sad; } - -static __inline void -SearchInterpolate(const uint8_t * const f_Ref, +static void +SearchInterpolate(const IMAGE * const f_Ref, const uint8_t * const f_RefH, const uint8_t * const f_RefV, const uint8_t * const f_RefHV, - const uint8_t * const b_Ref, + const IMAGE * const b_Ref, const uint8_t * const b_RefH, const uint8_t * const b_RefV, const uint8_t * const b_RefHV, @@ -1515,35 +1869,36 @@ { - 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; - fData->qpel_precision = bData.qpel_precision = 0; bData.qpel = fData->qpel; - bData.rounding = 0; + i = (x + y * fData->iEdgedWidth) * 16; + bData.bRef = fData->Ref = f_Ref->y + 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->y + i; + bData.RefH = fData->bRefH = b_RefH + i; + bData.RefV = fData->bRefV = b_RefV + i; + bData.RefHV = fData->bRefHV = b_RefHV + i; + bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; + 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_dx; @@ -1557,8 +1912,7 @@ CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); -//diamond. I wish we could use normal mainsearch functions (square, advdiamond) - +//diamond do { iDirection = 255; // forward MV moves @@ -1579,21 +1933,24 @@ } while (!(iDirection)); +//qpel refinement if (fData->qpel) { + if (*fData->iMinSAD > *best_sad + 500) return; CheckCandidate = CheckCandidateInt; fData->qpel_precision = bData.qpel_precision = 1; - get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 0); - get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 0); + 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; SubpelRefine(fData); + if (*fData->iMinSAD > *best_sad + 300) return; fData->currentQMV[2] = fData->currentQMV[0]; SubpelRefine(&bData); } - - *fData->iMinSAD += 2 * fData->lambda16; // two bits are needed to code interpolate mode. + + *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. if (*fData->iMinSAD < *best_sad) { *best_sad = *fData->iMinSAD; @@ -1635,16 +1992,15 @@ 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; VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ const int32_t TRB = time_pp - time_bp; const int32_t TRD = time_pp; - uint8_t * qimage; // some pre-inintialized data for the rest of the search @@ -1652,19 +2008,18 @@ int32_t iMinSAD; VECTOR currentMV[3]; VECTOR currentQMV[3]; + int32_t temp[8]; + 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] + 2; + Data.lambda16 = lambda_vec16[frame->quant]; Data.qpel = pParam->m_quarterpel; Data.rounding = 0; + Data.chroma = frame->motion_flags & PMV_CHROMA8; + Data.temp = temp; - 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; - + Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) // note: i==horizontal, j==vertical for (j = 0; j < pParam->mb_height; j++) { @@ -1682,6 +2037,8 @@ } Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; + Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; + Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; pMB->quant = frame->quant; /* direct search comes first, because it (1) checks for SKIP-mode @@ -1698,9 +2055,9 @@ &Data); if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } - + // forward search - SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, &frame->image, i, j, frame->motion_flags, frame->fcode, pParam, @@ -1708,7 +2065,7 @@ MODE_FORWARD, &Data); // backward search - SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, &frame->image, i, j, frame->motion_flags, frame->bcode, pParam, @@ -1716,9 +2073,8 @@ MODE_BACKWARD, &Data); // interpolate search comes last, because it uses data from forward and backward as prediction - - SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, + b_ref, b_refH->y, b_refV->y, b_refHV->y, &frame->image, i, j, frame->fcode, frame->bcode, @@ -1728,26 +2084,24 @@ 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, i, j, &Data); + switch (pMB->mode) { case MODE_FORWARD: f_count++; - if (pParam->m_quarterpel) f_predMV = pMB->qmvs[0]; - else f_predMV = pMB->mvs[0]; + f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; break; case MODE_BACKWARD: b_count++; - if (pParam->m_quarterpel) b_predMV = pMB->b_qmvs[0]; - else b_predMV = pMB->b_mvs[0]; + b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; break; case MODE_INTERPOLATE: i_count++; - if (pParam->m_quarterpel) { - f_predMV = pMB->qmvs[0]; - b_predMV = pMB->b_qmvs[0]; - } else { - f_predMV = pMB->mvs[0]; - b_predMV = pMB->b_mvs[0]; - } + f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; + b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; break; case MODE_DIRECT: case MODE_DIRECT_NO4V: @@ -1757,296 +2111,77 @@ } } } - 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; - - 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; - 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 (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); - else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); - - for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; - - if (pMB->dquant != NO_CHANGE) inter4v = 0; - - if (inter4v || 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) 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)) { - get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, - pParam->width, pParam->height, Data->iFcode, 0); - Data->qpel_precision = 1; - SubpelRefine(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 sumx, sumy, dx, dy; - - if(pParam->m_quarterpel) { - 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 (!