1 |
/************************************************************************** |
/***************************************************************************** |
2 |
* |
* |
3 |
* XVID MPEG-4 VIDEO CODEC |
* XVID MPEG-4 VIDEO CODEC |
4 |
* motion estimation |
* - Motion Estimation related code - |
5 |
* |
* |
6 |
* This program is an implementation of a part of one or more MPEG-4 |
* Copyright(C) 2002 Christoph Lampert <gruel@web.de> |
7 |
* Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
* 2002 Michael Militzer <michael@xvid.org> |
8 |
* to use this software module in hardware or software products are |
* 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net> |
|
* 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. |
|
9 |
* |
* |
10 |
* This program is free software; you can redistribute it and/or modify |
* This program is free software; you can redistribute it and/or modify |
11 |
* it under the terms of the GNU General Public License as published by |
* it under the terms of the GNU General Public License as published by |
19 |
* |
* |
20 |
* You should have received a copy of the GNU General Public License |
* You should have received a copy of the GNU General Public License |
21 |
* along with this program; if not, write to the Free Software |
* along with this program; if not, write to the Free Software |
22 |
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
|
* |
24 |
|
* $Id: motion_est.c,v 1.58.2.23 2003-07-24 13:09:01 Isibaar Exp $ |
25 |
* |
* |
26 |
*************************************************************************/ |
****************************************************************************/ |
27 |
|
|
28 |
#include <assert.h> |
#include <assert.h> |
29 |
#include <stdio.h> |
#include <stdio.h> |
30 |
#include <stdlib.h> |
#include <stdlib.h> |
31 |
#include <string.h> // memcpy |
#include <string.h> /* memcpy */ |
32 |
#include <math.h> // lrint |
#include <math.h> /* lrint */ |
33 |
|
|
34 |
#include "../encoder.h" |
#include "../encoder.h" |
35 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
40 |
#include "motion_est.h" |
#include "motion_est.h" |
41 |
#include "motion.h" |
#include "motion.h" |
42 |
#include "sad.h" |
#include "sad.h" |
43 |
|
#include "gmc.h" |
44 |
#include "../utils/emms.h" |
#include "../utils/emms.h" |
45 |
#include "../dct/fdct.h" |
#include "../dct/fdct.h" |
46 |
|
|
72 |
#define CHECK_CANDIDATE(X,Y,D) { \ |
#define CHECK_CANDIDATE(X,Y,D) { \ |
73 |
CheckCandidate((X),(Y), (D), &iDirection, data ); } |
CheckCandidate((X),(Y), (D), &iDirection, data ); } |
74 |
|
|
75 |
|
|
76 |
/***************************************************************************** |
/***************************************************************************** |
77 |
* Code |
* Code |
78 |
****************************************************************************/ |
****************************************************************************/ |
167 |
const uint32_t stride = data->iEdgedWidth/2; |
const uint32_t stride = data->iEdgedWidth/2; |
168 |
int offset = (dx>>1) + (dy>>1)*stride; |
int offset = (dx>>1) + (dy>>1)*stride; |
169 |
|
|
170 |
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently |
if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */ |
171 |
data->temp[5] = dx; data->temp[6] = dy; // backup |
data->temp[5] = dx; data->temp[6] = dy; /* backup */ |
172 |
|
|
173 |
switch (((dx & 1) << 1) | (dy & 1)) { |
switch (((dx & 1) << 1) | (dy & 1)) { |
174 |
case 0: |
case 0: |
191 |
sad += sad8(data->CurV, data->RefQ, stride); |
sad += sad8(data->CurV, data->RefQ, stride); |
192 |
break; |
break; |
193 |
} |
} |
194 |
data->temp[7] = sad; //backup, part 2 |
data->temp[7] = sad; /* backup, part 2 */ |
195 |
return sad; |
return sad; |
196 |
} |
} |
197 |
|
|
198 |
static __inline const uint8_t * |
static __inline const uint8_t * |
199 |
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data) |
200 |
{ |
{ |
201 |
// dir : 0 = forward, 1 = backward |
/* dir : 0 = forward, 1 = backward */ |
202 |
const uint8_t* const *direction = ( dir == 0 ? data->RefP : data->b_RefP ); |
const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP ); |
203 |
const int picture = ((x&1)<<1) | (y&1); |
const int picture = ((x&1)<<1) | (y&1); |
204 |
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; |
205 |
return direction[picture] + offset; |
return direction[picture] + offset; |
206 |
} |
} |
207 |
|
|
208 |
// 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 */ |
209 |
static __inline const uint8_t * |
static __inline const uint8_t * |
210 |
GetReference(const int x, const int y, const SearchData * const data) |
GetReference(const int x, const int y, const SearchData * const data) |
211 |
{ |
{ |
217 |
static uint8_t * |
static uint8_t * |
218 |
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) |
219 |
{ |
{ |
220 |
// create or find a qpel-precision reference picture; return pointer to it |
/* create or find a qpel-precision reference picture; return pointer to it */ |
221 |
uint8_t * Reference = data->RefQ + 16*dir; |
uint8_t * Reference = data->RefQ + 16*dir; |
222 |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
223 |
const uint32_t rounding = data->rounding; |
const uint32_t rounding = data->rounding; |
228 |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
229 |
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
230 |
switch( ((x&1)<<1) + (y&1) ) { |
switch( ((x&1)<<1) + (y&1) ) { |
231 |
case 3: // x and y in qpel resolution - the "corners" (top left/right and |
case 3: /* x and y in qpel resolution - the "corners" (top left/right and */ |
232 |
// bottom left/right) during qpel refinement |
/* bottom left/right) during qpel refinement */ |
233 |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
234 |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
235 |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
239 |
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); |
240 |
break; |
break; |
241 |
|
|
242 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
243 |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
244 |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
245 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
246 |
break; |
break; |
247 |
|
|
248 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
249 |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
250 |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; |
251 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
252 |
break; |
break; |
253 |
|
|
254 |
default: // pure halfpel position |
default: /* pure halfpel position */ |
255 |
return (uint8_t *) ref1; |
return (uint8_t *) ref1; |
256 |
|
|
257 |
} |
} |
261 |
static uint8_t * |
static uint8_t * |
262 |
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data) |
263 |
{ |
{ |
264 |
// create or find a qpel-precision reference picture; return pointer to it |
/* create or find a qpel-precision reference picture; return pointer to it */ |
265 |
uint8_t * Reference = data->RefQ + 16*dir; |
uint8_t * Reference = data->RefQ + 16*dir; |
266 |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
const uint32_t iEdgedWidth = data->iEdgedWidth; |
267 |
const uint32_t rounding = data->rounding; |
const uint32_t rounding = data->rounding; |
271 |
|
|
272 |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); |
273 |
switch( ((x&1)<<1) + (y&1) ) { |
switch( ((x&1)<<1) + (y&1) ) { |
274 |
case 3: // x and y in qpel resolution - the "corners" (top left/right and |
case 3: |
275 |
// bottom left/right) during qpel refinement |
/* |
276 |
|
* x and y in qpel resolution - the "corners" (top left/right and |
277 |
|
* bottom left/right) during qpel refinement |
278 |
|
*/ |
279 |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
280 |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
281 |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data); |
285 |
interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, 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); |
286 |
break; |
break; |
287 |
|
|
288 |
case 1: // x halfpel, y qpel - top or bottom during qpel refinement |
case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */ |
289 |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); |
290 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
291 |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
293 |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
294 |
break; |
break; |
295 |
|
|
296 |
case 2: // x qpel, y halfpel - left or right during qpel refinement |
case 2: /* x qpel, y halfpel - left or right during qpel refinement */ |
297 |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); |
298 |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); |
299 |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); |
301 |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); |
302 |
break; |
break; |
303 |
|
|
304 |
default: // pure halfpel position |
default: /* pure halfpel position */ |
305 |
return (uint8_t *) ref1; |
return (uint8_t *) ref1; |
306 |
} |
} |
307 |
return Reference; |
return Reference; |
324 |
Reference = GetReference(x, y, data); |
Reference = GetReference(x, y, data); |
325 |
current = data->currentMV; |
current = data->currentMV; |
326 |
xc = x; yc = y; |
xc = x; yc = y; |
327 |
} else { // x and y are in 1/4 precision |
} else { /* x and y are in 1/4 precision */ |
328 |
Reference = Interpolate16x16qpel(x, y, 0, data); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
329 |
xc = x/2; yc = y/2; //for chroma sad |
xc = x/2; yc = y/2; /* for chroma sad */ |
330 |
current = data->currentQMV; |
current = data->currentQMV; |
331 |
} |
} |
332 |
|
|
336 |
sad += (data->lambda16 * t * sad)>>10; |
sad += (data->lambda16 * t * sad)>>10; |
337 |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; |
338 |
|
|
339 |
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
if (data->chroma && sad < data->iMinSAD[0]) |
340 |
|
sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
341 |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
342 |
|
|
343 |
if (sad < data->iMinSAD[0]) { |
if (sad < data->iMinSAD[0]) { |
354 |
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } |
355 |
if (data->temp[4] < data->iMinSAD[4]) { |
if (data->temp[4] < data->iMinSAD[4]) { |
356 |
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } |
|
|
|
357 |
} |
} |
358 |
|
|
359 |
static void |
static void |
369 |
if (!data->qpel_precision) { |
if (!data->qpel_precision) { |
370 |
Reference = GetReference(x, y, data); |
Reference = GetReference(x, y, data); |
371 |
current = data->currentMV; |
current = data->currentMV; |
372 |
} else { // x and y are in 1/4 precision |
} else { /* x and y are in 1/4 precision */ |
373 |
Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
Reference = Interpolate8x8qpel(x, y, 0, 0, data); |
374 |
current = data->currentQMV; |
current = data->currentQMV; |
375 |
} |
} |
386 |
} |
} |
387 |
} |
} |
388 |
|
|
|
|
|
389 |
static void |
static void |
390 |
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
391 |
{ |
{ |
392 |
uint32_t t; |
uint32_t t; |
393 |
const uint8_t * Reference; |
const uint8_t * Reference; |
394 |
|
|
395 |
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero even value |
if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */ |
396 |
(x > data->max_dx) || (x < data->min_dx) |
(x > data->max_dx) || (x < data->min_dx) |
397 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
398 |
|
|
430 |
if ( (x > data->max_dx) || ( x < data->min_dx) |
if ( (x > data->max_dx) || ( x < data->min_dx) |
431 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
432 |
|
|
433 |
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 */ |
434 |
|
|
435 |
if (data->qpel_precision) { // x and y are in 1/4 precision |
if (data->qpel_precision) { /* x and y are in 1/4 precision */ |
436 |
Reference = Interpolate16x16qpel(x, y, 0, data); |
Reference = Interpolate16x16qpel(x, y, 0, data); |
437 |
current = data->currentQMV; |
current = data->currentQMV; |
438 |
xc = x/2; yc = y/2; |
xc = x/2; yc = y/2; |
447 |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
448 |
sad += (data->lambda16 * t * sad)>>10; |
sad += (data->lambda16 * t * sad)>>10; |
449 |
|
|
450 |
if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
if (data->chroma && sad < *data->iMinSAD) |
451 |
|
sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
452 |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
(yc >> 1) + roundtab_79[yc & 0x3], data); |
453 |
|
|
454 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
459 |
} |
} |
460 |
|
|
461 |
static void |
static void |
462 |
|
CheckCandidate16I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
