[svn] / branches / Isibaar / xvidcore / src / motion / motion_est.c Repository:
ViewVC logotype

Annotation of /branches/Isibaar/xvidcore/src/motion/motion_est.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 1004 - (view) (download)

1 : edgomez 851 /**************************************************************************
2 : Isibaar 3 *
3 : edgomez 851 * XVID MPEG-4 VIDEO CODEC
4 :     * motion estimation
5 : chl 259 *
6 : edgomez 851 * This program is an implementation of a part of one or more MPEG-4
7 :     * Video tools as specified in ISO/IEC 14496-2 standard. Those intending
8 :     * to use this software module in hardware or software products are
9 :     * advised that its use may infringe existing patents or copyrights, and
10 :     * any such use would be at such party's own risk. The original
11 :     * developer of this software module and his/her company, and subsequent
12 :     * editors and their companies, will have no liability for use of this
13 :     * software or modifications or derivatives thereof.
14 : chl 259 *
15 : edgomez 851 * This program is free software; you can redistribute it and/or modify
16 :     * it under the terms of the GNU General Public License as published by
17 :     * the Free Software Foundation; either version 2 of the License, or
18 :     * (at your option) any later version.
19 : chl 259 *
20 : edgomez 851 * This program is distributed in the hope that it will be useful,
21 :     * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 :     * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 :     * GNU General Public License for more details.
24 : chl 259 *
25 : edgomez 851 * You should have received a copy of the GNU General Public License
26 :     * along with this program; if not, write to the Free Software
27 :     * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 : chl 259 *
29 :     *************************************************************************/
30 :    
31 : Isibaar 3 #include <assert.h>
32 :     #include <stdio.h>
33 : chl 96 #include <stdlib.h>
34 : edgomez 851 #include <string.h> // memcpy
35 :     #include <math.h> // lrint
36 : Isibaar 3
37 :     #include "../encoder.h"
38 :     #include "../utils/mbfunctions.h"
39 :     #include "../prediction/mbprediction.h"
40 :     #include "../global.h"
41 :     #include "../utils/timer.h"
42 : edgomez 851 #include "../image/interpolate8x8.h"
43 :     #include "motion_est.h"
44 : suxen_drol 118 #include "motion.h"
45 : Isibaar 3 #include "sad.h"
46 : edgomez 851 #include "../utils/emms.h"
47 :     #include "../dct/fdct.h"
48 : Isibaar 3
49 : edgomez 872 /*****************************************************************************
50 :     * Modified rounding tables -- declared in motion.h
51 :     * Original tables see ISO spec tables 7-6 -> 7-9
52 :     ****************************************************************************/
53 :    
54 :     const uint32_t roundtab[16] =
55 :     {0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 };
56 :    
57 :     /* K = 4 */
58 :     const uint32_t roundtab_76[16] =
59 :     { 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 };
60 :    
61 :     /* K = 2 */
62 :     const uint32_t roundtab_78[8] =
63 :     { 0, 0, 1, 1, 0, 0, 0, 1 };
64 :    
65 :     /* K = 1 */
66 :     const uint32_t roundtab_79[4] =
67 :     { 0, 1, 0, 0 };
68 :    
69 : edgomez 851 #define INITIAL_SKIP_THRESH (10)
70 :     #define FINAL_SKIP_THRESH (50)
71 :     #define MAX_SAD00_FOR_SKIP (20)
72 :     #define MAX_CHROMA_SAD_FOR_SKIP (22)
73 : Isibaar 3
74 : edgomez 851 #define CHECK_CANDIDATE(X,Y,D) { \
75 :     CheckCandidate((X),(Y), (D), &iDirection, data ); }
76 : Isibaar 3
77 : edgomez 872 /*****************************************************************************
78 :     * Code
79 :     ****************************************************************************/
80 :    
81 : edgomez 851 static __inline uint32_t
82 :     d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
83 :     {
84 : syskin 973 int bits;
85 :     const int q = (1 << (iFcode - 1)) - 1;
86 :    
87 :     x <<= qpel;
88 :     y <<= qpel;
89 : edgomez 851 if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
90 : chl 141
91 : edgomez 851 x -= pred.x;
92 : syskin 973 bits = (x != 0 ? iFcode:0);
93 :     x = abs(x);
94 :     x += q;
95 :     x >>= (iFcode - 1);
96 :     bits += mvtab[x];
97 :    
98 : edgomez 851 y -= pred.y;
99 : syskin 973 bits += (y != 0 ? iFcode:0);
100 :     y = abs(y);
101 :     y += q;
102 :     y >>= (iFcode - 1);
103 :     bits += mvtab[y];
104 : chl 141
105 : syskin 973 return bits;
106 : edgomez 851 }
107 : chl 141
108 : syskin 976 static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by,
109 :     const SearchData * const data)
110 : edgomez 851 {
111 :     int sad;
112 :     const uint32_t stride = data->iEdgedWidth/2;
113 :     uint8_t * f_refu = data->RefQ,
114 :     * f_refv = data->RefQ + 8,
115 :     * b_refu = data->RefQ + 16,
116 :     * b_refv = data->RefQ + 24;
117 : syskin 976 int offset = (fx>>1) + (fy>>1)*stride;
118 : Isibaar 3
119 : edgomez 851 switch (((fx & 1) << 1) | (fy & 1)) {
120 :     case 0:
121 : syskin 976 f_refu = (uint8_t*)data->RefP[4] + offset;
122 :     f_refv = (uint8_t*)data->RefP[5] + offset;
123 : edgomez 851 break;
124 :     case 1:
125 : syskin 976 interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding);
126 :     interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding);
127 : edgomez 851 break;
128 :     case 2:
129 : syskin 976 interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding);
130 :     interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding);
131 : edgomez 851 break;
132 :     default:
133 : syskin 976 interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding);
134 :     interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding);
135 : edgomez 851 break;
136 :     }
137 : Isibaar 3
138 : syskin 976 offset = (bx>>1) + (by>>1)*stride;
139 : edgomez 851 switch (((bx & 1) << 1) | (by & 1)) {
140 :     case 0:
141 : syskin 976 b_refu = (uint8_t*)data->b_RefP[4] + offset;
142 :     b_refv = (uint8_t*)data->b_RefP[5] + offset;
143 : edgomez 851 break;
144 :     case 1:
145 : syskin 976 interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
