--- branches/dev-api-4/xvidcore/src/motion/motion_est.c 2003/03/26 14:56:49 949 +++ branches/dev-api-4/xvidcore/src/motion/motion_est.c 2003/06/09 13:55:56 1054 @@ -1,38 +1,35 @@ -/************************************************************************** +/***************************************************************************** + * + * XVID MPEG-4 VIDEO CODEC + * - Motion Estimation related code - * - * XVID MPEG-4 VIDEO CODEC - * motion estimation + * Copyright(C) 2002 Christoph Lampert + * 2002 Michael Militzer + * 2002-2003 Radoslaw Czyz * - * 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 - * to use this software module in hardware or software products are - * advised that its use may infringe existing patents or copyrights, and - * any such use would be at such party's own risk. The original - * developer of this software module and his/her company, and subsequent - * editors and their companies, will have no liability for use of this - * software or modifications or derivatives thereof. + * This program is free software ; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation ; either version 2 of the License, or + * (at your option) any later version. * - * This program is free software; you can redistribute it and/or modify - * it under the terms of the GNU General Public License as published by - * the Free Software Foundation; either version 2 of the License, or - * (at your option) any later version. + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY ; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. * - * This program is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * You should have received a copy of the GNU General Public License + * along with this program ; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA * - * You should have received a copy of the GNU General Public License - * along with this program; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * $Id: motion_est.c,v 1.58.2.17 2003-06-09 13:54:37 edgomez Exp $ * - *************************************************************************/ + ****************************************************************************/ #include #include #include -#include // memcpy -#include // lrint +#include /* memcpy */ +#include /* lrint */ #include "../encoder.h" #include "../utils/mbfunctions.h" @@ -81,33 +78,32 @@ 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; - x = qpel ? x<<1 : x; - y = qpel ? y<<1 : y; + int bits; + const int q = (1 << (iFcode - 1)) - 1; + + x <<= qpel; + y <<= qpel; if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } x -= pred.x; + bits = (x != 0 ? iFcode:0); + x = abs(x); + x += q; + x >>= (iFcode - 1); + bits += mvtab[x]; + y -= pred.y; + bits += (y != 0 ? iFcode:0); + y = abs(y); + y += q; + y >>= (iFcode - 1); + bits += mvtab[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) { - 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; + return bits; } -static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data) +static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by, + const SearchData * const data) { int sad; const uint32_t stride = data->iEdgedWidth/2; @@ -115,50 +111,44 @@ * f_refv = data->RefQ + 8, * b_refu = data->RefQ + 16, * b_refv = data->RefQ + 24; + int offset = (fx>>1) + (fy>>1)*stride; 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; + f_refu = (uint8_t*)data->RefP[4] + offset; + f_refv = (uint8_t*)data->RefP[5] + offset; 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); + interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding); + interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, 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); + interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding); + interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, 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); + interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding); + interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding); break; } + offset = (bx>>1) + (by>>1)*stride; 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; + b_refu = (uint8_t*)data->b_RefP[4] + offset; + b_refv = (uint8_t*)data->b_RefP[5] + offset; 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); + interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding); + interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, 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); + interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding); + interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, 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); + interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding); + interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding); break; } @@ -168,77 +158,64 @@ return sad; } - static int32_t -ChromaSAD(int dx, int dy, const SearchData * const data) +ChromaSAD(const int dx, const int dy, const SearchData * const data) { int sad; const uint32_t stride = data->iEdgedWidth/2; + int offset = (dx>>1) + (dy>>1)*stride; - 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 + if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */ + data->temp[5] = dx; data->temp[6] = dy; /* backup */ switch (((dx & 1) << 1) | (dy & 1)) { case 0: - 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); + sad = sad8(data->CurU, data->RefP[4] + offset, stride); + sad += sad8(data->CurV, data->RefP[5] + offset, stride); break; case 1: - dx = dx / 2; dy = (dy - 1) / 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); + sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride); + sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride); break; case 2: - dx = (dx - 1) / 2; dy = dy / 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); + sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride); + sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride); break; default: - dx = (dx - 1) / 2; dy = (dy - 1) / 2; - interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); + interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding); sad = sad8(data->CurU, data->RefQ, stride); - interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); + interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding); sad += sad8(data->CurV, data->RefQ, stride); break; } - data->temp[7] = sad; //backup, part 2 + data->temp[7] = sad; /* backup, part 2 */ return sad; } static __inline const uint8_t * 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) ) { - case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); - case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); - case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); - case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); - case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); - case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); - case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); - default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); - } + /* dir : 0 = forward, 1 = backward */ + const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP ); + const int picture = ((x&1)<<1) | (y&1); + const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; + return direction[picture] + offset; } -// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate +/* 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 - } + const int picture = ((x&1)<<1) | (y&1); + const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; + return data->RefP[picture] + offset; } 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 + /* create or find a qpel-precision reference picture; return pointer to it */ uint8_t * Reference = data->RefQ + 16*dir; const uint32_t iEdgedWidth = data->iEdgedWidth; const uint32_t rounding = data->rounding; @@ -249,32 +226,32 @@ 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 - return (uint8_t *) ref1; + 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); + ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); break; - case 1: // x halfpel, y qpel - top or bottom during qpel refinement + case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ 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 + case 2: /* x qpel, y halfpel - left or right during qpel refinement */ 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 = 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; - interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); - break; + default: /* pure halfpel position */ + return (uint8_t *) ref1; + } return Reference; } @@ -282,7 +259,7 @@ 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 + /* create or find a qpel-precision reference picture; return pointer to it */ uint8_t * Reference = data->RefQ + 16*dir; const uint32_t iEdgedWidth = data->iEdgedWidth; const uint32_t rounding = data->rounding; @@ -292,8 +269,11 @@ ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); switch( ((x&1)<<1) + (y&1) ) { - case 3: // x and y in qpel resolution - the "corners" (top left/right and - // bottom left/right) during qpel refinement + 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); @@ -303,7 +283,7 @@ 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 + case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ 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); @@ -311,7 +291,7 @@ interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); break; - case 2: // x qpel, y halfpel - left or right during qpel refinement + case 2: /* x qpel, y halfpel - left or right during qpel refinement */ 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); @@ -319,7 +299,7 @@ interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); break; - case 0: // pure halfpel position + default: /* pure halfpel position */ return (uint8_t *) ref1; } return Reference; @@ -342,9 +322,9 @@ Reference = GetReference(x, y, data); current = data->currentMV; xc = x; yc = y; - } else { // x and y are in 1/4 precision + } else { /* x and y are in 1/4 precision */ Reference = Interpolate16x16qpel(x, y, 0, data); - xc = x/2; yc = y/2; //for chroma sad + xc = x/2; yc = y/2; /* for chroma sad */ current = data->currentQMV; } @@ -355,7 +335,7 @@ data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], - (yc >> 1) + roundtab_79[yc & 0x3], data); + (yc >> 1) + roundtab_79[yc & 0x3], data); if (sad < data->iMinSAD[0]) { data->iMinSAD[0] = sad; @@ -371,7 +351,6 @@ data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } if (data->temp[4] < data->iMinSAD[4]) { data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } - } static void @@ -379,12 +358,18 @@ { int32_t sad; uint32_t t; const uint8_t * Reference; + VECTOR * current; 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); + if (!data->qpel_precision) { + Reference = GetReference(x, y, data); + current = data->currentMV; + } else { /* x and y are in 1/4 precision */ + Reference = Interpolate8x8qpel(x, y, 0, 0, data); + current = data->currentQMV; + } sad = sad8(data->Cur, Reference, data->iEdgedWidth); t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); @@ -393,19 +378,18 @@ if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; - data->currentMV->x = x; data->currentMV->y = y; + current->x = x; current->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 + if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */ (x > data->max_dx) || (x < data->min_dx) || (y > data->max_dy) || (y < data->min_dy) ) return; @@ -440,12 +424,12 @@ 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->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 + if (data->qpel_precision) { /* x and y are in 1/4 precision */ Reference = Interpolate16x16qpel(x, y, 0, data); current = data->currentQMV; xc = x/2; yc = y/2; @@ -473,14 +457,14 @@ 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 + /* maximum speed - for P/B/I decision */ int32_t sad; if ( (x > data->max_dx) || (x < data->min_dx) || (y > data->max_dy) || (y < data->min_dy) ) return; - sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), - data->iEdgedWidth, data->temp+1); + sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*(data->iEdgedWidth), + data->iEdgedWidth, data->temp+1); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; @@ -506,8 +490,9 @@ 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) { ReferenceF = GetReference(xf, yf, data); @@ -552,7 +537,7 @@ 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; @@ -565,10 +550,11 @@ 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 ((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; @@ -576,7 +562,7 @@ } 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 + mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */ b_mvs.x *= 2; b_mvs.y *= 2; } @@ -610,7 +596,7 @@ 