[svn] / branches / dev-api-4 / xvidcore / src / motion / motion_est.c Repository:
ViewVC logotype

Annotation of /branches/dev-api-4/xvidcore/src/motion/motion_est.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1053 - (view) (download)

1 : edgomez 851 /**************************************************************************
2 : Isibaar 3 *
3 : edgomez 851 * XVID MPEG-4 VIDEO CODEC
4 :     * motion estimation
5 : chl 259 *
6 : edgomez 851 * 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 : chl 259 *
15 : edgomez 851 * 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 : chl 259 *
20 : edgomez 851 * 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 : chl 259 *
25 : edgomez 851 * 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 : chl 259 *
29 :     *************************************************************************/
30 :    
31 : Isibaar 3 #include <assert.h>
32 :     #include <stdio.h>
33 : chl 96 #include <stdlib.h>
34 : edgomez 1053 #include <string.h> /* memcpy */
35 :     #include <math.h> /* lrint */
36 : Isibaar 3
37 :     #include "../encoder.h"
38 :     #include "../utils/mbfunctions.h"
39 :     #include "../prediction/mbprediction.h"
40 :     #include "../global.h"
41 :     #include "../utils/timer.h"
42 : edgomez 851 #include "../image/interpolate8x8.h"
43 :     #include "motion_est.h"
44 : suxen_drol 118 #include "motion.h"
45 : Isibaar 3 #include "sad.h"
46 : edgomez 851 #include "../utils/emms.h"
47 :     #include "../dct/fdct.h"
48 : Isibaar 3
49 : edgomez 872 /*****************************************************************************
50 :     * Modified rounding tables -- declared in motion.h
51 :     * Original tables see ISO spec tables 7-6 -> 7-9
52 :     ****************************************************************************/
53 :    
54 :     const uint32_t roundtab[16] =
55 :     {0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 };
56 :    
57 :     /* K = 4 */
58 :     const uint32_t roundtab_76[16] =
59 :     { 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 };
60 :    
61 :     /* K = 2 */
62 :     const uint32_t roundtab_78[8] =
63 :     { 0, 0, 1, 1, 0, 0, 0, 1 };
64 :    
65 :     /* K = 1 */
66 :     const uint32_t roundtab_79[4] =
67 :     { 0, 1, 0, 0 };
68 :    
69 : edgomez 851 #define INITIAL_SKIP_THRESH (10)
70 :     #define FINAL_SKIP_THRESH (50)
71 :     #define MAX_SAD00_FOR_SKIP (20)
72 :     #define MAX_CHROMA_SAD_FOR_SKIP (22)
73 : Isibaar 3
74 : edgomez 851 #define CHECK_CANDIDATE(X,Y,D) { \
75 :     CheckCandidate((X),(Y), (D), &iDirection, data ); }
76 : Isibaar 3
77 : edgomez 872 /*****************************************************************************
78 :     * Code
79 :     ****************************************************************************/
80 :    
81 : edgomez 851 static __inline uint32_t
82 :     d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
83 :     {
84 : edgomez 974 int bits;
85 :     const int q = (1 << (iFcode - 1)) - 1;
86 :    
87 :     x <<= qpel;
88 :     y <<= qpel;
89 : edgomez 851 if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
90 : chl 141
91 : edgomez 851 x -= pred.x;
92 : edgomez 974 bits = (x != 0 ? iFcode:0);
93 :     x = abs(x);
94 :     x += q;
95 :     x >>= (iFcode - 1);
96 :     bits += mvtab[x];
97 :    
98 : edgomez 851 y -= pred.y;
99 : edgomez 974 bits += (y != 0 ? iFcode:0);
100 :     y = abs(y);
101 :     y += q;
102 :     y >>= (iFcode - 1);
103 :     bits += mvtab[y];
104 : chl 141
105 : edgomez 974 return bits;
106 : edgomez 851 }
107 : chl 141
108 : edgomez 978 static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by,
109 :     const SearchData * const data)
110 : edgomez 851 {
111 :     int sad;
112 :     const uint32_t stride = data->iEdgedWidth/2;
113 :     uint8_t * f_refu = data->RefQ,
114 :     * f_refv = data->RefQ + 8,
115 :     * b_refu = data->RefQ + 16,
116 :     * b_refv = data->RefQ + 24;
117 : edgomez 978 int offset = (fx>>1) + (fy>>1)*stride;
118 : Isibaar 3
119 : edgomez 851 switch (((fx & 1) << 1) | (fy & 1)) {
120 :     case 0:
121 : edgomez 978 f_refu = (uint8_t*)data->RefP[4] + offset;
122 :     f_refv = (uint8_t*)data->RefP[5] + offset;
123 : edgomez 851 break;
124 :     case 1:
125 : edgomez 978 interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding);
126 :     interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding);
127 : edgomez 851 break;
128 :     case 2:
129 : edgomez 978 interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding);
130 :     interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding);
131 : edgomez 851 break;
132 :     default:
133 : edgomez 978 interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding);
134 :     interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding);
135 : edgomez 851 break;
136 :     }
137 : Isibaar 3
138 : edgomez 978 offset = (bx>>1) + (by>>1)*stride;
139 : edgomez 851 switch (((bx & 1) << 1) | (by & 1)) {
140 :     case 0:
141 : edgomez 978 b_refu = (uint8_t*)data->b_RefP[4] + offset;
142 :     b_refv = (uint8_t*)data->b_RefP[5] + offset;
143 : edgomez 851 break;
144 :     case 1:
145 : edgomez 978 interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
146 :     interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
147 : edgomez 851 break;
148 :     case 2:
149 : edgomez 978 interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
150 :     interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
151 : edgomez 851 break;
152 :     default:
153 : edgomez 978 interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
154 :     interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
155 : edgomez 851 break;
156 :     }
157 : Isibaar 3
158 : edgomez 851 sad = sad8bi(data->CurU, b_refu, f_refu, stride);
159 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
160 : Isibaar 3
161 : edgomez 851 return sad;
162 :     }
163 : Isibaar 3
164 : edgomez 851 static int32_t
165 : edgomez 978 ChromaSAD(const int dx, const int dy, const SearchData * const data)
166 : edgomez 851 {
167 :     int sad;
168 :     const uint32_t stride = data->iEdgedWidth/2;
169 : edgomez 978 int offset = (dx>>1) + (dy>>1)*stride;
170 : Isibaar 3
171 : edgomez 1053 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */
172 :     data->temp[5] = dx; data->temp[6] = dy; /* backup */
173 : Isibaar 3
174 : edgomez 851 switch (((dx & 1) << 1) | (dy & 1)) {
175 :     case 0:
176 : edgomez 978 sad = sad8(data->CurU, data->RefP[4] + offset, stride);
177 :     sad += sad8(data->CurV, data->RefP[5] + offset, stride);
178 : edgomez 851 break;
179 :     case 1:
180 : edgomez 978 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride);
181 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride);
182 : edgomez 851 break;
183 :     case 2:
184 : edgomez 978 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride);
185 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride);
186 : edgomez 851 break;
187 :     default:
188 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding);
189 : edgomez 851 sad = sad8(data->CurU, data->RefQ, stride);
190 :    
191 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding);
192 : edgomez 851 sad += sad8(data->CurV, data->RefQ, stride);
193 :     break;
194 :     }
195 : edgomez 1053 data->temp[7] = sad; /* backup, part 2 */
196 : edgomez 851 return sad;
197 : Isibaar 3 }
198 :    
199 : edgomez 851 static __inline const uint8_t *
200 :     GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
201 : Isibaar 3 {
202 : edgomez 1053 /* dir : 0 = forward, 1 = backward */
203 : edgomez 987 const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP );
204 : edgomez 978 const int picture = ((x&1)<<1) | (y&1);
205 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
206 :     return direction[picture] + offset;
207 : Isibaar 3 }
208 :    
209 : edgomez 1053 /* this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate */
210 : edgomez 851 static __inline const uint8_t *
211 :     GetReference(const int x, const int y, const SearchData * const data)
212 : Isibaar 3 {
213 : edgomez 978 const int picture = ((x&1)<<1) | (y&1);
214 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
215 :     return data->RefP[picture] + offset;
216 : Isibaar 3 }
217 : edgomez 851
218 :     static uint8_t *
219 :     Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
220 : Isibaar 3 {
221 : edgomez 1053 /* create or find a qpel-precision reference picture; return pointer to it */
222 : edgomez 851 uint8_t * Reference = data->RefQ + 16*dir;
223 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
224 :     const uint32_t rounding = data->rounding;
225 :     const int halfpel_x = x/2;
226 :     const int halfpel_y = y/2;
227 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
228 : suxen_drol 136
229 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
230 :     ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
231 :     switch( ((x&1)<<1) + (y&1) ) {
232 : edgomez 1053 case 3: /* x and y in qpel resolution - the "corners" (top left/right and */
233 :     /* bottom left/right) during qpel refinement */
234 : edgomez 963 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
235 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
236 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
237 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
238 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
239 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
240 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
241 : edgomez 851 break;
242 : chl 184
243 : edgomez 1053 case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
244 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
245 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
246 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
247 :     break;
248 : suxen_drol 136
249 : edgomez 1053 case 2: /* x qpel, y halfpel - left or right during qpel refinement */
250 : edgomez 851 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
251 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
252 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
253 :     break;
254 : edgomez 195
255 : edgomez 1053 default: /* pure halfpel position */
256 : edgomez 963 return (uint8_t *) ref1;
257 : edgomez 978
258 : edgomez 851 }
259 :     return Reference;
260 :     }
261 : Isibaar 3
262 : edgomez 851 static uint8_t *
263 :     Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
264 :     {
265 : edgomez 1053 /* create or find a qpel-precision reference picture; return pointer to it */
266 : edgomez 851 uint8_t * Reference = data->RefQ + 16*dir;
267 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
268 :     const uint32_t rounding = data->rounding;
269 :     const int halfpel_x = x/2;
270 :     const int halfpel_y = y/2;
271 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
272 : chl 326
273 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
274 :     switch( ((x&1)<<1) + (y&1) ) {
275 : edgomez 1053 case 3:
276 :     /*
277 :     * x and y in qpel resolution - the "corners" (top left/right and
278 :     * bottom left/right) during qpel refinement
279 :     */
280 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
281 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
282 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
283 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
284 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
285 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
286 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
287 :     break;
288 : chl 184
289 : edgomez 1053 case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
290 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
291 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
292 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
293 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
294 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
295 :     break;
296 : chl 184
297 : edgomez 1053 case 2: /* x qpel, y halfpel - left or right during qpel refinement */
