[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 1077 - (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 : chl 1077 * $Id: motion_est.c,v 1.58.2.20 2003-06-28 15:52:10 chl 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 : chl 1077 #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 :     CheckCandidate((X),(Y), (D), &iDirection, data ); }
74 : Isibaar 3
75 : edgomez 872 /*****************************************************************************
76 :     * Code
77 :     ****************************************************************************/
78 :    
79 : edgomez 851 static __inline uint32_t
80 :     d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
81 :     {
82 : edgomez 974 int bits;
83 :     const int q = (1 << (iFcode - 1)) - 1;
84 :    
85 :     x <<= qpel;
86 :     y <<= qpel;
87 : edgomez 851 if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
88 : chl 141
89 : edgomez 851 x -= pred.x;
90 : edgomez 974 bits = (x != 0 ? iFcode:0);
91 :     x = abs(x);
92 :     x += q;
93 :     x >>= (iFcode - 1);
94 :     bits += mvtab[x];
95 :    
96 : edgomez 851 y -= pred.y;
97 : edgomez 974 bits += (y != 0 ? iFcode:0);
98 :     y = abs(y);
99 :     y += q;
100 :     y >>= (iFcode - 1);
101 :     bits += mvtab[y];
102 : chl 141
103 : edgomez 974 return bits;
104 : edgomez 851 }
105 : chl 141
106 : edgomez 978 static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by,
107 :     const SearchData * const data)
108 : edgomez 851 {
109 :     int sad;
110 :     const uint32_t stride = data->iEdgedWidth/2;
111 :     uint8_t * f_refu = data->RefQ,
112 :     * f_refv = data->RefQ + 8,
113 :     * b_refu = data->RefQ + 16,
114 :     * b_refv = data->RefQ + 24;
115 : edgomez 978 int offset = (fx>>1) + (fy>>1)*stride;
116 : Isibaar 3
117 : edgomez 851 switch (((fx & 1) << 1) | (fy & 1)) {
118 :     case 0:
119 : edgomez 978 f_refu = (uint8_t*)data->RefP[4] + offset;
120 :     f_refv = (uint8_t*)data->RefP[5] + offset;
121 : edgomez 851 break;
122 :     case 1:
123 : edgomez 978 interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding);
124 :     interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding);
125 : edgomez 851 break;
126 :     case 2:
127 : edgomez 978 interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding);
128 :     interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding);
129 : edgomez 851 break;
130 :     default:
131 : edgomez 978 interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding);
132 :     interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding);
133 : edgomez 851 break;
134 :     }
135 : Isibaar 3
136 : edgomez 978 offset = (bx>>1) + (by>>1)*stride;
137 : edgomez 851 switch (((bx & 1) << 1) | (by & 1)) {
138 :     case 0:
139 : edgomez 978 b_refu = (uint8_t*)data->b_RefP[4] + offset;
140 :     b_refv = (uint8_t*)data->b_RefP[5] + offset;
141 : edgomez 851 break;
142 :     case 1:
143 : edgomez 978 interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
144 :     interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
145 : edgomez 851 break;
146 :     case 2:
147 : edgomez 978 interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
148 :     interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
149 : edgomez 851 break;
150 :     default:
151 : edgomez 978 interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
152 :     interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
153 : edgomez 851 break;
154 :     }
155 : Isibaar 3
156 : edgomez 851 sad = sad8bi(data->CurU, b_refu, f_refu, stride);
157 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
158 : Isibaar 3
159 : edgomez 851 return sad;
160 :     }
161 : Isibaar 3
162 : edgomez 851 static int32_t
163 : edgomez 978 ChromaSAD(const int dx, const int dy, const SearchData * const data)
164 : edgomez 851 {
165 :     int sad;
166 :     const uint32_t stride = data->iEdgedWidth/2;
167 : edgomez 978 int offset = (dx>>1) + (dy>>1)*stride;
168 : Isibaar 3
169 : edgomez 1053 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */
170 :     data->temp[5] = dx; data->temp[6] = dy; /* backup */
171 : Isibaar 3
172 : edgomez 851 switch (((dx & 1) << 1) | (dy & 1)) {
173 :     case 0:
174 : edgomez 978 sad = sad8(data->CurU, data->RefP[4] + offset, stride);
175 :     sad += sad8(data->CurV, data->RefP[5] + offset, stride);
176 : edgomez 851 break;
177 :     case 1:
178 : edgomez 978 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride);
179 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride);
180 : edgomez 851 break;
181 :     case 2:
182 : edgomez 978 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride);
183 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride);
184 : edgomez 851 break;
185 :     default:
186 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding);
187 : edgomez 851 sad = sad8(data->CurU, data->RefQ, stride);
188 :    
189 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding);
190 : edgomez 851 sad += sad8(data->CurV, data->RefQ, stride);
191 :     break;
192 :     }
193 : edgomez 1053 data->temp[7] = sad; /* backup, part 2 */
194 : edgomez 851 return sad;
195 : Isibaar 3 }
196 :    
197 : edgomez 851 static __inline const uint8_t *
198 :     GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
199 : Isibaar 3 {
200 : edgomez 1053 /* dir : 0 = forward, 1 = backward */
201 : edgomez 987 const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP );
202 : edgomez 978 const int picture = ((x&1)<<1) | (y&1);
203 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
204 :     return direction[picture] + offset;
205 : Isibaar 3 }
206 :    
207 : edgomez 1053 /* this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate */
208 : edgomez 851 static __inline const uint8_t *
209 :     GetReference(const int x, const int y, const SearchData * const data)
210 : Isibaar 3 {
211 : edgomez 978 const int picture = ((x&1)<<1) | (y&1);
212 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
213 :     return data->RefP[picture] + offset;
214 : Isibaar 3 }
215 : edgomez 851
216 :     static uint8_t *
217 :     Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
218 : Isibaar 3 {
219 : edgomez 1053 /* create or find a qpel-precision reference picture; return pointer to it */
220 : edgomez 851 uint8_t * Reference = data->RefQ + 16*dir;
221 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
222 :     const uint32_t rounding = data->rounding;
223 :     const int halfpel_x = x/2;
224 :     const int halfpel_y = y/2;
225 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
226 : suxen_drol 136
227 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
228 :     ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
229 :     switch( ((x&1)<<1) + (y&1) ) {
230 : edgomez 1053 case 3: /* x and y in qpel resolution - the "corners" (top left/right and */
231 :     /* bottom left/right) during qpel refinement */
232 : edgomez 963 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
233 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
234 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
235 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
236 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
237 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
238 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
239 : edgomez 851 break;
240 : chl 184
241 : edgomez 1053 case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
242 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
243 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
244 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
245 :     break;
246 : suxen_drol 136
247 : edgomez 1053 case 2: /* x qpel, y halfpel - left or right during qpel refinement */
248 : edgomez 851 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
249 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
250 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
251 :     break;
252 : edgomez 195
253 : edgomez 1053 default: /* pure halfpel position */
254 : edgomez 963 return (uint8_t *) ref1;
255 : edgomez 978
256 : edgomez 851 }
257 :     return Reference;
258 :     }
259 : Isibaar 3
260 : edgomez 851 static uint8_t *
261 :     Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
262 :     {
263 : edgomez 1053 /* create or find a qpel-precision reference picture; return pointer to it */
264 : edgomez 851 uint8_t * Reference = data->RefQ + 16*dir;
265 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
266 :     const uint32_t rounding = data->rounding;
267 :     const int halfpel_x = x/2;
268 :     const int halfpel_y = y/2;
269 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
270 : chl 326
271 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
272 :     switch( ((x&1)<<1) + (y&1) ) {
273 : edgomez 1053 case 3:
274 :     /*
275 :     * x and y in qpel resolution - the "corners" (top left/right and
276 :     * bottom left/right) during qpel refinement
277 :     */
278 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
279 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
280 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
281 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
282 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
283 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
284 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
285 :     break;
286 : chl 184
287 : edgomez 1053 case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
288 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
289 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
290 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
291 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
292 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
293 :     break;
294 : chl 184
295 : edgomez 1053 case 2: /* x qpel, y halfpel - left or right during qpel refinement */
296 : edgomez 851 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
297 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
298 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
299 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
300 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
301 :     break;
302 : edgomez 195
303 : edgomez 1053 default: /* pure halfpel position */
304 : edgomez 851 return (uint8_t *) ref1;
305 :     }
306 :     return Reference;
307 :     }
308 : edgomez 195
309 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS START */
310 : edgomez 195
311 : edgomez 851 static void
312 :     CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
313 :     {
314 :     int xc, yc;
315 :     const uint8_t * Reference;
316 :     VECTOR * current;
317 :     int32_t sad; uint32_t t;
318 : edgomez 195
319 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
320 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
321 : edgomez 195
322 : edgomez 851 if (!data->qpel_precision) {
323 :     Reference = GetReference(x, y, data);
324 :     current = data->currentMV;
325 :     xc = x; yc = y;
326 : edgomez 1053 } else { /* x and y are in 1/4 precision */
327 : edgomez 851 Reference = Interpolate16x16qpel(x, y, 0, data);
328 : edgomez 1053 xc = x/2; yc = y/2; /* for chroma sad */
329 : edgomez 851 current = data->currentQMV;
330 :     }
331 : edgomez 195
332 : edgomez 851 sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
333 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
334 : edgomez 195
335 : edgomez 851 sad += (data->lambda16 * t * sad)>>10;
336 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
337 : chl 175
338 : edgomez 851 if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
339 : edgomez 953 (yc >> 1) + roundtab_79[yc & 0x3], data);
340 : edgomez 195
341 : edgomez 851 if (sad < data->iMinSAD[0]) {
342 :     data->iMinSAD[0] = sad;
343 :     current[0].x = x; current[0].y = y;
344 :     *dir = Direction;
345 :     }
346 : edgomez 195
347 : edgomez 851 if (data->temp[1] < data->iMinSAD[1]) {
348 :     data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
349 :     if (data->temp[2] < data->iMinSAD[2]) {
350 :     data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
351 :     if (data->temp[3] < data->iMinSAD[3]) {
352 :     data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
353 :     if (data->temp[4] < data->iMinSAD[4]) {
354 :     data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
355 : Isibaar 3 }
356 :    
357 : edgomez 851 static void
358 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
359 :     {
360 :     int32_t sad; uint32_t t;
361 :     const uint8_t * Reference;
362 : edgomez 974 VECTOR * current;
363 : chl 326
364 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
365 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
366 : Isibaar 3
367 : edgomez 974 if (!data->qpel_precision) {
368 :     Reference = GetReference(x, y, data);
369 :     current = data->currentMV;
370 : edgomez 1053 } else { /* x and y are in 1/4 precision */
371 : edgomez 974 Reference = Interpolate8x8qpel(x, y, 0, 0, data);
372 :     current = data->currentQMV;
373 :     }
374 : Isibaar 3
375 : edgomez 851 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
376 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
377 : Isibaar 3
378 : edgomez 851 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
379 : Isibaar 3
380 : edgomez 851 if (sad < *(data->iMinSAD)) {
381 :     *(data->iMinSAD) = sad;
382 : edgomez 974 current->x = x; current->y = y;
383 : edgomez 851 *dir = Direction;
384 :     }
385 : Isibaar 3 }
386 :    
387 : edgomez 851 static void
388 :     CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
389 :     {
390 :     uint32_t t;
391 :     const uint8_t * Reference;
392 :    
393 : edgomez 1053 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */
394 : edgomez 851 (x > data->max_dx) || (x < data->min_dx)
395 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
396 :    
397 :     Reference = GetReference(x, y, data);
398 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
399 :    
400 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
401 :    
402 :     data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
403 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
404 :    
405 :     if (data->temp[0] < data->iMinSAD[0]) {
406 :     data->iMinSAD[0] = data->temp[0];
407 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
408 :     *dir = Direction; }
409 :    
410 :     if (data->temp[1] < data->iMinSAD[1]) {
411 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
412 :     if (data->temp[2] < data->iMinSAD[2]) {
413 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
414 :     if (data->temp[3] < data->iMinSAD[3]) {
415 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
416 :     if (data->temp[4] < data->iMinSAD[4]) {
417 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
418 : Isibaar 3 }
419 : edgomez 195
420 : edgomez 851 static void
421 :     CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
422 :     {
423 :     int32_t sad, xc, yc;
424 :     const uint8_t * Reference;
425 :     uint32_t t;
426 :     VECTOR * current;
427 :    
428 : edgomez 963 if ( (x > data->max_dx) || ( x < data->min_dx)
429 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
430 : edgomez 851
431 : edgomez 1053 if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */
432 : edgomez 851
433 : edgomez 1053 if (data->qpel_precision) { /* x and y are in 1/4 precision */
434 : edgomez 851 Reference = Interpolate16x16qpel(x, y, 0, data);
435 :     current = data->currentQMV;
436 :     xc = x/2; yc = y/2;
437 :     } else {
438 :     Reference = GetReference(x, y, data);
439 :     current = data->currentMV;
440 :     xc = x; yc = y;
441 :     }
442 :     t = d_mv_bits(x, y, data->predMV, data->iFcode,
443 :     data->qpel^data->qpel_precision, data->rrv);
444 :    
445 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
446 :     sad += (data->lambda16 * t * sad)>>10;
447 :    
448 :     if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
449 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
450 :    
451 :     if (sad < *(data->iMinSAD)) {
452 :     *(data->iMinSAD) = sad;
453 :     current->x = x; current->y = y;
454 :     *dir = Direction;
455 :     }
456 : chl 96 }
457 : Isibaar 3
458 : edgomez 851 static void
459 : chl 1077 CheckCandidate16I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
460 :     {
461 :     int sad;
462 :     // int xc, yc;
463 :     const uint8_t * Reference;
464 :     // VECTOR * current;
465 :    
466 :     if ( (x > data->max_dx) || ( x < data->min_dx)
467 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
468 :    
469 :     Reference = GetReference(x, y, data);
470 :     // xc = x; yc = y;
471 :    
472 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
473 :     // sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0);
474 :    
475 :     /* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
476 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
477 :     */
478 :    
479 :     if (sad < data->iMinSAD[0]) {
480 :     data->iMinSAD[0] = sad;
481 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
482 :     *dir = Direction;
483 :     }
484 :     }
485 :    
486 :     static void
487 : edgomez 851 CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
488 :     {
489 : edgomez 1053 /* maximum speed - for P/B/I decision */
490 : edgomez 851 int32_t sad;
491 :    
492 :     if ( (x > data->max_dx) || (x < data->min_dx)
493 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
494 :    
495 : edgomez 987 sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*(data->iEdgedWidth),
496 :     data->iEdgedWidth, data->temp+1);
497 : edgomez 851
498 :     if (sad < *(data->iMinSAD)) {
499 :     *(data->iMinSAD) = sad;
500 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
501 :     *dir = Direction;
502 :     }
503 :     if (data->temp[1] < data->iMinSAD[1]) {
504 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
505 :     if (data->temp[2] < data->iMinSAD[2]) {
506 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
507 :     if (data->temp[3] < data->iMinSAD[3]) {
508 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
509 :     if (data->temp[4] < data->iMinSAD[4]) {
510 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
511 :    
512 : Isibaar 3 }
513 :    
514 : edgomez 851 static void
515 :     CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
516 :     {
517 :     int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
518 :     uint32_t t;
519 :     const uint8_t *ReferenceF, *ReferenceB;
520 :     VECTOR *current;
521 :    
522 : edgomez 953 if ((xf > data->max_dx) || (xf < data->min_dx) ||
523 :     (yf > data->max_dy) || (yf < data->min_dy))
524 :     return;
525 : edgomez 851
526 :     if (!data->qpel_precision) {
527 :     ReferenceF = GetReference(xf, yf, data);
528 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
529 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
530 :     current = data->currentMV;
531 :     xcf = xf; ycf = yf;
532 :     xcb = xb; ycb = yb;
533 :     } else {
534 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
535 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
536 :     current = data->currentQMV;
537 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
538 :     xcf = xf/2; ycf = yf/2;
539 :     xcb = xb/2; ycb = yb/2;
540 :     }
541 :    
542 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
543 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
544 :    
545 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
546 :     sad += (data->lambda16 * t * sad)>>10;
547 :    
548 :     if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
549 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
550 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
551 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
552 :    
553 :     if (sad < *(data->iMinSAD)) {
554 :     *(data->iMinSAD) = sad;
555 :     current->x = xf; current->y = yf;
556 :     *dir = Direction;
557 :     }
558 : Isibaar 3 }
559 :    
560 : edgomez 851 static void
561 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
562 :     {
563 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
564 :     uint32_t k;
565 :     const uint8_t *ReferenceF;
566 :     const uint8_t *ReferenceB;
567 :     VECTOR mvs, b_mvs;
568 :    
569 : edgomez 953 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
570 : edgomez 851
571 :     for (k = 0; k < 4; k++) {
572 :     mvs.x = data->directmvF[k].x + x;
573 :     b_mvs.x = ((x == 0) ?