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) ? @@ -622,10 +608,10 @@ 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 ( (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); @@ -659,8 +645,8 @@ CheckCandidateBits16(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 bits = 0, sum; + int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; + int32_t bits = 0; VECTOR * current; const uint8_t * ptr; int i, cbp = 0, t, xc, yc; @@ -672,7 +658,7 @@ ptr = GetReference(x, y, data); current = data->currentMV; xc = x; yc = y; - } else { // x and y are in 1/4 precision + } else { /* x and y are in 1/4 precision */ ptr = Interpolate16x16qpel(x, y, 0, data); current = data->currentQMV; xc = x/2; yc = y/2; @@ -681,48 +667,31 @@ 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]); - } - } + bits += data->temp[i] = Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, i); } - bits += cbpy_tab[15-(cbp>>2)].len; - bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; + bits += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); + + bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; + + if (bits >= data->iMinSAD[0]) return; + + /* chroma */ + xc = (xc >> 1) + roundtab_79[xc & 0x3]; + yc = (yc >> 1) + roundtab_79[yc & 0x3]; + + /* chroma U */ + ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); + transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); + bits += Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 4); + if (bits >= data->iMinSAD[0]) return; + + /* chroma V */ + ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); + transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2); + bits += Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 5); + + bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; if (bits < data->iMinSAD[0]) { data->iMinSAD[0] = bits; @@ -744,11 +713,11 @@ 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; + int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; + int32_t bits; VECTOR * current; const uint8_t * ptr; - int cbp; + int cbp = 0; if ( (x > data->max_dx) || (x < data->min_dx) || (y > data->max_dy) || (y < data->min_dy) ) return; @@ -756,21 +725,14 @@ if (!data->qpel_precision) { ptr = GetReference(x, y, data); current = data->currentMV; - } else { // x and y are in 1/4 precision + } 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); + bits = Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 5); + bits += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); if (bits < data->iMinSAD[0]) { data->temp[0] = cbp; @@ -792,7 +754,7 @@ int iDirection; - for(;;) { //forever + for(;;) { /* forever */ iDirection = 0; if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); @@ -801,14 +763,14 @@ /* now we're doing diagonal checks near our candidate */ - if (iDirection) { //if anything found + if (iDirection) { /* if anything found */ bDirection = iDirection; iDirection = 0; x = data->currentMV->x; y = data->currentMV->y; - if (bDirection & 3) { //our candidate is left or right + 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 + } else { /* what remains here is up or down */ CHECK_CANDIDATE(x + iDiamondSize, y, 2); CHECK_CANDIDATE(x - iDiamondSize, y, 1); } @@ -817,7 +779,7 @@ bDirection += iDirection; x = data->currentMV->x; y = data->currentMV->y; } - } else { //about to quit, eh? not so fast.... + } else { /* about to quit, eh? not so fast.... */ switch (bDirection) { case 2: CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); @@ -855,14 +817,14 @@ 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 + 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 + if (!iDirection) break; /* ok, the end. really */ bDirection = iDirection; x = data->currentMV->x; y = data->currentMV->y; } @@ -907,14 +869,14 @@ /* now we're doing diagonal checks near our candidate */ - if (iDirection) { //checking if anything found + 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 + 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 + } else { /* what remains here is up or down */ CHECK_CANDIDATE(x + iDiamondSize, y, 2); CHECK_CANDIDATE(x - iDiamondSize, y, 1); } @@ -932,7 +894,7 @@ { /* Do a half-pel or q-pel refinement */ const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; - int iDirection; //only needed because macro expects it + int iDirection; /* only needed because macro expects it */ CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); @@ -950,21 +912,22 @@ const uint32_t stride, const uint32_t iQuant, int rrv) { + int offset = (x + y*stride)*8; if(!rrv) { - uint32_t sadC = sad8(current->u + x*8 + y*stride*8, - reference->u + x*8 + y*stride*8, stride); + uint32_t sadC = sad8(current->u + offset, + reference->u + offset, 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); + sadC += sad8(current->v + offset, + reference->v + offset, stride); if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; return 1; } else { - uint32_t sadC = sad16(current->u + x*16 + y*stride*16, - reference->u + x*16 + y*stride*16, stride, 256*4096); + uint32_t sadC = sad16(current->u + 2*offset, + reference->u + 2*offset, 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); + sadC += sad16(current->v + 2*offset, + reference->v + 2*offset, stride, 256*4096); if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; return 1; } @@ -979,6 +942,122 @@ pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; } +static __inline void +ModeDecision(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 uint32_t VopFlags, + const uint32_t VolFlags, + const IMAGE * const pCurrent, + const IMAGE * const pRef) +{ + int mode = MODE_INTER; + int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); + const uint32_t iQuant = pMB->quant; + + const int skip_possible = (!