[svn] / branches / dev-api-3 / xvidcore / src / motion / motion_est.c Repository:
ViewVC logotype

Annotation of /branches/dev-api-3/xvidcore/src/motion/motion_est.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 753 - (view) (download)

1 : Isibaar 3 /**************************************************************************
2 :     *
3 : chl 259 * XVID MPEG-4 VIDEO CODEC
4 : syskin 748 * motion estimation
5 : chl 259 *
6 :     * This program is an implementation of a part of one or more MPEG-4
7 :     * Video tools as specified in ISO/IEC 14496-2 standard. Those intending
8 :     * to use this software module in hardware or software products are
9 :     * advised that its use may infringe existing patents or copyrights, and
10 :     * any such use would be at such party's own risk. The original
11 :     * developer of this software module and his/her company, and subsequent
12 :     * editors and their companies, will have no liability for use of this
13 :     * software or modifications or derivatives thereof.
14 :     *
15 :     * This program is free software; you can redistribute it and/or modify
16 :     * it under the terms of the GNU General Public License as published by
17 :     * the Free Software Foundation; either version 2 of the License, or
18 :     * (at your option) any later version.
19 :     *
20 :     * This program is distributed in the hope that it will be useful,
21 :     * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 :     * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 :     * GNU General Public License for more details.
24 :     *
25 :     * You should have received a copy of the GNU General Public License
26 :     * along with this program; if not, write to the Free Software
27 :     * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 :     *
29 :     *************************************************************************/
30 :    
31 : Isibaar 3 #include <assert.h>
32 :     #include <stdio.h>
33 : chl 96 #include <stdlib.h>
34 : suxen_drol 721 #include <string.h> // memcpy
35 : Isibaar 3
36 :     #include "../encoder.h"
37 :     #include "../utils/mbfunctions.h"
38 :     #include "../prediction/mbprediction.h"
39 :     #include "../global.h"
40 :     #include "../utils/timer.h"
41 : Isibaar 580 #include "../image/interpolate8x8.h"
42 : chl 530 #include "motion_est.h"
43 : suxen_drol 118 #include "motion.h"
44 : Isibaar 3 #include "sad.h"
45 : Isibaar 539 #include "../utils/emms.h"
46 : Isibaar 3
47 : chl 530 #define INITIAL_SKIP_THRESH (10)
48 :     #define FINAL_SKIP_THRESH (50)
49 :     #define MAX_SAD00_FOR_SKIP (20)
50 : Isibaar 539 #define MAX_CHROMA_SAD_FOR_SKIP (22)
51 : Isibaar 3
52 : chl 530 #define CHECK_CANDIDATE(X,Y,D) { \
53 :     (*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); }
54 : Isibaar 3
55 : chl 530 static __inline int
56 : syskin 724 d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
57 : Isibaar 3 {
58 : chl 530 int xb, yb;
59 : syskin 712 if (qpel) { x *= 2; y *= 2;}
60 :     else if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
61 : syskin 748 x -= pred.x;
62 :     y -= pred.y;
63 : chl 530
64 : syskin 748 if (x) {
65 : chl 530 if (x < 0) x = -x;
66 :     x += (1 << (iFcode - 1)) - 1;
67 :     x >>= (iFcode - 1);
68 :     if (x > 32) x = 32;
69 :     xb = mvtab[x] + iFcode;
70 : syskin 748 } else xb = 1;
71 : Isibaar 3
72 : syskin 748 if (y) {
73 : chl 530 if (y < 0) y = -y;
74 :     y += (1 << (iFcode - 1)) - 1;
75 :     y >>= (iFcode - 1);
76 :     if (y > 32) y = 32;
77 :     yb = mvtab[y] + iFcode;
78 : syskin 748 } else yb = 1;
79 : chl 530 return xb + yb;
80 : Isibaar 3 }
81 :    
82 : syskin 748 static int32_t
83 : syskin 628 ChromaSAD(int dx, int dy, const SearchData * const data)
84 : Isibaar 3 {
85 : syskin 628 int sad;
86 : syskin 748 const uint32_t stride = data->iEdgedWidth/2;
87 : Isibaar 3
88 : syskin 712 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently
89 : syskin 748 data->temp[5] = dx; data->temp[6] = dy; // backup
90 : syskin 712
91 :     switch (((dx & 1) << 1) | (dy & 1)) {
92 : syskin 628 case 0:
93 : syskin 748 dx = dx / 2; dy = dy / 2;
94 :     sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride);
95 :     sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride);
96 : syskin 628 break;
97 :     case 1:
98 :     dx = dx / 2; dy = (dy - 1) / 2;
99 : syskin 748 sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride);
100 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride);
101 : syskin 628 break;
102 :     case 2:
103 :     dx = (dx - 1) / 2; dy = dy / 2;
104 : syskin 748 sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride);
105 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride);
106 : syskin 628 break;
107 :     default:
108 :     dx = (dx - 1) / 2; dy = (dy - 1) / 2;
109 : syskin 748 interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding);
110 :     sad = sad8(data->CurU, data->RefQ, stride);
111 :    
112 :     interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding);
113 :     sad += sad8(data->CurV, data->RefQ, stride);
114 : syskin 628 break;
115 : chl 530 }
116 : syskin 748 data->temp[7] = sad; //backup, part 2
117 : syskin 628 return sad;
118 :     }
119 : Isibaar 3
120 : syskin 663 static __inline const uint8_t *
121 : syskin 748 GetReferenceB(const int x, const int y, const int dir, const SearchData * const data)
122 : Isibaar 606 {
123 : syskin 663 // dir : 0 = forward, 1 = backward
124 :     switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) {
125 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
126 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
127 :     case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
128 :     case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
129 :     case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth);
130 :     case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
131 :     case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
132 :     default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
133 : Isibaar 606 }
134 :     }
135 :    
136 : syskin 748 // this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate
137 :     static __inline const uint8_t *
138 :     GetReference(const int x, const int y, const SearchData * const data)
139 :     {
140 :     switch ( ((x&1)<<1) | (y&1) ) {
141 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
142 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
143 :     case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
144 :     default : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
145 :     }
146 :     }
147 :    
148 :     static uint8_t *
149 : syskin 663 Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data)
150 : chl 530 {
151 : syskin 663 // create or find a qpel-precision reference picture; return pointer to it
152 : syskin 668 uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir;
153 : syskin 663 const int32_t iEdgedWidth = data->iEdgedWidth;
154 :     const uint32_t rounding = data->rounding;
155 :     const int halfpel_x = x/2;
156 :     const int halfpel_y = y/2;
157 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
158 : chl 326
159 : syskin 748 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
160 : syskin 663 ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
161 :     switch( ((x&1)<<1) + (y&1) ) {
162 :     case 0: // pure halfpel position
163 : syskin 740 return (uint8_t *) ref1;
164 : syskin 663 break;
165 : Isibaar 3
166 : syskin 663 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
167 : syskin 748 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
168 : syskin 663 ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
169 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
170 : syskin 663 break;
171 : Isibaar 3
172 : syskin 663 case 2: // x qpel, y halfpel - left or right during qpel refinement
173 : syskin 748 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
174 : syskin 663 ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
175 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
176 : syskin 663 break;
177 : Isibaar 3
178 : syskin 663 default: // x and y in qpel resolution - the "corners" (top left/right and
179 :     // bottom left/right) during qpel refinement
180 : syskin 748 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
181 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
182 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
183 : syskin 663 ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
184 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
185 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
186 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
187 :     break;
188 :     }
189 :     return Reference;
190 : Isibaar 3 }
191 :    
192 : syskin 748 static uint8_t *
193 : syskin 663 Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data)
194 : Isibaar 580 {
195 : syskin 663 // create or find a qpel-precision reference picture; return pointer to it
196 : syskin 668 uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir;
197 : syskin 663 const int32_t iEdgedWidth = data->iEdgedWidth;
198 :     const uint32_t rounding = data->rounding;
199 :     const int halfpel_x = x/2;
200 :     const int halfpel_y = y/2;
201 : Isibaar 580 const uint8_t *ref1, *ref2, *ref3, *ref4;
202 :    
203 : syskin 748 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
204 : syskin 663 switch( ((x&1)<<1) + (y&1) ) {
205 :     case 0: // pure halfpel position
206 : syskin 740 return (uint8_t *) ref1;
207 : Isibaar 580 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
208 : syskin 748 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
209 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
210 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
211 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
212 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
213 : Isibaar 580 break;
214 :    
215 :     case 2: // x qpel, y halfpel - left or right during qpel refinement
216 : syskin 748 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
217 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
218 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
219 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
220 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
221 : Isibaar 580 break;
222 :    
223 :     default: // x and y in qpel resolution - the "corners" (top left/right and
224 :     // bottom left/right) during qpel refinement
225 : syskin 748 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
226 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
227 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
228 : Isibaar 580 interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
229 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
230 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
231 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
232 :     break;
233 :     }
234 : syskin 663 return Reference;
235 :     }
236 :    
237 :     /* CHECK_CANDIATE FUNCTIONS START */
238 :    
239 : syskin 748 static void
240 : syskin 663 CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
241 :     {
242 :     int t, xc, yc;
243 :     const uint8_t * Reference;
244 :     VECTOR * current;
245 :    
246 :     if (( x > data->max_dx) || ( x < data->min_dx)
247 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
248 :    
249 :     if (data->qpel_precision) { // x and y are in 1/4 precision
250 :     Reference = Interpolate16x16qpel(x, y, 0, data);
251 :     xc = x/2; yc = y/2; //for chroma sad
252 :     current = data->currentQMV;
253 :     } else {
254 : syskin 748 Reference = GetReference(x, y, data);
255 : syskin 663 current = data->currentMV;
256 :     xc = x; yc = y;
257 :     }
258 : syskin 724 t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0);
259 : Isibaar 580
260 : syskin 663 data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
261 : Isibaar 580
262 : syskin 628 data->temp[0] += (data->lambda16 * t * data->temp[0])/1000;
263 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100;
264 : Isibaar 580