(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->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; - } - } 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 +static __inline void MEanalyzeMB ( const uint8_t * const pRef, const uint8_t * const pCur, const int x, const int y, const MBParam * const pParam, - const MACROBLOCK * const pMBs, - MACROBLOCK * const pMB, + MACROBLOCK * const pMBs, SearchData * const Data) { - int i = 255, mask; + int i, mask; VECTOR pmv[3]; - *(Data->iMinSAD) = MV_MAX_ERROR; + MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; + + for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; //median is only used as prediction. it doesn't have to be real if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; else if (x == 1) //left macroblock does not have any vector now Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median - else if (y == 1) // top macroblock don't have it's vector + else if (y == 1) // top macroblock doesn't have it's vector Data->predMV = (pMB - 1)->mvs[0]; // left instead of median else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, - pParam->width, pParam->height, Data->iFcode, pParam->m_quarterpel); + pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv); Data->Cur = pCur + (x + y * pParam->edged_width) * 16; Data->Ref = pRef + (x + y * pParam->edged_width) * 16; - + pmv[1].x = EVEN(pMB->mvs[0].x); pmv[1].y = EVEN(pMB->mvs[0].y); pmv[2].x = EVEN(Data->predMV.x); pmv[2].y = EVEN(Data->predMV.y); pmv[0].x = pmv[0].y = 0; - (*CheckCandidate)(0, 0, 255, &i, Data); + CheckCandidate32I(0, 0, 255, &i, Data); -//early skip for 0,0 - if (*Data->iMinSAD < MAX_SAD00_FOR_SKIP * 4) { - pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; - pMB->mode = MODE_NOT_CODED; - return 0; - } + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { - if (!(mask = make_mask(pmv, 1))) - (*CheckCandidate)(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); + if (!(mask = make_mask(pmv, 1))) + CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); + if (!(mask = make_mask(pmv, 2))) + CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); - if (*Data->iMinSAD > MAX_SAD00_FOR_SKIP * 4) // diamond only if needed - DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed + DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); - pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; - pMB->mode = MODE_INTER; - return *(Data->iMinSAD); + for (i = 0; i < 4; i++) { + MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width]; + MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; + MB->mode = MODE_INTER; + MB->sad16 = Data->iMinSAD[i+1]; + } + } } -#define INTRA_THRESH 1350 -#define INTER_THRESH 1200 - +#define INTRA_BIAS 2500 +#define INTRA_THRESH 1500 +#define INTER_THRESH 1400 int -MEanalysis( const IMAGE * const pRef, +MEanalysis( const IMAGE * const pRef, FRAMEINFO * const Current, MBParam * const pParam, int maxIntra, //maximum number if non-I frames int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame - int bCount) // number if B frames in a row + int bCount) // number of B frames in a row { uint32_t x, y, intra = 0; int sSAD = 0; @@ -2054,171 +2189,520 @@ const IMAGE * const pCurrent = &Current->image; int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; - VECTOR currentMV; - int32_t iMinSAD; + int32_t iMinSAD[5], temp[5]; + VECTOR currentMV[5]; SearchData Data; Data.iEdgedWidth = pParam->edged_width; - Data.currentMV = ¤tMV; - Data.iMinSAD = &iMinSAD; + Data.currentMV = currentMV; + Data.iMinSAD = iMinSAD; Data.iFcode = Current->fcode; - CheckCandidate = CheckCandidate16no4vI; + Data.rrv = Current->global_flags & XVID_REDUCED; + Data.temp = temp; + CheckCandidate = CheckCandidate32I; - if (intraCount < 10) // we're right after an I frame + if (intraCount != 0 && intraCount < 10) // we're right after an I frame IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); else if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; - InterThresh += 400 * (1 - bCount); - if (InterThresh < 200) InterThresh = 200; + if (InterThresh < 300) InterThresh = 300; if (sadInit) (*sadInit) (); - for (y = 1; y < pParam->mb_height-1; y++) { - for (x = 1; x < pParam->mb_width-1; x++) { - int sad, dev; - MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; - - sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, - pParam, pMBs, pMB, &Data); - - if (sad > IntraThresh) { - dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, - pParam->edged_width); - if (dev + IntraThresh < sad) { - pMB->mode = MODE_INTRA; - if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame + for (y = 1; y < pParam->mb_height-1; y += 2) { + for (x = 1; x < pParam->mb_width-1; x += 2) { + int i; + + if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; + + MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); + + for (i = 0; i < 4; i++) { + int dev; + MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; + if (pMB->sad16 > IntraThresh) { + dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, + pParam->edged_width); + if (dev + IntraThresh < pMB->sad16) { + pMB->mode = MODE_INTRA; + if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; + } } + sSAD += pMB->sad16; } - sSAD += sad; } } sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); - if (sSAD > InterThresh ) return 