[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 1115 - (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 1115 * $Id: motion_est.c,v 1.58.2.27 2003-08-07 15:42:50 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 : syskin 1084 #include "../global.h"
38 : Isibaar 3 #include "../utils/timer.h"
39 : edgomez 851 #include "../image/interpolate8x8.h"
40 :     #include "motion_est.h"
41 : suxen_drol 118 #include "motion.h"
42 : Isibaar 3 #include "sad.h"
43 : chl 1077 #include "gmc.h"
44 : edgomez 851 #include "../utils/emms.h"
45 :     #include "../dct/fdct.h"
46 : Isibaar 3
47 : edgomez 872 /*****************************************************************************
48 :     * Modified rounding tables -- declared in motion.h
49 :     * Original tables see ISO spec tables 7-6 -> 7-9
50 :     ****************************************************************************/
51 :    
52 :     const uint32_t roundtab[16] =
53 :     {0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 };
54 :    
55 :     /* K = 4 */
56 :     const uint32_t roundtab_76[16] =
57 :     { 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 };
58 :    
59 :     /* K = 2 */
60 :     const uint32_t roundtab_78[8] =
61 :     { 0, 0, 1, 1, 0, 0, 0, 1 };
62 :    
63 :     /* K = 1 */
64 :     const uint32_t roundtab_79[4] =
65 :     { 0, 1, 0, 0 };
66 :    
67 : edgomez 851 #define INITIAL_SKIP_THRESH (10)
68 :     #define FINAL_SKIP_THRESH (50)
69 :     #define MAX_SAD00_FOR_SKIP (20)
70 :     #define MAX_CHROMA_SAD_FOR_SKIP (22)
71 : Isibaar 3
72 : edgomez 851 #define CHECK_CANDIDATE(X,Y,D) { \
73 :     CheckCandidate((X),(Y), (D), &iDirection, data ); }
74 : Isibaar 3
75 : syskin 1084
76 : edgomez 872 /*****************************************************************************
77 :     * Code
78 :     ****************************************************************************/
79 :    
80 : edgomez 851 static __inline uint32_t
81 :     d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
82 :     {
83 : edgomez 974 int bits;
84 :     const int q = (1 << (iFcode - 1)) - 1;
85 :    
86 :     x <<= qpel;
87 :     y <<= qpel;
88 : edgomez 851 if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
89 : chl 141
90 : edgomez 851 x -= pred.x;
91 : edgomez 974 bits = (x != 0 ? iFcode:0);
92 :     x = abs(x);
93 :     x += q;
94 :     x >>= (iFcode - 1);
95 :     bits += mvtab[x];
96 :    
97 : edgomez 851 y -= pred.y;
98 : edgomez 974 bits += (y != 0 ? iFcode:0);
99 :     y = abs(y);
100 :     y += q;
101 :     y >>= (iFcode - 1);
102 :     bits += mvtab[y];
103 : chl 141
104 : edgomez 974 return bits;
105 : edgomez 851 }
106 : chl 141
107 : edgomez 978 static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by,
108 :     const SearchData * const data)
109 : edgomez 851 {
110 :     int sad;
111 :     const uint32_t stride = data->iEdgedWidth/2;
112 :     uint8_t * f_refu = data->RefQ,
113 :     * f_refv = data->RefQ + 8,
114 :     * b_refu = data->RefQ + 16,
115 :     * b_refv = data->RefQ + 24;
116 : edgomez 978 int offset = (fx>>1) + (fy>>1)*stride;
117 : Isibaar 3
118 : edgomez 851 switch (((fx & 1) << 1) | (fy & 1)) {
119 :     case 0:
120 : edgomez 978 f_refu = (uint8_t*)data->RefP[4] + offset;
121 :     f_refv = (uint8_t*)data->RefP[5] + offset;
122 : edgomez 851 break;
123 :     case 1:
124 : edgomez 978 interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding);
125 :     interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding);
126 : edgomez 851 break;
127 :     case 2:
128 : edgomez 978 interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding);
129 :     interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding);
130 : edgomez 851 break;
131 :     default:
132 : edgomez 978 interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding);
133 :     interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding);
134 : edgomez 851 break;
135 :     }
136 : Isibaar 3
137 : edgomez 978 offset = (bx>>1) + (by>>1)*stride;
138 : edgomez 851 switch (((bx & 1) << 1) | (by & 1)) {
139 :     case 0:
140 : edgomez 978 b_refu = (uint8_t*)data->b_RefP[4] + offset;
141 :     b_refv = (uint8_t*)data->b_RefP[5] + offset;
142 : edgomez 851 break;
143 :     case 1:
144 : edgomez 978 interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
145 :     interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
146 : edgomez 851 break;
147 :     case 2:
148 : edgomez 978 interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
149 :     interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
150 : edgomez 851 break;
151 :     default:
152 : edgomez 978 interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
153 :     interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
154 : edgomez 851 break;
155 :     }
156 : Isibaar 3
157 : edgomez 851 sad = sad8bi(data->CurU, b_refu, f_refu, stride);
158 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
159 : Isibaar 3
160 : edgomez 851 return sad;
161 :     }
162 : Isibaar 3
163 : edgomez 851 static int32_t
164 : edgomez 978 ChromaSAD(const int dx, const int dy, const SearchData * const data)
165 : edgomez 851 {
166 :     int sad;
167 :     const uint32_t stride = data->iEdgedWidth/2;
168 : edgomez 978 int offset = (dx>>1) + (dy>>1)*stride;
169 : Isibaar 3
170 : edgomez 1053 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; /* it has been checked recently */
171 :     data->temp[5] = dx; data->temp[6] = dy; /* backup */
172 : Isibaar 3
173 : edgomez 851 switch (((dx & 1) << 1) | (dy & 1)) {
174 :     case 0:
175 : edgomez 978 sad = sad8(data->CurU, data->RefP[4] + offset, stride);
176 :     sad += sad8(data->CurV, data->RefP[5] + offset, stride);
177 : edgomez 851 break;
178 :     case 1:
179 : edgomez 978 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride);
180 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride);
181 : edgomez 851 break;
182 :     case 2:
183 : edgomez 978 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride);
184 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride);
185 : edgomez 851 break;
186 :     default:
187 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding);
188 : edgomez 851 sad = sad8(data->CurU, data->RefQ, stride);
189 :    
190 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding);
191 : edgomez 851 sad += sad8(data->CurV, data->RefQ, stride);
192 :     break;
193 :     }
194 : edgomez 1053 data->temp[7] = sad; /* backup, part 2 */
195 : edgomez 851 return sad;
196 : Isibaar 3 }
197 :    
198 : edgomez 851 static __inline const uint8_t *
199 :     GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
200 : Isibaar 3 {
201 : edgomez 1053 /* dir : 0 = forward, 1 = backward */
202 : edgomez 987 const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP );
203 : edgomez 978 const int picture = ((x&1)<<1) | (y&1);
204 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
205 :     return direction[picture] + offset;
206 : Isibaar 3 }
207 :    
208 : edgomez 1053 /* this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate */
209 : edgomez 851 static __inline const uint8_t *
210 :     GetReference(const int x, const int y, const SearchData * const data)
211 : Isibaar 3 {
212 : edgomez 978 const int picture = ((x&1)<<1) | (y&1);
213 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
214 :     return data->RefP[picture] + offset;
215 : Isibaar 3 }
216 : edgomez 851
217 :     static uint8_t *
218 :     Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
219 : Isibaar 3 {
220 : edgomez 1053 /* create or find a qpel-precision reference picture; return pointer to it */
221 : edgomez 851 uint8_t * Reference = data->RefQ + 16*dir;
222 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
223 :     const uint32_t rounding = data->rounding;
224 :     const int halfpel_x = x/2;
225 :     const int halfpel_y = y/2;
226 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
227 : suxen_drol 136
228 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
229 :     ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
230 :     switch( ((x&1)<<1) + (y&1) ) {
231 : edgomez 1053 case 3: /* x and y in qpel resolution - the "corners" (top left/right and */
232 :     /* bottom left/right) during qpel refinement */
233 : edgomez 963 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
234 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
235 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
236 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
237 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
238 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
239 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
240 : edgomez 851 break;
241 : chl 184
242 : edgomez 1053 case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
243 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
244 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
245 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
246 :     break;
247 : suxen_drol 136
248 : edgomez 1053 case 2: /* x qpel, y halfpel - left or right during qpel refinement */
249 : edgomez 851 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
250 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
251 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
252 :     break;
253 : edgomez 195
254 : edgomez 1053 default: /* pure halfpel position */
255 : edgomez 963 return (uint8_t *) ref1;
256 : edgomez 978
257 : edgomez 851 }
258 :     return Reference;
259 :     }
260 : Isibaar 3
261 : edgomez 851 static uint8_t *
262 :     Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
263 :     {
264 : edgomez 1053 /* create or find a qpel-precision reference picture; return pointer to it */
265 : edgomez 851 uint8_t * Reference = data->RefQ + 16*dir;
266 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
267 :     const uint32_t rounding = data->rounding;
268 :     const int halfpel_x = x/2;
269 :     const int halfpel_y = y/2;
270 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
271 : chl 326
272 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
273 :     switch( ((x&1)<<1) + (y&1) ) {
274 : edgomez 1053 case 3:
275 :     /*
276 :     * x and y in qpel resolution - the "corners" (top left/right and
277 :     * bottom left/right) during qpel refinement
278 :     */
279 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
280 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
281 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
282 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
283 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
284 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
285 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
286 :     break;
287 : chl 184
288 : edgomez 1053 case 1: /* x halfpel, y qpel - top or bottom during qpel refinement */
289 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
290 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
291 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
292 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
293 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
294 :     break;
295 : chl 184
296 : edgomez 1053 case 2: /* x qpel, y halfpel - left or right during qpel refinement */
297 : edgomez 851 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
298 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
299 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
300 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
301 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
302 :     break;
303 : edgomez 195
304 : edgomez 1053 default: /* pure halfpel position */
305 : edgomez 851 return (uint8_t *) ref1;
306 :     }
307 :     return Reference;
308 :     }
309 : edgomez 195
310 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS START */
311 : edgomez 195
312 : edgomez 851 static void
313 :     CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
314 :     {
315 :     int xc, yc;
316 :     const uint8_t * Reference;
317 :     VECTOR * current;
318 :     int32_t sad; uint32_t t;
319 : edgomez 195
320 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
321 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
322 : edgomez 195
323 : edgomez 851 if (!data->qpel_precision) {
324 :     Reference = GetReference(x, y, data);
325 :     current = data->currentMV;
326 :     xc = x; yc = y;
327 : edgomez 1053 } else { /* x and y are in 1/4 precision */
328 : edgomez 851 Reference = Interpolate16x16qpel(x, y, 0, data);
329 : edgomez 1053 xc = x/2; yc = y/2; /* for chroma sad */
330 : edgomez 851 current = data->currentQMV;
331 :     }
332 : edgomez 195
333 : edgomez 851 sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
334 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
335 : edgomez 195
336 : edgomez 851 sad += (data->lambda16 * t * sad)>>10;
337 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
338 : chl 175
339 : syskin 1084 if (data->chroma && sad < data->iMinSAD[0])
340 :     sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
341 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
342 : edgomez 195
343 : edgomez 851 if (sad < data->iMinSAD[0]) {
344 :     data->iMinSAD[0] = sad;
345 :     current[0].x = x; current[0].y = y;
346 :     *dir = Direction;
347 :     }
348 : edgomez 195
349 : edgomez 851 if (data->temp[1] < data->iMinSAD[1]) {
350 :     data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
351 :     if (data->temp[2] < data->iMinSAD[2]) {
352 :     data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
353 :     if (data->temp[3] < data->iMinSAD[3]) {
354 :     data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
355 :     if (data->temp[4] < data->iMinSAD[4]) {
356 :     data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
357 : Isibaar 3 }
358 :    
359 : edgomez 851 static void
360 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
361 :     {
362 :     int32_t sad; uint32_t t;
363 :     const uint8_t * Reference;
364 : edgomez 974 VECTOR * current;
365 : chl 326
366 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
367 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
368 : Isibaar 3
369 : edgomez 974 if (!data->qpel_precision) {
370 :     Reference = GetReference(x, y, data);
371 :     current = data->currentMV;
372 : edgomez 1053 } else { /* x and y are in 1/4 precision */
373 : edgomez 974 Reference = Interpolate8x8qpel(x, y, 0, 0, data);
374 :     current = data->currentQMV;
375 :     }
376 : Isibaar 3
377 : edgomez 851 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
378 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
379 : Isibaar 3
380 : edgomez 851 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
381 : Isibaar 3
382 : edgomez 851 if (sad < *(data->iMinSAD)) {
383 :     *(data->iMinSAD) = sad;
384 : edgomez 974 current->x = x; current->y = y;
385 : edgomez 851 *dir = Direction;
386 :     }
387 : Isibaar 3 }
388 :    
389 : edgomez 851 static void
390 :     CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
391 :     {
392 :     uint32_t t;
393 :     const uint8_t * Reference;
394 :    
395 : edgomez 1053 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */
396 : edgomez 851 (x > data->max_dx) || (x < data->min_dx)
397 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
398 :    
399 :     Reference = GetReference(x, y, data);
400 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
401 :    
402 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
403 :    
404 :     data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
405 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
406 :    
407 :     if (data->temp[0] < data->iMinSAD[0]) {
408 :     data->iMinSAD[0] = data->temp[0];
409 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
410 :     *dir = Direction; }
411 :    
412 :     if (data->temp[1] < data->iMinSAD[1]) {
413 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
414 :     if (data->temp[2] < data->iMinSAD[2]) {
415 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