265 : syskin 753 if (data->chroma) data->temp[0] += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
266 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
267 : syskin 628
268 : Isibaar 580 if (data->temp[0] < data->iMinSAD[0]) {
269 :     data->iMinSAD[0] = data->temp[0];
270 : syskin 663 current[0].x = x; current[0].y = y;
271 :     *dir = Direction; }
272 : Isibaar 580
273 :     if (data->temp[1] < data->iMinSAD[1]) {
274 : syskin 663 data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; }
275 : Isibaar 580 if (data->temp[2] < data->iMinSAD[2]) {
276 : syskin 663 data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
277 : Isibaar 580 if (data->temp[3] < data->iMinSAD[3]) {
278 : syskin 663 data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
279 : Isibaar 580 if (data->temp[4] < data->iMinSAD[4]) {
280 : syskin 663 data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
281 :    
282 : Isibaar 580 }
283 :    
284 : syskin 748 static void
285 : syskin 704 CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
286 :     {
287 :     int t;
288 :     const uint8_t * Reference;
289 :    
290 :     if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value
291 :     ( x > data->max_dx) || ( x < data->min_dx)
292 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
293 :    
294 : syskin 748 Reference = GetReference(x, y, data);
295 : syskin 724 t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
296 : syskin 704
297 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
298 :    
299 :     data->temp[0] += (data->lambda16 * t * data->temp[0])/1000;
300 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100;
301 :    
302 :     if (data->temp[0] < data->iMinSAD[0]) {
303 :     data->iMinSAD[0] = data->temp[0];
304 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
305 :     *dir = Direction; }
306 :    
307 :     if (data->temp[1] < data->iMinSAD[1]) {
308 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
309 :     if (data->temp[2] < data->iMinSAD[2]) {
310 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
311 :     if (data->temp[3] < data->iMinSAD[3]) {
312 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
313 :     if (data->temp[4] < data->iMinSAD[4]) {
314 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
315 :     }
316 :    
317 : syskin 748 static void
318 : syskin 663 CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
319 : syskin 658 {
320 :     int32_t sad;
321 : syskin 663 const uint8_t * Reference;
322 :     int t;
323 :     VECTOR * current;
324 : syskin 658
325 :     if (( x > data->max_dx) || ( x < data->min_dx)
326 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
327 :    
328 : syskin 712 if (data->rrv)
329 : syskin 704 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) ) return; //non-zero integer value
330 :    
331 : syskin 663 if (data->qpel_precision) { // x and y are in 1/4 precision
332 :     Reference = Interpolate16x16qpel(x, y, 0, data);
333 :     current = data->currentQMV;
334 :     } else {
335 : syskin 748 Reference = GetReference(x, y, data);
336 : syskin 663 current = data->currentMV;
337 : syskin 658 }
338 : syskin 724 t = d_mv_bits(x, y, data->predMV, data->iFcode,
339 : syskin 748 data->qpel && !data->qpel_precision, data->rrv);
340 : syskin 658
341 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
342 : syskin 663 sad += (data->lambda16 * t * sad)/1000;
343 : syskin 658
344 : syskin 663 if (sad < *(data->iMinSAD)) {
345 :     *(data->iMinSAD) = sad;
346 :     current->x = x; current->y = y;
347 :     *dir = Direction; }
348 : syskin 658 }
349 :    
350 : syskin 748 static void
351 : syskin 739 CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
352 : Isibaar 539 {
353 : syskin 658 // maximum speed - for P/B/I decision
354 : Isibaar 539
355 :     if (( x > data->max_dx) || ( x < data->min_dx)
356 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
357 :    
358 : syskin 739 data->temp[0] = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
359 :     data->iEdgedWidth, data->temp+1);
360 :     if (data->temp[0] < *(data->iMinSAD)) {
361 :     *(data->iMinSAD) = data->temp[0];
362 : Isibaar 539 data->currentMV[0].x = x; data->currentMV[0].y = y;
363 :     *dir = Direction; }
364 : syskin 739 if (data->temp[1] < data->iMinSAD[1]) {
365 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
366 :     if (data->temp[2] < data->iMinSAD[2]) {
367 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
368 :     if (data->temp[3] < data->iMinSAD[3]) {
369 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
370 :     if (data->temp[4] < data->iMinSAD[4]) {
371 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
372 :    
373 : Isibaar 539 }
374 :    
375 : syskin 748 static void
376 : chl 530 CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
377 :     {
378 :     int32_t sad;
379 : syskin 663 int xb, yb, t;
380 : chl 530 const uint8_t *ReferenceF, *ReferenceB;
381 : syskin 663 VECTOR *current;
382 : Isibaar 3
383 : chl 530 if (( xf > data->max_dx) || ( xf < data->min_dx)
384 :     || ( yf > data->max_dy) || (yf < data->min_dy)) return;
385 : Isibaar 3
386 : syskin 663 if (data->qpel_precision) {
387 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
388 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
389 :     current = data->currentQMV;
390 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
391 :     } else {
392 : syskin 748 ReferenceF = GetReference(xf, yf, data);
393 : syskin 663 xb = data->currentMV[1].x; yb = data->currentMV[1].y;
394 : syskin 748 ReferenceB = GetReferenceB(xb, yb, 1, data);
395 : syskin 663 current = data->currentMV;
396 : chl 530 }
397 : edgomez 195
398 : syskin 724 t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0)
399 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel && !data->qpel_precision, 0);
400 : syskin 712
401 : syskin 628 sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
402 : syskin 663 sad += (data->lambda16 * t * sad)/1000;
403 : syskin 628
404 : chl 530 if (sad < *(data->iMinSAD)) {
405 :     *(data->iMinSAD) = sad;
406 : syskin 663 current->x = xf; current->y = yf;
407 : chl 530 *dir = Direction; }
408 : Isibaar 3 }
409 :    
410 : syskin 748 static void
411 : chl 530 CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
412 : Isibaar 3 {
413 : syskin 628 int32_t sad = 0;
414 : chl 530 int k;
415 :     const uint8_t *ReferenceF;
416 :     const uint8_t *ReferenceB;
417 :     VECTOR mvs, b_mvs;
418 : syskin 724 const VECTOR zeroMV={0,0};
419 : Isibaar 3
420 : chl 530 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
421 : Isibaar 3
422 : chl 530 for (k = 0; k < 4; k++) {
423 :     mvs.x = data->directmvF[k].x + x;
424 :     b_mvs.x = ((x == 0) ?
425 :     data->directmvB[k].x
426 :     : mvs.x - data->referencemv[k].x);
427 : Isibaar 3
428 : chl 530 mvs.y = data->directmvF[k].y + y;
429 :     b_mvs.y = ((y == 0) ?
430 :     data->directmvB[k].y
431 :     : mvs.y - data->referencemv[k].y);
432 :    
433 :     if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx )
434 :     || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy )
435 :     || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx )
436 :     || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return;
437 : Isibaar 3
438 : syskin 748 if (!data->qpel) {
439 :     mvs.x *= 2; mvs.y *= 2;
440 : syskin 663 b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway
441 : chl 530 }
442 : syskin 663 ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
443 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
444 : chl 530
445 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
446 : syskin 663 ReferenceF, ReferenceB,
447 : chl 530 data->iEdgedWidth);
448 :     if (sad > *(data->iMinSAD)) return;
449 : edgomez 78 }
450 : chl 530
451 : syskin 724 sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000;
452 : syskin 628
453 : chl 530 if (sad < *(data->iMinSAD)) {
454 :     *(data->iMinSAD) = sad;
455 :     data->currentMV->x = x; data->currentMV->y = y;
456 :     *dir = Direction; }
457 : Isibaar 3 }
458 :    
459 : syskin 748 static void
460 : chl 530 CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
461 : chl 96 {
462 : chl 530 int32_t sad;
463 :     const uint8_t *ReferenceF;
464 :     const uint8_t *ReferenceB;
465 :     VECTOR mvs, b_mvs;
466 : syskin 724 const VECTOR zeroMV = {0,0};
467 : chl 96
468 : syskin 576 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
469 : edgomez 195
470 : chl 530 mvs.x = data->directmvF[0].x + x;
471 :     b_mvs.x = ((x == 0) ?
472 :     data->directmvB[0].x
473 :     : mvs.x - data->referencemv[0].x);
474 : edgomez 195
475 : chl 530 mvs.y = data->directmvF[0].y + y;
476 :     b_mvs.y = ((y == 0) ?
477 :     data->directmvB[0].y
478 :     : mvs.y - data->referencemv[0].y);
479 :    
480 :     if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx )
481 :     || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy )
482 :     || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx )
483 :     || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return;
484 : chl 96
485 : syskin 748 if (!data->qpel) {
486 :     mvs.x *= 2; mvs.y *= 2;
487 : syskin 663 b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway
488 : syskin 712 }
489 : syskin 663 ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
490 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
491 : chl 96
492 : syskin 628 sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
493 : syskin 724 sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000;
494 : chl 96
495 : chl 530 if (sad < *(data->iMinSAD)) {
496 :     *(data->iMinSAD) = sad;
497 :     data->currentMV->x = x; data->currentMV->y = y;
498 :     *dir = Direction; }
499 :     }
500 : chl 96
501 : chl 530 static void
502 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
503 :     {
504 : syskin 628 int32_t sad; int t;
505 : chl 530 const uint8_t * Reference;
506 : edgomez 195
507 : chl 530 if (( x > data->max_dx) || ( x < data->min_dx)
508 :     || ( y > data->max_dy) || (y < data->min_dy)) return;
509 : chl 96
510 : syskin 663 if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data);
511 : syskin 748 else Reference = GetReference(x, y, data);
512 : edgomez 195
513 : chl 530 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
514 : syskin 724 t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0);
515 : edgomez 195
516 : syskin 628 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100;
517 : chl 96
518 : Isibaar 606 if (sad < *(data->iMinSAD)) {
519 :     *(data->iMinSAD) = sad;
520 :     data->currentMV->x = x; data->currentMV->y = y;
521 :     *dir = Direction; }
522 :     }
523 :    
524 : Isibaar 539 /* CHECK_CANDIATE FUNCTIONS END */
525 : chl 96
526 : chl 530 /* MAINSEARCH FUNCTIONS START */
527 : edgomez 195
528 : chl 530 static void
529 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
530 : chl 181 {
531 :    
532 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
533 :    
534 : syskin 712 int iDirection;
535 : edgomez 195
536 : syskin 712 for(;;) { //forever
537 :     iDirection = 0;
538 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
539 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
540 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
541 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
542 : chl 181
543 : syskin 712 /* now we're doing diagonal checks near our candidate */
544 : chl 181
545 : syskin 712 if (iDirection) { //checking if anything found
546 :     bDirection = iDirection;
547 :     iDirection = 0;
548 :     x = data->currentMV->x; y = data->currentMV->y;
549 :     if (bDirection & 3) { //our candidate is left or right
550 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
551 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
552 :     } else { // what remains here is up or down
553 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
554 : syskin 748 CHECK_CANDIDATE(x - iDiamondSize, y, 1);
555 : syskin 712 }
556 : chl 181
557 : syskin 712 if (iDirection) {
558 :     bDirection += iDirection;
559 : syskin 748 x = data->currentMV->x; y = data->currentMV->y;
560 : chl 181 }
561 : syskin 712 } else { //about to quit, eh? not so fast....
