[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 1135 - (view) (download)

1 : edgomez 1054 /*****************************************************************************
2 : Isibaar 3 *
3 : edgomez 1054 * XVID MPEG-4 VIDEO CODEC
4 :     * - Motion Estimation related code -
5 : chl 259 *
6 : edgomez 1054 * Copyright(C) 2002 Christoph Lampert <gruel@web.de>
7 :     * 2002 Michael Militzer <michael@xvid.org>
8 :     * 2002-2003 Radoslaw Czyz <xvid@syskin.cjb.net>
9 : chl 259 *
10 : edgomez 1054 * This program is free software ; you can redistribute it and/or modify
11 :     * it under the terms of the GNU General Public License as published by
12 :     * the Free Software Foundation ; either version 2 of the License, or
13 :     * (at your option) any later version.
14 : chl 259 *
15 : edgomez 1054 * This program is distributed in the hope that it will be useful,
16 :     * but WITHOUT ANY WARRANTY ; without even the implied warranty of
17 :     * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 :     * GNU General Public License for more details.
19 : chl 259 *
20 : edgomez 1054 * You should have received a copy of the GNU General Public License
21 :     * along with this program ; if not, write to the Free Software
22 :     * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 : chl 259 *
24 : syskin 1135 * $Id: motion_est.c,v 1.58.2.32 2003-08-29 13:47:21 syskin Exp $
25 : edgomez 1054 *
26 :     ****************************************************************************/
27 : chl 259
28 : Isibaar 3 #include <assert.h>
29 :     #include <stdio.h>
30 : chl 96 #include <stdlib.h>
31 : edgomez 1053 #include <string.h> /* memcpy */
32 :     #include <math.h> /* lrint */
33 : Isibaar 3
34 :     #include "../encoder.h"
35 :     #include "../utils/mbfunctions.h"
36 :     #include "../prediction/mbprediction.h"
37 : syskin 1084 #include "../global.h"
38 : Isibaar 3 #include "../utils/timer.h"
39 : edgomez 851 #include "../image/interpolate8x8.h"
40 :     #include "motion_est.h"
41 : suxen_drol 118 #include "motion.h"
42 : Isibaar 3 #include "sad.h"
43 : chl 1077 #include "gmc.h"
44 : edgomez 851 #include "../utils/emms.h"
45 :     #include "../dct/fdct.h"
46 : Isibaar 3
47 : edgomez 872 /*****************************************************************************
48 :     * Modified rounding tables -- declared in motion.h
49 :     * Original tables see ISO spec tables 7-6 -> 7-9
50 :     ****************************************************************************/
51 :    
52 :     const uint32_t roundtab[16] =
53 :     {0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 };
54 :    
55 :     /* K = 4 */
56 :     const uint32_t roundtab_76[16] =
57 :     { 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 };
58 :    
59 :     /* K = 2 */
60 :     const uint32_t roundtab_78[8] =
61 :     { 0, 0, 1, 1, 0, 0, 0, 1 };
62 :    
63 :     /* K = 1 */
64 :     const uint32_t roundtab_79[4] =
65 :     { 0, 1, 0, 0 };
66 :    
67 : edgomez 851 #define INITIAL_SKIP_THRESH (10)
68 :     #define FINAL_SKIP_THRESH (50)
69 :     #define MAX_SAD00_FOR_SKIP (20)
70 :     #define MAX_CHROMA_SAD_FOR_SKIP (22)
71 : Isibaar 3
72 : edgomez 851 #define CHECK_CANDIDATE(X,Y,D) { \
73 : syskin 1132 CheckCandidate((X),(Y), data, (D) ); }
74 : Isibaar 3
75 : syskin 1084
76 : edgomez 872 /*****************************************************************************
77 :     * Code
78 :     ****************************************************************************/
79 :    
80 : edgomez 851 static __inline uint32_t
81 :     d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
82 :     {
83 : edgomez 974 int bits;
84 :     const int q = (1 << (iFcode - 1)) - 1;
85 :    
86 :     x <<= qpel;
87 :     y <<= qpel;
88 : edgomez 851 if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
89 : chl 141
90 : edgomez 851 x -= pred.x;
91 : edgomez 974 bits = (x != 0 ? iFcode:0);
92 :     x = abs(x);
93 :     x += q;
94 :     x >>= (iFcode - 1);
95 :     bits += mvtab[x];
96 :    
97 : edgomez 851 y -= pred.y;
98 : edgomez 974 bits += (y != 0 ? iFcode:0);
99 :     y = abs(y);
100 :     y += q;
101 :     y >>= (iFcode - 1);
102 :     bits += mvtab[y];
103 : chl 141
104 : edgomez 974 return bits;
105 : edgomez 851 }
106 : chl 141
107 : edgomez 978 static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by,
108 :     const SearchData * const data)
109 : edgomez 851 {
110 :     int sad;
111 :     const uint32_t stride = data->iEdgedWidth/2;
112 : syskin 1129 uint8_t *f_refu, *f_refv, *b_refu, *b_refv;
113 :    
114 :     const INTERPOLATE8X8_PTR interpolate8x8_halfpel[] = {
115 :     NULL,
116 :     interpolate8x8_halfpel_v,
117 :     interpolate8x8_halfpel_h,
118 :     interpolate8x8_halfpel_hv
119 :     };
120 :    
121 : edgomez 978 int offset = (fx>>1) + (fy>>1)*stride;
122 : syskin 1129 int filter = ((fx & 1) << 1) | (fy & 1);
123 : Isibaar 3
124 : syskin 1129 if (filter != 0) {
125 :     f_refu = data->RefQ;
126 :     f_refv = data->RefQ + 8;
127 :     interpolate8x8_halfpel[filter](f_refu, data->RefP[4] + offset, stride, data->rounding);
128 :     interpolate8x8_halfpel[filter](f_refv, data->RefP[5] + offset, stride, data->rounding);
129 :     } else {
130 :     f_refu = (uint8_t*)data->RefP[4] + offset;
131 :     f_refv = (uint8_t*)data->RefP[5] + offset;
132 : edgomez 851 }
133 : Isibaar 3
134 : edgomez 978 offset = (bx>>1) + (by>>1)*stride;
135 : syskin 1129 filter = ((bx & 1) << 1) | (by & 1);
136 :    
137 :     if (filter != 0) {
138 :     b_refu = data->RefQ + 16;
139 :     b_refv = data->RefQ + 24;
140 :     interpolate8x8_halfpel[filter](b_refu, data->b_RefP[4] + offset, stride, data->rounding);
141 :     interpolate8x8_halfpel[filter](b_refv, data->b_RefP[5] + offset, stride, data->rounding);
142 :     } else {
143 :     b_refu = (uint8_t*)data->b_RefP[4] + offset;
144 :     b_refv = (uint8_t*)data->b_RefP[5] + offset;
145 : edgomez 851 }
146 : Isibaar 3
147 : edgomez 851 sad = sad8bi(data->CurU, b_refu, f_refu, stride);
148 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
149 : Isibaar 3
150 : edgomez 851 return sad;
151 :     }
152 : Isibaar 3
153 : edgomez 851 static int32_t
154 : edgomez 978 ChromaSAD(const int dx, const int dy, const SearchData * const data)
155 : edgomez 851 {
156 :     int sad;
157 :     const uint32_t stride = data->iEdgedWidth/2;
158 : edgomez 978 int offset = (dx>>1) + (dy>>1)*stride;
159 : syskin 1132 int next = 1;
160 : Isibaar 3
161 : edgomez 1053 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */
162 :     data->temp[5] = dx; data->temp[6] = dy; /* backup */
163 : Isibaar 3
164 : edgomez 851 switch (((dx & 1) << 1) | (dy & 1)) {
165 :     case 0:
166 : edgomez 978 sad = sad8(data->CurU, data->RefP[4] + offset, stride);
167 :     sad += sad8(data->CurV, data->RefP[5] + offset, stride);
168 : edgomez 851 break;
169 :     case 1:
170 : syskin 1132 next = stride;
171 : edgomez 851 case 2:
172 : syskin 1132 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + next, stride);
173 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + next, stride);
174 : edgomez 851 break;
175 :     default:
176 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding);
177 : edgomez 851 sad = sad8(data->CurU, data->RefQ, stride);
178 :    
179 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding);
180 : edgomez 851 sad += sad8(data->CurV, data->RefQ, stride);
181 :     break;
182 :     }
183 : edgomez 1053 data->temp[7] = sad; /* backup, part 2 */
184 : edgomez 851 return sad;
185 : Isibaar 3 }
186 :    
187 : edgomez 851 static __inline const uint8_t *
188 :     GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
189 : Isibaar 3 {
190 : edgomez 1053 /* dir : 0 = forward, 1 = backward */
191 : edgomez 987 const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP );
192 : edgomez 978 const int picture = ((x&1)<<1) | (y&1);
193 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
194 :     return direction[picture] + offset;
195 : Isibaar 3 }
196 :    
197 : edgomez 1053 /* this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate */
198 : edgomez 851 static __inline const uint8_t *
199 :     GetReference(const int x, const int y, const SearchData * const data)
200 : Isibaar 3 {
201 : edgomez 978 const int picture = ((x&1)<<1) | (y&1);
202 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
203 :     return data->RefP[picture] + offset;
204 : Isibaar 3 }
205 : edgomez 851
206 :     static uint8_t *
207 :     Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
208 : Isibaar 3 {
209 : edgomez 1053 /* create or find a qpel-precision reference picture; return pointer to it */
210 : edgomez 851 uint8_t * Reference = data->RefQ + 16*dir;
211 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
212 :     const uint32_t rounding = data->rounding;
213 :     const int halfpel_x = x/2;
214 :     const int halfpel_y = y/2;
215 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
216 : suxen_drol 136
217 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
218 :     ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
219 :     switch( ((x&1)<<1) + (y&1) ) {
220 : edgomez 1053 case 3: /* x and y in qpel resolution - the "corners" (top left/right and */
221 :     /* bottom left/right) during qpel refinement */
222 : edgomez 963 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
223 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
224 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
225 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
226 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
227 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
228 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
229 : edgomez 851 break;
230 : chl 184
231 : edgomez 1053 case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
232 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
233 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
234 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
235 :     break;
236 : suxen_drol 136
237 : edgomez 1053 case 2: /* x qpel, y halfpel - left or right during qpel refinement */
238 : edgomez 851 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
239 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
240 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
241 :     break;
242 : edgomez 195
243 : edgomez 1053 default: /* pure halfpel position */
244 : edgomez 963 return (uint8_t *) ref1;
245 : edgomez 978
246 : edgomez 851 }
247 :     return Reference;
248 :     }
249 : Isibaar 3
250 : edgomez 851 static uint8_t *
251 :     Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
252 :     {
253 : edgomez 1053 /* create or find a qpel-precision reference picture; return pointer to it */
254 : edgomez 851 uint8_t * Reference = data->RefQ + 16*dir;
255 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
256 :     const uint32_t rounding = data->rounding;
257 :     const int halfpel_x = x/2;
258 :     const int halfpel_y = y/2;
259 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
260 : chl 326
261 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
262 :     switch( ((x&1)<<1) + (y&1) ) {
263 : edgomez 1053 case 3:
264 :     /*
265 :     * x and y in qpel resolution - the "corners" (top left/right and
266 :     * bottom left/right) during qpel refinement
267 :     */
268 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
269 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
270 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
271 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
272 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
273 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
274 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
275 :     break;
276 : chl 184
277 : edgomez 1053 case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
278 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
279 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
280 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
281 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
282 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
283 :     break;
284 : chl 184
285 : edgomez 1053 case 2: /* x qpel, y halfpel - left or right during qpel refinement */
286 : edgomez 851 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
287 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
288 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
289 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
290 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
291 :     break;
292 : edgomez 195
293 : syskin 1129
294 : edgomez 1053 default: /* pure halfpel position */
295 : edgomez 851 return (uint8_t *) ref1;
296 :     }
297 :     return Reference;
298 :     }
299 : edgomez 195
300 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS START */
301 : edgomez 195
302 : edgomez 851 static void
303 : syskin 1132 CheckCandidate16(const int x, const int y, const SearchData * const data, const int Direction)
304 : edgomez 851 {
305 :     int xc, yc;
306 :     const uint8_t * Reference;
307 :     VECTOR * current;
308 :     int32_t sad; uint32_t t;
309 : edgomez 195
310 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
311 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
312 : edgomez 195
313 : edgomez 851 if (!data->qpel_precision) {
314 :     Reference = GetReference(x, y, data);
315 :     current = data->currentMV;
316 :     xc = x; yc = y;
317 : edgomez 1053 } else { /* x and y are in 1/4 precision */
318 : edgomez 851 Reference = Interpolate16x16qpel(x, y, 0, data);
319 : edgomez 1053 xc = x/2; yc = y/2; /* for chroma sad */
320 : edgomez 851 current = data->currentQMV;
321 :     }
322 : edgomez 195
323 : syskin 1132 sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp);
324 : edgomez 851 t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
325 : edgomez 195
326 : edgomez 851 sad += (data->lambda16 * t * sad)>>10;
327 : syskin 1132 data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10;
328 : chl 175
329 : syskin 1084 if (data->chroma && sad < data->iMinSAD[0])
330 :     sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
331 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
332 : edgomez 195
333 : edgomez 851 if (sad < data->iMinSAD[0]) {
334 :     data->iMinSAD[0] = sad;
335 :     current[0].x = x; current[0].y = y;
336 : syskin 1132 *data->dir = Direction;
337 : edgomez 851 }
338 : edgomez 195
339 : syskin 1132 if (data->temp[0] < data->iMinSAD[1]) {
340 :     data->iMinSAD[1] = data->temp[0]; current[1].x = x; current[1].y = y; }
341 :     if (data->temp[1] < data->iMinSAD[2]) {
342 :     data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
343 :     if (data->temp[2] < data->iMinSAD[3]) {
344 :     data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
345 :     if (data->temp[3] < data->iMinSAD[4]) {
346 :     data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
347 : Isibaar 3 }
348 :    
349 : edgomez 851 static void
350 : syskin 1132 CheckCandidate8(const int x, const int y, const SearchData * const data, const int Direction)
351 : edgomez 851 {
352 :     int32_t sad; uint32_t t;
353 :     const uint8_t * Reference;
354 : edgomez 974 VECTOR * current;
355 : chl 326
356 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
357 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
358 : Isibaar 3
359 : edgomez 974 if (!data->qpel_precision) {
360 :     Reference = GetReference(x, y, data);
361 :     current = data->currentMV;
362 : edgomez 1053 } else { /* x and y are in 1/4 precision */
363 : edgomez 974 Reference = Interpolate8x8qpel(x, y, 0, 0, data);
364 :     current = data->currentQMV;
365 :     }
366 : Isibaar 3
367 : edgomez 851 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
368 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
369 : Isibaar 3
370 : edgomez 851 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
371 : Isibaar 3
372 : edgomez 851 if (sad < *(data->iMinSAD)) {
373 :     *(data->iMinSAD) = sad;
374 : edgomez 974 current->x = x; current->y = y;
375 : syskin 1132 *data->dir = Direction;
376 : edgomez 851 }
377 : Isibaar 3 }
378 :    
379 : edgomez 851 static void
380 : syskin 1132 CheckCandidate32(const int x, const int y, const SearchData * const data, const int Direction)
381 : edgomez 851 {
382 :     uint32_t t;
383 :     const uint8_t * Reference;
384 : syskin 1132 int sad;
385 : edgomez 851
386 : edgomez 1053 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */
387 : edgomez 851 (x > data->max_dx) || (x < data->min_dx)
388 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
389 :    
390 :     Reference = GetReference(x, y, data);
391 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
392 :    
393 : syskin 1132 sad = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp);
394 : edgomez 851
395 : syskin 1132 sad += (data->lambda16 * t * sad) >> 10;
396 :     data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10;
397 : edgomez 851
398 : syskin 1132 if (sad < data->iMinSAD[0]) {
399 :     data->iMinSAD[0] = sad;
400 : edgomez 851 data->currentMV[0].x = x; data->currentMV[0].y = y;
401 : syskin 1132 *data->dir = Direction;
402 :     }
403 : edgomez 851
404 : syskin 1132 if (data->temp[0] < data->iMinSAD[1]) {
405 :     data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
406 :     if (data->temp[1] < data->iMinSAD[2]) {
407 :     data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
408 :     if (data->temp[2] < data->iMinSAD[3]) {
409 :     data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
410 :     if (data->temp[3] < data->iMinSAD[4]) {
411 :     data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
412 : Isibaar 3 }
413 : edgomez 195
414 : edgomez 851 static void
415 : syskin 1132 CheckCandidate16no4v(const int x, const int y, const SearchData * const data, const int Direction)
416 : edgomez 851 {
417 :     int32_t sad, xc, yc;
418 :     const uint8_t * Reference;
419 :     uint32_t t;
420 :     VECTOR * current;
421 :    
422 : edgomez 963 if ( (x > data->max_dx) || ( x < data->min_dx)
423 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
424 : edgomez 851
425 : edgomez 1053 if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */
426 : edgomez 851
427 : edgomez 1053 if (data->qpel_precision) { /* x and y are in 1/4 precision */
428 : edgomez 851 Reference = Interpolate16x16qpel(x, y, 0, data);
429 :     current = data->currentQMV;
430 :     xc = x/2; yc = y/2;
431 :     } else {
432 :     Reference = GetReference(x, y, data);
433 :     current = data->currentMV;
434 :     xc = x; yc = y;
435 :     }
436 :     t = d_mv_bits(x, y, data->predMV, data->iFcode,
437 :     data->qpel^data->qpel_precision, data->rrv);
438 :    
439 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
440 :     sad += (data->lambda16 * t * sad)>>10;
441 :    
442 : syskin 1084 if (data->chroma && sad < *data->iMinSAD)
443 :     sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
444 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
445 : edgomez 851
446 :     if (sad < *(data->iMinSAD)) {
447 :     *(data->iMinSAD) = sad;
448 :     current->x = x; current->y = y;
449 : syskin 1132 *data->dir = Direction;
450 : edgomez 851 }
451 : chl 96 }
452 : Isibaar 3
453 : edgomez 851 static void
454 : syskin 1132 CheckCandidate16I(const int x, const int y, const SearchData * const data, const int Direction)
455 : chl 1077 {
456 :     int sad;
457 :     // int xc, yc;
458 :     const uint8_t * Reference;
459 :     // VECTOR * current;
460 :    
461 :     if ( (x > data->max_dx) || ( x < data->min_dx)
462 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
463 :    
464 :     Reference = GetReference(x, y, data);
465 :     // xc = x; yc = y;
466 :    
467 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
468 :     // sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0);
469 :    
470 :     /* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
471 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
472 :     */
473 :    
474 :     if (sad < data->iMinSAD[0]) {
475 :     data->iMinSAD[0] = sad;
476 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
477 : syskin 1132 *data->dir = Direction;
478 : chl 1077 }
479 :     }
480 :    
481 :     static void
482 : syskin 1132 CheckCandidate32I(const int x, const int y, const SearchData * const data, const int Direction)
483 : edgomez 851 {
484 : edgomez 1053 /* maximum speed - for P/B/I decision */
485 : edgomez 851 int32_t sad;
486 :    
487 :     if ( (x > data->max_dx) || (x < data->min_dx)
488 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
489 :    
490 : syskin 1135 sad = sad32v_c(data->Cur, data->RefP[0] + x + y*((int)data->iEdgedWidth),
491 : syskin 1132 data->iEdgedWidth, data->temp);
492 : edgomez 851
493 :     if (sad < *(data->iMinSAD)) {
494 :     *(data->iMinSAD) = sad;
495 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
496 : syskin 1132 *data->dir = Direction;
497 : edgomez 851 }
498 : syskin 1132 if (data->temp[0] < data->iMinSAD[1]) {
499 :     data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
500 :     if (data->temp[1] < data->iMinSAD[2]) {
501 :     data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
502 :     if (data->temp[2] < data->iMinSAD[3]) {
503 :     data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
504 :     if (data->temp[3] < data->iMinSAD[4]) {
505 :     data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
506 : edgomez 851
507 : Isibaar 3 }
508 :    
509 : edgomez 851 static void
510 : syskin 1132 CheckCandidateInt(const int xf, const int yf, const SearchData * const data, const int Direction)
511 : edgomez 851 {
512 :     int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
513 :     uint32_t t;
514 :     const uint8_t *ReferenceF, *ReferenceB;
515 :     VECTOR *current;
516 :    
517 : edgomez 953 if ((xf > data->max_dx) || (xf < data->min_dx) ||
518 :     (yf > data->max_dy) || (yf < data->min_dy))
519 :     return;
520 : edgomez 851
521 :     if (!data->qpel_precision) {
522 :     ReferenceF = GetReference(xf, yf, data);
523 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
524 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
525 :     current = data->currentMV;
526 :     xcf = xf; ycf = yf;
527 :     xcb = xb; ycb = yb;
528 :     } else {
529 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
530 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
531 :     current = data->currentQMV;
532 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
533 :     xcf = xf/2; ycf = yf/2;
534 :     xcb = xb/2; ycb = yb/2;
535 :     }
536 :    
537 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
538 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
539 :    
540 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
541 :     sad += (data->lambda16 * t * sad)>>10;
542 :    
543 : syskin 1084 if (data->chroma && sad < *data->iMinSAD)
544 :     sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
545 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
546 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
547 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
548 : edgomez 851
549 :     if (sad < *(data->iMinSAD)) {
550 :     *(data->iMinSAD) = sad;
551 :     current->x = xf; current->y = yf;
552 : syskin 1132 *data->dir = Direction;
553 : edgomez 851 }
554 : Isibaar 3 }
555 :    
556 : edgomez 851 static void
557 : syskin 1132 CheckCandidateDirect(const int x, const int y, const SearchData * const data, const int Direction)
558 : edgomez 851 {
559 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
560 :     uint32_t k;
561 :     const uint8_t *ReferenceF;
562 :     const uint8_t *ReferenceB;
563 :     VECTOR mvs, b_mvs;
564 :    
565 : edgomez 953 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
566 : edgomez 851
567 :     for (k = 0; k < 4; k++) {
568 :     mvs.x = data->directmvF[k].x + x;
569 :     b_mvs.x = ((x == 0) ?
570 :     data->directmvB[k].x
571 :     : mvs.x - data->referencemv[k].x);
572 :    
573 :     mvs.y = data->directmvF[k].y + y;
574 :     b_mvs.y = ((y == 0) ?
575 :     data->directmvB[k].y
576 :     : mvs.y - data->referencemv[k].y);
577 :    
578 : edgomez 953 if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) ||
579 :     (mvs.y > data->max_dy) || (mvs.y < data->min_dy) ||
580 :     (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) ||
581 :     (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) )
582 :     return;
583 : edgomez 851
584 :     if (data->qpel) {
585 :     xcf += mvs.x/2; ycf += mvs.y/2;
586 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
587 :     } else {
588 :     xcf += mvs.x; ycf += mvs.y;
589 :     xcb += b_mvs.x; ycb += b_mvs.y;
590 : edgomez 1053 mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */
591 : edgomez 851 b_mvs.x *= 2; b_mvs.y *= 2;
592 :     }
593 :    
594 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
595 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
596 :    
597 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
598 :     ReferenceF, ReferenceB, data->iEdgedWidth);
599 :     if (sad > *(data->iMinSAD)) return;
600 :     }
601 :    
602 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
603 :    
604 : syskin 1084 if (data->chroma && sad < *data->iMinSAD)
605 :     sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
606 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
607 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
608 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
609 : edgomez 851
610 :     if (sad < *(data->iMinSAD)) {
611 :     *(data->iMinSAD) = sad;
612 :     data->currentMV->x = x; data->currentMV->y = y;
613 : syskin 1132 *data->dir = Direction;
614 : edgomez 851 }
615 : Isibaar 3 }
616 :    
617 : edgomez 851 static void
618 : syskin 1132 CheckCandidateDirectno4v(const int x, const int y, const SearchData * const data, const int Direction)
619 : Isibaar 3 {
620 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
621 :     const uint8_t *ReferenceF;
622 :     const uint8_t *ReferenceB;
623 :     VECTOR mvs, b_mvs;
624 : Isibaar 3
625 : edgomez 963 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
626 : Isibaar 3
627 : edgomez 851 mvs.x = data->directmvF[0].x + x;
628 :     b_mvs.x = ((x == 0) ?
629 :     data->directmvB[0].x
630 :     : mvs.x - data->referencemv[0].x);
631 : Isibaar 3
632 : edgomez 851 mvs.y = data->directmvF[0].y + y;
633 :     b_mvs.y = ((y == 0) ?
634 :     data->directmvB[0].y
635 :     : mvs.y - data->referencemv[0].y);
636 : Isibaar 3
637 : edgomez 963 if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx)
638 :     || (mvs.y > data->max_dy) || (mvs.y < data->min_dy)
639 :     || (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx)
640 :     || (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return;
641 : Isibaar 3
642 : edgomez 851 if (data->qpel) {
643 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
644 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
645 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
646 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
647 :     } else {
648 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
649 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
650 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
651 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
652 :     }
653 :    
654 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
655 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
656 :    
657 : syskin 1084 if (data->chroma && sad < *data->iMinSAD)
658 :     sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
659 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
660 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
661 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
662 : edgomez 851
663 :     if (sad < *(data->iMinSAD)) {
664 :     *(data->iMinSAD) = sad;
665 :     data->currentMV->x = x; data->currentMV->y = y;
666 : syskin 1132 *data->dir = Direction;
667 : edgomez 851 }
668 : Isibaar 3 }
669 :    
670 : edgomez 851
671 :     static void
672 : syskin 1132 CheckCandidateRD16(const int x, const int y, const SearchData * const data, const int Direction)
673 : Isibaar 3 {
674 :    
675 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
676 : syskin 1108 int32_t rd = 0;
677 : edgomez 851 VECTOR * current;
678 :     const uint8_t * ptr;
679 :     int i, cbp = 0, t, xc, yc;
680 : edgomez 195
681 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
682 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
683 : edgomez 195
684 : edgomez 851 if (!data->qpel_precision) {
685 :     ptr = GetReference(x, y, data);
686 :     current = data->currentMV;
687 :     xc = x; yc = y;
688 : edgomez 1053 } else { /* x and y are in 1/4 precision */
689 : edgomez 851 ptr = Interpolate16x16qpel(x, y, 0, data);
690 :     current = data->currentQMV;
691 :     xc = x/2; yc = y/2;
692 :     }
693 : Isibaar 3
694 : edgomez 851 for(i = 0; i < 4; i++) {
695 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
696 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
697 : syskin 1108 rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i);
698 : edgomez 851 }
699 : Isibaar 3
700 : syskin 1108 rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
701 : edgomez 195
702 : syskin 1071 if (data->temp[0] + t < data->iMinSAD[1]) {
703 : chl 1115 data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) | (cbp&32); }
704 : syskin 1071 if (data->temp[1] < data->iMinSAD[2]) {
705 : chl 1115 data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; data->cbp[1] = (data->cbp[1]&~16) | (cbp&16); }
706 : syskin 1071 if (data->temp[2] < data->iMinSAD[3]) {
707 : chl 1115 data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; data->cbp[1] = (data->cbp[1]&~8) | (cbp&8); }
708 : syskin 1071 if (data->temp[3] < data->iMinSAD[4]) {
709 : chl 1115 data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | (cbp&4); }
710 : syskin 1071
711 : syskin 1108 rd += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
712 : edgomez 851
713 : syskin 1108 if (rd >= data->iMinSAD[0]) return;
714 : edgomez 851
715 : edgomez 1053 /* chroma */
716 : edgomez 1022 xc = (xc >> 1) + roundtab_79[xc & 0x3];
717 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
718 : edgomez 851
719 : edgomez 1053 /* chroma U */
720 : syskin 1084 ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
721 :     transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2);
722 : syskin 1108 rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4);
723 :     if (rd >= data->iMinSAD[0]) return;
724 : edgomez 1022
725 : edgomez 1053 /* chroma V */
726 : syskin 1084 ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
727 :     transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2);
728 : syskin 1108 rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
729 : edgomez 1022
730 : syskin 1108 rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
731 : edgomez 851
732 : syskin 1108 if (rd < data->iMinSAD[0]) {
733 :     data->iMinSAD[0] = rd;
734 : edgomez 851 current[0].x = x; current[0].y = y;
735 : syskin 1132 *data->dir = Direction;
736 : syskin 1084 *data->cbp = cbp;
737 : edgomez 851 }
738 : syskin 1071 }
739 : edgomez 851
740 :     static void
741 : syskin 1132 CheckCandidateRD8(const int x, const int y, const SearchData * const data, const int Direction)
742 : chl 96 {
743 :    
744 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
745 : syskin 1108 int32_t rd;
746 : edgomez 851 VECTOR * current;
747 :     const uint8_t * ptr;
748 : edgomez 1022 int cbp = 0;
749 : edgomez 195
750 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
751 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
752 : edgomez 195
753 : edgomez 851 if (!data->qpel_precision) {
754 :     ptr = GetReference(x, y, data);
755 :     current = data->currentMV;
756 : edgomez 1053 } else { /* x and y are in 1/4 precision */
757 : edgomez 851 ptr = Interpolate8x8qpel(x, y, 0, 0, data);
758 :     current = data->currentQMV;
759 :     }
760 : chl 96
761 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
762 : syskin 1108 rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
763 :     rd += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
764 : chl 96
765 : syskin 1108 if (rd < data->iMinSAD[0]) {
766 : syskin 1084 *data->cbp = cbp;
767 : syskin 1108 data->iMinSAD[0] = rd;
768 : edgomez 851 current[0].x = x; current[0].y = y;
769 : syskin 1132 *data->dir = Direction;
770 : edgomez 851 }
771 :     }
772 : chl 96
773 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
774 : edgomez 195
775 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
776 : chl 96
777 : edgomez 851 static void
778 : syskin 1132 AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate)
779 : edgomez 851 {
780 : edgomez 195
781 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
782 : edgomez 195
783 : syskin 1132 unsigned int * const iDirection = data->dir;
784 : edgomez 195
785 : edgomez 1053 for(;;) { /* forever */