1; //P frame +// if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP; + if (sSAD > InterThresh ) return P_VOP; emms(); - return 0; // B frame + return B_VOP; } -int -FindFcode( const MBParam * const pParam, - const FRAMEINFO * const current) + +static WARPPOINTS +GlobalMotionEst(const MACROBLOCK * const pMBs, + const MBParam * const pParam, + const FRAMEINFO * const current, + const FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV ) { - uint32_t x, y; - int max = 0, min = 0, i; - 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; + const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs + const int deltay=8; + const int grad=512; // lower bound for deviation in MB - if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; - if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; - } + WARPPOINTS gmc; + + uint32_t mx, my; + + int MBh = pParam->mb_height; + int MBw = pParam->mb_width; + + int *MBmask= calloc(MBh*MBw,sizeof(int)); + double DtimesF[4] = { 0.,0., 0., 0. }; + double sol[4] = { 0., 0., 0., 0. }; + double a,b,c,n,denom; + double meanx,meany; + int num,oldnum; + + if (!MBmask) { fprintf(stderr,"Mem error\n"); return gmc;} + +// filter mask of all blocks + + for (my = 1; my < MBh-1; my++) + for (mx = 1; mx < MBw-1; mx++) + { + const int mbnum = mx + my * MBw; + const MACROBLOCK *pMB = &pMBs[mbnum]; + const VECTOR mv = pMB->mvs[0]; + + if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) + continue; + + if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) ) + && ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) ) + && ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) ) + && ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) ) + MBmask[mbnum]=1; + } + + for (my = 1; my < MBh-1; my++) + for (mx = 1; mx < MBw-1; mx++) + { + const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; + + const int mbnum = mx + my * MBw; + if (!MBmask[mbnum]) + continue; + + if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= grad ) + MBmask[mbnum] = 0; + if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= grad ) + MBmask[mbnum] = 0; + + } + + emms(); + + do { /* until convergence */ + + a = b = c = n = 0; + DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; + for (my = 0; my < MBh; my++) + for (mx = 0; mx < MBw; mx++) + { + const int mbnum = mx + my * MBw; + const MACROBLOCK *pMB = &pMBs[mbnum]; + const VECTOR mv = pMB->mvs[0]; + + if (!MBmask[mbnum]) + continue; + + n++; + a += 16*mx+8; + b += 16*my+8; + c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); + + DtimesF[0] += (double)mv.x; + DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); + DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); + DtimesF[3] += (double)mv.y; + } + + denom = a*a+b*b-c*n; + +/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ +/* D'*E*F has been calculated in the same loop as matrix */ + + sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; + sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; + sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; + sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; + + sol[0] /= denom; + sol[1] /= denom; + sol[2] /= denom; + sol[3] /= denom; + + meanx = meany = 0.; + oldnum = 0; + for (my = 0; my < MBh; my++) + for (mx = 0; mx < MBw; mx++) + { + const int mbnum = mx + my * MBw; + const MACROBLOCK *pMB = &pMBs[mbnum]; + const VECTOR mv = pMB->mvs[0]; + + if (!MBmask[mbnum]) + continue; + + oldnum++; + meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); + meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ); } + + if (4*meanx > oldnum) /* better fit than 0.25 is useless */ + meanx /= oldnum; + else + meanx = 0.25; + + if (4*meany > oldnum) + meany /= oldnum; + else + meany = 0.25; + +/* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]); + fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum); +*/ + num = 0; + for (my = 0; my < MBh; my++) + for (mx = 0; mx < MBw; mx++) + { + const int mbnum = mx + my * MBw; + const MACROBLOCK *pMB = &pMBs[mbnum]; + const VECTOR mv = pMB->mvs[0]; + + if (!MBmask[mbnum]) + continue; + + if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) + || ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) + MBmask[mbnum]=0; + else + num++; + } + + } while ( (oldnum != num) && (num>=4) ); + + if (num < 4) + { + gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0; + } else { + + gmc.duv[0].x=(int)(sol[0]+0.5); + gmc.duv[0].y=(int)(sol[3]+0.5); + + gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); + gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); + + gmc.duv[2].x=0; + gmc.duv[2].y=0; } +// fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y); - min = -min; - max += 1; - if (min > max) max = min; - if (pParam->m_quarterpel) max *= 2; + free(MBmask); - for (i = 1; (max > 32 << (i - 1)); i++); - return i; + return gmc; } -static void -CheckGMC(int x, int y, const int dir, int * iDirection, - const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, - const MBParam * const pParam) +// functions which perform BITS-based search/bitcount + +static int +CountMBBitsInter(SearchData * const Data, + const MACROBLOCK * const pMBs, const int x, const int y, + const MBParam * const pParam, + const uint32_t MotionFlags) { - uint32_t mx, my, a, count = 0; + int i, iDirection; + int32_t bsad[5]; + + CheckCandidate = CheckCandidateBits16; + + if (Data->qpel) { + for(i = 0; i < 5; i++) { + Data->currentMV[i].x = Data->currentQMV[i].x/2; + Data->currentMV[i].y = Data->currentQMV[i].y/2; + } + Data->qpel_precision = 1; + CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); + + //checking if this vector is perfect. if it is, we stop. + if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) + return 0; //quick stop + + if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search + for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); + Data->qpel_precision = 0; + if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) + CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); + } + + } else { // not qpel + + CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); + //checking if this vector is perfect. if it is, we stop. + if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) { + return 0; //inter + } + } + + if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); + + if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data); + + if (Data->qpel) { + if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search + for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { + Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match + Data->currentQMV[i].y = 2 * Data->currentMV[i].y; + } - for (my = 1; my < pParam->mb_height-1; my++) - for (mx = 1; mx < pParam->mb_width-1; mx++) { - VECTOR mv; - const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; - if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; - mv = pMB->mvs[0]; - a = ABS(mv.x - x) + ABS(mv.y - y); - if (a < 6) count += 6 - a; + // preparing for qpel-precision search + 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, 1, 0); } + if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data); + } - if (count > *bestcount) { - *bestcount = count; - *iDirection = dir; - GMC->x = x; GMC->y = y; + if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction + VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; + if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) + CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); } + return Data->iMinSAD[0]; } -static VECTOR -GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) +static int +CountMBBitsInter4v(const SearchData * const Data, + MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, + const int x, const int y, + const MBParam * const pParam, const uint32_t MotionFlags, + const VECTOR * const backup) { - uint32_t count, bestcount = 0; - int x, y; - VECTOR gmc = {0,0}; - int step, min_x, max_x, min_y, max_y; - uint32_t mx, my; - int iDirection, bDirection; + int cbp = 0, bits = 0, t = 0, i, iDirection; + SearchData Data2, *Data8 = &Data2; + int sumx = 0, sumy = 0; + int16_t in[64], coeff[64]; - min_x = min_y = -32<iMinSAD = Data->iMinSAD + i + 1; + Data8->currentMV = Data->currentMV + i + 1; + Data8->currentQMV = Data->currentQMV + i + 1; + Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->Ref = Data->Ref + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->RefH = Data->RefH + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->RefV = Data->RefV + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->RefHV = Data->RefHV + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + + if(Data->qpel) { + Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); + if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y, + Data8->predMV, Data8->iFcode, 0, 0); + } else { + Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i); + if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y, + Data8->predMV, Data8->iFcode, 0, 0); + } + + get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, + pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); + + *Data8->iMinSAD += t; + + Data8->qpel_precision = Data8->qpel; + // checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) + if (Data8->qpel) { + if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y)) + CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); + } else { + if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y)) + CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); + } + + if (Data8->qpel) { + if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows + int32_t s = *Data8->iMinSAD; + Data8->currentMV->x = Data8->currentQMV->x/2; + Data8->currentMV->y = Data8->currentQMV->y/2; + Data8->qpel_precision = 0; + get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, + pParam->width, pParam->height, Data8->iFcode - 1, 0, 0); + + if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) + CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); + + if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS) + SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); + + if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); + + if(s > *Data8->iMinSAD) { //we have found a better match + Data8->currentQMV->x = 2*Data8->currentMV->x; + Data8->currentQMV->y = 2*Data8->currentMV->y; + } + + Data8->qpel_precision = 1; + get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, + pParam->width, pParam->height, Data8->iFcode, 1, 0); -//step1: let's find a rough camera panning - for (step = 32; step >= 2; step /= 2) { - bestcount = 0; - for (y = min_y; y <= max_y; y += step) - for (x = min_x ; x <= max_x; x += step) { - count = 0; - //for all macroblocks - for (my = 1; my < pParam->mb_height-1; my++) - for (mx = 1; mx < pParam->mb_width-1; mx++) { - const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; - VECTOR mv; - - if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) - continue; - - mv = pMB->mvs[0]; - if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ - count++; - } - if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } } - min_x = gmc.x - step; - max_x = gmc.x + step; - min_y = gmc.y - step; - max_y = gmc.y + step; - - } - - if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) - gmc.x = gmc.y = 0; //no camara pan, no GMC - -// step2: let's refine camera panning using gradiend-descent approach. -// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) - bestcount = 0; - CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); - do { - x = gmc.x; y = gmc.y; - bDirection = iDirection; iDirection = 0; - if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); - if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); - if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); - if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); + if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8); - } while (iDirection); + } else // not qpel + if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement - if (pParam->m_quarterpel) { - gmc.x *= 2; - gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ + //checking vector equal to predicion + if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) { + const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; + if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y)) + CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); + } + + bits += *Data8->iMinSAD; + if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V + + // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else + if(Data->qpel) { + pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; + pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; + pMB->qmvs[i] = *Data8->currentQMV; + sumx += Data8->currentQMV->x/2; + sumy += Data8->currentQMV->y/2; + } else { + pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x; + pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y; + sumx += Data8->currentMV->x; + sumy += Data8->currentMV->y; + } + pMB->mvs[i] = *Data8->currentMV; + pMB->sad8[i] = 4 * *Data8->iMinSAD; + if (Data8->temp[0]) cbp |= 1 << (5 - i); + } + + if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma + const uint8_t * ptr; + sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; + sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; + + //chroma U + ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCU, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); + transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); + fdct(in); + if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); + else i = quant4_inter(coeff, in, Data->lambda16); + if (i > 0) { + bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + cbp |= 1 << (5 - 4); + } + + if (bits < *Data->iMinSAD) { // still possible + //chroma V + ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCV, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); + transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); + fdct(in); + if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); + else i = quant4_inter(coeff, in, Data->lambda16); + if (i > 0) { + bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + cbp |= 1 << (5 - 5); + } + bits += cbpy_tab[15-(cbp>>2)].len; + bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; + } } - return gmc; + return bits; +} + + +static int +CountMBBitsIntra(const SearchData * const Data) +{ + int bits = 1; //this one is ac/dc prediction flag. always 1. + int cbp = 0, i, t, dc = 0, b_dc = 1024; + const uint32_t iQuant = Data->lambda16; + int16_t in[64], coeff[64]; + + for(i = 0; i < 4; i++) { + uint32_t iDcScaler = get_dc_scaler(iQuant, 1); + + int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); + transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); + fdct(in); + b_dc = dc; + dc = in[0]; + in[0] -= b_dc; + if (Data->lambda8 == 0) quant_intra_c(coeff, in, iQuant, iDcScaler); + else quant4_intra_c(coeff, in, iQuant, iDcScaler); + + b_dc = dc; + dc = coeff[0]; + if (i != 0) coeff[0] -= b_dc; + + bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;; + Data->temp[i] = t; + if (t != 0) cbp |= 1 << (5 - i); + if (bits >= Data->iMinSAD[0]) break; + } + + if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma + uint32_t iDcScaler = get_dc_scaler(iQuant, 0); + //chroma U + transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); + fdct(in); + in[0] -= 1024; + if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); + else quant4_intra(coeff, in, iQuant, iDcScaler); + + bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; + if (t != 0) cbp |= 1 << (5 - 4); + Data->temp[4] = t; + + if (bits < Data->iMinSAD[0]) { + //chroma V + transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); + fdct(in); + in[0] -= 1024; + if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); + else quant4_intra(coeff, in, iQuant, iDcScaler); + + bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; + if (t != 0) cbp |= 1 << (5 - 5); + + Data->temp[5] = t; + + bits += t = cbpy_tab[cbp>>2].len; + Data->temp[6] = t; + + bits += t = mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; + Data->temp[7] = t; + + } + } + + return bits; }