416 :     if (data->temp[3] < data->iMinSAD[3]) {
417 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
418 :     if (data->temp[4] < data->iMinSAD[4]) {
419 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
420 : Isibaar 3 }
421 : edgomez 195
422 : edgomez 851 static void
423 :     CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
424 :     {
425 :     int32_t sad, xc, yc;
426 :     const uint8_t * Reference;
427 :     uint32_t t;
428 :     VECTOR * current;
429 :    
430 : edgomez 963 if ( (x > data->max_dx) || ( x < data->min_dx)
431 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
432 : edgomez 851
433 : edgomez 1053 if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; /* non-zero even value */
434 : edgomez 851
435 : edgomez 1053 if (data->qpel_precision) { /* x and y are in 1/4 precision */
436 : edgomez 851 Reference = Interpolate16x16qpel(x, y, 0, data);
437 :     current = data->currentQMV;
438 :     xc = x/2; yc = y/2;
439 :     } else {
440 :     Reference = GetReference(x, y, data);
441 :     current = data->currentMV;
442 :     xc = x; yc = y;
443 :     }
444 :     t = d_mv_bits(x, y, data->predMV, data->iFcode,
445 :     data->qpel^data->qpel_precision, data->rrv);
446 :    
447 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
448 :     sad += (data->lambda16 * t * sad)>>10;
449 :    
450 : syskin 1084 if (data->chroma && sad < *data->iMinSAD)
451 :     sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
452 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
453 : edgomez 851
454 :     if (sad < *(data->iMinSAD)) {
455 :     *(data->iMinSAD) = sad;
456 :     current->x = x; current->y = y;
457 :     *dir = Direction;
458 :     }
459 : chl 96 }
460 : Isibaar 3
461 : edgomez 851 static void
462 : chl 1077 CheckCandidate16I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
463 :     {
464 :     int sad;
465 :     // int xc, yc;
466 :     const uint8_t * Reference;
467 :     // VECTOR * current;
468 :    
469 :     if ( (x > data->max_dx) || ( x < data->min_dx)
470 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
471 :    
472 :     Reference = GetReference(x, y, data);
473 :     // xc = x; yc = y;
474 :    
475 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
476 :     // sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0);
477 :    
478 :     /* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
479 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
480 :     */
481 :    
482 :     if (sad < data->iMinSAD[0]) {
483 :     data->iMinSAD[0] = sad;
484 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
485 :     *dir = Direction;
486 :     }
487 :     }
488 :    
489 :     static void
490 : edgomez 851 CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
491 :     {
492 : edgomez 1053 /* maximum speed - for P/B/I decision */
493 : edgomez 851 int32_t sad;
494 :    
495 :     if ( (x > data->max_dx) || (x < data->min_dx)
496 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
497 :    
498 : syskin 1084 sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*((int)data->iEdgedWidth),
499 : edgomez 987 data->iEdgedWidth, data->temp+1);
500 : edgomez 851
501 :     if (sad < *(data->iMinSAD)) {
502 :     *(data->iMinSAD) = sad;
503 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
504 :     *dir = Direction;
505 :     }
506 :     if (data->temp[1] < data->iMinSAD[1]) {
507 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
508 :     if (data->temp[2] < data->iMinSAD[2]) {
509 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
510 :     if (data->temp[3] < data->iMinSAD[3]) {
511 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
512 :     if (data->temp[4] < data->iMinSAD[4]) {
513 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
514 :    
515 : Isibaar 3 }
516 :    
517 : edgomez 851 static void
518 :     CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
519 :     {
520 :     int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
521 :     uint32_t t;
522 :     const uint8_t *ReferenceF, *ReferenceB;
523 :     VECTOR *current;
524 :    
525 : edgomez 953 if ((xf > data->max_dx) || (xf < data->min_dx) ||
526 :     (yf > data->max_dy) || (yf < data->min_dy))
527 :     return;
528 : edgomez 851
529 :     if (!data->qpel_precision) {
530 :     ReferenceF = GetReference(xf, yf, data);
531 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
532 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
533 :     current = data->currentMV;
534 :     xcf = xf; ycf = yf;
535 :     xcb = xb; ycb = yb;
536 :     } else {
537 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
538 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
539 :     current = data->currentQMV;
540 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
541 :     xcf = xf/2; ycf = yf/2;
542 :     xcb = xb/2; ycb = yb/2;
543 :     }
544 :    
545 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
546 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
547 :    
548 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
549 :     sad += (data->lambda16 * t * sad)>>10;
550 :    
551 : syskin 1084 if (data->chroma && sad < *data->iMinSAD)
552 :     sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
553 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
554 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
555 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
556 : edgomez 851
557 :     if (sad < *(data->iMinSAD)) {
558 :     *(data->iMinSAD) = sad;
559 :     current->x = xf; current->y = yf;
560 :     *dir = Direction;
561 :     }
562 : Isibaar 3 }
563 :    
564 : edgomez 851 static void
565 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
566 :     {
567 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
568 :     uint32_t k;
569 :     const uint8_t *ReferenceF;
570 :     const uint8_t *ReferenceB;
571 :     VECTOR mvs, b_mvs;
572 :    
573 : edgomez 953 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
574 : edgomez 851
575 :     for (k = 0; k < 4; k++) {
576 :     mvs.x = data->directmvF[k].x + x;
577 :     b_mvs.x = ((x == 0) ?
578 :     data->directmvB[k].x
579 :     : mvs.x - data->referencemv[k].x);
580 :    
581 :     mvs.y = data->directmvF[k].y + y;
582 :     b_mvs.y = ((y == 0) ?
583 :     data->directmvB[k].y
584 :     : mvs.y - data->referencemv[k].y);
585 :    
586 : edgomez 953 if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) ||
587 :     (mvs.y > data->max_dy) || (mvs.y < data->min_dy) ||
588 :     (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) ||
589 :     (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) )
590 :     return;
591 : edgomez 851
592 :     if (data->qpel) {
593 :     xcf += mvs.x/2; ycf += mvs.y/2;
594 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
595 :     } else {
596 :     xcf += mvs.x; ycf += mvs.y;
597 :     xcb += b_mvs.x; ycb += b_mvs.y;
598 : edgomez 1053 mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */
599 : edgomez 851 b_mvs.x *= 2; b_mvs.y *= 2;
600 :     }
601 :    
602 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
603 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
604 :    
605 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
606 :     ReferenceF, ReferenceB, data->iEdgedWidth);
607 :     if (sad > *(data->iMinSAD)) return;
608 :     }
609 :    
610 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
611 :    
612 : syskin 1084 if (data->chroma && sad < *data->iMinSAD)
613 :     sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
614 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
615 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
616 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
617 : edgomez 851
618 :     if (sad < *(data->iMinSAD)) {
619 :     *(data->iMinSAD) = sad;
620 :     data->currentMV->x = x; data->currentMV->y = y;
621 :     *dir = Direction;
622 :     }
623 : Isibaar 3 }
624 :    
625 : edgomez 851 static void
626 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
627 : Isibaar 3 {
628 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
629 :     const uint8_t *ReferenceF;
630 :     const uint8_t *ReferenceB;
631 :     VECTOR mvs, b_mvs;
632 : Isibaar 3
633 : edgomez 963 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
634 : Isibaar 3
635 : edgomez 851 mvs.x = data->directmvF[0].x + x;
636 :     b_mvs.x = ((x == 0) ?
637 :     data->directmvB[0].x
638 :     : mvs.x - data->referencemv[0].x);
639 : Isibaar 3
640 : edgomez 851 mvs.y = data->directmvF[0].y + y;
641 :     b_mvs.y = ((y == 0) ?
642 :     data->directmvB[0].y
643 :     : mvs.y - data->referencemv[0].y);
644 : Isibaar 3
645 : edgomez 963 if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx)
646 :     || (mvs.y > data->max_dy) || (mvs.y < data->min_dy)
647 :     || (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx)
648 :     || (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return;
649 : Isibaar 3
650 : edgomez 851 if (data->qpel) {
651 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
652 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
653 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
654 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
655 :     } else {
656 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
657 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
658 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
659 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
660 :     }
661 :    
662 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
663 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
664 :    
665 : syskin 1084 if (data->chroma && sad < *data->iMinSAD)
666 :     sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
667 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
668 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
669 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
670 : edgomez 851
671 :     if (sad < *(data->iMinSAD)) {
672 :     *(data->iMinSAD) = sad;
673 :     data->currentMV->x = x; data->currentMV->y = y;
674 :     *dir = Direction;
675 :     }
676 : Isibaar 3 }
677 :    
678 : edgomez 851
679 :     static void
680 : syskin 1108 CheckCandidateRD16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
681 : Isibaar 3 {
682 :    
683 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
684 : syskin 1108 int32_t rd = 0;
685 : edgomez 851 VECTOR * current;
686 :     const uint8_t * ptr;
687 :     int i, cbp = 0, t, xc, yc;
688 : edgomez 195
689 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
690 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
691 : edgomez 195
692 : edgomez 851 if (!data->qpel_precision) {
693 :     ptr = GetReference(x, y, data);
694 :     current = data->currentMV;
695 :     xc = x; yc = y;
696 : edgomez 1053 } else { /* x and y are in 1/4 precision */
697 : edgomez 851 ptr = Interpolate16x16qpel(x, y, 0, data);
698 :     current = data->currentQMV;
699 :     xc = x/2; yc = y/2;
700 :     }
701 : Isibaar 3
702 : edgomez 851 for(i = 0; i < 4; i++) {
703 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
704 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
705 : syskin 1108 rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i);
706 : edgomez 851 }
707 : Isibaar 3
708 : syskin 1108 rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
709 : edgomez 195
710 : syskin 1071 if (data->temp[0] + t < data->iMinSAD[1]) {
711 : chl 1115 data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) | (cbp&32); }
712 : syskin 1071 if (data->temp[1] < data->iMinSAD[2]) {
713 : chl 1115 data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; data->cbp[1] = (data->cbp[1]&~16) | (cbp&16); }
714 : syskin 1071 if (data->temp[2] < data->iMinSAD[3]) {
715 : chl 1115 data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; data->cbp[1] = (data->cbp[1]&~8) | (cbp&8); }
716 : syskin 1071 if (data->temp[3] < data->iMinSAD[4]) {
717 : chl 1115 data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | (cbp&4); }
718 : syskin 1071
719 : syskin 1108 rd += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
720 : edgomez 851
721 : syskin 1108 if (rd >= data->iMinSAD[0]) return;
722 : edgomez 851
723 : edgomez 1053 /* chroma */
724 : edgomez 1022 xc = (xc >> 1) + roundtab_79[xc & 0x3];
725 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
726 : edgomez 851
727 : edgomez 1053 /* chroma U */
728 : syskin 1084 ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
729 :     transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2);
730 : syskin 1108 rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4);
731 :     if (rd >= data->iMinSAD[0]) return;
732 : edgomez 1022
733 : edgomez 1053 /* chroma V */
734 : syskin 1084 ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
735 :     transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2);
736 : syskin 1108 rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
737 : edgomez 1022
738 : syskin 1108 rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
739 : edgomez 851
740 : syskin 1108 if (rd < data->iMinSAD[0]) {
741 :     data->iMinSAD[0] = rd;
742 : edgomez 851 current[0].x = x; current[0].y = y;
743 :     *dir = Direction;
744 : syskin 1084 *data->cbp = cbp;
745 : edgomez 851 }
746 : syskin 1071 }
747 : edgomez 851
748 :     static void
749 : syskin 1108 CheckCandidateRD8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
750 : chl 96 {
751 :    
752 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
753 : syskin 1108 int32_t rd;
754 : edgomez 851 VECTOR * current;
755 :     const uint8_t * ptr;
756 : edgomez 1022 int cbp = 0;
757 : edgomez 195
758 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
759 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
760 : edgomez 195
761 : edgomez 851 if (!data->qpel_precision) {
762 :     ptr = GetReference(x, y, data);
763 :     current = data->currentMV;
764 : edgomez 1053 } else { /* x and y are in 1/4 precision */
765 : edgomez 851 ptr = Interpolate8x8qpel(x, y, 0, 0, data);
766 :     current = data->currentQMV;
767 :     }
768 : chl 96
769 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
770 : syskin 1108 rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5);
771 :     rd += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
772 : chl 96
773 : syskin 1108 if (rd < data->iMinSAD[0]) {
774 : syskin 1084 *data->cbp = cbp;
775 : syskin 1108 data->iMinSAD[0] = rd;
776 : edgomez 851 current[0].x = x; current[0].y = y;
777 :     *dir = Direction;
778 :     }
779 :     }
780 : chl 96
781 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
782 : edgomez 195
783 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
784 : chl 96
785 : edgomez 851 static void
786 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
787 :     {
788 : edgomez 195
789 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
790 : edgomez 195
791 : edgomez 851 int iDirection;
792 : edgomez 195
793 : edgomez 1053 for(;;) { /* forever */
794 : edgomez 851 iDirection = 0;
795 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
796 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
797 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
798 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
799 : edgomez 195
800 : edgomez 851 /* now we're doing diagonal checks near our candidate */
801 : edgomez 195
802 : edgomez 1053 if (iDirection) { /* if anything found */
803 : edgomez 851 bDirection = iDirection;
804 :     iDirection = 0;
805 :     x = data->currentMV->x; y = data->currentMV->y;
806 : edgomez 1053 if (bDirection & 3) { /* our candidate is left or right */
807 : edgomez 851 CHECK_CANDIDATE(x, y + iDiamondSize, 8);
808 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
809 : edgomez 1053 } else { /* what remains here is up or down */
810 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y, 2);
811 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
812 :     }
813 : edgomez 195
814 : edgomez 851 if (iDirection) {