786 : syskin 1132 *iDirection = 0;
787 : edgomez 851 if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
788 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
789 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
790 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
791 : edgomez 195
792 : edgomez 851 /* now we're doing diagonal checks near our candidate */
793 : edgomez 195
794 : syskin 1132 if (*iDirection) { /* if anything found */
795 :     bDirection = *iDirection;
796 :     *iDirection = 0;
797 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
798 : edgomez 1053 if (bDirection & 3) { /* our candidate is left or right */
799 : edgomez 851 CHECK_CANDIDATE(x, y + iDiamondSize, 8);
800 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
801 : edgomez 1053 } else { /* what remains here is up or down */
802 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y, 2);
803 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
804 :     }
805 : edgomez 195
806 : syskin 1132 if (*iDirection) {
807 :     bDirection += *iDirection;
808 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
809 :     }
810 : edgomez 1053 } else { /* about to quit, eh? not so fast.... */
811 : edgomez 851 switch (bDirection) {
812 :     case 2:
813 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
814 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
815 : edgomez 195 break;
816 : edgomez 851 case 1:
817 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
818 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
819 : edgomez 195 break;
820 : edgomez 851 case 2 + 4:
821 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
822 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
823 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
824 :     break;
825 :     case 4:
826 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
827 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
828 :     break;
829 : edgomez 195 case 8:
830 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
831 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
832 : edgomez 195 break;
833 : edgomez 851 case 1 + 4:
834 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
835 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
836 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
837 : edgomez 195 break;
838 : edgomez 851 case 2 + 8:
839 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
840 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
841 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
842 :     break;
843 :     case 1 + 8:
844 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
845 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
846 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
847 :     break;
848 : edgomez 1053 default: /* 1+2+4+8 == we didn't find anything at all */
849 : edgomez 851 CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
850 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
851 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
852 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
853 :     break;
854 : chl 96 }
855 : syskin 1132 if (!*iDirection) break; /* ok, the end. really */
856 :     bDirection = *iDirection;
857 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
858 : chl 345 }
859 : edgomez 195 }
860 : chl 96 }
861 :    
862 : edgomez 851 static void
863 : syskin 1132 SquareSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate)
864 : chl 96 {
865 : syskin 1132 unsigned int * const iDirection = data->dir;
866 : edgomez 195
867 : edgomez 851 do {
868 : syskin 1132 *iDirection = 0;
869 : edgomez 851 if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
870 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
871 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
872 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
873 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
874 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
875 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
876 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
877 : chl 96
878 : syskin 1132 bDirection = *iDirection;
879 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
880 : syskin 1132 } while (*iDirection);
881 : chl 96 }
882 :    
883 : edgomez 851 static void
884 : syskin 1132 DiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate)
885 : chl 181 {
886 :    
887 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
888 :    
889 : syskin 1132 unsigned int * const iDirection = data->dir;
890 : edgomez 195
891 : edgomez 851 do {
892 : syskin 1132 *iDirection = 0;
893 : edgomez 851 if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
894 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
895 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
896 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
897 :    
898 :     /* now we're doing diagonal checks near our candidate */
899 :    
900 : syskin 1132 if (*iDirection) { /* checking if anything found */
901 :     bDirection = *iDirection;
902 :     *iDirection = 0;
903 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
904 : edgomez 1053 if (bDirection & 3) { /* our candidate is left or right */
905 : edgomez 851 CHECK_CANDIDATE(x, y + iDiamondSize, 8);
906 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
907 : edgomez 1053 } else { /* what remains here is up or down */
908 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y, 2);
909 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
910 :     }
911 : syskin 1132 bDirection += *iDirection;
912 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
913 :     }
914 :     }
915 : syskin 1132 while (*iDirection);
916 : edgomez 851 }
917 : chl 181
918 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
919 : chl 181
920 : edgomez 851 static void
921 : syskin 1132 SubpelRefine(const SearchData * const data, CheckFunc * const CheckCandidate)
922 : edgomez 851 {
923 :     /* Do a half-pel or q-pel refinement */
924 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
925 : chl 181
926 : edgomez 851 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
927 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
928 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
929 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
930 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
931 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
932 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
933 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
934 :     }
935 : chl 181
936 : edgomez 851 static __inline int
937 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
938 :     const int x, const int y,
939 :     const uint32_t stride, const uint32_t iQuant, int rrv)
940 : chl 181
941 : edgomez 851 {
942 : edgomez 963 int offset = (x + y*stride)*8;
943 : edgomez 851 if(!rrv) {
944 : edgomez 963 uint32_t sadC = sad8(current->u + offset,
945 :     reference->u + offset, stride);
946 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
947 : edgomez 963 sadC += sad8(current->v + offset,
948 :     reference->v + offset, stride);
949 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
950 :     return 1;
951 : chl 181
952 : edgomez 851 } else {
953 : edgomez 963 uint32_t sadC = sad16(current->u + 2*offset,
954 :     reference->u + 2*offset, stride, 256*4096);
955 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
956 : edgomez 963 sadC += sad16(current->v + 2*offset,
957 :     reference->v + 2*offset, stride, 256*4096);
958 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
959 :     return 1;
960 : chl 181 }
961 : edgomez 851 }
962 : edgomez 504
963 : edgomez 851 static __inline void
964 : Isibaar 1095 ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad)
965 :     {
966 :     pMB->mode = MODE_INTER;
967 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
968 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
969 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
970 :     }
971 :    
972 :     static __inline void
973 : edgomez 1022 ModeDecision(SearchData * const Data,
974 :     MACROBLOCK * const pMB,
975 :     const MACROBLOCK * const pMBs,
976 :     const int x, const int y,
977 :     const MBParam * const pParam,
978 :     const uint32_t MotionFlags,
979 :     const uint32_t VopFlags,
980 :     const uint32_t VolFlags,
981 :     const IMAGE * const pCurrent,
982 : syskin 1084 const IMAGE * const pRef,
983 :     const IMAGE * const vGMC,
984 :     const int coding_type)
985 : edgomez 1022 {
986 :     int mode = MODE_INTER;
987 : syskin 1084 int mcsel = 0;
988 : edgomez 1022 int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
989 :     const uint32_t iQuant = pMB->quant;
990 :    
991 : syskin 1084 const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0);
992 : edgomez 1022
993 : syskin 1084 pMB->mcsel = 0;
994 :    
995 : edgomez 1107 if (!(VopFlags & XVID_VOP_MODEDECISION_RD)) { /* normal, fast, SAD-based mode decision */
996 : edgomez 1022 int sad;
997 :     int InterBias = MV16_INTER_BIAS;
998 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
999 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1000 :     mode = MODE_INTER;
1001 :     sad = Data->iMinSAD[0];
1002 :     } else {
1003 :     mode = MODE_INTER4V;
1004 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1005 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1006 :     Data->iMinSAD[0] = sad;
1007 :     }
1008 :    
1009 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
1010 :     if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP))
1011 :     if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH)
1012 :     if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) {
1013 :     mode = MODE_NOT_CODED;
1014 :     sad = 0;
1015 :     }
1016 :    
1017 : syskin 1084 /* mcsel */
1018 :     if (coding_type == S_VOP) {
1019 :    
1020 :     int32_t iSAD = sad16(Data->Cur,
1021 :     vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536);
1022 :    
1023 :     if (Data->chroma) {
1024 :     iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
1025 :     iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
1026 :     }
1027 :    
1028 :     if (iSAD <= sad) { /* mode decision GMC */
1029 :     mode = MODE_INTER;
1030 :     mcsel = 1;
1031 :     sad = iSAD;
1032 :     }
1033 :    
1034 :     }
1035 :    
1036 : edgomez 1022 /* intra decision */
1037 :    
1038 : edgomez 1053 if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */
1039 : edgomez 1022 if (y != 0)
1040 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1041 :     if (x != 0)
1042 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1043 :    
1044 : syskin 1084 if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */
1045 : edgomez 1022 if (Data->rrv) InterBias *= 4;
1046 :    
1047 : syskin 1084 if (InterBias < sad) {
1048 : edgomez 1022 int32_t deviation;
1049 : syskin 1084 if (!Data->rrv)
1050 :     deviation = dev16(Data->Cur, Data->iEdgedWidth);
1051 :     else
1052 :     deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */
1053 :     dev16(Data->Cur+16, Data->iEdgedWidth) +
1054 :     dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) +
1055 :     dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth);
1056 : edgomez 1022
1057 :     if (deviation < (sad - InterBias)) mode = MODE_INTRA;
1058 :     }
1059 :    
1060 : syskin 1084 pMB->cbp = 63;
1061 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
1062 :    
1063 : syskin 1108 } else { /* Rate-Distortion */
1064 : edgomez 1022
1065 : syskin 1108 int min_rd, intra_rd, i, cbp, c[2] = {0, 0};
1066 : edgomez 1022 VECTOR backup[5], *v;
1067 :     Data->iQuant = iQuant;
1068 : syskin 1084 Data->cbp = c;
1069 : edgomez 1022
1070 :     v = Data->qpel ? Data->currentQMV : Data->currentMV;
1071 :     for (i = 0; i < 5; i++) {
1072 :     Data->iMinSAD[i] = 256*4096;
1073 :     backup[i] = v[i];
1074 :     }
1075 :    
1076 : syskin 1108 min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags);
1077 : syskin 1084 cbp = *Data->cbp;
1078 :    
1079 :     if (coding_type == S_VOP) {
1080 : syskin 1108 int gmc_rd;
1081 :     *Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */
1082 :     gmc_rd = findRDgmc(Data, vGMC, x, y);
1083 :     if (gmc_rd < min_rd) {
1084 : syskin 1084 mcsel = 1;
1085 : syskin 1108 *Data->iMinSAD = min_rd = gmc_rd;
1086 : syskin 1084 mode = MODE_INTER;
1087 :     cbp = *Data->cbp;
1088 : edgomez 1022 }
1089 : syskin 1084 }
1090 : edgomez 1022
1091 : syskin 1084 if (inter4v) {
1092 : syskin 1108 int v4_rd;
1093 :     v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1094 :     if (v4_rd < min_rd) {
1095 :     Data->iMinSAD[0] = min_rd = v4_rd;
1096 : syskin 1084 mode = MODE_INTER4V;
1097 :     cbp = *Data->cbp;
1098 :     }
1099 :     }
1100 : edgomez 1022
1101 : syskin 1108 intra_rd = findRDintra(Data);
1102 :     if (intra_rd < min_rd) {
1103 :     *Data->iMinSAD = min_rd = intra_rd;
1104 : syskin 1084 mode = MODE_INTRA;
1105 : edgomez 1022 }
1106 : syskin 1084
1107 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0;
1108 :     pMB->cbp = cbp;
1109 : edgomez 1022 }
1110 :    
1111 :     if (Data->rrv) {
1112 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1113 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1114 :     }
1115 :    
1116 : syskin 1084 if (mode == MODE_INTER && mcsel == 0) {
1117 : edgomez 1022 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1118 :    
1119 :     if(Data->qpel) {
1120 :     pMB->qmvs[0] = pMB->qmvs[1]
1121 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1122 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1123 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1124 :     } else {
1125 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1126 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1127 :     }
1128 :    
1129 : syskin 1084 } else if (mode == MODE_INTER ) { // but mcsel == 1
1130 :    
1131 :     pMB->mcsel = 1;
1132 :     if (Data->qpel) {
1133 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv;
1134 :     pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2;
1135 :     pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2;
1136 :     } else
1137 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv;
1138 :    
1139 :     } else
1140 :     if (mode == MODE_INTER4V) ; /* anything here? */
1141 :     else /* INTRA, NOT_CODED */
1142 : syskin 1108 ZeroMacroblockP(pMB, 0);
1143 : edgomez 1022
1144 :     pMB->mode = mode;
1145 :     }
1146 :    
1147 : edgomez 851 bool
1148 :     MotionEstimation(MBParam * const pParam,
1149 :     FRAMEINFO * const current,
1150 :     FRAMEINFO * const reference,
1151 :     const IMAGE * const pRefH,
1152 :     const IMAGE * const pRefV,
1153 :     const IMAGE * const pRefHV,
1154 : chl 1077 const IMAGE * const pGMC,
1155 : edgomez 851 const uint32_t iLimit)
1156 : chl 181 {
1157 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
1158 :     const IMAGE *const pCurrent = &current->image;
1159 :     const IMAGE *const pRef = &reference->image;
1160 : chl 181
1161 : edgomez 851 uint32_t mb_width = pParam->mb_width;
1162 :     uint32_t mb_height = pParam->mb_height;
1163 :     const uint32_t iEdgedWidth = pParam->edged_width;
1164 : suxen_drol 890 const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
1165 : chl 181
1166 : edgomez 851 uint32_t x, y;
1167 :     uint32_t iIntra = 0;
1168 : syskin 1108 int32_t sad00;
1169 : chl 1077 int skip_thresh = INITIAL_SKIP_THRESH * \
1170 : edgomez 953 (current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \
1171 : edgomez 1107 (current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1);
1172 : chl 181
1173 : edgomez 1053 /* some pre-initialized thingies for SearchP */
1174 : syskin 1132 int32_t temp[8]; uint32_t dir;
1175 : edgomez 851 VECTOR currentMV[5];
1176 :     VECTOR currentQMV[5];
1177 :     int32_t iMinSAD[5];
1178 : syskin 1075 DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE);
1179 : edgomez 851 SearchData Data;
1180 :     memset(&Data, 0, sizeof(SearchData));
1181 :     Data.iEdgedWidth = iEdgedWidth;
1182 :     Data.currentMV = currentMV;
1183 :     Data.currentQMV = currentQMV;
1184 :     Data.iMinSAD = iMinSAD;
1185 :     Data.temp = temp;
1186 : syskin 1132 Data.dir = &dir;
1187 : edgomez 851 Data.iFcode = current->fcode;
1188 :     Data.rounding = pParam->m_rounding_type;
1189 : edgomez 963 Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0);
1190 : edgomez 1107 Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP;