574 :     data->directmvB[k].x
575 :     : mvs.x - data->referencemv[k].x);
576 :    
577 :     mvs.y = data->directmvF[k].y + y;
578 :     b_mvs.y = ((y == 0) ?
579 :     data->directmvB[k].y
580 :     : mvs.y - data->referencemv[k].y);
581 :    
582 : edgomez 953 if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) ||
583 :     (mvs.y > data->max_dy) || (mvs.y < data->min_dy) ||
584 :     (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) ||
585 :     (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) )
586 :     return;
587 : edgomez 851
588 :     if (data->qpel) {
589 :     xcf += mvs.x/2; ycf += mvs.y/2;
590 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
591 :     } else {
592 :     xcf += mvs.x; ycf += mvs.y;
593 :     xcb += b_mvs.x; ycb += b_mvs.y;
594 : edgomez 1053 mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */
595 : edgomez 851 b_mvs.x *= 2; b_mvs.y *= 2;
596 :     }
597 :    
598 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
599 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
600 :    
601 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
602 :     ReferenceF, ReferenceB, data->iEdgedWidth);
603 :     if (sad > *(data->iMinSAD)) return;
604 :     }
605 :    
606 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
607 :    
608 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
609 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
610 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
611 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
612 :    
613 :     if (sad < *(data->iMinSAD)) {
614 :     *(data->iMinSAD) = sad;
615 :     data->currentMV->x = x; data->currentMV->y = y;
616 :     *dir = Direction;
617 :     }
618 : Isibaar 3 }
619 :    
620 : edgomez 851 static void
621 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
622 : Isibaar 3 {
623 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
624 :     const uint8_t *ReferenceF;
625 :     const uint8_t *ReferenceB;
626 :     VECTOR mvs, b_mvs;
627 : Isibaar 3
628 : edgomez 963 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
629 : Isibaar 3
630 : edgomez 851 mvs.x = data->directmvF[0].x + x;
631 :     b_mvs.x = ((x == 0) ?
632 :     data->directmvB[0].x
633 :     : mvs.x - data->referencemv[0].x);
634 : Isibaar 3
635 : edgomez 851 mvs.y = data->directmvF[0].y + y;
636 :     b_mvs.y = ((y == 0) ?
637 :     data->directmvB[0].y
638 :     : mvs.y - data->referencemv[0].y);
639 : Isibaar 3
640 : edgomez 963 if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx)
641 :     || (mvs.y > data->max_dy) || (mvs.y < data->min_dy)
642 :     || (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx)
643 :     || (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return;
644 : Isibaar 3
645 : edgomez 851 if (data->qpel) {
646 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
647 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
648 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
649 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
650 :     } else {
651 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
652 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
653 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
654 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
655 :     }
656 :    
657 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
658 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
659 :    
660 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
661 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
662 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
663 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
664 :    
665 :     if (sad < *(data->iMinSAD)) {
666 :     *(data->iMinSAD) = sad;
667 :     data->currentMV->x = x; data->currentMV->y = y;
668 :     *dir = Direction;
669 :     }
670 : Isibaar 3 }
671 :    
672 : edgomez 851
673 :     static void
674 :     CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
675 : Isibaar 3 {
676 :    
677 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
678 : edgomez 1022 int32_t bits = 0;
679 : edgomez 851 VECTOR * current;
680 :     const uint8_t * ptr;
681 :     int i, cbp = 0, t, xc, yc;
682 : edgomez 195
683 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
684 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
685 : edgomez 195
686 : edgomez 851 if (!data->qpel_precision) {
687 :     ptr = GetReference(x, y, data);
688 :     current = data->currentMV;
689 :     xc = x; yc = y;
690 : edgomez 1053 } else { /* x and y are in 1/4 precision */
691 : edgomez 851 ptr = Interpolate16x16qpel(x, y, 0, data);
692 :     current = data->currentQMV;
693 :     xc = x/2; yc = y/2;
694 :     }
695 : Isibaar 3
696 : edgomez 851 for(i = 0; i < 4; i++) {
697 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
698 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
699 : syskin 1075 bits += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i);
700 : edgomez 851 }
701 : Isibaar 3
702 : edgomez 1023 bits += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
703 : edgomez 195
704 : syskin 1071 if (data->temp[0] + t < data->iMinSAD[1]) {
705 :     data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
706 :     if (data->temp[1] < data->iMinSAD[2]) {
707 :     data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
708 :     if (data->temp[2] < data->iMinSAD[3]) {
709 :     data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
710 :     if (data->temp[3] < data->iMinSAD[4]) {
711 :     data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
712 :    
713 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
714 : edgomez 851
715 : edgomez 1022 if (bits >= data->iMinSAD[0]) return;
716 : edgomez 851
717 : edgomez 1053 /* chroma */
718 : edgomez 1022 xc = (xc >> 1) + roundtab_79[xc & 0x3];
719 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
720 : edgomez 851
721 : edgomez 1053 /* chroma U */
722 : edgomez 1022 ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
723 :     transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
724 : syskin 1075 bits += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4);
725 : edgomez 1022 if (bits >= data->iMinSAD[0]) return;
726 :    
727 : edgomez 1053 /* chroma V */
728 : edgomez 1022 ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
729 :     transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2);
730 : syskin 1075 bits += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
731 : edgomez 1022
732 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
733 : edgomez 851
734 :     if (bits < data->iMinSAD[0]) {
735 :     data->iMinSAD[0] = bits;
736 :     current[0].x = x; current[0].y = y;
737 :     *dir = Direction;
738 :     }
739 : syskin 1071 }
740 : edgomez 851
741 :     static void
742 :     CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
743 : chl 96 {
744 :    
745 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
746 : edgomez 1022 int32_t bits;
747 : edgomez 851 VECTOR * current;
748 :     const uint8_t * ptr;
749 : edgomez 1022 int cbp = 0;
750 : edgomez 195
751 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
752 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
753 : edgomez 195
754 : edgomez 851 if (!data->qpel_precision) {
755 :     ptr = GetReference(x, y, data);
756 :     current = data->currentMV;
757 : edgomez 1053 } else { /* x and y are in 1/4 precision */
758 : edgomez 851 ptr = Interpolate8x8qpel(x, y, 0, 0, data);
759 :     current = data->currentQMV;
760 :     }
761 : chl 96
762 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
763 : syskin 1075 bits = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
764 : edgomez 1023 bits += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
765 : chl 96
766 : edgomez 851 if (bits < data->iMinSAD[0]) {
767 :     data->temp[0] = cbp;
768 :     data->iMinSAD[0] = bits;
769 :     current[0].x = x; current[0].y = y;
770 :     *dir = Direction;
771 :     }
772 :     }
773 : chl 96
774 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
775 : edgomez 195
776 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
777 : chl 96
778 : edgomez 851 static void
779 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
780 :     {
781 : edgomez 195
782 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
783 : edgomez 195
784 : edgomez 851 int iDirection;
785 : edgomez 195
786 : edgomez 1053 for(;;) { /* forever */
787 : edgomez 851 iDirection = 0;
788 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
789 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
790 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
791 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
792 : edgomez 195
793 : edgomez 851 /* now we're doing diagonal checks near our candidate */
794 : edgomez 195
795 : edgomez 1053 if (iDirection) { /* if anything found */
796 : edgomez 851 bDirection = iDirection;
797 :     iDirection = 0;
798 :     x = data->currentMV->x; y = data->currentMV->y;
799 : edgomez 1053 if (bDirection & 3) { /* our candidate is left or right */
800 : edgomez 851 CHECK_CANDIDATE(x, y + iDiamondSize, 8);
801 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
802 : edgomez 1053 } else { /* what remains here is up or down */
803 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y, 2);
804 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
805 :     }
806 : edgomez 195
807 : edgomez 851 if (iDirection) {
808 :     bDirection += iDirection;
809 :     x = data->currentMV->x; y = data->currentMV->y;
810 :     }
811 : edgomez 1053 } else { /* about to quit, eh? not so fast.... */
812 : edgomez 851 switch (bDirection) {
813 :     case 2:
814 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
815 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
816 : edgomez 195 break;
817 : edgomez 851 case 1:
818 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
819 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
820 : edgomez 195 break;
821 : edgomez 851 case 2 + 4:
822 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
823 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
824 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
825 :     break;
826 :     case 4:
827 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
828 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
829 :     break;
830 : edgomez 195 case 8:
831 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
832 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
833 : edgomez 195 break;
834 : edgomez 851 case 1 + 4:
835 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
836 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
837 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
838 : edgomez 195 break;
839 : edgomez 851 case 2 + 8:
840 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
841 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
842 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
843 :     break;
844 :     case 1 + 8:
845 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
846 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
847 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
848 :     break;
849 : edgomez 1053 default: /* 1+2+4+8 == we didn't find anything at all */
850 : edgomez 851 CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
851 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
852 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
853 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
854 :     break;
855 : chl 96 }
856 : edgomez 1053 if (!iDirection) break; /* ok, the end. really */
857 : edgomez 851 bDirection = iDirection;
858 :     x = data->currentMV->x; y = data->currentMV->y;
859 : chl 345 }
860 : edgomez 195 }
861 : chl 96 }
862 :    
863 : edgomez 851 static void
864 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
865 : chl 96 {
866 : edgomez 851 int iDirection;
867 : edgomez 195
868 : edgomez 851 do {
869 :     iDirection = 0;
870 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
871 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
872 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
873 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
874 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
875 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
876 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
877 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
878 : chl 96
879 : edgomez 851 bDirection = iDirection;
880 :     x = data->currentMV->x; y = data->currentMV->y;
881 :     } while (iDirection);
882 : chl 96 }
883 :    
884 : edgomez 851 static void
885 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
886 : chl 181 {
887 :    
888 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
889 :    
890 : edgomez 851 int iDirection;
891 : edgomez 195
892 : edgomez 851 do {
893 :     iDirection = 0;
894 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
895 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
896 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
897 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
898 :    
899 :     /* now we're doing diagonal checks near our candidate */
900 :    
901 : edgomez 1053 if (iDirection) { /* checking if anything found */
902 : edgomez 851 bDirection = iDirection;
903 : chl 181 iDirection = 0;
904 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
905 : edgomez 1053 if (bDirection & 3) { /* our candidate is left or right */
906 : edgomez 851 CHECK_CANDIDATE(x, y + iDiamondSize, 8);
907 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
908 : edgomez 1053 } else { /* what remains here is up or down */
909 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y, 2);
910 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
911 :     }
912 :     bDirection += iDirection;
913 :     x = data->currentMV->x; y = data->currentMV->y;
914 :     }
915 :     }
916 :     while (iDirection);
917 :     }
918 : chl 181
919 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
920 : chl 181
921 : edgomez 851 static void
922 :     SubpelRefine(const SearchData * const data)
923 :     {
924 :     /* Do a half-pel or q-pel refinement */
925 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
926 : edgomez 1053 int iDirection; /* only needed because macro expects it */
927 : chl 181
928 : edgomez 851 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
929 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
930 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