463 |
|
{ |
464 |
|
int sad; |
465 |
|
// int xc, yc; |
466 |
|
const uint8_t * Reference; |
467 |
|
// VECTOR * current; |
468 |
|
|
469 |
|
if ( (x > data->max_dx) || ( x < data->min_dx) |
470 |
|
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
471 |
|
|
472 |
|
Reference = GetReference(x, y, data); |
473 |
|
// xc = x; yc = y; |
474 |
|
|
475 |
|
sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); |
476 |
|
// sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0); |
477 |
|
|
478 |
|
/* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], |
479 |
|
(yc >> 1) + roundtab_79[yc & 0x3], data); |
480 |
|
*/ |
481 |
|
|
482 |
|
if (sad < data->iMinSAD[0]) { |
483 |
|
data->iMinSAD[0] = sad; |
484 |
|
data->currentMV[0].x = x; data->currentMV[0].y = y; |
485 |
|
*dir = Direction; |
486 |
|
} |
487 |
|
} |
488 |
|
|
489 |
|
static void |
490 |
CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
491 |
{ |
{ |
492 |
// maximum speed - for P/B/I decision |
/* maximum speed - for P/B/I decision */ |
493 |
int32_t sad; |
int32_t sad; |
494 |
|
|
495 |
if ( (x > data->max_dx) || (x < data->min_dx) |
if ( (x > data->max_dx) || (x < data->min_dx) |
496 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
497 |
|
|
498 |
sad = sad32v_c(data->Cur, data->RefP[0] + x/2 + (y/2)*(data->iEdgedWidth), |
sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*((int)data->iEdgedWidth), |
499 |
data->iEdgedWidth, data->temp+1); |
data->iEdgedWidth, data->temp+1); |
500 |
|
|
501 |
if (sad < *(data->iMinSAD)) { |
if (sad < *(data->iMinSAD)) { |
548 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
549 |
sad += (data->lambda16 * t * sad)>>10; |
sad += (data->lambda16 * t * sad)>>10; |
550 |
|
|
551 |
if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
if (data->chroma && sad < *data->iMinSAD) |
552 |
|
sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], |
553 |
(ycf >> 1) + roundtab_79[ycf & 0x3], |
(ycf >> 1) + roundtab_79[ycf & 0x3], |
554 |
(xcb >> 1) + roundtab_79[xcb & 0x3], |
(xcb >> 1) + roundtab_79[xcb & 0x3], |
555 |
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
(ycb >> 1) + roundtab_79[ycb & 0x3], data); |
595 |
} else { |
} else { |
596 |
xcf += mvs.x; ycf += mvs.y; |
xcf += mvs.x; ycf += mvs.y; |
597 |
xcb += b_mvs.x; ycb += b_mvs.y; |
xcb += b_mvs.x; ycb += b_mvs.y; |
598 |
mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway |
mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */ |
599 |
b_mvs.x *= 2; b_mvs.y *= 2; |
b_mvs.x *= 2; b_mvs.y *= 2; |
600 |
} |
} |
601 |
|
|
609 |
|
|
610 |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
611 |
|
|
612 |
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
if (data->chroma && sad < *data->iMinSAD) |
613 |
|
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
614 |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
615 |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
616 |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
662 |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); |
663 |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; |
664 |
|
|
665 |
if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
if (data->chroma && sad < *data->iMinSAD) |
666 |
|
sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], |
667 |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
(ycf >> 3) + roundtab_76[ycf & 0xf], |
668 |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
(xcb >> 3) + roundtab_76[xcb & 0xf], |
669 |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
(ycb >> 3) + roundtab_76[ycb & 0xf], data); |
681 |
{ |
{ |
682 |
|
|
683 |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
684 |
int32_t bits = 0, sum; |
int32_t bits = 0; |
685 |
VECTOR * current; |
VECTOR * current; |
686 |
const uint8_t * ptr; |
const uint8_t * ptr; |
687 |
int i, cbp = 0, t, xc, yc; |
int i, cbp = 0, t, xc, yc; |
693 |
ptr = GetReference(x, y, data); |
ptr = GetReference(x, y, data); |
694 |
current = data->currentMV; |
current = data->currentMV; |
695 |
xc = x; yc = y; |
xc = x; yc = y; |
696 |
} else { // x and y are in 1/4 precision |
} else { /* x and y are in 1/4 precision */ |
697 |
ptr = Interpolate16x16qpel(x, y, 0, data); |
ptr = Interpolate16x16qpel(x, y, 0, data); |
698 |
current = data->currentQMV; |
current = data->currentQMV; |
699 |
xc = x/2; yc = y/2; |
xc = x/2; yc = y/2; |
702 |
for(i = 0; i < 4; i++) { |
for(i = 0; i < 4; i++) { |
703 |
int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); |
int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); |
704 |
transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); |
transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); |
705 |
fdct(in); |
bits += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i); |
|
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; |
|
706 |
} |
} |
707 |
|
|
708 |
bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
bits += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
709 |
|
|
710 |
|
if (data->temp[0] + t < data->iMinSAD[1]) { |
711 |
|
data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) | cbp&32; } |
712 |
|
if (data->temp[1] < data->iMinSAD[2]) { |
713 |
|
data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; data->cbp[1] = (data->cbp[1]&~16) | cbp&16; } |
714 |
|
if (data->temp[2] < data->iMinSAD[3]) { |
715 |
|
data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; data->cbp[1] = (data->cbp[1]&~8) | cbp&8; } |
716 |
|
if (data->temp[3] < data->iMinSAD[4]) { |
717 |
|
data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | cbp&4; } |
718 |
|
|
719 |
|
bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
720 |
|
|
721 |
|
if (bits >= data->iMinSAD[0]) return; |
722 |
|
|
723 |
if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma |
/* chroma */ |
724 |
xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
xc = (xc >> 1) + roundtab_79[xc & 0x3]; |
725 |
yc = (yc >> 1) + roundtab_79[yc & 0x3]; |
yc = (yc >> 1) + roundtab_79[yc & 0x3]; |
726 |
|
|
727 |
//chroma U |
/* chroma U */ |
728 |
ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
729 |
transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); |
transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2); |
730 |
fdct(in); |
bits += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4); |
731 |
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
if (bits >= data->iMinSAD[0]) return; |
732 |
else sum = quant4_inter(coeff, in, data->lambda16); |
|
733 |
if (sum > 0) { |
/* chroma V */ |
734 |
cbp |= 1 << (5 - 4); |
ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); |
735 |
bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); |
transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2); |
736 |
} |
bits += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
|
|
|
|
if (bits < data->iMinSAD[0]) { |
|
|
//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); |
|
|
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]); |
|
|
} |
|
|
} |
|
|
} |
|
737 |
|
|
738 |
bits += xvid_cbpy_tab[15-(cbp>>2)].len; |
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
|
bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
|
739 |
|
|
740 |
if (bits < data->iMinSAD[0]) { |
if (bits < data->iMinSAD[0]) { |
741 |
data->iMinSAD[0] = bits; |
data->iMinSAD[0] = bits; |
742 |
current[0].x = x; current[0].y = y; |
current[0].x = x; current[0].y = y; |
743 |
*dir = Direction; |
*dir = Direction; |
744 |
|
*data->cbp = cbp; |
745 |
} |
} |
|
|
|
|
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; } |
|
|
|
|
746 |
} |
} |
747 |
|
|
748 |
static void |
static void |
749 |
CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) |
750 |
{ |
{ |
751 |
|
|
752 |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; |
753 |
int32_t sum, bits; |
int32_t bits; |
754 |
VECTOR * current; |
VECTOR * current; |
755 |
const uint8_t * ptr; |
const uint8_t * ptr; |
756 |
int cbp; |
int cbp = 0; |
757 |
|
|
758 |
if ( (x > data->max_dx) || (x < data->min_dx) |
if ( (x > data->max_dx) || (x < data->min_dx) |
759 |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
|| (y > data->max_dy) || (y < data->min_dy) ) return; |
761 |
if (!data->qpel_precision) { |
if (!data->qpel_precision) { |
762 |
ptr = GetReference(x, y, data); |
ptr = GetReference(x, y, data); |
763 |
current = data->currentMV; |
current = data->currentMV; |
764 |
} else { // x and y are in 1/4 precision |
} else { /* x and y are in 1/4 precision */ |
765 |
ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
ptr = Interpolate8x8qpel(x, y, 0, 0, data); |
766 |
current = data->currentQMV; |
current = data->currentQMV; |
767 |
} |
} |
768 |
|
|
769 |
transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); |
770 |
fdct(in); |
bits = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); |
771 |
if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); |
bits += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); |
|
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); |
|
772 |
|
|
773 |
if (bits < data->iMinSAD[0]) { |
if (bits < data->iMinSAD[0]) { |
774 |
data->temp[0] = cbp; |
*data->cbp = cbp; |
775 |
data->iMinSAD[0] = bits; |
data->iMinSAD[0] = bits; |
776 |
current[0].x = x; current[0].y = y; |
current[0].x = x; current[0].y = y; |
777 |
*dir = Direction; |
*dir = Direction; |
790 |
|
|
791 |
int iDirection; |
int iDirection; |
792 |
|
|
793 |
for(;;) { //forever |
for(;;) { /* forever */ |
794 |
iDirection = 0; |
iDirection = 0; |
795 |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
796 |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
799 |
|
|
800 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
801 |
|
|
802 |
if (iDirection) { //if anything found |
if (iDirection) { /* if anything found */ |
803 |
bDirection = iDirection; |
bDirection = iDirection; |
804 |
iDirection = 0; |
iDirection = 0; |
805 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
806 |
if (bDirection & 3) { //our candidate is left or right |
if (bDirection & 3) { /* our candidate is left or right */ |
807 |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
808 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
809 |
} else { // what remains here is up or down |
} else { /* what remains here is up or down */ |
810 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
811 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
812 |
} |
} |
815 |
bDirection += iDirection; |
bDirection += iDirection; |
816 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
817 |
} |
} |
818 |
} else { //about to quit, eh? not so fast.... |
} else { /* about to quit, eh? not so fast.... */ |
819 |
switch (bDirection) { |
switch (bDirection) { |
820 |
case 2: |
case 2: |
821 |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
853 |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
854 |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
855 |
break; |
break; |
856 |
default: //1+2+4+8 == we didn't find anything at all |
default: /* 1+2+4+8 == we didn't find anything at all */ |
857 |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); |
858 |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); |
859 |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); |
860 |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); |
861 |
break; |
break; |
862 |
} |
} |
863 |
if (!iDirection) break; //ok, the end. really |
if (!iDirection) break; /* ok, the end. really */ |
864 |
bDirection = iDirection; |
bDirection = iDirection; |
865 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
866 |
} |
} |
905 |
|
|
906 |
/* now we're doing diagonal checks near our candidate */ |
/* now we're doing diagonal checks near our candidate */ |
907 |
|
|
908 |
if (iDirection) { //checking if anything found |
if (iDirection) { /* checking if anything found */ |
909 |
bDirection = iDirection; |
bDirection = iDirection; |
910 |
iDirection = 0; |
iDirection = 0; |
911 |
x = data->currentMV->x; y = data->currentMV->y; |
x = data->currentMV->x; y = data->currentMV->y; |
912 |
if (bDirection & 3) { //our candidate is left or right |
if (bDirection & 3) { /* our candidate is left or right */ |
913 |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
CHECK_CANDIDATE(x, y + iDiamondSize, 8); |
914 |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
CHECK_CANDIDATE(x, y - iDiamondSize, 4); |
915 |
} else { // what remains here is up or down |
} else { /* what remains here is up or down */ |
916 |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
CHECK_CANDIDATE(x + iDiamondSize, y, 2); |
917 |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
CHECK_CANDIDATE(x - iDiamondSize, y, 1); |
918 |
} |
} |
930 |
{ |
{ |
931 |
/* Do a half-pel or q-pel refinement */ |