562 :     switch (bDirection) {
563 :     case 2:
564 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
565 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
566 :     break;
567 :     case 1:
568 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
569 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
570 :     break;
571 :     case 2 + 4:
572 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
573 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
574 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
575 :     break;
576 :     case 4:
577 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
578 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
579 :     break;
580 :     case 8:
581 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
582 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
583 :     break;
584 :     case 1 + 4:
585 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
586 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
587 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
588 :     break;
589 :     case 2 + 8:
590 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
591 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
592 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
593 :     break;
594 :     case 1 + 8:
595 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
596 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
597 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
598 :     break;
599 :     default: //1+2+4+8 == we didn't find anything at all
600 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
601 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
602 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
603 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
604 :     break;
605 :     }
606 :     if (!iDirection) break; //ok, the end. really
607 :     bDirection = iDirection;
608 :     x = data->currentMV->x; y = data->currentMV->y;
609 : chl 181 }
610 : syskin 712 }
611 : chl 181 }
612 :    
613 : chl 530 static void
614 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
615 : chl 326 {
616 : chl 530 int iDirection;
617 : chl 326
618 : chl 530 do {
619 :     iDirection = 0;
620 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
621 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
622 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
623 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
624 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
625 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
626 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
627 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
628 : chl 326
629 : chl 530 bDirection = iDirection;
630 :     x = data->currentMV->x; y = data->currentMV->y;
631 :     } while (iDirection);
632 : chl 346 }
633 :    
634 : chl 530 static void
635 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
636 : chl 346 {
637 :    
638 : chl 181 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
639 :    
640 : syskin 712 int iDirection;
641 : edgomez 195
642 : syskin 712 do {
643 :     iDirection = 0;
644 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
645 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
646 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
647 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
648 : chl 181
649 : syskin 712 /* now we're doing diagonal checks near our candidate */
650 : chl 181
651 : syskin 712 if (iDirection) { //checking if anything found
652 :     bDirection = iDirection;
653 :     iDirection = 0;
654 :     x = data->currentMV->x; y = data->currentMV->y;
655 :     if (bDirection & 3) { //our candidate is left or right
656 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
657 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
658 :     } else { // what remains here is up or down
659 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
660 : syskin 748 CHECK_CANDIDATE(x - iDiamondSize, y, 1);
661 : chl 181 }
662 : syskin 712 bDirection += iDirection;
663 :     x = data->currentMV->x; y = data->currentMV->y;
664 : chl 181 }
665 : syskin 712 }
666 :     while (iDirection);
667 : chl 181 }
668 :    
669 : chl 530 /* MAINSEARCH FUNCTIONS END */
670 : chl 181
671 : chl 530 /* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */
672 :    
673 :     static void
674 : syskin 663 SubpelRefine(const SearchData * const data)
675 : chl 96 {
676 : syskin 663 /* Do a half-pel or q-pel refinement */
677 :     VECTOR backupMV;
678 : chl 530 int iDirection; //not needed
679 : chl 96
680 : syskin 663 if (data->qpel_precision)
681 :     backupMV = *(data->currentQMV);
682 :     else backupMV = *(data->currentMV);
683 : edgomez 195
684 : syskin 748 CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0);
685 : Isibaar 580 CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0);
686 : syskin 748 CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0);
687 :     CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0);
688 :     CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0);
689 : Isibaar 580 CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0);
690 :     CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0);
691 : syskin 748 CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0);
692 : Isibaar 580 }
693 :    
694 : chl 530 static __inline int
695 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
696 :     const int x, const int y,
697 : syskin 704 const uint32_t iEdgedWidth, const uint32_t iQuant, int rrv)
698 : chl 96
699 : Isibaar 3 {
700 : chl 530 /* keep repeating checks for all b-frames before this P frame,
701 :     to make sure that SKIP is possible (todo)
702 :     how: if skip is not possible set sad00 to a very high value */
703 : syskin 704 if(rrv) {
704 :     uint32_t sadC = sad16(current->u + x*16 + y*(iEdgedWidth/2)*16,
705 :     reference->u + x*16 + y*(iEdgedWidth/2)*16, iEdgedWidth/2, 256*4096);
706 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
707 :     sadC += sad16(current->v + (x + y*(iEdgedWidth/2))*16,
708 :     reference->v + (x + y*(iEdgedWidth/2))*16, iEdgedWidth/2, 256*4096);
709 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
710 :     return 1;
711 :     } else {
712 :     uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8,
713 :     reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2);
714 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
715 :     sadC += sad8(current->v + (x + y*(iEdgedWidth/2))*8,
716 :     reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
717 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
718 :     return 1;
719 :     }
720 : chl 530 }
721 : edgomez 195
722 : chl 530 static __inline void
723 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
724 :     {
725 :     pMB->mode = MODE_NOT_CODED;
726 : Isibaar 539 pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0;
727 :     pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0;
728 : syskin 628
729 : Isibaar 580 pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0;
730 :     pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0;
731 :    
732 : chl 530 pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
733 : Isibaar 3 }
734 :    
735 : syskin 628 bool
736 : chl 530 MotionEstimation(MBParam * const pParam,
737 :     FRAMEINFO * const current,
738 :     FRAMEINFO * const reference,
739 :     const IMAGE * const pRefH,
740 :     const IMAGE * const pRefV,
741 :     const IMAGE * const pRefHV,
742 :     const uint32_t iLimit)
743 : Isibaar 3 {
744 : chl 530 MACROBLOCK *const pMBs = current->mbs;
745 :     const IMAGE *const pCurrent = &current->image;
746 :     const IMAGE *const pRef = &reference->image;
747 : Isibaar 3
748 : chl 530 const VECTOR zeroMV = { 0, 0 };
749 : Isibaar 3
750 : syskin 704 uint32_t mb_width = pParam->mb_width;
751 :     uint32_t mb_height = pParam->mb_height;
752 : suxen_drol 702
753 : chl 530 uint32_t x, y;
754 :     uint32_t iIntra = 0;
755 : syskin 628 int32_t InterBias, quant = current->quant, sad00;
756 : Isibaar 580 uint8_t *qimage;
757 : edgomez 195
758 : Isibaar 539 // some pre-initialized thingies for SearchP
759 : syskin 712 int32_t temp[8];
760 : Isibaar 539 VECTOR currentMV[5];
761 : Isibaar 580 VECTOR currentQMV[5];
762 : Isibaar 539 int32_t iMinSAD[5];
763 :     SearchData Data;
764 : syskin 704 memset(&Data, 0, sizeof(SearchData));
765 : Isibaar 539 Data.iEdgedWidth = pParam->edged_width;
766 :     Data.currentMV = currentMV;
767 : Isibaar 580 Data.currentQMV = currentQMV;
768 : Isibaar 539 Data.iMinSAD = iMinSAD;
769 :     Data.temp = temp;
770 :     Data.iFcode = current->fcode;
771 : Isibaar 580 Data.rounding = pParam->m_rounding_type;
772 : syskin 628 Data.qpel = pParam->m_quarterpel;
773 :     Data.chroma = current->global_flags & XVID_ME_COLOUR;
774 : syskin 704 Data.rrv = current->global_flags & XVID_REDUCED;
775 : Isibaar 539
776 : syskin 704 if ((current->global_flags & XVID_REDUCED)) {
777 : suxen_drol 702 mb_width = (pParam->width + 31) / 32;
778 :     mb_height = (pParam->height + 31) / 32;
779 : syskin 704 Data.qpel = Data.chroma = 0;
780 : suxen_drol 702 }
781 :    
782 : Isibaar 580 if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL)
783 : syskin 594 return 1; // allocate some mem for qpel interpolated blocks
784 : Isibaar 580 // somehow this is dirty since I think we shouldn't use malloc outside
785 :     // encoder_create() - so please fix me!
786 : syskin 628 Data.RefQ = qimage;
787 : chl 530 if (sadInit) (*sadInit) ();
788 : edgomez 195
789 : suxen_drol 702 for (y = 0; y < mb_height; y++) {
790 :     for (x = 0; x < mb_width; x++) {
791 : chl 530 MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
792 : syskin 628
793 : syskin 748 if (Data.rrv) pMB->sad16 =
794 : syskin 704 sad32v_c(pCurrent->y + (x + y * pParam->edged_width) * 32,
795 :     pRef->y + (x + y * pParam->edged_width) * 32,
796 :     pParam->edged_width, pMB->sad8 );
797 :    
798 : syskin 748 else pMB->sad16 =
799 : syskin 704 sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16,
800 : chl 530 pRef->y + (x + y * pParam->edged_width) * 16,
801 :     pParam->edged_width, pMB->sad8 );
802 : edgomez 195
803 : syskin 628 if (Data.chroma) {
804 :     pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8,
805 :     pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2);
806 :    
807 :     pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8,
808 :     pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2);
809 :     }
810 :    
811 :     sad00 = pMB->sad16; //if no gmc; else sad00 = (..)