931 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
932 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
933 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
934 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
935 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
936 :     }
937 : chl 181
938 : edgomez 851 static __inline int
939 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
940 :     const int x, const int y,
941 :     const uint32_t stride, const uint32_t iQuant, int rrv)
942 : chl 181
943 : edgomez 851 {
944 : edgomez 963 int offset = (x + y*stride)*8;
945 : edgomez 851 if(!rrv) {
946 : edgomez 963 uint32_t sadC = sad8(current->u + offset,
947 :     reference->u + offset, stride);
948 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
949 : edgomez 963 sadC += sad8(current->v + offset,
950 :     reference->v + offset, stride);
951 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
952 :     return 1;
953 : chl 181
954 : edgomez 851 } else {
955 : edgomez 963 uint32_t sadC = sad16(current->u + 2*offset,
956 :     reference->u + 2*offset, stride, 256*4096);
957 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
958 : edgomez 963 sadC += sad16(current->v + 2*offset,
959 :     reference->v + 2*offset, stride, 256*4096);
960 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
961 :     return 1;
962 : chl 181 }
963 : edgomez 851 }
964 : edgomez 504
965 : edgomez 851 static __inline void
966 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
967 :     {
968 :     pMB->mode = MODE_NOT_CODED;
969 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
970 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
971 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
972 : chl 181 }
973 :    
974 : edgomez 1022 static __inline void
975 :     ModeDecision(SearchData * const Data,
976 :     MACROBLOCK * const pMB,
977 :     const MACROBLOCK * const pMBs,
978 :     const int x, const int y,
979 :     const MBParam * const pParam,
980 :     const uint32_t MotionFlags,
981 :     const uint32_t VopFlags,
982 :     const uint32_t VolFlags,
983 :     const IMAGE * const pCurrent,
984 :     const IMAGE * const pRef)
985 :     {
986 :     int mode = MODE_INTER;
987 :     int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
988 :     const uint32_t iQuant = pMB->quant;
989 :    
990 :     const int skip_possible = (!(VolFlags & XVID_VOL_GMC)) && (pMB->dquant == 0);
991 :    
992 : edgomez 1053 if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) { /* normal, fast, SAD-based mode decision */
993 : edgomez 1022 int sad;
994 :     int InterBias = MV16_INTER_BIAS;
995 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
996 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
997 :     mode = MODE_INTER;
998 :     sad = Data->iMinSAD[0];
999 :     } else {
1000 :     mode = MODE_INTER4V;
1001 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1002 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1003 :     Data->iMinSAD[0] = sad;
1004 :     }
1005 :    
1006 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
1007 :     if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP))
1008 :     if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH)
1009 :     if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) {
1010 :     mode = MODE_NOT_CODED;
1011 :     sad = 0;
1012 :     }
1013 :    
1014 :     /* intra decision */
1015 :    
1016 : edgomez 1053 if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */
1017 : edgomez 1022 if (y != 0)
1018 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1019 :     if (x != 0)
1020 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1021 :    
1022 : edgomez 1053 if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? */
1023 : edgomez 1022 if (Data->rrv) InterBias *= 4;
1024 :    
1025 :     if (InterBias < pMB->sad16) {
1026 :     int32_t deviation;
1027 :     if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
1028 :     else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
1029 :     dev16(Data->Cur+16, Data->iEdgedWidth) +
1030 :     dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) +
1031 :     dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth);
1032 :    
1033 :     if (deviation < (sad - InterBias)) mode = MODE_INTRA;
1034 :     }
1035 :    
1036 : edgomez 1053 } else { /* BITS */
1037 : edgomez 1022
1038 :     int bits, intra, i;
1039 :     VECTOR backup[5], *v;
1040 :     Data->iQuant = iQuant;
1041 :    
1042 :     v = Data->qpel ? Data->currentQMV : Data->currentMV;
1043 :     for (i = 0; i < 5; i++) {
1044 :     Data->iMinSAD[i] = 256*4096;
1045 :     backup[i] = v[i];
1046 :     }
1047 :    
1048 :     bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1049 :     if (bits == 0)
1050 : edgomez 1053 mode = MODE_INTER; /* quick stop */
1051 : edgomez 1022 else {
1052 :     if (inter4v) {
1053 :     int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1054 :     if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; }
1055 :     }
1056 :    
1057 :     intra = CountMBBitsIntra(Data);
1058 :    
1059 :     if (intra < bits) { *Data->iMinSAD = bits = intra; mode = MODE_INTRA; }
1060 :     }
1061 :     }
1062 :    
1063 :     if (Data->rrv) {
1064 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1065 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1066 :     }
1067 :    
1068 :     if (mode == MODE_INTER) {
1069 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1070 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1071 :    
1072 :     if(Data->qpel) {
1073 :     pMB->qmvs[0] = pMB->qmvs[1]
1074 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1075 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1076 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1077 :     } else {
1078 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1079 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1080 :     }
1081 :    
1082 :     } else if (mode == MODE_INTER4V)
1083 :     pMB->sad16 = Data->iMinSAD[0];
1084 : edgomez 1053 else /* INTRA, NOT_CODED */
1085 : edgomez 1022 SkipMacroblockP(pMB, 0);
1086 :    
1087 :     pMB->mode = mode;
1088 :     }
1089 :    
1090 : edgomez 851 bool
1091 :     MotionEstimation(MBParam * const pParam,
1092 :     FRAMEINFO * const current,
1093 :     FRAMEINFO * const reference,
1094 :     const IMAGE * const pRefH,
1095 :     const IMAGE * const pRefV,
1096 :     const IMAGE * const pRefHV,
1097 : chl 1077 const IMAGE * const pGMC,
1098 : edgomez 851 const uint32_t iLimit)
1099 : chl 181 {
1100 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
1101 :     const IMAGE *const pCurrent = &current->image;
1102 :     const IMAGE *const pRef = &reference->image;
1103 : chl 181
1104 : edgomez 851 uint32_t mb_width = pParam->mb_width;
1105 :     uint32_t mb_height = pParam->mb_height;
1106 :     const uint32_t iEdgedWidth = pParam->edged_width;
1107 : suxen_drol 890 const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
1108 : chl 181
1109 : edgomez 851 uint32_t x, y;
1110 :     uint32_t iIntra = 0;
1111 : edgomez 953 int32_t quant = current->quant, sad00;
1112 : chl 1077 int skip_thresh = INITIAL_SKIP_THRESH * \
1113 : edgomez 953 (current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \
1114 :     (current->vop_flags & XVID_VOP_MODEDECISION_BITS ? 2:1);
1115 : chl 181
1116 : edgomez 1053 /* some pre-initialized thingies for SearchP */
1117 : edgomez 851 int32_t temp[8];
1118 :     VECTOR currentMV[5];
1119 :     VECTOR currentQMV[5];
1120 :     int32_t iMinSAD[5];
1121 : syskin 1075 DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE);
1122 : edgomez 851 SearchData Data;
1123 :     memset(&Data, 0, sizeof(SearchData));
1124 :     Data.iEdgedWidth = iEdgedWidth;
1125 :     Data.currentMV = currentMV;
1126 :     Data.currentQMV = currentQMV;
1127 :     Data.iMinSAD = iMinSAD;
1128 :     Data.temp = temp;
1129 :     Data.iFcode = current->fcode;
1130 :     Data.rounding = pParam->m_rounding_type;
1131 : edgomez 963 Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0);
1132 : edgomez 949 Data.chroma = MotionFlags & XVID_ME_CHROMA16;
1133 : chl 1077 Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0;
1134 : edgomez 962 Data.dctSpace = dct_space;
1135 : edgomez 1022 Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT);
1136 : edgomez 195
1137 : edgomez 949 if ((current->vop_flags & XVID_VOP_REDUCED)) {
1138 : edgomez 851 mb_width = (pParam->width + 31) / 32;
1139 :     mb_height = (pParam->height + 31) / 32;
1140 :     Data.qpel = 0;
1141 :     }
1142 : chl 181
1143 : edgomez 1053 Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */
1144 : edgomez 851 if (sadInit) (*sadInit) ();
1145 : chl 181
1146 : edgomez 851 for (y = 0; y < mb_height; y++) {
1147 :     for (x = 0; x < mb_width; x++) {
1148 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1149 : chl 181
1150 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1151 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1152 :     pRef->y + (x + y * iEdgedWidth) * 16,
1153 :     pParam->edged_width, pMB->sad8 );
1154 : chl 181
1155 : edgomez 851 else pMB->sad16 =
1156 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1157 :     pRef->y + (x + y * iEdgedWidth) * 32,
1158 :     pParam->edged_width, pMB->sad8 );
1159 : chl 181
1160 : edgomez 851 if (Data.chroma) {
1161 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1162 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1163 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1164 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1165 :     pMB->sad16 += Data.temp[7];
1166 :     }
1167 :    
1168 :     sad00 = pMB->sad16;
1169 :    
1170 : edgomez 953 if (pMB->dquant != 0) {
1171 :     quant += DQtab[pMB->dquant];
1172 :     if (quant > 31) quant = 31;
1173 :     else if (quant < 1) quant = 1;
1174 :     }
1175 : edgomez 1022 pMB->quant = quant;
1176 : edgomez 953
1177 : edgomez 1053 /* initial skip decision */
1178 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1179 : edgomez 949 if (!(current->vol_flags & XVID_VOL_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1180 : edgomez 953 if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh)
1181 : edgomez 851 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1182 :     SkipMacroblockP(pMB, sad00);
1183 :     continue;
1184 :     }
1185 :     }
1186 :    
1187 :     SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1188 : edgomez 1022 y, MotionFlags, current->vop_flags, current->vol_flags,
1189 :     &Data, pParam, pMBs, reference->mbs, pMB);
1190 : edgomez 851
1191 : edgomez 1022 ModeDecision(&Data, pMB, pMBs, x, y, pParam,
1192 :     MotionFlags, current->vop_flags, current->vol_flags,
1193 :     pCurrent, pRef);
1194 :    
1195 : edgomez 851 if (pMB->mode == MODE_INTRA)
1196 :     if (++iIntra > iLimit) return 1;
1197 : chl 181 }
1198 :     }
1199 : edgomez 851
1200 : chl 1077 // if (current->vol_flags & XVID_VOL_GMC ) /* GMC only for S(GMC)-VOPs */
1201 :     // {
1202 :     // current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1203 :     // }
1204 : edgomez 851 return 0;
1205 : chl 181 }
1206 :    
1207 :    
1208 : edgomez 851 static __inline int
1209 :     make_mask(const VECTOR * const pmv, const int i)
1210 : chl 96 {
1211 : edgomez 851 int mask = 255, j;
1212 :     for (j = 0; j < i; j++) {
1213 : edgomez 1053 if (MVequal(pmv[i], pmv[j])) return 0; /* same vector has been checked already */
1214 : edgomez 851 if (pmv[i].x == pmv[j].x) {
1215 :     if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1216 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1217 :     } else
1218 :     if (pmv[i].y == pmv[j].y) {
1219 :     if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1220 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1221 :     }
1222 :     }
1223 :     return mask;
1224 :     }
1225 : edgomez 195
1226 : edgomez 851 static __inline void
1227 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1228 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1229 :     {
1230 : edgomez 1053 /* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */
1231 : edgomez 851 if (rrv) { iWcount /= 2; iHcount /= 2; }
1232 : edgomez 195
1233 : edgomez 1053 if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */
1234 : edgomez 851 pmv[5].x = EVEN(pmv[3].x);
1235 :     pmv[5].y = EVEN(pmv[3].y);
1236 :     } else pmv[5].x = pmv[5].y = 0;
1237 : chl 96
1238 : edgomez 1053 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */
1239 : edgomez 851 else pmv[3].x = pmv[3].y = 0;
1240 : chl 96
1241 : edgomez 1053 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */
1242 : edgomez 851 else pmv[4].x = pmv[4].y = 0;
1243 : Isibaar 3
1244 : edgomez 1053 /* [1] median prediction */
1245 : edgomez 851 pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1246 : edgomez 195
1247 : 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 */
1248 : edgomez 195
1249 : edgomez 1053 pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */
1250 : edgomez 851 pmv[2].y = EVEN(prevMB->mvs[0].y);
1251 :    
1252 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1253 : edgomez 1053 pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */
1254 : edgomez 851 pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1255 :     } else pmv[6].x = pmv[6].y = 0;
1256 :    
1257 :     if (rrv) {
1258 :     int i;
1259 :     for (i = 0; i < 7; i++) {
1260 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1261 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1262 :     }
1263 :     }
1264 : Isibaar 3 }
1265 :    
1266 : edgomez 851 static void
1267 :     SearchP(const IMAGE * const pRef,
1268 :     const uint8_t * const pRefH,
1269 :     const uint8_t * const pRefV,
1270 :     const uint8_t * const pRefHV,
1271 :     const IMAGE * const pCur,
1272 :     const int x,
1273 :     const int y,
1274 :     const uint32_t MotionFlags,
1275 : suxen_drol 890 const uint32_t VopFlags,
1276 : edgomez 1022 const uint32_t VolFlags,
1277 : edgomez 851 SearchData * const Data,
1278 :     const MBParam * const pParam,
1279 :     const MACROBLOCK * const pMBs,
1280 :     const MACROBLOCK * const prevMBs,
1281 :     MACROBLOCK * const pMB)
1282 :     {
1283 : Isibaar 3
1284 : edgomez 851 int i, iDirection = 255, mask, threshA;
1285 :     VECTOR pmv[7];
1286 : edgomez 1022 int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
1287 : Isibaar 3
1288 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1289 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1290 : edgomez 195