/* Do a half-pel or q-pel refinement */ |
932 |
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; |
933 |
int iDirection; //only needed because macro expects it |
int iDirection; /* only needed because macro expects it */ |
934 |
|
|
935 |
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); |
936 |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); |
978 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
979 |
} |
} |
980 |
|
|
981 |
|
static __inline void |
982 |
|
ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad) |
983 |
|
{ |
984 |
|
pMB->mode = MODE_INTER; |
985 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV; |
986 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV; |
987 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
988 |
|
} |
989 |
|
|
990 |
|
static __inline void |
991 |
|
ModeDecision(SearchData * const Data, |
992 |
|
MACROBLOCK * const pMB, |
993 |
|
const MACROBLOCK * const pMBs, |
994 |
|
const int x, const int y, |
995 |
|
const MBParam * const pParam, |
996 |
|
const uint32_t MotionFlags, |
997 |
|
const uint32_t VopFlags, |
998 |
|
const uint32_t VolFlags, |
999 |
|
const IMAGE * const pCurrent, |
1000 |
|
const IMAGE * const pRef, |
1001 |
|
const IMAGE * const vGMC, |
1002 |
|
const int coding_type) |
1003 |
|
{ |
1004 |
|
int mode = MODE_INTER; |
1005 |
|
int mcsel = 0; |
1006 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
1007 |
|
const uint32_t iQuant = pMB->quant; |
1008 |
|
|
1009 |
|
const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0); |
1010 |
|
|
1011 |
|
pMB->mcsel = 0; |
1012 |
|
|
1013 |
|
if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) { /* normal, fast, SAD-based mode decision */ |
1014 |
|
int sad; |
1015 |
|
int InterBias = MV16_INTER_BIAS; |
1016 |
|
if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + |
1017 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { |
1018 |
|
mode = MODE_INTER; |
1019 |
|
sad = Data->iMinSAD[0]; |
1020 |
|
} else { |
1021 |
|
mode = MODE_INTER4V; |
1022 |
|
sad = Data->iMinSAD[1] + Data->iMinSAD[2] + |
1023 |
|
Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; |
1024 |
|
Data->iMinSAD[0] = sad; |
1025 |
|
} |
1026 |
|
|
1027 |
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
1028 |
|
if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) |
1029 |
|
if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) |
1030 |
|
if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { |
1031 |
|
mode = MODE_NOT_CODED; |
1032 |
|
sad = 0; |
1033 |
|
} |
1034 |
|
|
1035 |
|
/* mcsel */ |
1036 |
|
if (coding_type == S_VOP) { |
1037 |
|
|
1038 |
|
int32_t iSAD = sad16(Data->Cur, |
1039 |
|
vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); |
1040 |
|
|
1041 |
|
if (Data->chroma) { |
1042 |
|
iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1043 |
|
iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); |
1044 |
|
} |
1045 |
|
|
1046 |
|
if (iSAD <= sad) { /* mode decision GMC */ |
1047 |
|
mode = MODE_INTER; |
1048 |
|
mcsel = 1; |
1049 |
|
sad = iSAD; |
1050 |
|
} |
1051 |
|
|
1052 |
|
} |
1053 |
|
|
1054 |
|
/* intra decision */ |
1055 |
|
|
1056 |
|
if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */ |
1057 |
|
if (y != 0) |
1058 |
|
if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; |
1059 |
|
if (x != 0) |
1060 |
|
if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; |
1061 |
|
|
1062 |
|
if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ |
1063 |
|
if (Data->rrv) InterBias *= 4; |
1064 |
|
|
1065 |
|
if (InterBias < sad) { |
1066 |
|
int32_t deviation; |
1067 |
|
if (!Data->rrv) |
1068 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth); |
1069 |
|
else |
1070 |
|
deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */ |
1071 |
|
dev16(Data->Cur+16, Data->iEdgedWidth) + |
1072 |
|
dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + |
1073 |
|
dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); |
1074 |
|
|
1075 |
|
if (deviation < (sad - InterBias)) mode = MODE_INTRA; |
1076 |
|
} |
1077 |
|
|
1078 |
|
pMB->cbp = 63; |
1079 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; |
1080 |
|
|
1081 |
|
} else { /* BITS */ |
1082 |
|
|
1083 |
|
int bits, intra, i, cbp, c[2] = {0, 0}; |
1084 |
|
VECTOR backup[5], *v; |
1085 |
|
Data->iQuant = iQuant; |
1086 |
|
Data->cbp = c; |
1087 |
|
|
1088 |
|
v = Data->qpel ? Data->currentQMV : Data->currentMV; |
1089 |
|
for (i = 0; i < 5; i++) { |
1090 |
|
Data->iMinSAD[i] = 256*4096; |
1091 |
|
backup[i] = v[i]; |
1092 |
|
} |
1093 |
|
|
1094 |
|
bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags); |
1095 |
|
cbp = *Data->cbp; |
1096 |
|
|
1097 |
|
if (coding_type == S_VOP) { |
1098 |
|
int bits_gmc; |
1099 |
|
*Data->iMinSAD = bits += BITS_MULT*1; /* mcsel */ |
1100 |
|
bits_gmc = CountMBBitsGMC(Data, vGMC, x, y); |
1101 |
|
if (bits_gmc < bits) { |
1102 |
|
mcsel = 1; |
1103 |
|
*Data->iMinSAD = bits = bits_gmc; |
1104 |
|
mode = MODE_INTER; |
1105 |
|
cbp = *Data->cbp; |
1106 |
|
} |
1107 |
|
} |
1108 |
|
|
1109 |
|
if (inter4v) { |
1110 |
|
int bits_4v; |
1111 |
|
bits_4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); |
1112 |
|
if (bits_4v < bits) { |
1113 |
|
Data->iMinSAD[0] = bits = bits_4v; |
1114 |
|
mode = MODE_INTER4V; |
1115 |
|
cbp = *Data->cbp; |
1116 |
|
} |
1117 |
|
} |
1118 |
|
|
1119 |
|
intra = CountMBBitsIntra(Data); |
1120 |
|
if (intra < bits) { |
1121 |
|
*Data->iMinSAD = bits = intra; |
1122 |
|
mode = MODE_INTRA; |
1123 |
|
} |
1124 |
|
|
1125 |
|
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; |
1126 |
|
pMB->cbp = cbp; |
1127 |
|
} |
1128 |
|
|
1129 |
|
if (Data->rrv) { |
1130 |
|
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
1131 |
|
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
1132 |
|
} |
1133 |
|
|
1134 |
|
if (mode == MODE_INTER && mcsel == 0) { |
1135 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
1136 |
|
|
1137 |
|
if(Data->qpel) { |
1138 |
|
pMB->qmvs[0] = pMB->qmvs[1] |
1139 |
|
= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; |
1140 |
|
pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; |
1141 |
|
pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; |
1142 |
|
} else { |
1143 |
|
pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; |
1144 |
|
pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; |
1145 |
|
} |
1146 |
|
|
1147 |
|
} else if (mode == MODE_INTER ) { // but mcsel == 1 |
1148 |
|
|
1149 |
|
pMB->mcsel = 1; |
1150 |
|
if (Data->qpel) { |
1151 |
|
pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv; |
1152 |
|
pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2; |
1153 |
|
pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2; |
1154 |
|
} else |
1155 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv; |
1156 |
|
|
1157 |
|
} else |
1158 |
|
if (mode == MODE_INTER4V) ; /* anything here? */ |
1159 |
|
else /* INTRA, NOT_CODED */ |
1160 |
|
SkipMacroblockP(pMB, 0); |
1161 |
|
|
1162 |
|
pMB->mode = mode; |
1163 |
|
} |
1164 |
|
|
1165 |
bool |
bool |
1166 |
MotionEstimation(MBParam * const pParam, |
MotionEstimation(MBParam * const pParam, |
1167 |
FRAMEINFO * const current, |
FRAMEINFO * const current, |
1169 |
const IMAGE * const pRefH, |
const IMAGE * const pRefH, |
1170 |
const IMAGE * const pRefV, |
const IMAGE * const pRefV, |
1171 |
const IMAGE * const pRefHV, |
const IMAGE * const pRefHV, |
1172 |
|
const IMAGE * const pGMC, |
1173 |
const uint32_t iLimit) |
const uint32_t iLimit) |
1174 |
{ |
{ |
1175 |
MACROBLOCK *const pMBs = current->mbs; |
MACROBLOCK *const pMBs = current->mbs; |
1184 |
uint32_t x, y; |
uint32_t x, y; |
1185 |
uint32_t iIntra = 0; |
uint32_t iIntra = 0; |
1186 |
int32_t quant = current->quant, sad00; |
int32_t quant = current->quant, sad00; |
1187 |
int skip_thresh = \ |
int skip_thresh = INITIAL_SKIP_THRESH * \ |
|
INITIAL_SKIP_THRESH * \ |
|
1188 |
(current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \ |
(current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \ |
1189 |
(current->vop_flags & XVID_VOP_MODEDECISION_BITS ? 2:1); |
(current->vop_flags & XVID_VOP_MODEDECISION_BITS ? 2:1); |
1190 |
|
|
1191 |
// some pre-initialized thingies for SearchP |
/* some pre-initialized thingies for SearchP */ |
1192 |
int32_t temp[8]; |
int32_t temp[8]; |
1193 |
VECTOR currentMV[5]; |
VECTOR currentMV[5]; |
1194 |
VECTOR currentQMV[5]; |
VECTOR currentQMV[5]; |
1195 |
int32_t iMinSAD[5]; |
int32_t iMinSAD[5]; |
1196 |
DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE); |
DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE); |
1197 |
SearchData Data; |
SearchData Data; |
1198 |
memset(&Data, 0, sizeof(SearchData)); |
memset(&Data, 0, sizeof(SearchData)); |
1199 |
Data.iEdgedWidth = iEdgedWidth; |
Data.iEdgedWidth = iEdgedWidth; |
1205 |
Data.rounding = pParam->m_rounding_type; |
Data.rounding = pParam->m_rounding_type; |
1206 |
Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0); |
Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0); |
1207 |
Data.chroma = MotionFlags & XVID_ME_CHROMA16; |
Data.chroma = MotionFlags & XVID_ME_CHROMA16; |
1208 |
Data.rrv = (current->vop_flags & XVID_VOP_REDUCED ? 1:0); |
Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0; |
1209 |
Data.dctSpace = dct_space; |
Data.dctSpace = dct_space; |
1210 |
|
Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); |
1211 |
|
|
1212 |
if ((current->vop_flags & XVID_VOP_REDUCED)) { |
if ((current->vop_flags & XVID_VOP_REDUCED)) { |
1213 |
mb_width = (pParam->width + 31) / 32; |
mb_width = (pParam->width + 31) / 32; |
1215 |
Data.qpel = 0; |
Data.qpel = 0; |
1216 |
} |
} |
1217 |
|
|
1218 |
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) */ |
1219 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
1220 |
|
|
1221 |
for (y = 0; y < mb_height; y++) { |
for (y = 0; y < mb_height; y++) { |
1247 |
if (quant > 31) quant = 31; |
if (quant > 31) quant = 31; |
1248 |
else if (quant < 1) quant = 1; |
else if (quant < 1) quant = 1; |
1249 |
} |
} |
1250 |
|
pMB->quant = quant; |
1251 |
|
|
1252 |
pMB->quant = current->quant; |
/* initial skip decision */ |
|
|
|
|
//initial skip decision |
|
1253 |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
/* no early skip for GMC (global vector = skip vector is unknown!) */ |
1254 |
if (!(current->vol_flags & XVID_VOL_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ |
if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */ |
1255 |
if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh) |
if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh) |
1256 |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { |
1257 |
SkipMacroblockP(pMB, sad00); |
SkipMacroblockP(pMB, sad00); |
1259 |
} |
} |
1260 |
} |
} |
1261 |
|
|
1262 |
|
if ((current->vop_flags & XVID_VOP_CARTOON) && |
1263 |
|
(sad00 < pMB->quant * 4 * skip_thresh)) { /* favorize (0,0) vector for cartoons */ |
1264 |
|
ZeroMacroblockP(pMB, sad00); |
1265 |
|
continue; |
1266 |
|
} |
1267 |
|
|
1268 |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, |
1269 |
y, MotionFlags, current->vol_flags, pMB->quant, |
y, MotionFlags, current->vop_flags, current->vol_flags, |
1270 |
&Data, pParam, pMBs, reference->mbs, |
&Data, pParam, pMBs, reference->mbs, pMB); |
1271 |
current->vop_flags & XVID_VOP_INTER4V, pMB); |
|
1272 |
|
ModeDecision(&Data, pMB, pMBs, x, y, pParam, |
1273 |
|
MotionFlags, current->vop_flags, current->vol_flags, |
1274 |
|
pCurrent, pRef, pGMC, current->coding_type); |
1275 |
|
|
|
/* final skip decision, a.k.a. "the vector you found, really that good?" */ |
|
|
if (!(current->vol_flags & XVID_VOL_GMC || current->vop_flags & XVID_VOP_MODEDECISION_BITS)) { |
|
|
if ( pMB->dquant == 0 && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) { |
|
|
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); |
|
|
} |
|
|
} |
|
1276 |
if (pMB->mode == MODE_INTRA) |
if (pMB->mode == MODE_INTRA) |
1277 |
if (++iIntra > iLimit) return 1; |
if (++iIntra > iLimit) return 1; |
1278 |
} |
} |
1279 |
} |
} |
1280 |
|
|
1281 |
if (current->vol_flags & XVID_VOL_GMC ) /* GMC only for S(GMC)-VOPs */ |
// if (current->vol_flags & XVID_VOL_GMC ) /* GMC only for S(GMC)-VOPs */ |
1282 |
current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); |
// { |
1283 |
|
// current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); |
1284 |
|
// } |
1285 |
return 0; |
return 0; |
1286 |
} |
} |
1287 |
|
|
1291 |
{ |
{ |
1292 |
int mask = 255, j; |
int mask = 255, j; |
1293 |
for (j = 0; j < i; j++) { |
for (j = 0; j < i; j++) { |
1294 |
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 */ |
1295 |
if (pmv[i].x == pmv[j].x) { |
if (pmv[i].x == pmv[j].x) { |
1296 |
if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4; |