1191 : chl 1077 Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0;
1192 : edgomez 962 Data.dctSpace = dct_space;
1193 : edgomez 1022 Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT);
1194 : edgomez 195
1195 : edgomez 949 if ((current->vop_flags & XVID_VOP_REDUCED)) {
1196 : edgomez 851 mb_width = (pParam->width + 31) / 32;
1197 :     mb_height = (pParam->height + 31) / 32;
1198 :     Data.qpel = 0;
1199 :     }
1200 : chl 181
1201 : edgomez 1053 Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */
1202 : edgomez 851 if (sadInit) (*sadInit) ();
1203 : chl 181
1204 : edgomez 851 for (y = 0; y < mb_height; y++) {
1205 :     for (x = 0; x < mb_width; x++) {
1206 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1207 : chl 181
1208 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1209 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1210 :     pRef->y + (x + y * iEdgedWidth) * 16,
1211 :     pParam->edged_width, pMB->sad8 );
1212 : chl 181
1213 : edgomez 851 else pMB->sad16 =
1214 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1215 :     pRef->y + (x + y * iEdgedWidth) * 32,
1216 :     pParam->edged_width, pMB->sad8 );
1217 : chl 181
1218 : edgomez 851 if (Data.chroma) {
1219 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1220 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1221 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1222 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1223 :     pMB->sad16 += Data.temp[7];
1224 :     }
1225 :    
1226 :     sad00 = pMB->sad16;
1227 :    
1228 : edgomez 1053 /* initial skip decision */
1229 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1230 : syskin 1084 if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */
1231 : edgomez 953 if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh)
1232 : edgomez 851 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1233 : syskin 1108 ZeroMacroblockP(pMB, sad00);
1234 :     pMB->mode = MODE_NOT_CODED;
1235 : edgomez 851 continue;
1236 :     }
1237 :     }
1238 :    
1239 : Isibaar 1095 if ((current->vop_flags & XVID_VOP_CARTOON) &&
1240 :     (sad00 < pMB->quant * 4 * skip_thresh)) { /* favorize (0,0) vector for cartoons */
1241 :     ZeroMacroblockP(pMB, sad00);
1242 :     continue;
1243 :     }
1244 :    
1245 : edgomez 851 SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1246 : edgomez 1022 y, MotionFlags, current->vop_flags, current->vol_flags,
1247 :     &Data, pParam, pMBs, reference->mbs, pMB);
1248 : edgomez 851
1249 : edgomez 1022 ModeDecision(&Data, pMB, pMBs, x, y, pParam,
1250 :     MotionFlags, current->vop_flags, current->vol_flags,
1251 : syskin 1084 pCurrent, pRef, pGMC, current->coding_type);
1252 : edgomez 1022
1253 : edgomez 851 if (pMB->mode == MODE_INTRA)
1254 :     if (++iIntra > iLimit) return 1;
1255 : chl 181 }
1256 :     }
1257 : edgomez 851
1258 :     return 0;
1259 : chl 181 }
1260 :    
1261 : syskin 1132 /* check if given vector is equal to any vector checked before */
1262 :     static __inline int
1263 :     vector_repeats(const VECTOR * const pmv, const int i)
1264 :     {
1265 :     unsigned int j;
1266 :     for (j = 0; j < i; j++)
1267 :     if (MVequal(pmv[i], pmv[j])) return 1; /* same vector has been checked already */
1268 :     return 0;
1269 :     }
1270 : chl 181
1271 : syskin 1132 /* make a binary mask that prevents diamonds/squares
1272 :     from checking a vector which has been checked as a prediction */
1273 : edgomez 851 static __inline int
1274 : syskin 1132 make_mask(const VECTOR * const pmv, const int i, const int current)
1275 : chl 96 {
1276 : syskin 1132 unsigned int mask = 255, j;
1277 : edgomez 851 for (j = 0; j < i; j++) {
1278 : syskin 1132 if (pmv[current].x == pmv[j].x) {
1279 :     if (pmv[current].y == pmv[j].y + iDiamondSize) mask &= ~4;
1280 :     else if (pmv[current].y == pmv[j].y - iDiamondSize) mask &= ~8;
1281 : edgomez 851 } else
1282 : syskin 1132 if (pmv[current].y == pmv[j].y) {
1283 :     if (pmv[current].x == pmv[j].x + iDiamondSize) mask &= ~1;
1284 :     else if (pmv[current].x == pmv[j].x - iDiamondSize) mask &= ~2;
1285 : edgomez 851 }
1286 :     }
1287 :     return mask;
1288 :     }
1289 : edgomez 195
1290 : edgomez 851 static __inline void
1291 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1292 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1293 :     {
1294 : edgomez 1053 /* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */
1295 : edgomez 851 if (rrv) { iWcount /= 2; iHcount /= 2; }
1296 : edgomez 195
1297 : edgomez 1053 if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */
1298 : edgomez 851 pmv[5].x = EVEN(pmv[3].x);
1299 :     pmv[5].y = EVEN(pmv[3].y);
1300 :     } else pmv[5].x = pmv[5].y = 0;
1301 : chl 96
1302 : edgomez 1053 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */
1303 : edgomez 851 else pmv[3].x = pmv[3].y = 0;
1304 : chl 96
1305 : edgomez 1053 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */
1306 : edgomez 851 else pmv[4].x = pmv[4].y = 0;
1307 : Isibaar 3
1308 : edgomez 1053 /* [1] median prediction */
1309 : edgomez 851 pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1310 : edgomez 195
1311 : 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 */
1312 : edgomez 195
1313 : edgomez 1053 pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */
1314 : edgomez 851 pmv[2].y = EVEN(prevMB->mvs[0].y);
1315 :    
1316 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1317 : edgomez 1053 pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */
1318 : edgomez 851 pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1319 :     } else pmv[6].x = pmv[6].y = 0;
1320 :    
1321 :     if (rrv) {
1322 :     int i;
1323 :     for (i = 0; i < 7; i++) {
1324 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1325 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1326 :     }
1327 :     }
1328 : Isibaar 3 }
1329 :    
1330 : edgomez 851 static void
1331 :     SearchP(const IMAGE * const pRef,
1332 :     const uint8_t * const pRefH,
1333 :     const uint8_t * const pRefV,
1334 :     const uint8_t * const pRefHV,
1335 :     const IMAGE * const pCur,
1336 :     const int x,
1337 :     const int y,
1338 :     const uint32_t MotionFlags,
1339 : suxen_drol 890 const uint32_t VopFlags,
1340 : edgomez 1022 const uint32_t VolFlags,
1341 : edgomez 851 SearchData * const Data,
1342 :     const MBParam * const pParam,
1343 :     const MACROBLOCK * const pMBs,
1344 :     const MACROBLOCK * const prevMBs,
1345 :     MACROBLOCK * const pMB)
1346 :     {
1347 : Isibaar 3
1348 : syskin 1132 int i, threshA;
1349 : edgomez 851 VECTOR pmv[7];
1350 : edgomez 1022 int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
1351 : syskin 1132 CheckFunc * CheckCandidate;
1352 : Isibaar 3
1353 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
1354 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv);
1355 : edgomez 195
1356 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1357 : Isibaar 3
1358 : edgomez 1053 Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */
1359 : edgomez 851 i = Data->rrv ? 2 : 1;
1360 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1361 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1362 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1363 : chl 326
1364 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1365 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1366 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1367 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1368 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1369 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1370 : edgomez 195
1371 : edgomez 1022 Data->lambda16 = lambda_vec16[pMB->quant];
1372 :     Data->lambda8 = lambda_vec8[pMB->quant];
1373 : edgomez 851 Data->qpel_precision = 0;
1374 : syskin 1132 *Data->dir = 0;
1375 : edgomez 195
1376 : edgomez 963 memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1377 : chl 169
1378 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1379 :     else Data->predMV = pmv[0];
1380 : Isibaar 3
1381 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1382 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1383 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1384 :     Data->iMinSAD[2] = pMB->sad8[1];
1385 :     Data->iMinSAD[3] = pMB->sad8[2];
1386 :     Data->iMinSAD[4] = pMB->sad8[3];
1387 : chl 169
1388 : edgomez 1107 if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) {
1389 : edgomez 1053 threshA = Data->temp[0]; /* that's where we keep this SAD atm */
1390 : edgomez 851 if (threshA < 512) threshA = 512;
1391 :     else if (threshA > 1024) threshA = 1024;
1392 :     } else
1393 :     threshA = 512;
1394 : chl 169
1395 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1396 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1397 : chl 169
1398 : edgomez 851 if (!Data->rrv) {
1399 : syskin 1132 if (inter4v) CheckCandidate = CheckCandidate16;
1400 : edgomez 1053 else CheckCandidate = CheckCandidate16no4v; /* for extra speed */
1401 : edgomez 851 } else CheckCandidate = CheckCandidate32;
1402 : chl 169
1403 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1404 : chl 169
1405 : syskin 1132 for (i = 1; i < 7; i++)
1406 :     if (!vector_repeats(pmv, i)) {
1407 :     CheckCandidate(pmv[i].x, pmv[i].y, Data, i);
1408 :     if (Data->iMinSAD[0] <= threshA) { i++; break; }
1409 :     }
1410 : chl 169
1411 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1412 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1413 : edgomez 1022 (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16)))
1414 :     inter4v = 0;
1415 : edgomez 851 else {
1416 : Isibaar 3
1417 : edgomez 851 MainSearchFunc * MainSearchPtr;
1418 : syskin 1132 int mask = make_mask(pmv, i, *Data->dir); // all vectors pmv[0..i-1] have been checked
1419 :    
1420 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1421 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1422 : edgomez 851 else MainSearchPtr = DiamondSearch;
1423 : Isibaar 3
1424 : syskin 1132 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate);
1425 : Isibaar 3
1426 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1427 :     note that this search is/might be done in halfpel positions,
1428 :     which makes it more different than the diamond above */
1429 : chl 169
1430 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH16) {
1431 : edgomez 851 int32_t bSAD;
1432 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1433 :     if (Data->rrv) {
1434 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1435 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1436 :     }
1437 :     if (!(MVequal(startMV, backupMV))) {
1438 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1439 : edgomez 195
1440 : syskin 1132 CheckCandidate(startMV.x, startMV.y, Data, 255);
1441 :     MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate);
1442 : edgomez 851 if (bSAD < Data->iMinSAD[0]) {
1443 :     Data->currentMV[0] = backupMV;
1444 :     Data->iMinSAD[0] = bSAD; }
1445 :     }
1446 : chl 169
1447 : edgomez 851 backupMV = Data->currentMV[0];
1448 :     startMV.x = startMV.y = 1;
1449 :     if (!(MVequal(startMV, backupMV))) {
1450 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1451 : chl 169
1452 : syskin 1132 CheckCandidate(startMV.x, startMV.y, Data, 255);
1453 :     MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate);
1454 : edgomez 851 if (bSAD < Data->iMinSAD[0]) {
1455 :     Data->currentMV[0] = backupMV;
1456 : syskin 1132 Data->iMinSAD[0] = bSAD;
1457 :     }
1458 : edgomez 195 }
1459 : edgomez 851 }
1460 :     }
1461 : edgomez 195
1462 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE16)
1463 : syskin 1132 SubpelRefine(Data, CheckCandidate);
1464 : edgomez 195
1465 : edgomez 851 for(i = 0; i < 5; i++) {
1466 : edgomez 1053 Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */
1467 : edgomez 851 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1468 :     }
1469 : Isibaar 3
1470 : edgomez 1023 if (Data->qpel) {
1471 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
1472 :     pParam->width, pParam->height, Data->iFcode, 2, 0);
1473 : edgomez 987 Data->qpel_precision = 1;
1474 : edgomez 1023 if (MotionFlags & XVID_ME_QUARTERPELREFINE16)
1475 : syskin 1132 SubpelRefine(Data, CheckCandidate);
1476 : edgomez 953 }
1477 : Isibaar 3
1478 : edgomez 1023 if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30)
1479 : edgomez 1022 inter4v = 0;
1480 : Isibaar 3
1481 : edgomez 987 if (inter4v) {
1482 : edgomez 851 SearchData Data8;
1483 : edgomez 1053 memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */
1484 : Isibaar 3
1485 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1486 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1487 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1488 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1489 : chl 181
1490 : edgomez 1107 if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) {
1491 : edgomez 1053 /* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */
1492 : edgomez 851 int sumx = 0, sumy = 0;
1493 : Isibaar 3
1494 : edgomez 1022 if (Data->qpel)
1495 :     for (i = 1; i < 5; i++) {
1496 :     sumx += Data->currentQMV[i].x/2;
1497 :     sumy += Data->currentQMV[i].y/2;
1498 :     }
1499 :     else
1500 :     for (i = 1; i < 5; i++) {
1501 :     sumx += Data->currentMV[i].x;
1502 :     sumy += Data->currentMV[i].y;
1503 :     }
1504 : chl 259
1505 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1506 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1507 :     }
1508 : edgomez 1022 } else Data->iMinSAD[1] = 4096*256;
1509 : edgomez 851 }
1510 :    
1511 :     static void
1512 :     Search8(const SearchData * const OldData,
1513 :     const int x, const int y,
1514 :     const uint32_t MotionFlags,
1515 :     const MBParam * const pParam,
1516 :     MACROBLOCK * const pMB,
1517 :     const MACROBLOCK * const pMBs,
1518 :     const int block,
1519 :     SearchData * const Data)
1520 :     {
1521 :     int i = 0;
1522 : syskin 1132 CheckFunc * CheckCandidate;
1523 : edgomez 851 Data->iMinSAD = OldData->iMinSAD + 1 + block;
1524 :     Data->currentMV = OldData->currentMV + 1 + block;
1525 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1526 :    
1527 :     if(Data->qpel) {
1528 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1529 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1530 :     Data->predMV, Data->iFcode, 0, 0);
1531 :     } else {
1532 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1533 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1534 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1535 : Isibaar 3 }
1536 :    
1537 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1538 : Isibaar 3
1539 : edgomez 949 if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) {
1540 : Isibaar 3
1541 : edgomez 978 if (Data->rrv) i = 16; else i = 8;
1542 : Isibaar 3
1543 : edgomez 978 Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1544 :     Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1545 :     Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1546 :     Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1547 :    
1548 :     Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1549 : edgomez 851 Data->qpel_precision = 0;
1550 : Isibaar 3
1551 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3,
1552 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv);
1553 : Isibaar 3
1554 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1555 :     else CheckCandidate = CheckCandidate16no4v;
1556 : Isibaar 3
1557 : edgomez 1107 if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) {
1558 : edgomez 1053 int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */
1559 : Isibaar 3
1560 : edgomez 851 MainSearchFunc *MainSearchPtr;
1561 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch;
1562 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1563 : edgomez 851 else MainSearchPtr = DiamondSearch;
1564 : Isibaar 3
1565 : syskin 1132 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidate);
1566 : Isibaar 3
1567 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1568 : edgomez 1053 Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1569 : edgomez 851 Data->currentQMV->y = 2 * Data->currentMV->y;
1570 :     }
1571 :     }
1572 : edgomez 195
1573 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8) {
1574 : edgomez 1053 int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */
1575 : edgomez 195