815 :     bDirection += iDirection;
816 :     x = data->currentMV->x; y = data->currentMV->y;
817 :     }
818 : edgomez 1053 } else { /* about to quit, eh? not so fast.... */
819 : edgomez 851 switch (bDirection) {
820 :     case 2:
821 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
822 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
823 : edgomez 195 break;
824 : edgomez 851 case 1:
825 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
826 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
827 : edgomez 195 break;
828 : edgomez 851 case 2 + 4:
829 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
830 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
831 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
832 :     break;
833 :     case 4:
834 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
835 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
836 :     break;
837 : edgomez 195 case 8:
838 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
839 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
840 : edgomez 195 break;
841 : edgomez 851 case 1 + 4:
842 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
843 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
844 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
845 : edgomez 195 break;
846 : edgomez 851 case 2 + 8:
847 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
848 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
849 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
850 :     break;
851 :     case 1 + 8:
852 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
853 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
854 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
855 :     break;
856 : edgomez 1053 default: /* 1+2+4+8 == we didn't find anything at all */
857 : edgomez 851 CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
858 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
859 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
860 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
861 :     break;
862 : chl 96 }
863 : edgomez 1053 if (!iDirection) break; /* ok, the end. really */
864 : edgomez 851 bDirection = iDirection;
865 :     x = data->currentMV->x; y = data->currentMV->y;
866 : chl 345 }
867 : edgomez 195 }
868 : chl 96 }
869 :    
870 : edgomez 851 static void
871 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
872 : chl 96 {
873 : edgomez 851 int iDirection;
874 : edgomez 195
875 : edgomez 851 do {
876 :     iDirection = 0;
877 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
878 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
879 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
880 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
881 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
882 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
883 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
884 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
885 : chl 96
886 : edgomez 851 bDirection = iDirection;
887 :     x = data->currentMV->x; y = data->currentMV->y;
888 :     } while (iDirection);
889 : chl 96 }
890 :    
891 : edgomez 851 static void
892 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
893 : chl 181 {
894 :    
895 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
896 :    
897 : edgomez 851 int iDirection;
898 : edgomez 195
899 : edgomez 851 do {
900 :     iDirection = 0;
901 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
902 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
903 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
904 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
905 :    
906 :     /* now we're doing diagonal checks near our candidate */
907 :    
908 : edgomez 1053 if (iDirection) { /* checking if anything found */
909 : edgomez 851 bDirection = iDirection;
910 : chl 181 iDirection = 0;
911 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
912 : edgomez 1053 if (bDirection & 3) { /* our candidate is left or right */
913 : edgomez 851 CHECK_CANDIDATE(x, y + iDiamondSize, 8);
914 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
915 : edgomez 1053 } else { /* what remains here is up or down */
916 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y, 2);
917 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
918 :     }
919 :     bDirection += iDirection;
920 :     x = data->currentMV->x; y = data->currentMV->y;
921 :     }
922 :     }
923 :     while (iDirection);
924 :     }
925 : chl 181
926 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
927 : chl 181
928 : edgomez 851 static void
929 :     SubpelRefine(const SearchData * const data)
930 :     {
931 :     /* Do a half-pel or q-pel refinement */
932 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
933 : edgomez 1053 int iDirection; /* only needed because macro expects it */
934 : chl 181
935 : edgomez 851 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
936 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
937 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
938 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
939 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
940 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
941 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
942 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
943 :     }
944 : chl 181
945 : edgomez 851 static __inline int
946 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
947 :     const int x, const int y,
948 :     const uint32_t stride, const uint32_t iQuant, int rrv)
949 : chl 181
950 : edgomez 851 {
951 : edgomez 963 int offset = (x + y*stride)*8;
952 : edgomez 851 if(!rrv) {
953 : edgomez 963 uint32_t sadC = sad8(current->u + offset,
954 :     reference->u + offset, stride);
955 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
956 : edgomez 963 sadC += sad8(current->v + offset,
957 :     reference->v + offset, stride);
958 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
959 :     return 1;
960 : chl 181
961 : edgomez 851 } else {
962 : edgomez 963 uint32_t sadC = sad16(current->u + 2*offset,
963 :     reference->u + 2*offset, stride, 256*4096);
964 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
965 : edgomez 963 sadC += sad16(current->v + 2*offset,
966 :     reference->v + 2*offset, stride, 256*4096);
967 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
968 :     return 1;
969 : chl 181 }
970 : edgomez 851 }
971 : edgomez 504
972 : edgomez 851 static __inline void
973 : Isibaar 1095 ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad)
974 :     {
975 :     pMB->mode = MODE_INTER;
976 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
977 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
978 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
979 :     }
980 :    
981 :     static __inline void
982 : edgomez 1022 ModeDecision(SearchData * const Data,
983 :     MACROBLOCK * const pMB,
984 :     const MACROBLOCK * const pMBs,
985 :     const int x, const int y,
986 :     const MBParam * const pParam,
987 :     const uint32_t MotionFlags,
988 :     const uint32_t VopFlags,
989 :     const uint32_t VolFlags,
990 :     const IMAGE * const pCurrent,
991 : syskin 1084 const IMAGE * const pRef,
992 :     const IMAGE * const vGMC,
993 :     const int coding_type)
994 : edgomez 1022 {
995 :     int mode = MODE_INTER;
996 : syskin 1084 int mcsel = 0;
997 : edgomez 1022 int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
998 :     const uint32_t iQuant = pMB->quant;
999 :    
1000 : syskin 1084 const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0);
1001 : edgomez 1022
1002 : syskin 1084 pMB->mcsel = 0;
1003 :    
1004 : edgomez 1107 if (!(VopFlags & XVID_VOP_MODEDECISION_RD)) { /* normal, fast, SAD-based mode decision */
1005 : edgomez 1022 int sad;
1006 :     int InterBias = MV16_INTER_BIAS;
1007 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1008 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1009 :     mode = MODE_INTER;
1010 :     sad = Data->iMinSAD[0];
1011 :     } else {
1012 :     mode = MODE_INTER4V;
1013 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1014 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1015 :     Data->iMinSAD[0] = sad;
1016 :     }
1017 :    
1018 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
1019 :     if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP))
1020 :     if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH)
1021 :     if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) {
1022 :     mode = MODE_NOT_CODED;
1023 :     sad = 0;
1024 :     }
1025 :    
1026 : syskin 1084 /* mcsel */
1027 :     if (coding_type == S_VOP) {
1028 :    
1029 :     int32_t iSAD = sad16(Data->Cur,
1030 :     vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536);
1031 :    
1032 :     if (Data->chroma) {
1033 :     iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
1034 :     iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
1035 :     }
1036 :    
1037 :     if (iSAD <= sad) { /* mode decision GMC */
1038 :     mode = MODE_INTER;
1039 :     mcsel = 1;
1040 :     sad = iSAD;
1041 :     }
1042 :    
1043 :     }
1044 :    
1045 : edgomez 1022 /* intra decision */
1046 :    
1047 : edgomez 1053 if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */
1048 : edgomez 1022 if (y != 0)
1049 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1050 :     if (x != 0)
1051 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1052 :    
1053 : syskin 1084 if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */
1054 : edgomez 1022 if (Data->rrv) InterBias *= 4;
1055 :    
1056 : syskin 1084 if (InterBias < sad) {
1057 : edgomez 1022 int32_t deviation;
1058 : syskin 1084 if (!Data->rrv)
1059 :     deviation = dev16(Data->Cur, Data->iEdgedWidth);
1060 :     else
1061 :     deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */
1062 :     dev16(Data->Cur+16, Data->iEdgedWidth) +
1063 :     dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) +
1064 :     dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth);
1065 : edgomez 1022
1066 :     if (deviation < (sad - InterBias)) mode = MODE_INTRA;
1067 :     }
1068 :    
1069 : syskin 1084 pMB->cbp = 63;
1070 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
1071 :    
1072 : syskin 1108 } else { /* Rate-Distortion */
1073 : edgomez 1022
1074 : syskin 1108 int min_rd, intra_rd, i, cbp, c[2] = {0, 0};
1075 : edgomez 1022 VECTOR backup[5], *v;
1076 :     Data->iQuant = iQuant;
1077 : syskin 1084 Data->cbp = c;
1078 : edgomez 1022
1079 :     v = Data->qpel ? Data->currentQMV : Data->currentMV;
1080 :     for (i = 0; i < 5; i++) {
1081 :     Data->iMinSAD[i] = 256*4096;
1082 :     backup[i] = v[i];
1083 :     }
1084 :    
1085 : syskin 1108 min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags);
1086 : syskin 1084 cbp = *Data->cbp;
1087 :    
1088 :     if (coding_type == S_VOP) {
1089 : syskin 1108 int gmc_rd;
1090 :     *Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */
1091 :     gmc_rd = findRDgmc(Data, vGMC, x, y);
1092 :     if (gmc_rd < min_rd) {
1093 : syskin 1084 mcsel = 1;
1094 : syskin 1108 *Data->iMinSAD = min_rd = gmc_rd;
1095 : syskin 1084 mode = MODE_INTER;
1096 :     cbp = *Data->cbp;
1097 : edgomez 1022 }
1098 : syskin 1084 }
1099 : edgomez 1022
1100 : syskin 1084 if (inter4v) {
1101 : syskin 1108 int v4_rd;
1102 :     v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1103 :     if (v4_rd < min_rd) {
1104 :     Data->iMinSAD[0] = min_rd = v4_rd;
1105 : syskin 1084 mode = MODE_INTER4V;
1106 :     cbp = *Data->cbp;
1107 :     }
1108 :     }
1109 : edgomez 1022
1110 : syskin 1108 intra_rd = findRDintra(Data);
1111 :     if (intra_rd < min_rd) {
1112 :     *Data->iMinSAD = min_rd = intra_rd;
1113 : syskin 1084 mode = MODE_INTRA;
1114 : edgomez 1022 }
1115 : syskin 1084
1116 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0;
1117 :     pMB->cbp = cbp;
1118 : edgomez 1022 }
1119 :    
1120 :     if (Data->rrv) {
1121 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1122 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1123 :     }
1124 :    
1125 : syskin 1084 if (mode == MODE_INTER && mcsel == 0) {
1126 : edgomez 1022 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1127 :    
1128 :     if(Data->qpel) {
1129 :     pMB->qmvs[0] = pMB->qmvs[1]
1130 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1131 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1132 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1133 :     } else {
1134 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1135 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1136 :     }
1137 :    
1138 : syskin 1084 } else if (mode == MODE_INTER ) { // but mcsel == 1
1139 :    
1140 :     pMB->mcsel = 1;
1141 :     if (Data->qpel) {
1142 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv;
1143 :     pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2;
1144 :     pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2;
1145 :     } else
1146 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv;
1147 :    
1148 :     } else
1149 :     if (mode == MODE_INTER4V) ; /* anything here? */
1150 :     else /* INTRA, NOT_CODED */
1151 : syskin 1108 ZeroMacroblockP(pMB, 0);
1152 : edgomez 1022
1153 :     pMB->mode = mode;
1154 :     }
1155 :    
1156 : edgomez 851 bool
1157 :     MotionEstimation(MBParam * const pParam,
1158 :     FRAMEINFO * const current,
1159 :     FRAMEINFO * const reference,
1160 :     const IMAGE * const pRefH,
1161 :     const IMAGE * const pRefV,
1162 :     const IMAGE * const pRefHV,
1163 : chl 1077 const IMAGE * const pGMC,
1164 : edgomez 851 const uint32_t iLimit)
1165 : chl 181 {
1166 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
1167 :     const IMAGE *const pCurrent = &current->image;
1168 :     const IMAGE *const pRef = &reference->image;
1169 : chl 181
1170 : edgomez 851 uint32_t mb_width = pParam->mb_width;
1171 :     uint32_t mb_height = pParam->mb_height;
1172 :     const uint32_t iEdgedWidth = pParam->edged_width;
1173 : suxen_drol 890 const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
1174 : chl 181
1175 : edgomez 851 uint32_t x, y;
1176 :     uint32_t iIntra = 0;
1177 : syskin 1108 int32_t sad00;
1178 : chl 1077 int skip_thresh = INITIAL_SKIP_THRESH * \
1179 : edgomez 953 (current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \
1180 : edgomez 1107 (current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1);
1181 : chl 181
1182 : edgomez 1053 /* some pre-initialized thingies for SearchP */
1183 : edgomez 851 int32_t temp[8];
1184 :     VECTOR currentMV[5];
1185 :     VECTOR currentQMV[5];
1186 :     int32_t iMinSAD[5];
1187 : syskin 1075 DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE);
1188 : edgomez 851 SearchData Data;
1189 :     memset(&Data, 0, sizeof(SearchData));
1190 :     Data.iEdgedWidth = iEdgedWidth;
1191 :     Data.currentMV = currentMV;
1192 :     Data.currentQMV = currentQMV;
1193 :     Data.iMinSAD = iMinSAD;
1194 :     Data.temp = temp;
1195 :     Data.iFcode = current->fcode;
1196 :     Data.rounding = pParam->m_rounding_type;
1197 : edgomez 963 Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0);
1198 : edgomez 1107 Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP;
1199 : chl 1077 Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0;
1200 : edgomez 962 Data.dctSpace = dct_space;
1201 : edgomez 1022 Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT);
1202 : edgomez 195
1203 : edgomez 949 if ((current->vop_flags & XVID_VOP_REDUCED)) {
1204 : edgomez 851 mb_width = (pParam->width + 31) / 32;
1205 :     mb_height = (pParam->height + 31) / 32;
1206 :     Data.qpel = 0;
1207 :     }
1208 : chl 181
1209 : edgomez 1053 Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */
1210 : edgomez 851 if (sadInit) (*sadInit) ();
1211 : chl 181
1212 : edgomez 851 for (y = 0; y < mb_height; y++) {
1213 :     for (x = 0; x < mb_width; x++) {