1297 |
else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; |
else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; |
1308 |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, |
1309 |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
int iHcount, const MACROBLOCK * const prevMB, int rrv) |
1310 |
{ |
{ |
1311 |
|
/* 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 |
|
1312 |
if (rrv) { iWcount /= 2; iHcount /= 2; } |
if (rrv) { iWcount /= 2; iHcount /= 2; } |
1313 |
|
|
1314 |
if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour |
if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */ |
1315 |
pmv[5].x = EVEN(pmv[3].x); |
pmv[5].x = EVEN(pmv[3].x); |
1316 |
pmv[5].y = EVEN(pmv[3].y); |
pmv[5].y = EVEN(pmv[3].y); |
1317 |
} else pmv[5].x = pmv[5].y = 0; |
} else pmv[5].x = pmv[5].y = 0; |
1318 |
|
|
1319 |
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 */ |
1320 |
else pmv[3].x = pmv[3].y = 0; |
else pmv[3].x = pmv[3].y = 0; |
1321 |
|
|
1322 |
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 */ |
1323 |
else pmv[4].x = pmv[4].y = 0; |
else pmv[4].x = pmv[4].y = 0; |
1324 |
|
|
1325 |
// [1] median prediction |
/* [1] median prediction */ |
1326 |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); |
1327 |
|
|
1328 |
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 */ |
1329 |
|
|
1330 |
pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame |
pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */ |
1331 |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
pmv[2].y = EVEN(prevMB->mvs[0].y); |
1332 |
|
|
1333 |
if ((x < iWcount-1) && (y < iHcount-1)) { |
if ((x < iWcount-1) && (y < iHcount-1)) { |
1334 |
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 */ |
1335 |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); |
1336 |
} else pmv[6].x = pmv[6].y = 0; |
} else pmv[6].x = pmv[6].y = 0; |
1337 |
|
|
1344 |
} |
} |
1345 |
} |
} |
1346 |
|
|
|
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 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; |
|
|
} |
|
|
|
|
|
/* 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; |
|
|
} |
|
|
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 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; return MODE_INTRA; } |
|
|
|
|
|
return mode; |
|
|
} |
|
|
} |
|
|
|
|
1347 |
static void |
static void |
1348 |
SearchP(const IMAGE * const pRef, |
SearchP(const IMAGE * const pRef, |
1349 |
const uint8_t * const pRefH, |
const uint8_t * const pRefH, |
1354 |
const int y, |
const int y, |
1355 |
const uint32_t MotionFlags, |
const uint32_t MotionFlags, |
1356 |
const uint32_t VopFlags, |
const uint32_t VopFlags, |
1357 |
const uint32_t iQuant, |
const uint32_t VolFlags, |
1358 |
SearchData * const Data, |
SearchData * const Data, |
1359 |
const MBParam * const pParam, |
const MBParam * const pParam, |
1360 |
const MACROBLOCK * const pMBs, |
const MACROBLOCK * const pMBs, |
1361 |
const MACROBLOCK * const prevMBs, |
const MACROBLOCK * const prevMBs, |
|
int inter4v, |
|
1362 |
MACROBLOCK * const pMB) |
MACROBLOCK * const pMB) |
1363 |
{ |
{ |
1364 |
|
|
1365 |
int i, iDirection = 255, mask, threshA; |
int i, iDirection = 255, mask, threshA; |
1366 |
VECTOR pmv[7]; |
VECTOR pmv[7]; |
1367 |
|
int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); |
1368 |
|
|
1369 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1370 |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
1371 |
|
|
1372 |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); |
1373 |
|
|
1374 |
Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache |
Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */ |
1375 |
i = Data->rrv ? 2 : 1; |
i = Data->rrv ? 2 : 1; |
1376 |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; |
1377 |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1384 |
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1385 |
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; |
1386 |
|
|
1387 |
Data->lambda16 = lambda_vec16[iQuant]; |
Data->lambda16 = lambda_vec16[pMB->quant]; |
1388 |
Data->lambda8 = lambda_vec8[iQuant]; |
Data->lambda8 = lambda_vec8[pMB->quant]; |
1389 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
1390 |
|
|
|
if (pMB->dquant != 0) inter4v = 0; |
|
|
|
|
1391 |
memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
memset(Data->currentMV, 0, 5*sizeof(VECTOR)); |
1392 |
|
|
1393 |
if (Data->qpel) 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); |
1400 |
Data->iMinSAD[3] = pMB->sad8[2]; |
Data->iMinSAD[3] = pMB->sad8[2]; |
1401 |
Data->iMinSAD[4] = pMB->sad8[3]; |
Data->iMinSAD[4] = pMB->sad8[3]; |
1402 |
|
|
1403 |
if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) || (x | y)) { |
if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) && (x | y)) { |
1404 |
threshA = Data->temp[0]; // that's where we keep this SAD atm |
threshA = Data->temp[0]; /* that's where we keep this SAD atm */ |
1405 |
if (threshA < 512) threshA = 512; |
if (threshA < 512) threshA = 512; |
1406 |
else if (threshA > 1024) threshA = 1024; |
else if (threshA > 1024) threshA = 1024; |
1407 |
} else |
} else |
1412 |
|
|
1413 |
if (!Data->rrv) { |
if (!Data->rrv) { |
1414 |
if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; |
if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; |
1415 |
else CheckCandidate = CheckCandidate16no4v; //for extra speed |
else CheckCandidate = CheckCandidate16no4v; /* for extra speed */ |
1416 |
} else CheckCandidate = CheckCandidate32; |
} else CheckCandidate = CheckCandidate32; |
1417 |
|
|
1418 |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ |
1425 |
|
|
1426 |
if ((Data->iMinSAD[0] <= threshA) || |
if ((Data->iMinSAD[0] <= threshA) || |
1427 |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && |
1428 |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { |
(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) |
1429 |
if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) inter4v = 0; } |
inter4v = 0; |
1430 |
else { |
else { |
1431 |
|
|
1432 |
MainSearchFunc * MainSearchPtr; |
MainSearchFunc * MainSearchPtr; |
1472 |
} |
} |
1473 |
|
|
1474 |
if (MotionFlags & XVID_ME_HALFPELREFINE16) |
if (MotionFlags & XVID_ME_HALFPELREFINE16) |
|
if ((!(MotionFlags & XVID_ME_HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant) |
|
1475 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1476 |
|
|
1477 |
for(i = 0; i < 5; i++) { |
for(i = 0; i < 5; i++) { |
1478 |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors |
Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */ |
1479 |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
1480 |
} |
} |
1481 |
|
|
1482 |
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) { |
if (Data->qpel) { |
|
|
|
1483 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
1484 |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
|
|
|
|
if ((!(MotionFlags & XVID_ME_QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) { |
|
1485 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
1486 |
|
if (MotionFlags & XVID_ME_QUARTERPELREFINE16) |
1487 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1488 |
} |
} |
|
} |
|
1489 |
|
|
1490 |
if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0; |
if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30) |
1491 |
|
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 |
|
1492 |
|
|
1493 |
|
if (inter4v) { |
1494 |
SearchData Data8; |
SearchData Data8; |
1495 |
memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data |
memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */ |
1496 |
|
|
1497 |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); |
1498 |
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); |
1500 |
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); |
1501 |
|
|
1502 |
if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_BITS))) { |
if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_BITS))) { |
1503 |
// 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 */ |
1504 |
int sumx = 0, sumy = 0; |
int sumx = 0, sumy = 0; |
|
const int div = 1 + Data->qpel; |
|
|
const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; |
|
1505 |
|
|
1506 |
for (i = 0; i < 4; i++) { |
if (Data->qpel) |
1507 |
sumx += mv[i].x / div; |
for (i = 1; i < 5; i++) { |
1508 |
sumy += mv[i].y / div; |
sumx += Data->currentQMV[i].x/2; |
1509 |
|
sumy += Data->currentQMV[i].y/2; |
1510 |
|
} |
1511 |
|
else |
1512 |
|
for (i = 1; i < 5; i++) { |
1513 |
|
sumx += Data->currentMV[i].x; |
1514 |
|
sumy += Data->currentMV[i].y; |
1515 |
} |
} |
1516 |
|
|
1517 |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], |
1518 |
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
(sumy >> 3) + roundtab_76[sumy & 0xf], Data); |
1519 |
} |
} |
1520 |
|
} else Data->iMinSAD[1] = 4096*256; |
1521 |
} |
} |
1522 |
|
|
1523 |
inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, VopFlags); |
static void |
1524 |
|
Search8(const SearchData * const OldData, |
1525 |
if (Data->rrv) { |
const int x, const int y, |
1526 |
Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); |
const uint32_t MotionFlags, |
1527 |
Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); |
const MBParam * const pParam, |
1528 |
} |
MACROBLOCK * const pMB, |
1529 |
|
const MACROBLOCK * const pMBs, |
1530 |
if (inter4v == MODE_INTER) { |
const int block, |
1531 |
pMB->mode = MODE_INTER; |
SearchData * const Data) |
1532 |
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
{ |
1533 |
pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; |
int i = 0; |
1534 |
|
Data->iMinSAD = OldData->iMinSAD + 1 + block; |
1535 |
if(Data->qpel) { |
Data->currentMV = OldData->currentMV + 1 + block; |
1536 |
pMB->qmvs[0] = pMB->qmvs[1] |
Data->currentQMV = OldData->currentQMV + 1 + block; |
|
= 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; |
|
|
} |
|
|
|
|
|
} |
|
|
|
|
|
static void |
|
|
Search8(const SearchData * const OldData, |
|
|
const int x, const int y, |
|
|
const uint32_t MotionFlags, |
|
|
const MBParam * const pParam, |
|
|
MACROBLOCK * const pMB, |
|
|
const MACROBLOCK * const pMBs, |
|
|
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; |
|
1537 |
|
|
1538 |
if(Data->qpel) { |
if(Data->qpel) { |
1539 |
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); |
1566 |
else CheckCandidate = CheckCandidate16no4v; |
else CheckCandidate = CheckCandidate16no4v; |
1567 |
|
|
1568 |
if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_BITS))) { |
if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_BITS))) { |
1569 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
1570 |
|
|
1571 |
MainSearchFunc *MainSearchPtr; |
MainSearchFunc *MainSearchPtr; |
1572 |
if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch; |
if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch; |
1576 |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); |
MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); |
1577 |
|
|
1578 |
if(*(Data->iMinSAD) < temp_sad) { |
if(*(Data->iMinSAD) < temp_sad) { |
1579 |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
1580 |
Data->currentQMV->y = 2 * Data->currentMV->y; |
Data->currentQMV->y = 2 * Data->currentMV->y; |
1581 |
} |
} |
1582 |
} |
} |
1583 |
|
|
1584 |
if (MotionFlags & XVID_ME_HALFPELREFINE8) { |
if (MotionFlags & XVID_ME_HALFPELREFINE8) { |
1585 |
int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD |
int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ |
1586 |
|
|
1587 |
SubpelRefine(Data); // perform halfpel refine of current best vector |
SubpelRefine(Data); /* perform halfpel refine of current best vector */ |
1588 |
|
|
1589 |
if(*(Data->iMinSAD) < temp_sad) { // we have found a better match |
if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */ |
1590 |
Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector |
Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ |
1591 |
Data->currentQMV->y = 2 * Data->currentMV->y; |
Data->currentQMV->y = 2 * Data->currentMV->y; |
1592 |
} |
} |
1593 |
} |
} |
1634 |
const uint32_t mode_curr) |
const uint32_t mode_curr) |
1635 |
{ |
{ |
1636 |
|
|
1637 |
// [0] is prediction |
/* [0] is prediction */ |
1638 |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); |
1639 |
|
|
1640 |
pmv[1].x = pmv[1].y = 0; // [1] is zero |
pmv[1].x = pmv[1].y = 0; /* [1] is zero */ |
1641 |
|
|
1642 |
pmv[2] = ChoosePred(pMB, mode_curr); |
pmv[2] = ChoosePred(pMB, mode_curr); |
1643 |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); |
1644 |
|
|
1645 |
if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour |
if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */ |
1646 |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); |