(VolFlags & XVID_VOL_GMC)) && (pMB->dquant == 0); + + if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) { /* normal, fast, SAD-based mode decision */ + 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 = MODE_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; + } + + /* final skip decision, a.k.a. "the vector you found, really that good?" */ + if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) + if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) + if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { + mode = MODE_NOT_CODED; + sad = 0; + } + + /* 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; /* dev8(chroma) ??? */ + 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+16, Data->iEdgedWidth) + + dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + + dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); + + if (deviation < (sad - InterBias)) mode = MODE_INTRA; + } + + } else { /* BITS */ + + int bits, intra, i; + VECTOR backup[5], *v; + Data->iQuant = iQuant; + + 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) + mode = MODE_INTER; /* quick stop */ + else { + if (inter4v) { + int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); + if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; } + } + + intra = CountMBBitsIntra(Data); + + if (intra < bits) { *Data->iMinSAD = bits = intra; mode = MODE_INTRA; } + } + } + + 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 (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; + 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 if (mode == MODE_INTER4V) + pMB->sad16 = Data->iMinSAD[0]; + else /* INTRA, NOT_CODED */ + SkipMacroblockP(pMB, 0); + + pMB->mode = mode; +} + bool MotionEstimation(MBParam * const pParam, FRAMEINFO * const current, @@ -999,13 +1078,18 @@ uint32_t x, y; uint32_t iIntra = 0; - int32_t sad00; + int32_t quant = current->quant, sad00; + int skip_thresh = \ + INITIAL_SKIP_THRESH * \ + (current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \ + (current->vop_flags & XVID_VOP_MODEDECISION_BITS ? 2:1); - // some pre-initialized thingies for SearchP + /* some pre-initialized thingies for SearchP */ int32_t temp[8]; VECTOR currentMV[5]; VECTOR currentQMV[5]; int32_t iMinSAD[5]; + DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE); SearchData Data; memset(&Data, 0, sizeof(SearchData)); Data.iEdgedWidth = iEdgedWidth; @@ -1015,9 +1099,11 @@ Data.temp = temp; Data.iFcode = current->fcode; Data.rounding = pParam->m_rounding_type; - Data.qpel = current->vol_flags & XVID_VOL_QUARTERPEL; + Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0); Data.chroma = MotionFlags & XVID_ME_CHROMA16; - Data.rrv = current->vop_flags & XVID_VOP_REDUCED; + Data.rrv = (current->vop_flags & XVID_VOP_REDUCED ? 1:0); + Data.dctSpace = dct_space; + Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); if ((current->vop_flags & XVID_VOP_REDUCED)) { mb_width = (pParam->width + 31) / 32; @@ -1025,7 +1111,7 @@ Data.qpel = 0; } - Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) + Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */ if (sadInit) (*sadInit) (); for (y = 0; y < mb_height; y++) { @@ -1052,10 +1138,17 @@ sad00 = pMB->sad16; -//initial skip decision -/* no early skip for GMC (global vector = skip vector is unknown!) */ + if (pMB->dquant != 0) { + quant += DQtab[pMB->dquant]; + if (quant > 31) quant = 31; + else if (quant < 1) quant = 1; + } + pMB->quant = quant; + + /* initial skip decision */ + /* no early skip for GMC (global vector = skip vector is unknown!) */ if (!(current->vol_flags & XVID_VOL_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ - if (pMB->dquant == 0 && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) + if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh) if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { SkipMacroblockP(pMB, sad00); continue; @@ -1063,24 +1156,13 @@ } SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, - y, MotionFlags, current->vol_flags, pMB->quant, - &Data, pParam, pMBs, reference->mbs, - current->vop_flags & XVID_VOP_INTER4V, pMB); - -/* final skip decision, a.k.a. "the vector you found, really that good?" */ - if (!(current->vol_flags & XVID_VOL_GMC)) { - if ( pMB->dquant == 0 && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) { - if (!(current->vop_flags & XVID_VOP_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 + y, MotionFlags, current->vop_flags, current->vol_flags, + &Data, pParam, pMBs, reference->mbs, pMB); + + ModeDecision(&Data, pMB, pMBs, x, y, pParam, + MotionFlags, current->vop_flags, current->vol_flags, + pCurrent, pRef); - } - } - } if (pMB->mode == MODE_INTRA) if (++iIntra > iLimit) return 1; } @@ -1099,7 +1181,7 @@ { int mask = 255, j; for (j = 0; j < i; j++) { - if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already + 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; else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; @@ -1116,31 +1198,30 @@ 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 + /* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */ if (rrv) { iWcount /= 2; iHcount /= 2; } - if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour + if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */ pmv[5].x = EVEN(pmv[3].x); pmv[5].y = EVEN(pmv[3].y); } else pmv[5].x = pmv[5].y = 0; - if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour + 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 + 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; - // [1] median prediction + /* [1] median prediction */ pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); - pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask + pmv[0].x = pmv[0].y = 0; /* [0] is zero; not used in the loop (checked before) but needed here for make_mask */ - pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame + pmv[2].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)) { - pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame + pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */ pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); } else pmv[6].x = pmv[6].y = 0; @@ -1153,87 +1234,6 @@ } } -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 VopFlags) -{ - - int mode = MODE_INTER; - - if (!