1214 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1215 : chl 181
1216 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1217 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1218 :     pRef->y + (x + y * iEdgedWidth) * 16,
1219 :     pParam->edged_width, pMB->sad8 );
1220 : chl 181
1221 : edgomez 851 else pMB->sad16 =
1222 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1223 :     pRef->y + (x + y * iEdgedWidth) * 32,
1224 :     pParam->edged_width, pMB->sad8 );
1225 : chl 181
1226 : edgomez 851 if (Data.chroma) {
1227 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1228 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1229 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1230 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1231 :     pMB->sad16 += Data.temp[7];
1232 :     }
1233 :    
1234 :     sad00 = pMB->sad16;
1235 :    
1236 : edgomez 1053 /* initial skip decision */
1237 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1238 : syskin 1084 if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */
1239 : edgomez 953 if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh)
1240 : edgomez 851 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1241 : syskin 1108 ZeroMacroblockP(pMB, sad00);
1242 :     pMB->mode = MODE_NOT_CODED;
1243 : edgomez 851 continue;
1244 :     }
1245 :     }
1246 :    
1247 : Isibaar 1095 if ((current->vop_flags & XVID_VOP_CARTOON) &&
1248 :     (sad00 < pMB->quant * 4 * skip_thresh)) { /* favorize (0,0) vector for cartoons */
1249 :     ZeroMacroblockP(pMB, sad00);
1250 :     continue;
1251 :     }
1252 :    
1253 : edgomez 851 SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1254 : edgomez 1022 y, MotionFlags, current->vop_flags, current->vol_flags,
1255 :     &Data, pParam, pMBs, reference->mbs, pMB);
1256 : edgomez 851
1257 : edgomez 1022 ModeDecision(&Data, pMB, pMBs, x, y, pParam,
1258 :     MotionFlags, current->vop_flags, current->vol_flags,
1259 : syskin 1084 pCurrent, pRef, pGMC, current->coding_type);
1260 : edgomez 1022
1261 : edgomez 851 if (pMB->mode == MODE_INTRA)
1262 :     if (++iIntra > iLimit) return 1;
1263 : chl 181 }
1264 :     }
1265 : edgomez 851
1266 :     return 0;
1267 : chl 181 }
1268 :    
1269 :    
1270 : edgomez 851 static __inline int
1271 :     make_mask(const VECTOR * const pmv, const int i)
1272 : chl 96 {
1273 : edgomez 851 int mask = 255, j;
1274 :     for (j = 0; j < i; j++) {
1275 : edgomez 1053 if (MVequal(pmv[i], pmv[j])) return 0; /* same vector has been checked already */
1276 : edgomez 851 if (pmv[i].x == pmv[j].x) {
1277 :     if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1278 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1279 :     } else
1280 :     if (pmv[i].y == pmv[j].y) {
1281 :     if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1282 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1283 :     }
1284 :     }
1285 :     return mask;
1286 :     }
1287 : edgomez 195
1288 : edgomez 851 static __inline void
1289 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1290 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1291 :     {
1292 : edgomez 1053 /* this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself */
1293 : edgomez 851 if (rrv) { iWcount /= 2; iHcount /= 2; }
1294 : edgomez 195
1295 : edgomez 1053 if ( (y != 0) && (x < (iWcount-1)) ) { /* [5] top-right neighbour */
1296 : edgomez 851 pmv[5].x = EVEN(pmv[3].x);
1297 :     pmv[5].y = EVEN(pmv[3].y);
1298 :     } else pmv[5].x = pmv[5].y = 0;
1299 : chl 96
1300 : edgomez 1053 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }/* pmv[3] is left neighbour */
1301 : edgomez 851 else pmv[3].x = pmv[3].y = 0;
1302 : chl 96
1303 : edgomez 1053 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }/* [4] top neighbour */
1304 : edgomez 851 else pmv[4].x = pmv[4].y = 0;
1305 : Isibaar 3
1306 : edgomez 1053 /* [1] median prediction */
1307 : edgomez 851 pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1308 : edgomez 195
1309 : 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 */
1310 : edgomez 195
1311 : edgomez 1053 pmv[2].x = EVEN(prevMB->mvs[0].x); /* [2] is last frame */
1312 : edgomez 851 pmv[2].y = EVEN(prevMB->mvs[0].y);
1313 :    
1314 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1315 : edgomez 1053 pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); /* [6] right-down neighbour in last frame */
1316 : edgomez 851 pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1317 :     } else pmv[6].x = pmv[6].y = 0;
1318 :    
1319 :     if (rrv) {
1320 :     int i;
1321 :     for (i = 0; i < 7; i++) {
1322 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1323 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1324 :     }
1325 :     }
1326 : Isibaar 3 }
1327 :    
1328 : edgomez 851 static void
1329 :     SearchP(const IMAGE * const pRef,
1330 :     const uint8_t * const pRefH,
1331 :     const uint8_t * const pRefV,
1332 :     const uint8_t * const pRefHV,
1333 :     const IMAGE * const pCur,
1334 :     const int x,
1335 :     const int y,
1336 :     const uint32_t MotionFlags,
1337 : suxen_drol 890 const uint32_t VopFlags,
1338 : edgomez 1022 const uint32_t VolFlags,
1339 : edgomez 851 SearchData * const Data,
1340 :     const MBParam * const pParam,
1341 :     const MACROBLOCK * const pMBs,
1342 :     const MACROBLOCK * const prevMBs,
1343 :     MACROBLOCK * const pMB)
1344 :     {
1345 : Isibaar 3
1346 : edgomez 851 int i, iDirection = 255, mask, threshA;
1347 :     VECTOR pmv[7];
1348 : edgomez 1022 int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
1349 : Isibaar 3
1350 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1351 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1352 : edgomez 195
1353 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1354 : Isibaar 3
1355 : edgomez 1053 Data->temp[5] = Data->temp[6] = 0; /* chroma-sad cache */
1356 : edgomez 851 i = Data->rrv ? 2 : 1;
1357 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1358 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1359 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1360 : chl 326
1361 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1362 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1363 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1364 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1365 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1366 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1367 : edgomez 195
1368 : edgomez 1022 Data->lambda16 = lambda_vec16[pMB->quant];
1369 :     Data->lambda8 = lambda_vec8[pMB->quant];
1370 : edgomez 851 Data->qpel_precision = 0;
1371 : edgomez 195
1372 : edgomez 963 memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1373 : chl 169
1374 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1375 :     else Data->predMV = pmv[0];
1376 : Isibaar 3
1377 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1378 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1379 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1380 :     Data->iMinSAD[2] = pMB->sad8[1];
1381 :     Data->iMinSAD[3] = pMB->sad8[2];
1382 :     Data->iMinSAD[4] = pMB->sad8[3];
1383 : chl 169
1384 : edgomez 1107 if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) {
1385 : edgomez 1053 threshA = Data->temp[0]; /* that's where we keep this SAD atm */
1386 : edgomez 851 if (threshA < 512) threshA = 512;
1387 :     else if (threshA > 1024) threshA = 1024;
1388 :     } else
1389 :     threshA = 512;
1390 : chl 169
1391 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1392 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1393 : chl 169
1394 : edgomez 851 if (!Data->rrv) {
1395 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1396 : edgomez 1053 else CheckCandidate = CheckCandidate16no4v; /* for extra speed */
1397 : edgomez 851 } else CheckCandidate = CheckCandidate32;
1398 : chl 169
1399 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1400 : chl 169
1401 : edgomez 851 for (i = 1; i < 7; i++) {
1402 :     if (!(mask = make_mask(pmv, i)) ) continue;
1403 :     CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1404 :     if (Data->iMinSAD[0] <= threshA) break;
1405 :     }
1406 : chl 169
1407 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1408 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1409 : edgomez 1022 (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16)))
1410 :     inter4v = 0;
1411 : edgomez 851 else {
1412 : Isibaar 3
1413 : edgomez 851 MainSearchFunc * MainSearchPtr;
1414 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1415 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1416 : edgomez 851 else MainSearchPtr = DiamondSearch;
1417 : Isibaar 3
1418 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1419 : Isibaar 3
1420 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1421 :     note that this search is/might be done in halfpel positions,
1422 :     which makes it more different than the diamond above */
1423 : chl 169
1424 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH16) {
1425 : edgomez 851 int32_t bSAD;
1426 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1427 :     if (Data->rrv) {
1428 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1429 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1430 :     }
1431 :     if (!(MVequal(startMV, backupMV))) {
1432 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1433 : edgomez 195
1434 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1435 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1436 :     if (bSAD < Data->iMinSAD[0]) {
1437 :     Data->currentMV[0] = backupMV;
1438 :     Data->iMinSAD[0] = bSAD; }
1439 :     }
1440 : chl 169
1441 : edgomez 851 backupMV = Data->currentMV[0];
1442 :     startMV.x = startMV.y = 1;
1443 :     if (!(MVequal(startMV, backupMV))) {
1444 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1445 : chl 169
1446 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1447 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1448 :     if (bSAD < Data->iMinSAD[0]) {
1449 :     Data->currentMV[0] = backupMV;
1450 :     Data->iMinSAD[0] = bSAD; }
1451 : edgomez 195 }
1452 : edgomez 851 }
1453 :     }
1454 : edgomez 195
1455 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE16)
1456 : edgomez 851 SubpelRefine(Data);
1457 : edgomez 195
1458 : edgomez 851 for(i = 0; i < 5; i++) {
1459 : edgomez 1053 Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */
1460 : edgomez 851 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1461 :     }
1462 : Isibaar 3
1463 : edgomez 1023 if (Data->qpel) {
1464 : edgomez 953 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1465 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1466 : edgomez 987 Data->qpel_precision = 1;
1467 : edgomez 1023 if (MotionFlags & XVID_ME_QUARTERPELREFINE16)
1468 :     SubpelRefine(Data);
1469 : edgomez 953 }
1470 : Isibaar 3
1471 : edgomez 1023 if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30)
1472 : edgomez 1022 inter4v = 0;
1473 : Isibaar 3
1474 : edgomez 987 if (inter4v) {
1475 : edgomez 851 SearchData Data8;
1476 : edgomez 1053 memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */
1477 : Isibaar 3
1478 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1479 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1480 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1481 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1482 : chl 181
1483 : edgomez 1107 if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) {
1484 : edgomez 1053 /* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */
1485 : edgomez 851 int sumx = 0, sumy = 0;
1486 : Isibaar 3
1487 : edgomez 1022 if (Data->qpel)
1488 :     for (i = 1; i < 5; i++) {
1489 :     sumx += Data->currentQMV[i].x/2;
1490 :     sumy += Data->currentQMV[i].y/2;
1491 :     }
1492 :     else
1493 :     for (i = 1; i < 5; i++) {
1494 :     sumx += Data->currentMV[i].x;
1495 :     sumy += Data->currentMV[i].y;
1496 :     }
1497 : chl 259
1498 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1499 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1500 :     }
1501 : edgomez 1022 } else Data->iMinSAD[1] = 4096*256;
1502 : edgomez 851 }
1503 :    
1504 :     static void
1505 :     Search8(const SearchData * const OldData,
1506 :     const int x, const int y,
1507 :     const uint32_t MotionFlags,
1508 :     const MBParam * const pParam,
1509 :     MACROBLOCK * const pMB,
1510 :     const MACROBLOCK * const pMBs,
1511 :     const int block,
1512 :     SearchData * const Data)
1513 :     {
1514 :     int i = 0;
1515 :     Data->iMinSAD = OldData->iMinSAD + 1 + block;
1516 :     Data->currentMV = OldData->currentMV + 1 + block;
1517 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1518 :    
1519 :     if(Data->qpel) {
1520 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1521 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1522 :     Data->predMV, Data->iFcode, 0, 0);
1523 :     } else {
1524 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1525 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1526 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1527 : Isibaar 3 }
1528 :    
1529 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1530 : Isibaar 3
1531 : edgomez 949 if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) {
1532 : Isibaar 3
1533 : edgomez 978 if (Data->rrv) i = 16; else i = 8;
1534 : Isibaar 3
1535 : edgomez 978 Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1536 :     Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1537 :     Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1538 :     Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1539 :    
1540 :     Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1541 : edgomez 851 Data->qpel_precision = 0;
1542 : Isibaar 3
1543 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1544 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1545 : Isibaar 3
1546 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1547 :     else CheckCandidate = CheckCandidate16no4v;
1548 : Isibaar 3
1549 : edgomez 1107 if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) {
1550 : edgomez 1053 int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */
1551 : Isibaar 3
1552 : edgomez 851 MainSearchFunc *MainSearchPtr;
1553 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch;
1554 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1555 : edgomez 851 else MainSearchPtr = DiamondSearch;
1556 : Isibaar 3
1557 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1558 : Isibaar 3
1559 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1560 : edgomez 1053 Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1561 : edgomez 851 Data->currentQMV->y = 2 * Data->currentMV->y;
1562 :     }
1563 :     }
1564 : edgomez 195
1565 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8) {
1566 : edgomez 1053 int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */
1567 : edgomez 195
1568 : edgomez 1053 SubpelRefine(Data); /* perform halfpel refine of current best vector */
1569 : Isibaar 3
1570 : edgomez 1053 if(*(Data->iMinSAD) < temp_sad) { /* we have found a better match */
1571 :     Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */
1572 : edgomez 851 Data->currentQMV->y = 2 * Data->currentMV->y;
1573 :     }
1574 :     }
1575 : Isibaar 3
1576 : edgomez 949 if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) {
1577 : edgomez 851 Data->qpel_precision = 1;
1578 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1579 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1580 :     SubpelRefine(Data);
1581 :     }
1582 :     }
1583 : edgomez 195
1584 : edgomez 851 if (Data->rrv) {
1585 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1586 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1587 :     }
1588 :    
1589 :     if(Data->qpel) {
1590 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1591 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1592 :     pMB->qmvs[block] = *Data->currentQMV;