1647 |
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); |
1648 |
} else pmv[3].x = pmv[3].y = 0; |
} else pmv[3].x = pmv[3].y = 0; |
1688 |
*Data->iMinSAD = MV_MAX_ERROR; |
*Data->iMinSAD = MV_MAX_ERROR; |
1689 |
Data->iFcode = iFcode; |
Data->iFcode = iFcode; |
1690 |
Data->qpel_precision = 0; |
Data->qpel_precision = 0; |
1691 |
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 */ |
1692 |
|
|
1693 |
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; |
1694 |
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; |
Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; |
1710 |
Data->currentMV->x = Data->currentMV->y = 0; |
Data->currentMV->x = Data->currentMV->y = 0; |
1711 |
CheckCandidate = CheckCandidate16no4v; |
CheckCandidate = CheckCandidate16no4v; |
1712 |
|
|
1713 |
// main loop. checking all predictions |
/* main loop. checking all predictions */ |
1714 |
for (i = 0; i < 7; i++) { |
for (i = 0; i < 7; i++) { |
1715 |
if (!(mask = make_mask(pmv, i)) ) continue; |
if (!(mask = make_mask(pmv, i)) ) continue; |
1716 |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); |
1733 |
SubpelRefine(Data); |
SubpelRefine(Data); |
1734 |
} |
} |
1735 |
|
|
1736 |
// three bits are needed to code backward mode. four for forward |
/* three bits are needed to code backward mode. four for forward */ |
1737 |
|
|
1738 |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; |
1739 |
else *Data->iMinSAD += 3 * Data->lambda16; |
else *Data->iMinSAD += 3 * Data->lambda16; |
1757 |
} |
} |
1758 |
|
|
1759 |
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; |
1760 |
else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search |
else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */ |
1761 |
} |
} |
1762 |
|
|
1763 |
static void |
static void |
1773 |
const int div = 1 + Data->qpel; |
const int div = 1 + Data->qpel; |
1774 |
int k; |
int k; |
1775 |
const uint32_t stride = Data->iEdgedWidth/2; |
const uint32_t stride = Data->iEdgedWidth/2; |
1776 |
//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 */ |
1777 |
|
|
1778 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1779 |
dy += Data->directmvF[k].y / div; |
dy += Data->directmvF[k].y / div; |
1792 |
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, |
1793 |
stride); |
stride); |
1794 |
|
|
1795 |
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 */ |
1796 |
|
|
1797 |
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
sum += sad8bi(pCur->v + 8*x + 8 * y * stride, |
1798 |
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, |
1800 |
stride); |
stride); |
1801 |
|
|
1802 |
if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { |
1803 |
pMB->mode = MODE_DIRECT_NONE_MV; //skipped |
pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */ |
1804 |
for (k = 0; k < 4; k++) { |
for (k = 0; k < 4; k++) { |
1805 |
pMB->qmvs[k] = pMB->mvs[k]; |
pMB->qmvs[k] = pMB->mvs[k]; |
1806 |
pMB->b_qmvs[k] = pMB->b_mvs[k]; |
pMB->b_qmvs[k] = pMB->b_mvs[k]; |
1864 |
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx) |
1865 |
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
| (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) { |
1866 |
|
|
1867 |
*best_sad = 256*4096; // in that case, we won't use direct mode |
*best_sad = 256*4096; /* in that case, we won't use direct mode */ |
1868 |
pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" |
pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */ |
1869 |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; |
1870 |
return 256*4096; |
return 256*4096; |
1871 |
} |
} |
1882 |
|
|
1883 |
CheckCandidate(0, 0, 255, &k, Data); |
CheckCandidate(0, 0, 255, &k, Data); |
1884 |
|
|
1885 |
// initial (fast) skip decision |
/* initial (fast) skip decision */ |
1886 |
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { |
if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) { |
1887 |
//possible skip |
/* possible skip */ |
1888 |
if (Data->chroma) { |
if (Data->chroma) { |
1889 |
pMB->mode = MODE_DIRECT_NONE_MV; |
pMB->mode = MODE_DIRECT_NONE_MV; |
1890 |
return *Data->iMinSAD; // skip. |
return *Data->iMinSAD; /* skip. */ |
1891 |
} else { |
} else { |
1892 |
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); |
1893 |
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. |
if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */ |
1894 |
} |
} |
1895 |
} |
} |
1896 |
|
|
1897 |
*Data->iMinSAD += Data->lambda16; |
*Data->iMinSAD += Data->lambda16; |
1898 |
skip_sad = *Data->iMinSAD; |
skip_sad = *Data->iMinSAD; |
1899 |
|
|
1900 |
// DIRECT MODE DELTA VECTOR SEARCH. |
/* |
1901 |
// This has to be made more effective, but at the moment I'm happy it's running at all |
* DIRECT MODE DELTA VECTOR SEARCH. |
1902 |
|
* This has to be made more effective, but at the moment I'm happy it's running at all |
1903 |
|
*/ |
1904 |
|
|
1905 |
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; |
1906 |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; |
1913 |
*best_sad = *Data->iMinSAD; |
*best_sad = *Data->iMinSAD; |
1914 |
|
|
1915 |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT; |
1916 |
else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation |
else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */ |
1917 |
|
|
1918 |
pMB->pmvs[3] = *Data->currentMV; |
pMB->pmvs[3] = *Data->currentMV; |
1919 |
|
|
1971 |
SearchData bData; |
SearchData bData; |
1972 |
|
|
1973 |
fData->qpel_precision = 0; |
fData->qpel_precision = 0; |
1974 |
memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data |
memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */ |
1975 |
*fData->iMinSAD = 4096*256; |
*fData->iMinSAD = 4096*256; |
1976 |
bData.currentMV++; bData.currentQMV++; |
bData.currentMV++; bData.currentQMV++; |
1977 |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; |
2010 |
|
|
2011 |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); |
2012 |
|
|
2013 |
//diamond |
/* diamond */ |
2014 |
do { |
do { |
2015 |
iDirection = 255; |
iDirection = 255; |
2016 |
// forward MV moves |
/* forward MV moves */ |
2017 |
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
i = fData->currentMV[0].x; j = fData->currentMV[0].y; |
2018 |
|
|
2019 |
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
CheckCandidateInt(i + 1, j, 0, &iDirection, fData); |
2021 |
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
CheckCandidateInt(i - 1, j, 0, &iDirection, fData); |
2022 |
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
CheckCandidateInt(i, j - 1, 0, &iDirection, fData); |
2023 |
|
|
2024 |
// backward MV moves |
/* backward MV moves */ |
2025 |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
i = fData->currentMV[1].x; j = fData->currentMV[1].y; |
2026 |
fData->currentMV[2] = fData->currentMV[0]; |
fData->currentMV[2] = fData->currentMV[0]; |
2027 |
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); |
2031 |
|
|
2032 |
} while (!(iDirection)); |
} while (!(iDirection)); |
2033 |
|
|
2034 |
//qpel refinement |
/* qpel refinement */ |
2035 |
if (fData->qpel) { |
if (fData->qpel) { |
2036 |
if (*fData->iMinSAD > *best_sad + 500) return; |
if (*fData->iMinSAD > *best_sad + 500) return; |
2037 |
CheckCandidate = CheckCandidateInt; |
CheckCandidate = CheckCandidateInt; |
2048 |
SubpelRefine(&bData); |
SubpelRefine(&bData); |
2049 |
} |
} |
2050 |
|
|
2051 |
*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. */ |
2052 |
|
|
2053 |
if (*fData->iMinSAD < *best_sad) { |
if (*fData->iMinSAD < *best_sad) { |
2054 |
*best_sad = *fData->iMinSAD; |
*best_sad = *fData->iMinSAD; |
2076 |
FRAMEINFO * const frame, |
FRAMEINFO * const frame, |
2077 |
const int32_t time_bp, |
const int32_t time_bp, |
2078 |
const int32_t time_pp, |
const int32_t time_pp, |
2079 |
// forward (past) reference |
/* forward (past) reference */ |
2080 |
const MACROBLOCK * const f_mbs, |
const MACROBLOCK * const f_mbs, |
2081 |
const IMAGE * const f_ref, |
const IMAGE * const f_ref, |
2082 |
const IMAGE * const f_refH, |
const IMAGE * const f_refH, |
2083 |
const IMAGE * const f_refV, |
const IMAGE * const f_refV, |
2084 |
const IMAGE * const f_refHV, |
const IMAGE * const f_refHV, |
2085 |
// backward (future) reference |
/* backward (future) reference */ |
2086 |
const FRAMEINFO * const b_reference, |
const FRAMEINFO * const b_reference, |
2087 |
const IMAGE * const b_ref, |
const IMAGE * const b_ref, |
2088 |
const IMAGE * const b_refH, |
const IMAGE * const b_refH, |
2100 |
const int32_t TRB = time_pp - time_bp; |
const int32_t TRB = time_pp - time_bp; |
2101 |
const int32_t TRD = time_pp; |
const int32_t TRD = time_pp; |
2102 |
|
|
2103 |
// some pre-inintialized data for the rest of the search |
/* some pre-inintialized data for the rest of the search */ |
2104 |
|
|
2105 |
SearchData Data; |
SearchData Data; |
2106 |
int32_t iMinSAD; |
int32_t iMinSAD; |
2117 |
Data.chroma = frame->motion_flags & XVID_ME_CHROMA8; |
Data.chroma = frame->motion_flags & XVID_ME_CHROMA8; |
2118 |
Data.temp = temp; |
Data.temp = temp; |
2119 |
|
|
2120 |
Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) |
Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */ |
2121 |
// note: i==horizontal, j==vertical |
|
2122 |
|
/* note: i==horizontal, j==vertical */ |
2123 |
for (j = 0; j < pParam->mb_height; j++) { |
for (j = 0; j < pParam->mb_height; j++) { |
2124 |
|
|
2125 |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ |
2155 |
|
|
2156 |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } |
2157 |
|
|
2158 |
// forward search |
/* forward search */ |
2159 |
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2160 |
&frame->image, i, j, |
&frame->image, i, j, |
2161 |
frame->motion_flags, |
frame->motion_flags, |
2163 |
pMB, &f_predMV, &best_sad, |
pMB, &f_predMV, &best_sad, |
2164 |
MODE_FORWARD, &Data); |
MODE_FORWARD, &Data); |
2165 |
|
|
2166 |
// backward search |
/* backward search */ |
2167 |
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2168 |
&frame->image, i, j, |
&frame->image, i, j, |
2169 |
frame->motion_flags, |
frame->motion_flags, |
2171 |
pMB, &b_predMV, &best_sad, |
pMB, &b_predMV, &best_sad, |
2172 |
MODE_BACKWARD, &Data); |
MODE_BACKWARD, &Data); |
2173 |
|
|
2174 |
// 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 */ |
2175 |
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y, |
2176 |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
b_ref, b_refH->y, b_refV->y, b_refHV->y, |
2177 |
&frame->image, |
&frame->image, |
2183 |
pMB, &best_sad, |
pMB, &best_sad, |
2184 |
&Data); |
&Data); |
2185 |
|
|
2186 |
// final skip decision |
/* final skip decision */ |
2187 |
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2) |
if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2) |
2188 |
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
&& ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) |
2189 |
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data); |
2229 |
|
|
2230 |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; |
2231 |
|
|
2232 |
//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 */ |
2233 |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; |
2234 |
else |
else |
2235 |
if (x == 1) //left macroblock does not have any vector now |
if (x == 1) /* left macroblock does not have any vector now */ |
2236 |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median |
Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */ |
2237 |
else if (y == 1) // top macroblock doesn't have it's vector |
else if (y == 1) /* top macroblock doesn't have it's vector */ |
2238 |
Data->predMV = (pMB - 1)->mvs[0]; // left instead of median |
Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */ |
2239 |
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median |
else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */ |
2240 |
|
|
2241 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2242 |
pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0); |
pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0); |
2259 |
if (!(mask = make_mask(pmv, 2))) |
if (!(mask = make_mask(pmv, 2))) |
2260 |
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); |
2261 |
|
|
2262 |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) // diamond only if needed |
if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) /* diamond only if needed */ |
2263 |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); |
2264 |
} |
} |
2265 |
|
|
2271 |
} |
} |
2272 |
} |
} |
2273 |
|
|
2274 |
#define INTRA_THRESH 2400 |
#define INTRA_THRESH 2200 |
2275 |
#define INTER_THRESH 1300 |
#define INTER_THRESH 50 |
2276 |
|
#define INTRA_THRESH2 95 |
2277 |
|
|
2278 |
int |
int |
2279 |
MEanalysis( const IMAGE * const pRef, |