(VopFlags & XVID_VOP_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->vol_flags & XVID_VOL_MPEGQUANT)?1:0; - - 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 SearchP(const IMAGE * const pRef, const uint8_t * const pRefH, @@ -1244,44 +1244,41 @@ const int y, const uint32_t MotionFlags, const uint32_t VopFlags, - const uint32_t iQuant, + const uint32_t VolFlags, SearchData * const Data, const MBParam * const pParam, const MACROBLOCK * const pMBs, const MACROBLOCK * const prevMBs, - int inter4v, MACROBLOCK * const pMB) { int i, iDirection = 255, mask, threshA; VECTOR pmv[7]; + int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); 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); get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); - Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache + 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->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; + Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; - Data->lambda16 = lambda_vec16[iQuant]; - Data->lambda8 = lambda_vec8[iQuant]; + Data->lambda16 = lambda_vec16[pMB->quant]; + Data->lambda8 = lambda_vec8[pMB->quant]; Data->qpel_precision = 0; - if (pMB->dquant != 0) inter4v = 0; - - for(i = 0; i < 5; i++) - Data->currentMV[i].x = Data->currentMV[i].y = 0; + memset(Data->currentMV, 0, 5*sizeof(VECTOR)); if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); else Data->predMV = pmv[0]; @@ -1294,7 +1291,7 @@ Data->iMinSAD[4] = pMB->sad8[3]; if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) || (x | y)) { - threshA = Data->temp[0]; // that's where we keep this SAD atm + threshA = Data->temp[0]; /* that's where we keep this SAD atm */ if (threshA < 512) threshA = 512; else if (threshA > 1024) threshA = 1024; } else @@ -1305,7 +1302,7 @@ if (!Data->rrv) { if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; - else CheckCandidate = CheckCandidate16no4v; //for extra speed + else CheckCandidate = CheckCandidate16no4v; /* for extra speed */ } else CheckCandidate = CheckCandidate32; /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ @@ -1318,8 +1315,8 @@ 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))) { - if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) inter4v = 0; } + (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) + inter4v = 0; else { MainSearchFunc * MainSearchPtr; @@ -1365,32 +1362,27 @@ } if (MotionFlags & XVID_ME_HALFPELREFINE16) - if ((!(MotionFlags & XVID_ME_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].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */ Data->currentQMV[i].y = 2 * Data->currentMV[i].y; } - if (MotionFlags & XVID_ME_QUARTERPELREFINE16) - if ((!(MotionFlags & XVID_ME_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); - + if (Data->qpel) { + 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; + if (MotionFlags & XVID_ME_QUARTERPELREFINE16) SubpelRefine(Data); - } - - if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0; + } - if (inter4v && (!(VopFlags & XVID_VOP_MODEDECISION_BITS) || - (!(MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS)) || (!(MotionFlags & XVID_ME_HALFPELREFINE8_BITS)) || - ((!(MotionFlags & XVID_ME_EXTSEARCH_BITS)) && (!(MotionFlags&XVID_ME_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)pMB->quant * 30) + inter4v = 0; + if (inter4v) { SearchData Data8; - memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data + 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); @@ -1398,51 +1390,24 @@ Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); if ((Data->chroma) && (!(VopFlags & XVID_VOP_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 + /* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */ 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; - } + if (Data->qpel) + for (i = 1; i < 5; i++) { + sumx += Data->currentQMV[i].x/2; + sumy += Data->currentQMV[i].y/2; + } + else + for (i = 1; i < 5; i++) { + sumx += Data->currentMV[i].x; + sumy += Data->currentMV[i].y; + } Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], (sumy >> 3) + roundtab_76[sumy & 0xf], Data); } - } - - inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, VopFlags); - - if (Data->rrv) { - Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); - Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); - } - - if (inter4v == 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; - 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 if (inter4v == MODE_INTER4V) { - pMB->mode = MODE_INTER4V; - pMB->sad16 = Data->iMinSAD[0]; - } else { // INTRA mode - SkipMacroblockP(pMB, 0); // not skip, but similar enough - pMB->mode = MODE_INTRA; - } - + } else Data->iMinSAD[1] = 4096*256; } static void @@ -1473,14 +1438,15 @@ *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) { - if (Data->rrv) i = 2; else i = 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)); + if (Data->rrv) i = 16; else i = 8; - Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1)); + + Data->Cur = OldData->Cur + i * ((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, @@ -1490,7 +1456,7 @@ else CheckCandidate = CheckCandidate16no4v; if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_BITS))) { - int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD + int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ MainSearchFunc *MainSearchPtr; if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch; @@ -1500,18 +1466,18 @@ 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 + Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ Data->currentQMV->y = 2 * Data->currentMV->y; } } if (MotionFlags & XVID_ME_HALFPELREFINE8) { - int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD + int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ - SubpelRefine(Data); // perform halfpel refine of current best vector + SubpelRefine(Data); /* perform halfpel refine of current best vector */ - if(*(Data->iMinSAD) < temp_sad) { // we have found a better match - Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector + if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */ + Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ Data->currentQMV->y = 2 * Data->currentMV->y; } } @@ -1558,15 +1524,15 @@ const uint32_t mode_curr) { - // [0] is prediction + /* [0] is prediction */ pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); - pmv[1].x = pmv[1].y = 0; // [1] is zero + pmv[1].x = pmv[1].y = 0; /* [1] is zero */ pmv[2] = ChoosePred(pMB, mode_curr); pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); - if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour + 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); } else pmv[3].x = pmv[3].y = 0; @@ -1612,14 +1578,14 @@ *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->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */ - 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->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; + Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; + Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16; + Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16; + Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8; + Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8; Data->predMV = *predMV; @@ -1634,7 +1600,7 @@ Data->currentMV->x = Data->currentMV->y = 0; CheckCandidate = CheckCandidate16no4v; -// main loop. checking all predictions + /* main loop. checking all predictions */ for (i = 0; i < 7; i++) { if (!