298 : edgomez 851 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
299 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
300 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
301 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
302 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
303 :     break;
304 : edgomez 195
305 : edgomez 1053 default: /* pure halfpel position */
306 : edgomez 851 return (uint8_t *) ref1;
307 :     }
308 :     return Reference;
309 :     }
310 : edgomez 195
311 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS START */
312 : edgomez 195
313 : edgomez 851 static void
314 :     CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
315 :     {
316 :     int xc, yc;
317 :     const uint8_t * Reference;
318 :     VECTOR * current;
319 :     int32_t sad; uint32_t t;
320 : edgomez 195
321 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
322 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
323 : edgomez 195
324 : edgomez 851 if (!data->qpel_precision) {
325 :     Reference = GetReference(x, y, data);
326 :     current = data->currentMV;
327 :     xc = x; yc = y;
328 : edgomez 1053 } else { /* x and y are in 1/4 precision */
329 : edgomez 851 Reference = Interpolate16x16qpel(x, y, 0, data);
330 : edgomez 1053 xc = x/2; yc = y/2; /* for chroma sad */
331 : edgomez 851 current = data->currentQMV;
332 :     }
333 : edgomez 195
334 : edgomez 851 sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
335 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
336 : edgomez 195
337 : edgomez 851 sad += (data->lambda16 * t * sad)>>10;
338 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
339 : chl 175
340 : edgomez 851 if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
341 : edgomez 953 (yc >> 1) + roundtab_79[yc & 0x3], data);
342 : edgomez 195
343 : edgomez 851 if (sad < data->iMinSAD[0]) {
344 :     data->iMinSAD[0] = sad;
345 :     current[0].x = x; current[0].y = y;
346 :     *dir = Direction;
347 :     }
348 : edgomez 195
349 : edgomez 851 if (data->temp[1] < data->iMinSAD[1]) {
350 :     data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
351 :     if (data->temp[2] < data->iMinSAD[2]) {
352 :     data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
353 :     if (data->temp[3] < data->iMinSAD[3]) {
354 :     data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
355 :     if (data->temp[4] < data->iMinSAD[4]) {
356 :     data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
357 : Isibaar 3 }
358 :    
359 : edgomez 851 static void
360 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
361 :     {
362 :     int32_t sad; uint32_t t;
363 :     const uint8_t * Reference;
364 : edgomez 974 VECTOR * current;
365 : chl 326
366 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
367 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
368 : Isibaar 3
369 : edgomez 974 if (!data->qpel_precision) {
370 :     Reference = GetReference(x, y, data);
371 :     current = data->currentMV;
372 : edgomez 1053 } else { /* x and y are in 1/4 precision */
373 : edgomez 974 Reference = Interpolate8x8qpel(x, y, 0, 0, data);
374 :     current = data->currentQMV;
375 :     }
376 : Isibaar 3
377 : edgomez 851 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
378 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
379 : Isibaar 3
380 : edgomez 851 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
381 : Isibaar 3
382 : edgomez 851 if (sad < *(data->iMinSAD)) {
383 :     *(data->iMinSAD) = sad;
384 : edgomez 974 current->x = x; current->y = y;
385 : edgomez 851 *dir = Direction;
386 :     }
387 : Isibaar 3 }
388 :    
389 : edgomez 851 static void
390 :     CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
391 :     {
392 :     uint32_t t;
393 :     const uint8_t * Reference;
394 :    
395 : edgomez 1053 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */
396 : edgomez 851 (x > data->max_dx) || (x < data->min_dx)
397 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
398 :    
399 :     Reference = GetReference(x, y, data);
400 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
401 :    
402 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
403 :    
404 :     data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
405 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
406 :    
407 :     if (data->temp[0] < data->iMinSAD[0]) {
408 :     data->iMinSAD[0] = data->temp[0];
409 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
410 :     *dir = Direction; }
411 :    
412 :     if (data->temp[1] < data->iMinSAD[1]) {
413 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
414 :     if (data->temp[2] < data->iMinSAD[2]) {
415 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
416 :     if (data->temp[3] < data->iMinSAD[3]) {
417 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
418 :     if (data->temp[4] < data->iMinSAD[4]) {
419 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
420 : Isibaar 3 }
421 : edgomez 195
422 : edgomez 851 static void
423 :     CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
424 :     {
425 :     int32_t sad, xc, yc;
426 :     const uint8_t * Reference;
427 :     uint32_t t;
428 :     VECTOR * current;
429 :    
430 : edgomez 963 if ( (x > data->max_dx) || ( x < data->min_dx)
431 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
432 : edgomez 851
433 : edgomez 1053 if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */
434 : edgomez 851
435 : edgomez 1053 if (data->qpel_precision) { /* x and y are in 1/4 precision */
436 : edgomez 851 Reference = Interpolate16x16qpel(x, y, 0, data);
437 :     current = data->currentQMV;
438 :     xc = x/2; yc = y/2;
439 :     } else {
440 :     Reference = GetReference(x, y, data);
441 :     current = data->currentMV;
442 :     xc = x; yc = y;
443 :     }
444 :     t = d_mv_bits(x, y, data->predMV, data->iFcode,
445 :     data->qpel^data->qpel_precision, data->rrv);
446 :    
447 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
448 :     sad += (data->lambda16 * t * sad)>>10;
449 :    
450 :     if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
451 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
452 :    
453 :     if (sad < *(data->iMinSAD)) {
454 :     *(data->iMinSAD) = sad;
455 :     current->x = x; current->y = y;
456 :     *dir = Direction;
457 :     }
458 : chl 96 }
459 : Isibaar 3
460 : edgomez 851 static void
461 :     CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
462 :     {
463 : edgomez 1053 /* maximum speed - for P/B/I decision */
464 : edgomez 851 int32_t sad;
465 :    
466 :     if ( (x > data->max_dx) || (x < data->min_dx)
467 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
468 :    
469 : edgomez 987 sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*(data->iEdgedWidth),
470 :     data->iEdgedWidth, data->temp+1);
471 : edgomez 851
472 :     if (sad < *(data->iMinSAD)) {
473 :     *(data->iMinSAD) = sad;
474 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
475 :     *dir = Direction;
476 :     }
477 :     if (data->temp[1] < data->iMinSAD[1]) {
478 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
479 :     if (data->temp[2] < data->iMinSAD[2]) {
480 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
481 :     if (data->temp[3] < data->iMinSAD[3]) {
482 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
483 :     if (data->temp[4] < data->iMinSAD[4]) {
484 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
485 :    
486 : Isibaar 3 }
487 :    
488 : edgomez 851 static void
489 :     CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
490 :     {
491 :     int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
492 :     uint32_t t;
493 :     const uint8_t *ReferenceF, *ReferenceB;
494 :     VECTOR *current;
495 :    
496 : edgomez 953 if ((xf > data->max_dx) || (xf < data->min_dx) ||
497 :     (yf > data->max_dy) || (yf < data->min_dy))
498 :     return;
499 : edgomez 851
500 :     if (!data->qpel_precision) {
501 :     ReferenceF = GetReference(xf, yf, data);
502 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
503 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
504 :     current = data->currentMV;
505 :     xcf = xf; ycf = yf;
506 :     xcb = xb; ycb = yb;
507 :     } else {
508 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
509 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
510 :     current = data->currentQMV;
511 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
512 :     xcf = xf/2; ycf = yf/2;
513 :     xcb = xb/2; ycb = yb/2;
514 :     }
515 :    
516 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
517 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
518 :    
519 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
520 :     sad += (data->lambda16 * t * sad)>>10;
521 :    
522 :     if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
523 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
524 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
525 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
526 :    
527 :     if (sad < *(data->iMinSAD)) {
528 :     *(data->iMinSAD) = sad;
529 :     current->x = xf; current->y = yf;
530 :     *dir = Direction;
531 :     }
532 : Isibaar 3 }
533 :    
534 : edgomez 851 static void
535 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
536 :     {
537 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
538 :     uint32_t k;
539 :     const uint8_t *ReferenceF;
540 :     const uint8_t *ReferenceB;
541 :     VECTOR mvs, b_mvs;
542 :    
543 : edgomez 953 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
544 : edgomez 851
545 :     for (k = 0; k < 4; k++) {
546 :     mvs.x = data->directmvF[k].x + x;
547 :     b_mvs.x = ((x == 0) ?
548 :     data->directmvB[k].x
549 :     : mvs.x - data->referencemv[k].x);
550 :    
551 :     mvs.y = data->directmvF[k].y + y;
552 :     b_mvs.y = ((y == 0) ?
553 :     data->directmvB[k].y
554 :     : mvs.y - data->referencemv[k].y);
555 :    
556 : edgomez 953 if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) ||
557 :     (mvs.y > data->max_dy) || (mvs.y < data->min_dy) ||
558 :     (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) ||
559 :     (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) )
560 :     return;
561 : edgomez 851
562 :     if (data->qpel) {
563 :     xcf += mvs.x/2; ycf += mvs.y/2;
564 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
565 :     } else {
566 :     xcf += mvs.x; ycf += mvs.y;
567 :     xcb += b_mvs.x; ycb += b_mvs.y;
568 : edgomez 1053 mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */
569 : edgomez 851 b_mvs.x *= 2; b_mvs.y *= 2;
570 :     }
571 :    
572 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
573 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
574 :    
575 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
576 :     ReferenceF, ReferenceB, data->iEdgedWidth);
577 :     if (sad > *(data->iMinSAD)) return;
578 :     }
579 :    
580 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
581 :    
582 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
583 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
584 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
585 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
586 :    
587 :     if (sad < *(data->iMinSAD)) {
588 :     *(data->iMinSAD) = sad;
589 :     data->currentMV->x = x; data->currentMV->y = y;
590 :     *dir = Direction;
591 :     }
592 : Isibaar 3 }
593 :    
594 : edgomez 851 static void
595 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
596 : Isibaar 3 {
597 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
598 :     const uint8_t *ReferenceF;
599 :     const uint8_t *ReferenceB;
600 :     VECTOR mvs, b_mvs;
601 : Isibaar 3
602 : edgomez 963 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
603 : Isibaar 3
604 : edgomez 851 mvs.x = data->directmvF[0].x + x;
605 :     b_mvs.x = ((x == 0) ?
606 :     data->directmvB[0].x
607 :     : mvs.x - data->referencemv[0].x);
608 : Isibaar 3
609 : edgomez 851 mvs.y = data->directmvF[0].y + y;
610 :     b_mvs.y = ((y == 0) ?