1593 : edgomez 195 } else {
1594 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1595 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1596 : edgomez 78 }
1597 : edgomez 851
1598 :     pMB->mvs[block] = *Data->currentMV;
1599 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1600 : Isibaar 3 }
1601 :    
1602 : edgomez 851 /* motion estimation for B-frames */
1603 : chl 345
1604 : edgomez 851 static __inline VECTOR
1605 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1606 :     {
1607 :     /* the stupidiest function ever */
1608 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1609 :     }
1610 : chl 345
1611 : edgomez 851 static void __inline
1612 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1613 :     const uint32_t iWcount,
1614 :     const MACROBLOCK * const pMB,
1615 :     const uint32_t mode_curr)
1616 : chl 345 {
1617 :    
1618 : edgomez 1053 /* [0] is prediction */
1619 : edgomez 851 pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1620 : chl 345
1621 : edgomez 1053 pmv[1].x = pmv[1].y = 0; /* [1] is zero */
1622 : chl 345
1623 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1624 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1625 : chl 345
1626 : edgomez 1053 if ((y != 0)&&(x != (int)(iWcount+1))) { /* [3] top-right neighbour */
1627 : edgomez 851 pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1628 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1629 :     } else pmv[3].x = pmv[3].y = 0;
1630 : chl 345
1631 : edgomez 851 if (y != 0) {
1632 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1633 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1634 :     } else pmv[4].x = pmv[4].y = 0;
1635 : chl 345
1636 : edgomez 851 if (x != 0) {
1637 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1638 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1639 :     } else pmv[5].x = pmv[5].y = 0;
1640 : chl 345
1641 : edgomez 851 if (x != 0 && y != 0) {
1642 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1643 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1644 :     } else pmv[6].x = pmv[6].y = 0;
1645 :     }
1646 : chl 345
1647 :    
1648 : edgomez 851 /* search backward or forward */
1649 :     static void
1650 :     SearchBF( const IMAGE * const pRef,
1651 :     const uint8_t * const pRefH,
1652 :     const uint8_t * const pRefV,
1653 :     const uint8_t * const pRefHV,
1654 :     const IMAGE * const pCur,
1655 :     const int x, const int y,
1656 :     const uint32_t MotionFlags,
1657 :     const uint32_t iFcode,
1658 :     const MBParam * const pParam,
1659 :     MACROBLOCK * const pMB,
1660 :     const VECTOR * const predMV,
1661 :     int32_t * const best_sad,
1662 :     const int32_t mode_current,
1663 :     SearchData * const Data)
1664 :     {
1665 : chl 345
1666 : edgomez 851 int i, iDirection = 255, mask;
1667 :     VECTOR pmv[7];
1668 :     MainSearchFunc *MainSearchPtr;
1669 :     *Data->iMinSAD = MV_MAX_ERROR;
1670 :     Data->iFcode = iFcode;
1671 :     Data->qpel_precision = 0;
1672 : edgomez 1053 Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */
1673 : chl 345
1674 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16;
1675 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16;
1676 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16;
1677 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16;
1678 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1679 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1680 : chl 345
1681 : edgomez 851 Data->predMV = *predMV;
1682 : chl 345
1683 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1684 :     pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1685 : chl 345
1686 : edgomez 851 pmv[0] = Data->predMV;
1687 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1688 : chl 345
1689 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1690 : chl 345
1691 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1692 :     CheckCandidate = CheckCandidate16no4v;
1693 : chl 345
1694 : edgomez 1053 /* main loop. checking all predictions */
1695 : edgomez 851 for (i = 0; i < 7; i++) {
1696 :     if (!(mask = make_mask(pmv, i)) ) continue;
1697 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1698 : chl 345 }
1699 :    
1700 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1701 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1702 : edgomez 851 else MainSearchPtr = DiamondSearch;
1703 : chl 345
1704 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1705 : chl 345
1706 : edgomez 851 SubpelRefine(Data);
1707 : chl 345
1708 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1709 :     Data->currentQMV->x = 2*Data->currentMV->x;
1710 :     Data->currentQMV->y = 2*Data->currentMV->y;
1711 :     Data->qpel_precision = 1;
1712 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1713 :     pParam->width, pParam->height, iFcode, 1, 0);
1714 :     SubpelRefine(Data);
1715 :     }
1716 : chl 345
1717 : edgomez 1053 /* three bits are needed to code backward mode. four for forward */
1718 : Isibaar 3
1719 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1720 :     else *Data->iMinSAD += 3 * Data->lambda16;
1721 : edgomez 195
1722 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1723 :     *best_sad = *Data->iMinSAD;
1724 :     pMB->mode = mode_current;
1725 :     if (Data->qpel) {
1726 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1727 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1728 :     if (mode_current == MODE_FORWARD)
1729 :     pMB->qmvs[0] = *Data->currentQMV;
1730 :     else
1731 :     pMB->b_qmvs[0] = *Data->currentQMV;
1732 :     } else {
1733 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1734 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1735 :     }
1736 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1737 :     else pMB->b_mvs[0] = *Data->currentMV;
1738 :     }
1739 : edgomez 195
1740 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1741 : edgomez 1053 else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */
1742 : Isibaar 3 }
1743 :    
1744 : edgomez 851 static void
1745 :     SkipDecisionB(const IMAGE * const pCur,
1746 :     const IMAGE * const f_Ref,
1747 :     const IMAGE * const b_Ref,
1748 :     MACROBLOCK * const pMB,
1749 :     const uint32_t x, const uint32_t y,
1750 :     const SearchData * const Data)
1751 :     {
1752 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1753 :     int32_t sum;
1754 :     const int div = 1 + Data->qpel;
1755 :     int k;
1756 :     const uint32_t stride = Data->iEdgedWidth/2;
1757 : edgomez 1053 /* this is not full chroma compensation, only it's fullpel approximation. should work though */
1758 : Isibaar 3
1759 : edgomez 851 for (k = 0; k < 4; k++) {
1760 :     dy += Data->directmvF[k].y / div;
1761 : edgomez 963 dx += Data->directmvF[k].x / div;
1762 :     b_dy += Data->directmvB[k].y / div;
1763 :     b_dx += Data->directmvB[k].x / div;
1764 : edgomez 851 }
1765 : Isibaar 3
1766 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1767 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1768 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1769 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1770 : Isibaar 3
1771 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1772 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1773 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1774 :     stride);
1775 : Isibaar 3
1776 : syskin 1112 if (sum >= MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; /* no skip */
1777 : Isibaar 3
1778 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1779 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1780 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1781 :     stride);
1782 : edgomez 195
1783 : syskin 1112 if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1784 : edgomez 1053 pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */
1785 : edgomez 978 for (k = 0; k < 4; k++) {
1786 : edgomez 963 pMB->qmvs[k] = pMB->mvs[k];
1787 :     pMB->b_qmvs[k] = pMB->b_mvs[k];
1788 :     }
1789 :     }
1790 : edgomez 851 }
1791 : Isibaar 3
1792 : edgomez 851 static __inline uint32_t
1793 :     SearchDirect(const IMAGE * const f_Ref,
1794 :     const uint8_t * const f_RefH,
1795 :     const uint8_t * const f_RefV,
1796 :     const uint8_t * const f_RefHV,
1797 :     const IMAGE * const b_Ref,
1798 :     const uint8_t * const b_RefH,
1799 :     const uint8_t * const b_RefV,
1800 :     const uint8_t * const b_RefHV,
1801 :     const IMAGE * const pCur,
1802 :     const int x, const int y,
1803 :     const uint32_t MotionFlags,
1804 :     const int32_t TRB, const int32_t TRD,
1805 :     const MBParam * const pParam,
1806 :     MACROBLOCK * const pMB,
1807 :     const MACROBLOCK * const b_mb,
1808 :     int32_t * const best_sad,
1809 :     SearchData * const Data)
1810 : Isibaar 3
1811 : edgomez 851 {
1812 :     int32_t skip_sad;
1813 :     int k = (x + Data->iEdgedWidth*y) * 16;
1814 :     MainSearchFunc *MainSearchPtr;
1815 : Isibaar 3
1816 : edgomez 851 *Data->iMinSAD = 256*4096;
1817 : edgomez 978 Data->RefP[0] = f_Ref->y + k;
1818 :     Data->RefP[2] = f_RefH + k;
1819 :     Data->RefP[1] = f_RefV + k;
1820 :     Data->RefP[3] = f_RefHV + k;
1821 :     Data->b_RefP[0] = b_Ref->y + k;
1822 :     Data->b_RefP[2] = b_RefH + k;
1823 :     Data->b_RefP[1] = b_RefV + k;
1824 :     Data->b_RefP[3] = b_RefHV + k;
1825 :     Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1826 :     Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1827 :     Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1828 :     Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1829 : edgomez 195
1830 : edgomez 851 k = Data->qpel ? 4 : 2;
1831 :     Data->max_dx = k * (pParam->width - x * 16);
1832 :     Data->max_dy = k * (pParam->height - y * 16);
1833 :     Data->min_dx = -k * (16 + x * 16);
1834 :     Data->min_dy = -k * (16 + y * 16);
1835 : chl 181
1836 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1837 :     Data->qpel_precision = 0;
1838 : edgomez 170
1839 : edgomez 851 for (k = 0; k < 4; k++) {
1840 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1841 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1842 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1843 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1844 : Isibaar 3
1845 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1846 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1847 :    
1848 : edgomez 1053 *best_sad = 256*4096; /* in that case, we won't use direct mode */
1849 :     pMB->mode = MODE_DIRECT; /* just to make sure it doesn't say "MODE_DIRECT_NONE_MV" */
1850 : edgomez 851 pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1851 :     return 256*4096;
1852 :     }
1853 :     if (b_mb->mode != MODE_INTER4V) {
1854 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1855 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1856 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1857 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1858 :     break;
1859 :     }
1860 : edgomez 195 }
1861 : Isibaar 3
1862 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1863 : Isibaar 3
1864 : edgomez 851 CheckCandidate(0, 0, 255, &k, Data);
1865 : edgomez 195
1866 : edgomez 1053 /* initial (fast) skip decision */
1867 : edgomez 1023 if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) {
1868 : edgomez 1053 /* possible skip */
1869 : edgomez 851 if (Data->chroma) {
1870 :     pMB->mode = MODE_DIRECT_NONE_MV;
1871 : edgomez 1053 return *Data->iMinSAD; /* skip. */
1872 : edgomez 851 } else {
1873 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1874 : edgomez 1053 if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */
1875 : edgomez 851 }
1876 : Isibaar 3 }
1877 :    
1878 : edgomez 953 *Data->iMinSAD += Data->lambda16;
1879 : edgomez 851 skip_sad = *Data->iMinSAD;
1880 : syskin 1084
1881 : edgomez 1053 /*
1882 :     * DIRECT MODE DELTA VECTOR SEARCH.
1883 :     * This has to be made more effective, but at the moment I'm happy it's running at all
1884 :     */
1885 : edgomez 195
1886 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1887 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1888 : edgomez 851 else MainSearchPtr = DiamondSearch;
1889 : Isibaar 3
1890 : edgomez 851 MainSearchPtr(0, 0, Data, 255);
1891 : Isibaar 3
1892 : edgomez 851 SubpelRefine(Data);
1893 : chl 181
1894 : edgomez 851 *best_sad = *Data->iMinSAD;
1895 : chl 181
1896 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1897 : edgomez 1053 else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */
1898 : chl 181
1899 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1900 : edgomez 195
1901 : edgomez 851 for (k = 0; k < 4; k++) {
1902 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1903 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1904 :     ? Data->directmvB[k].x
1905 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1906 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1907 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1908 :     ? Data->directmvB[k].y
1909 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1910 :     if (Data->qpel) {
1911 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1912 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1913 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1914 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1915 :     }
1916 : edgomez 195
1917 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1918 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1919 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1920 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1921 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1922 :     break;
1923 :     }
1924 : edgomez 78 }
1925 : edgomez 851 return skip_sad;
1926 :     }
1927 : Isibaar 3
1928 : edgomez 851 static void
1929 :     SearchInterpolate(const IMAGE * const f_Ref,
1930 :     const uint8_t * const f_RefH,
1931 :     const uint8_t * const f_RefV,
1932 :     const uint8_t * const f_RefHV,
1933 :     const IMAGE * const b_Ref,
1934 :     const uint8_t * const b_RefH,
1935 :     const uint8_t * const b_RefV,
1936 :     const uint8_t * const b_RefHV,
1937 :     const IMAGE * const pCur,
1938 :     const int x, const int y,
1939 :     const uint32_t fcode,
1940 :     const uint32_t bcode,
1941 :     const uint32_t MotionFlags,
1942 :     const MBParam * const pParam,
1943 :     const VECTOR * const f_predMV,
1944 :     const VECTOR * const b_predMV,
1945 :     MACROBLOCK * const pMB,
1946 :     int32_t * const best_sad,
1947 :     SearchData * const fData)
1948 : chl 96
1949 : edgomez 851 {
1950 : chl 169
1951 : edgomez 851 int iDirection, i, j;
1952 :     SearchData bData;
1953 : chl 169
1954 : edgomez 851 fData->qpel_precision = 0;
1955 : edgomez 1053 memcpy(&bData, fData, sizeof(SearchData)); /* quick copy of common data */
1956 : edgomez 851 *fData->iMinSAD = 4096*256;
1957 :     bData.currentMV++; bData.currentQMV++;
1958 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1959 : chl 169
1960 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1961 : chl 169
1962 : edgomez 978 bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i;
1963 :     bData.b_RefP[2] = fData->RefP[2] = f_RefH + i;
1964 :     bData.b_RefP[1] = fData->RefP[1] = f_RefV + i;
1965 :     bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i;
1966 :     bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i;
1967 :     bData.RefP[2] = fData->b_RefP[2] = b_RefH + i;
1968 :     bData.RefP[1] = fData->b_RefP[1] = b_RefV + i;
1969 :     bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i;
1970 :     bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1971 :     bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1972 :     bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1973 :     bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1974 : chl 169
1975 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1976 :     fData->bpredMV = bData.predMV = *b_predMV;