812 :    
813 : chl 530 if (!(current->global_flags & XVID_LUMIMASKING)) {
814 :     pMB->dquant = NO_CHANGE;
815 : syskin 748 pMB->quant = current->quant;
816 : syskin 601 } else {
817 : syskin 574 if (pMB->dquant != NO_CHANGE) {
818 :     quant += DQtab[pMB->dquant];
819 :     if (quant > 31) quant = 31;
820 :     else if (quant < 1) quant = 1;
821 :     }
822 : syskin 601 pMB->quant = quant;
823 :     }
824 : chl 181
825 : syskin 628 //initial skip decision
826 : chl 619 /* no early skip for GMC (global vector = skip vector is unknown!) */
827 :     if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */
828 : syskin 704 if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) )
829 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) {
830 : chl 619 SkipMacroblockP(pMB, sad00);
831 :     continue;
832 :     }
833 :     }
834 : syskin 628
835 :     SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
836 : chl 530 y, current->motion_flags, pMB->quant,
837 : Isibaar 539 &Data, pParam, pMBs, reference->mbs,
838 : chl 530 current->global_flags & XVID_INTER4V, pMB);
839 : edgomez 195
840 : chl 530 /* final skip decision, a.k.a. "the vector you found, really that good?" */
841 : syskin 748 if (current->coding_type == P_VOP) {
842 : chl 619 if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP)
843 : syskin 704 && ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) )
844 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) {
845 : chl 619 SkipMacroblockP(pMB, sad00);
846 :     continue;
847 :     }
848 :     }
849 : syskin 628
850 : chl 530 /* finally, intra decision */
851 : Isibaar 3
852 : chl 530 InterBias = MV16_INTER_BIAS;
853 : syskin 639 if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work
854 : chl 530 if (y != 0)
855 : syskin 639 if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
856 : chl 530 if (x != 0)
857 : syskin 639 if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
858 : syskin 628
859 :     if (Data.chroma) InterBias += 50; // to compensate bigger SAD
860 : syskin 704 if (Data.rrv) InterBias *= 4; //??
861 : Isibaar 3
862 : chl 530 if (InterBias < pMB->sad16) {
863 : syskin 704 int32_t deviation;
864 :     if (Data.rrv) {
865 :     deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 32,
866 :     pParam->edged_width)
867 :     + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16,
868 :     pParam->edged_width)
869 :     + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * pParam->edged_width,
870 :     pParam->edged_width)
871 :     + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * (pParam->edged_width+1),
872 :     pParam->edged_width);
873 :     } else
874 :     deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16,
875 : chl 530 pParam->edged_width);
876 : Isibaar 3
877 : syskin 704 if (deviation < (pMB->sad16 - InterBias)) {
878 : Isibaar 580 if (++iIntra >= iLimit) { free(qimage); return 1; }
879 : chl 530 pMB->mode = MODE_INTRA;
880 : Isibaar 539 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] =
881 :     pMB->mvs[3] = zeroMV;
882 : Isibaar 580 pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] =
883 :     pMB->qmvs[3] = zeroMV;
884 : chl 530 pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] =
885 :     pMB->sad8[3] = 0;
886 :     }
887 :     }
888 :     }
889 : Isibaar 3 }
890 : Isibaar 580 free(qimage);
891 : chl 619
892 :     if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */
893 : syskin 628 current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode );
894 : chl 619 else
895 :     current->GMC_MV = zeroMV;
896 :    
897 : syskin 628 return 0;
898 : chl 530 }
899 : Isibaar 3
900 : edgomez 195
901 : chl 530 #define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16)
902 : Isibaar 3
903 : chl 530 static __inline int
904 :     make_mask(const VECTOR * const pmv, const int i)
905 :     {
906 : Isibaar 539 int mask = 255, j;
907 : chl 530 for (j = 0; j < i; j++) {
908 :     if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
909 :     if (pmv[i].x == pmv[j].x) {
910 :     if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; }
911 :     if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; }
912 : syskin 748 } else
913 : chl 530 if (pmv[i].y == pmv[j].y) {
914 :     if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; }
915 :     if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; }
916 :     }
917 : Isibaar 3 }
918 : chl 530 return mask;
919 :     }
920 : edgomez 195
921 : syskin 748 static __inline void
922 : syskin 704 PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
923 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
924 : chl 530 {
925 : edgomez 195
926 : chl 530 //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
927 : syskin 704 if (rrv) { iWcount /= 2; iHcount /= 2; }
928 : edgomez 195
929 : syskin 704 if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
930 : chl 530 pmv[5].x = EVEN(pmv[3].x);
931 : syskin 748 pmv[5].y = EVEN(pmv[3].y);
932 : Isibaar 539 } else pmv[5].x = pmv[5].y = 0;
933 : chl 169
934 : chl 530 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
935 :     else pmv[3].x = pmv[3].y = 0;
936 : Isibaar 3
937 : chl 530 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
938 :     else pmv[4].x = pmv[4].y = 0;
939 : chl 169
940 : chl 530 // [1] median prediction
941 : syskin 704 if (rrv) { //median is in halfzero-precision
942 :     pmv[1].x = RRV_MV_SCALEUP(pmv[0].x);
943 :     pmv[1].y = RRV_MV_SCALEUP(pmv[0].y);
944 :     } else { pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); }
945 : chl 169
946 : chl 530 pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask
947 : chl 169
948 : chl 530 pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
949 :     pmv[2].y = EVEN(prevMB->mvs[0].y);
950 : chl 169
951 : syskin 704 if ((x < iWcount-1) && (y < iHcount-1)) {
952 : chl 530 pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
953 : syskin 748 pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
954 : Isibaar 539 } else pmv[6].x = pmv[6].y = 0;
955 : syskin 704
956 :     if (rrv) {
957 :     int i;
958 :     for (i = 0; i < 7; i++) {
959 :     pmv[i].x = RRV_MV_SCALEDOWN(pmv[i].x);
960 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // a trick
961 :     }
962 :     }
963 : chl 530 }
964 : chl 169
965 : chl 530 static void
966 : syskin 628 SearchP(const IMAGE * const pRef,
967 : chl 530 const uint8_t * const pRefH,
968 :     const uint8_t * const pRefV,
969 :     const uint8_t * const pRefHV,
970 :     const IMAGE * const pCur,
971 :     const int x,
972 :     const int y,
973 :     const uint32_t MotionFlags,
974 :     const uint32_t iQuant,
975 : Isibaar 539 SearchData * const Data,
976 : chl 530 const MBParam * const pParam,
977 :     const MACROBLOCK * const pMBs,
978 :     const MACROBLOCK * const prevMBs,
979 :     int inter4v,
980 :     MACROBLOCK * const pMB)
981 :     {
982 : chl 169
983 : chl 530 int i, iDirection = 255, mask, threshA;
984 : Isibaar 539 VECTOR pmv[7];
985 : Isibaar 3
986 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
987 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
988 : Isibaar 3
989 : syskin 712 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)()
990 : syskin 628
991 : syskin 712 Data->temp[5] = Data->temp[7] = 256*4096; // to reset chroma-sad cache
992 : syskin 704 if (Data->rrv) i = 2; else i = 1;
993 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
994 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
995 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
996 : Isibaar 3
997 : syskin 704 Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
998 :     Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
999 :     Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1000 :     Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1001 :     Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1002 :     Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1003 :    
1004 : syskin 601 Data->lambda16 = lambda_vec16[iQuant];
1005 : Isibaar 606 Data->lambda8 = lambda_vec8[iQuant];
1006 : syskin 663 Data->qpel_precision = 0;
1007 : chl 169
1008 : edgomez 568 if (pMB->dquant != NO_CHANGE) inter4v = 0;
1009 : edgomez 195
1010 : Isibaar 580 for(i = 0; i < 5; i++)
1011 :     Data->currentMV[i].x = Data->currentMV[i].y = 0;
1012 : Isibaar 539
1013 : syskin 663 if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1014 :     else Data->predMV = pmv[0];
1015 : Isibaar 606
1016 : syskin 724 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1017 : syskin 628 Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000;
1018 :     Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100;
1019 : Isibaar 539 Data->iMinSAD[2] = pMB->sad8[1];
1020 :     Data->iMinSAD[3] = pMB->sad8[2];
1021 :     Data->iMinSAD[4] = pMB->sad8[3];
1022 :    
1023 : chl 530 if ((x == 0) && (y == 0)) threshA = 512;
1024 :     else {
1025 : edgomez 568 threshA = Data->temp[0]; // that's when we keep this SAD atm
1026 : chl 530 if (threshA < 512) threshA = 512;
1027 :     if (threshA > 1024) threshA = 1024; }
1028 : Isibaar 3
1029 : chl 530 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1030 : syskin 704 prevMBs + x + y * pParam->mb_width, Data->rrv);
1031 : edgomez 195
1032 : syskin 704 if (Data->rrv) CheckCandidate = CheckCandidate32;
1033 :     else if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16;
1034 :     else CheckCandidate = CheckCandidate16no4v; //for extra speed
1035 : edgomez 568
1036 : chl 530 /* main loop. checking all predictions */
1037 :    
1038 :     for (i = 1; i < 7; i++) {
1039 :     if (!(mask = make_mask(pmv, i)) ) continue;
1040 : edgomez 568 (*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1041 : Isibaar 539 if (Data->iMinSAD[0] <= threshA) break;
1042 : chl 530 }
1043 :    
1044 : Isibaar 539 if ((Data->iMinSAD[0] <= threshA) ||
1045 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1046 :     (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1047 : chl 530 inter4v = 0;
1048 : Isibaar 539 } else {
1049 : Isibaar 3
1050 : syskin 748 MainSearchFunc * MainSearchPtr;
1051 : Isibaar 539 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1052 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1053 :     else MainSearchPtr = DiamondSearch;
1054 : chl 181
1055 : Isibaar 539 (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1056 : Isibaar 3
1057 : chl 530 /* extended search, diamond starting in 0,0 and in prediction.
1058 :     note that this search is/might be done in halfpel positions,
1059 :     which makes it more different than the diamond above */
1060 : chl 259
1061 : Isibaar 539 if (MotionFlags & PMV_EXTSEARCH16) {
1062 :     int32_t bSAD;
1063 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1064 : syskin 704 if (Data->rrv) {
1065 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1066 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1067 : syskin 748 } else
1068 : syskin 704 if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel?