611 :     data->directmvB[0].y
612 :     : mvs.y - data->referencemv[0].y);
613 : Isibaar 3
614 : edgomez 963 if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx)
615 :     || (mvs.y > data->max_dy) || (mvs.y < data->min_dy)
616 :     || (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx)
617 :     || (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return;
618 : Isibaar 3
619 : edgomez 851 if (data->qpel) {
620 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
621 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
622 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
623 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
624 :     } else {
625 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
626 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
627 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
628 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
629 :     }
630 :    
631 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
632 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
633 :    
634 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
635 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
636 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
637 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
638 :    
639 :     if (sad < *(data->iMinSAD)) {
640 :     *(data->iMinSAD) = sad;
641 :     data->currentMV->x = x; data->currentMV->y = y;
642 :     *dir = Direction;
643 :     }
644 : Isibaar 3 }
645 :    
646 : edgomez 851
647 :     static void
648 :     CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
649 : Isibaar 3 {
650 :    
651 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
652 : edgomez 1022 int32_t bits = 0;
653 : edgomez 851 VECTOR * current;
654 :     const uint8_t * ptr;
655 :     int i, cbp = 0, t, xc, yc;
656 : edgomez 195
657 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
658 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
659 : edgomez 195
660 : edgomez 851 if (!data->qpel_precision) {
661 :     ptr = GetReference(x, y, data);
662 :     current = data->currentMV;
663 :     xc = x; yc = y;
664 : edgomez 1053 } else { /* x and y are in 1/4 precision */
665 : edgomez 851 ptr = Interpolate16x16qpel(x, y, 0, data);
666 :     current = data->currentQMV;
667 :     xc = x/2; yc = y/2;
668 :     }
669 : Isibaar 3
670 : edgomez 851 for(i = 0; i < 4; i++) {
671 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
672 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
673 : edgomez 1023 bits += data->temp[i] = Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, i);
674 : edgomez 851 }
675 : Isibaar 3
676 : edgomez 1023 bits += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
677 : edgomez 195
678 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
679 : edgomez 851
680 : edgomez 1022 if (bits >= data->iMinSAD[0]) return;
681 : edgomez 851
682 : edgomez 1053 /* chroma */
683 : edgomez 1022 xc = (xc >> 1) + roundtab_79[xc & 0x3];
684 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
685 : edgomez 851
686 : edgomez 1053 /* chroma U */
687 : edgomez 1022 ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
688 :     transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
689 : edgomez 1023 bits += Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 4);
690 : edgomez 1022 if (bits >= data->iMinSAD[0]) return;
691 :    
692 : edgomez 1053 /* chroma V */
693 : edgomez 1022 ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
694 :     transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2);
695 : edgomez 1023 bits += Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 5);
696 : edgomez 1022
697 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
698 : edgomez 851
699 :     if (bits < data->iMinSAD[0]) {
700 :     data->iMinSAD[0] = bits;
701 :     current[0].x = x; current[0].y = y;
702 :     *dir = Direction;
703 :     }
704 :    
705 :     if (data->temp[0] + t < data->iMinSAD[1]) {
706 :     data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
707 :     if (data->temp[1] < data->iMinSAD[2]) {
708 :     data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
709 :     if (data->temp[2] < data->iMinSAD[3]) {
710 :     data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
711 :     if (data->temp[3] < data->iMinSAD[4]) {
712 :     data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
713 :    
714 : Isibaar 3 }
715 : edgomez 851 static void
716 :     CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
717 : chl 96 {
718 :    
719 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
720 : edgomez 1022 int32_t bits;
721 : edgomez 851 VECTOR * current;
722 :     const uint8_t * ptr;
723 : edgomez 1022 int cbp = 0;
724 : edgomez 195
725 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
726 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
727 : edgomez 195
728 : edgomez 851 if (!data->qpel_precision) {
729 :     ptr = GetReference(x, y, data);
730 :     current = data->currentMV;
731 : edgomez 1053 } else { /* x and y are in 1/4 precision */
732 : edgomez 851 ptr = Interpolate8x8qpel(x, y, 0, 0, data);
733 :     current = data->currentQMV;
734 :     }
735 : chl 96
736 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
737 : edgomez 1023 bits = Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 5);
738 :     bits += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
739 : chl 96
740 : edgomez 851 if (bits < data->iMinSAD[0]) {
741 :     data->temp[0] = cbp;
742 :     data->iMinSAD[0] = bits;
743 :     current[0].x = x; current[0].y = y;
744 :     *dir = Direction;
745 :     }
746 :     }
747 : chl 96
748 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
749 : edgomez 195
750 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
751 : chl 96
752 : edgomez 851 static void
753 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
754 :     {
755 : edgomez 195
756 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
757 : edgomez 195
758 : edgomez 851 int iDirection;
759 : edgomez 195
760 : edgomez 1053 for(;;) { /* forever */
761 : edgomez 851 iDirection = 0;
762 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
763 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
764 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
765 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
766 : edgomez 195
767 : edgomez 851 /* now we're doing diagonal checks near our candidate */
768 : edgomez 195
769 : edgomez 1053 if (iDirection) { /* if anything found */
770 : edgomez 851 bDirection = iDirection;
771 :     iDirection = 0;
772 :     x = data->currentMV->x; y = data->currentMV->y;
773 : edgomez 1053 if (bDirection & 3) { /* our candidate is left or right */
774 : edgomez 851 CHECK_CANDIDATE(x, y + iDiamondSize, 8);
775 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
776 : edgomez 1053 } else { /* what remains here is up or down */
777 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y, 2);
778 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
779 :     }
780 : edgomez 195
781 : edgomez 851 if (iDirection) {
782 :     bDirection += iDirection;
783 :     x = data->currentMV->x; y = data->currentMV->y;
784 :     }
785 : edgomez 1053 } else { /* about to quit, eh? not so fast.... */
786 : edgomez 851 switch (bDirection) {
787 :     case 2:
788 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
789 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
790 : edgomez 195 break;
791 : edgomez 851 case 1:
792 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
793 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
794 : edgomez 195 break;
795 : edgomez 851 case 2 + 4:
796 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
797 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
798 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
799 :     break;
800 :     case 4:
801 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
802 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
803 :     break;
804 : edgomez 195 case 8:
805 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
806 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
807 : edgomez 195 break;
808 : edgomez 851 case 1 + 4:
809 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
810 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
811 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
812 : edgomez 195 break;
813 : edgomez 851 case 2 + 8:
814 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
815 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
816 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
817 :     break;
818 :     case 1 + 8:
819 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
820 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
821 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
822 :     break;
823 : edgomez 1053 default: /* 1+2+4+8 == we didn't find anything at all */
824 : edgomez 851 CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
825 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
826 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
827 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
828 :     break;
829 : chl 96 }
830 : edgomez 1053 if (!iDirection) break; /* ok, the end. really */
831 : edgomez 851 bDirection = iDirection;
832 :     x = data->currentMV->x; y = data->currentMV->y;
833 : chl 345 }
834 : edgomez 195 }
835 : chl 96 }
836 :    
837 : edgomez 851 static void
838 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
839 : chl 96 {
840 : edgomez 851 int iDirection;
841 : edgomez 195
842 : edgomez 851 do {
843 :     iDirection = 0;
844 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
845 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
846 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
847 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
848 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
849 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
850 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
851 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
852 : chl 96
853 : edgomez 851 bDirection = iDirection;
854 :     x = data->currentMV->x; y = data->currentMV->y;
855 :     } while (iDirection);
856 : chl 96 }
857 :    
858 : edgomez 851 static void
859 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
860 : chl 181 {
861 :    
862 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
863 :    
864 : edgomez 851 int iDirection;
865 : edgomez 195
866 : edgomez 851 do {
867 :     iDirection = 0;
868 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
869 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
870 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
871 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
872 :    
873 :     /* now we're doing diagonal checks near our candidate */
874 :    
875 : edgomez 1053 if (iDirection) { /* checking if anything found */
876 : edgomez 851 bDirection = iDirection;
877 : chl 181 iDirection = 0;
878 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
879 : edgomez 1053 if (bDirection & 3) { /* our candidate is left or right */
880 : edgomez 851 CHECK_CANDIDATE(x, y + iDiamondSize, 8);
881 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
882 : edgomez 1053 } else { /* what remains here is up or down */
883 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y, 2);
884 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
885 :     }
886 :     bDirection += iDirection;
887 :     x = data->currentMV->x; y = data->currentMV->y;
888 :     }
889 :     }
890 :     while (iDirection);
891 :     }
892 : chl 181
893 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
894 : chl 181
895 : edgomez 851 static void
896 :     SubpelRefine(const SearchData * const data)
897 :     {
898 :     /* Do a half-pel or q-pel refinement */
899 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
900 : edgomez 1053 int iDirection; /* only needed because macro expects it */
901 : chl 181
902 : edgomez 851 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
903 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
904 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
905 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
906 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
907 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
908 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
909 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
910 :     }
911 : chl 181
912 : edgomez 851 static __inline int
913 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
914 :     const int x, const int y,
915 :     const uint32_t stride, const uint32_t iQuant, int rrv)
916 : chl 181
917 : edgomez 851 {
918 : edgomez 963 int offset = (x + y*stride)*8;
919 : edgomez 851 if(!rrv) {
920 : edgomez 963 uint32_t sadC = sad8(current->u + offset,
921 :     reference->u + offset, stride);
922 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
923 : edgomez 963 sadC += sad8(current->v + offset,
924 :     reference->v + offset, stride);
925 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
926 :     return 1;
927 : chl 181
928 : edgomez 851 } else {
929 : edgomez 963 uint32_t sadC = sad16(current->u + 2*offset,
930 :     reference->u + 2*offset, stride, 256*4096);
931 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
932 : edgomez 963 sadC += sad16(current->v + 2*offset,
933 :     reference->v + 2*offset, stride, 256*4096);
934 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
935 :     return 1;
936 : chl 181 }
937 : edgomez 851 }
938 : edgomez 504
939 : edgomez 851 static __inline void
940 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
941 :     {
942 :     pMB->mode = MODE_NOT_CODED;
943 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
944 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
945 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
946 : chl 181 }
947 :    
948 : edgomez 1022 static __inline void
949 :     ModeDecision(SearchData * const Data,
950 :     MACROBLOCK * const pMB,
951 :     const MACROBLOCK * const pMBs,
952 :     const int x, const int y,
953 :     const MBParam * const pParam,
954 :     const uint32_t MotionFlags,
955 :     const uint32_t VopFlags,
956 :     const uint32_t VolFlags,
957 :     const IMAGE * const pCurrent,
958 :     const IMAGE * const pRef)
959 :     {
960 :     int mode = MODE_INTER;
961 :     int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
962 :     const uint32_t iQuant = pMB->quant;
963 :    
964 :     const int skip_possible = (!(VolFlags & XVID_VOL_GMC)) && (pMB->dquant == 0);
965 :    
966 : edgomez 1053 if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) { /* normal, fast, SAD-based mode decision */
967 : edgomez 1022 int sad;
968 :     int InterBias = MV16_INTER_BIAS;
969 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
970 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
971 :     mode = MODE_INTER;
972 :     sad = Data->iMinSAD[0];
973 :     } else {
974 :     mode = MODE_INTER4V;
975 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
976 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
977 :     Data->iMinSAD[0] = sad;
978 :     }
979 :    
980 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
981 :     if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP))
982 :     if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH)
983 :     if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) {
984 :     mode = MODE_NOT_CODED;
985 :     sad = 0;
986 :     }
987 :    
988 :     /* intra decision */
989 :    
990 : edgomez 1053 if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */
991 : edgomez 1022 if (y != 0)
992 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
993 :     if (x != 0)
994 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
995 :    
996 : edgomez 1053 if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? */
997 : edgomez 1022 if (Data->rrv) InterBias *= 4;
998 :    
999 :     if (InterBias < pMB->sad16) {
1000 :     int32_t deviation;
1001 :     if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
1002 :     else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
1003 :     dev16(Data->Cur+16, Data->iEdgedWidth) +
1004 :     dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) +