1977 :     fData->currentMV[0] = fData->currentMV[2];
1978 : Isibaar 3
1979 : 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);
1980 :     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);
1981 : Isibaar 3
1982 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1983 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1984 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1985 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1986 : chl 169
1987 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1988 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1989 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1990 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1991 : Isibaar 3
1992 : edgomez 851 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1993 : chl 169
1994 : edgomez 1053 /* diamond */
1995 : edgomez 851 do {
1996 :     iDirection = 255;
1997 : edgomez 1053 /* forward MV moves */
1998 : edgomez 851 i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1999 : chl 169
2000 : edgomez 851 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
2001 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
2002 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
2003 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
2004 : edgomez 195
2005 : edgomez 1053 /* backward MV moves */
2006 : edgomez 851 i = fData->currentMV[1].x; j = fData->currentMV[1].y;
2007 :     fData->currentMV[2] = fData->currentMV[0];
2008 :     CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
2009 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
2010 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
2011 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
2012 : Isibaar 3
2013 : edgomez 851 } while (!(iDirection));
2014 : chl 140
2015 : edgomez 1053 /* qpel refinement */
2016 : edgomez 851 if (fData->qpel) {
2017 :     if (*fData->iMinSAD > *best_sad + 500) return;
2018 :     CheckCandidate = CheckCandidateInt;
2019 :     fData->qpel_precision = bData.qpel_precision = 1;
2020 :     get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
2021 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
2022 :     fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
2023 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
2024 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
2025 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
2026 :     SubpelRefine(fData);
2027 :     if (*fData->iMinSAD > *best_sad + 300) return;
2028 :     fData->currentQMV[2] = fData->currentQMV[0];
2029 :     SubpelRefine(&bData);
2030 :     }
2031 : chl 140
2032 : edgomez 1053 *fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */
2033 : Isibaar 3
2034 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
2035 :     *best_sad = *fData->iMinSAD;
2036 :     pMB->mvs[0] = fData->currentMV[0];
2037 :     pMB->b_mvs[0] = fData->currentMV[1];
2038 :     pMB->mode = MODE_INTERPOLATE;
2039 :     if (fData->qpel) {
2040 :     pMB->qmvs[0] = fData->currentQMV[0];
2041 :     pMB->b_qmvs[0] = fData->currentQMV[1];
2042 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
2043 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
2044 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
2045 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
2046 :     } else {
2047 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
2048 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
2049 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
2050 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
2051 :     }
2052 : edgomez 78 }
2053 : edgomez 851 }
2054 : Isibaar 3
2055 : edgomez 851 void
2056 :     MotionEstimationBVOP(MBParam * const pParam,
2057 :     FRAMEINFO * const frame,
2058 :     const int32_t time_bp,
2059 :     const int32_t time_pp,
2060 : edgomez 1053 /* forward (past) reference */
2061 : edgomez 851 const MACROBLOCK * const f_mbs,
2062 :     const IMAGE * const f_ref,
2063 :     const IMAGE * const f_refH,
2064 :     const IMAGE * const f_refV,
2065 :     const IMAGE * const f_refHV,
2066 : edgomez 1053 /* backward (future) reference */
2067 : edgomez 851 const FRAMEINFO * const b_reference,
2068 :     const IMAGE * const b_ref,
2069 :     const IMAGE * const b_refH,
2070 :     const IMAGE * const b_refV,
2071 :     const IMAGE * const b_refHV)
2072 :     {
2073 :     uint32_t i, j;
2074 :     int32_t best_sad;
2075 :     uint32_t skip_sad;
2076 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2077 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
2078 : Isibaar 3
2079 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2080 : Isibaar 3
2081 : edgomez 851 const int32_t TRB = time_pp - time_bp;
2082 :     const int32_t TRD = time_pp;
2083 : edgomez 195
2084 : edgomez 1053 /* some pre-inintialized data for the rest of the search */
2085 : Isibaar 3
2086 : edgomez 851 SearchData Data;
2087 :     int32_t iMinSAD;
2088 :     VECTOR currentMV[3];
2089 :     VECTOR currentQMV[3];
2090 :     int32_t temp[8];
2091 :     memset(&Data, 0, sizeof(SearchData));
2092 :     Data.iEdgedWidth = pParam->edged_width;
2093 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2094 :     Data.iMinSAD = &iMinSAD;
2095 :     Data.lambda16 = lambda_vec16[frame->quant];
2096 : syskin 1112 Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0;
2097 : edgomez 851 Data.rounding = 0;
2098 : edgomez 1107 Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP;
2099 : edgomez 851 Data.temp = temp;
2100 : Isibaar 3
2101 : edgomez 1053 Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */
2102 :    
2103 :     /* note: i==horizontal, j==vertical */
2104 : edgomez 851 for (j = 0; j < pParam->mb_height; j++) {
2105 : edgomez 195
2106 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2107 : Isibaar 3
2108 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2109 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2110 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2111 : edgomez 195
2112 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2113 :     if (b_reference->coding_type != S_VOP)
2114 :     if (b_mb->mode == MODE_NOT_CODED) {
2115 :     pMB->mode = MODE_NOT_CODED;
2116 :     continue;
2117 :     }
2118 : Isibaar 3
2119 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2120 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2121 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2122 : edgomez 953 pMB->quant = frame->quant;
2123 : Isibaar 3
2124 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2125 :     and (2) sets very good predictions for forward and backward search */
2126 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2127 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2128 :     &frame->image,
2129 :     i, j,
2130 :     frame->motion_flags,
2131 :     TRB, TRD,
2132 :     pParam,
2133 :     pMB, b_mb,
2134 :     &best_sad,
2135 :     &Data);
2136 : Isibaar 3
2137 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2138 : edgomez 195
2139 : edgomez 1053 /* forward search */
2140 : edgomez 851 SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2141 :     &frame->image, i, j,
2142 :     frame->motion_flags,
2143 :     frame->fcode, pParam,
2144 :     pMB, &f_predMV, &best_sad,
2145 :     MODE_FORWARD, &Data);
2146 : edgomez 195
2147 : edgomez 1053 /* backward search */
2148 : edgomez 851 SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2149 :     &frame->image, i, j,
2150 :     frame->motion_flags,
2151 :     frame->bcode, pParam,
2152 :     pMB, &b_predMV, &best_sad,
2153 :     MODE_BACKWARD, &Data);
2154 :    
2155 : edgomez 1053 /* interpolate search comes last, because it uses data from forward and backward as prediction */
2156 : edgomez 851 SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2157 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2158 :     &frame->image,
2159 :     i, j,
2160 :     frame->fcode, frame->bcode,
2161 :     frame->motion_flags,
2162 :     pParam,
2163 :     &f_predMV, &b_predMV,
2164 :     pMB, &best_sad,
2165 :     &Data);
2166 :    
2167 : edgomez 1053 /* final skip decision */
2168 : edgomez 851 if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2169 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2170 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2171 :    
2172 :     switch (pMB->mode) {
2173 :     case MODE_FORWARD:
2174 :     f_count++;
2175 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2176 :     break;
2177 :     case MODE_BACKWARD:
2178 :     b_count++;
2179 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2180 :     break;
2181 :     case MODE_INTERPOLATE:
2182 :     i_count++;
2183 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2184 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2185 :     break;
2186 :     case MODE_DIRECT:
2187 :     case MODE_DIRECT_NO4V:
2188 :     d_count++;
2189 :     default:
2190 :     break;
2191 :     }
2192 :     }
2193 :     }
2194 : Isibaar 3 }
2195 : chl 96
2196 : edgomez 851 static __inline void
2197 :     MEanalyzeMB ( const uint8_t * const pRef,
2198 :     const uint8_t * const pCur,
2199 :     const int x,
2200 :     const int y,
2201 :     const MBParam * const pParam,
2202 :     MACROBLOCK * const pMBs,
2203 :     SearchData * const Data)
2204 : chl 96 {
2205 :    
2206 : edgomez 851 int i, mask;
2207 : edgomez 953 int quarterpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0;
2208 : edgomez 851 VECTOR pmv[3];
2209 : edgomez 953 MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2210 : chl 96
2211 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2212 : chl 96
2213 : edgomez 1053 /* median is only used as prediction. it doesn't have to be real */
2214 : edgomez 851 if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2215 :     else
2216 : edgomez 1053 if (x == 1) /* left macroblock does not have any vector now */
2217 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; /* top instead of median */
2218 :     else if (y == 1) /* top macroblock doesn't have it's vector */
2219 :     Data->predMV = (pMB - 1)->mvs[0]; /* left instead of median */
2220 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); /* else median */
2221 : edgomez 195
2222 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2223 : syskin 1084 pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0);
2224 : edgomez 195
2225 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2226 : edgomez 978 Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16;
2227 : chl 96
2228 : edgomez 851 pmv[1].x = EVEN(pMB->mvs[0].x);
2229 :     pmv[1].y = EVEN(pMB->mvs[0].y);
2230 :     pmv[2].x = EVEN(Data->predMV.x);
2231 :     pmv[2].y = EVEN(Data->predMV.y);
2232 :     pmv[0].x = pmv[0].y = 0;
2233 : edgomez 195
2234 : edgomez 851 CheckCandidate32I(0, 0, 255, &i, Data);
2235 : edgomez 195
2236 : edgomez 953 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) {
2237 : chl 96
2238 : edgomez 851 if (!(mask = make_mask(pmv, 1)))
2239 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2240 :     if (!(mask = make_mask(pmv, 2)))
2241 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2242 : chl 96
2243 : edgomez 1053 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) /* diamond only if needed */
2244 : edgomez 851 DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2245 : edgomez 953 }
2246 : edgomez 851
2247 : edgomez 953 for (i = 0; i < 4; i++) {
2248 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2249 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2250 :     MB->mode = MODE_INTER;
2251 :     MB->sad16 = Data->iMinSAD[i+1];
2252 : chl 96 }
2253 : edgomez 851 }
2254 : chl 96
2255 : syskin 1075 #define INTRA_THRESH 2200
2256 :     #define INTER_THRESH 50
2257 :     #define INTRA_THRESH2 95
2258 : chl 96
2259 : edgomez 851 int
2260 :     MEanalysis( const IMAGE * const pRef,
2261 : edgomez 953 const FRAMEINFO * const Current,
2262 :     const MBParam * const pParam,
2263 : syskin 1075 const int maxIntra, //maximum number if non-I frames
2264 :     const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2265 : syskin 1084 const int bCount, // number of B frames in a row
2266 : edgomez 953 const int b_thresh)
2267 : edgomez 851 {
2268 :     uint32_t x, y, intra = 0;
2269 :     int sSAD = 0;
2270 :     MACROBLOCK * const pMBs = Current->mbs;
2271 :     const IMAGE * const pCurrent = &Current->image;
2272 : syskin 1075 int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh;
2273 : chl 1077 int blocks = 0;
2274 : syskin 1075 int complexity = 0;
2275 : chl 96
2276 : edgomez 851 int32_t iMinSAD[5], temp[5];
2277 :     VECTOR currentMV[5];
2278 :     SearchData Data;
2279 :     Data.iEdgedWidth = pParam->edged_width;
2280 :     Data.currentMV = currentMV;
2281 :     Data.iMinSAD = iMinSAD;
2282 :     Data.iFcode = Current->fcode;
2283 :     Data.temp = temp;
2284 :     CheckCandidate = CheckCandidate32I;
2285 : chl 96
2286 : edgomez 1023 if (intraCount != 0) {
2287 : syskin 1075 if (intraCount < 10) // we're right after an I frame
2288 : edgomez 1023 IntraThresh += 15* (intraCount - 10) * (intraCount - 10);
2289 :     else
2290 : syskin 1075 if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2291 : edgomez 1023 IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra;
2292 :     }
2293 : chl 96
2294 : syskin 1075 InterThresh -= 12 * bCount;
2295 :     if (InterThresh < 15 + b_thresh) InterThresh = 15 + b_thresh;
2296 : chl 326
2297 : edgomez 851 if (sadInit) (*sadInit) ();
2298 : chl 96
2299 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2300 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2301 :     int i;
2302 : syskin 1075 blocks += 10;
2303 : edgomez 195
2304 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2305 : syskin 1075 else { //extrapolation of the vector found for last frame
2306 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].x =
2307 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2308 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].y =
2309 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2310 :     }
2311 : edgomez 195
2312 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2313 : edgomez 195
2314 : edgomez 851 for (i = 0; i < 4; i++) {
2315 :     int dev;
2316 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2317 : syskin 1075 dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2318 :     pParam->edged_width);
2319 : syskin 1084
2320 : syskin 1108 complexity += MAX(dev, 300);
2321 : syskin 1075 if (dev + IntraThresh < pMB->sad16) {
2322 :     pMB->mode = MODE_INTRA;
2323 :     if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2324 : edgomez 851 }
2325 : edgomez 953
2326 : syskin 1084 if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0)
2327 : syskin 1075 if (dev > 500 && pMB->sad16 < 1000)
2328 :     sSAD += 1000;
2329 :    
2330 : syskin 1108 sSAD += (dev < 3000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */
2331 : edgomez 851 }
2332 :     }
2333 : edgomez 195 }
2334 : syskin 1075 complexity >>= 7;
2335 : edgomez 953
2336 : syskin 1075 sSAD /= complexity + 4*blocks;
2337 : edgomez 953
2338 : syskin 1108 if (intraCount > 80 && sSAD > INTRA_THRESH2 ) return I_VOP;
2339 : edgomez 851 if (sSAD > InterThresh ) return P_VOP;
2340 :     emms();
2341 :     return B_VOP;
2342 :     }
2343 : chl 96
2344 :    
2345 : edgomez 1053 /* functions which perform BITS-based search/bitcount */
2346 : chl 96
2347 : edgomez 851 static int
2348 : syskin 1108 findRDinter(SearchData * const Data,
2349 :     const MACROBLOCK * const pMBs, const int x, const int y,
2350 :     const MBParam * const pParam,
2351 :     const uint32_t MotionFlags)
2352 : chl 96 {
2353 : edgomez 851 int i, iDirection;
2354 :     int32_t bsad[5];
2355 : chl 141
2356 : syskin 1108 CheckCandidate = CheckCandidateRD16;
2357 : chl 96
2358 : edgomez 851 if (Data->qpel) {
2359 :     for(i = 0; i < 5; i++) {
2360 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2361 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2362 :     }
2363 :     Data->qpel_precision = 1;