1291 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1292 : Isibaar 3
1293 : edgomez 1053 Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */
1294 : edgomez 851 i = Data->rrv ? 2 : 1;
1295 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1296 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1297 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1298 : chl 326
1299 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1300 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1301 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1302 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1303 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1304 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1305 : edgomez 195
1306 : edgomez 1022 Data->lambda16 = lambda_vec16[pMB->quant];
1307 :     Data->lambda8 = lambda_vec8[pMB->quant];
1308 : edgomez 851 Data->qpel_precision = 0;
1309 : edgomez 195
1310 : edgomez 963 memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1311 : chl 169
1312 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1313 :     else Data->predMV = pmv[0];
1314 : Isibaar 3
1315 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1316 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1317 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1318 :     Data->iMinSAD[2] = pMB->sad8[1];
1319 :     Data->iMinSAD[3] = pMB->sad8[2];
1320 :     Data->iMinSAD[4] = pMB->sad8[3];
1321 : chl 169
1322 : syskin 1071 if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) && (x | y)) {
1323 : edgomez 1053 threshA = Data->temp[0]; /* that's where we keep this SAD atm */
1324 : edgomez 851 if (threshA < 512) threshA = 512;
1325 :     else if (threshA > 1024) threshA = 1024;
1326 :     } else
1327 :     threshA = 512;
1328 : chl 169
1329 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1330 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1331 : chl 169
1332 : edgomez 851 if (!Data->rrv) {
1333 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1334 : edgomez 1053 else CheckCandidate = CheckCandidate16no4v; /* for extra speed */
1335 : edgomez 851 } else CheckCandidate = CheckCandidate32;
1336 : chl 169
1337 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1338 : chl 169
1339 : edgomez 851 for (i = 1; i < 7; i++) {
1340 :     if (!(mask = make_mask(pmv, i)) ) continue;
1341 :     CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1342 :     if (Data->iMinSAD[0] <= threshA) break;
1343 :     }
1344 : chl 169
1345 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1346 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1347 : edgomez 1022 (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16)))
1348 :     inter4v = 0;
1349 : edgomez 851 else {
1350 : Isibaar 3
1351 : edgomez 851 MainSearchFunc * MainSearchPtr;
1352 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1353 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1354 : edgomez 851 else MainSearchPtr = DiamondSearch;
1355 : Isibaar 3
1356 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1357 : Isibaar 3
1358 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1359 :     note that this search is/might be done in halfpel positions,
1360 :     which makes it more different than the diamond above */
1361 : chl 169
1362 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH16) {
1363 : edgomez 851 int32_t bSAD;
1364 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1365 :     if (Data->rrv) {
1366 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1367 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1368 :     }
1369 :     if (!(MVequal(startMV, backupMV))) {
1370 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1371 : edgomez 195
1372 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1373 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1374 :     if (bSAD < Data->iMinSAD[0]) {
1375 :     Data->currentMV[0] = backupMV;
1376 :     Data->iMinSAD[0] = bSAD; }
1377 :     }
1378 : chl 169
1379 : edgomez 851 backupMV = Data->currentMV[0];
1380 :     startMV.x = startMV.y = 1;
1381 :     if (!(MVequal(startMV, backupMV))) {
1382 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1383 : chl 169
1384 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1385 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1386 :     if (bSAD < Data->iMinSAD[0]) {
1387 :     Data->currentMV[0] = backupMV;
1388 :     Data->iMinSAD[0] = bSAD; }
1389 : edgomez 195 }
1390 : edgomez 851 }
1391 :     }
1392 : edgomez 195
1393 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE16)
1394 : edgomez 851 SubpelRefine(Data);
1395 : edgomez 195
1396 : edgomez 851 for(i = 0; i < 5; i++) {
1397 : edgomez 1053 Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */
1398 : edgomez 851 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1399 :     }
1400 : Isibaar 3
1401 : edgomez 1023 if (Data->qpel) {
1402 : edgomez 953 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1403 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1404 : edgomez 987 Data->qpel_precision = 1;
1405 : edgomez 1023 if (MotionFlags & XVID_ME_QUARTERPELREFINE16)
1406 :     SubpelRefine(Data);
1407 : edgomez 953 }
1408 : Isibaar 3
1409 : edgomez 1023 if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30)
1410 : edgomez 1022 inter4v = 0;
1411 : Isibaar 3
1412 : edgomez 987 if (inter4v) {
1413 : edgomez 851 SearchData Data8;
1414 : edgomez 1053 memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */
1415 : Isibaar 3
1416 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1417 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1418 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1419 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1420 : chl 181
1421 : edgomez 949 if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_BITS))) {
1422 : edgomez 1053 /* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */
1423 : edgomez 851 int sumx = 0, sumy = 0;
1424 : Isibaar 3
1425 : edgomez 1022 if (Data->qpel)
1426 :     for (i = 1; i < 5; i++) {
1427 :     sumx += Data->currentQMV[i].x/2;
1428 :     sumy += Data->currentQMV[i].y/2;
1429 :     }
1430 :     else
1431 :     for (i = 1; i < 5; i++) {
1432 :     sumx += Data->currentMV[i].x;
1433 :     sumy += Data->currentMV[i].y;
1434 :     }
1435 : chl 259
1436 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1437 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1438 :     }
1439 : edgomez 1022 } else Data->iMinSAD[1] = 4096*256;
1440 : edgomez 851 }
1441 :    
1442 :     static void
1443 :     Search8(const SearchData * const OldData,
1444 :     const int x, const int y,
1445 :     const uint32_t MotionFlags,
1446 :     const MBParam * const pParam,
1447 :     MACROBLOCK * const pMB,
1448 :     const MACROBLOCK * const pMBs,
1449 :     const int block,
1450 :     SearchData * const Data)
1451 :     {
1452 :     int i = 0;
1453 :     Data->iMinSAD = OldData->iMinSAD + 1 + block;
1454 :     Data->currentMV = OldData->currentMV + 1 + block;
1455 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1456 :    
1457 :     if(Data->qpel) {
1458 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1459 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1460 :     Data->predMV, Data->iFcode, 0, 0);
1461 :     } else {
1462 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1463 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1464 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1465 : Isibaar 3 }
1466 :    
1467 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1468 : Isibaar 3
1469 : edgomez 949 if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) {
1470 : Isibaar 3
1471 : edgomez 978 if (Data->rrv) i = 16; else i = 8;
1472 : Isibaar 3
1473 : edgomez 978 Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1474 :     Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1475 :     Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1476 :     Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1477 :    
1478 :     Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1479 : edgomez 851 Data->qpel_precision = 0;
1480 : Isibaar 3
1481 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1482 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1483 : Isibaar 3
1484 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1485 :     else CheckCandidate = CheckCandidate16no4v;
1486 : Isibaar 3
1487 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_BITS))) {
1488 : edgomez 1053 int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */
1489 : Isibaar 3
1490 : edgomez 851 MainSearchFunc *MainSearchPtr;
1491 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch;
1492 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1493 : edgomez 851 else MainSearchPtr = DiamondSearch;
1494 : Isibaar 3
1495 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1496 : Isibaar 3
1497 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1498 : edgomez 1053 Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1499 : edgomez 851 Data->currentQMV->y = 2 * Data->currentMV->y;
1500 :     }
1501 :     }
1502 : edgomez 195
1503 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8) {
1504 : edgomez 1053 int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */
1505 : edgomez 195
1506 : edgomez 1053 SubpelRefine(Data); /* perform halfpel refine of current best vector */
1507 : Isibaar 3
1508 : edgomez 1053 if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */
1509 :     Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1510 : edgomez 851 Data->currentQMV->y = 2 * Data->currentMV->y;
1511 :     }
1512 :     }
1513 : Isibaar 3
1514 : edgomez 949 if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) {
1515 : edgomez 851 Data->qpel_precision = 1;
1516 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1517 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1518 :     SubpelRefine(Data);
1519 :     }
1520 :     }
1521 : edgomez 195
1522 : edgomez 851 if (Data->rrv) {
1523 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1524 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1525 :     }
1526 :    
1527 :     if(Data->qpel) {
1528 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1529 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1530 :     pMB->qmvs[block] = *Data->currentQMV;
1531 : edgomez 195 } else {
1532 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1533 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1534 : edgomez 78 }
1535 : edgomez 851
1536 :     pMB->mvs[block] = *Data->currentMV;
1537 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1538 : Isibaar 3 }
1539 :    
1540 : edgomez 851 /* motion estimation for B-frames */
1541 : chl 345
1542 : edgomez 851 static __inline VECTOR
1543 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1544 :     {
1545 :     /* the stupidiest function ever */
1546 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1547 :     }
1548 : chl 345
1549 : edgomez 851 static void __inline
1550 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1551 :     const uint32_t iWcount,
1552 :     const MACROBLOCK * const pMB,
1553 :     const uint32_t mode_curr)
1554 : chl 345 {
1555 :    
1556 : edgomez 1053 /* [0] is prediction */
1557 : edgomez 851 pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1558 : chl 345
1559 : edgomez 1053 pmv[1].x = pmv[1].y = 0; /* [1] is zero */
1560 : chl 345
1561 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1562 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1563 : chl 345
1564 : edgomez 1053 if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */
1565 : edgomez 851 pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1566 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1567 :     } else pmv[3].x = pmv[3].y = 0;
1568 : chl 345
1569 : edgomez 851 if (y != 0) {
1570 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1571 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1572 :     } else pmv[4].x = pmv[4].y = 0;
1573 : chl 345
1574 : edgomez 851 if (x != 0) {
1575 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1576 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1577 :     } else pmv[5].x = pmv[5].y = 0;
1578 : chl 345
1579 : edgomez 851 if (x != 0 && y != 0) {
1580 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1581 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1582 :     } else pmv[6].x = pmv[6].y = 0;
1583 :     }
1584 : chl 345
1585 :    
1586 : edgomez 851 /* search backward or forward */
1587 :     static void
1588 :     SearchBF( const IMAGE * const pRef,
1589 :     const uint8_t * const pRefH,
1590 :     const uint8_t * const pRefV,
1591 :     const uint8_t * const pRefHV,
1592 :     const IMAGE * const pCur,
1593 :     const int x, const int y,
1594 :     const uint32_t MotionFlags,
1595 :     const uint32_t iFcode,
1596 :     const MBParam * const pParam,
1597 :     MACROBLOCK * const pMB,
1598 :     const VECTOR * const predMV,
1599 :     int32_t * const best_sad,
1600 :     const int32_t mode_current,
1601 :     SearchData * const Data)
1602 :     {
1603 : chl 345
1604 : edgomez 851 int i, iDirection = 255, mask;
1605 :     VECTOR pmv[7];
1606 :     MainSearchFunc *MainSearchPtr;
1607 :     *Data->iMinSAD = MV_MAX_ERROR;
1608 :     Data->iFcode = iFcode;
1609 :     Data->qpel_precision = 0;
1610 : edgomez 1053 Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */
1611 : chl 345
1612 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16;
1613 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16;
1614 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16;
1615 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16;
1616 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1617 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1618 : chl 345
1619 : edgomez 851 Data->predMV = *predMV;
1620 : chl 345
1621 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1622 :     pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1623 : chl 345
1624 : edgomez 851 pmv[0] = Data->predMV;
1625 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1626 : chl 345
1627 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1628 : chl 345
1629 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1630 :     CheckCandidate = CheckCandidate16no4v;
1631 : chl 345
1632 : edgomez 1053 /* main loop. checking all predictions */
1633 : edgomez 851 for (i = 0; i < 7; i++) {
1634 :     if (!(mask = make_mask(pmv, i)) ) continue;