1005 :     dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth);
1006 :    
1007 :     if (deviation < (sad - InterBias)) mode = MODE_INTRA;
1008 :     }
1009 :    
1010 : edgomez 1053 } else { /* BITS */
1011 : edgomez 1022
1012 :     int bits, intra, i;
1013 :     VECTOR backup[5], *v;
1014 :     Data->iQuant = iQuant;
1015 :    
1016 :     v = Data->qpel ? Data->currentQMV : Data->currentMV;
1017 :     for (i = 0; i < 5; i++) {
1018 :     Data->iMinSAD[i] = 256*4096;
1019 :     backup[i] = v[i];
1020 :     }
1021 :    
1022 :     bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1023 :     if (bits == 0)
1024 : edgomez 1053 mode = MODE_INTER; /* quick stop */
1025 : edgomez 1022 else {
1026 :     if (inter4v) {
1027 :     int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1028 :     if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; }
1029 :     }
1030 :    
1031 :     intra = CountMBBitsIntra(Data);
1032 :    
1033 :     if (intra < bits) { *Data->iMinSAD = bits = intra; mode = MODE_INTRA; }
1034 :     }
1035 :     }
1036 :    
1037 :     if (Data->rrv) {
1038 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1039 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1040 :     }
1041 :    
1042 :     if (mode == MODE_INTER) {
1043 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1044 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1045 :    
1046 :     if(Data->qpel) {
1047 :     pMB->qmvs[0] = pMB->qmvs[1]
1048 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1049 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1050 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1051 :     } else {
1052 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1053 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1054 :     }
1055 :    
1056 :     } else if (mode == MODE_INTER4V)
1057 :     pMB->sad16 = Data->iMinSAD[0];
1058 : edgomez 1053 else /* INTRA, NOT_CODED */
1059 : edgomez 1022 SkipMacroblockP(pMB, 0);
1060 :    
1061 :     pMB->mode = mode;
1062 :     }
1063 :    
1064 : edgomez 851 bool
1065 :     MotionEstimation(MBParam * const pParam,
1066 :     FRAMEINFO * const current,
1067 :     FRAMEINFO * const reference,
1068 :     const IMAGE * const pRefH,
1069 :     const IMAGE * const pRefV,
1070 :     const IMAGE * const pRefHV,
1071 :     const uint32_t iLimit)
1072 : chl 181 {
1073 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
1074 :     const IMAGE *const pCurrent = &current->image;
1075 :     const IMAGE *const pRef = &reference->image;
1076 : chl 181
1077 : edgomez 851 uint32_t mb_width = pParam->mb_width;
1078 :     uint32_t mb_height = pParam->mb_height;
1079 :     const uint32_t iEdgedWidth = pParam->edged_width;
1080 : suxen_drol 890 const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
1081 : chl 181
1082 : edgomez 851 uint32_t x, y;
1083 :     uint32_t iIntra = 0;
1084 : edgomez 953 int32_t quant = current->quant, sad00;
1085 :     int skip_thresh = \
1086 :     INITIAL_SKIP_THRESH * \
1087 :     (current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \
1088 :     (current->vop_flags & XVID_VOP_MODEDECISION_BITS ? 2:1);
1089 : chl 181
1090 : edgomez 1053 /* some pre-initialized thingies for SearchP */
1091 : edgomez 851 int32_t temp[8];
1092 :     VECTOR currentMV[5];
1093 :     VECTOR currentQMV[5];
1094 :     int32_t iMinSAD[5];
1095 : edgomez 962 DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE);
1096 : edgomez 851 SearchData Data;
1097 :     memset(&Data, 0, sizeof(SearchData));
1098 :     Data.iEdgedWidth = iEdgedWidth;
1099 :     Data.currentMV = currentMV;
1100 :     Data.currentQMV = currentQMV;
1101 :     Data.iMinSAD = iMinSAD;
1102 :     Data.temp = temp;
1103 :     Data.iFcode = current->fcode;
1104 :     Data.rounding = pParam->m_rounding_type;
1105 : edgomez 963 Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0);
1106 : edgomez 949 Data.chroma = MotionFlags & XVID_ME_CHROMA16;
1107 : edgomez 963 Data.rrv = (current->vop_flags & XVID_VOP_REDUCED ? 1:0);
1108 : edgomez 962 Data.dctSpace = dct_space;
1109 : edgomez 1022 Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT);
1110 : edgomez 195
1111 : edgomez 949 if ((current->vop_flags & XVID_VOP_REDUCED)) {
1112 : edgomez 851 mb_width = (pParam->width + 31) / 32;
1113 :     mb_height = (pParam->height + 31) / 32;
1114 :     Data.qpel = 0;
1115 :     }
1116 : chl 181
1117 : edgomez 1053 Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */
1118 : edgomez 851 if (sadInit) (*sadInit) ();
1119 : chl 181
1120 : edgomez 851 for (y = 0; y < mb_height; y++) {
1121 :     for (x = 0; x < mb_width; x++) {
1122 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1123 : chl 181
1124 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1125 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1126 :     pRef->y + (x + y * iEdgedWidth) * 16,
1127 :     pParam->edged_width, pMB->sad8 );
1128 : chl 181
1129 : edgomez 851 else pMB->sad16 =
1130 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1131 :     pRef->y + (x + y * iEdgedWidth) * 32,
1132 :     pParam->edged_width, pMB->sad8 );
1133 : chl 181
1134 : edgomez 851 if (Data.chroma) {
1135 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1136 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1137 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1138 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1139 :     pMB->sad16 += Data.temp[7];
1140 :     }
1141 :    
1142 :     sad00 = pMB->sad16;
1143 :    
1144 : edgomez 953 if (pMB->dquant != 0) {
1145 :     quant += DQtab[pMB->dquant];
1146 :     if (quant > 31) quant = 31;
1147 :     else if (quant < 1) quant = 1;
1148 :     }
1149 : edgomez 1022 pMB->quant = quant;
1150 : edgomez 953
1151 : edgomez 1053 /* initial skip decision */
1152 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1153 : edgomez 949 if (!(current->vol_flags & XVID_VOL_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1154 : edgomez 953 if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh)
1155 : edgomez 851 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1156 :     SkipMacroblockP(pMB, sad00);
1157 :     continue;
1158 :     }
1159 :     }
1160 :    
1161 :     SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1162 : edgomez 1022 y, MotionFlags, current->vop_flags, current->vol_flags,
1163 :     &Data, pParam, pMBs, reference->mbs, pMB);
1164 : edgomez 851
1165 : edgomez 1022 ModeDecision(&Data, pMB, pMBs, x, y, pParam,
1166 :     MotionFlags, current->vop_flags, current->vol_flags,
1167 :     pCurrent, pRef);
1168 :    
1169 : edgomez 851 if (pMB->mode == MODE_INTRA)
1170 :     if (++iIntra > iLimit) return 1;
1171 : chl 181 }
1172 :     }
1173 : edgomez 851
1174 : edgomez 949 if (current->vol_flags & XVID_VOL_GMC ) /* GMC only for S(GMC)-VOPs */
1175 : edgomez 987 {
1176 : edgomez 851 current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1177 : edgomez 987 }
1178 : edgomez 851 return 0;
1179 : chl 181 }
1180 :    
1181 :    
1182 : edgomez 851 static __inline int
1183 :     make_mask(const VECTOR * const pmv, const int i)
1184 : chl 96 {
1185 : edgomez 851 int mask = 255, j;
1186 :     for (j = 0; j < i; j++) {
1187 : edgomez 1053 if (MVequal(pmv[i], pmv[j])) return 0; /* same vector has been checked already */
1188 : edgomez 851 if (pmv[i].x == pmv[j].x) {
1189 :     if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1190 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1191 :     } else
1192 :     if (pmv[i].y == pmv[j].y) {
1193 :     if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1194 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1195 :     }
1196 :     }
1197 :     return mask;
1198 :     }
1199 : edgomez 195
1200 : edgomez 851 static __inline void
1201 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1202 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1203 :     {
1204 : edgomez 1053 /* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */
1205 : edgomez 851 if (rrv) { iWcount /= 2; iHcount /= 2; }
1206 : edgomez 195
1207 : edgomez 1053 if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */
1208 : edgomez 851 pmv[5].x = EVEN(pmv[3].x);
1209 :     pmv[5].y = EVEN(pmv[3].y);
1210 :     } else pmv[5].x = pmv[5].y = 0;
1211 : chl 96
1212 : edgomez 1053 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */
1213 : edgomez 851 else pmv[3].x = pmv[3].y = 0;
1214 : chl 96
1215 : edgomez 1053 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */
1216 : edgomez 851 else pmv[4].x = pmv[4].y = 0;
1217 : Isibaar 3
1218 : edgomez 1053 /* [1] median prediction */
1219 : edgomez 851 pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1220 : edgomez 195
1221 : edgomez 1053 pmv[0].x = pmv[0].y = 0; /* [0] is zero; not used in the loop (checked before) but needed here for make_mask */
1222 : edgomez 195
1223 : edgomez 1053 pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */
1224 : edgomez 851 pmv[2].y = EVEN(prevMB->mvs[0].y);
1225 :    
1226 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1227 : edgomez 1053 pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */
1228 : edgomez 851 pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1229 :     } else pmv[6].x = pmv[6].y = 0;
1230 :    
1231 :     if (rrv) {
1232 :     int i;
1233 :     for (i = 0; i < 7; i++) {
1234 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1235 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1236 :     }
1237 :     }
1238 : Isibaar 3 }
1239 :    
1240 : edgomez 851 static void
1241 :     SearchP(const IMAGE * const pRef,
1242 :     const uint8_t * const pRefH,
1243 :     const uint8_t * const pRefV,
1244 :     const uint8_t * const pRefHV,
1245 :     const IMAGE * const pCur,
1246 :     const int x,
1247 :     const int y,
1248 :     const uint32_t MotionFlags,
1249 : suxen_drol 890 const uint32_t VopFlags,
1250 : edgomez 1022 const uint32_t VolFlags,
1251 : edgomez 851 SearchData * const Data,
1252 :     const MBParam * const pParam,
1253 :     const MACROBLOCK * const pMBs,
1254 :     const MACROBLOCK * const prevMBs,
1255 :     MACROBLOCK * const pMB)
1256 :     {
1257 : Isibaar 3
1258 : edgomez 851 int i, iDirection = 255, mask, threshA;
1259 :     VECTOR pmv[7];
1260 : edgomez 1022 int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
1261 : Isibaar 3
1262 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1263 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1264 : edgomez 195
1265 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1266 : Isibaar 3
1267 : edgomez 1053 Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */
1268 : edgomez 851 i = Data->rrv ? 2 : 1;
1269 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1270 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1271 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1272 : chl 326
1273 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1274 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1275 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1276 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1277 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1278 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1279 : edgomez 195
1280 : edgomez 1022 Data->lambda16 = lambda_vec16[pMB->quant];
1281 :     Data->lambda8 = lambda_vec8[pMB->quant];
1282 : edgomez 851 Data->qpel_precision = 0;
1283 : edgomez 195
1284 : edgomez 963 memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1285 : chl 169
1286 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1287 :     else Data->predMV = pmv[0];
1288 : Isibaar 3
1289 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1290 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1291 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1292 :     Data->iMinSAD[2] = pMB->sad8[1];
1293 :     Data->iMinSAD[3] = pMB->sad8[2];
1294 :     Data->iMinSAD[4] = pMB->sad8[3];
1295 : chl 169
1296 : edgomez 949 if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) || (x | y)) {
1297 : edgomez 1053 threshA = Data->temp[0]; /* that's where we keep this SAD atm */
1298 : edgomez 851 if (threshA < 512) threshA = 512;
1299 :     else if (threshA > 1024) threshA = 1024;
1300 :     } else
1301 :     threshA = 512;
1302 : chl 169
1303 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1304 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1305 : chl 169
1306 : edgomez 851 if (!Data->rrv) {
1307 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1308 : edgomez 1053 else CheckCandidate = CheckCandidate16no4v; /* for extra speed */
1309 : edgomez 851 } else CheckCandidate = CheckCandidate32;
1310 : chl 169
1311 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1312 : chl 169
1313 : edgomez 851 for (i = 1; i < 7; i++) {
1314 :     if (!(mask = make_mask(pmv, i)) ) continue;
1315 :     CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1316 :     if (Data->iMinSAD[0] <= threshA) break;
1317 :     }
1318 : chl 169
1319 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1320 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1321 : edgomez 1022 (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16)))
1322 :     inter4v = 0;
1323 : edgomez 851 else {
1324 : Isibaar 3
1325 : edgomez 851 MainSearchFunc * MainSearchPtr;
1326 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1327 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1328 : edgomez 851 else MainSearchPtr = DiamondSearch;
1329 : Isibaar 3
1330 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1331 : Isibaar 3
1332 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1333 :     note that this search is/might be done in halfpel positions,
1334 :     which makes it more different than the diamond above */
1335 : chl 169
1336 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH16) {
1337 : edgomez 851 int32_t bSAD;
1338 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1339 :     if (Data->rrv) {
1340 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1341 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1342 :     }
1343 :     if (!(MVequal(startMV, backupMV))) {
1344 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1345 : edgomez 195
1346 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1347 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1348 :     if (bSAD < Data->iMinSAD[0]) {
1349 :     Data->currentMV[0] = backupMV;
1350 :     Data->iMinSAD[0] = bSAD; }
1351 :     }
1352 : chl 169
1353 : edgomez 851 backupMV = Data->currentMV[0];
1354 :     startMV.x = startMV.y = 1;
1355 :     if (!(MVequal(startMV, backupMV))) {
1356 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1357 : chl 169
1358 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1359 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1360 :     if (bSAD < Data->iMinSAD[0]) {
1361 :     Data->currentMV[0] = backupMV;
1362 :     Data->iMinSAD[0] = bSAD; }
1363 : edgomez 195 }
1364 : edgomez 851 }
1365 :     }
1366 : edgomez 195
1367 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE16)
1368 : edgomez 851 SubpelRefine(Data);
1369 : edgomez 195
1370 : edgomez 851 for(i = 0; i < 5; i++) {
1371 : edgomez 1053 Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */
1372 : edgomez 851 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1373 :     }
1374 : Isibaar 3
1375 : edgomez 1023 if (Data->qpel) {
1376 : edgomez 953 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1377 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1378 : edgomez 987 Data->qpel_precision = 1;
1379 : edgomez 1023 if (MotionFlags & XVID_ME_QUARTERPELREFINE16)
1380 :     SubpelRefine(Data);
1381 : edgomez 953 }
1382 : Isibaar 3
1383 : edgomez 1023 if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30)
1384 : edgomez 1022 inter4v = 0;
1385 : Isibaar 3
1386 : edgomez 987 if (inter4v) {
1387 : edgomez 851 SearchData Data8;
1388 : edgomez 1053 memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */
1389 : Isibaar 3
1390 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1391 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1392 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1393 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1394 : chl 181
1395 : edgomez 949 if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_BITS))) {
1396 : edgomez 1053 /* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */
1397 : edgomez 851 int sumx = 0, sumy = 0;
1398 : Isibaar 3
1399 : edgomez 1022 if (Data->qpel)
1400 :     for (i = 1; i < 5; i++) {
1401 :     sumx += Data->currentQMV[i].x/2;
1402 :     sumy += Data->currentQMV[i].y/2;
1403 :     }
1404 :     else
1405 :     for (i = 1; i < 5; i++) {
1406 :     sumx += Data->currentMV[i].x;
1407 :     sumy += Data->currentMV[i].y;
1408 :     }
1409 : chl 259
1410 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1411 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1412 :     }
1413 : edgomez 1022 } else Data->iMinSAD[1] = 4096*256;
1414 : edgomez 851 }
1415 :    
1416 :     static void
1417 :     Search8(const SearchData * const OldData,
1418 :     const int x, const int y,
1419 :     const uint32_t MotionFlags,
1420 :     const MBParam * const pParam,
1421 :     MACROBLOCK * const pMB,
1422 :     const MACROBLOCK * const pMBs,
1423 :     const int block,
1424 :     SearchData * const Data)
1425 :     {
1426 :     int i = 0;
1427 :     Data->iMinSAD = OldData->iMinSAD + 1 + block;
1428 :     Data->currentMV = OldData->currentMV + 1 + block;
1429 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1430 :    
1431 :     if(Data->qpel) {
1432 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1433 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1434 :     Data->predMV, Data->iFcode, 0, 0);
1435 :     } else {
1436 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1437 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1438 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1439 : Isibaar 3 }
1440 :    
1441 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1442 : Isibaar 3
1443 : edgomez 949 if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) {
1444 : Isibaar 3
1445 : edgomez 978 if (Data->rrv) i = 16; else i = 8;
1446 : Isibaar 3
1447 : edgomez 978 Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1448 :     Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1449 :     Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1450 :     Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1451 :    
1452 :     Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1453 : edgomez 851 Data->qpel_precision = 0;
1454 : Isibaar 3
1455 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1456 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1457 : Isibaar 3
1458 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1459 :     else CheckCandidate = CheckCandidate16no4v;
1460 : Isibaar 3
1461 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_BITS))) {
1462 : edgomez 1053 int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */
1463 : Isibaar 3
1464 : edgomez 851 MainSearchFunc *MainSearchPtr;
1465 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch;
1466 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1467 : edgomez 851 else MainSearchPtr = DiamondSearch;
1468 : Isibaar 3
1469 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1470 : Isibaar 3
1471 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1472 : edgomez 1053 Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1473 : edgomez 851 Data->currentQMV->y = 2 * Data->currentMV->y;
1474 :     }
1475 :     }
1476 : edgomez 195
1477 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8) {
1478 : edgomez 1053 int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */
1479 : edgomez 195
1480 : edgomez 1053 SubpelRefine(Data); /* perform halfpel refine of current best vector */
1481 : Isibaar 3
1482 : edgomez 1053 if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */
1483 :     Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1484 : edgomez 851 Data->currentQMV->y = 2 * Data->currentMV->y;
1485 :     }
1486 :     }
1487 : Isibaar 3
1488 : edgomez 949 if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) {
1489 : edgomez 851 Data->qpel_precision = 1;
1490 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1491 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1492 :     SubpelRefine(Data);
1493 :     }
1494 :     }
1495 : edgomez 195
1496 : edgomez 851 if (Data->rrv) {
1497 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1498 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1499 :     }
1500 :    
1501 :     if(Data->qpel) {
1502 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1503 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1504 :     pMB->qmvs[block] = *Data->currentQMV;
1505 : edgomez 195 } else {
1506 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1507 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1508 : edgomez 78 }
1509 : edgomez 851
1510 :     pMB->mvs[block] = *Data->currentMV;
1511 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1512 : Isibaar 3 }
1513 :    
1514 : edgomez 851 /* motion estimation for B-frames */
1515 : chl 345
1516 : edgomez 851 static __inline VECTOR
1517 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1518 :     {
1519 :     /* the stupidiest function ever */
1520 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1521 :     }
1522 : chl 345
1523 : edgomez 851 static void __inline
1524 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1525 :     const uint32_t iWcount,
1526 :     const MACROBLOCK * const pMB,
1527 :     const uint32_t mode_curr)
1528 : chl 345 {
1529 :    
1530 : edgomez 1053 /* [0] is prediction */
1531 : edgomez 851 pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1532 : chl 345
1533 : edgomez 1053 pmv[1].x = pmv[1].y = 0; /* [1] is zero */
1534 : chl 345
1535 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1536 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1537 : chl 345
1538 : edgomez 1053 if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */
1539 : edgomez 851 pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1540 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1541 :     } else pmv[3].x = pmv[3].y = 0;
1542 : chl 345
1543 : edgomez 851 if (y != 0) {
1544 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1545 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1546 :     } else pmv[4].x = pmv[4].y = 0;
1547 : chl 345
1548 : edgomez 851 if (x != 0) {
1549 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1550 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1551 :     } else pmv[5].x = pmv[5].y = 0;
1552 : chl 345
1553 : edgomez 851 if (x != 0 && y != 0) {
1554 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1555 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1556 :     } else pmv[6].x = pmv[6].y = 0;
1557 :     }
1558 : chl 345
1559 :    
1560 : edgomez 851 /* search backward or forward */
1561 :     static void
1562 :     SearchBF( const IMAGE * const pRef,
1563 :     const uint8_t * const pRefH,
1564 :     const uint8_t * const pRefV,
1565 :     const uint8_t * const pRefHV,
1566 :     const IMAGE * const pCur,
1567 :     const int x, const int y,
1568 :     const uint32_t MotionFlags,
1569 :     const uint32_t iFcode,
1570 :     const MBParam * const pParam,
1571 :     MACROBLOCK * const pMB,
1572 :     const VECTOR * const predMV,
1573 :     int32_t * const best_sad,
1574 :     const int32_t mode_current,
1575 :     SearchData * const Data)
1576 :     {
1577 : chl 345
1578 : edgomez 851 int i, iDirection = 255, mask;
1579 :     VECTOR pmv[7];
1580 :     MainSearchFunc *MainSearchPtr;
1581 :     *Data->iMinSAD = MV_MAX_ERROR;
1582 :     Data->iFcode = iFcode;
1583 :     Data->qpel_precision = 0;
1584 : edgomez 1053 Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */
1585 : chl 345
1586 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16;
1587 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16;
1588 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16;
1589 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16;
1590 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1591 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1592 : chl 345
1593 : edgomez 851 Data->predMV = *predMV;
1594 : chl 345
1595 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1596 :     pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1597 : chl 345
1598 : edgomez 851 pmv[0] = Data->predMV;
1599 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1600 : chl 345
1601 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1602 : chl 345
1603 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1604 :     CheckCandidate = CheckCandidate16no4v;
1605 : chl 345
1606 : edgomez 1053 /* main loop. checking all predictions */
1607 : edgomez 851 for (i = 0; i < 7; i++) {
1608 :     if (!(mask = make_mask(pmv, i)) ) continue;
1609 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1610 : chl 345 }
1611 :    
1612 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1613 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1614 : edgomez 851 else MainSearchPtr = DiamondSearch;
1615 : chl 345
1616 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1617 : chl 345
1618 : edgomez 851 SubpelRefine(Data);
1619 : chl 345
1620 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1621 :     Data->currentQMV->x = 2*Data->currentMV->x;
1622 :     Data->currentQMV->y = 2*Data->currentMV->y;
1623 :     Data->qpel_precision = 1;
1624 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1625 :     pParam->width, pParam->height, iFcode, 1, 0);
1626 :     SubpelRefine(Data);
1627 :     }
1628 : chl 345
1629 : edgomez 1053 /* three bits are needed to code backward mode. four for forward */
1630 : Isibaar 3
1631 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1632 :     else *Data->iMinSAD += 3 * Data->lambda16;
1633 : edgomez 195
1634 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1635 :     *best_sad = *Data->iMinSAD;
1636 :     pMB->mode = mode_current;
1637 :     if (Data->qpel) {
1638 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1639 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1640 :     if (mode_current == MODE_FORWARD)
1641 :     pMB->qmvs[0] = *Data->currentQMV;
1642 :     else
1643 :     pMB->b_qmvs[0] = *Data->currentQMV;
1644 :     } else {
1645 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1646 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1647 :     }
1648 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1649 :     else pMB->b_mvs[0] = *Data->currentMV;
1650 :     }
1651 : edgomez 195
1652 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1653 : edgomez 1053 else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */
1654 : Isibaar 3 }
1655 :    
1656 : edgomez 851 static void
1657 :     SkipDecisionB(const IMAGE * const pCur,
1658 :     const IMAGE * const f_Ref,
1659 :     const IMAGE * const b_Ref,
1660 :     MACROBLOCK * const pMB,
1661 :     const uint32_t x, const uint32_t y,
1662 :     const SearchData * const Data)
1663 :     {
1664 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1665 :     int32_t sum;
1666 :     const int div = 1 + Data->qpel;
1667 :     int k;
1668 :     const uint32_t stride = Data->iEdgedWidth/2;
1669 : edgomez 1053 /* this is not full chroma compensation, only it's fullpel approximation. should work though */
1670 : Isibaar 3
1671 : edgomez 851 for (k = 0; k < 4; k++) {
1672 :     dy += Data->directmvF[k].y / div;
1673 : edgomez 963 dx += Data->directmvF[k].x / div;
1674 :     b_dy += Data->directmvB[k].y / div;
1675 :     b_dx += Data->directmvB[k].x / div;
1676 : edgomez 851 }
1677 : Isibaar 3
1678 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1679 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1680 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1681 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1682 : Isibaar 3
1683 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1684 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1685 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1686 :     stride);
1687 : Isibaar 3
1688 : edgomez 1053 if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; /* no skip */
1689 : Isibaar 3
1690 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1691 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1692 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1693 :     stride);
1694 : edgomez 195
1695 : edgomez 963 if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1696 : edgomez 1053 pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */
1697 : edgomez 978 for (k = 0; k < 4; k++) {
1698 : edgomez 963 pMB->qmvs[k] = pMB->mvs[k];
1699 :     pMB->b_qmvs[k] = pMB->b_mvs[k];
1700 :     }
1701 :     }
1702 : edgomez 851 }
1703 : Isibaar 3
1704 : edgomez 851 static __inline uint32_t
1705 :     SearchDirect(const IMAGE * const f_Ref,
1706 :     const uint8_t * const f_RefH,
1707 :     const uint8_t * const f_RefV,
1708 :     const uint8_t * const f_RefHV,
1709 :     const IMAGE * const b_Ref,
1710 :     const uint8_t * const b_RefH,
1711 :     const uint8_t * const b_RefV,
1712 :     const uint8_t * const b_RefHV,
1713 :     const IMAGE * const pCur,
1714 :     const int x, const int y,
1715 :     const uint32_t MotionFlags,
1716 :     const int32_t TRB, const int32_t TRD,
1717 :     const MBParam * const pParam,
1718 :     MACROBLOCK * const pMB,
1719 :     const MACROBLOCK * const b_mb,
1720 :     int32_t * const best_sad,
1721 :     SearchData * const Data)
1722 : Isibaar 3
1723 : edgomez 851 {
1724 :     int32_t skip_sad;
1725 :     int k = (x + Data->iEdgedWidth*y) * 16;
1726 :     MainSearchFunc *MainSearchPtr;
1727 : Isibaar 3
1728 : edgomez 851 *Data->iMinSAD = 256*4096;
1729 : edgomez 978 Data->RefP[0] = f_Ref->y + k;
1730 :     Data->RefP[2] = f_RefH + k;
1731 :     Data->RefP[1] = f_RefV + k;
1732 :     Data->RefP[3] = f_RefHV + k;
1733 :     Data->b_RefP[0] = b_Ref->y + k;
1734 :     Data->b_RefP[2] = b_RefH + k;
1735 :     Data->b_RefP[1] = b_RefV + k;
1736 :     Data->b_RefP[3] = b_RefHV + k;
1737 :     Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1738 :     Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1739 :     Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1740 :     Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1741 : edgomez 195
1742 : edgomez 851 k = Data->qpel ? 4 : 2;
1743 :     Data->max_dx = k * (pParam->width - x * 16);
1744 :     Data->max_dy = k * (pParam->height - y * 16);
1745 :     Data->min_dx = -k * (16 + x * 16);
1746 :     Data->min_dy = -k * (16 + y * 16);
1747 : chl 181
1748 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1749 :     Data->qpel_precision = 0;
1750 : edgomez 170
1751 : edgomez 851 for (k = 0; k < 4; k++) {
1752 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1753 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1754 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1755 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1756 : Isibaar 3
1757 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1758 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1759 :    
1760 : edgomez 1053 *best_sad = 256*4096; /* in that case, we won't use direct mode */
1761 :     pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */
1762 : edgomez 851 pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1763 :     return 256*4096;
1764 :     }
1765 :     if (b_mb->mode != MODE_INTER4V) {
1766 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1767 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1768 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1769 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1770 :     break;
1771 :     }
1772 : edgomez 195 }
1773 : Isibaar 3
1774 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1775 : Isibaar 3
1776 : edgomez 851 CheckCandidate(0, 0, 255, &k, Data);
1777 : edgomez 195
1778 : edgomez 1053 /* initial (fast) skip decision */
1779 : edgomez 1023 if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) {
1780 : edgomez 1053 /* possible skip */
1781 : edgomez 851 if (Data->chroma) {
1782 :     pMB->mode = MODE_DIRECT_NONE_MV;
1783 : edgomez 1053 return *Data->iMinSAD; /* skip. */
1784 : edgomez 851 } else {
1785 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1786 : edgomez 1053 if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */
1787 : edgomez 851 }
1788 : Isibaar 3 }
1789 :    
1790 : edgomez 953 *Data->iMinSAD += Data->lambda16;
1791 : edgomez 851 skip_sad = *Data->iMinSAD;
1792 : edgomez 1053
1793 :     /*
1794 :     * DIRECT MODE DELTA VECTOR SEARCH.