2364 : syskin 1108 CheckCandidateRD16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2365 : chl 96
2366 : edgomez 1107 if (MotionFlags & (XVID_ME_HALFPELREFINE16_RD | XVID_ME_EXTSEARCH_RD)) { /* we have to prepare for halfpixel-precision search */
2367 : edgomez 851 for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2368 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2369 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2370 :     Data->qpel_precision = 0;
2371 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2372 : syskin 1108 CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2373 : edgomez 851 }
2374 : edgomez 195
2375 : edgomez 1053 } else { /* not qpel */
2376 : edgomez 195
2377 : syskin 1108 CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2378 : edgomez 851 }
2379 : chl 96
2380 : edgomez 1107 if (MotionFlags&XVID_ME_EXTSEARCH_RD) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2381 : chl 96
2382 : edgomez 1107 if (MotionFlags&XVID_ME_HALFPELREFINE16_RD) SubpelRefine(Data);
2383 : chl 96
2384 : edgomez 851 if (Data->qpel) {
2385 : edgomez 1107 if (MotionFlags&(XVID_ME_EXTSEARCH_RD | XVID_ME_HALFPELREFINE16_RD)) { /* there was halfpel-precision search */
2386 : edgomez 851 for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2387 : edgomez 1053 Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* we have found a better match */
2388 : edgomez 851 Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2389 :     }
2390 : edgomez 195
2391 : edgomez 1053 /* preparing for qpel-precision search */
2392 : edgomez 851 Data->qpel_precision = 1;
2393 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2394 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
2395 :     }
2396 : edgomez 1107 if (MotionFlags&XVID_ME_QUARTERPELREFINE16_RD) SubpelRefine(Data);
2397 : edgomez 851 }
2398 : chl 96
2399 : edgomez 1107 if (MotionFlags&XVID_ME_CHECKPREDICTION_RD) { /* let's check vector equal to prediction */
2400 : edgomez 851 VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2401 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2402 : syskin 1108 CheckCandidateRD16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2403 : edgomez 851 }
2404 :     return Data->iMinSAD[0];
2405 :     }
2406 : chl 96
2407 : edgomez 851 static int
2408 : syskin 1108 findRDinter4v(const SearchData * const Data,
2409 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2410 :     const int x, const int y,
2411 :     const MBParam * const pParam, const uint32_t MotionFlags,
2412 :     const VECTOR * const backup)
2413 : edgomez 851 {
2414 : chl 96
2415 : edgomez 851 int cbp = 0, bits = 0, t = 0, i, iDirection;
2416 :     SearchData Data2, *Data8 = &Data2;
2417 :     int sumx = 0, sumy = 0;
2418 : edgomez 962 int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64;
2419 : edgomez 1022 uint8_t * ptr;
2420 : chl 96
2421 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2422 : syskin 1108 CheckCandidate = CheckCandidateRD8;
2423 : chl 96
2424 : edgomez 1053 for (i = 0; i < 4; i++) { /* for all luma blocks */
2425 : edgomez 1022
2426 : edgomez 851 Data8->iMinSAD = Data->iMinSAD + i + 1;
2427 :     Data8->currentMV = Data->currentMV + i + 1;
2428 :     Data8->currentQMV = Data->currentQMV + i + 1;
2429 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2430 : edgomez 978 Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2431 :     Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2432 :     Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2433 :     Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2434 : syskin 1084 *Data8->cbp = (Data->cbp[1] & (1<<(5-i))) ? 1:0; // copy corresponding cbp bit
2435 : edgomez 195
2436 : edgomez 851 if(Data->qpel) {
2437 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2438 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2439 :     Data8->predMV, Data8->iFcode, 0, 0);
2440 :     } else {
2441 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2442 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2443 :     Data8->predMV, Data8->iFcode, 0, 0);
2444 :     }
2445 : chl 96
2446 : 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,
2447 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2448 : chl 96
2449 : edgomez 1023 *Data8->iMinSAD += BITS_MULT*t;
2450 : chl 96
2451 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2452 : edgomez 1053 /* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */
2453 : edgomez 987 {
2454 :     VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV;
2455 : edgomez 1022 if (!MVequal (*v, backup[i+1]) )
2456 : syskin 1108 CheckCandidateRD8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2457 : edgomez 851 }
2458 : edgomez 195
2459 : edgomez 851 if (Data8->qpel) {
2460 : edgomez 1107 if (MotionFlags&XVID_ME_HALFPELREFINE8_RD || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_RD)) { /* halfpixel motion search follows */
2461 : edgomez 851 int32_t s = *Data8->iMinSAD;
2462 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2463 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2464 :     Data8->qpel_precision = 0;
2465 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2466 :     pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2467 : edgomez 195
2468 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2469 : syskin 1108 CheckCandidateRD8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2470 : edgomez 195
2471 : edgomez 1107 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD)
2472 : edgomez 851 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2473 : chl 96
2474 : edgomez 1107 if (MotionFlags & XVID_ME_HALFPELREFINE8_RD)
2475 : edgomez 1022 SubpelRefine(Data8);
2476 : chl 96
2477 : edgomez 1053 if(s > *Data8->iMinSAD) { /* we have found a better match */
2478 : edgomez 851 Data8->currentQMV->x = 2*Data8->currentMV->x;
2479 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2480 :     }
2481 : chl 96
2482 : edgomez 851 Data8->qpel_precision = 1;
2483 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2484 :     pParam->width, pParam->height, Data8->iFcode, 1, 0);
2485 : chl 96
2486 : edgomez 851 }
2487 : edgomez 1107 if (MotionFlags & XVID_ME_QUARTERPELREFINE8_RD) SubpelRefine(Data8);
2488 : edgomez 195
2489 : edgomez 1053 } else { /* not qpel */
2490 : edgomez 851
2491 : edgomez 1107 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) /* extsearch */
2492 : edgomez 1022 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2493 :    
2494 : edgomez 1107 if (MotionFlags & XVID_ME_HALFPELREFINE8_RD)
2495 : edgomez 1053 SubpelRefine(Data8); /* halfpel refinement */
2496 : edgomez 1022 }
2497 :    
2498 : edgomez 1053 /* checking vector equal to predicion */
2499 : edgomez 1107 if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_RD) {
2500 : edgomez 851 const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2501 : edgomez 1022 if (!MVequal(*v, Data8->predMV))
2502 : syskin 1108 CheckCandidateRD8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2503 : chl 141 }
2504 : edgomez 195
2505 : edgomez 851 bits += *Data8->iMinSAD;
2506 : edgomez 1053 if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */
2507 : edgomez 851
2508 : edgomez 1053 /* MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else */
2509 : edgomez 851 if(Data->qpel) {
2510 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2511 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2512 :     pMB->qmvs[i] = *Data8->currentQMV;
2513 :     sumx += Data8->currentQMV->x/2;
2514 :     sumy += Data8->currentQMV->y/2;
2515 :     } else {
2516 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2517 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2518 :     sumx += Data8->currentMV->x;
2519 :     sumy += Data8->currentMV->y;
2520 : chl 141 }
2521 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2522 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2523 : syskin 1084 if (Data8->cbp[0]) cbp |= 1 << (5 - i);
2524 : chl 141
2525 : syskin 1084 } /* end - for all luma blocks */
2526 : chl 141
2527 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
2528 : chl 141
2529 : edgomez 1053 /* let's check chroma */
2530 : edgomez 1022 sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2531 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2532 : chl 96
2533 : edgomez 1053 /* chroma U */
2534 : edgomez 1022 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2535 :     transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2536 : syskin 1075 bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4);
2537 : edgomez 1022
2538 :     if (bits >= *Data->iMinSAD) return bits;
2539 :    
2540 : edgomez 1053 /* chroma V */
2541 : edgomez 1022 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2542 :     transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2543 : syskin 1075 bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5);
2544 : edgomez 1022
2545 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2546 : edgomez 1022
2547 : syskin 1084 *Data->cbp = cbp;
2548 : edgomez 851 return bits;
2549 :     }
2550 : chl 96
2551 : edgomez 851 static int
2552 : syskin 1108 findRDintra(const SearchData * const Data)
2553 : edgomez 851 {
2554 : edgomez 1053 int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */
2555 : edgomez 1023 int cbp = 0, i, dc = 0;
2556 : edgomez 962 int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2557 : edgomez 195
2558 : edgomez 851 for(i = 0; i < 4; i++) {
2559 :     int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2560 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2561 : syskin 1075 bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc);
2562 : chl 141
2563 : edgomez 1022 if (bits >= Data->iMinSAD[0]) return bits;
2564 : edgomez 851 }
2565 : chl 181
2566 : edgomez 1023 bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len;
2567 : chl 96
2568 : edgomez 1053 /*chroma U */
2569 : edgomez 1022 transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2570 : syskin 1075 bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc);
2571 : syskin 1084
2572 : edgomez 1022 if (bits >= Data->iMinSAD[0]) return bits;
2573 :    
2574 : edgomez 1053 /* chroma V */
2575 : edgomez 1022 transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2576 : syskin 1075 bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc);
2577 : edgomez 1022
2578 : edgomez 1023 bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2579 : edgomez 1022
2580 : edgomez 851 return bits;
2581 : chl 96 }
2582 : chl 1077
2583 : syskin 1084 static int
2584 : syskin 1108 findRDgmc(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y)
2585 : syskin 1084 {
2586 :     int bits = BITS_MULT*1; /* this one is mcsel */
2587 :     int cbp = 0, i;
2588 :     int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2589 : chl 1077
2590 : syskin 1084 for(i = 0; i < 4; i++) {
2591 :     int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2592 :     transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16*(x+y*Data->iEdgedWidth), Data->iEdgedWidth);
2593 :     bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i);
2594 :     if (bits >= Data->iMinSAD[0]) return bits;
2595 :     }
2596 : chl 1077
2597 : syskin 1084 bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len;
2598 : chl 1077
2599 : syskin 1084 /*chroma U */
2600 :     transfer_8to16subro(in, Data->CurU, vGMC->u + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2);
2601 :     bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4);
2602 : chl 1077
2603 : syskin 1084 if (bits >= Data->iMinSAD[0]) return bits;
2604 :    
2605 :     /* chroma V */
2606 :     transfer_8to16subro(in, Data->CurV , vGMC->v + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2);
2607 :     bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5);
2608 :    
2609 :     bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
2610 :    
2611 :     *Data->cbp = cbp;
2612 :    
2613 :     return bits;
2614 :     }
2615 :    
2616 :    
2617 :    
2618 :    
2619 : chl 1077 static __inline void
2620 :     GMEanalyzeMB ( const uint8_t * const pCur,
2621 :     const uint8_t * const pRef,
2622 :     const uint8_t * const pRefH,
2623 :     const uint8_t * const pRefV,
2624 :     const uint8_t * const pRefHV,
2625 :     const int x,
2626 :     const int y,
2627 :     const MBParam * const pParam,
2628 :     MACROBLOCK * const pMBs,
2629 :     SearchData * const Data)
2630 :     {
2631 :    
2632 :     int i=0;
2633 :     MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2634 :    
2635 :     Data->iMinSAD[0] = MV_MAX_ERROR;
2636 :    
2637 : syskin 1084 Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0);
2638 : chl 1077
2639 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2640 : syskin 1084 pParam->width, pParam->height, 16, 0, 0);
2641 : chl 1077
2642 :     Data->Cur = pCur + 16*(x + y * pParam->edged_width);
2643 :     Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width);
2644 :     Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width);
2645 :     Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width);
2646 :     Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width);
2647 :    
2648 :     Data->currentMV[0].x = Data->currentMV[0].y = 0;
2649 :     CheckCandidate16I(0, 0, 255, &i, Data);
2650 :    
2651 :     if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) )
2652 :     CheckCandidate16I(Data->predMV.x, Data->predMV.y, 255, &i, Data);
2653 :    
2654 : syskin 1084 AdvDiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255);
2655 : chl 1077
2656 : syskin 1084 SubpelRefine(Data);
2657 :    
2658 :    
2659 : chl 1077 /* for QPel halfpel positions are worse than in halfpel mode :( */
2660 :     /* if (Data->qpel) {
2661 :     Data->currentQMV->x = 2*Data->currentMV->x;
2662 :     Data->currentQMV->y = 2*Data->currentMV->y;
2663 :     Data->qpel_precision = 1;
2664 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2665 :     pParam->width, pParam->height, iFcode, 1, 0);
2666 :     SubpelRefine(Data);
2667 :     }
2668 :     */
2669 :    
2670 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
2671 :     pMB->sad16 = Data->iMinSAD[0];
2672 : syskin 1084 pMB->mode = MODE_INTER;
2673 :     pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0);
2674 : chl 1077 return;
2675 :     }
2676 :    
2677 :     void
2678 :     GMEanalysis(const MBParam * const pParam,
2679 :     const FRAMEINFO * const current,
2680 :     const FRAMEINFO * const reference,
2681 :     const IMAGE * const pRefH,
2682 :     const IMAGE * const pRefV,
2683 :     const IMAGE * const pRefHV)
2684 :     {
2685 :     uint32_t x, y;
2686 :     MACROBLOCK * const pMBs = current->mbs;
2687 :     const IMAGE * const pCurrent = &current->image;
2688 :     const IMAGE * const pReference = &reference->image;
2689 :    
2690 :     int32_t iMinSAD[5], temp[5];
2691 :     VECTOR currentMV[5];
2692 :     SearchData Data;
2693 :     memset(&Data, 0, sizeof(SearchData));
2694 : syskin 1084
2695 : chl 1077 Data.iEdgedWidth = pParam->edged_width;
2696 :     Data.rounding = pParam->m_rounding_type;
2697 : syskin 1084
2698 : chl 1077 Data.currentMV = &currentMV[0];
2699 :     Data.iMinSAD = &iMinSAD[0];
2700 :     Data.iFcode = current->fcode;
2701 :     Data.temp = temp;
2702 :    
2703 :     CheckCandidate = CheckCandidate16I;
2704 : syskin 1084
2705 : chl 1077 if (sadInit) (*sadInit) ();
2706 :    
2707 :     for (y = 0; y < pParam->mb_height; y ++) {
2708 :     for (x = 0; x < pParam->mb_width; x ++) {
2709 :    
2710 :     GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data);
2711 :     }
2712 :     }
2713 :     return;
2714 :     }
2715 :    
2716 :    
2717 :     WARPPOINTS
2718 :     GlobalMotionEst(MACROBLOCK * const pMBs,
2719 :     const MBParam * const pParam,
2720 :     const FRAMEINFO * const current,
2721 :     const FRAMEINFO * const reference,
2722 :     const IMAGE * const pRefH,
2723 :     const IMAGE * const pRefV,
2724 :     const IMAGE * const pRefHV)