(mask = make_mask(pmv, i)) ) continue; CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); @@ -1657,7 +1623,7 @@ SubpelRefine(Data); } -// three bits are needed to code backward mode. four for forward + /* three bits are needed to code backward mode. four for forward */ if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; else *Data->iMinSAD += 3 * Data->lambda16; @@ -1681,7 +1647,7 @@ } if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; - else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search + else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */ } static void @@ -1697,13 +1663,13 @@ 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 + /* 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; + dx += Data->directmvF[k].x / div; + b_dy += Data->directmvB[k].y / div; + b_dx += Data->directmvB[k].x / div; } dy = (dy >> 3) + roundtab_76[dy & 0xf]; @@ -1716,14 +1682,20 @@ 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 + 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 + if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { + pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */ + for (k = 0; k < 4; k++) { + pMB->qmvs[k] = pMB->mvs[k]; + pMB->b_qmvs[k] = pMB->b_mvs[k]; + } + } } static __inline uint32_t @@ -1751,18 +1723,18 @@ 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; + Data->RefP[0] = f_Ref->y + k; + Data->RefP[2] = f_RefH + k; + Data->RefP[1] = f_RefV + k; + Data->RefP[3] = f_RefHV + k; + Data->b_RefP[0] = b_Ref->y + k; + Data->b_RefP[2] = b_RefH + k; + Data->b_RefP[1] = b_RefV + k; + Data->b_RefP[3] = b_RefHV + k; + Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; + Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; + Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; + Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; k = Data->qpel ? 4 : 2; Data->max_dx = k * (pParam->width - x * 16); @@ -1782,8 +1754,8 @@ 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" + *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 256*4096; } @@ -1800,22 +1772,25 @@ CheckCandidate(0, 0, 255, &k, Data); -// initial (fast) skip decision - if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { - //possible skip + /* initial (fast) skip decision */ + if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) { + /* possible skip */ if (Data->chroma) { pMB->mode = MODE_DIRECT_NONE_MV; - return *Data->iMinSAD; // skip. + 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. + if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */ } } + *Data->iMinSAD += Data->lambda16; skip_sad = *Data->iMinSAD; - -// DIRECT MODE DELTA VECTOR SEARCH. -// This has to be made more effective, but at the moment I'm happy it's running at all + + /* + * 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 & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; @@ -1828,7 +1803,7 @@ *best_sad = *Data->iMinSAD; if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; - else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation + else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */ pMB->pmvs[3] = *Data->currentMV; @@ -1886,25 +1861,25 @@ SearchData bData; fData->qpel_precision = 0; - memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data + 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; 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.b_RefP[0] = fData->RefP[0] = f_Ref->y + i; + bData.b_RefP[2] = fData->RefP[2] = f_RefH + i; + bData.b_RefP[1] = fData->RefP[1] = f_RefV + i; + bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i; + bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i; + bData.RefP[2] = fData->b_RefP[2] = b_RefH + i; + bData.RefP[1] = fData->b_RefP[1] = b_RefV + i; + bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i; + bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; bData.bpredMV = fData->predMV = *f_predMV; fData->bpredMV = bData.predMV = *b_predMV; @@ -1925,10 +1900,10 @@ CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); -//diamond + /* diamond */ do { iDirection = 255; - // forward MV moves + /* forward MV moves */ i = fData->currentMV[0].x; j = fData->currentMV[0].y; CheckCandidateInt(i + 1, j, 0, &iDirection, fData); @@ -1936,7 +1911,7 @@ CheckCandidateInt(i - 1, j, 0, &iDirection, fData); CheckCandidateInt(i, j - 1, 0, &iDirection, fData); - // backward MV moves + /* backward MV moves */ i = fData->currentMV[1].x; j = fData->currentMV[1].y; fData->currentMV[2] = fData->currentMV[0]; CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); @@ -1946,7 +1921,7 @@ } while (!(iDirection)); -//qpel refinement + /* qpel refinement */ if (fData->qpel) { if (*fData->iMinSAD > *best_sad + 500) return; CheckCandidate = CheckCandidateInt; @@ -1963,7 +1938,7 @@ SubpelRefine(&bData); } - *fData->iMinSAD += (2+3) * 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; @@ -1991,13 +1966,13 @@ FRAMEINFO * const frame, const int32_t time_bp, const int32_t time_pp, - // forward (past) reference + /* forward (past) reference */ const MACROBLOCK * const f_mbs, const IMAGE * const f_ref, const IMAGE * const f_refH, const IMAGE * const f_refV, const IMAGE * const f_refHV, - // backward (future) reference + /* backward (future) reference */ const FRAMEINFO * const b_reference, const IMAGE * const b_ref, const IMAGE * const b_refH, @@ -2015,7 +1990,7 @@ const int32_t TRB = time_pp - time_bp; const int32_t TRD = time_pp; -// some pre-inintialized data for the rest of the search + /* some pre-inintialized data for the rest of the search */ SearchData Data; int32_t iMinSAD; @@ -2032,8 +2007,9 @@ Data.chroma = frame->motion_flags & XVID_ME_CHROMA8; Data.temp = temp; - Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) - // note: i==horizontal, j==vertical + 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++) { f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ @@ -2052,6 +2028,7 @@ 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 and (2) sets very good predictions for forward and backward search */ @@ -2068,7 +2045,7 @@ if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } - // forward search + /* forward search */ SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, &frame->image, i, j, frame->motion_flags, @@ -2076,7 +2053,7 @@ pMB, &f_predMV, &best_sad, MODE_FORWARD, &Data); - // backward search + /* backward search */ SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, &frame->image, i, j, frame->motion_flags, @@ -2084,7 +2061,7 @@ pMB, &b_predMV, &best_sad, MODE_BACKWARD, &Data); - // interpolate search comes last, because it uses data from forward and backward as prediction + /* interpolate search comes last, because it uses data from forward and backward as prediction */ 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, @@ -2096,7 +2073,7 @@ pMB, &best_sad, &Data); -// final skip decision + /* 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); @@ -2136,25 +2113,26 @@ { int i, mask; + int quarterpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0; VECTOR pmv[3]; - MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; + MACROBLOCK * const 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 + /* 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 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 + 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 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->vol_flags&XVID_VOL_QUARTERPEL?1:0), 0, Data->rrv); + pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0); Data->Cur = pCur + (x + y * pParam->edged_width) * 16; - Data->Ref = pRef + (x + y * pParam->edged_width) * 16; + Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; pmv[1].x = EVEN(pMB->mvs[0].x); pmv[1].y = EVEN(pMB->mvs[0].y); @@ -2164,42 +2142,43 @@ CheckCandidate32I(0, 0, 255, &i, Data); - if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) { 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 > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) /* diamond only if needed */ DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); + } - 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]; - } + 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_BIAS 2500 -#define INTRA_THRESH 1500 -#define INTER_THRESH 1400 +#define INTRA_THRESH 1700 +#define INTER_THRESH 1200 int 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 of B frames in a row + const FRAMEINFO * const Current, + const MBParam * const pParam, + const int maxIntra, /* maximum number if non-I frames */ + const int intraCount, /* number of non-I frames after last I frame; 0 if we force P/B frame */ + const int bCount, /* number of B frames in a row */ + const int b_thresh) { uint32_t x, y, intra = 0; int sSAD = 0; MACROBLOCK * const pMBs = Current->mbs; const IMAGE * const pCurrent = &Current->image; - int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; + int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh; + int s = 0, blocks = 0; int32_t iMinSAD[5], temp[5]; VECTOR currentMV[5]; @@ -2208,26 +2187,34 @@ Data.currentMV = currentMV; Data.iMinSAD = iMinSAD; Data.iFcode = Current->fcode; - Data.rrv = Current->vop_flags & XVID_VOP_REDUCED; Data.temp = temp; CheckCandidate = CheckCandidate32I; - 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; + if (intraCount != 0) { + if (intraCount < 10) /* we're right after an I frame */ + IntraThresh += 15* (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 - 8*(maxIntra - intraCount)))/maxIntra; + } - InterThresh += 400 * (1 - bCount); - if (InterThresh < 300) InterThresh = 300; + InterThresh -= (350 - 8*b_thresh) * bCount; + if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh; if (sadInit) (*sadInit) (); for (y = 1; y < pParam->mb_height-1; y += 2) { for (x = 1; x < pParam->mb_width-1; x += 2) { int i; + blocks += 4; if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; + else { /* extrapolation of the vector found for last frame */ + pMBs[x + y * pParam->mb_width].mvs[0].x = + (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount; + pMBs[x + y * pParam->mb_width].mvs[0].y = + (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount; + } MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); @@ -2239,19 +2226,26 @@ 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; + if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP; } } + if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++; + sSAD += pMB->sad16; } } } - sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); -// if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP; + + sSAD /= blocks; + + if (b_thresh < 20) { + s = (10*s) / blocks; + if (s > 4) sSAD += (s - 2) * (40 - 2*b_thresh); /* static block - looks bad when in bframe... */ + } + if (sSAD > InterThresh ) return P_VOP; emms(); return B_VOP; - } @@ -2265,9 +2259,9 @@ const IMAGE * const pRefHV ) { - const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs + 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 + const int grad=512; /* lower bound for deviation in MB */ WARPPOINTS gmc; @@ -2283,13 +2277,13 @@ double meanx,meany; int num,oldnum; - if (!MBmask) { fprintf(stderr,"Mem error\n"); - gmc.duv[0].x= gmc.duv[0].y = - gmc.duv[1].x= gmc.duv[1].y = + if (!MBmask) { fprintf(stderr,"Mem error\n"); + 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; return gmc; } -// filter mask of all blocks + /* filter mask of all blocks */ for (my = 1; my < (uint32_t)MBh-1; my++) for (mx = 1; mx < (uint32_t)MBw-1; mx++) @@ -2301,10 +2295,10 @@ 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) ) ) + 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; } @@ -2353,7 +2347,7 @@ denom = a*a+b*b-c*n; -/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ +/* 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]; @@ -2379,8 +2373,8 @@ 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 ); + meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); + meany += fabs(( 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 */ @@ -2407,8 +2401,8 @@ 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 ) ) + if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) + || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) MBmask[mbnum]=0; else num++; @@ -2430,14 +2424,14 @@ 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); +/* 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); */ free(MBmask); return gmc; } -// functions which perform BITS-based search/bitcount +/* functions