1795 :     * This has to be made more effective, but at the moment I'm happy it's running at all
1796 :     */
1797 : edgomez 195
1798 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1799 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1800 : edgomez 851 else MainSearchPtr = DiamondSearch;
1801 : Isibaar 3
1802 : edgomez 851 MainSearchPtr(0, 0, Data, 255);
1803 : Isibaar 3
1804 : edgomez 851 SubpelRefine(Data);
1805 : chl 181
1806 : edgomez 851 *best_sad = *Data->iMinSAD;
1807 : chl 181
1808 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1809 : edgomez 1053 else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */
1810 : chl 181
1811 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1812 : edgomez 195
1813 : edgomez 851 for (k = 0; k < 4; k++) {
1814 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1815 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1816 :     ? Data->directmvB[k].x
1817 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1818 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1819 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1820 :     ? Data->directmvB[k].y
1821 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1822 :     if (Data->qpel) {
1823 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1824 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1825 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1826 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1827 :     }
1828 : edgomez 195
1829 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1830 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1831 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1832 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1833 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1834 :     break;
1835 :     }
1836 : edgomez 78 }
1837 : edgomez 851 return skip_sad;
1838 :     }
1839 : Isibaar 3
1840 : edgomez 851 static void
1841 :     SearchInterpolate(const IMAGE * const f_Ref,
1842 :     const uint8_t * const f_RefH,
1843 :     const uint8_t * const f_RefV,
1844 :     const uint8_t * const f_RefHV,
1845 :     const IMAGE * const b_Ref,
1846 :     const uint8_t * const b_RefH,
1847 :     const uint8_t * const b_RefV,
1848 :     const uint8_t * const b_RefHV,
1849 :     const IMAGE * const pCur,
1850 :     const int x, const int y,
1851 :     const uint32_t fcode,
1852 :     const uint32_t bcode,
1853 :     const uint32_t MotionFlags,
1854 :     const MBParam * const pParam,
1855 :     const VECTOR * const f_predMV,
1856 :     const VECTOR * const b_predMV,
1857 :     MACROBLOCK * const pMB,
1858 :     int32_t * const best_sad,
1859 :     SearchData * const fData)
1860 : chl 96
1861 : edgomez 851 {
1862 : chl 169
1863 : edgomez 851 int iDirection, i, j;
1864 :     SearchData bData;
1865 : chl 169
1866 : edgomez 851 fData->qpel_precision = 0;
1867 : edgomez 1053 memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */
1868 : edgomez 851 *fData->iMinSAD = 4096*256;
1869 :     bData.currentMV++; bData.currentQMV++;
1870 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1871 : chl 169
1872 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1873 : chl 169
1874 : edgomez 978 bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i;
1875 :     bData.b_RefP[2] = fData->RefP[2] = f_RefH + i;
1876 :     bData.b_RefP[1] = fData->RefP[1] = f_RefV + i;
1877 :     bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i;
1878 :     bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i;
1879 :     bData.RefP[2] = fData->b_RefP[2] = b_RefH + i;
1880 :     bData.RefP[1] = fData->b_RefP[1] = b_RefV + i;
1881 :     bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i;
1882 :     bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1883 :     bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1884 :     bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1885 :     bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1886 : chl 169
1887 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1888 :     fData->bpredMV = bData.predMV = *b_predMV;
1889 :     fData->currentMV[0] = fData->currentMV[2];
1890 : Isibaar 3
1891 : edgomez 851 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);
1892 :     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);
1893 : Isibaar 3
1894 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1895 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1896 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1897 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1898 : chl 169
1899 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1900 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1901 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1902 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1903 : Isibaar 3
1904 : edgomez 851 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1905 : chl 169
1906 : edgomez 1053 /* diamond */
1907 : edgomez 851 do {
1908 :     iDirection = 255;
1909 : edgomez 1053 /* forward MV moves */
1910 : edgomez 851 i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1911 : chl 169
1912 : edgomez 851 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1913 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1914 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1915 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1916 : edgomez 195
1917 : edgomez 1053 /* backward MV moves */
1918 : edgomez 851 i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1919 :     fData->currentMV[2] = fData->currentMV[0];
1920 :     CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1921 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1922 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1923 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1924 : Isibaar 3
1925 : edgomez 851 } while (!(iDirection));
1926 : chl 140
1927 : edgomez 1053 /* qpel refinement */
1928 : edgomez 851 if (fData->qpel) {
1929 :     if (*fData->iMinSAD > *best_sad + 500) return;
1930 :     CheckCandidate = CheckCandidateInt;
1931 :     fData->qpel_precision = bData.qpel_precision = 1;
1932 :     get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1933 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1934 :     fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1935 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1936 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1937 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1938 :     SubpelRefine(fData);
1939 :     if (*fData->iMinSAD > *best_sad + 300) return;
1940 :     fData->currentQMV[2] = fData->currentQMV[0];
1941 :     SubpelRefine(&bData);
1942 :     }
1943 : chl 140
1944 : edgomez 1053 *fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */
1945 : Isibaar 3
1946 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
1947 :     *best_sad = *fData->iMinSAD;
1948 :     pMB->mvs[0] = fData->currentMV[0];
1949 :     pMB->b_mvs[0] = fData->currentMV[1];
1950 :     pMB->mode = MODE_INTERPOLATE;
1951 :     if (fData->qpel) {
1952 :     pMB->qmvs[0] = fData->currentQMV[0];
1953 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1954 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1955 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1956 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1957 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1958 :     } else {
1959 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1960 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1961 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1962 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1963 :     }
1964 : edgomez 78 }
1965 : edgomez 851 }
1966 : Isibaar 3
1967 : edgomez 851 void
1968 :     MotionEstimationBVOP(MBParam * const pParam,
1969 :     FRAMEINFO * const frame,
1970 :     const int32_t time_bp,
1971 :     const int32_t time_pp,
1972 : edgomez 1053 /* forward (past) reference */
1973 : edgomez 851 const MACROBLOCK * const f_mbs,
1974 :     const IMAGE * const f_ref,
1975 :     const IMAGE * const f_refH,
1976 :     const IMAGE * const f_refV,
1977 :     const IMAGE * const f_refHV,
1978 : edgomez 1053 /* backward (future) reference */
1979 : edgomez 851 const FRAMEINFO * const b_reference,
1980 :     const IMAGE * const b_ref,
1981 :     const IMAGE * const b_refH,
1982 :     const IMAGE * const b_refV,
1983 :     const IMAGE * const b_refHV)
1984 :     {
1985 :     uint32_t i, j;
1986 :     int32_t best_sad;
1987 :     uint32_t skip_sad;
1988 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
1989 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
1990 : Isibaar 3
1991 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
1992 : Isibaar 3
1993 : edgomez 851 const int32_t TRB = time_pp - time_bp;
1994 :     const int32_t TRD = time_pp;
1995 : edgomez 195
1996 : edgomez 1053 /* some pre-inintialized data for the rest of the search */
1997 : Isibaar 3
1998 : edgomez 851 SearchData Data;
1999 :     int32_t iMinSAD;
2000 :     VECTOR currentMV[3];
2001 :     VECTOR currentQMV[3];
2002 :     int32_t temp[8];
2003 :     memset(&Data, 0, sizeof(SearchData));
2004 :     Data.iEdgedWidth = pParam->edged_width;
2005 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2006 :     Data.iMinSAD = &iMinSAD;
2007 :     Data.lambda16 = lambda_vec16[frame->quant];
2008 : edgomez 949 Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL;
2009 : edgomez 851 Data.rounding = 0;
2010 : edgomez 949 Data.chroma = frame->motion_flags & XVID_ME_CHROMA8;
2011 : edgomez 851 Data.temp = temp;
2012 : Isibaar 3
2013 : edgomez 1053 Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */
2014 :    
2015 :     /* note: i==horizontal, j==vertical */
2016 : edgomez 851 for (j = 0; j < pParam->mb_height; j++) {
2017 : edgomez 195
2018 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2019 : Isibaar 3
2020 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2021 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2022 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2023 : edgomez 195
2024 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2025 :     if (b_reference->coding_type != S_VOP)
2026 :     if (b_mb->mode == MODE_NOT_CODED) {
2027 :     pMB->mode = MODE_NOT_CODED;
2028 :     continue;
2029 :     }
2030 : Isibaar 3
2031 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2032 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2033 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2034 : edgomez 953 pMB->quant = frame->quant;
2035 : Isibaar 3
2036 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2037 :     and (2) sets very good predictions for forward and backward search */
2038 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2039 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2040 :     &frame->image,
2041 :     i, j,
2042 :     frame->motion_flags,
2043 :     TRB, TRD,
2044 :     pParam,
2045 :     pMB, b_mb,
2046 :     &best_sad,
2047 :     &Data);
2048 : Isibaar 3
2049 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2050 : edgomez 195
2051 : edgomez 1053 /* forward search */
2052 : edgomez 851 SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2053 :     &frame->image, i, j,
2054 :     frame->motion_flags,
2055 :     frame->fcode, pParam,
2056 :     pMB, &f_predMV, &best_sad,
2057 :     MODE_FORWARD, &Data);
2058 : edgomez 195
2059 : edgomez 1053 /* backward search */
2060 : edgomez 851 SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2061 :     &frame->image, i, j,
2062 :     frame->motion_flags,
2063 :     frame->bcode, pParam,
2064 :     pMB, &b_predMV, &best_sad,
2065 :     MODE_BACKWARD, &Data);
2066 :    
2067 : edgomez 1053 /* interpolate search comes last, because it uses data from forward and backward as prediction */
2068 : edgomez 851 SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2069 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2070 :     &frame->image,
2071 :     i, j,
2072 :     frame->fcode, frame->bcode,
2073 :     frame->motion_flags,
2074 :     pParam,
2075 :     &f_predMV, &b_predMV,
2076 :     pMB, &best_sad,
2077 :     &Data);
2078 :    
2079 : edgomez 1053 /* final skip decision */
2080 : edgomez 851 if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2081 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2082 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2083 :    