MEanalysis( const IMAGE * const pRef, |
2288 |
int sSAD = 0; |
int sSAD = 0; |
2289 |
MACROBLOCK * const pMBs = Current->mbs; |
MACROBLOCK * const pMBs = Current->mbs; |
2290 |
const IMAGE * const pCurrent = &Current->image; |
const IMAGE * const pCurrent = &Current->image; |
2291 |
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh; |
int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh; |
2292 |
int s = 0, blocks = 0; |
int blocks = 0; |
2293 |
|
int complexity = 0; |
2294 |
|
|
2295 |
int32_t iMinSAD[5], temp[5]; |
int32_t iMinSAD[5], temp[5]; |
2296 |
VECTOR currentMV[5]; |
VECTOR currentMV[5]; |
2302 |
Data.temp = temp; |
Data.temp = temp; |
2303 |
CheckCandidate = CheckCandidate32I; |
CheckCandidate = CheckCandidate32I; |
2304 |
|
|
2305 |
if (intraCount != 0 && intraCount < 10) // we're right after an I frame |
if (intraCount != 0) { |
2306 |
IntraThresh += 8 * (intraCount - 10) * (intraCount - 10); |
if (intraCount < 10) // we're right after an I frame |
2307 |
|
IntraThresh += 15* (intraCount - 10) * (intraCount - 10); |
2308 |
else |
else |
2309 |
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec |
2310 |
IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; |
IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra; |
2311 |
|
} |
2312 |
|
|
2313 |
InterThresh -= (350 - 8*b_thresh) * bCount; |
InterThresh -= 12 * bCount; |
2314 |
if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh; |
if (InterThresh < 15 + b_thresh) InterThresh = 15 + b_thresh; |
2315 |
|
|
2316 |
if (sadInit) (*sadInit) (); |
if (sadInit) (*sadInit) (); |
2317 |
|
|
2318 |
for (y = 1; y < pParam->mb_height-1; y += 2) { |
for (y = 1; y < pParam->mb_height-1; y += 2) { |
2319 |
for (x = 1; x < pParam->mb_width-1; x += 2) { |
for (x = 1; x < pParam->mb_width-1; x += 2) { |
2320 |
int i; |
int i; |
2321 |
blocks += 4; |
blocks += 10; |
2322 |
|
|
2323 |
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; |
2324 |
else { //extrapolation of the vector found for last frame |
else { //extrapolation of the vector found for last frame |
2333 |
for (i = 0; i < 4; i++) { |
for (i = 0; i < 4; i++) { |
2334 |
int dev; |
int dev; |
2335 |
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width]; |
|
if (pMB->sad16 > IntraThresh) { |
|
2336 |
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, |
2337 |
pParam->edged_width); |
pParam->edged_width); |
2338 |
|
|
2339 |
|
complexity += dev; |
2340 |
if (dev + IntraThresh < pMB->sad16) { |
if (dev + IntraThresh < pMB->sad16) { |
2341 |
pMB->mode = MODE_INTRA; |
pMB->mode = MODE_INTRA; |
2342 |
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; |
2343 |
} |
} |
2344 |
} |
|
2345 |
if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++; |
if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) |
2346 |
|
if (dev > 500 && pMB->sad16 < 1000) |
2347 |
|
sSAD += 1000; |
2348 |
|
|
2349 |
sSAD += pMB->sad16; |
sSAD += pMB->sad16; |
2350 |
} |
} |
2351 |
} |
} |
2352 |
} |
} |
2353 |
|
complexity >>= 7; |
2354 |
|
|
2355 |
sSAD /= blocks; |
sSAD /= complexity + 4*blocks; |
|
s = (10*s) / blocks; |
|
|
|
|
|
if (s > 4) sSAD += (s - 3) * (300 - 2*b_thresh); //static block - looks bad when in bframe... |
|
2356 |
|
|
2357 |
|
if (intraCount > 12 && sSAD > INTRA_THRESH2 ) return I_VOP; |
2358 |
if (sSAD > InterThresh ) return P_VOP; |
if (sSAD > InterThresh ) return P_VOP; |
2359 |
emms(); |
emms(); |
2360 |
return B_VOP; |
return B_VOP; |
2361 |
} |
} |
2362 |
|
|
2363 |
|
|
2364 |
static WARPPOINTS |
/* functions which perform BITS-based search/bitcount */ |
|
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 ) |
|
|
{ |
|
|
|
|
|
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 |
|
|
|
|
|
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"); |
|
|
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 |
|
|
|
|
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
|
|
for (mx = 1; mx < (uint32_t)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 < (uint32_t)MBh-1; my++) |
|
|
for (mx = 1; mx < (uint32_t)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) <= (uint32_t)grad ) |
|
|
MBmask[mbnum] = 0; |
|
|
if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)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 < (uint32_t)MBh; my++) |
|
|
for (mx = 0; mx < (uint32_t)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 < (uint32_t)MBh; my++) |
|
|
for (mx = 0; mx < (uint32_t)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 += 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 */ |
|
|
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 < (uint32_t)MBh; my++) |
|
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
|
|
{ |
|
|
const int mbnum = mx + my * MBw; |
|
|
const MACROBLOCK *pMB = &pMBs[mbnum]; |
|
|
const VECTOR mv = pMB->mvs[0]; |
|
|
|
|
|
if (!MBmask[mbnum]) |
|
|
continue; |
|
|
|
|
|
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++; |
|
|
} |
|
|
|
|
|
} 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); |
|
|
|
|
|
free(MBmask); |
|
|
|
|
|
return gmc; |
|
|
} |
|
|
|
|
|
// functions which perform BITS-based search/bitcount |
|
2365 |
|
|
2366 |
static int |
static int |
2367 |
CountMBBitsInter(SearchData * const Data, |
CountMBBitsInter(SearchData * const Data, |
2382 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
2383 |
CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); |
CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); |
2384 |
|
|
2385 |
//checking if this vector is perfect. if it is, we stop. |
if (MotionFlags & (XVID_ME_HALFPELREFINE16_BITS | XVID_ME_EXTSEARCH_BITS)) { /* we have to prepare for halfpixel-precision search */ |
|
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 |
|
2386 |
for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; |
2387 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2388 |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); |
2391 |
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
2392 |
} |
} |
2393 |
|
|
2394 |
} else { // not qpel |
} else { /* not qpel */ |
2395 |
|
|
2396 |
CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); |
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 |
|
|
} |
|
2397 |
} |
} |
2398 |
|
|
2399 |
if (MotionFlags&XVID_ME_EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
if (MotionFlags&XVID_ME_EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); |
2401 |
if (MotionFlags&XVID_ME_HALFPELREFINE16_BITS) SubpelRefine(Data); |
if (MotionFlags&XVID_ME_HALFPELREFINE16_BITS) SubpelRefine(Data); |
2402 |
|
|
2403 |
if (Data->qpel) { |
if (Data->qpel) { |
2404 |
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 */ |
2405 |
for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { |
2406 |
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 */ |
2407 |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
Data->currentQMV[i].y = 2 * Data->currentMV[i].y; |
2408 |
} |
} |
2409 |
|
|
2410 |
// preparing for qpel-precision search |
/* preparing for qpel-precision search */ |
2411 |
Data->qpel_precision = 1; |
Data->qpel_precision = 1; |
2412 |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2413 |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
pParam->width, pParam->height, Data->iFcode, 1, 0); |
2415 |
if (MotionFlags&XVID_ME_QUARTERPELREFINE16_BITS) SubpelRefine(Data); |
if (MotionFlags&XVID_ME_QUARTERPELREFINE16_BITS) SubpelRefine(Data); |
2416 |
} |
} |
2417 |
|
|
2418 |
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 */ |
2419 |
VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; |
2420 |
if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) |
if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) |
2421 |
CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); |
CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); |
2423 |
return Data->iMinSAD[0]; |
return Data->iMinSAD[0]; |
2424 |
} |
} |
2425 |
|
|
|
|
|
2426 |
static int |
static int |
2427 |
CountMBBitsInter4v(const SearchData * const Data, |
CountMBBitsInter4v(const SearchData * const Data, |
2428 |
MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, |
MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, |
2435 |
SearchData Data2, *Data8 = &Data2; |
SearchData Data2, *Data8 = &Data2; |
2436 |
int sumx = 0, sumy = 0; |
int sumx = 0, sumy = 0; |
2437 |
int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; |
int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; |
2438 |
|
uint8_t * ptr; |
2439 |
|
|
2440 |
memcpy(Data8, Data, sizeof(SearchData)); |
memcpy(Data8, Data, sizeof(SearchData)); |
2441 |
CheckCandidate = CheckCandidateBits8; |
CheckCandidate = CheckCandidateBits8; |
2442 |
|
|
2443 |
for (i = 0; i < 4; i++) { |
for (i = 0; i < 4; i++) { /* for all luma blocks */ |
2444 |
|
|
2445 |
Data8->iMinSAD = Data->iMinSAD + i + 1; |
Data8->iMinSAD = Data->iMinSAD + i + 1; |
2446 |
Data8->currentMV = Data->currentMV + i + 1; |
Data8->currentMV = Data->currentMV + i + 1; |
2447 |
Data8->currentQMV = Data->currentQMV + i + 1; |
Data8->currentQMV = Data->currentQMV + i + 1; |
2450 |
Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2451 |
Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2452 |
Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2453 |
|
*Data8->cbp = (Data->cbp[1] & (1<<(5-i))) ? 1:0; // copy corresponding cbp bit |
2454 |
|
// *Data8->cbp = 1; |
2455 |
|
|
2456 |
if(Data->qpel) { |
if(Data->qpel) { |
2457 |
Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); |
Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); |
2466 |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, |
2467 |
pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); |
pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); |
2468 |
|
|
2469 |
*Data8->iMinSAD += t; |
*Data8->iMinSAD += BITS_MULT*t; |
2470 |
|
|
2471 |
Data8->qpel_precision = Data8->qpel; |
Data8->qpel_precision = Data8->qpel; |
2472 |
// checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) |
/* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */ |
2473 |
if (Data8->qpel) { |
{ |
2474 |
if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y)) |
VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; |
2475 |
CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
if (!MVequal (*v, backup[i+1]) ) |
|
} else { |
|
|
if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y)) |
|
2476 |
CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); |
2477 |
} |
} |
2478 |
|
|
2479 |
if (Data8->qpel) { |
if (Data8->qpel) { |
2480 |
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 */ |
2481 |
int32_t s = *Data8->iMinSAD; |
int32_t s = *Data8->iMinSAD; |
2482 |
Data8->currentMV->x = Data8->currentQMV->x/2; |
Data8->currentMV->x = Data8->currentQMV->x/2; |
2483 |
Data8->currentMV->y = Data8->currentQMV->y/2; |
Data8->currentMV->y = Data8->currentQMV->y/2; |
2491 |
if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) |
if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) |
2492 |
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
2493 |
|
|
2494 |
if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8); |
if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) |
2495 |
|
SubpelRefine(Data8); |
2496 |
|
|
2497 |
if(s > *Data8->iMinSAD) { //we have found a better match |
if(s > *Data8->iMinSAD) { /* we have found a better match */ |
2498 |
Data8->currentQMV->x = 2*Data8->currentMV->x; |
Data8->currentQMV->x = 2*Data8->currentMV->x; |
2499 |
Data8->currentQMV->y = 2*Data8->currentMV->y; |
Data8->currentQMV->y = 2*Data8->currentMV->y; |
2500 |
} |
} |
2506 |
} |
} |
2507 |
if (MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS) SubpelRefine(Data8); |
if (MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS) SubpelRefine(Data8); |
2508 |
|
|
2509 |
} else // not qpel |
} else { /* not qpel */ |
2510 |
if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement |
|
2511 |
|
if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) /* extsearch */ |
2512 |
|
SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); |
2513 |
|
|
2514 |
|
if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) |
2515 |
|
SubpelRefine(Data8); /* halfpel refinement */ |
2516 |
|
} |
2517 |
|
|
2518 |
//checking vector equal to predicion |
/* checking vector equal to predicion */ |
2519 |
if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_BITS) { |
if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_BITS) { |
2520 |
const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; |
2521 |
if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y)) |
if (!MVequal(*v, Data8->predMV)) |
2522 |
CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); |
CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); |
2523 |
} |
} |