1576 : syskin 1132 SubpelRefine(Data, CheckCandidate); /* perform halfpel refine of current best vector */
1577 : Isibaar 3
1578 : edgomez 1053 if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */
1579 :     Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1580 : edgomez 851 Data->currentQMV->y = 2 * Data->currentMV->y;
1581 :     }
1582 :     }
1583 : Isibaar 3
1584 : edgomez 949 if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) {
1585 : edgomez 851 Data->qpel_precision = 1;
1586 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3,
1587 :     pParam->width, pParam->height, Data->iFcode, 2, 0);
1588 : syskin 1132 SubpelRefine(Data, CheckCandidate);
1589 : edgomez 851 }
1590 :     }
1591 : edgomez 195
1592 : edgomez 851 if (Data->rrv) {
1593 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1594 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1595 :     }
1596 :    
1597 :     if(Data->qpel) {
1598 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1599 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1600 :     pMB->qmvs[block] = *Data->currentQMV;
1601 : edgomez 195 } else {
1602 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1603 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1604 : edgomez 78 }
1605 : edgomez 851
1606 :     pMB->mvs[block] = *Data->currentMV;
1607 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1608 : Isibaar 3 }
1609 :    
1610 : edgomez 851 /* motion estimation for B-frames */
1611 : chl 345
1612 : edgomez 851 static __inline VECTOR
1613 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1614 :     {
1615 :     /* the stupidiest function ever */
1616 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1617 :     }
1618 : chl 345
1619 : edgomez 851 static void __inline
1620 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1621 :     const uint32_t iWcount,
1622 :     const MACROBLOCK * const pMB,
1623 :     const uint32_t mode_curr)
1624 : chl 345 {
1625 :    
1626 : edgomez 1053 /* [0] is prediction */
1627 : edgomez 851 pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1628 : chl 345
1629 : edgomez 1053 pmv[1].x = pmv[1].y = 0; /* [1] is zero */
1630 : chl 345
1631 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1632 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1633 : chl 345
1634 : edgomez 1053 if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */
1635 : edgomez 851 pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1636 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1637 :     } else pmv[3].x = pmv[3].y = 0;
1638 : chl 345
1639 : edgomez 851 if (y != 0) {
1640 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1641 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1642 :     } else pmv[4].x = pmv[4].y = 0;
1643 : chl 345
1644 : edgomez 851 if (x != 0) {
1645 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1646 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1647 :     } else pmv[5].x = pmv[5].y = 0;
1648 : chl 345
1649 : edgomez 851 if (x != 0 && y != 0) {
1650 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1651 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1652 :     } else pmv[6].x = pmv[6].y = 0;
1653 :     }
1654 : chl 345
1655 :    
1656 : edgomez 851 /* search backward or forward */
1657 :     static void
1658 :     SearchBF( const IMAGE * const pRef,
1659 :     const uint8_t * const pRefH,
1660 :     const uint8_t * const pRefV,
1661 :     const uint8_t * const pRefHV,
1662 :     const IMAGE * const pCur,
1663 :     const int x, const int y,
1664 :     const uint32_t MotionFlags,
1665 :     const uint32_t iFcode,
1666 :     const MBParam * const pParam,
1667 :     MACROBLOCK * const pMB,
1668 :     const VECTOR * const predMV,
1669 :     int32_t * const best_sad,
1670 :     const int32_t mode_current,
1671 :     SearchData * const Data)
1672 :     {
1673 : chl 345
1674 : syskin 1135 int i;
1675 : edgomez 851 VECTOR pmv[7];
1676 :     MainSearchFunc *MainSearchPtr;
1677 :     *Data->iMinSAD = MV_MAX_ERROR;
1678 :     Data->iFcode = iFcode;
1679 :     Data->qpel_precision = 0;
1680 : edgomez 1053 Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */
1681 : chl 345
1682 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16;
1683 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16;
1684 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16;
1685 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16;
1686 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1687 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1688 : chl 345
1689 : edgomez 851 Data->predMV = *predMV;
1690 : chl 345
1691 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
1692 :     pParam->width, pParam->height, iFcode - Data->qpel, 1, 0);
1693 : chl 345
1694 : edgomez 851 pmv[0] = Data->predMV;
1695 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1696 : chl 345
1697 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1698 : chl 345
1699 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1700 : chl 345
1701 : edgomez 1053 /* main loop. checking all predictions */
1702 : syskin 1132 for (i = 0; i < 7; i++)
1703 :     if (!vector_repeats(pmv, i) )
1704 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, Data, i);
1705 : chl 345
1706 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1707 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1708 : edgomez 851 else MainSearchPtr = DiamondSearch;
1709 : chl 345
1710 : syskin 1135 if (*Data->iMinSAD > 512) {
1711 :     unsigned int mask = make_mask(pmv, 7, *Data->dir);
1712 :     MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v);
1713 :     }
1714 : chl 345
1715 : syskin 1132 SubpelRefine(Data, CheckCandidate16no4v);
1716 :    
1717 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1718 :     Data->currentQMV->x = 2*Data->currentMV->x;
1719 :     Data->currentQMV->y = 2*Data->currentMV->y;
1720 :     Data->qpel_precision = 1;
1721 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
1722 :     pParam->width, pParam->height, iFcode, 2, 0);
1723 : syskin 1132 SubpelRefine(Data, CheckCandidate16no4v);
1724 : edgomez 851 }
1725 : chl 345
1726 : edgomez 1053 /* three bits are needed to code backward mode. four for forward */
1727 : Isibaar 3
1728 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1729 :     else *Data->iMinSAD += 3 * Data->lambda16;
1730 : edgomez 195
1731 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1732 :     *best_sad = *Data->iMinSAD;
1733 :     pMB->mode = mode_current;
1734 :     if (Data->qpel) {
1735 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1736 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1737 :     if (mode_current == MODE_FORWARD)
1738 :     pMB->qmvs[0] = *Data->currentQMV;
1739 :     else
1740 :     pMB->b_qmvs[0] = *Data->currentQMV;
1741 :     } else {
1742 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1743 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1744 :     }
1745 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1746 :     else pMB->b_mvs[0] = *Data->currentMV;
1747 :     }
1748 : edgomez 195
1749 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1750 : edgomez 1053 else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */
1751 : Isibaar 3 }
1752 :    
1753 : edgomez 851 static void
1754 :     SkipDecisionB(const IMAGE * const pCur,
1755 :     const IMAGE * const f_Ref,
1756 :     const IMAGE * const b_Ref,
1757 :     MACROBLOCK * const pMB,
1758 :     const uint32_t x, const uint32_t y,
1759 :     const SearchData * const Data)
1760 :     {
1761 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1762 :     int32_t sum;
1763 :     int k;
1764 :     const uint32_t stride = Data->iEdgedWidth/2;
1765 : edgomez 1053 /* this is not full chroma compensation, only it's fullpel approximation. should work though */
1766 : Isibaar 3
1767 : edgomez 851 for (k = 0; k < 4; k++) {
1768 : syskin 1135 dy += Data->directmvF[k].y >> Data->qpel;
1769 :     dx += Data->directmvF[k].x >> Data->qpel;
1770 :     b_dy += Data->directmvB[k].y >> Data->qpel;
1771 :     b_dx += Data->directmvB[k].x >> Data->qpel;
1772 : edgomez 851 }
1773 : Isibaar 3
1774 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1775 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1776 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1777 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1778 : Isibaar 3
1779 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1780 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1781 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1782 :     stride);
1783 : Isibaar 3
1784 : syskin 1135 if (sum >= MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) return; /* no skip */
1785 : Isibaar 3
1786 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1787 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1788 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1789 :     stride);
1790 : edgomez 195
1791 : syskin 1135 if (sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) {
1792 : edgomez 1053 pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */
1793 : edgomez 978 for (k = 0; k < 4; k++) {
1794 : edgomez 963 pMB->qmvs[k] = pMB->mvs[k];
1795 :     pMB->b_qmvs[k] = pMB->b_mvs[k];
1796 :     }
1797 :     }
1798 : edgomez 851 }
1799 : Isibaar 3
1800 : edgomez 851 static __inline uint32_t
1801 :     SearchDirect(const IMAGE * const f_Ref,
1802 :     const uint8_t * const f_RefH,
1803 :     const uint8_t * const f_RefV,
1804 :     const uint8_t * const f_RefHV,
1805 :     const IMAGE * const b_Ref,
1806 :     const uint8_t * const b_RefH,
1807 :     const uint8_t * const b_RefV,
1808 :     const uint8_t * const b_RefHV,
1809 :     const IMAGE * const pCur,
1810 :     const int x, const int y,
1811 :     const uint32_t MotionFlags,
1812 :     const int32_t TRB, const int32_t TRD,
1813 :     const MBParam * const pParam,
1814 :     MACROBLOCK * const pMB,
1815 :     const MACROBLOCK * const b_mb,
1816 :     int32_t * const best_sad,
1817 :     SearchData * const Data)
1818 : Isibaar 3
1819 : edgomez 851 {
1820 :     int32_t skip_sad;
1821 :     int k = (x + Data->iEdgedWidth*y) * 16;
1822 :     MainSearchFunc *MainSearchPtr;
1823 : syskin 1132 CheckFunc * CheckCandidate;
1824 : Isibaar 3
1825 : edgomez 851 *Data->iMinSAD = 256*4096;
1826 : edgomez 978 Data->RefP[0] = f_Ref->y + k;
1827 :     Data->RefP[2] = f_RefH + k;
1828 :     Data->RefP[1] = f_RefV + k;
1829 :     Data->RefP[3] = f_RefHV + k;
1830 :     Data->b_RefP[0] = b_Ref->y + k;
1831 :     Data->b_RefP[2] = b_RefH + k;
1832 :     Data->b_RefP[1] = b_RefV + k;
1833 :     Data->b_RefP[3] = b_RefHV + k;
1834 :     Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1835 :     Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1836 :     Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1837 :     Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1838 : edgomez 195
1839 : edgomez 851 k = Data->qpel ? 4 : 2;
1840 :     Data->max_dx = k * (pParam->width - x * 16);
1841 :     Data->max_dy = k * (pParam->height - y * 16);
1842 :     Data->min_dx = -k * (16 + x * 16);
1843 :     Data->min_dy = -k * (16 + y * 16);
1844 : chl 181
1845 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1846 :     Data->qpel_precision = 0;
1847 : edgomez 170
1848 : edgomez 851 for (k = 0; k < 4; k++) {
1849 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1850 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1851 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1852 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1853 : Isibaar 3
1854 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1855 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1856 :    
1857 : edgomez 1053 *best_sad = 256*4096; /* in that case, we won't use direct mode */
1858 :     pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */
1859 : edgomez 851 pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1860 :     return 256*4096;
1861 :     }
1862 :     if (b_mb->mode != MODE_INTER4V) {
1863 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1864 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1865 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1866 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1867 :     break;
1868 :     }
1869 : edgomez 195 }
1870 : Isibaar 3
1871 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1872 : Isibaar 3
1873 : syskin 1132 CheckCandidate(0, 0, Data, 255);
1874 : edgomez 195
1875 : edgomez 1053 /* initial (fast) skip decision */
1876 : syskin 1135 if (*Data->iMinSAD < Data->iQuant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) {
1877 : edgomez 1053 /* possible skip */
1878 : edgomez 851 if (Data->chroma) {
1879 :     pMB->mode = MODE_DIRECT_NONE_MV;
1880 : edgomez 1053 return *Data->iMinSAD; /* skip. */
1881 : edgomez 851 } else {
1882 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1883 : edgomez 1053 if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */
1884 : edgomez 851 }
1885 : Isibaar 3 }
1886 :    
1887 : edgomez 953 *Data->iMinSAD += Data->lambda16;
1888 : edgomez 851 skip_sad = *Data->iMinSAD;
1889 : syskin 1084
1890 : edgomez 1053 /*
1891 :     * DIRECT MODE DELTA VECTOR SEARCH.
1892 :     * This has to be made more effective, but at the moment I'm happy it's running at all
1893 :     */
1894 : edgomez 195
1895 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1896 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1897 : edgomez 851 else MainSearchPtr = DiamondSearch;
1898 : Isibaar 3
1899 : syskin 1132 MainSearchPtr(0, 0, Data, 255, CheckCandidate);
1900 : Isibaar 3
1901 : syskin 1132 SubpelRefine(Data, CheckCandidate);
1902 : chl 181
1903 : edgomez 851 *best_sad = *Data->iMinSAD;
1904 : chl 181
1905 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1906 : edgomez 1053 else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */
1907 : chl 181
1908 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1909 : edgomez 195
1910 : edgomez 851 for (k = 0; k < 4; k++) {
1911 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1912 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1913 :     ? Data->directmvB[k].x
1914 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1915 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1916 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1917 :     ? Data->directmvB[k].y
1918 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1919 :     if (Data->qpel) {
1920 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1921 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1922 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1923 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1924 :     }
1925 : edgomez 195
1926 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1927 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1928 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1929 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1930 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1931 :     break;
1932 :     }
1933 : edgomez 78 }
1934 : edgomez 851 return skip_sad;
1935 :     }
1936 : Isibaar 3
1937 : edgomez 851 static void
1938 :     SearchInterpolate(const IMAGE * const f_Ref,
1939 :     const uint8_t * const f_RefH,
1940 :     const uint8_t * const f_RefV,
1941 :     const uint8_t * const f_RefHV,
1942 :     const IMAGE * const b_Ref,
1943 :     const uint8_t * const b_RefH,
1944 :     const uint8_t * const b_RefV,
1945 :     const uint8_t * const b_RefHV,
1946 :     const IMAGE * const pCur,
1947 :     const int x, const int y,
1948 :     const uint32_t fcode,
1949 :     const uint32_t bcode,
1950 :     const uint32_t MotionFlags,
1951 :     const MBParam * const pParam,
1952 :     const VECTOR * const f_predMV,
1953 :     const VECTOR * const b_predMV,
1954 :     MACROBLOCK * const pMB,
1955 :     int32_t * const best_sad,
1956 :     SearchData * const fData)
1957 : chl 96
1958 : edgomez 851 {
1959 : chl 169
1960 : syskin 1132 int i, j;
1961 : edgomez 851 SearchData bData;
1962 : chl 169
1963 : edgomez 851 fData->qpel_precision = 0;
1964 : edgomez 1053 memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */
1965 : edgomez 851 *fData->iMinSAD = 4096*256;
1966 :     bData.currentMV++; bData.currentQMV++;
1967 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1968 : chl 169
1969 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1970 : chl 169
1971 : edgomez 978 bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i;
1972 :     bData.b_RefP[2] = fData->RefP[2] = f_RefH + i;