2084 :     switch (pMB->mode) {
2085 :     case MODE_FORWARD:
2086 :     f_count++;
2087 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2088 :     break;
2089 :     case MODE_BACKWARD:
2090 :     b_count++;
2091 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2092 :     break;
2093 :     case MODE_INTERPOLATE:
2094 :     i_count++;
2095 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2096 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2097 :     break;
2098 :     case MODE_DIRECT:
2099 :     case MODE_DIRECT_NO4V:
2100 :     d_count++;
2101 :     default:
2102 :     break;
2103 :     }
2104 :     }
2105 :     }
2106 : Isibaar 3 }
2107 : chl 96
2108 : edgomez 851 static __inline void
2109 :     MEanalyzeMB ( const uint8_t * const pRef,
2110 :     const uint8_t * const pCur,
2111 :     const int x,
2112 :     const int y,
2113 :     const MBParam * const pParam,
2114 :     MACROBLOCK * const pMBs,
2115 :     SearchData * const Data)
2116 : chl 96 {
2117 :    
2118 : edgomez 851 int i, mask;
2119 : edgomez 953 int quarterpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0;
2120 : edgomez 851 VECTOR pmv[3];
2121 : edgomez 953 MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2122 : chl 96
2123 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2124 : chl 96
2125 : edgomez 1053 /* median is only used as prediction. it doesn't have to be real */
2126 : edgomez 851 if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2127 :     else
2128 : edgomez 1053 if (x == 1) /* left macroblock does not have any vector now */
2129 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */
2130 :     else if (y == 1) /* top macroblock doesn't have it's vector */
2131 :     Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */
2132 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */
2133 : edgomez 195
2134 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2135 : edgomez 953 pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0);
2136 : edgomez 195
2137 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2138 : edgomez 978 Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16;
2139 : chl 96
2140 : edgomez 851 pmv[1].x = EVEN(pMB->mvs[0].x);
2141 :     pmv[1].y = EVEN(pMB->mvs[0].y);
2142 :     pmv[2].x = EVEN(Data->predMV.x);
2143 :     pmv[2].y = EVEN(Data->predMV.y);
2144 :     pmv[0].x = pmv[0].y = 0;
2145 : edgomez 195
2146 : edgomez 851 CheckCandidate32I(0, 0, 255, &i, Data);
2147 : edgomez 195
2148 : edgomez 953 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) {
2149 : chl 96
2150 : edgomez 851 if (!(mask = make_mask(pmv, 1)))
2151 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2152 :     if (!(mask = make_mask(pmv, 2)))
2153 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2154 : chl 96
2155 : edgomez 1053 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) /* diamond only if needed */
2156 : edgomez 851 DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2157 : edgomez 953 }
2158 : edgomez 851
2159 : edgomez 953 for (i = 0; i < 4; i++) {
2160 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2161 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2162 :     MB->mode = MODE_INTER;
2163 :     MB->sad16 = Data->iMinSAD[i+1];
2164 : chl 96 }
2165 : edgomez 851 }
2166 : chl 96
2167 : edgomez 1023 #define INTRA_THRESH 1700
2168 : edgomez 987 #define INTER_THRESH 1200
2169 : chl 96
2170 : edgomez 851 int
2171 :     MEanalysis( const IMAGE * const pRef,
2172 : edgomez 953 const FRAMEINFO * const Current,
2173 :     const MBParam * const pParam,
2174 : edgomez 1053 const int maxIntra, /* maximum number if non-I frames */
2175 :     const int intraCount, /* number of non-I frames after last I frame; 0 if we force P/B frame */
2176 :     const int bCount, /* number of B frames in a row */
2177 : edgomez 953 const int b_thresh)
2178 : edgomez 851 {
2179 :     uint32_t x, y, intra = 0;
2180 :     int sSAD = 0;
2181 :     MACROBLOCK * const pMBs = Current->mbs;
2182 :     const IMAGE * const pCurrent = &Current->image;
2183 : edgomez 953 int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh;
2184 :     int s = 0, blocks = 0;
2185 : chl 96
2186 : edgomez 851 int32_t iMinSAD[5], temp[5];
2187 :     VECTOR currentMV[5];
2188 :     SearchData Data;
2189 :     Data.iEdgedWidth = pParam->edged_width;
2190 :     Data.currentMV = currentMV;
2191 :     Data.iMinSAD = iMinSAD;
2192 :     Data.iFcode = Current->fcode;
2193 :     Data.temp = temp;
2194 :     CheckCandidate = CheckCandidate32I;
2195 : chl 96
2196 : edgomez 1023 if (intraCount != 0) {
2197 : edgomez 1053 if (intraCount < 10) /* we're right after an I frame */
2198 : edgomez 1023 IntraThresh += 15* (intraCount - 10) * (intraCount - 10);
2199 :     else
2200 : edgomez 1053 if ( 5*(maxIntra - intraCount) < maxIntra) /* we're close to maximum. 2 sec when max is 10 sec */
2201 : edgomez 1023 IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra;
2202 :     }
2203 : chl 96
2204 : edgomez 953 InterThresh -= (350 - 8*b_thresh) * bCount;
2205 :     if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh;
2206 : chl 326
2207 : edgomez 851 if (sadInit) (*sadInit) ();
2208 : chl 96
2209 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2210 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2211 :     int i;
2212 : edgomez 953 blocks += 4;
2213 : edgomez 195
2214 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2215 : edgomez 1053 else { /* extrapolation of the vector found for last frame */
2216 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].x =
2217 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2218 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].y =
2219 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2220 :     }
2221 : edgomez 195
2222 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2223 : edgomez 195
2224 : edgomez 851 for (i = 0; i < 4; i++) {
2225 :     int dev;
2226 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2227 :     if (pMB->sad16 > IntraThresh) {
2228 :     dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2229 :     pParam->edged_width);
2230 :     if (dev + IntraThresh < pMB->sad16) {
2231 :     pMB->mode = MODE_INTRA;
2232 : edgomez 953 if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2233 : edgomez 851 }
2234 :     }
2235 : edgomez 953 if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++;
2236 :    
2237 : edgomez 851 sSAD += pMB->sad16;
2238 :     }
2239 :     }
2240 : edgomez 195 }
2241 : edgomez 953
2242 :     sSAD /= blocks;
2243 :    
2244 : edgomez 1023 if (b_thresh < 20) {
2245 :     s = (10*s) / blocks;
2246 : edgomez 1053 if (s > 4) sSAD += (s - 2) * (40 - 2*b_thresh); /* static block - looks bad when in bframe... */
2247 : edgomez 1023 }
2248 : edgomez 953
2249 : edgomez 851 if (sSAD > InterThresh ) return P_VOP;
2250 :     emms();
2251 :     return B_VOP;
2252 :     }
2253 : chl 96
2254 :    
2255 : edgomez 851 static WARPPOINTS
2256 :     GlobalMotionEst(const MACROBLOCK * const pMBs,
2257 :     const MBParam * const pParam,
2258 :     const FRAMEINFO * const current,
2259 :     const FRAMEINFO * const reference,
2260 :     const IMAGE * const pRefH,
2261 :     const IMAGE * const pRefV,
2262 :     const IMAGE * const pRefHV )
2263 :     {
2264 : chl 96
2265 : edgomez 1053 const int deltax=8; /* upper bound for difference between a MV and it's neighbour MVs */
2266 : edgomez 851 const int deltay=8;
2267 : edgomez 1053 const int grad=512; /* lower bound for deviation in MB */
2268 : chl 96
2269 : edgomez 851 WARPPOINTS gmc;
2270 : chl 96
2271 : edgomez 851 uint32_t mx, my;
2272 : chl 96
2273 : edgomez 851 int MBh = pParam->mb_height;
2274 :     int MBw = pParam->mb_width;
2275 : chl 96
2276 : edgomez 851 int *MBmask= calloc(MBh*MBw,sizeof(int));
2277 :     double DtimesF[4] = { 0.,0., 0., 0. };
2278 :     double sol[4] = { 0., 0., 0., 0. };
2279 :     double a,b,c,n,denom;
2280 :     double meanx,meany;
2281 :     int num,oldnum;
2282 : edgomez 195
2283 : edgomez 978 if (!MBmask) { fprintf(stderr,"Mem error\n");
2284 :     gmc.duv[0].x= gmc.duv[0].y =
2285 :     gmc.duv[1].x= gmc.duv[1].y =
2286 : chl 871 gmc.duv[2].x= gmc.duv[2].y = 0;
2287 :     return gmc; }
2288 : edgomez 195
2289 : edgomez 1053 /* filter mask of all blocks */
2290 : edgomez 851
2291 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2292 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2293 : edgomez 851 {
2294 :     const int mbnum = mx + my * MBw;
2295 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2296 :     const VECTOR mv = pMB->mvs[0];
2297 :    
2298 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2299 :     continue;
2300 :    
2301 : edgomez 982 if ( ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2302 :     && ( (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2303 :     && ( (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2304 :     && ( (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2305 : edgomez 851 MBmask[mbnum]=1;
2306 : chl 96 }
2307 :    
2308 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2309 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2310 : edgomez 851 {
2311 :     const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx;
2312 : chl 96
2313 : edgomez 851 const int mbnum = mx + my * MBw;
2314 :     if (!MBmask[mbnum])
2315 :     continue;
2316 :    
2317 : edgomez 876 if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad )
2318 : edgomez 851 MBmask[mbnum] = 0;
2319 : edgomez 876 if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad )
2320 : edgomez 851 MBmask[mbnum] = 0;
2321 :    
2322 : edgomez 195 }
2323 : chl 96
2324 : edgomez 851 emms();
2325 : chl 96
2326 : edgomez 851 do { /* until convergence */
2327 : chl 96
2328 : edgomez 851 a = b = c = n = 0;
2329 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2330 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2331 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2332 : edgomez 851 {
2333 :     const int mbnum = mx + my * MBw;
2334 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2335 :     const VECTOR mv = pMB->mvs[0];
2336 : chl 96
2337 : edgomez 851 if (!MBmask[mbnum])
2338 :     continue;
2339 : chl 96
2340 : edgomez 851 n++;
2341 :     a += 16*mx+8;
2342 :     b += 16*my+8;
2343 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2344 : chl 96
2345 : edgomez 851 DtimesF[0] += (double)mv.x;
2346 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2347 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2348 :     DtimesF[3] += (double)mv.y;
2349 :     }
2350 : chl 96
2351 : edgomez 851 denom = a*a+b*b-c*n;
2352 : chl 96
2353 : edgomez 978 /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2354 : edgomez 851 /* D'*E*F has been calculated in the same loop as matrix */
2355 : chl 96
2356 : edgomez 851 sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2357 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2358 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2359 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2360 : chl 96
2361 : edgomez 851 sol[0] /= denom;
2362 :     sol[1] /= denom;
2363 :     sol[2] /= denom;
2364 :     sol[3] /= denom;
2365 : chl 96
2366 : edgomez 851 meanx = meany = 0.;
2367 :     oldnum = 0;
2368 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2369 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2370 : edgomez 851 {
2371 :     const int mbnum = mx + my * MBw;
2372 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2373 :     const VECTOR mv = pMB->mvs[0];
2374 : chl 96
2375 : edgomez 851 if (!MBmask[mbnum])
2376 :     continue;
2377 : chl 184
2378 : edgomez 851 oldnum++;
2379 : edgomez 982 meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x );
2380 :     meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y );
2381 : edgomez 851 }
2382 : chl 96
2383 : edgomez 851 if (4*meanx > oldnum) /* better fit than 0.25 is useless */
2384 :     meanx /= oldnum;
2385 :     else
2386 :     meanx = 0.25;
2387 : edgomez 195
2388 : edgomez 851 if (4*meany > oldnum)
2389 :     meany /= oldnum;
2390 :     else
2391 :     meany = 0.25;
2392 : chl 96
2393 : edgomez 851 /* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2394 :     fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2395 :     */
2396 :     num = 0;
2397 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2398 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2399 : edgomez 851 {