2524 |
|
|
2525 |
bits += *Data8->iMinSAD; |
bits += *Data8->iMinSAD; |
2526 |
if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V |
if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */ |
2527 |
|
|
2528 |
// 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 */ |
2529 |
if(Data->qpel) { |
if(Data->qpel) { |
2530 |
pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; |
pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; |
2531 |
pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; |
pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; |
2540 |
} |
} |
2541 |
pMB->mvs[i] = *Data8->currentMV; |
pMB->mvs[i] = *Data8->currentMV; |
2542 |
pMB->sad8[i] = 4 * *Data8->iMinSAD; |
pMB->sad8[i] = 4 * *Data8->iMinSAD; |
2543 |
if (Data8->temp[0]) cbp |= 1 << (5 - i); |
if (Data8->cbp[0]) cbp |= 1 << (5 - i); |
|
} |
|
2544 |
|
|
2545 |
if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma |
} /* end - for all luma blocks */ |
2546 |
const uint8_t * ptr; |
|
2547 |
|
bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
2548 |
|
|
2549 |
|
/* let's check chroma */ |
2550 |
sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; |
2551 |
sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; |
2552 |
|
|
2553 |
//chroma U |
/* chroma U */ |
2554 |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2555 |
transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); |
2556 |
fdct(in); |
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); |
|
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); |
|
|
} |
|
2557 |
|
|
2558 |
if (bits < *Data->iMinSAD) { // still possible |
if (bits >= *Data->iMinSAD) return bits; |
2559 |
//chroma V |
|
2560 |
|
/* chroma V */ |
2561 |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); |
2562 |
transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); |
2563 |
fdct(in); |
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); |
|
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 += xvid_cbpy_tab[15-(cbp>>2)].len; |
|
|
bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
|
|
} |
|
|
} |
|
2564 |
|
|
2565 |
|
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; |
2566 |
|
|
2567 |
|
*Data->cbp = cbp; |
2568 |
return bits; |
return bits; |
2569 |
} |
} |
2570 |
|
|
|
|
|
2571 |
static int |
static int |
2572 |
CountMBBitsIntra(const SearchData * const Data) |
CountMBBitsIntra(const SearchData * const Data) |
2573 |
{ |
{ |
2574 |
int bits = 1; //this one is ac/dc prediction flag. always 1. |
int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */ |
2575 |
int cbp = 0, i, t, dc = 1024, b_dc; |
int cbp = 0, i, dc = 0; |
|
const uint32_t iQuant = Data->lambda16; |
|
2576 |
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
|
uint32_t iDcScaler = get_dc_scaler(iQuant, 1);; |
|
2577 |
|
|
2578 |
for(i = 0; i < 4; i++) { |
for(i = 0; i < 4; i++) { |
2579 |
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2580 |
transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); |
transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); |
2581 |
fdct(in); |
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc); |
2582 |
b_dc = in[0]; |
|
2583 |
in[0] -= dc; |
if (bits >= Data->iMinSAD[0]) return bits; |
2584 |
dc = b_dc; |
} |
2585 |
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
|
2586 |
else quant4_intra(coeff, in, iQuant, iDcScaler); |
bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len; |
2587 |
|
|
2588 |
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;; |
/*chroma U */ |
|
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 |
|
|
iDcScaler = get_dc_scaler(iQuant, 0); |
|
|
//chroma U |
|
2589 |
transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); |
2590 |
fdct(in); |
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc); |
|
in[0] -= 1024; |
|
|
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
|
|
else quant4_intra(coeff, in, iQuant, iDcScaler); |
|
2591 |
|
|
2592 |
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
if (bits >= Data->iMinSAD[0]) return bits; |
|
if (t != 0) cbp |= 1 << (5 - 4); |
|
2593 |
|
|
2594 |
if (bits < Data->iMinSAD[0]) { |
/* chroma V */ |
|
//chroma V |
|
2595 |
transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); |
2596 |
fdct(in); |
bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc); |
|
in[0] -= 1024; |
|
|
if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); |
|
|
else quant4_intra(coeff, in, iQuant, iDcScaler); |
|
2597 |
|
|
2598 |
bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; |
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
|
if (t != 0) cbp |= 1 << (5 - 5); |
|
2599 |
|
|
2600 |
bits += xvid_cbpy_tab[cbp>>2].len; |
return bits; |
|
bits += mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; |
|
2601 |
} |
} |
2602 |
|
|
2603 |
|
static int |
2604 |
|
CountMBBitsGMC(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y) |
2605 |
|
{ |
2606 |
|
int bits = BITS_MULT*1; /* this one is mcsel */ |
2607 |
|
int cbp = 0, i; |
2608 |
|
int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; |
2609 |
|
|
2610 |
|
for(i = 0; i < 4; i++) { |
2611 |
|
int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); |
2612 |
|
transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16*(x+y*Data->iEdgedWidth), Data->iEdgedWidth); |
2613 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i); |
2614 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2615 |
} |
} |
2616 |
|
|
2617 |
|
bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; |
2618 |
|
|
2619 |
|
/*chroma U */ |
2620 |
|
transfer_8to16subro(in, Data->CurU, vGMC->u + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); |
2621 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); |
2622 |
|
|
2623 |
|
if (bits >= Data->iMinSAD[0]) return bits; |
2624 |
|
|
2625 |
|
/* chroma V */ |
2626 |
|
transfer_8to16subro(in, Data->CurV , vGMC->v + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); |
2627 |
|
bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); |
2628 |
|
|
2629 |
|
bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; |
2630 |
|
|
2631 |
|
*Data->cbp = cbp; |
2632 |
|
|
2633 |
return bits; |
return bits; |
2634 |
} |
} |
2635 |
|
|
2636 |
|
|
2637 |
|
|
2638 |
|
|
2639 |
|
static __inline void |
2640 |
|
GMEanalyzeMB ( const uint8_t * const pCur, |
2641 |
|
const uint8_t * const pRef, |
2642 |
|
const uint8_t * const pRefH, |
2643 |
|
const uint8_t * const pRefV, |
2644 |
|
const uint8_t * const pRefHV, |
2645 |
|
const int x, |
2646 |
|
const int y, |
2647 |
|
const MBParam * const pParam, |
2648 |
|
MACROBLOCK * const pMBs, |
2649 |
|
SearchData * const Data) |
2650 |
|
{ |
2651 |
|
|
2652 |
|
int i=0; |
2653 |
|
MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; |
2654 |
|
|
2655 |
|
Data->iMinSAD[0] = MV_MAX_ERROR; |
2656 |
|
|
2657 |
|
Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); |
2658 |
|
|
2659 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2660 |
|
pParam->width, pParam->height, 16, 0, 0); |
2661 |
|
|
2662 |
|
Data->Cur = pCur + 16*(x + y * pParam->edged_width); |
2663 |
|
Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width); |
2664 |
|
Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width); |
2665 |
|
Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width); |
2666 |
|
Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width); |
2667 |
|
|
2668 |
|
Data->currentMV[0].x = Data->currentMV[0].y = 0; |
2669 |
|
CheckCandidate16I(0, 0, 255, &i, Data); |
2670 |
|
|
2671 |
|
if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) ) |
2672 |
|
CheckCandidate16I(Data->predMV.x, Data->predMV.y, 255, &i, Data); |
2673 |
|
|
2674 |
|
AdvDiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); |
2675 |
|
|
2676 |
|
SubpelRefine(Data); |
2677 |
|
|
2678 |
|
|
2679 |
|
/* for QPel halfpel positions are worse than in halfpel mode :( */ |
2680 |
|
/* if (Data->qpel) { |
2681 |
|
Data->currentQMV->x = 2*Data->currentMV->x; |
2682 |
|
Data->currentQMV->y = 2*Data->currentMV->y; |
2683 |
|
Data->qpel_precision = 1; |
2684 |
|
get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, |
2685 |
|
pParam->width, pParam->height, iFcode, 1, 0); |
2686 |
|
SubpelRefine(Data); |
2687 |
|
} |
2688 |
|
*/ |
2689 |
|
|
2690 |
|
pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; |
2691 |
|
pMB->sad16 = Data->iMinSAD[0]; |
2692 |
|
pMB->mode = MODE_INTER; |
2693 |
|
pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0); |
2694 |
|
return; |
2695 |
|
} |
2696 |
|
|
2697 |
|
void |
2698 |
|
GMEanalysis(const MBParam * const pParam, |
2699 |
|
const FRAMEINFO * const current, |
2700 |
|
const FRAMEINFO * const reference, |
2701 |
|
const IMAGE * const pRefH, |
2702 |
|
const IMAGE * const pRefV, |
2703 |
|
const IMAGE * const pRefHV) |
2704 |
|
{ |
2705 |
|
uint32_t x, y; |
2706 |
|
MACROBLOCK * const pMBs = current->mbs; |
2707 |
|
const IMAGE * const pCurrent = ¤t->image; |
2708 |
|
const IMAGE * const pReference = &reference->image; |
2709 |
|
|
2710 |
|
int32_t iMinSAD[5], temp[5]; |
2711 |
|
VECTOR currentMV[5]; |
2712 |
|
SearchData Data; |
2713 |
|
memset(&Data, 0, sizeof(SearchData)); |
2714 |
|
|
2715 |
|
Data.iEdgedWidth = pParam->edged_width; |
2716 |
|
Data.rounding = pParam->m_rounding_type; |
2717 |
|
|
2718 |
|
Data.currentMV = ¤tMV[0]; |
2719 |
|
Data.iMinSAD = &iMinSAD[0]; |
2720 |
|
Data.iFcode = current->fcode; |
2721 |
|
Data.temp = temp; |
2722 |
|
|
2723 |
|
CheckCandidate = CheckCandidate16I; |
2724 |
|
|
2725 |
|
if (sadInit) (*sadInit) (); |
2726 |
|
|
2727 |
|
for (y = 0; y < pParam->mb_height; y ++) { |
2728 |
|
for (x = 0; x < pParam->mb_width; x ++) { |
2729 |
|
|
2730 |
|
GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data); |
2731 |
|
} |
2732 |
|
} |
2733 |
|
return; |
2734 |
|
} |
2735 |
|
|
2736 |
|
|
2737 |
|
WARPPOINTS |
2738 |
|
GlobalMotionEst(MACROBLOCK * const pMBs, |
2739 |
|
const MBParam * const pParam, |
2740 |
|
const FRAMEINFO * const current, |
2741 |
|
const FRAMEINFO * const reference, |
2742 |
|
const IMAGE * const pRefH, |
2743 |
|
const IMAGE * const pRefV, |
2744 |
|
const IMAGE * const pRefHV) |
2745 |
|
{ |
2746 |
|
|
2747 |
|
const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs |
2748 |
|
const int deltay=8; |
2749 |
|
const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks) |
2750 |
|
const unsigned int grady=512; |
2751 |
|
|
2752 |
|
double sol[4] = { 0., 0., 0., 0. }; |
2753 |
|
|
2754 |
|
WARPPOINTS gmc; |
2755 |
|
|
2756 |
|
uint32_t mx, my; |
2757 |
|
|
2758 |
|
int MBh = pParam->mb_height; |
2759 |
|
int MBw = pParam->mb_width; |
2760 |
|
const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */ |
2761 |
|
const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */ |
2762 |
|
|
2763 |
|
int num=0; |
2764 |
|
int oldnum; |
2765 |
|
|
2766 |
|
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; |
2767 |
|
|
2768 |
|
GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV); |
2769 |
|
|
2770 |
|
/* block based ME isn't done, yet, so do a quick presearch */ |
2771 |
|
|
2772 |
|
// filter mask of all blocks |
2773 |
|
|
2774 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2775 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2776 |
|
{ |
2777 |
|
const int mbnum = mx + my * MBw; |
2778 |
|
pMBs[mbnum].mcsel = 0; |
2779 |
|
} |
2780 |
|
|
2781 |
|
|
2782 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ |
2783 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
2784 |
|
{ |
2785 |
|
const int mbnum = mx + my * MBw; |
2786 |
|
MACROBLOCK *const pMB = &pMBs[mbnum]; |
2787 |
|
const VECTOR mv = pMB->mvs[0]; |
2788 |
|
|
2789 |
|
/* don't use object boundaries */ |
2790 |
|
if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) |
2791 |
|
&& (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) |
2792 |
|
&& (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) |
2793 |
|
&& (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) |
2794 |
|
&& (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) |
2795 |
|
&& (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) |
2796 |
|
&& (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) |
2797 |
|
&& (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) |
2798 |
|
{ const int iEdgedWidth = pParam->edged_width; |
2799 |
|
const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); |
2800 |
|
if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) |
2801 |
|
&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) |
2802 |
|
{ pMB->mcsel = 1; |
2803 |
|
num++; |
2804 |
|
} |
2805 |
|
|
2806 |
|
/* only use "structured" blocks */ |
2807 |
|
} |
2808 |
|
} |
2809 |
|
emms(); |
2810 |
|
|
2811 |
|
/* further filtering would be possible, but during iteration, remaining |