1973 :     bData.b_RefP[1] = fData->RefP[1] = f_RefV + i;
1974 :     bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i;
1975 :     bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i;
1976 :     bData.RefP[2] = fData->b_RefP[2] = b_RefH + i;
1977 :     bData.RefP[1] = fData->b_RefP[1] = b_RefV + i;
1978 :     bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i;
1979 :     bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1980 :     bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1981 :     bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1982 :     bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1983 : syskin 1132 bData.dir = fData->dir;
1984 : chl 169
1985 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1986 :     fData->bpredMV = bData.predMV = *b_predMV;
1987 :     fData->currentMV[0] = fData->currentMV[2];
1988 : Isibaar 3
1989 : syskin 1129 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode - fData->qpel, 1, 0);
1990 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode - fData->qpel, 1, 0);
1991 : Isibaar 3
1992 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1993 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1994 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1995 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1996 : chl 169
1997 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1998 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1999 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
2000 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
2001 : Isibaar 3
2002 : syskin 1132 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, fData, 255);
2003 : chl 169
2004 : edgomez 1053 /* diamond */
2005 : edgomez 851 do {
2006 : syskin 1132 *fData->dir = 255;
2007 : edgomez 1053 /* forward MV moves */
2008 : edgomez 851 i = fData->currentMV[0].x; j = fData->currentMV[0].y;
2009 : chl 169
2010 : syskin 1132 CheckCandidateInt(i + 1, j, fData, 0);
2011 :     CheckCandidateInt(i, j + 1, fData, 0);
2012 :     CheckCandidateInt(i - 1, j, fData, 0);
2013 :     CheckCandidateInt(i, j - 1, fData, 0);
2014 : edgomez 195
2015 : edgomez 1053 /* backward MV moves */
2016 : edgomez 851 i = fData->currentMV[1].x; j = fData->currentMV[1].y;
2017 :     fData->currentMV[2] = fData->currentMV[0];
2018 : syskin 1132 CheckCandidateInt(i + 1, j, &bData, 0);
2019 :     CheckCandidateInt(i, j + 1, &bData, 0);
2020 :     CheckCandidateInt(i - 1, j, &bData, 0);
2021 :     CheckCandidateInt(i, j - 1, &bData, 0);
2022 : Isibaar 3
2023 : syskin 1132 } while (!(*fData->dir));
2024 : chl 140
2025 : edgomez 1053 /* qpel refinement */
2026 : edgomez 851 if (fData->qpel) {
2027 :     if (*fData->iMinSAD > *best_sad + 500) return;
2028 :     fData->qpel_precision = bData.qpel_precision = 1;
2029 : syskin 1129 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode, 2, 0);
2030 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode, 2, 0);
2031 : edgomez 851 fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
2032 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
2033 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
2034 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
2035 : syskin 1132 SubpelRefine(fData, CheckCandidateInt);
2036 : edgomez 851 if (*fData->iMinSAD > *best_sad + 300) return;
2037 :     fData->currentQMV[2] = fData->currentQMV[0];
2038 : syskin 1132 SubpelRefine(&bData, CheckCandidateInt);
2039 : edgomez 851 }
2040 : chl 140
2041 : edgomez 1053 *fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */
2042 : Isibaar 3
2043 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
2044 :     *best_sad = *fData->iMinSAD;
2045 :     pMB->mvs[0] = fData->currentMV[0];
2046 :     pMB->b_mvs[0] = fData->currentMV[1];
2047 :     pMB->mode = MODE_INTERPOLATE;
2048 :     if (fData->qpel) {
2049 :     pMB->qmvs[0] = fData->currentQMV[0];
2050 :     pMB->b_qmvs[0] = fData->currentQMV[1];
2051 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
2052 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
2053 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
2054 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
2055 :     } else {
2056 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
2057 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
2058 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
2059 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
2060 :     }
2061 : edgomez 78 }
2062 : edgomez 851 }
2063 : Isibaar 3
2064 : edgomez 851 void
2065 :     MotionEstimationBVOP(MBParam * const pParam,
2066 :     FRAMEINFO * const frame,
2067 :     const int32_t time_bp,
2068 :     const int32_t time_pp,
2069 : edgomez 1053 /* forward (past) reference */
2070 : edgomez 851 const MACROBLOCK * const f_mbs,
2071 :     const IMAGE * const f_ref,
2072 :     const IMAGE * const f_refH,
2073 :     const IMAGE * const f_refV,
2074 :     const IMAGE * const f_refHV,
2075 : edgomez 1053 /* backward (future) reference */
2076 : edgomez 851 const FRAMEINFO * const b_reference,
2077 :     const IMAGE * const b_ref,
2078 :     const IMAGE * const b_refH,
2079 :     const IMAGE * const b_refV,
2080 :     const IMAGE * const b_refHV)
2081 :     {
2082 :     uint32_t i, j;
2083 :     int32_t best_sad;
2084 :     uint32_t skip_sad;
2085 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2086 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
2087 : Isibaar 3
2088 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2089 : Isibaar 3
2090 : edgomez 851 const int32_t TRB = time_pp - time_bp;
2091 :     const int32_t TRD = time_pp;
2092 : edgomez 195
2093 : edgomez 1053 /* some pre-inintialized data for the rest of the search */
2094 : Isibaar 3
2095 : edgomez 851 SearchData Data;
2096 :     int32_t iMinSAD;
2097 : syskin 1132 uint32_t dir;
2098 : edgomez 851 VECTOR currentMV[3];
2099 :     VECTOR currentQMV[3];
2100 :     int32_t temp[8];
2101 :     memset(&Data, 0, sizeof(SearchData));
2102 :     Data.iEdgedWidth = pParam->edged_width;
2103 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2104 :     Data.iMinSAD = &iMinSAD;
2105 : syskin 1135 Data.lambda16 = lambda_vec16[MAX(frame->quant-2, 2)];
2106 : syskin 1112 Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0;
2107 : edgomez 851 Data.rounding = 0;
2108 : edgomez 1107 Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP;
2109 : edgomez 851 Data.temp = temp;
2110 : syskin 1132 Data.dir = &dir;
2111 : syskin 1135 Data.iQuant = frame->quant;
2112 : Isibaar 3
2113 : edgomez 1053 Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */
2114 :    
2115 :     /* note: i==horizontal, j==vertical */
2116 : edgomez 851 for (j = 0; j < pParam->mb_height; j++) {
2117 : edgomez 195
2118 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2119 : Isibaar 3
2120 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2121 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2122 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2123 : edgomez 195
2124 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2125 :     if (b_reference->coding_type != S_VOP)
2126 :     if (b_mb->mode == MODE_NOT_CODED) {
2127 :     pMB->mode = MODE_NOT_CODED;
2128 :     continue;
2129 :     }
2130 : Isibaar 3
2131 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2132 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2133 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2134 : Isibaar 3
2135 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2136 :     and (2) sets very good predictions for forward and backward search */
2137 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2138 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2139 :     &frame->image,
2140 :     i, j,
2141 :     frame->motion_flags,
2142 :     TRB, TRD,
2143 :     pParam,
2144 :     pMB, b_mb,
2145 :     &best_sad,
2146 :     &Data);
2147 : Isibaar 3
2148 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2149 : edgomez 195
2150 : edgomez 1053 /* forward search */
2151 : edgomez 851 SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2152 :     &frame->image, i, j,
2153 :     frame->motion_flags,
2154 :     frame->fcode, pParam,
2155 :     pMB, &f_predMV, &best_sad,
2156 :     MODE_FORWARD, &Data);
2157 : edgomez 195
2158 : edgomez 1053 /* backward search */
2159 : edgomez 851 SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2160 :     &frame->image, i, j,
2161 :     frame->motion_flags,
2162 :     frame->bcode, pParam,
2163 :     pMB, &b_predMV, &best_sad,
2164 :     MODE_BACKWARD, &Data);
2165 :    
2166 : edgomez 1053 /* interpolate search comes last, because it uses data from forward and backward as prediction */
2167 : edgomez 851 SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2168 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2169 :     &frame->image,
2170 :     i, j,
2171 :     frame->fcode, frame->bcode,
2172 :     frame->motion_flags,
2173 :     pParam,
2174 :     &f_predMV, &b_predMV,
2175 :     pMB, &best_sad,
2176 :     &Data);
2177 :    
2178 : edgomez 1053 /* final skip decision */
2179 : syskin 1135 if ( (skip_sad < Data.iQuant * MAX_SAD00_FOR_SKIP * 2)
2180 : edgomez 851 && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2181 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2182 :    
2183 :     switch (pMB->mode) {
2184 :     case MODE_FORWARD:
2185 :     f_count++;
2186 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2187 :     break;
2188 :     case MODE_BACKWARD:
2189 :     b_count++;
2190 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2191 :     break;
2192 :     case MODE_INTERPOLATE:
2193 :     i_count++;
2194 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2195 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2196 :     break;
2197 :     case MODE_DIRECT:
2198 :     case MODE_DIRECT_NO4V:
2199 :     d_count++;
2200 :     default:
2201 :     break;
2202 :     }
2203 :     }
2204 :     }
2205 : Isibaar 3 }
2206 : chl 96
2207 : edgomez 851 static __inline void
2208 :     MEanalyzeMB ( const uint8_t * const pRef,
2209 :     const uint8_t * const pCur,
2210 :     const int x,
2211 :     const int y,
2212 :     const MBParam * const pParam,
2213 :     MACROBLOCK * const pMBs,
2214 :     SearchData * const Data)
2215 : chl 96 {
2216 :    
2217 : syskin 1132 int i;
2218 : edgomez 851 VECTOR pmv[3];
2219 : edgomez 953 MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2220 : chl 96
2221 : syskin 1135 unsigned int simplicity = 0;
2222 :    
2223 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2224 : chl 96
2225 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2226 : syskin 1135 pParam->width, pParam->height, Data->iFcode - Data->qpel - 1, 0, 0);
2227 : edgomez 195
2228 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2229 : edgomez 978 Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16;
2230 : chl 96
2231 : syskin 1135 pmv[0].x = pMB->mvs[0].x;
2232 :     pmv[0].y = pMB->mvs[0].y;
2233 : edgomez 195
2234 : syskin 1135 CheckCandidate32I(pmv[0].x, pmv[0].y, Data, 0);
2235 : edgomez 195
2236 : syskin 1135 if (*Data->iMinSAD > 200) {
2237 : chl 96
2238 : syskin 1135 pmv[1].x = pmv[1].y = 0;
2239 :    
2240 :     /* median is only used as prediction. it doesn't have to be real */
2241 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2242 :     else
2243 :     if (x == 1) /* left macroblock does not have any vector now */
2244 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */
2245 :     else if (y == 1) /* top macroblock doesn't have it's vector */
2246 :     Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */
2247 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */
2248 :    
2249 :     pmv[2].x = Data->predMV.x;
2250 :     pmv[2].y = Data->predMV.y;
2251 :    
2252 : syskin 1133 if (!vector_repeats(pmv, 1))
2253 : syskin 1132 CheckCandidate32I(pmv[1].x, pmv[1].y, Data, 1);
2254 : syskin 1133 if (!vector_repeats(pmv, 2))
2255 : syskin 1132 CheckCandidate32I(pmv[2].x, pmv[2].y, Data, 2);
2256 : chl 96
2257 : syskin 1135 if (*Data->iMinSAD > 500) { /* diamond only if needed */
2258 : syskin 1133 unsigned int mask = make_mask(pmv, 3, *Data->dir);
2259 : syskin 1134 DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate32I);
2260 : syskin 1135 } else simplicity++;
2261 : edgomez 851
2262 : syskin 1135 if (*Data->iMinSAD > 500) /* refinement from 2-pixel to 1-pixel */
2263 :     SubpelRefine(Data, CheckCandidate32I);
2264 :     else simplicity++;
2265 :     } else simplicity++;
2266 :    
2267 : edgomez 953 for (i = 0; i < 4; i++) {
2268 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2269 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2270 :     MB->mode = MODE_INTER;
2271 : syskin 1135 /* if we skipped some search steps, we have to assume that SAD would be lower with them */
2272 :     MB->sad16 = Data->iMinSAD[i+1] - (simplicity<<7);
2273 : chl 96 }
2274 : edgomez 851 }
2275 : chl 96
2276 : syskin 1075 #define INTRA_THRESH 2200
2277 : syskin 1135 #define INTER_THRESH 40
2278 : syskin 1075 #define INTRA_THRESH2 95
2279 : chl 96
2280 : edgomez 851 int
2281 :     MEanalysis( const IMAGE * const pRef,
2282 : edgomez 953 const FRAMEINFO * const Current,
2283 :     const MBParam * const pParam,
2284 : syskin 1075 const int maxIntra, //maximum number if non-I frames
2285 :     const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2286 : syskin 1084 const int bCount, // number of B frames in a row
2287 : edgomez 953 const int b_thresh)
2288 : edgomez 851 {
2289 :     uint32_t x, y, intra = 0;
2290 :     int sSAD = 0;
2291 :     MACROBLOCK * const pMBs = Current->mbs;
2292 :     const IMAGE * const pCurrent = &Current->image;
2293 : syskin 1075 int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh;
2294 : chl 1077 int blocks = 0;
2295 : syskin 1075 int complexity = 0;
2296 : chl 96
2297 : edgomez 851 int32_t iMinSAD[5], temp[5];
2298 : syskin 1132 uint32_t dir;
2299 : edgomez 851 VECTOR currentMV[5];
2300 :     SearchData Data;
2301 :     Data.iEdgedWidth = pParam->edged_width;
2302 :     Data.currentMV = currentMV;
2303 :     Data.iMinSAD = iMinSAD;
2304 :     Data.iFcode = Current->fcode;
2305 :     Data.temp = temp;
2306 : syskin 1132 Data.dir = &dir;
2307 :     Data.qpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0;
2308 : syskin 1135 Data.qpel_precision = 0;
2309 : chl 96
2310 : edgomez 1023 if (intraCount != 0) {
2311 : syskin 1075 if (intraCount < 10) // we're right after an I frame
2312 : edgomez 1023 IntraThresh += 15* (intraCount - 10) * (intraCount - 10);
2313 :     else
2314 : syskin 1075 if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2315 : edgomez 1023 IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra;
2316 :     }
2317 : chl 96
2318 : syskin 1135 InterThresh -= 20 * bCount;
2319 :     if (InterThresh < 10 + b_thresh) InterThresh = 10 + b_thresh;
2320 : chl 326
2321 : edgomez 851 if (sadInit) (*sadInit) ();
2322 : chl 96
2323 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2324 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2325 :     int i;
2326 : syskin 1075 blocks += 10;
2327 : edgomez 195
2328 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2329 : syskin 1075 else { //extrapolation of the vector found for last frame
2330 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].x =
2331 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2332 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].y =
2333 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2334 :     }
2335 : edgomez 195
2336 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2337 : edgomez 195
2338 : edgomez 851 for (i = 0; i < 4; i++) {
2339 :     int dev;
2340 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2341 : syskin 1075 dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2342 :     pParam->edged_width);
2343 : syskin 1084
2344 : syskin 1108 complexity += MAX(dev, 300);
2345 : syskin 1075 if (dev + IntraThresh < pMB->sad16) {
2346 :     pMB->mode = MODE_INTRA;
2347 :     if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2348 : edgomez 851 }
2349 : edgomez 953
2350 : syskin 1084 if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0)
2351 : syskin 1135 if (dev > 1000 && pMB->sad16 < 1000)
2352 : syskin 1075 sSAD += 1000;
2353 :    
2354 : syskin 1135 sSAD += (dev < 4000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */
2355 : edgomez 851 }
2356 :     }
2357 : edgomez 195 }
2358 : syskin 1075 complexity >>= 7;
2359 : edgomez 953
2360 : syskin 1075 sSAD /= complexity + 4*blocks;
2361 : edgomez 953
2362 : syskin 1108 if (intraCount > 80 && sSAD > INTRA_THRESH2 ) return I_VOP;