1069 :     startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y);
1070 : Isibaar 539 if (!(MVequal(startMV, backupMV))) {
1071 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1072 : Isibaar 3
1073 : Isibaar 606 (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data);
1074 : Isibaar 539 (*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
1075 :     if (bSAD < Data->iMinSAD[0]) {
1076 :     Data->currentMV[0] = backupMV;
1077 :     Data->iMinSAD[0] = bSAD; }
1078 :     }
1079 : Isibaar 3
1080 : Isibaar 539 backupMV = Data->currentMV[0];
1081 : syskin 704 if (!MotionFlags & PMV_HALFPELREFINE16 || Data->rrv) startMV.x = startMV.y = 0;
1082 :     else startMV.x = startMV.y = 1;
1083 : Isibaar 539 if (!(MVequal(startMV, backupMV))) {
1084 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1085 : edgomez 195
1086 : Isibaar 606 (*CheckCandidate)(startMV.x, startMV.y, 255, &iDirection, Data);
1087 : Isibaar 539 (*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
1088 :     if (bSAD < Data->iMinSAD[0]) {
1089 :     Data->currentMV[0] = backupMV;
1090 :     Data->iMinSAD[0] = bSAD; }
1091 :     }
1092 : Isibaar 3 }
1093 :     }
1094 :    
1095 : syskin 663 if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data);
1096 : Isibaar 3
1097 : Isibaar 580 for(i = 0; i < 5; i++) {
1098 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1099 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1100 :     }
1101 :    
1102 : syskin 748 if((Data->qpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) {
1103 : Isibaar 580
1104 : syskin 663 Data->qpel_precision = 1;
1105 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1106 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1107 : syskin 601
1108 : syskin 663 SubpelRefine(Data);
1109 : Isibaar 580 }
1110 :    
1111 : syskin 628 if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0;
1112 : Isibaar 539 if (inter4v) {
1113 :     SearchData Data8;
1114 :     Data8.iFcode = Data->iFcode;
1115 : Isibaar 606 Data8.lambda8 = Data->lambda8;
1116 : Isibaar 539 Data8.iEdgedWidth = Data->iEdgedWidth;
1117 : Isibaar 606 Data8.RefQ = Data->RefQ;
1118 : syskin 628 Data8.qpel = Data->qpel;
1119 : syskin 704 Data8.rrv = Data->rrv;
1120 : Isibaar 539 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1121 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1122 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1123 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1124 : syskin 628
1125 :     if (Data->chroma) {
1126 : syskin 753 int sumx, sumy;
1127 : syskin 628
1128 :     if(pParam->m_quarterpel) {
1129 : syskin 753 sumx = pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2;
1130 : syskin 663 sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2;
1131 :     } else {
1132 :     sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x;
1133 :     sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y;
1134 :     }
1135 : syskin 628
1136 : syskin 753 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1137 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1138 : syskin 628 }
1139 : Isibaar 539 }
1140 : Isibaar 3
1141 : syskin 704 if (Data->rrv) {
1142 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1143 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1144 :     }
1145 : syskin 740
1146 : chl 530 if (!(inter4v) ||
1147 : syskin 748 (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1148 : Isibaar 539 Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) {
1149 : chl 530 // INTER MODE
1150 :     pMB->mode = MODE_INTER;
1151 : Isibaar 539 pMB->mvs[0] = pMB->mvs[1]
1152 :     = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1153 : Isibaar 3
1154 : chl 530 pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] =
1155 : Isibaar 539 pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1156 : Isibaar 3
1157 : Isibaar 580 if(pParam->m_quarterpel) {
1158 : syskin 663 pMB->qmvs[0] = pMB->qmvs[1]
1159 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1160 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1161 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1162 : syskin 628 } else {
1163 : Isibaar 580 pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1164 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1165 :     }
1166 : edgomez 195 } else {
1167 : syskin 748 // INTER4V MODE; all other things are already set in Search8
1168 : chl 530 pMB->mode = MODE_INTER4V;
1169 : syskin 748 pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] +
1170 : Isibaar 539 Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant;
1171 : edgomez 78 }
1172 : Isibaar 3 }
1173 :    
1174 : chl 530 static void
1175 :     Search8(const SearchData * const OldData,
1176 :     const int x, const int y,
1177 :     const uint32_t MotionFlags,
1178 :     const MBParam * const pParam,
1179 :     MACROBLOCK * const pMB,
1180 :     const MACROBLOCK * const pMBs,
1181 : Isibaar 539 const int block,
1182 :     SearchData * const Data)
1183 : chl 345 {
1184 : syskin 704 int i = 0;
1185 : Isibaar 539 Data->iMinSAD = OldData->iMinSAD + 1 + block;
1186 :     Data->currentMV = OldData->currentMV + 1 + block;
1187 : Isibaar 580 Data->currentQMV = OldData->currentQMV + 1 + block;
1188 : chl 345
1189 : syskin 601 if(pParam->m_quarterpel) {
1190 : syskin 704 Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1191 : syskin 724 if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1192 :     Data->predMV, Data->iFcode, 0, 0);
1193 : syskin 712
1194 : Isibaar 606 } else {
1195 : syskin 704 Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1196 :     if (block != 0) {
1197 : syskin 724 if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1198 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1199 : syskin 704 }
1200 : Isibaar 606 }
1201 : chl 345
1202 : syskin 704 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100;
1203 :    
1204 : chl 530 if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) {
1205 : syskin 704 if (Data->rrv) i = 2; else i = 1;
1206 : chl 345
1207 : syskin 704 Data->Ref = OldData->Ref + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1208 :     Data->RefH = OldData->RefH + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1209 :     Data->RefV = OldData->RefV + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1210 :     Data->RefHV = OldData->RefHV + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1211 : chl 345
1212 : syskin 704 Data->Cur = OldData->Cur + i*8 * ((block&1) + pParam->edged_width*(block>>1));
1213 : syskin 663 Data->qpel_precision = 0;
1214 : chl 345
1215 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1216 :     pParam->width, pParam->height, OldData->iFcode - Data->qpel, 0, Data->rrv);
1217 :    
1218 :     if (Data->rrv) CheckCandidate = CheckCandidate16no4v;
1219 :     else CheckCandidate = CheckCandidate8;
1220 :    
1221 : chl 530 if (MotionFlags & PMV_EXTSEARCH8) {
1222 : Isibaar 580 int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1223 :    
1224 : chl 530 MainSearchFunc *MainSearchPtr;
1225 :     if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1226 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1227 :     else MainSearchPtr = DiamondSearch;
1228 : chl 345
1229 : Isibaar 580 (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255);
1230 : chl 345
1231 : syskin 601 if(*(Data->iMinSAD) < temp_sad) {
1232 : Isibaar 580 Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1233 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1234 :     }
1235 :     }
1236 :    
1237 :     if (MotionFlags & PMV_HALFPELREFINE8) {
1238 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1239 :    
1240 : syskin 663 SubpelRefine(Data); // perform halfpel refine of current best vector
1241 : Isibaar 580
1242 :     if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1243 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1244 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1245 :     }
1246 :     }
1247 :    
1248 : syskin 748 if(Data->qpel) {
1249 : syskin 601 if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) &&
1250 :     (MotionFlags & PMV_QUARTERPELREFINE8)) {
1251 : syskin 663 Data->qpel_precision = 1;
1252 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1253 :     pParam->width, pParam->height, OldData->iFcode, 1, 0);
1254 : syskin 663 SubpelRefine(Data);
1255 : syskin 601 }
1256 : Isibaar 580 }
1257 : chl 345 }
1258 : chl 530
1259 : syskin 704 if (Data->rrv) {
1260 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1261 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1262 :     }
1263 :    
1264 : syskin 712 if(Data->qpel) {
1265 : syskin 663 pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1266 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1267 :     pMB->qmvs[block] = *(Data->currentQMV);
1268 : syskin 704 } else {
1269 : Isibaar 580 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1270 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1271 :     }
1272 :    
1273 : Isibaar 539 pMB->mvs[block] = *(Data->currentMV);
1274 : syskin 601 pMB->sad8[block] = 4 * (*Data->iMinSAD);
1275 : chl 345 }
1276 :    
1277 : chl 530 /* B-frames code starts here */
1278 : chl 345
1279 : chl 530 static __inline VECTOR
1280 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1281 : Isibaar 3 {
1282 : chl 530 /* the stupidiest function ever */
1283 :     if (mode == MODE_FORWARD) return pMB->mvs[0];
1284 :     else return pMB->b_mvs[0];
1285 : Isibaar 3 }
1286 :    
1287 : chl 530 static void __inline
1288 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1289 :     const uint32_t iWcount,
1290 :     const MACROBLOCK * const pMB,
1291 :     const uint32_t mode_curr)
1292 : Isibaar 3 {
1293 :    
1294 : chl 530 // [0] is prediction
1295 :     pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1296 : Isibaar 3
1297 : chl 530 pmv[1].x = pmv[1].y = 0; // [1] is zero
1298 : Isibaar 3
1299 : chl 530 pmv[2] = ChoosePred(pMB, mode_curr);
1300 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1301 : edgomez 195
1302 : chl 530 if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1303 :     pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1304 : syskin 748 pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1305 : Isibaar 539 } else pmv[3].x = pmv[3].y = 0;
1306 : Isibaar 3
1307 : chl 530 if (y != 0) {
1308 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1309 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1310 :     } else pmv[4].x = pmv[4].y = 0;
1311 : Isibaar 3
1312 : chl 530 if (x != 0) {
1313 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1314 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1315 :     } else pmv[5].x = pmv[5].y = 0;
1316 : Isibaar 3
1317 : chl 530 if ((x != 0)&&(y != 0)) {
1318 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1319 :     pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1320 :     } else pmv[6].x = pmv[6].y = 0;
1321 : edgomez 195
1322 : chl 530 // more?