1635 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1636 : chl 345 }
1637 :    
1638 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1639 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1640 : edgomez 851 else MainSearchPtr = DiamondSearch;
1641 : chl 345
1642 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1643 : chl 345
1644 : edgomez 851 SubpelRefine(Data);
1645 : chl 345
1646 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1647 :     Data->currentQMV->x = 2*Data->currentMV->x;
1648 :     Data->currentQMV->y = 2*Data->currentMV->y;
1649 :     Data->qpel_precision = 1;
1650 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1651 :     pParam->width, pParam->height, iFcode, 1, 0);
1652 :     SubpelRefine(Data);
1653 :     }
1654 : chl 345
1655 : edgomez 1053 /* three bits are needed to code backward mode. four for forward */
1656 : Isibaar 3
1657 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1658 :     else *Data->iMinSAD += 3 * Data->lambda16;
1659 : edgomez 195
1660 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1661 :     *best_sad = *Data->iMinSAD;
1662 :     pMB->mode = mode_current;
1663 :     if (Data->qpel) {
1664 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1665 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1666 :     if (mode_current == MODE_FORWARD)
1667 :     pMB->qmvs[0] = *Data->currentQMV;
1668 :     else
1669 :     pMB->b_qmvs[0] = *Data->currentQMV;
1670 :     } else {
1671 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1672 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1673 :     }
1674 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1675 :     else pMB->b_mvs[0] = *Data->currentMV;
1676 :     }
1677 : edgomez 195
1678 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1679 : edgomez 1053 else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */
1680 : Isibaar 3 }
1681 :    
1682 : edgomez 851 static void
1683 :     SkipDecisionB(const IMAGE * const pCur,
1684 :     const IMAGE * const f_Ref,
1685 :     const IMAGE * const b_Ref,
1686 :     MACROBLOCK * const pMB,
1687 :     const uint32_t x, const uint32_t y,
1688 :     const SearchData * const Data)
1689 :     {
1690 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1691 :     int32_t sum;
1692 :     const int div = 1 + Data->qpel;
1693 :     int k;
1694 :     const uint32_t stride = Data->iEdgedWidth/2;
1695 : edgomez 1053 /* this is not full chroma compensation, only it's fullpel approximation. should work though */
1696 : Isibaar 3
1697 : edgomez 851 for (k = 0; k < 4; k++) {
1698 :     dy += Data->directmvF[k].y / div;
1699 : edgomez 963 dx += Data->directmvF[k].x / div;
1700 :     b_dy += Data->directmvB[k].y / div;
1701 :     b_dx += Data->directmvB[k].x / div;
1702 : edgomez 851 }
1703 : Isibaar 3
1704 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1705 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1706 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1707 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1708 : Isibaar 3
1709 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1710 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1711 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1712 :     stride);
1713 : Isibaar 3
1714 : edgomez 1053 if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; /* no skip */
1715 : Isibaar 3
1716 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1717 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1718 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1719 :     stride);
1720 : edgomez 195
1721 : edgomez 963 if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1722 : edgomez 1053 pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */
1723 : edgomez 978 for (k = 0; k < 4; k++) {
1724 : edgomez 963 pMB->qmvs[k] = pMB->mvs[k];
1725 :     pMB->b_qmvs[k] = pMB->b_mvs[k];
1726 :     }
1727 :     }
1728 : edgomez 851 }
1729 : Isibaar 3
1730 : edgomez 851 static __inline uint32_t
1731 :     SearchDirect(const IMAGE * const f_Ref,
1732 :     const uint8_t * const f_RefH,
1733 :     const uint8_t * const f_RefV,
1734 :     const uint8_t * const f_RefHV,
1735 :     const IMAGE * const b_Ref,
1736 :     const uint8_t * const b_RefH,
1737 :     const uint8_t * const b_RefV,
1738 :     const uint8_t * const b_RefHV,
1739 :     const IMAGE * const pCur,
1740 :     const int x, const int y,
1741 :     const uint32_t MotionFlags,
1742 :     const int32_t TRB, const int32_t TRD,
1743 :     const MBParam * const pParam,
1744 :     MACROBLOCK * const pMB,
1745 :     const MACROBLOCK * const b_mb,
1746 :     int32_t * const best_sad,
1747 :     SearchData * const Data)
1748 : Isibaar 3
1749 : edgomez 851 {
1750 :     int32_t skip_sad;
1751 :     int k = (x + Data->iEdgedWidth*y) * 16;
1752 :     MainSearchFunc *MainSearchPtr;
1753 : Isibaar 3
1754 : edgomez 851 *Data->iMinSAD = 256*4096;
1755 : edgomez 978 Data->RefP[0] = f_Ref->y + k;
1756 :     Data->RefP[2] = f_RefH + k;
1757 :     Data->RefP[1] = f_RefV + k;
1758 :     Data->RefP[3] = f_RefHV + k;
1759 :     Data->b_RefP[0] = b_Ref->y + k;
1760 :     Data->b_RefP[2] = b_RefH + k;
1761 :     Data->b_RefP[1] = b_RefV + k;
1762 :     Data->b_RefP[3] = b_RefHV + k;
1763 :     Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1764 :     Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1765 :     Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1766 :     Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1767 : edgomez 195
1768 : edgomez 851 k = Data->qpel ? 4 : 2;
1769 :     Data->max_dx = k * (pParam->width - x * 16);
1770 :     Data->max_dy = k * (pParam->height - y * 16);
1771 :     Data->min_dx = -k * (16 + x * 16);
1772 :     Data->min_dy = -k * (16 + y * 16);
1773 : chl 181
1774 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1775 :     Data->qpel_precision = 0;
1776 : edgomez 170
1777 : edgomez 851 for (k = 0; k < 4; k++) {
1778 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1779 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1780 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1781 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1782 : Isibaar 3
1783 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1784 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1785 :    
1786 : edgomez 1053 *best_sad = 256*4096; /* in that case, we won't use direct mode */
1787 :     pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */
1788 : edgomez 851 pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1789 :     return 256*4096;
1790 :     }
1791 :     if (b_mb->mode != MODE_INTER4V) {
1792 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1793 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1794 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1795 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1796 :     break;
1797 :     }
1798 : edgomez 195 }
1799 : Isibaar 3
1800 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1801 : Isibaar 3
1802 : edgomez 851 CheckCandidate(0, 0, 255, &k, Data);
1803 : edgomez 195
1804 : edgomez 1053 /* initial (fast) skip decision */
1805 : edgomez 1023 if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) {
1806 : edgomez 1053 /* possible skip */
1807 : edgomez 851 if (Data->chroma) {
1808 :     pMB->mode = MODE_DIRECT_NONE_MV;
1809 : edgomez 1053 return *Data->iMinSAD; /* skip. */
1810 : edgomez 851 } else {
1811 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1812 : edgomez 1053 if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */
1813 : edgomez 851 }
1814 : Isibaar 3 }
1815 :    
1816 : edgomez 953 *Data->iMinSAD += Data->lambda16;
1817 : edgomez 851 skip_sad = *Data->iMinSAD;
1818 : edgomez 1053
1819 :     /*
1820 :     * DIRECT MODE DELTA VECTOR SEARCH.
1821 :     * This has to be made more effective, but at the moment I'm happy it's running at all
1822 :     */
1823 : edgomez 195
1824 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1825 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1826 : edgomez 851 else MainSearchPtr = DiamondSearch;
1827 : Isibaar 3
1828 : edgomez 851 MainSearchPtr(0, 0, Data, 255);
1829 : Isibaar 3
1830 : edgomez 851 SubpelRefine(Data);
1831 : chl 181
1832 : edgomez 851 *best_sad = *Data->iMinSAD;
1833 : chl 181
1834 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1835 : edgomez 1053 else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */
1836 : chl 181
1837 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1838 : edgomez 195
1839 : edgomez 851 for (k = 0; k < 4; k++) {
1840 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1841 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1842 :     ? Data->directmvB[k].x
1843 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1844 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1845 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1846 :     ? Data->directmvB[k].y
1847 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1848 :     if (Data->qpel) {
1849 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1850 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1851 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1852 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1853 :     }
1854 : edgomez 195
1855 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1856 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1857 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1858 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1859 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1860 :     break;
1861 :     }
1862 : edgomez 78 }
1863 : edgomez 851 return skip_sad;
1864 :     }
1865 : Isibaar 3
1866 : edgomez 851 static void
1867 :     SearchInterpolate(const IMAGE * const f_Ref,
1868 :     const uint8_t * const f_RefH,
1869 :     const uint8_t * const f_RefV,
1870 :     const uint8_t * const f_RefHV,
1871 :     const IMAGE * const b_Ref,
1872 :     const uint8_t * const b_RefH,
1873 :     const uint8_t * const b_RefV,
1874 :     const uint8_t * const b_RefHV,
1875 :     const IMAGE * const pCur,
1876 :     const int x, const int y,
1877 :     const uint32_t fcode,
1878 :     const uint32_t bcode,
1879 :     const uint32_t MotionFlags,
1880 :     const MBParam * const pParam,
1881 :     const VECTOR * const f_predMV,
1882 :     const VECTOR * const b_predMV,
1883 :     MACROBLOCK * const pMB,
1884 :     int32_t * const best_sad,
1885 :     SearchData * const fData)
1886 : chl 96
1887 : edgomez 851 {
1888 : chl 169
1889 : edgomez 851 int iDirection, i, j;
1890 :     SearchData bData;
1891 : chl 169
1892 : edgomez 851 fData->qpel_precision = 0;
1893 : edgomez 1053 memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */
1894 : edgomez 851 *fData->iMinSAD = 4096*256;
1895 :     bData.currentMV++; bData.currentQMV++;
1896 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1897 : chl 169
1898 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1899 : chl 169
1900 : edgomez 978 bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i;
1901 :     bData.b_RefP[2] = fData->RefP[2] = f_RefH + i;
1902 :     bData.b_RefP[1] = fData->RefP[1] = f_RefV + i;
1903 :     bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i;
1904 :     bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i;
1905 :     bData.RefP[2] = fData->b_RefP[2] = b_RefH + i;
1906 :     bData.RefP[1] = fData->b_RefP[1] = b_RefV + i;
1907 :     bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i;
1908 :     bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1909 :     bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1910 :     bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1911 :     bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1912 : chl 169
1913 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1914 :     fData->bpredMV = bData.predMV = *b_predMV;
1915 :     fData->currentMV[0] = fData->currentMV[2];
1916 : Isibaar 3
1917 : edgomez 851 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0);
1918 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0);
1919 : Isibaar 3
1920 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1921 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1922 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1923 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1924 : chl 169
1925 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1926 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1927 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1928 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1929 : Isibaar 3
1930 : edgomez 851 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1931 : chl 169
1932 : edgomez 1053 /* diamond */
1933 : edgomez 851 do {
1934 :     iDirection = 255;
1935 : edgomez 1053 /* forward MV moves */
1936 : edgomez 851 i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1937 : chl 169
1938 : edgomez 851 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1939 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1940 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1941 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1942 : edgomez 195
1943 : edgomez 1053 /* backward MV moves */
1944 : edgomez 851 i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1945 :     fData->currentMV[2] = fData->currentMV[0];
1946 :     CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1947 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1948 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1949 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1950 : Isibaar 3
1951 : edgomez 851 } while (!(iDirection));
1952 : chl 140
1953 : edgomez 1053 /* qpel refinement */
1954 : edgomez 851 if (fData->qpel) {
1955 :     if (*fData->iMinSAD > *best_sad + 500) return;
1956 :     CheckCandidate = CheckCandidateInt;
1957 :     fData->qpel_precision = bData.qpel_precision = 1;
1958 :     get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1959 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1960 :     fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1961 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1962 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1963 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1964 :     SubpelRefine(fData);
1965 :     if (*fData->iMinSAD > *best_sad + 300) return;
1966 :     fData->currentQMV[2] = fData->currentQMV[0];
1967 :     SubpelRefine(&bData);
1968 :     }
1969 : chl 140
1970 : edgomez 1053 *fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */
1971 : Isibaar 3