2725 :     {
2726 :    
2727 : syskin 1084 const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2728 :     const int deltay=8;
2729 : chl 1077 const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks)
2730 : syskin 1084 const unsigned int grady=512;
2731 : chl 1077
2732 :     double sol[4] = { 0., 0., 0., 0. };
2733 : syskin 1084
2734 : chl 1077 WARPPOINTS gmc;
2735 :    
2736 :     uint32_t mx, my;
2737 :    
2738 :     int MBh = pParam->mb_height;
2739 :     int MBw = pParam->mb_width;
2740 :     const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */
2741 :     const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */
2742 :    
2743 :     int num=0;
2744 :     int oldnum;
2745 :    
2746 :     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;
2747 :    
2748 :     GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV);
2749 :    
2750 :     /* block based ME isn't done, yet, so do a quick presearch */
2751 :    
2752 :     // filter mask of all blocks
2753 :    
2754 :     for (my = 0; my < (uint32_t)MBh; my++)
2755 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2756 :     {
2757 :     const int mbnum = mx + my * MBw;
2758 :     pMBs[mbnum].mcsel = 0;
2759 :     }
2760 : syskin 1084
2761 :    
2762 :     for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */
2763 : chl 1077 for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */
2764 :     {
2765 :     const int mbnum = mx + my * MBw;
2766 :     MACROBLOCK *const pMB = &pMBs[mbnum];
2767 :     const VECTOR mv = pMB->mvs[0];
2768 :    
2769 :     /* don't use object boundaries */
2770 : syskin 1084 if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax)
2771 : chl 1077 && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay)
2772 : syskin 1084 && (abs(mv.x - (pMB+1)->mvs[0].x) < deltax)
2773 : chl 1077 && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay)
2774 : syskin 1084 && (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax)
2775 : chl 1077 && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay)
2776 : syskin 1084 && (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax)
2777 : chl 1077 && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) )
2778 :     { const int iEdgedWidth = pParam->edged_width;
2779 :     const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx);
2780 :     if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx )
2781 :     && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) )
2782 : syskin 1084 { pMB->mcsel = 1;
2783 : chl 1077 num++;
2784 :     }
2785 :    
2786 :     /* only use "structured" blocks */
2787 :     }
2788 :     }
2789 :     emms();
2790 :    
2791 : syskin 1084 /* further filtering would be possible, but during iteration, remaining
2792 : chl 1077 outliers usually are removed, too */
2793 :    
2794 :     if (num>= minblocks)
2795 :     do { /* until convergence */
2796 :     double DtimesF[4];
2797 :     double a,b,c,n,invdenom;
2798 :     double meanx,meany;
2799 :    
2800 :     a = b = c = n = 0;
2801 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2802 :     for (my = 1; my < (uint32_t)MBh-1; my++)
2803 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2804 :     {
2805 :     const int mbnum = mx + my * MBw;
2806 :     const VECTOR mv = pMBs[mbnum].mvs[0];
2807 :    
2808 :     if (!pMBs[mbnum].mcsel)
2809 :     continue;
2810 :    
2811 :     n++;
2812 :     a += 16*mx+8;
2813 :     b += 16*my+8;
2814 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2815 :    
2816 :     DtimesF[0] += (double)mv.x;
2817 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2818 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2819 :     DtimesF[3] += (double)mv.y;
2820 :     }
2821 : syskin 1084
2822 : chl 1077 invdenom = a*a+b*b-c*n;
2823 :    
2824 :     /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2825 :     /* D'*E*F has been calculated in the same loop as matrix */
2826 :    
2827 :     sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2828 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2829 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2830 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2831 :    
2832 :     sol[0] /= invdenom;
2833 :     sol[1] /= invdenom;
2834 :     sol[2] /= invdenom;
2835 :     sol[3] /= invdenom;
2836 :    
2837 :     meanx = meany = 0.;
2838 :     oldnum = 0;
2839 :     for (my = 1; my < (uint32_t)MBh-1; my++)
2840 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2841 :     {
2842 :     const int mbnum = mx + my * MBw;
2843 :     const VECTOR mv = pMBs[mbnum].mvs[0];
2844 :    
2845 :     if (!pMBs[mbnum].mcsel)
2846 :     continue;
2847 :    
2848 :     oldnum++;
2849 :     meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x );
2850 :     meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y );
2851 :     }
2852 :    
2853 :     if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */
2854 :     meanx /= oldnum;
2855 :     else
2856 :     meanx = 0.25;
2857 :    
2858 :     if (4*meany > oldnum)
2859 :     meany /= oldnum;
2860 :     else
2861 :     meany = 0.25;
2862 :    
2863 :     num = 0;
2864 :     for (my = 0; my < (uint32_t)MBh; my++)
2865 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2866 :     {
2867 :     const int mbnum = mx + my * MBw;
2868 :     const VECTOR mv = pMBs[mbnum].mvs[0];
2869 :    
2870 :     if (!pMBs[mbnum].mcsel)
2871 :     continue;
2872 :    
2873 :     if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx )
2874 :     || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) )
2875 :     pMBs[mbnum].mcsel=0;
2876 :     else
2877 :     num++;
2878 :     }
2879 :    
2880 :     } while ( (oldnum != num) && (num>= minblocks) );
2881 :    
2882 :     if (num < minblocks)
2883 :     {
2884 :     const int iEdgedWidth = pParam->edged_width;
2885 :     num = 0;
2886 :    
2887 :     /* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw);
2888 : syskin 1084 */
2889 : chl 1077 gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0;
2890 :    
2891 : edgomez 1107 if (!(current->motion_flags & XVID_ME_GME_REFINE))
2892 : chl 1077 return gmc;
2893 : syskin 1084
2894 :     for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */
2895 : chl 1077 for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */
2896 :     {
2897 :     const int mbnum = mx + my * MBw;
2898 :     MACROBLOCK *const pMB = &pMBs[mbnum];
2899 :     const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx);
2900 :     if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx )
2901 :     && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) )
2902 : syskin 1084 { pMB->mcsel = 1;
2903 : chl 1077 gmc.duv[0].x += pMB->mvs[0].x;
2904 :     gmc.duv[0].y += pMB->mvs[0].y;
2905 :     num++;
2906 :     }
2907 :     }
2908 : syskin 1084
2909 : chl 1077 if (gmc.duv[0].x)
2910 :     gmc.duv[0].x /= num;
2911 : syskin 1084 if (gmc.duv[0].y)
2912 : chl 1077 gmc.duv[0].y /= num;
2913 :     } else {
2914 :    
2915 :     gmc.duv[0].x=(int)(sol[0]+0.5);
2916 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2917 :    
2918 :     gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2919 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2920 :    
2921 :     gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */
2922 :     gmc.duv[2].y=gmc.duv[1].x;
2923 :     }
2924 :     if (num>maxblocks)
2925 :     { for (my = 1; my < (uint32_t)MBh-1; my++)
2926 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2927 :     {
2928 :     const int mbnum = mx + my * MBw;
2929 :     if (pMBs[mbnum-1].mcsel)
2930 :     pMBs[mbnum].mcsel=0;
2931 : syskin 1084 else
2932 : chl 1077 if (pMBs[mbnum-MBw].mcsel)
2933 :     pMBs[mbnum].mcsel=0;
2934 :     }
2935 :     }
2936 :     return gmc;
2937 :     }
2938 :    
2939 :     int
2940 :     GlobalMotionEstRefine(
2941 :     WARPPOINTS *const startwp,
2942 :     MACROBLOCK * const pMBs,
2943 :     const MBParam * const pParam,
2944 :     const FRAMEINFO * const current,
2945 :     const FRAMEINFO * const reference,
2946 :     const IMAGE * const pCurr,
2947 :     const IMAGE * const pRef,
2948 :     const IMAGE * const pRefH,
2949 :     const IMAGE * const pRefV,
2950 :     const IMAGE * const pRefHV)
2951 :     {
2952 :     uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width);
2953 :     WARPPOINTS bestwp=*startwp;
2954 :     WARPPOINTS centerwp,currwp;
2955 :     int gmcminSAD=0;
2956 :     int gmcSAD=0;
2957 :     int direction;
2958 :     // int mx,my;
2959 :    
2960 :     /* use many blocks... */
2961 :     /* for (my = 0; my < (uint32_t)pParam->mb_height; my++)
2962 :     for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++)
2963 :     {
2964 :     const int mbnum = mx + my * pParam->mb_width;
2965 : syskin 1084 pMBs[mbnum].mcsel=1;
2966 : chl 1077 }
2967 :     */
2968 :    
2969 :     /* or rather don't use too many blocks... */
2970 :     /*
2971 :     for (my = 1; my < (uint32_t)MBh-1; my++)
2972 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2973 :     {
2974 :     const int mbnum = mx + my * MBw;
2975 :     if (MBmask[mbnum-1])
2976 :     MBmask[mbnum-1]=0;
2977 : syskin 1084 else
2978 : chl 1077 if (MBmask[mbnum-MBw])
2979 :     MBmask[mbnum-1]=0;
2980 : syskin 1084
2981 : chl 1077 }
2982 :     */
2983 :     gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
2984 :    
2985 : syskin 1084 if ( (reference->coding_type == S_VOP)
2986 :     && ( (reference->warp.duv[1].x != bestwp.duv[1].x)
2987 :     || (reference->warp.duv[1].y != bestwp.duv[1].y)
2988 :     || (reference->warp.duv[0].x != bestwp.duv[0].x)
2989 :     || (reference->warp.duv[0].y != bestwp.duv[0].y)
2990 :     || (reference->warp.duv[2].x != bestwp.duv[2].x)
2991 :     || (reference->warp.duv[2].y != bestwp.duv[2].y) ) )
2992 : chl 1077 {
2993 : syskin 1084 gmcSAD = globalSAD(&reference->warp, pParam, pMBs,
2994 : chl 1077 current, pRef, pCurr, GMCblock);
2995 : syskin 1084
2996 : chl 1077 if (gmcSAD < gmcminSAD)
2997 :     { bestwp = reference->warp;
2998 :     gmcminSAD = gmcSAD;
2999 :     }
3000 :     }
3001 :    
3002 :     do {
3003 :     direction = 0;
3004 :     centerwp = bestwp;
3005 :    
3006 : syskin 1084 currwp = centerwp;
3007 :    
3008 : chl 1077 currwp.duv[0].x--;
3009 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3010 :     if (gmcSAD < gmcminSAD)
3011 :     { bestwp = currwp;
3012 :     gmcminSAD = gmcSAD;
3013 : syskin 1084 direction = 1;
3014 : chl 1077 }
3015 : syskin 1084 else
3016 :     {
3017 : chl 1077 currwp = centerwp; currwp.duv[0].x++;
3018 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3019 :     if (gmcSAD < gmcminSAD)
3020 :     { bestwp = currwp;
3021 :     gmcminSAD = gmcSAD;
3022 : syskin 1084 direction = 2;
3023 : chl 1077 }
3024 :     }
3025 :     if (direction) continue;
3026 :    
3027 :     currwp = centerwp; currwp.duv[0].y--;
3028 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3029 :     if (gmcSAD < gmcminSAD)
3030 :     { bestwp = currwp;
3031 :     gmcminSAD = gmcSAD;
3032 : syskin 1084 direction = 4;
3033 :     }
3034 :     else
3035 : chl 1077 {
3036 :     currwp = centerwp; currwp.duv[0].y++;
3037 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3038 :     if (gmcSAD < gmcminSAD)
3039 :     { bestwp = currwp;
3040 :     gmcminSAD = gmcSAD;
3041 : syskin 1084 direction = 8;
3042 : chl 1077 }
3043 : syskin 1084 }
3044 : chl 1077 if (direction) continue;
3045 : syskin 1084
3046 : chl 1077 currwp = centerwp; currwp.duv[1].x++;
3047 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3048 :     if (gmcSAD < gmcminSAD)
3049 :     { bestwp = currwp;
3050 :     gmcminSAD = gmcSAD;
3051 : syskin 1084 direction = 32;
3052 :     }
3053 : chl 1077 currwp.duv[2].y++;
3054 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3055 :     if (gmcSAD < gmcminSAD)
3056 :     { bestwp = currwp;
3057 :     gmcminSAD = gmcSAD;
3058 : syskin 1084 direction = 1024;
3059 :     }
3060 : chl 1077
3061 :     currwp = centerwp; currwp.duv[1].x--;
3062 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3063 :     if (gmcSAD < gmcminSAD)
3064 :     { bestwp = currwp;
3065 :     gmcminSAD = gmcSAD;
3066 : syskin 1084 direction = 16;
3067 :     }
3068 :     else
3069 : chl 1077 {
3070 :     currwp = centerwp; currwp.duv[1].x++;
3071 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3072 :     if (gmcSAD < gmcminSAD)
3073 :     { bestwp = currwp;
3074 :     gmcminSAD = gmcSAD;
3075 : syskin 1084 direction = 32;
3076 : chl 1077 }
3077 : syskin 1084 }
3078 : chl 1077 if (direction) continue;
3079 :    
3080 :    
3081 :     currwp = centerwp; currwp.duv[1].y--;
3082 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3083 :     if (gmcSAD < gmcminSAD)
3084 :     { bestwp = currwp;
3085 :     gmcminSAD = gmcSAD;
3086 : syskin 1084 direction = 64;
3087 :     }
3088 :     else
3089 : chl 1077 {
3090 :     currwp = centerwp; currwp.duv[1].y++;
3091 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3092 :     if (gmcSAD < gmcminSAD)
3093 :     { bestwp = currwp;
3094 :     gmcminSAD = gmcSAD;
3095 : syskin 1084 direction = 128;
3096 : chl 1077 }
3097 :     }
3098 :     if (direction) continue;
3099 : syskin 1084
3100 : chl 1077 currwp = centerwp; currwp.duv[2].x--;
3101 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3102 :     if (gmcSAD < gmcminSAD)
3103 :     { bestwp = currwp;
3104 :     gmcminSAD = gmcSAD;
3105 : syskin 1084 direction = 256;
3106 :     }
3107 :     else
3108 : chl 1077 {
3109 :     currwp = centerwp; currwp.duv[2].x++;
3110 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3111 :     if (gmcSAD < gmcminSAD)
3112 :     { bestwp = currwp;
3113 :     gmcminSAD = gmcSAD;
3114 : syskin 1084 direction = 512;
3115 : chl 1077 }
3116 :     }
3117 :     if (direction) continue;
3118 : syskin 1084
3119 : chl 1077 currwp = centerwp; currwp.duv[2].y--;
3120 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3121 :     if (gmcSAD < gmcminSAD)
3122 :     { bestwp = currwp;
3123 :     gmcminSAD = gmcSAD;
3124 : syskin 1084 direction = 1024;
3125 :     }
3126 : chl 1077 else
3127 : syskin 1084 {
3128 : chl 1077 currwp = centerwp; currwp.duv[2].y++;
3129 :     gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock);
3130 :     if (gmcSAD < gmcminSAD)
3131 :     { bestwp = currwp;
3132 :     gmcminSAD = gmcSAD;
3133 : syskin 1084 direction = 2048;
3134 : chl 1077 }
3135 :     }
3136 :     } while (direction);
3137 :     free(GMCblock);
3138 : syskin 1084
3139 : chl 1077 *startwp = bestwp;
3140 :    
3141 :     return gmcminSAD;
3142 :     }
3143 :    
3144 :     int
3145 : syskin 1084 globalSAD(const WARPPOINTS *const wp,
3146 :     const MBParam * const pParam,
3147 : chl 1077 const MACROBLOCK * const pMBs,
3148 :     const FRAMEINFO * const current,
3149 :     const IMAGE * const pRef,
3150 :     const IMAGE * const pCurr,
3151 :     uint8_t *const GMCblock)
3152 :     {
3153 :     NEW_GMC_DATA gmc_data;
3154 :     int iSAD, gmcSAD=0;
3155 :     int num=0;
3156 :     unsigned int mx, my;
3157 : syskin 1084
3158 : chl 1077 generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data);
3159 : syskin 1084
3160 : chl 1077 for (my = 0; my < (uint32_t)pParam->mb_height; my++)
3161 :     for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) {
3162 :    
3163 :     const int mbnum = mx + my * pParam->mb_width;
3164 :     const int iEdgedWidth = pParam->edged_width;
3165 : syskin 1084
3166 : chl 1077 if (!pMBs[mbnum].mcsel)
3167 :     continue;
3168 : syskin 1084
3169 :     gmc_data.predict_16x16(&gmc_data, GMCblock,
3170 : chl 1077 pRef->y,
3171 : syskin 1084 iEdgedWidth,
3172 :     iEdgedWidth,
3173 :     mx, my,
3174 : chl 1077 pParam->m_rounding_type);
3175 :    
3176 :     iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx),
3177 : syskin 1084 GMCblock , iEdgedWidth, 65536);
3178 : chl 1077 iSAD -= pMBs[mbnum].sad16;
3179 : syskin 1084
3180 :     if (iSAD<0)
3181 : chl 1077 gmcSAD += iSAD;
3182 :     num++;
3183 :     }
3184 :     return gmcSAD;
3185 :     }
3186 :    

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