1323 :     }
1324 : chl 181
1325 : edgomez 170
1326 : chl 530 /* search backward or forward, for b-frames */
1327 :     static void
1328 :     SearchBF( const uint8_t * const pRef,
1329 :     const uint8_t * const pRefH,
1330 :     const uint8_t * const pRefV,
1331 :     const uint8_t * const pRefHV,
1332 :     const IMAGE * const pCur,
1333 :     const int x, const int y,
1334 :     const uint32_t MotionFlags,
1335 :     const uint32_t iFcode,
1336 :     const MBParam * const pParam,
1337 :     MACROBLOCK * const pMB,
1338 :     const VECTOR * const predMV,
1339 :     int32_t * const best_sad,
1340 : h 545 const int32_t mode_current,
1341 :     SearchData * const Data)
1342 : chl 530 {
1343 : Isibaar 3
1344 : chl 530 const int32_t iEdgedWidth = pParam->edged_width;
1345 :    
1346 : syskin 740 int i, iDirection = 255, mask;
1347 : h 545 VECTOR pmv[7];
1348 : chl 530 MainSearchFunc *MainSearchPtr;
1349 : h 545 *Data->iMinSAD = MV_MAX_ERROR;
1350 :     Data->iFcode = iFcode;
1351 : syskin 663 Data->qpel_precision = 0;
1352 : Isibaar 3
1353 : h 545 Data->Ref = pRef + (x + y * iEdgedWidth) * 16;
1354 :     Data->RefH = pRefH + (x + y * iEdgedWidth) * 16;
1355 :     Data->RefV = pRefV + (x + y * iEdgedWidth) * 16;
1356 :     Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16;
1357 : Isibaar 3
1358 : h 545 Data->predMV = *predMV;
1359 : edgomez 195
1360 : h 545 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1361 : syskin 712 pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1362 : Isibaar 3
1363 : h 545 pmv[0] = Data->predMV;
1364 : syskin 658 if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1365 : h 545 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1366 : edgomez 195
1367 : h 545 Data->currentMV->x = Data->currentMV->y = 0;
1368 : chl 530 CheckCandidate = CheckCandidate16no4v;
1369 : chl 181
1370 : chl 530 // main loop. checking all predictions
1371 : syskin 740 for (i = 0; i < 7; i++) {
1372 : chl 530 if (!(mask = make_mask(pmv, i)) ) continue;
1373 : h 545 CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1374 : edgomez 78 }
1375 : Isibaar 3
1376 : chl 530 if (MotionFlags & PMV_USESQUARES16)
1377 :     MainSearchPtr = SquareSearch;
1378 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16)
1379 :     MainSearchPtr = AdvDiamondSearch;
1380 :     else MainSearchPtr = DiamondSearch;
1381 : chl 96
1382 : syskin 740 (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1383 : chl 169
1384 : syskin 663 SubpelRefine(Data);
1385 : syskin 658
1386 : syskin 740 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1387 : syskin 658 Data->currentQMV->x = 2*Data->currentMV->x;
1388 :     Data->currentQMV->y = 2*Data->currentMV->y;
1389 : syskin 663 Data->qpel_precision = 1;
1390 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1391 :     pParam->width, pParam->height, iFcode, 1, 0);
1392 : syskin 663 SubpelRefine(Data);
1393 : syskin 658 }
1394 : chl 169
1395 : chl 530 // three bits are needed to code backward mode. four for forward
1396 : syskin 740
1397 : syskin 628 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1398 :     else *Data->iMinSAD += 3 * Data->lambda16;
1399 : chl 169
1400 : h 545 if (*Data->iMinSAD < *best_sad) {
1401 :     *best_sad = *Data->iMinSAD;
1402 : chl 530 pMB->mode = mode_current;
1403 : syskin 658 if (Data->qpel) {
1404 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1405 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1406 : syskin 748 if (mode_current == MODE_FORWARD)
1407 : syskin 658 pMB->qmvs[0] = *Data->currentQMV;
1408 : syskin 748 else
1409 : syskin 658 pMB->b_qmvs[0] = *Data->currentQMV;
1410 :     } else {
1411 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1412 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1413 :     }
1414 : syskin 739 if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1415 :     else pMB->b_mvs[0] = *Data->currentMV;
1416 :     }
1417 :     if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1418 :     else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
1419 : syskin 658
1420 : chl 530 }
1421 : chl 169
1422 : syskin 748 static void
1423 : syskin 712 SkipDecisionB(const IMAGE * const pCur,
1424 :     const IMAGE * const f_Ref,
1425 : syskin 748 const IMAGE * const b_Ref,
1426 : syskin 712 MACROBLOCK * const pMB,
1427 :     const uint32_t quant,
1428 :     const uint32_t x, const uint32_t y,
1429 :     const SearchData * const Data)
1430 :     {
1431 :     int dx, dy, b_dx, b_dy;
1432 :     uint32_t sum;
1433 :     //this is not full chroma compensation, only it's fullpel approximation. should work though
1434 :     if (Data->qpel) {
1435 :     dy = Data->directmvF[0].y/2 + Data->directmvF[1].y/2 +
1436 :     Data->directmvF[2].y/2 + Data->directmvF[3].y/2;
1437 :    
1438 :     dx = Data->directmvF[0].x/2 + Data->directmvF[1].x/2 +
1439 :     Data->directmvF[2].x/2 + Data->directmvF[3].x/2;
1440 :    
1441 :     b_dy = Data->directmvB[0].y/2 + Data->directmvB[1].y/2 +
1442 :     Data->directmvB[2].y/2 + Data->directmvB[3].y/2;
1443 :    
1444 :     b_dx = Data->directmvB[0].x/2 + Data->directmvB[1].x/2 +
1445 :     Data->directmvB[2].x/2 + Data->directmvB[3].x/2;
1446 :    
1447 :     } else {
1448 :     dy = Data->directmvF[0].y + Data->directmvF[1].y +
1449 :     Data->directmvF[2].y + Data->directmvF[3].y;
1450 :    
1451 :     dx = Data->directmvF[0].x + Data->directmvF[1].x +
1452 :     Data->directmvF[2].x + Data->directmvF[3].x;
1453 :    
1454 :     b_dy = Data->directmvB[0].y + Data->directmvB[1].y +
1455 :     Data->directmvB[2].y + Data->directmvB[3].y;
1456 :    
1457 :     b_dx = Data->directmvB[0].x + Data->directmvB[1].x +
1458 :     Data->directmvB[2].x + Data->directmvB[3].x;
1459 :     }
1460 :    
1461 :    
1462 :     dy = (dy >> 3) + roundtab_76[dy & 0xf];
1463 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1464 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1465 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1466 :    
1467 :     sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2),
1468 :     f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2,
1469 :     b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2,
1470 :     Data->iEdgedWidth/2);
1471 :     sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2),
1472 :     f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2,
1473 :     b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2,
1474 :     Data->iEdgedWidth/2);
1475 :    
1476 :     if (sum < 2*MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1477 :     }
1478 :    
1479 :    
1480 :    
1481 :     static __inline uint32_t
1482 : Isibaar 539 SearchDirect(const IMAGE * const f_Ref,
1483 : chl 530 const uint8_t * const f_RefH,
1484 :     const uint8_t * const f_RefV,
1485 :     const uint8_t * const f_RefHV,
1486 : Isibaar 539 const IMAGE * const b_Ref,
1487 : chl 530 const uint8_t * const b_RefH,
1488 :     const uint8_t * const b_RefV,
1489 :     const uint8_t * const b_RefHV,
1490 :     const IMAGE * const pCur,
1491 :     const int x, const int y,
1492 :     const uint32_t MotionFlags,
1493 :     const int32_t TRB, const int32_t TRD,
1494 :     const MBParam * const pParam,
1495 :     MACROBLOCK * const pMB,
1496 :     const MACROBLOCK * const b_mb,
1497 : h 545 int32_t * const best_sad,
1498 :     SearchData * const Data)
1499 : edgomez 195
1500 : chl 530 {
1501 : h 545 int32_t skip_sad;
1502 : chl 530 int k;
1503 : h 545
1504 : chl 530 MainSearchFunc *MainSearchPtr;
1505 : Isibaar 3
1506 : h 545 *Data->iMinSAD = 256*4096;
1507 : chl 140
1508 : h 545 Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16;
1509 :     Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16;
1510 :     Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16;
1511 :     Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16;
1512 :     Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16;
1513 :     Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16;
1514 :     Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16;
1515 :     Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16;
1516 : chl 140
1517 : h 545 Data->max_dx = 2 * pParam->width - 2 * (x) * 16;
1518 :     Data->max_dy = 2 * pParam->height - 2 * (y) * 16;
1519 :     Data->min_dx = -(2 * 16 + 2 * (x) * 16);
1520 :     Data->min_dy = -(2 * 16 + 2 * (y) * 16);
1521 : syskin 658 if (Data->qpel) { //we measure in qpixels
1522 :     Data->max_dx *= 2;
1523 :     Data->max_dy *= 2;
1524 :     Data->min_dx *= 2;
1525 :     Data->min_dy *= 2;
1526 :     Data->referencemv = b_mb->qmvs;
1527 :     } else Data->referencemv = b_mb->mvs;
1528 : syskin 672 Data->qpel_precision = 0; // it's a trick. it's 1 not 0, but we need 0 here
1529 : Isibaar 3
1530 : chl 530 for (k = 0; k < 4; k++) {
1531 : h 545 pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1532 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1533 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1534 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1535 : Isibaar 3
1536 : h 545 if ( ( pMB->b_mvs[k].x > Data->max_dx ) || ( pMB->b_mvs[k].x < Data->min_dx )
1537 :     || ( pMB->b_mvs[k].y > Data->max_dy ) || ( pMB->b_mvs[k].y < Data->min_dy )) {
1538 : Isibaar 3
1539 : Isibaar 539 *best_sad = 256*4096; // in that case, we won't use direct mode
1540 :     pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1541 :     pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1542 : syskin 712 return 256*4096;
1543 : Isibaar 539 }
1544 :     if (b_mb->mode != MODE_INTER4V) {
1545 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1546 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1547 : h 545 Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1548 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1549 : Isibaar 539 break;
1550 :     }
1551 :     }
1552 : Isibaar 3
1553 : syskin 663 if (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect;
1554 :     else CheckCandidate = CheckCandidateDirectno4v;
1555 : syskin 658
1556 : h 545 (*CheckCandidate)(0, 0, 255, &k, Data);
1557 : Isibaar 3
1558 : syskin 712 // initial (fast) skip decision
1559 :     if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH*2) {
1560 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data->chroma, Data); //possible skip - checking chroma
1561 :     if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1562 : Isibaar 539 }
1563 :    
1564 : h 545 skip_sad = *Data->iMinSAD;
1565 : edgomez 195
1566 : chl 530 // DIRECT MODE DELTA VECTOR SEARCH.