1972 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
1973 :     *best_sad = *fData->iMinSAD;
1974 :     pMB->mvs[0] = fData->currentMV[0];
1975 :     pMB->b_mvs[0] = fData->currentMV[1];
1976 :     pMB->mode = MODE_INTERPOLATE;
1977 :     if (fData->qpel) {
1978 :     pMB->qmvs[0] = fData->currentQMV[0];
1979 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1980 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1981 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1982 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1983 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1984 :     } else {
1985 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1986 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1987 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1988 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1989 :     }
1990 : edgomez 78 }
1991 : edgomez 851 }
1992 : Isibaar 3
1993 : edgomez 851 void
1994 :     MotionEstimationBVOP(MBParam * const pParam,
1995 :     FRAMEINFO * const frame,
1996 :     const int32_t time_bp,
1997 :     const int32_t time_pp,
1998 : edgomez 1053 /* forward (past) reference */
1999 : edgomez 851 const MACROBLOCK * const f_mbs,
2000 :     const IMAGE * const f_ref,
2001 :     const IMAGE * const f_refH,
2002 :     const IMAGE * const f_refV,
2003 :     const IMAGE * const f_refHV,
2004 : edgomez 1053 /* backward (future) reference */
2005 : edgomez 851 const FRAMEINFO * const b_reference,
2006 :     const IMAGE * const b_ref,
2007 :     const IMAGE * const b_refH,
2008 :     const IMAGE * const b_refV,
2009 :     const IMAGE * const b_refHV)
2010 :     {
2011 :     uint32_t i, j;
2012 :     int32_t best_sad;
2013 :     uint32_t skip_sad;
2014 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2015 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
2016 : Isibaar 3
2017 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2018 : Isibaar 3
2019 : edgomez 851 const int32_t TRB = time_pp - time_bp;
2020 :     const int32_t TRD = time_pp;
2021 : edgomez 195
2022 : edgomez 1053 /* some pre-inintialized data for the rest of the search */
2023 : Isibaar 3
2024 : edgomez 851 SearchData Data;
2025 :     int32_t iMinSAD;
2026 :     VECTOR currentMV[3];
2027 :     VECTOR currentQMV[3];
2028 :     int32_t temp[8];
2029 :     memset(&Data, 0, sizeof(SearchData));
2030 :     Data.iEdgedWidth = pParam->edged_width;
2031 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2032 :     Data.iMinSAD = &iMinSAD;
2033 :     Data.lambda16 = lambda_vec16[frame->quant];
2034 : edgomez 949 Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL;
2035 : edgomez 851 Data.rounding = 0;
2036 : edgomez 949 Data.chroma = frame->motion_flags & XVID_ME_CHROMA8;
2037 : edgomez 851 Data.temp = temp;
2038 : Isibaar 3
2039 : edgomez 1053 Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */
2040 :    
2041 :     /* note: i==horizontal, j==vertical */
2042 : edgomez 851 for (j = 0; j < pParam->mb_height; j++) {
2043 : edgomez 195
2044 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2045 : Isibaar 3
2046 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2047 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2048 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2049 : edgomez 195
2050 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2051 :     if (b_reference->coding_type != S_VOP)
2052 :     if (b_mb->mode == MODE_NOT_CODED) {
2053 :     pMB->mode = MODE_NOT_CODED;
2054 :     continue;
2055 :     }
2056 : Isibaar 3
2057 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2058 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2059 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2060 : edgomez 953 pMB->quant = frame->quant;
2061 : Isibaar 3
2062 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2063 :     and (2) sets very good predictions for forward and backward search */
2064 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2065 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2066 :     &frame->image,
2067 :     i, j,
2068 :     frame->motion_flags,
2069 :     TRB, TRD,
2070 :     pParam,
2071 :     pMB, b_mb,
2072 :     &best_sad,
2073 :     &Data);
2074 : Isibaar 3
2075 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2076 : edgomez 195
2077 : edgomez 1053 /* forward search */
2078 : edgomez 851 SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2079 :     &frame->image, i, j,
2080 :     frame->motion_flags,
2081 :     frame->fcode, pParam,
2082 :     pMB, &f_predMV, &best_sad,
2083 :     MODE_FORWARD, &Data);
2084 : edgomez 195
2085 : edgomez 1053 /* backward search */
2086 : edgomez 851 SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2087 :     &frame->image, i, j,
2088 :     frame->motion_flags,
2089 :     frame->bcode, pParam,
2090 :     pMB, &b_predMV, &best_sad,
2091 :     MODE_BACKWARD, &Data);
2092 :    
2093 : edgomez 1053 /* interpolate search comes last, because it uses data from forward and backward as prediction */
2094 : edgomez 851 SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2095 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2096 :     &frame->image,
2097 :     i, j,
2098 :     frame->fcode, frame->bcode,
2099 :     frame->motion_flags,
2100 :     pParam,
2101 :     &f_predMV, &b_predMV,
2102 :     pMB, &best_sad,
2103 :     &Data);
2104 :    
2105 : edgomez 1053 /* final skip decision */
2106 : edgomez 851 if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2107 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2108 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2109 :    
2110 :     switch (pMB->mode) {
2111 :     case MODE_FORWARD:
2112 :     f_count++;
2113 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2114 :     break;
2115 :     case MODE_BACKWARD:
2116 :     b_count++;
2117 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2118 :     break;
2119 :     case MODE_INTERPOLATE:
2120 :     i_count++;
2121 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2122 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2123 :     break;
2124 :     case MODE_DIRECT:
2125 :     case MODE_DIRECT_NO4V:
2126 :     d_count++;
2127 :     default:
2128 :     break;
2129 :     }
2130 :     }
2131 :     }
2132 : Isibaar 3 }
2133 : chl 96
2134 : edgomez 851 static __inline void
2135 :     MEanalyzeMB ( const uint8_t * const pRef,
2136 :     const uint8_t * const pCur,
2137 :     const int x,
2138 :     const int y,
2139 :     const MBParam * const pParam,
2140 :     MACROBLOCK * const pMBs,
2141 :     SearchData * const Data)
2142 : chl 96 {
2143 :    
2144 : edgomez 851 int i, mask;
2145 : edgomez 953 int quarterpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0;
2146 : edgomez 851 VECTOR pmv[3];
2147 : edgomez 953 MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2148 : chl 96
2149 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2150 : chl 96
2151 : edgomez 1053 /* median is only used as prediction. it doesn't have to be real */
2152 : edgomez 851 if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2153 :     else
2154 : edgomez 1053 if (x == 1) /* left macroblock does not have any vector now */
2155 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */
2156 :     else if (y == 1) /* top macroblock doesn't have it's vector */
2157 :     Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */
2158 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */
2159 : edgomez 195
2160 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2161 : edgomez 953 pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0);
2162 : edgomez 195
2163 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2164 : edgomez 978 Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16;
2165 : chl 96
2166 : edgomez 851 pmv[1].x = EVEN(pMB->mvs[0].x);
2167 :     pmv[1].y = EVEN(pMB->mvs[0].y);
2168 :     pmv[2].x = EVEN(Data->predMV.x);
2169 :     pmv[2].y = EVEN(Data->predMV.y);
2170 :     pmv[0].x = pmv[0].y = 0;
2171 : edgomez 195
2172 : edgomez 851 CheckCandidate32I(0, 0, 255, &i, Data);
2173 : edgomez 195
2174 : edgomez 953 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) {
2175 : chl 96
2176 : edgomez 851 if (!(mask = make_mask(pmv, 1)))
2177 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2178 :     if (!(mask = make_mask(pmv, 2)))
2179 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2180 : chl 96
2181 : edgomez 1053 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) /* diamond only if needed */
2182 : edgomez 851 DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2183 : edgomez 953 }
2184 : edgomez 851
2185 : edgomez 953 for (i = 0; i < 4; i++) {
2186 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2187 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2188 :     MB->mode = MODE_INTER;
2189 :     MB->sad16 = Data->iMinSAD[i+1];
2190 : chl 96 }
2191 : edgomez 851 }
2192 : chl 96
2193 : syskin 1075 #define INTRA_THRESH 2200
2194 :     #define INTER_THRESH 50
2195 :     #define INTRA_THRESH2 95
2196 : chl 96
2197 : edgomez 851 int
2198 :     MEanalysis( const IMAGE * const pRef,
2199 : edgomez 953 const FRAMEINFO * const Current,
2200 :     const MBParam * const pParam,
2201 : syskin 1075 const int maxIntra, //maximum number if non-I frames
2202 :     const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2203 :     const int bCount, // number of B frames in a row
2204 : edgomez 953 const int b_thresh)
2205 : edgomez 851 {
2206 :     uint32_t x, y, intra = 0;
2207 :     int sSAD = 0;
2208 :     MACROBLOCK * const pMBs = Current->mbs;
2209 :     const IMAGE * const pCurrent = &Current->image;
2210 : syskin 1075 int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh;
2211 : chl 1077 int blocks = 0;
2212 : syskin 1075 int complexity = 0;
2213 : chl 96
2214 : edgomez 851 int32_t iMinSAD[5], temp[5];
2215 :     VECTOR currentMV[5];
2216 :     SearchData Data;
2217 :     Data.iEdgedWidth = pParam->edged_width;
2218 :     Data.currentMV = currentMV;
2219 :     Data.iMinSAD = iMinSAD;
2220 :     Data.iFcode = Current->fcode;
2221 :     Data.temp = temp;
2222 :     CheckCandidate = CheckCandidate32I;
2223 : chl 96
2224 : syskin 1075
2225 : edgomez 1023 if (intraCount != 0) {
2226 : syskin 1075 if (intraCount < 10) // we're right after an I frame
2227 : edgomez 1023 IntraThresh += 15* (intraCount - 10) * (intraCount - 10);
2228 :     else
2229 : syskin 1075 if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2230 : edgomez 1023 IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra;
2231 :     }
2232 : chl 96
2233 : syskin 1075 InterThresh -= 12 * bCount;
2234 :     if (InterThresh < 15 + b_thresh) InterThresh = 15 + b_thresh;
2235 : chl 326
2236 : edgomez 851 if (sadInit) (*sadInit) ();
2237 : chl 96
2238 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2239 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2240 :     int i;
2241 : syskin 1075 blocks += 10;
2242 : edgomez 195
2243 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2244 : syskin 1075 else { //extrapolation of the vector found for last frame
2245 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].x =
2246 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2247 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].y =
2248 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2249 :     }
2250 : edgomez 195
2251 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2252 : edgomez 195
2253 : edgomez 851 for (i = 0; i < 4; i++) {
2254 :     int dev;
2255 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2256 : syskin 1075 dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2257 :     pParam->edged_width);
2258 :    
2259 :     complexity += dev;
2260 :     if (dev + IntraThresh < pMB->sad16) {
2261 :     pMB->mode = MODE_INTRA;
2262 :     if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2263 : edgomez 851 }
2264 : edgomez 953
2265 : syskin 1075 if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0)
2266 :     if (dev > 500 && pMB->sad16 < 1000)
2267 :     sSAD += 1000;
2268 :    
2269 : edgomez 851 sSAD += pMB->sad16;
2270 :     }
2271 :     }
2272 : edgomez 195 }
2273 : syskin 1075 complexity >>= 7;
2274 : edgomez 953
2275 : syskin 1075 sSAD /= complexity + 4*blocks;
2276 : edgomez 953
2277 : syskin 1075 if (intraCount > 12 && sSAD > INTRA_THRESH2 ) return I_VOP;
2278 : edgomez 851 if (sSAD > InterThresh ) return P_VOP;
2279 :     emms();
2280 :     return B_VOP;
2281 :     }
2282 : chl 96
2283 :    
2284 : edgomez 1053 /* functions which perform BITS-based search/bitcount */
2285 : chl 96
2286 : edgomez 851 static int
2287 :     CountMBBitsInter(SearchData * const Data,
2288 :     const MACROBLOCK * const pMBs, const int x, const int y,
2289 :     const MBParam * const pParam,
2290 :     const uint32_t MotionFlags)
2291 : chl 96 {
2292 : edgomez 851 int i, iDirection;
2293 :     int32_t bsad[5];
2294 : chl 141
2295 : edgomez 851 CheckCandidate = CheckCandidateBits16;
2296 : chl 96
2297 : edgomez 851 if (Data->qpel) {
2298 :     for(i = 0; i < 5; i++) {
2299 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2300 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2301 :     }
2302 :     Data->qpel_precision = 1;
2303 :     CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2304 : chl 96
2305 : edgomez 1053 if (MotionFlags & (XVID_ME_HALFPELREFINE16_BITS | XVID_ME_EXTSEARCH_BITS)) { /* we have to prepare for halfpixel-precision search */
2306 : edgomez 851 for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2307 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2308 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2309 :     Data->qpel_precision = 0;
2310 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2311 :     CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2312 :     }
2313 : edgomez 195
2314 : edgomez 1053 } else { /* not qpel */
2315 : edgomez 195
2316 : edgomez 851 CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2317 :     }
2318 : chl 96
2319 : edgomez 949 if (MotionFlags&XVID_ME_EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2320 : chl 96
2321 : edgomez 949 if (MotionFlags&XVID_ME_HALFPELREFINE16_BITS) SubpelRefine(Data);
2322 : chl 96
2323 : edgomez 851 if (Data->qpel) {
2324 : edgomez 1053 if (MotionFlags&(XVID_ME_EXTSEARCH_BITS | XVID_ME_HALFPELREFINE16_BITS)) { /* there was halfpel-precision search */
2325 : edgomez 851 for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2326 : edgomez 1053 Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* we have found a better match */