2400 :     const int mbnum = mx + my * MBw;
2401 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2402 :     const VECTOR mv = pMB->mvs[0];
2403 : chl 96
2404 : edgomez 851 if (!MBmask[mbnum])
2405 :     continue;
2406 : chl 96
2407 : edgomez 982 if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx )
2408 : edgomez 987 || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) )
2409 : edgomez 851 MBmask[mbnum]=0;
2410 :     else
2411 :     num++;
2412 : chl 96 }
2413 : edgomez 195
2414 : edgomez 851 } while ( (oldnum != num) && (num>=4) );
2415 : edgomez 195
2416 : edgomez 851 if (num < 4)
2417 :     {
2418 :     gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0;
2419 :     } else {
2420 : edgomez 195
2421 : edgomez 851 gmc.duv[0].x=(int)(sol[0]+0.5);
2422 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2423 : chl 96
2424 : edgomez 851 gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2425 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2426 : chl 96
2427 : edgomez 851 gmc.duv[2].x=0;
2428 :     gmc.duv[2].y=0;
2429 :     }
2430 : edgomez 1053 /* fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y); */
2431 : chl 96
2432 : edgomez 851 free(MBmask);
2433 : chl 96
2434 : edgomez 851 return gmc;
2435 : chl 96 }
2436 :    
2437 : edgomez 1053 /* functions which perform BITS-based search/bitcount */
2438 : chl 96
2439 : edgomez 851 static int
2440 :     CountMBBitsInter(SearchData * const Data,
2441 :     const MACROBLOCK * const pMBs, const int x, const int y,
2442 :     const MBParam * const pParam,
2443 :     const uint32_t MotionFlags)
2444 : chl 96 {
2445 : edgomez 851 int i, iDirection;
2446 :     int32_t bsad[5];
2447 : chl 141
2448 : edgomez 851 CheckCandidate = CheckCandidateBits16;
2449 : chl 96
2450 : edgomez 851 if (Data->qpel) {
2451 :     for(i = 0; i < 5; i++) {
2452 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2453 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2454 :     }
2455 :     Data->qpel_precision = 1;
2456 :     CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2457 : chl 96
2458 : edgomez 1053 if (MotionFlags & (XVID_ME_HALFPELREFINE16_BITS | XVID_ME_EXTSEARCH_BITS)) { /* we have to prepare for halfpixel-precision search */
2459 : edgomez 851 for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2460 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2461 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2462 :     Data->qpel_precision = 0;
2463 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2464 :     CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2465 :     }
2466 : edgomez 195
2467 : edgomez 1053 } else { /* not qpel */
2468 : edgomez 195
2469 : edgomez 851 CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2470 :     }
2471 : chl 96
2472 : edgomez 949 if (MotionFlags&XVID_ME_EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2473 : chl 96
2474 : edgomez 949 if (MotionFlags&XVID_ME_HALFPELREFINE16_BITS) SubpelRefine(Data);
2475 : chl 96
2476 : edgomez 851 if (Data->qpel) {
2477 : edgomez 1053 if (MotionFlags&(XVID_ME_EXTSEARCH_BITS | XVID_ME_HALFPELREFINE16_BITS)) { /* there was halfpel-precision search */
2478 : edgomez 851 for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2479 : edgomez 1053 Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* we have found a better match */
2480 : edgomez 851 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2481 :     }
2482 : edgomez 195
2483 : edgomez 1053 /* preparing for qpel-precision search */
2484 : edgomez 851 Data->qpel_precision = 1;
2485 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2486 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
2487 :     }
2488 : edgomez 949 if (MotionFlags&XVID_ME_QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2489 : edgomez 851 }
2490 : chl 96
2491 : edgomez 1053 if (MotionFlags&XVID_ME_CHECKPREDICTION_BITS) { /* let's check vector equal to prediction */
2492 : edgomez 851 VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2493 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2494 :     CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2495 :     }
2496 :     return Data->iMinSAD[0];
2497 :     }
2498 : chl 96
2499 : edgomez 851 static int
2500 :     CountMBBitsInter4v(const SearchData * const Data,
2501 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2502 :     const int x, const int y,
2503 :     const MBParam * const pParam, const uint32_t MotionFlags,
2504 :     const VECTOR * const backup)
2505 :     {
2506 : chl 96
2507 : edgomez 851 int cbp = 0, bits = 0, t = 0, i, iDirection;
2508 :     SearchData Data2, *Data8 = &Data2;
2509 :     int sumx = 0, sumy = 0;
2510 : edgomez 962 int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64;
2511 : edgomez 1022 uint8_t * ptr;
2512 : chl 96
2513 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2514 :     CheckCandidate = CheckCandidateBits8;
2515 : chl 96
2516 : edgomez 1053 for (i = 0; i < 4; i++) { /* for all luma blocks */
2517 : edgomez 1022
2518 : edgomez 851 Data8->iMinSAD = Data->iMinSAD + i + 1;
2519 :     Data8->currentMV = Data->currentMV + i + 1;
2520 :     Data8->currentQMV = Data->currentQMV + i + 1;
2521 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2522 : edgomez 978 Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2523 :     Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2524 :     Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2525 :     Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2526 : edgomez 195
2527 : edgomez 851 if(Data->qpel) {
2528 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2529 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2530 :     Data8->predMV, Data8->iFcode, 0, 0);
2531 :     } else {
2532 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2533 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2534 :     Data8->predMV, Data8->iFcode, 0, 0);
2535 :     }
2536 : chl 96
2537 : edgomez 851 get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2538 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2539 : chl 96
2540 : edgomez 1023 *Data8->iMinSAD += BITS_MULT*t;
2541 : chl 96
2542 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2543 : edgomez 1053 /* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */
2544 : edgomez 987 {
2545 :     VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV;
2546 : edgomez 1022 if (!MVequal (*v, backup[i+1]) )
2547 : edgomez 851 CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2548 :     }
2549 : edgomez 195
2550 : edgomez 851 if (Data8->qpel) {
2551 : edgomez 1053 if (MotionFlags&XVID_ME_HALFPELREFINE8_BITS || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_BITS)) { /* halfpixel motion search follows */
2552 : edgomez 851 int32_t s = *Data8->iMinSAD;
2553 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2554 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2555 :     Data8->qpel_precision = 0;
2556 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2557 :     pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2558 : edgomez 195
2559 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2560 :     CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2561 : edgomez 195
2562 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS)
2563 : edgomez 851 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2564 : chl 96
2565 : edgomez 1022 if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS)
2566 :     SubpelRefine(Data8);
2567 : chl 96
2568 : edgomez 1053 if(s > *Data8->iMinSAD) { /* we have found a better match */
2569 : edgomez 851 Data8->currentQMV->x = 2*Data8->currentMV->x;
2570 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2571 :     }
2572 : chl 96
2573 : edgomez 851 Data8->qpel_precision = 1;
2574 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2575 :     pParam->width, pParam->height, Data8->iFcode, 1, 0);
2576 : chl 96
2577 : edgomez 851 }
2578 : edgomez 949 if (MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2579 : edgomez 195
2580 : edgomez 1053 } else { /* not qpel */
2581 : edgomez 851
2582 : edgomez 1053 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) /* extsearch */
2583 : edgomez 1022 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2584 :    
2585 :     if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS)
2586 : edgomez 1053 SubpelRefine(Data8); /* halfpel refinement */
2587 : edgomez 1022 }
2588 :    
2589 : edgomez 1053 /* checking vector equal to predicion */
2590 : edgomez 949 if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_BITS) {
2591 : edgomez 851 const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2592 : edgomez 1022 if (!MVequal(*v, Data8->predMV))
2593 : edgomez 851 CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2594 : chl 141 }
2595 : edgomez 195
2596 : edgomez 851 bits += *Data8->iMinSAD;
2597 : edgomez 1053 if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */
2598 : edgomez 851
2599 : edgomez 1053 /* MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else */
2600 : edgomez 851 if(Data->qpel) {
2601 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2602 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2603 :     pMB->qmvs[i] = *Data8->currentQMV;
2604 :     sumx += Data8->currentQMV->x/2;
2605 :     sumy += Data8->currentQMV->y/2;
2606 :     } else {
2607 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2608 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2609 :     sumx += Data8->currentMV->x;
2610 :     sumy += Data8->currentMV->y;
2611 : chl 141 }
2612 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2613 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2614 :     if (Data8->temp[0]) cbp |= 1 << (5 - i);
2615 : chl 141
2616 : edgomez 1053 } /* /for all luma blocks */
2617 : chl 141
2618 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
2619 : chl 141
2620 : edgomez 1053 /* let's check chroma */
2621 : edgomez 1022 sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2622 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2623 : chl 96
2624 : edgomez 1053 /* chroma U */
2625 : edgomez 1022 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2626 :     transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2627 : edgomez 1023 bits += Block_CalcBits(coeff, in, Data->iQuant, Data->quant_type, &cbp, 4);
2628 : edgomez 1022
2629 :     if (bits >= *Data->iMinSAD) return bits;
2630 :    
2631 : edgomez 1053 /* chroma V */
2632 : edgomez 1022 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2633 :     transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2634 : edgomez 1023 bits += Block_CalcBits(coeff, in, Data->iQuant, Data->quant_type, &cbp, 5);
2635 : edgomez 1022
2636 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2637 : edgomez 1022
2638 : edgomez 851 return bits;
2639 :     }
2640 : chl 96
2641 : edgomez 851 static int
2642 :     CountMBBitsIntra(const SearchData * const Data)
2643 :     {
2644 : edgomez 1053 int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */
2645 : edgomez 1023 int cbp = 0, i, dc = 0;
2646 : edgomez 962 int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2647 : edgomez 195
2648 : edgomez 851 for(i = 0; i < 4; i++) {
2649 :     int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2650 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2651 : edgomez 1023 bits += Block_CalcBitsIntra(coeff, in, Data->iQuant, Data->quant_type, &cbp, i, &dc);
2652 : chl 141
2653 : edgomez 1022 if (bits >= Data->iMinSAD[0]) return bits;
2654 : edgomez 851 }
2655 : chl 181
2656 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len;
2657 : chl 96
2658 : edgomez 1053 /*chroma U */
2659 : edgomez 1022 transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2660 : edgomez 1023 bits += Block_CalcBitsIntra(coeff, in, Data->iQuant, Data->quant_type, &cbp, 4, &dc);
2661 :    
2662 : edgomez 1022 if (bits >= Data->iMinSAD[0]) return bits;
2663 :    
2664 : edgomez 1053 /* chroma V */
2665 : edgomez 1022 transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2666 : edgomez 1023 bits += Block_CalcBitsIntra(coeff, in, Data->iQuant, Data->quant_type, &cbp, 5, &dc);
2667 : edgomez 1022
2668 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2669 : edgomez 1022
2670 : edgomez 851 return bits;
2671 : chl 96 }

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