1567 :     // This has to be made more effective, but at the moment I'm happy it's running at all
1568 : Isibaar 3
1569 : chl 530 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1570 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1571 :     else MainSearchPtr = DiamondSearch;
1572 : edgomez 195
1573 : h 545 (*MainSearchPtr)(0, 0, Data, 255);
1574 : Isibaar 3
1575 : syskin 663 SubpelRefine(Data);
1576 : Isibaar 3
1577 : h 545 *best_sad = *Data->iMinSAD;
1578 : Isibaar 3
1579 : syskin 743 if (b_mb->mode == MODE_INTER4V || Data->qpel) pMB->mode = MODE_DIRECT;
1580 : syskin 672 else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1581 : edgomez 195
1582 : h 545 pMB->pmvs[3] = *Data->currentMV;
1583 : edgomez 195
1584 : chl 530 for (k = 0; k < 4; k++) {
1585 : h 545 pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1586 : syskin 658 pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1587 : h 545 ? Data->directmvB[k].x
1588 : syskin 658 :pMB->mvs[k].x - Data->referencemv[k].x);
1589 : h 545 pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1590 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1591 :     ? Data->directmvB[k].y
1592 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1593 : syskin 658 if (Data->qpel) {
1594 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1595 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1596 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1597 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1598 :     }
1599 :    
1600 : chl 530 if (b_mb->mode != MODE_INTER4V) {
1601 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1602 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1603 : syskin 658 pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1604 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1605 : chl 530 break;
1606 :     }
1607 :     }
1608 : h 545 return skip_sad;
1609 : Isibaar 3 }
1610 : chl 96
1611 : h 545
1612 : syskin 712 static void
1613 : chl 530 SearchInterpolate(const uint8_t * const f_Ref,
1614 :     const uint8_t * const f_RefH,
1615 :     const uint8_t * const f_RefV,
1616 :     const uint8_t * const f_RefHV,
1617 :     const uint8_t * const b_Ref,
1618 :     const uint8_t * const b_RefH,
1619 :     const uint8_t * const b_RefV,
1620 :     const uint8_t * const b_RefHV,
1621 :     const IMAGE * const pCur,
1622 :     const int x, const int y,
1623 :     const uint32_t fcode,
1624 :     const uint32_t bcode,
1625 :     const uint32_t MotionFlags,
1626 :     const MBParam * const pParam,
1627 :     const VECTOR * const f_predMV,
1628 :     const VECTOR * const b_predMV,
1629 :     MACROBLOCK * const pMB,
1630 : h 545 int32_t * const best_sad,
1631 :     SearchData * const fData)
1632 : chl 530
1633 : chl 96 {
1634 :    
1635 : chl 530 int iDirection, i, j;
1636 : h 545 SearchData bData;
1637 : chl 96
1638 : syskin 712 fData->qpel_precision = 0;
1639 :     memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1640 :     *fData->iMinSAD = 4096*256;
1641 : syskin 739 bData.currentMV++; bData.currentQMV++;
1642 : h 545 fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1643 : edgomez 195
1644 : syskin 712 i = (x + y * fData->iEdgedWidth) * 16;
1645 :     bData.bRef = fData->Ref = f_Ref + i;
1646 :     bData.bRefH = fData->RefH = f_RefH + i;
1647 :     bData.bRefV = fData->RefV = f_RefV + i;
1648 :     bData.bRefHV = fData->RefHV = f_RefHV + i;
1649 :     bData.Ref = fData->bRef = b_Ref + i;
1650 :     bData.RefH = fData->bRefH = b_RefH + i;
1651 :     bData.RefV = fData->bRefV = b_RefV + i;
1652 :     bData.RefHV = fData->bRefHV = b_RefHV + i;
1653 : chl 96
1654 : h 545 bData.bpredMV = fData->predMV = *f_predMV;
1655 :     fData->bpredMV = bData.predMV = *b_predMV;
1656 : syskin 661 fData->currentMV[0] = fData->currentMV[2];
1657 : chl 96
1658 : syskin 712 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0);
1659 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0);
1660 :    
1661 : h 545 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1662 : syskin 675 if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1663 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1664 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1665 : Isibaar 539
1666 : h 545 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1667 : syskin 675 if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1668 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1669 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1670 : Isibaar 539
1671 : h 545 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1672 : chl 96
1673 : chl 530 //diamond. I wish we could use normal mainsearch functions (square, advdiamond)
1674 : chl 96
1675 : chl 530 do {
1676 :     iDirection = 255;
1677 :     // forward MV moves
1678 : h 545 i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1679 : chl 96
1680 : h 545 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1681 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1682 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1683 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1684 : chl 96
1685 : chl 530 // backward MV moves
1686 : h 545 i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1687 :     fData->currentMV[2] = fData->currentMV[0];
1688 : Isibaar 539 CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1689 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1690 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1691 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1692 : chl 96
1693 : chl 530 } while (!(iDirection));
1694 : edgomez 195
1695 : syskin 658 if (fData->qpel) {
1696 : syskin 740 if (*fData->iMinSAD > *best_sad + 500) return;
1697 : syskin 675 CheckCandidate = CheckCandidateInt;
1698 : syskin 663 fData->qpel_precision = bData.qpel_precision = 1;
1699 : syskin 712 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1700 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1701 : syskin 658 fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1702 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1703 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1704 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1705 : syskin 663 SubpelRefine(fData);
1706 : syskin 740 if (*fData->iMinSAD > *best_sad + 300) return;
1707 : syskin 658 fData->currentQMV[2] = fData->currentQMV[0];
1708 : syskin 663 SubpelRefine(&bData);
1709 : syskin 658 }
1710 : syskin 676
1711 : syskin 740 *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode.
1712 : syskin 658
1713 : h 545 if (*fData->iMinSAD < *best_sad) {
1714 :     *best_sad = *fData->iMinSAD;
1715 :     pMB->mvs[0] = fData->currentMV[0];
1716 :     pMB->b_mvs[0] = fData->currentMV[1];
1717 : chl 530 pMB->mode = MODE_INTERPOLATE;
1718 : syskin 658 if (fData->qpel) {
1719 :     pMB->qmvs[0] = fData->currentQMV[0];
1720 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1721 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1722 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1723 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1724 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1725 :     } else {
1726 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1727 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1728 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1729 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1730 :     }
1731 : chl 530 }
1732 :     }
1733 : chl 96
1734 : chl 530 void
1735 :     MotionEstimationBVOP(MBParam * const pParam,
1736 :     FRAMEINFO * const frame,
1737 :     const int32_t time_bp,
1738 :     const int32_t time_pp,
1739 :     // forward (past) reference
1740 :     const MACROBLOCK * const f_mbs,
1741 :     const IMAGE * const f_ref,
1742 :     const IMAGE * const f_refH,
1743 :     const IMAGE * const f_refV,
1744 :     const IMAGE * const f_refHV,
1745 :     // backward (future) reference
1746 : syskin 644 const FRAMEINFO * const b_reference,
1747 : chl 530 const IMAGE * const b_ref,
1748 :     const IMAGE * const b_refH,
1749 :     const IMAGE * const b_refV,
1750 :     const IMAGE * const b_refHV)
1751 :     {
1752 :     uint32_t i, j;
1753 : syskin 712 int32_t best_sad;
1754 :     uint32_t skip_sad;
1755 : chl 530 int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
1756 :     static const VECTOR zeroMV={0,0};
1757 : syskin 644 const MACROBLOCK * const b_mbs = b_reference->mbs;
1758 : chl 96
1759 : chl 530 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
1760 : chl 96
1761 : chl 530 const int32_t TRB = time_pp - time_bp;
1762 :     const int32_t TRD = time_pp;
1763 : syskin 658 uint8_t * qimage;
1764 : chl 96
1765 : h 545 // some pre-inintialized data for the rest of the search
1766 :    
1767 :     SearchData Data;
1768 :     int32_t iMinSAD;
1769 :     VECTOR currentMV[3];
1770 : syskin 658 VECTOR currentQMV[3];
1771 : syskin 704 memset(&Data, 0, sizeof(SearchData));
1772 : h 545 Data.iEdgedWidth = pParam->edged_width;
1773 : syskin 658 Data.currentMV = currentMV; Data.currentQMV = currentQMV;
1774 : h 545 Data.iMinSAD = &iMinSAD;
1775 : syskin 712 Data.lambda16 = lambda_vec16[frame->quant];
1776 :     Data.chroma = frame->quant;
1777 : syskin 658 Data.qpel = pParam->m_quarterpel;
1778 : syskin 663 Data.rounding = 0;
1779 : h 545
1780 : syskin 658 if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL)
1781 :     return; // allocate some mem for qpel interpolated blocks
1782 :     // somehow this is dirty since I think we shouldn't use malloc outside
1783 :     // encoder_create() - so please fix me!