2327 : edgomez 851 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2328 :     }
2329 : edgomez 195
2330 : edgomez 1053 /* preparing for qpel-precision search */
2331 : edgomez 851 Data->qpel_precision = 1;
2332 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2333 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
2334 :     }
2335 : edgomez 949 if (MotionFlags&XVID_ME_QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2336 : edgomez 851 }
2337 : chl 96
2338 : edgomez 1053 if (MotionFlags&XVID_ME_CHECKPREDICTION_BITS) { /* let's check vector equal to prediction */
2339 : edgomez 851 VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2340 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2341 :     CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2342 :     }
2343 :     return Data->iMinSAD[0];
2344 :     }
2345 : chl 96
2346 : edgomez 851 static int
2347 :     CountMBBitsInter4v(const SearchData * const Data,
2348 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2349 :     const int x, const int y,
2350 :     const MBParam * const pParam, const uint32_t MotionFlags,
2351 :     const VECTOR * const backup)
2352 :     {
2353 : chl 96
2354 : edgomez 851 int cbp = 0, bits = 0, t = 0, i, iDirection;
2355 :     SearchData Data2, *Data8 = &Data2;
2356 :     int sumx = 0, sumy = 0;
2357 : edgomez 962 int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64;
2358 : edgomez 1022 uint8_t * ptr;
2359 : chl 96
2360 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2361 :     CheckCandidate = CheckCandidateBits8;
2362 : chl 96
2363 : edgomez 1053 for (i = 0; i < 4; i++) { /* for all luma blocks */
2364 : edgomez 1022
2365 : edgomez 851 Data8->iMinSAD = Data->iMinSAD + i + 1;
2366 :     Data8->currentMV = Data->currentMV + i + 1;
2367 :     Data8->currentQMV = Data->currentQMV + i + 1;
2368 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2369 : edgomez 978 Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2370 :     Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2371 :     Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2372 :     Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2373 : edgomez 195
2374 : edgomez 851 if(Data->qpel) {
2375 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2376 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2377 :     Data8->predMV, Data8->iFcode, 0, 0);
2378 :     } else {
2379 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2380 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2381 :     Data8->predMV, Data8->iFcode, 0, 0);
2382 :     }
2383 : chl 96
2384 : edgomez 851 get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2385 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2386 : chl 96
2387 : edgomez 1023 *Data8->iMinSAD += BITS_MULT*t;
2388 : chl 96
2389 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2390 : edgomez 1053 /* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */
2391 : edgomez 987 {
2392 :     VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV;
2393 : edgomez 1022 if (!MVequal (*v, backup[i+1]) )
2394 : edgomez 851 CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2395 :     }
2396 : edgomez 195
2397 : edgomez 851 if (Data8->qpel) {
2398 : edgomez 1053 if (MotionFlags&XVID_ME_HALFPELREFINE8_BITS || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_BITS)) { /* halfpixel motion search follows */
2399 : edgomez 851 int32_t s = *Data8->iMinSAD;
2400 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2401 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2402 :     Data8->qpel_precision = 0;
2403 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2404 :     pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2405 : edgomez 195
2406 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2407 :     CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2408 : edgomez 195
2409 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS)
2410 : edgomez 851 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2411 : chl 96
2412 : edgomez 1022 if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS)
2413 :     SubpelRefine(Data8);
2414 : chl 96
2415 : edgomez 1053 if(s > *Data8->iMinSAD) { /* we have found a better match */
2416 : edgomez 851 Data8->currentQMV->x = 2*Data8->currentMV->x;
2417 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2418 :     }
2419 : chl 96
2420 : edgomez 851 Data8->qpel_precision = 1;
2421 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2422 :     pParam->width, pParam->height, Data8->iFcode, 1, 0);
2423 : chl 96
2424 : edgomez 851 }
2425 : edgomez 949 if (MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2426 : edgomez 195
2427 : edgomez 1053 } else { /* not qpel */
2428 : edgomez 851
2429 : edgomez 1053 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) /* extsearch */
2430 : edgomez 1022 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2431 :    
2432 :     if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS)
2433 : edgomez 1053 SubpelRefine(Data8); /* halfpel refinement */
2434 : edgomez 1022 }
2435 :    
2436 : edgomez 1053 /* checking vector equal to predicion */
2437 : edgomez 949 if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_BITS) {
2438 : edgomez 851 const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2439 : edgomez 1022 if (!MVequal(*v, Data8->predMV))
2440 : edgomez 851 CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2441 : chl 141 }
2442 : edgomez 195
2443 : edgomez 851 bits += *Data8->iMinSAD;
2444 : edgomez 1053 if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */
2445 : edgomez 851
2446 : edgomez 1053 /* MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else */
2447 : edgomez 851 if(Data->qpel) {
2448 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2449 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2450 :     pMB->qmvs[i] = *Data8->currentQMV;
2451 :     sumx += Data8->currentQMV->x/2;
2452 :     sumy += Data8->currentQMV->y/2;
2453 :     } else {
2454 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2455 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2456 :     sumx += Data8->currentMV->x;
2457 :     sumy += Data8->currentMV->y;
2458 : chl 141 }
2459 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2460 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2461 :     if (Data8->temp[0]) cbp |= 1 << (5 - i);
2462 : chl 141
2463 : edgomez 1053 } /* /for all luma blocks */
2464 : chl 141
2465 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
2466 : chl 141
2467 : edgomez 1053 /* let's check chroma */
2468 : edgomez 1022 sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2469 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2470 : chl 96
2471 : edgomez 1053 /* chroma U */
2472 : edgomez 1022 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2473 :     transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2474 : syskin 1075 bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4);
2475 : edgomez 1022
2476 :     if (bits >= *Data->iMinSAD) return bits;
2477 :    
2478 : edgomez 1053 /* chroma V */
2479 : edgomez 1022 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2480 :     transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2481 : syskin 1075 bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5);
2482 : edgomez 1022
2483 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2484 : edgomez 1022
2485 : edgomez 851 return bits;
2486 :     }
2487 : chl 96
2488 : edgomez 851 static int
2489 :     CountMBBitsIntra(const SearchData * const Data)
2490 :     {
2491 : edgomez 1053 int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */
2492 : edgomez 1023 int cbp = 0, i, dc = 0;
2493 : edgomez 962 int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2494 : edgomez 195
2495 : edgomez 851 for(i = 0; i < 4; i++) {
2496 :     int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2497 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2498 : syskin 1075 bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc);
2499 : chl 141
2500 : edgomez 1022 if (bits >= Data->iMinSAD[0]) return bits;
2501 : edgomez 851 }
2502 : chl 181
2503 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len;
2504 : chl 96
2505 : edgomez 1053 /*chroma U */
2506 : edgomez 1022 transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2507 : syskin 1075 bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc);
2508 : edgomez 1023
2509 : edgomez 1022 if (bits >= Data->iMinSAD[0]) return bits;
2510 :    
2511 : edgomez 1053 /* chroma V */
2512 : edgomez 1022 transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2513 : syskin 1075 bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc);
2514 : edgomez 1022
2515 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2516 : edgomez 1022
2517 : edgomez 851 return bits;
2518 : chl 96 }
2519 : chl 1077
2520 :    
2521 :    
2522 :    
2523 :    
2524 :     static __inline void
2525 :     GMEanalyzeMB ( const uint8_t * const pCur,
2526 :     const uint8_t * const pRef,
2527 :     const uint8_t * const pRefH,
2528 :     const uint8_t * const pRefV,
2529 :     const uint8_t * const pRefHV,
2530 :     const int x,
2531 :     const int y,
2532 :     const MBParam * const pParam,
2533 :     MACROBLOCK * const pMBs,
2534 :     SearchData * const Data)
2535 :     {
2536 :    
2537 :     int i=0;
2538 :     // VECTOR pmv[3];
2539 :     MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2540 :    
2541 :     Data->iMinSAD[0] = MV_MAX_ERROR;
2542 :    
2543 :     //median is only used as prediction. it doesn't have to be real
2544 :     if (x == 0 && y == 0)
2545 :     Data->predMV.x = Data->predMV.y = 0;
2546 :     else
2547 :     if (x == 0) //left macroblock does not have any vector now
2548 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2549 :     else if (y == 0) // top macroblock doesn't have it's vector
2550 :     Data->predMV = (pMB-1)->mvs[0]; // left instead of median
2551 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2552 :    
2553 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2554 :     pParam->width, pParam->height, Data->iFcode - ((pParam->vol_flags & XVID_VOL_QUARTERPEL)?1:0), 0, 0);
2555 :    
2556 :     Data->Cur = pCur + 16*(x + y * pParam->edged_width);
2557 :     Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width);
2558 :     Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width);
2559 :     Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width);
2560 :     Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width);
2561 :    
2562 :     Data->currentMV[0].x = Data->currentMV[0].y = 0;
2563 :     CheckCandidate16I(0, 0, 255, &i, Data);
2564 :    
2565 :     if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) )
2566 :     CheckCandidate16I(Data->predMV.x, Data->predMV.y, 255, &i, Data);
2567 :    
2568 :     if (Data->iMinSAD[0] > 256 /*4 * MAX_SAD00_FOR_SKIP*/) // diamond only if needed
2569 :     DiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255);
2570 :    
2571 :     SubpelRefine(Data);
2572 :    
2573 :    
2574 :     /* for QPel halfpel positions are worse than in halfpel mode :( */
2575 :     /* if (Data->qpel) {
2576 :     Data->currentQMV->x = 2*Data->currentMV->x;
2577 :     Data->currentQMV->y = 2*Data->currentMV->y;
2578 :     Data->qpel_precision = 1;
2579 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2580 :     pParam->width, pParam->height, iFcode, 1, 0);
2581 :     SubpelRefine(Data);
2582 :     }
2583 :     */
2584 :    
2585 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
2586 :     pMB->sad16 = Data->iMinSAD[0];
2587 :     pMB->sad16 += d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0);
2588 :     return;
2589 :     }
2590 :    
2591 :     void
2592 :     GMEanalysis(const MBParam * const pParam,
2593 :     const FRAMEINFO * const current,
2594 :     const FRAMEINFO * const reference,
2595 :     const IMAGE * const pRefH,
2596 :     const IMAGE * const pRefV,
2597 :     const IMAGE * const pRefHV)
2598 :     {
2599 :     uint32_t x, y;
2600 :     MACROBLOCK * const pMBs = current->mbs;
2601 :     const IMAGE * const pCurrent = &current->image;
2602 :     const IMAGE * const pReference = &reference->image;
2603 :    
2604 :     int32_t iMinSAD[5], temp[5];
2605 :     VECTOR currentMV[5];
2606 :     SearchData Data;
2607 :     memset(&Data, 0, sizeof(SearchData));
2608 :    
2609 :     Data.iEdgedWidth = pParam->edged_width;
2610 :     Data.qpel = ((pParam->vol_flags & XVID_VOL_QUARTERPEL)?1:0);
2611 :     Data.qpel_precision = 0;
2612 :     Data.rounding = pParam->m_rounding_type;
2613 :     Data.chroma = current->motion_flags & XVID_ME_CHROMA16;
2614 :     Data.rrv = current->vop_flags & XVID_VOL_REDUCED_ENABLE;
2615 :    
2616 :     Data.currentMV = &currentMV[0];
2617 :     Data.iMinSAD = &iMinSAD[0];
2618 :     Data.iFcode = current->fcode;
2619 :     Data.temp = temp;
2620 :     Data.RefP[0] = pReference->y;
2621 :     Data.RefP[1] = pRefV->y;
2622 :     Data.RefP[2] = pRefH->y;
2623 :     Data.RefP[3] = pRefHV->y;
2624 :    
2625 :     CheckCandidate = CheckCandidate16I;
2626 :    
2627 :     if (sadInit) (*sadInit) ();
2628 :    
2629 :     for (y = 0; y < pParam->mb_height; y ++) {
2630 :     for (x = 0; x < pParam->mb_width; x ++) {
2631 :    
2632 :     GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data);
2633 :     }
2634 :     }
2635 :     return;
2636 :     }
2637 :    
2638 :    
2639 :     WARPPOINTS
2640 :     GlobalMotionEst(MACROBLOCK * const pMBs,
2641 :     const MBParam * const pParam,
2642 :     const FRAMEINFO * const current,
2643 :     const FRAMEINFO * const reference,
2644 :     const IMAGE * const pRefH,
2645 :     const IMAGE * const pRefV,
2646 :     const IMAGE * const pRefHV)
2647 :     {
2648 :    
2649 :     const unsigned int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2650 :     const unsigned int deltay=8;
2651 :     const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks)
2652 :     const unsigned int grady=512;
2653 :    
2654 :     double sol[4] = { 0., 0., 0., 0. };
2655 :    
2656 :     WARPPOINTS gmc;
2657 :    
2658 :     uint32_t mx, my;
2659 :    
2660 :     int MBh = pParam->mb_height;
2661 :     int MBw = pParam->mb_width;
2662 :     const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */
2663 :     const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */
2664 :    
2665 :     int num=0;
2666 :     int oldnum;
2667 :    
2668 :     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;
2669 :    