2812 |
|
outliers usually are removed, too */ |
2813 |
|
|
2814 |
|
if (num>= minblocks) |
2815 |
|
do { /* until convergence */ |
2816 |
|
double DtimesF[4]; |
2817 |
|
double a,b,c,n,invdenom; |
2818 |
|
double meanx,meany; |
2819 |
|
|
2820 |
|
a = b = c = n = 0; |
2821 |
|
DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; |
2822 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2823 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2824 |
|
{ |
2825 |
|
const int mbnum = mx + my * MBw; |
2826 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
2827 |
|
|
2828 |
|
if (!pMBs[mbnum].mcsel) |
2829 |
|
continue; |
2830 |
|
|
2831 |
|
n++; |
2832 |
|
a += 16*mx+8; |
2833 |
|
b += 16*my+8; |
2834 |
|
c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); |
2835 |
|
|
2836 |
|
DtimesF[0] += (double)mv.x; |
2837 |
|
DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); |
2838 |
|
DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); |
2839 |
|
DtimesF[3] += (double)mv.y; |
2840 |
|
} |
2841 |
|
|
2842 |
|
invdenom = a*a+b*b-c*n; |
2843 |
|
|
2844 |
|
/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ |
2845 |
|
/* D'*E*F has been calculated in the same loop as matrix */ |
2846 |
|
|
2847 |
|
sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; |
2848 |
|
sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; |
2849 |
|
sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; |
2850 |
|
sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; |
2851 |
|
|
2852 |
|
sol[0] /= invdenom; |
2853 |
|
sol[1] /= invdenom; |
2854 |
|
sol[2] /= invdenom; |
2855 |
|
sol[3] /= invdenom; |
2856 |
|
|
2857 |
|
meanx = meany = 0.; |
2858 |
|
oldnum = 0; |
2859 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2860 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2861 |
|
{ |
2862 |
|
const int mbnum = mx + my * MBw; |
2863 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
2864 |
|
|
2865 |
|
if (!pMBs[mbnum].mcsel) |
2866 |
|
continue; |
2867 |
|
|
2868 |
|
oldnum++; |
2869 |
|
meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ); |
2870 |
|
meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ); |
2871 |
|
} |
2872 |
|
|
2873 |
|
if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */ |
2874 |
|
meanx /= oldnum; |
2875 |
|
else |
2876 |
|
meanx = 0.25; |
2877 |
|
|
2878 |
|
if (4*meany > oldnum) |
2879 |
|
meany /= oldnum; |
2880 |
|
else |
2881 |
|
meany = 0.25; |
2882 |
|
|
2883 |
|
num = 0; |
2884 |
|
for (my = 0; my < (uint32_t)MBh; my++) |
2885 |
|
for (mx = 0; mx < (uint32_t)MBw; mx++) |
2886 |
|
{ |
2887 |
|
const int mbnum = mx + my * MBw; |
2888 |
|
const VECTOR mv = pMBs[mbnum].mvs[0]; |
2889 |
|
|
2890 |
|
if (!pMBs[mbnum].mcsel) |
2891 |
|
continue; |
2892 |
|
|
2893 |
|
if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx ) |
2894 |
|
|| ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) ) |
2895 |
|
pMBs[mbnum].mcsel=0; |
2896 |
|
else |
2897 |
|
num++; |
2898 |
|
} |
2899 |
|
|
2900 |
|
} while ( (oldnum != num) && (num>= minblocks) ); |
2901 |
|
|
2902 |
|
if (num < minblocks) |
2903 |
|
{ |
2904 |
|
const int iEdgedWidth = pParam->edged_width; |
2905 |
|
num = 0; |
2906 |
|
|
2907 |
|
/* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw); |
2908 |
|
*/ |
2909 |
|
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; |
2910 |
|
|
2911 |
|
if (!(current->motion_flags & XVID_GME_REFINE)) |
2912 |
|
return gmc; |
2913 |
|
|
2914 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ |
2915 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ |
2916 |
|
{ |
2917 |
|
const int mbnum = mx + my * MBw; |
2918 |
|
MACROBLOCK *const pMB = &pMBs[mbnum]; |
2919 |
|
const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); |
2920 |
|
if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) |
2921 |
|
&& (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) |
2922 |
|
{ pMB->mcsel = 1; |
2923 |
|
gmc.duv[0].x += pMB->mvs[0].x; |
2924 |
|
gmc.duv[0].y += pMB->mvs[0].y; |
2925 |
|
num++; |
2926 |
|
} |
2927 |
|
} |
2928 |
|
|
2929 |
|
if (gmc.duv[0].x) |
2930 |
|
gmc.duv[0].x /= num; |
2931 |
|
if (gmc.duv[0].y) |
2932 |
|
gmc.duv[0].y /= num; |
2933 |
|
} else { |
2934 |
|
|
2935 |
|
gmc.duv[0].x=(int)(sol[0]+0.5); |
2936 |
|
gmc.duv[0].y=(int)(sol[3]+0.5); |
2937 |
|
|
2938 |
|
gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); |
2939 |
|
gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); |
2940 |
|
|
2941 |
|
gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */ |
2942 |
|
gmc.duv[2].y=gmc.duv[1].x; |
2943 |
|
} |
2944 |
|
if (num>maxblocks) |
2945 |
|
{ for (my = 1; my < (uint32_t)MBh-1; my++) |
2946 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2947 |
|
{ |
2948 |
|
const int mbnum = mx + my * MBw; |
2949 |
|
if (pMBs[mbnum-1].mcsel) |
2950 |
|
pMBs[mbnum].mcsel=0; |
2951 |
|
else |
2952 |
|
if (pMBs[mbnum-MBw].mcsel) |
2953 |
|
pMBs[mbnum].mcsel=0; |
2954 |
|
} |
2955 |
|
} |
2956 |
|
return gmc; |
2957 |
|
} |
2958 |
|
|
2959 |
|
int |
2960 |
|
GlobalMotionEstRefine( |
2961 |
|
WARPPOINTS *const startwp, |
2962 |
|
MACROBLOCK * const pMBs, |
2963 |
|
const MBParam * const pParam, |
2964 |
|
const FRAMEINFO * const current, |
2965 |
|
const FRAMEINFO * const reference, |
2966 |
|
const IMAGE * const pCurr, |
2967 |
|
const IMAGE * const pRef, |
2968 |
|
const IMAGE * const pRefH, |
2969 |
|
const IMAGE * const pRefV, |
2970 |
|
const IMAGE * const pRefHV) |
2971 |
|
{ |
2972 |
|
uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width); |
2973 |
|
WARPPOINTS bestwp=*startwp; |
2974 |
|
WARPPOINTS centerwp,currwp; |
2975 |
|
int gmcminSAD=0; |
2976 |
|
int gmcSAD=0; |
2977 |
|
int direction; |
2978 |
|
// int mx,my; |
2979 |
|
|
2980 |
|
/* use many blocks... */ |
2981 |
|
/* for (my = 0; my < (uint32_t)pParam->mb_height; my++) |
2982 |
|
for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) |
2983 |
|
{ |
2984 |
|
const int mbnum = mx + my * pParam->mb_width; |
2985 |
|
pMBs[mbnum].mcsel=1; |
2986 |
|
} |
2987 |
|
*/ |
2988 |
|
|
2989 |
|
/* or rather don't use too many blocks... */ |
2990 |
|
/* |
2991 |
|
for (my = 1; my < (uint32_t)MBh-1; my++) |
2992 |
|
for (mx = 1; mx < (uint32_t)MBw-1; mx++) |
2993 |
|
{ |
2994 |
|
const int mbnum = mx + my * MBw; |
2995 |
|
if (MBmask[mbnum-1]) |
2996 |
|
MBmask[mbnum-1]=0; |
2997 |
|
else |
2998 |
|
if (MBmask[mbnum-MBw]) |
2999 |
|
MBmask[mbnum-1]=0; |
3000 |
|
|
3001 |
|
} |
3002 |
|
*/ |
3003 |
|
gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3004 |
|
|
3005 |
|
if ( (reference->coding_type == S_VOP) |
3006 |
|
&& ( (reference->warp.duv[1].x != bestwp.duv[1].x) |
3007 |
|
|| (reference->warp.duv[1].y != bestwp.duv[1].y) |
3008 |
|
|| (reference->warp.duv[0].x != bestwp.duv[0].x) |
3009 |
|
|| (reference->warp.duv[0].y != bestwp.duv[0].y) |
3010 |
|
|| (reference->warp.duv[2].x != bestwp.duv[2].x) |
3011 |
|
|| (reference->warp.duv[2].y != bestwp.duv[2].y) ) ) |
3012 |
|
{ |
3013 |
|
gmcSAD = globalSAD(&reference->warp, pParam, pMBs, |
3014 |
|
current, pRef, pCurr, GMCblock); |
3015 |
|
|
3016 |
|
if (gmcSAD < gmcminSAD) |
3017 |
|
{ bestwp = reference->warp; |
3018 |
|
gmcminSAD = gmcSAD; |
3019 |
|
} |
3020 |
|
} |
3021 |
|
|
3022 |
|
do { |
3023 |
|
direction = 0; |
3024 |
|
centerwp = bestwp; |
3025 |
|
|
3026 |
|
currwp = centerwp; |
3027 |
|
|
3028 |
|
currwp.duv[0].x--; |
3029 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3030 |
|
if (gmcSAD < gmcminSAD) |
3031 |
|
{ bestwp = currwp; |
3032 |
|
gmcminSAD = gmcSAD; |
3033 |
|
direction = 1; |
3034 |
|
} |
3035 |
|
else |
3036 |
|
{ |
3037 |
|
currwp = centerwp; currwp.duv[0].x++; |
3038 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3039 |
|
if (gmcSAD < gmcminSAD) |
3040 |
|
{ bestwp = currwp; |
3041 |
|
gmcminSAD = gmcSAD; |
3042 |
|
direction = 2; |
3043 |
|
} |
3044 |
|
} |
3045 |
|
if (direction) continue; |
3046 |
|
|
3047 |
|
currwp = centerwp; currwp.duv[0].y--; |
3048 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3049 |
|
if (gmcSAD < gmcminSAD) |
3050 |
|
{ bestwp = currwp; |
3051 |
|
gmcminSAD = gmcSAD; |
3052 |
|
direction = 4; |
3053 |
|
} |
3054 |
|
else |
3055 |
|
{ |
3056 |
|
currwp = centerwp; currwp.duv[0].y++; |
3057 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3058 |
|
if (gmcSAD < gmcminSAD) |
3059 |
|
{ bestwp = currwp; |
3060 |
|
gmcminSAD = gmcSAD; |
3061 |
|
direction = 8; |
3062 |
|
} |
3063 |
|
} |
3064 |
|
if (direction) continue; |
3065 |
|
|
3066 |
|
currwp = centerwp; currwp.duv[1].x++; |
3067 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3068 |
|
if (gmcSAD < gmcminSAD) |
3069 |
|
{ bestwp = currwp; |
3070 |
|
gmcminSAD = gmcSAD; |
3071 |
|
direction = 32; |
3072 |
|
} |
3073 |
|
currwp.duv[2].y++; |
3074 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3075 |
|
if (gmcSAD < gmcminSAD) |
3076 |
|
{ bestwp = currwp; |
3077 |
|
gmcminSAD = gmcSAD; |
3078 |
|
direction = 1024; |
3079 |
|
} |
3080 |
|
|
3081 |
|
currwp = centerwp; currwp.duv[1].x--; |
3082 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3083 |
|
if (gmcSAD < gmcminSAD) |
3084 |
|
{ bestwp = currwp; |
3085 |
|
gmcminSAD = gmcSAD; |
3086 |
|
direction = 16; |
3087 |
|
} |
3088 |
|
else |
3089 |
|
{ |
3090 |
|
currwp = centerwp; currwp.duv[1].x++; |
3091 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3092 |
|
if (gmcSAD < gmcminSAD) |
3093 |
|
{ bestwp = currwp; |
3094 |
|
gmcminSAD = gmcSAD; |
3095 |
|
direction = 32; |
3096 |
|
} |
3097 |
|
} |
3098 |
|
if (direction) continue; |
3099 |
|
|
3100 |
|
|
3101 |
|
currwp = centerwp; currwp.duv[1].y--; |
3102 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3103 |
|
if (gmcSAD < gmcminSAD) |
3104 |
|
{ bestwp = currwp; |
3105 |
|
gmcminSAD = gmcSAD; |
3106 |
|
direction = 64; |
3107 |
|
} |
3108 |
|
else |
3109 |
|
{ |
3110 |
|
currwp = centerwp; currwp.duv[1].y++; |
3111 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3112 |
|
if (gmcSAD < gmcminSAD) |
3113 |
|
{ bestwp = currwp; |
3114 |
|
gmcminSAD = gmcSAD; |
3115 |
|
direction = 128; |
3116 |
|
} |
3117 |
|
} |
3118 |
|
if (direction) continue; |
3119 |
|
|
3120 |
|
currwp = centerwp; currwp.duv[2].x--; |
3121 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3122 |
|
if (gmcSAD < gmcminSAD) |
3123 |
|
{ bestwp = currwp; |
3124 |
|
gmcminSAD = gmcSAD; |
3125 |
|
direction = 256; |
3126 |
|
} |
3127 |
|
else |
3128 |
|
{ |
3129 |
|
currwp = centerwp; currwp.duv[2].x++; |
3130 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3131 |
|
if (gmcSAD < gmcminSAD) |
3132 |
|
{ bestwp = currwp; |
3133 |
|
gmcminSAD = gmcSAD; |
3134 |
|
direction = 512; |
3135 |
|
} |
3136 |
|
} |
3137 |
|
if (direction) continue; |
3138 |
|
|
3139 |
|
currwp = centerwp; currwp.duv[2].y--; |
3140 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3141 |
|
if (gmcSAD < gmcminSAD) |
3142 |
|
{ bestwp = currwp; |
3143 |
|
gmcminSAD = gmcSAD; |
3144 |
|
direction = 1024; |
3145 |
|
} |
3146 |
|
else |
3147 |
|
{ |
3148 |
|
currwp = centerwp; currwp.duv[2].y++; |
3149 |
|
gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); |
3150 |
|
if (gmcSAD < gmcminSAD) |
3151 |
|
{ bestwp = currwp; |
3152 |
|
gmcminSAD = gmcSAD; |
3153 |
|
direction = 2048; |
3154 |
|
} |
3155 |
|
} |
3156 |
|
} while (direction); |
3157 |
|
free(GMCblock); |
3158 |
|
|
3159 |
|
*startwp = bestwp; |
3160 |
|
|
3161 |
|
return gmcminSAD; |
3162 |
|
} |
3163 |
|
|
3164 |
|
int |
3165 |
|
globalSAD(const WARPPOINTS *const wp, |
3166 |
|
const MBParam * const pParam, |
3167 |
|
const MACROBLOCK * const pMBs, |
3168 |
|
const FRAMEINFO * const current, |
3169 |
|
const IMAGE * const pRef, |
3170 |
|
const IMAGE * const pCurr, |
3171 |
|
uint8_t *const GMCblock) |
3172 |
|
{ |
3173 |
|
NEW_GMC_DATA gmc_data; |
3174 |
|
int iSAD, gmcSAD=0; |
3175 |
|
int num=0; |
3176 |
|
unsigned int mx, my; |
3177 |
|
|
3178 |
|
generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data); |
3179 |
|
|
3180 |
|
for (my = 0; my < (uint32_t)pParam->mb_height; my++) |
3181 |
|
for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) { |
3182 |
|
|
3183 |
|
const int mbnum = mx + my * pParam->mb_width; |
3184 |
|
const int iEdgedWidth = pParam->edged_width; |
3185 |
|
|
3186 |
|
if (!pMBs[mbnum].mcsel) |
3187 |
|
continue; |
3188 |
|
|
3189 |
|
gmc_data.predict_16x16(&gmc_data, GMCblock, |
3190 |
|
pRef->y, |
3191 |
|
iEdgedWidth, |
3192 |
|
iEdgedWidth, |
3193 |
|
mx, my, |
3194 |
|
pParam->m_rounding_type); |
3195 |
|
|
3196 |
|
iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx), |
3197 |
|
GMCblock , iEdgedWidth, 65536); |
3198 |
|
iSAD -= pMBs[mbnum].sad16; |
3199 |
|
|
3200 |
|
if (iSAD<0) |
3201 |
|
gmcSAD += iSAD; |
3202 |
|
num++; |
3203 |
|
} |
3204 |
|
return gmcSAD; |
3205 |
|
} |
3206 |
|
|