which perform BITS-based search/bitcount */ static int CountMBBitsInter(SearchData * const Data, @@ -2458,11 +2452,7 @@ 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 & (XVID_ME_HALFPELREFINE16_BITS | XVID_ME_EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search + if (MotionFlags & (XVID_ME_HALFPELREFINE16_BITS | XVID_ME_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); @@ -2471,13 +2461,9 @@ CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); } - } else { // not qpel + } 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&XVID_ME_EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); @@ -2485,13 +2471,13 @@ if (MotionFlags&XVID_ME_HALFPELREFINE16_BITS) SubpelRefine(Data); if (Data->qpel) { - if (MotionFlags&(XVID_ME_EXTSEARCH_BITS | XVID_ME_HALFPELREFINE16_BITS)) { // there was halfpel-precision search + if (MotionFlags&(XVID_ME_EXTSEARCH_BITS | XVID_ME_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].x = 2 * Data->currentMV[i].x; /* we have found a better match */ Data->currentQMV[i].y = 2 * Data->currentMV[i].y; } - // preparing for qpel-precision search + /* 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); @@ -2499,7 +2485,7 @@ if (MotionFlags&XVID_ME_QUARTERPELREFINE16_BITS) SubpelRefine(Data); } - if (MotionFlags&XVID_ME_CHECKPREDICTION_BITS) { //let's check vector equal to prediction + if (MotionFlags&XVID_ME_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); @@ -2507,7 +2493,6 @@ return Data->iMinSAD[0]; } - static int CountMBBitsInter4v(const SearchData * const Data, MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, @@ -2519,20 +2504,22 @@ int cbp = 0, bits = 0, t = 0, i, iDirection; SearchData Data2, *Data8 = &Data2; int sumx = 0, sumy = 0; - int16_t in[64], coeff[64]; + int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; + uint8_t * ptr; memcpy(Data8, Data, sizeof(SearchData)); CheckCandidate = CheckCandidateBits8; - for (i = 0; i < 4; i++) { + for (i = 0; i < 4; i++) { /* for all luma blocks */ + Data8->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); + Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); if(Data->qpel) { Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); @@ -2547,20 +2534,18 @@ 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->iMinSAD += BITS_MULT*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)) + /* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */ + { + VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; + if (!MVequal (*v, backup[i+1]) ) CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); } if (Data8->qpel) { - if (MotionFlags&XVID_ME_HALFPELREFINE8_BITS || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_BITS)) { // halfpixel motion search follows + if (MotionFlags&XVID_ME_HALFPELREFINE8_BITS || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_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; @@ -2574,9 +2559,10 @@ if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); - if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8); + if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) + SubpelRefine(Data8); - if(s > *Data8->iMinSAD) { //we have found a better match + if(s > *Data8->iMinSAD) { /* we have found a better match */ Data8->currentQMV->x = 2*Data8->currentMV->x; Data8->currentQMV->y = 2*Data8->currentMV->y; } @@ -2588,20 +2574,26 @@ } if (MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS) SubpelRefine(Data8); - } else // not qpel - if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement + } else { /* not qpel */ + + if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) /* extsearch */ + SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); + + if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) + SubpelRefine(Data8); /* halfpel refinement */ + } - //checking vector equal to predicion + /* checking vector equal to predicion */ if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_BITS) { const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; - if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y)) + if (!MVequal(*v, Data8->predMV)) CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); } bits += *Data8->iMinSAD; - if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V + if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */ - // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else + /* 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; @@ -2617,108 +2609,60 @@ 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; - } - } + } /* /for all luma blocks */ + + bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; + + /* let's check chroma */ + sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; + sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; + + /* chroma U */ + ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); + transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); + bits += Block_CalcBits(coeff, in, Data->iQuant, Data->quant_type, &cbp, 4); + + if (bits >= *Data->iMinSAD) return bits; + + /* chroma V */ + ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); + transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); + bits += Block_CalcBits(coeff, in, Data->iQuant, Data->quant_type, &cbp, 5); + + bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; 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]; + int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */ + int cbp = 0, i, dc = 0; + int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 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); + bits += Block_CalcBitsIntra(coeff, in, Data->iQuant, Data->quant_type, &cbp, i, &dc); - Data->temp[5] = t; + if (bits >= Data->iMinSAD[0]) return bits; + } - bits += t = cbpy_tab[cbp>>2].len; - Data->temp[6] = t; + bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len; - bits += t = mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; - Data->temp[7] = t; + /*chroma U */ + transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); + bits += Block_CalcBitsIntra(coeff, in, Data->iQuant, Data->quant_type, &cbp, 4, &dc); + + if (bits >= Data->iMinSAD[0]) return bits; - } - } + /* chroma V */ + transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); + bits += Block_CalcBitsIntra(coeff, in, Data->iQuant, Data->quant_type, &cbp, 5, &dc); + + bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; return bits; }