2670 :     GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV);
2671 :    
2672 :     /* block based ME isn't done, yet, so do a quick presearch */
2673 :    
2674 :     // filter mask of all blocks
2675 :    
2676 :     for (my = 0; my < (uint32_t)MBh; my++)
2677 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2678 :     {
2679 :     const int mbnum = mx + my * MBw;
2680 :     pMBs[mbnum].mcsel = 0;
2681 :     }
2682 :    
2683 :    
2684 :     for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */
2685 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */
2686 :     {
2687 :     const int mbnum = mx + my * MBw;
2688 :     MACROBLOCK *const pMB = &pMBs[mbnum];
2689 :     const VECTOR mv = pMB->mvs[0];
2690 :    
2691 :     /* don't use object boundaries */
2692 :     if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax)
2693 :     && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay)
2694 :     && (abs(mv.x - (pMB+1)->mvs[0].x) < deltax)
2695 :     && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay)
2696 :     && (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax)
2697 :     && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay)
2698 :     && (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax)
2699 :     && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) )
2700 :     { const int iEdgedWidth = pParam->edged_width;
2701 :     const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx);
2702 :     if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx )
2703 :     && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) )
2704 :     { pMB->mcsel = 1;
2705 :     num++;
2706 :     }
2707 :    
2708 :     /* only use "structured" blocks */
2709 :     }
2710 :     }
2711 :     emms();
2712 :    
2713 :     /* further filtering would be possible, but during iteration, remaining
2714 :     outliers usually are removed, too */
2715 :    
2716 :     if (num>= minblocks)
2717 :     do { /* until convergence */
2718 :     double DtimesF[4];
2719 :     double a,b,c,n,invdenom;
2720 :     double meanx,meany;
2721 :    
2722 :     a = b = c = n = 0;
2723 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2724 :     for (my = 1; my < (uint32_t)MBh-1; my++)
2725 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2726 :     {
2727 :     const int mbnum = mx + my * MBw;
2728 :     const VECTOR mv = pMBs[mbnum].mvs[0];
2729 :    
2730 :     if (!pMBs[mbnum].mcsel)
2731 :     continue;
2732 :    
2733 :     n++;
2734 :     a += 16*mx+8;
2735 :     b += 16*my+8;
2736 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2737 :    
2738 :     DtimesF[0] += (double)mv.x;
2739 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2740 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2741 :     DtimesF[3] += (double)mv.y;
2742 :     }
2743 :    
2744 :     invdenom = a*a+b*b-c*n;
2745 :    
2746 :     /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2747 :     /* D'*E*F has been calculated in the same loop as matrix */
2748 :    
2749 :     sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2750 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2751 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2752 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2753 :    
2754 :     sol[0] /= invdenom;
2755 :     sol[1] /= invdenom;
2756 :     sol[2] /= invdenom;
2757 :     sol[3] /= invdenom;
2758 :    
2759 :     meanx = meany = 0.;
2760 :     oldnum = 0;
2761 :     for (my = 1; my < (uint32_t)MBh-1; my++)
2762 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2763 :     {
2764 :     const int mbnum = mx + my * MBw;
2765 :     const VECTOR mv = pMBs[mbnum].mvs[0];
2766 :    
2767 :     if (!pMBs[mbnum].mcsel)
2768 :     continue;
2769 :    
2770 :     oldnum++;
2771 :     meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x );
2772 :     meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y );
2773 :     }
2774 :    
2775 :     if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */
2776 :     meanx /= oldnum;
2777 :     else
2778 :     meanx = 0.25;
2779 :    
2780 :     if (4*meany > oldnum)
2781 :     meany /= oldnum;
2782 :     else
2783 :     meany = 0.25;
2784 :    
2785 :     num = 0;
2786 :     for (my = 0; my < (uint32_t)MBh; my++)
2787 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2788 :     {
2789 :     const int mbnum = mx + my * MBw;
2790 :     const VECTOR mv = pMBs[mbnum].mvs[0];
2791 :    
2792 :     if (!pMBs[mbnum].mcsel)
2793 :     continue;
2794 :    
2795 :     if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx )
2796 :     || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) )
2797 :     pMBs[mbnum].mcsel=0;
2798 :     else
2799 :     num++;
2800 :     }
2801 :    
2802 :     } while ( (oldnum != num) && (num>= minblocks) );
2803 :    
2804 :     if (num < minblocks)
2805 :     {
2806 :     const int iEdgedWidth = pParam->edged_width;
2807 :     num = 0;
2808 :    
2809 :     /* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw);
2810 :     */
2811 :     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;
2812 :    
2813 :     if (!(current->motion_flags & XVID_GME_REFINE))
2814 :     return gmc;
2815 :    
2816 :     for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */
2817 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */
2818 :     {
2819 :     const int mbnum = mx + my * MBw;
2820 :     MACROBLOCK *const pMB = &pMBs[mbnum];
2821 :     const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx);
2822 :     if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx )
2823 :     && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) )
2824 :     { pMB->mcsel = 1;
2825 :     gmc.duv[0].x += pMB->mvs[0].x;
2826 :     gmc.duv[0].y += pMB->mvs[0].y;
2827 :     num++;
2828 :     }
2829 :     }
2830 :    
2831 :     if (gmc.duv[0].x)
2832 :     gmc.duv[0].x /= num;
2833 :     if (gmc.duv[0].y)
2834 :     gmc.duv[0].y /= num;
2835 :     } else {
2836 :    
2837 :     gmc.duv[0].x=(int)(sol[0]+0.5);
2838 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2839 :    
2840 :     gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2841 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2842 :    
2843 :     gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */
2844 :     gmc.duv[2].y=gmc.duv[1].x;
2845 :     }
2846 :     if (num>maxblocks)
2847 :     { for (my = 1; my < (uint32_t)MBh-1; my++)
2848 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2849 :     {
2850 :     const int mbnum = mx + my * MBw;
2851 :     if (pMBs[mbnum-1].mcsel)
2852 :     pMBs[mbnum].mcsel=0;
2853 :     else
2854 :     if (pMBs[mbnum-MBw].mcsel)
2855 :     pMBs[mbnum].mcsel=0;
2856 :     }
2857 :     }
2858 :     return gmc;
2859 :     }
2860 :    
2861 :     int
2862 :     GlobalMotionEstRefine(
2863 :     WARPPOINTS *const startwp,
2864 :     MACROBLOCK * const pMBs,
2865 :     const MBParam * const pParam,
2866 :     const FRAMEINFO * const current,
2867 :     const FRAMEINFO * const reference,
2868 :     const IMAGE * const pCurr,
2869 :     const IMAGE * const pRef,
2870 :     const IMAGE * const pRefH,
2871 :     const IMAGE * const pRefV,
2872 :     const IMAGE * const pRefHV)
2873 :     {
2874 :     uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width);
2875 :     WARPPOINTS bestwp=*startwp;
2876 :     WARPPOINTS centerwp,currwp;
2877 :     int gmcminSAD=0;
2878 :     int gmcSAD=0;
2879 :     int direction;
2880 :     // int mx,my;
2881 :    
2882 :     /* use many blocks... */
2883 :     /* for (my = 0; my < (uint32_t)pParam->mb_height; my++)
2884 :     for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++)
2885 :     {
2886 :     const int mbnum = mx + my * pParam->mb_width;
2887 :     pMBs[mbnum].mcsel=1;
2888 :     }
2889 :     */
2890 :    
2891 :     /* or rather don't use too many blocks... */
2892 :     /*
2893 :     for (my = 1; my < (uint32_t)MBh-1; my++)
2894 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2895 :     {
2896 :     const int mbnum = mx + my * MBw;
2897 :     if (MBmask[mbnum-1])
2898 :     MBmask[mbnum-1]=0;
2899 :     else
2900 :     if (MBmask[mbnum-MBw])
2901 :     MBmask[mbnum-1]=0;
2902 :    
2903 :     }
2904 :     */
2905 :     gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2906 :    
2907 :     if ( (reference->coding_type == S_VOP)
2908 :     && ( (reference->warp.duv[1].x != bestwp.duv[1].x)
2909 :     || (reference->warp.duv[1].y != bestwp.duv[1].y)
2910 :     || (reference->warp.duv[0].x != bestwp.duv[0].x)
2911 :     || (reference->warp.duv[0].y != bestwp.duv[0].y)
2912 :     || (reference->warp.duv[2].x != bestwp.duv[2].x)
2913 :     || (reference->warp.duv[2].y != bestwp.duv[2].y) ) )
2914 :     {
2915 :     gmcSAD = globalSAD(&reference->warp, pParam, pMBs,
2916 :     current, pRef, pCurr, GMCblock);
2917 :    
2918 :     if (gmcSAD < gmcminSAD)
2919 :     { bestwp = reference->warp;
2920 :     gmcminSAD = gmcSAD;
2921 :     }
2922 :     }
2923 :    
2924 :     do {
2925 :     direction = 0;
2926 :     centerwp = bestwp;
2927 :    
2928 :     currwp = centerwp;
2929 :    
2930 :     currwp.duv[0].x--;
2931 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2932 :     if (gmcSAD < gmcminSAD)
2933 :     { bestwp = currwp;
2934 :     gmcminSAD = gmcSAD;
2935 :     direction = 1;
2936 :     }
2937 :     else
2938 :     {
2939 :     currwp = centerwp; currwp.duv[0].x++;
2940 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2941 :     if (gmcSAD < gmcminSAD)
2942 :     { bestwp = currwp;
2943 :     gmcminSAD = gmcSAD;
2944 :     direction = 2;
2945 :     }
2946 :     }
2947 :     if (direction) continue;
2948 :    
2949 :     currwp = centerwp; currwp.duv[0].y--;
2950 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2951 :     if (gmcSAD < gmcminSAD)
2952 :     { bestwp = currwp;
2953 :     gmcminSAD = gmcSAD;
2954 :     direction = 4;
2955 :     }
2956 :     else
2957 :     {
2958 :     currwp = centerwp; currwp.duv[0].y++;
2959 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2960 :     if (gmcSAD < gmcminSAD)
2961 :     { bestwp = currwp;
2962 :     gmcminSAD = gmcSAD;
2963 :     direction = 8;
2964 :     }
2965 :     }
2966 :     if (direction) continue;
2967 :    
2968 :     currwp = centerwp; currwp.duv[1].x++;
2969 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2970 :     if (gmcSAD < gmcminSAD)
2971 :     { bestwp = currwp;
2972 :     gmcminSAD = gmcSAD;
2973 :     direction = 32;
2974 :     }
2975 :     currwp.duv[2].y++;
2976 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2977 :     if (gmcSAD < gmcminSAD)
2978 :     { bestwp = currwp;
2979 :     gmcminSAD = gmcSAD;
2980 :     direction = 1024;
2981 :     }
2982 :    
2983 :     currwp = centerwp; currwp.duv[1].x--;
2984 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2985 :     if (gmcSAD < gmcminSAD)
2986 :     { bestwp = currwp;
2987 :     gmcminSAD = gmcSAD;
2988 :     direction = 16;
2989 :     }
2990 :     else
2991 :     {
2992 :     currwp = centerwp; currwp.duv[1].x++;
2993 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2994 :     if (gmcSAD < gmcminSAD)
2995 :     { bestwp = currwp;
2996 :     gmcminSAD = gmcSAD;
2997 :     direction = 32;
2998 :     }
2999 :     }
3000 :     if (direction) continue;
3001 :    
3002 :    
3003 :     currwp = centerwp; currwp.duv[1].y--;
3004 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3005 :     if (gmcSAD < gmcminSAD)
3006 :     { bestwp = currwp;
3007 :     gmcminSAD = gmcSAD;
3008 :     direction = 64;
3009 :     }
3010 :     else
3011 :     {
3012 :     currwp = centerwp; currwp.duv[1].y++;
3013 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3014 :     if (gmcSAD < gmcminSAD)
3015 :     { bestwp = currwp;
3016 :     gmcminSAD = gmcSAD;
3017 :     direction = 128;
3018 :     }
3019 :     }
3020 :     if (direction) continue;
3021 :    
3022 :     currwp = centerwp; currwp.duv[2].x--;
3023 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3024 :     if (gmcSAD < gmcminSAD)
3025 :     { bestwp = currwp;
3026 :     gmcminSAD = gmcSAD;
3027 :     direction = 256;
3028 :     }
3029 :     else
3030 :     {
3031 :     currwp = centerwp; currwp.duv[2].x++;
3032 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3033 :     if (gmcSAD < gmcminSAD)
3034 :     { bestwp = currwp;
3035 :     gmcminSAD = gmcSAD;
3036 :     direction = 512;
3037 :     }
3038 :     }
3039 :     if (direction) continue;
3040 :    
3041 :     currwp = centerwp; currwp.duv[2].y--;
3042 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3043 :     if (gmcSAD < gmcminSAD)
3044 :     { bestwp = currwp;
3045 :     gmcminSAD = gmcSAD;
3046 :     direction = 1024;
3047 :     }
3048 :     else
3049 :     {
3050 :     currwp = centerwp; currwp.duv[2].y++;
3051 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3052 :     if (gmcSAD < gmcminSAD)
3053 :     { bestwp = currwp;
3054 :     gmcminSAD = gmcSAD;
3055 :     direction = 2048;
3056 :     }
3057 :     }
3058 :     } while (direction);
3059 :     free(GMCblock);
3060 :    
3061 :     *startwp = bestwp;
3062 :    
3063 :     return gmcminSAD;
3064 :     }
3065 :    
3066 :     int
3067 :     globalSAD(const WARPPOINTS *const wp,
3068 :     const MBParam * const pParam,
3069 :     const MACROBLOCK * const pMBs,
3070 :     const FRAMEINFO * const current,
3071 :     const IMAGE * const pRef,
3072 :     const IMAGE * const pCurr,
3073 :     uint8_t *const GMCblock)
3074 :     {
3075 :     NEW_GMC_DATA gmc_data;
3076 :     int iSAD, gmcSAD=0;
3077 :     int num=0;
3078 :     unsigned int mx, my;
3079 :    
3080 :     generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data);
3081 :    
3082 :     for (my = 0; my < (uint32_t)pParam->mb_height; my++)
3083 :     for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) {
3084 :    
3085 :     const int mbnum = mx + my * pParam->mb_width;
3086 :     const int iEdgedWidth = pParam->edged_width;
3087 :    
3088 :     if (!pMBs[mbnum].mcsel)
3089 :     continue;
3090 :    
3091 :     gmc_data.predict_16x16(&gmc_data, GMCblock,
3092 :     pRef->y,
3093 :     iEdgedWidth,
3094 :     iEdgedWidth,
3095 :     mx, my,
3096 :     pParam->m_rounding_type);
3097 :    
3098 :     iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx),
3099 :     GMCblock , iEdgedWidth, 65536);
3100 :     iSAD -= pMBs[mbnum].sad16;
3101 :    
3102 :     if (iSAD<0)
3103 :     gmcSAD += iSAD;
3104 :     num++;
3105 :     }
3106 :     return gmcSAD;
3107 :     }
3108 :    

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