2363 : edgomez 851 if (sSAD > InterThresh ) return P_VOP;
2364 :     emms();
2365 :     return B_VOP;
2366 :     }
2367 : chl 96
2368 :    
2369 : edgomez 1053 /* functions which perform BITS-based search/bitcount */
2370 : chl 96
2371 : edgomez 851 static int
2372 : syskin 1108 findRDinter(SearchData * const Data,
2373 :     const MACROBLOCK * const pMBs, const int x, const int y,
2374 :     const MBParam * const pParam,
2375 :     const uint32_t MotionFlags)
2376 : chl 96 {
2377 : syskin 1132 int i;
2378 : edgomez 851 int32_t bsad[5];
2379 : chl 141
2380 : edgomez 851 if (Data->qpel) {
2381 :     for(i = 0; i < 5; i++) {
2382 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2383 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2384 :     }
2385 :     Data->qpel_precision = 1;
2386 : syskin 1132 CheckCandidateRD16(Data->currentQMV[0].x, Data->currentQMV[0].y, Data, 255);
2387 : chl 96
2388 : edgomez 1107 if (MotionFlags & (XVID_ME_HALFPELREFINE16_RD | XVID_ME_EXTSEARCH_RD)) { /* we have to prepare for halfpixel-precision search */
2389 : edgomez 851 for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2390 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2391 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv);
2392 : edgomez 851 Data->qpel_precision = 0;
2393 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2394 : syskin 1132 CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255);
2395 : edgomez 851 }
2396 : edgomez 195
2397 : edgomez 1053 } else { /* not qpel */
2398 : edgomez 195
2399 : syskin 1132 CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255);
2400 : edgomez 851 }
2401 : chl 96
2402 : syskin 1132 if (MotionFlags&XVID_ME_EXTSEARCH_RD)
2403 :     SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidateRD16);
2404 : chl 96
2405 : syskin 1132 if (MotionFlags&XVID_ME_HALFPELREFINE16_RD)
2406 :     SubpelRefine(Data, CheckCandidateRD16);
2407 : chl 96
2408 : edgomez 851 if (Data->qpel) {
2409 : edgomez 1107 if (MotionFlags&(XVID_ME_EXTSEARCH_RD | XVID_ME_HALFPELREFINE16_RD)) { /* there was halfpel-precision search */
2410 : edgomez 851 for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2411 : edgomez 1053 Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* we have found a better match */
2412 : edgomez 851 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2413 :     }
2414 : edgomez 195
2415 : edgomez 1053 /* preparing for qpel-precision search */
2416 : edgomez 851 Data->qpel_precision = 1;
2417 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2418 :     pParam->width, pParam->height, Data->iFcode, 2, 0);
2419 : edgomez 851 }
2420 : syskin 1132 if (MotionFlags&XVID_ME_QUARTERPELREFINE16_RD)
2421 :     SubpelRefine(Data, CheckCandidateRD16);
2422 : edgomez 851 }
2423 : chl 96
2424 : edgomez 1107 if (MotionFlags&XVID_ME_CHECKPREDICTION_RD) { /* let's check vector equal to prediction */
2425 : edgomez 851 VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2426 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2427 : syskin 1132 CheckCandidateRD16(Data->predMV.x, Data->predMV.y, Data, 255);
2428 : edgomez 851 }
2429 :     return Data->iMinSAD[0];
2430 :     }
2431 : chl 96
2432 : edgomez 851 static int
2433 : syskin 1108 findRDinter4v(const SearchData * const Data,
2434 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2435 :     const int x, const int y,
2436 :     const MBParam * const pParam, const uint32_t MotionFlags,
2437 :     const VECTOR * const backup)
2438 : edgomez 851 {
2439 : chl 96
2440 : syskin 1132 int cbp = 0, bits = 0, t = 0, i;
2441 : edgomez 851 SearchData Data2, *Data8 = &Data2;
2442 :     int sumx = 0, sumy = 0;
2443 : edgomez 962 int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64;
2444 : edgomez 1022 uint8_t * ptr;
2445 : chl 96
2446 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2447 : syskin 1132
2448 : edgomez 1053 for (i = 0; i < 4; i++) { /* for all luma blocks */
2449 : edgomez 1022
2450 : edgomez 851 Data8->iMinSAD = Data->iMinSAD + i + 1;
2451 :     Data8->currentMV = Data->currentMV + i + 1;
2452 :     Data8->currentQMV = Data->currentQMV + i + 1;
2453 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2454 : edgomez 978 Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2455 :     Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2456 :     Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2457 :     Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2458 : syskin 1084 *Data8->cbp = (Data->cbp[1] & (1<<(5-i))) ? 1:0; // copy corresponding cbp bit
2459 : edgomez 195
2460 : edgomez 851 if(Data->qpel) {
2461 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2462 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2463 :     Data8->predMV, Data8->iFcode, 0, 0);
2464 :     } else {
2465 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2466 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2467 :     Data8->predMV, Data8->iFcode, 0, 0);
2468 :     }
2469 : chl 96
2470 : syskin 1129 get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3,
2471 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel+1, 0);
2472 : chl 96
2473 : edgomez 1023 *Data8->iMinSAD += BITS_MULT*t;
2474 : chl 96
2475 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2476 : edgomez 1053 /* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */
2477 : edgomez 987 {
2478 :     VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV;
2479 : edgomez 1022 if (!MVequal (*v, backup[i+1]) )
2480 : syskin 1132 CheckCandidateRD8(backup[i+1].x, backup[i+1].y, Data8, 255);
2481 : edgomez 851 }
2482 : edgomez 195
2483 : edgomez 851 if (Data8->qpel) {
2484 : edgomez 1107 if (MotionFlags&XVID_ME_HALFPELREFINE8_RD || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_RD)) { /* halfpixel motion search follows */
2485 : edgomez 851 int32_t s = *Data8->iMinSAD;
2486 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2487 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2488 :     Data8->qpel_precision = 0;
2489 : syskin 1129 get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3,
2490 :     pParam->width, pParam->height, Data8->iFcode - 1, 1, 0);
2491 : edgomez 195
2492 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2493 : syskin 1132 CheckCandidateRD8(Data8->currentMV->x, Data8->currentMV->y, Data8, 255);
2494 : edgomez 195
2495 : edgomez 1107 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD)
2496 : syskin 1132 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8);
2497 : chl 96
2498 : edgomez 1107 if (MotionFlags & XVID_ME_HALFPELREFINE8_RD)
2499 : syskin 1132 SubpelRefine(Data8, CheckCandidateRD8);
2500 : chl 96
2501 : edgomez 1053 if(s > *Data8->iMinSAD) { /* we have found a better match */
2502 : edgomez 851 Data8->currentQMV->x = 2*Data8->currentMV->x;
2503 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2504 :     }
2505 : chl 96
2506 : edgomez 851 Data8->qpel_precision = 1;
2507 : syskin 1129 get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3,
2508 :     pParam->width, pParam->height, Data8->iFcode, 2, 0);
2509 : chl 96
2510 : edgomez 851 }
2511 : syskin 1132 if (MotionFlags & XVID_ME_QUARTERPELREFINE8_RD)
2512 :     SubpelRefine(Data8, CheckCandidateRD8);
2513 : edgomez 195
2514 : edgomez 1053 } else { /* not qpel */
2515 : edgomez 851
2516 : edgomez 1107 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) /* extsearch */
2517 : syskin 1132 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8);
2518 : edgomez 1022
2519 : edgomez 1107 if (MotionFlags & XVID_ME_HALFPELREFINE8_RD)
2520 : syskin 1132 SubpelRefine(Data8, CheckCandidateRD8); /* halfpel refinement */
2521 : edgomez 1022 }
2522 :    
2523 : edgomez 1053 /* checking vector equal to predicion */
2524 : edgomez 1107 if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_RD) {
2525 : edgomez 851 const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2526 : edgomez 1022 if (!MVequal(*v, Data8->predMV))
2527 : syskin 1132 CheckCandidateRD8(Data8->predMV.x, Data8->predMV.y, Data8, 255);
2528 : chl 141 }
2529 : edgomez 195
2530 : edgomez 851 bits += *Data8->iMinSAD;
2531 : edgomez 1053 if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */
2532 : edgomez 851
2533 : edgomez 1053 /* MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else */
2534 : edgomez 851 if(Data->qpel) {
2535 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2536 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2537 :     pMB->qmvs[i] = *Data8->currentQMV;
2538 :     sumx += Data8->currentQMV->x/2;
2539 :     sumy += Data8->currentQMV->y/2;
2540 :     } else {
2541 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2542 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2543 :     sumx += Data8->currentMV->x;
2544 :     sumy += Data8->currentMV->y;
2545 : chl 141 }
2546 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2547 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2548 : syskin 1084 if (Data8->cbp[0]) cbp |= 1 << (5 - i);
2549 : chl 141
2550 : syskin 1084 } /* end - for all luma blocks */
2551 : chl 141
2552 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
2553 : chl 141
2554 : edgomez 1053 /* let's check chroma */
2555 : edgomez 1022 sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2556 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2557 : chl 96
2558 : edgomez 1053 /* chroma U */
2559 : edgomez 1022 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2560 :     transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2561 : syskin 1075 bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4);
2562 : edgomez 1022
2563 :     if (bits >= *Data->iMinSAD) return bits;
2564 :    
2565 : edgomez 1053 /* chroma V */
2566 : edgomez 1022 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2567 :     transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2568 : syskin 1075 bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5);
2569 : edgomez 1022
2570 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2571 : edgomez 1022
2572 : syskin 1084 *Data->cbp = cbp;
2573 : edgomez 851 return bits;
2574 :     }
2575 : chl 96
2576 : edgomez 851 static int
2577 : syskin 1108 findRDintra(const SearchData * const Data)
2578 : edgomez 851 {
2579 : edgomez 1053 int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */
2580 : edgomez 1023 int cbp = 0, i, dc = 0;
2581 : edgomez 962 int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2582 : edgomez 195
2583 : edgomez 851 for(i = 0; i < 4; i++) {
2584 :     int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2585 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2586 : syskin 1075 bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc);
2587 : chl 141
2588 : edgomez 1022 if (bits >= Data->iMinSAD[0]) return bits;
2589 : edgomez 851 }
2590 : chl 181
2591 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len;
2592 : chl 96
2593 : edgomez 1053 /*chroma U */
2594 : edgomez 1022 transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2595 : syskin 1075 bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc);
2596 : syskin 1084
2597 : edgomez 1022 if (bits >= Data->iMinSAD[0]) return bits;
2598 :    
2599 : edgomez 1053 /* chroma V */
2600 : edgomez 1022 transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2601 : syskin 1075 bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc);
2602 : edgomez 1022
2603 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2604 : edgomez 1022
2605 : edgomez 851 return bits;
2606 : chl 96 }
2607 : chl 1077
2608 : syskin 1084 static int
2609 : syskin 1108 findRDgmc(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y)
2610 : syskin 1084 {
2611 :     int bits = BITS_MULT*1; /* this one is mcsel */
2612 :     int cbp = 0, i;
2613 :     int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2614 : chl 1077
2615 : syskin 1084 for(i = 0; i < 4; i++) {
2616 :     int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2617 :     transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16*(x+y*Data->iEdgedWidth), Data->iEdgedWidth);
2618 :     bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i);
2619 :     if (bits >= Data->iMinSAD[0]) return bits;
2620 :     }
2621 : chl 1077
2622 : syskin 1084 bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
2623 : chl 1077
2624 : syskin 1084 /*chroma U */
2625 :     transfer_8to16subro(in, Data->CurU, vGMC->u + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2);
2626 :     bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4);
2627 : chl 1077
2628 : syskin 1084 if (bits >= Data->iMinSAD[0]) return bits;
2629 :    
2630 :     /* chroma V */
2631 :     transfer_8to16subro(in, Data->CurV , vGMC->v + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2);
2632 :     bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5);
2633 :    
2634 :     bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
2635 :    
2636 :     *Data->cbp = cbp;
2637 :    
2638 :     return bits;
2639 :     }
2640 :    
2641 :    
2642 :    
2643 :    
2644 : chl 1077 static __inline void
2645 :     GMEanalyzeMB ( const uint8_t * const pCur,
2646 :     const uint8_t * const pRef,
2647 :     const uint8_t * const pRefH,
2648 :     const uint8_t * const pRefV,
2649 :     const uint8_t * const pRefHV,
2650 :     const int x,
2651 :     const int y,
2652 :     const MBParam * const pParam,
2653 :     MACROBLOCK * const pMBs,
2654 :     SearchData * const Data)
2655 :     {
2656 :    
2657 :     int i=0;
2658 :     MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2659 :    
2660 :     Data->iMinSAD[0] = MV_MAX_ERROR;
2661 :    
2662 : syskin 1084 Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0);
2663 : chl 1077
2664 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2665 :     pParam->width, pParam->height, 16, 1, 0);
2666 : chl 1077
2667 :     Data->Cur = pCur + 16*(x + y * pParam->edged_width);
2668 :     Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width);
2669 :     Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width);
2670 :     Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width);
2671 :     Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width);
2672 :    
2673 :     Data->currentMV[0].x = Data->currentMV[0].y = 0;
2674 : syskin 1132 CheckCandidate16I(0, 0, Data, 255);
2675 : chl 1077
2676 :     if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) )
2677 : syskin 1132 CheckCandidate16I(Data->predMV.x, Data->predMV.y, Data, 255);
2678 : chl 1077
2679 : syskin 1132 DiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255, CheckCandidate16I);
2680 : chl 1077
2681 : syskin 1132 SubpelRefine(Data, CheckCandidate16I);
2682 : syskin 1084
2683 :    
2684 : chl 1077 /* for QPel halfpel positions are worse than in halfpel mode :( */
2685 :     /* if (Data->qpel) {
2686 :     Data->currentQMV->x = 2*Data->currentMV->x;
2687 :     Data->currentQMV->y = 2*Data->currentMV->y;
2688 :     Data->qpel_precision = 1;
2689 : syskin 1129 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
2690 :     pParam->width, pParam->height, iFcode, 2, 0);
2691 : chl 1077 SubpelRefine(Data);
2692 :     }
2693 :     */
2694 :    
2695 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
2696 :     pMB->sad16 = Data->iMinSAD[0];
2697 : syskin 1084 pMB->mode = MODE_INTER;
2698 :     pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0);
2699 : chl 1077 return;
2700 :     }
2701 :    
2702 :     void
2703 :     GMEanalysis(const MBParam * const pParam,
2704 :     const FRAMEINFO * const current,
2705 :     const FRAMEINFO * const reference,
2706 :     const IMAGE * const pRefH,
2707 :     const IMAGE * const pRefV,
2708 :     const IMAGE * const pRefHV)
2709 :     {
2710 :     uint32_t x, y;
2711 :     MACROBLOCK * const pMBs = current->mbs;
2712 :     const IMAGE * const pCurrent = &current->image;
2713 :     const IMAGE * const pReference = &reference->image;
2714 :    
2715 :     int32_t iMinSAD[5], temp[5];
2716 :     VECTOR currentMV[5];
2717 : syskin 1132 uint32_t dir;
2718 : chl 1077 SearchData Data;
2719 :     memset(&Data, 0, sizeof(SearchData));
2720 : syskin 1084
2721 : chl 1077 Data.iEdgedWidth = pParam->edged_width;
2722 :     Data.rounding = pParam->m_rounding_type;
2723 : syskin 1084
2724 : chl 1077 Data.currentMV = &currentMV[0];
2725 :     Data.iMinSAD = &iMinSAD[0];
2726 :     Data.iFcode = current->fcode;
2727 :     Data.temp = temp;
2728 : syskin 1132 Data.dir = &dir;
2729 : chl 1077
2730 :     if (sadInit) (*sadInit) ();
2731 :    
2732 :     for (y = 0; y < pParam->mb_height; y ++) {