1784 :     Data.RefQ = qimage;
1785 :    
1786 : chl 530 // note: i==horizontal, j==vertical
1787 :     for (j = 0; j < pParam->mb_height; j++) {
1788 : chl 96
1789 : chl 530 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
1790 : edgomez 195
1791 : chl 530 for (i = 0; i < pParam->mb_width; i++) {
1792 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
1793 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
1794 : edgomez 195
1795 : chl 619 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
1796 : syskin 644 if (b_reference->coding_type != S_VOP)
1797 :     if (b_mb->mode == MODE_NOT_CODED) {
1798 :     pMB->mode = MODE_NOT_CODED;
1799 :     continue;
1800 :     }
1801 : chl 96
1802 : h 545 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
1803 : syskin 601 pMB->quant = frame->quant;
1804 : syskin 628
1805 : chl 530 /* direct search comes first, because it (1) checks for SKIP-mode
1806 :     and (2) sets very good predictions for forward and backward search */
1807 : Isibaar 539 skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
1808 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
1809 : chl 530 &frame->image,
1810 :     i, j,
1811 :     frame->motion_flags,
1812 :     TRB, TRD,
1813 :     pParam,
1814 :     pMB, b_mb,
1815 : h 545 &best_sad,
1816 :     &Data);
1817 : chl 96
1818 : Isibaar 539 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
1819 : syskin 739
1820 : chl 530 // forward search
1821 :     SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y,
1822 :     &frame->image, i, j,
1823 :     frame->motion_flags,
1824 : h 545 frame->fcode, pParam,
1825 : chl 530 pMB, &f_predMV, &best_sad,
1826 : h 545 MODE_FORWARD, &Data);
1827 : chl 96
1828 : chl 530 // backward search
1829 :     SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y,
1830 :     &frame->image, i, j,
1831 :     frame->motion_flags,
1832 : h 545 frame->bcode, pParam,
1833 : chl 530 pMB, &b_predMV, &best_sad,
1834 : h 545 MODE_BACKWARD, &Data);
1835 : chl 96
1836 : chl 530 // interpolate search comes last, because it uses data from forward and backward as prediction
1837 :     SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y,
1838 :     b_ref->y, b_refH->y, b_refV->y, b_refHV->y,
1839 :     &frame->image,
1840 :     i, j,
1841 :     frame->fcode, frame->bcode,
1842 :     frame->motion_flags,
1843 : h 545 pParam,
1844 : chl 530 &f_predMV, &b_predMV,
1845 : h 545 pMB, &best_sad,
1846 :     &Data);
1847 : syskin 662
1848 : syskin 712 // final skip decision
1849 :     if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2)
1850 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
1851 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB,frame->quant, i, j, &Data);
1852 :    
1853 : chl 530 switch (pMB->mode) {
1854 :     case MODE_FORWARD:
1855 :     f_count++;
1856 : syskin 712 if (Data.qpel) f_predMV = pMB->qmvs[0];
1857 : syskin 658 else f_predMV = pMB->mvs[0];
1858 : chl 530 break;
1859 :     case MODE_BACKWARD:
1860 :     b_count++;
1861 : syskin 712 if (Data.qpel) b_predMV = pMB->b_qmvs[0];
1862 : syskin 658 else b_predMV = pMB->b_mvs[0];
1863 : chl 530 break;
1864 :     case MODE_INTERPOLATE:
1865 :     i_count++;
1866 : syskin 712 if (Data.qpel) {
1867 : syskin 658 f_predMV = pMB->qmvs[0];
1868 :     b_predMV = pMB->b_qmvs[0];
1869 : syskin 748 } else {
1870 : syskin 658 f_predMV = pMB->mvs[0];
1871 :     b_predMV = pMB->b_mvs[0];
1872 :     }
1873 : chl 530 break;
1874 :     case MODE_DIRECT:
1875 :     case MODE_DIRECT_NO4V:
1876 :     d_count++;
1877 :     default:
1878 :     break;
1879 : chl 96 }
1880 :     }
1881 :     }
1882 : syskin 658 free(qimage);
1883 : chl 96 }
1884 :    
1885 : syskin 739 static __inline void
1886 : Isibaar 539 MEanalyzeMB ( const uint8_t * const pRef,
1887 :     const uint8_t * const pCur,
1888 :     const int x,
1889 :     const int y,
1890 :     const MBParam * const pParam,
1891 : syskin 739 MACROBLOCK * const pMBs,
1892 : h 545 SearchData * const Data)
1893 : Isibaar 539 {
1894 : chl 317
1895 : syskin 739 int i, mask;
1896 : h 545 VECTOR pmv[3];
1897 : syskin 739 MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width];
1898 :    
1899 :     for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
1900 : suxen_drol 118
1901 : syskin 640 //median is only used as prediction. it doesn't have to be real
1902 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
1903 :     else
1904 :     if (x == 1) //left macroblock does not have any vector now
1905 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
1906 : syskin 748 else if (y == 1) // top macroblock doesn't have it's vector
1907 : syskin 640 Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
1908 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
1909 :    
1910 : h 545 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1911 : syskin 739 pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv);
1912 : chl 326
1913 : h 545 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
1914 :     Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
1915 : syskin 739
1916 : Isibaar 539 pmv[1].x = EVEN(pMB->mvs[0].x);
1917 :     pmv[1].y = EVEN(pMB->mvs[0].y);
1918 : syskin 640 pmv[2].x = EVEN(Data->predMV.x);
1919 :     pmv[2].y = EVEN(Data->predMV.y);
1920 :     pmv[0].x = pmv[0].y = 0;
1921 : chl 326
1922 : syskin 739 CheckCandidate32I(0, 0, 255, &i, Data);
1923 : syskin 640
1924 : syskin 739 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) {
1925 : syskin 640
1926 : syskin 739 if (!(mask = make_mask(pmv, 1)))
1927 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
1928 :     if (!(mask = make_mask(pmv, 2)))
1929 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
1930 : Isibaar 539
1931 : syskin 739 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed
1932 :     DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
1933 : Isibaar 539
1934 : syskin 739 for (i = 0; i < 4; i++) {
1935 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1) * pParam->mb_width)];
1936 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
1937 :     MB->mode = MODE_INTER;
1938 :     MB->sad16 = Data->iMinSAD[i+1];
1939 :     }
1940 :     }
1941 : Isibaar 539 }
1942 :    
1943 : syskin 739 #define INTRA_BIAS 2500
1944 :     #define INTRA_THRESH 1500
1945 :     #define INTER_THRESH 1400
1946 : Isibaar 539
1947 : syskin 644
1948 : Isibaar 539 int
1949 :     MEanalysis( const IMAGE * const pRef,
1950 : syskin 644 FRAMEINFO * const Current,
1951 : Isibaar 539 MBParam * const pParam,
1952 : syskin 644 int maxIntra, //maximum number if non-I frames
1953 :     int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
1954 :     int bCount) // number if B frames in a row
1955 : Isibaar 539 {
1956 :     uint32_t x, y, intra = 0;
1957 :     int sSAD = 0;
1958 : syskin 644 MACROBLOCK * const pMBs = Current->mbs;
1959 :     const IMAGE * const pCurrent = &Current->image;
1960 :     int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH;
1961 : syskin 739 const VECTOR zeroMV = {0,0};
1962 : h 545
1963 : syskin 739 int32_t iMinSAD[5], temp[5];
1964 :     VECTOR currentMV[5];
1965 : h 545 SearchData Data;
1966 :     Data.iEdgedWidth = pParam->edged_width;
1967 : syskin 739 Data.currentMV = currentMV;
1968 :     Data.iMinSAD = iMinSAD;
1969 : syskin 644 Data.iFcode = Current->fcode;
1970 : syskin 739 Data.rrv = Current->global_flags & XVID_REDUCED;
1971 :     Data.temp = temp;
1972 :     CheckCandidate = CheckCandidate32I;
1973 : h 545
1974 : syskin 658 if (intraCount < 10) // we're right after an I frame
1975 :     IntraThresh += 4 * (intraCount - 10) * (intraCount - 10);
1976 : syskin 644 else
1977 :     if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
1978 :     IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
1979 :    
1980 : syskin 672 InterThresh += 400 * (1 - bCount);
1981 : syskin 739 if (InterThresh < 300) InterThresh = 300;
1982 : syskin 644
1983 : Isibaar 539 if (sadInit) (*sadInit) ();
1984 :    
1985 : syskin 739 for (y = 1; y < pParam->mb_height-1; y+=2) {
1986 :     for (x = 1; x < pParam->mb_width-1; x+=2) {
1987 :     int i;
1988 : Isibaar 539
1989 : syskin 739 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
1990 :    
1991 :     MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
1992 : Isibaar 539
1993 : syskin 739 for (i = 0; i < 4; i++) {
1994 :     int dev;
1995 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + y+(i>>1) * pParam->mb_width];
1996 :     if (pMB->sad16 > IntraThresh) {
1997 :     dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1))* pParam->edged_width) * 16,
1998 :     pParam->edged_width);
1999 :     if (dev + IntraThresh < pMB->sad16) {
2000 :     pMB->mode = MODE_INTRA;
2001 :     if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP;
2002 :     }
2003 : Isibaar 539 }
2004 : syskin 739 sSAD += pMB->sad16;
2005 : syskin 644 }
2006 : suxen_drol 118 }
2007 :     }
2008 : syskin 704 sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2);
2009 : syskin 739 if (sSAD > IntraThresh + INTRA_BIAS ) return I_VOP;
2010 : suxen_drol 709 if (sSAD > InterThresh ) return P_VOP;
2011 : Isibaar 539 emms();
2012 : suxen_drol 709 return B_VOP;
2013 : syskin 712
2014 : suxen_drol 118 }
2015 : syskin 576
2016 : syskin 628 static void
2017 : syskin 748 CheckGMC(int x, int y, const int dir, int * iDirection,
2018 : syskin 628 const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC,
2019 :     const MBParam * const pParam)
2020 :     {
2021 :     uint32_t mx, my, a, count = 0;
2022 : chl 619
2023 : syskin 628 for (my = 1; my < pParam->mb_height-1; my++)
2024 :     for (mx = 1; mx < pParam->mb_width-1; mx++) {
2025 :     VECTOR mv;
2026 :     const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width];
2027 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue;
2028 :     mv = pMB->mvs[0];
2029 :     a = ABS(mv.x - x) + ABS(mv.y - y);
2030 :     if (a < 6) count += 6 - a;
2031 :     }
2032 : chl 619
2033 : syskin 628 if (count > *bestcount) {
2034 :     *bestcount = count;
2035 :     *iDirection = dir;
2036 : syskin 748 GMC->x = x; GMC->y = y;
2037 : syskin 628 }
2038 :     }
2039 :    
2040 :    
2041 : suxen_drol 721 static VECTOR
2042 : chl 619 GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode)
2043 :     {
2044 :    
2045 : syskin 628 uint32_t count, bestcount = 0;
2046 : chl 619 int x, y;
2047 : syskin 628 VECTOR gmc = {0,0};
2048 : chl 619 int step, min_x, max_x, min_y, max_y;
2049 :     uint32_t mx, my;
2050 : syskin 628 int iDirection, bDirection;
2051 : chl 619
2052 :     min_x = min_y = -32<<iFcode;
2053 :     max_x = max_y = 32<<iFcode;
2054 :    
2055 : syskin 628 //step1: let's find a rough camera panning
2056 : chl 619 for (step = 32; step >= 2; step /= 2) {
2057 :     bestcount = 0;
2058 :     for (y = min_y; y <= max_y; y += step)
2059 :     for (x = min_x ; x <= max_x; x += step) {
2060 :     count = 0;
2061 :     //for all macroblocks
2062 :     for (my = 1; my < pParam->mb_height-1; my++)
2063 :     for (mx = 1; mx < pParam->mb_width-1; mx++) {
2064 :     const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width];
2065 :     VECTOR mv;
2066 :    
2067 : syskin 748 if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2068 : chl 619 continue;
2069 :    
2070 :     mv = pMB->mvs[0];
2071 :     if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */
2072 :     count++;
2073 :     }
2074 :     if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; }
2075 :     }
2076 :     min_x = gmc.x - step;
2077 :     max_x = gmc.x + step;
2078 :     min_y = gmc.y - step;
2079 :     max_y = gmc.y + step;
2080 :     }
2081 : syskin 628
2082 : syskin 748 if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10)
2083 : syskin 628 gmc.x = gmc.y = 0; //no camara pan, no GMC
2084 :    
2085 : syskin 748 // step2: let's refine camera panning using gradiend-descent approach
2086 : syskin 628 // TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond)
2087 :     bestcount = 0;
2088 :     CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam);
2089 :     do {
2090 :     x = gmc.x; y = gmc.y;
2091 :     bDirection = iDirection; iDirection = 0;
2092 :     if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
2093 :     if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
2094 :     if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam);
2095 :     if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam);
2096 :    
2097 :     } while (iDirection);
2098 :    
2099 : chl 619 if (pParam->m_quarterpel) {
2100 :     gmc.x *= 2;
2101 :     gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */
2102 :     }
2103 :    
2104 :     return gmc;
2105 :     }

No admin address has been configured
ViewVC Help
Powered by ViewVC 1.0.4