2733 :     for (x = 0; x < pParam->mb_width; x ++) {
2734 :     GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data);
2735 :     }
2736 :     }
2737 :     return;
2738 :     }
2739 :    
2740 :    
2741 :     WARPPOINTS
2742 :     GlobalMotionEst(MACROBLOCK * const pMBs,
2743 :     const MBParam * const pParam,
2744 :     const FRAMEINFO * const current,
2745 :     const FRAMEINFO * const reference,
2746 :     const IMAGE * const pRefH,
2747 :     const IMAGE * const pRefV,
2748 :     const IMAGE * const pRefHV)
2749 :     {
2750 :    
2751 : syskin 1084 const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2752 :     const int deltay=8;
2753 : chl 1077 const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks)
2754 : syskin 1084 const unsigned int grady=512;
2755 : chl 1077
2756 :     double sol[4] = { 0., 0., 0., 0. };
2757 : syskin 1084
2758 : chl 1077 WARPPOINTS gmc;
2759 :    
2760 :     uint32_t mx, my;
2761 :    
2762 :     int MBh = pParam->mb_height;
2763 :     int MBw = pParam->mb_width;
2764 :     const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */
2765 :     const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */
2766 :    
2767 :     int num=0;
2768 :     int oldnum;
2769 :    
2770 :     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;
2771 :    
2772 :     GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV);
2773 :    
2774 :     /* block based ME isn't done, yet, so do a quick presearch */
2775 :    
2776 :     // filter mask of all blocks
2777 :    
2778 :     for (my = 0; my < (uint32_t)MBh; my++)
2779 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2780 :     {
2781 :     const int mbnum = mx + my * MBw;
2782 :     pMBs[mbnum].mcsel = 0;
2783 :     }
2784 : syskin 1084
2785 :    
2786 :     for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */
2787 : chl 1077 for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */
2788 :     {
2789 :     const int mbnum = mx + my * MBw;
2790 :     MACROBLOCK *const pMB = &pMBs[mbnum];
2791 :     const VECTOR mv = pMB->mvs[0];
2792 :    
2793 :     /* don't use object boundaries */
2794 : syskin 1084 if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax)
2795 : chl 1077 && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay)
2796 : syskin 1084 && (abs(mv.x - (pMB+1)->mvs[0].x) < deltax)
2797 : chl 1077 && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay)
2798 : syskin 1084 && (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax)
2799 : chl 1077 && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay)
2800 : syskin 1084 && (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax)
2801 : chl 1077 && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) )
2802 :     { const int iEdgedWidth = pParam->edged_width;
2803 :     const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx);
2804 :     if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx )
2805 :     && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) )
2806 : syskin 1084 { pMB->mcsel = 1;
2807 : chl 1077 num++;
2808 :     }
2809 :    
2810 :     /* only use "structured" blocks */
2811 :     }
2812 :     }
2813 :     emms();
2814 :    
2815 : syskin 1084 /* further filtering would be possible, but during iteration, remaining
2816 : chl 1077 outliers usually are removed, too */
2817 :    
2818 :     if (num>= minblocks)
2819 :     do { /* until convergence */
2820 :     double DtimesF[4];
2821 :     double a,b,c,n,invdenom;
2822 :     double meanx,meany;
2823 :    
2824 :     a = b = c = n = 0;
2825 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2826 :     for (my = 1; my < (uint32_t)MBh-1; my++)
2827 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2828 :     {
2829 :     const int mbnum = mx + my * MBw;
2830 :     const VECTOR mv = pMBs[mbnum].mvs[0];
2831 :    
2832 :     if (!pMBs[mbnum].mcsel)
2833 :     continue;
2834 :    
2835 :     n++;
2836 :     a += 16*mx+8;
2837 :     b += 16*my+8;
2838 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2839 :    
2840 :     DtimesF[0] += (double)mv.x;
2841 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2842 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2843 :     DtimesF[3] += (double)mv.y;
2844 :     }
2845 : syskin 1084
2846 : chl 1077 invdenom = a*a+b*b-c*n;
2847 :    
2848 :     /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2849 :     /* D'*E*F has been calculated in the same loop as matrix */
2850 :    
2851 :     sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2852 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2853 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2854 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2855 :    
2856 :     sol[0] /= invdenom;
2857 :     sol[1] /= invdenom;
2858 :     sol[2] /= invdenom;
2859 :     sol[3] /= invdenom;
2860 :    
2861 :     meanx = meany = 0.;
2862 :     oldnum = 0;
2863 :     for (my = 1; my < (uint32_t)MBh-1; my++)
2864 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2865 :     {
2866 :     const int mbnum = mx + my * MBw;
2867 :     const VECTOR mv = pMBs[mbnum].mvs[0];
2868 :    
2869 :     if (!pMBs[mbnum].mcsel)
2870 :     continue;
2871 :    
2872 :     oldnum++;
2873 :     meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x );
2874 :     meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y );
2875 :     }
2876 :    
2877 :     if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */
2878 :     meanx /= oldnum;
2879 :     else
2880 :     meanx = 0.25;
2881 :    
2882 :     if (4*meany > oldnum)
2883 :     meany /= oldnum;
2884 :     else
2885 :     meany = 0.25;
2886 :    
2887 :     num = 0;
2888 :     for (my = 0; my < (uint32_t)MBh; my++)
2889 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2890 :     {
2891 :     const int mbnum = mx + my * MBw;
2892 :     const VECTOR mv = pMBs[mbnum].mvs[0];
2893 :    
2894 :     if (!pMBs[mbnum].mcsel)
2895 :     continue;
2896 :    
2897 :     if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx )
2898 :     || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) )
2899 :     pMBs[mbnum].mcsel=0;
2900 :     else
2901 :     num++;
2902 :     }
2903 :    
2904 :     } while ( (oldnum != num) && (num>= minblocks) );
2905 :    
2906 :     if (num < minblocks)
2907 :     {
2908 :     const int iEdgedWidth = pParam->edged_width;
2909 :     num = 0;
2910 :    
2911 :     /* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw);
2912 : syskin 1084 */
2913 : chl 1077 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;
2914 :    
2915 : edgomez 1107 if (!(current->motion_flags & XVID_ME_GME_REFINE))
2916 : chl 1077 return gmc;
2917 : syskin 1084
2918 :     for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */
2919 : chl 1077 for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */
2920 :     {
2921 :     const int mbnum = mx + my * MBw;
2922 :     MACROBLOCK *const pMB = &pMBs[mbnum];
2923 :     const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx);
2924 :     if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx )
2925 :     && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) )
2926 : syskin 1084 { pMB->mcsel = 1;
2927 : chl 1077 gmc.duv[0].x += pMB->mvs[0].x;
2928 :     gmc.duv[0].y += pMB->mvs[0].y;
2929 :     num++;
2930 :     }
2931 :     }
2932 : syskin 1084
2933 : chl 1077 if (gmc.duv[0].x)
2934 :     gmc.duv[0].x /= num;
2935 : syskin 1084 if (gmc.duv[0].y)
2936 : chl 1077 gmc.duv[0].y /= num;
2937 :     } else {
2938 :    
2939 :     gmc.duv[0].x=(int)(sol[0]+0.5);
2940 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2941 :    
2942 :     gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2943 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2944 :    
2945 :     gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */
2946 :     gmc.duv[2].y=gmc.duv[1].x;
2947 :     }
2948 :     if (num>maxblocks)
2949 :     { for (my = 1; my < (uint32_t)MBh-1; my++)
2950 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2951 :     {
2952 :     const int mbnum = mx + my * MBw;
2953 :     if (pMBs[mbnum-1].mcsel)
2954 :     pMBs[mbnum].mcsel=0;
2955 : syskin 1084 else
2956 : chl 1077 if (pMBs[mbnum-MBw].mcsel)
2957 :     pMBs[mbnum].mcsel=0;
2958 :     }
2959 :     }
2960 :     return gmc;
2961 :     }
2962 :    
2963 :     int
2964 :     GlobalMotionEstRefine(
2965 :     WARPPOINTS *const startwp,
2966 :     MACROBLOCK * const pMBs,
2967 :     const MBParam * const pParam,
2968 :     const FRAMEINFO * const current,
2969 :     const FRAMEINFO * const reference,
2970 :     const IMAGE * const pCurr,
2971 :     const IMAGE * const pRef,
2972 :     const IMAGE * const pRefH,
2973 :     const IMAGE * const pRefV,
2974 :     const IMAGE * const pRefHV)
2975 :     {
2976 :     uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width);
2977 :     WARPPOINTS bestwp=*startwp;
2978 :     WARPPOINTS centerwp,currwp;
2979 :     int gmcminSAD=0;
2980 :     int gmcSAD=0;
2981 :     int direction;
2982 :     // int mx,my;
2983 :    
2984 :     /* use many blocks... */
2985 :     /* for (my = 0; my < (uint32_t)pParam->mb_height; my++)
2986 :     for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++)
2987 :     {
2988 :     const int mbnum = mx + my * pParam->mb_width;
2989 : syskin 1084 pMBs[mbnum].mcsel=1;
2990 : chl 1077 }
2991 :     */
2992 :    
2993 :     /* or rather don't use too many blocks... */
2994 :     /*
2995 :     for (my = 1; my < (uint32_t)MBh-1; my++)
2996 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2997 :     {
2998 :     const int mbnum = mx + my * MBw;
2999 :     if (MBmask[mbnum-1])
3000 :     MBmask[mbnum-1]=0;
3001 : syskin 1084 else
3002 : chl 1077 if (MBmask[mbnum-MBw])
3003 :     MBmask[mbnum-1]=0;
3004 : syskin 1084
3005 : chl 1077 }
3006 :     */
3007 :     gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3008 :    
3009 : syskin 1084 if ( (reference->coding_type == S_VOP)
3010 :     && ( (reference->warp.duv[1].x != bestwp.duv[1].x)
3011 :     || (reference->warp.duv[1].y != bestwp.duv[1].y)
3012 :     || (reference->warp.duv[0].x != bestwp.duv[0].x)
3013 :     || (reference->warp.duv[0].y != bestwp.duv[0].y)
3014 :     || (reference->warp.duv[2].x != bestwp.duv[2].x)
3015 :     || (reference->warp.duv[2].y != bestwp.duv[2].y) ) )
3016 : chl 1077 {
3017 : syskin 1084 gmcSAD = globalSAD(&reference->warp, pParam, pMBs,
3018 : chl 1077 current, pRef, pCurr, GMCblock);
3019 : syskin 1084
3020 : chl 1077 if (gmcSAD < gmcminSAD)
3021 :     { bestwp = reference->warp;
3022 :     gmcminSAD = gmcSAD;
3023 :     }
3024 :     }
3025 :    
3026 :     do {
3027 :     direction = 0;
3028 :     centerwp = bestwp;
3029 :    
3030 : syskin 1084 currwp = centerwp;
3031 :    
3032 : chl 1077 currwp.duv[0].x--;
3033 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3034 :     if (gmcSAD < gmcminSAD)
3035 :     { bestwp = currwp;
3036 :     gmcminSAD = gmcSAD;
3037 : syskin 1084 direction = 1;
3038 : chl 1077 }
3039 : syskin 1084 else
3040 :     {
3041 : chl 1077 currwp = centerwp; currwp.duv[0].x++;
3042 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3043 :     if (gmcSAD < gmcminSAD)
3044 :     { bestwp = currwp;
3045 :     gmcminSAD = gmcSAD;
3046 : syskin 1084 direction = 2;
3047 : chl 1077 }
3048 :     }
3049 :     if (direction) continue;
3050 :    
3051 :     currwp = centerwp; currwp.duv[0].y--;
3052 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3053 :     if (gmcSAD < gmcminSAD)
3054 :     { bestwp = currwp;
3055 :     gmcminSAD = gmcSAD;
3056 : syskin 1084 direction = 4;
3057 :     }
3058 :     else
3059 : chl 1077 {
3060 :     currwp = centerwp; currwp.duv[0].y++;
3061 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3062 :     if (gmcSAD < gmcminSAD)
3063 :     { bestwp = currwp;
3064 :     gmcminSAD = gmcSAD;
3065 : syskin 1084 direction = 8;
3066 : chl 1077 }
3067 : syskin 1084 }
3068 : chl 1077 if (direction) continue;
3069 : syskin 1084
3070 : chl 1077 currwp = centerwp; currwp.duv[1].x++;
3071 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3072 :     if (gmcSAD < gmcminSAD)
3073 :     { bestwp = currwp;
3074 :     gmcminSAD = gmcSAD;
3075 : syskin 1084 direction = 32;
3076 :     }
3077 : chl 1077 currwp.duv[2].y++;
3078 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3079 :     if (gmcSAD < gmcminSAD)
3080 :     { bestwp = currwp;
3081 :     gmcminSAD = gmcSAD;
3082 : syskin 1084 direction = 1024;
3083 :     }
3084 : chl 1077
3085 :     currwp = centerwp; currwp.duv[1].x--;
3086 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3087 :     if (gmcSAD < gmcminSAD)
3088 :     { bestwp = currwp;
3089 :     gmcminSAD = gmcSAD;
3090 : syskin 1084 direction = 16;
3091 :     }
3092 :     else
3093 : chl 1077 {
3094 :     currwp = centerwp; currwp.duv[1].x++;
3095 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3096 :     if (gmcSAD < gmcminSAD)
3097 :     { bestwp = currwp;
3098 :     gmcminSAD = gmcSAD;
3099 : syskin 1084 direction = 32;
3100 : chl 1077 }
3101 : syskin 1084 }
3102 : chl 1077 if (direction) continue;
3103 :    
3104 :    
3105 :     currwp = centerwp; currwp.duv[1].y--;
3106 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3107 :     if (gmcSAD < gmcminSAD)
3108 :     { bestwp = currwp;
3109 :     gmcminSAD = gmcSAD;
3110 : syskin 1084 direction = 64;
3111 :     }
3112 :     else
3113 : chl 1077 {
3114 :     currwp = centerwp; currwp.duv[1].y++;
3115 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3116 :     if (gmcSAD < gmcminSAD)
3117 :     { bestwp = currwp;
3118 :     gmcminSAD = gmcSAD;
3119 : syskin 1084 direction = 128;
3120 : chl 1077 }
3121 :     }
3122 :     if (direction) continue;
3123 : syskin 1084
3124 : chl 1077 currwp = centerwp; currwp.duv[2].x--;
3125 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3126 :     if (gmcSAD < gmcminSAD)
3127 :     { bestwp = currwp;
3128 :     gmcminSAD = gmcSAD;
3129 : syskin 1084 direction = 256;
3130 :     }
3131 :     else
3132 : chl 1077 {
3133 :     currwp = centerwp; currwp.duv[2].x++;
3134 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3135 :     if (gmcSAD < gmcminSAD)
3136 :     { bestwp = currwp;
3137 :     gmcminSAD = gmcSAD;
3138 : syskin 1084 direction = 512;
3139 : chl 1077 }
3140 :     }
3141 :     if (direction) continue;
3142 : syskin 1084
3143 : chl 1077 currwp = centerwp; currwp.duv[2].y--;
3144 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3145 :     if (gmcSAD < gmcminSAD)
3146 :     { bestwp = currwp;
3147 :     gmcminSAD = gmcSAD;
3148 : syskin 1084 direction = 1024;
3149 :     }
3150 : chl 1077 else
3151 : syskin 1084 {
3152 : chl 1077 currwp = centerwp; currwp.duv[2].y++;
3153 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3154 :     if (gmcSAD < gmcminSAD)
3155 :     { bestwp = currwp;
3156 :     gmcminSAD = gmcSAD;
3157 : syskin 1084 direction = 2048;
3158 : chl 1077 }
3159 :     }
3160 :     } while (direction);
3161 :     free(GMCblock);
3162 : syskin 1084
3163 : chl 1077 *startwp = bestwp;
3164 :    
3165 :     return gmcminSAD;
3166 :     }
3167 :    
3168 :     int
3169 : syskin 1084 globalSAD(const WARPPOINTS *const wp,
3170 :     const MBParam * const pParam,
3171 : chl 1077 const MACROBLOCK * const pMBs,
3172 :     const FRAMEINFO * const current,
3173 :     const IMAGE * const pRef,
3174 :     const IMAGE * const pCurr,
3175 :     uint8_t *const GMCblock)
3176 :     {
3177 :     NEW_GMC_DATA gmc_data;
3178 :     int iSAD, gmcSAD=0;
3179 :     int num=0;
3180 :     unsigned int mx, my;
3181 : syskin 1084
3182 : chl 1077 generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data);
3183 : syskin 1084
3184 : chl 1077 for (my = 0; my < (uint32_t)pParam->mb_height; my++)
3185 :     for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) {
3186 :    
3187 :     const int mbnum = mx + my * pParam->mb_width;
3188 :     const int iEdgedWidth = pParam->edged_width;
3189 : syskin 1084
3190 : chl 1077 if (!pMBs[mbnum].mcsel)
3191 :     continue;
3192 : syskin 1084
3193 :     gmc_data.predict_16x16(&gmc_data, GMCblock,
3194 : chl 1077 pRef->y,
3195 : syskin 1084 iEdgedWidth,
3196 :     iEdgedWidth,
3197 :     mx, my,
3198 : chl 1077 pParam->m_rounding_type);
3199 :    
3200 :     iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx),
3201 : syskin 1084 GMCblock , iEdgedWidth, 65536);
3202 : chl 1077 iSAD -= pMBs[mbnum].sad16;
3203 : syskin 1084
3204 :     if (iSAD<0)
3205 : chl 1077 gmcSAD += iSAD;
3206 :     num++;
3207 :     }
3208 :     return gmcSAD;
3209 :     }
3210 :    

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