146 :     interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
147 : edgomez 851 break;
148 :     case 2:
149 : syskin 976 interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
150 :     interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
151 : edgomez 851 break;
152 :     default:
153 : syskin 976 interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
154 :     interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
155 : edgomez 851 break;
156 :     }
157 : Isibaar 3
158 : edgomez 851 sad = sad8bi(data->CurU, b_refu, f_refu, stride);
159 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
160 : Isibaar 3
161 : edgomez 851 return sad;
162 :     }
163 : Isibaar 3
164 : edgomez 851 static int32_t
165 : syskin 976 ChromaSAD(const int dx, const int dy, const SearchData * const data)
166 : edgomez 851 {
167 :     int sad;
168 :     const uint32_t stride = data->iEdgedWidth/2;
169 : syskin 976 int offset = (dx>>1) + (dy>>1)*stride;
170 : Isibaar 3
171 : edgomez 851 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently
172 :     data->temp[5] = dx; data->temp[6] = dy; // backup
173 : Isibaar 3
174 : edgomez 851 switch (((dx & 1) << 1) | (dy & 1)) {
175 :     case 0:
176 : syskin 976 sad = sad8(data->CurU, data->RefP[4] + offset, stride);
177 :     sad += sad8(data->CurV, data->RefP[5] + offset, stride);
178 : edgomez 851 break;
179 :     case 1:
180 : syskin 976 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride);
181 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride);
182 : edgomez 851 break;
183 :     case 2:
184 : syskin 976 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride);
185 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride);
186 : edgomez 851 break;
187 :     default:
188 : syskin 976 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding);
189 : edgomez 851 sad = sad8(data->CurU, data->RefQ, stride);
190 :    
191 : syskin 976 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding);
192 : edgomez 851 sad += sad8(data->CurV, data->RefQ, stride);
193 :     break;
194 :     }
195 :     data->temp[7] = sad; //backup, part 2
196 :     return sad;
197 : Isibaar 3 }
198 :    
199 : edgomez 851 static __inline const uint8_t *
200 :     GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
201 : Isibaar 3 {
202 : edgomez 851 // dir : 0 = forward, 1 = backward
203 : syskin 985 const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP );
204 : syskin 976 const int picture = ((x&1)<<1) | (y&1);
205 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
206 :     return direction[picture] + offset;
207 : Isibaar 3 }
208 :    
209 : edgomez 851 // this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate
210 :     static __inline const uint8_t *
211 :     GetReference(const int x, const int y, const SearchData * const data)
212 : Isibaar 3 {
213 : syskin 976 const int picture = ((x&1)<<1) | (y&1);
214 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
215 :     return data->RefP[picture] + offset;
216 : Isibaar 3 }
217 : edgomez 851
218 :     static uint8_t *
219 :     Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
220 : Isibaar 3 {
221 : edgomez 851 // create or find a qpel-precision reference picture; return pointer to it
222 :     uint8_t * Reference = data->RefQ + 16*dir;
223 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
224 :     const uint32_t rounding = data->rounding;
225 :     const int halfpel_x = x/2;
226 :     const int halfpel_y = y/2;
227 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
228 : suxen_drol 136
229 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
230 :     ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
231 :     switch( ((x&1)<<1) + (y&1) ) {
232 : edgomez 963 case 3: // x and y in qpel resolution - the "corners" (top left/right and
233 :     // bottom left/right) during qpel refinement
234 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
235 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
236 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
237 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
238 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
239 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
240 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
241 : edgomez 851 break;
242 : chl 184
243 : edgomez 851 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
244 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
245 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
246 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
247 :     break;
248 : suxen_drol 136
249 : edgomez 851 case 2: // x qpel, y halfpel - left or right during qpel refinement
250 :     ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
251 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
252 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
253 :     break;
254 : edgomez 195
255 : edgomez 963 default: // pure halfpel position
256 :     return (uint8_t *) ref1;
257 : syskin 976
258 : edgomez 851 }
259 :     return Reference;
260 :     }
261 : Isibaar 3
262 : edgomez 851 static uint8_t *
263 :     Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
264 :     {
265 :     // create or find a qpel-precision reference picture; return pointer to it
266 :     uint8_t * Reference = data->RefQ + 16*dir;
267 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
268 :     const uint32_t rounding = data->rounding;
269 :     const int halfpel_x = x/2;
270 :     const int halfpel_y = y/2;
271 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
272 : chl 326
273 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
274 :     switch( ((x&1)<<1) + (y&1) ) {
275 :     case 3: // x and y in qpel resolution - the "corners" (top left/right and
276 : syskin 976 // bottom left/right) during qpel refinement
277 : edgomez 851 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
278 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
279 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
280 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
281 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
282 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
283 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
284 :     break;
285 : chl 184
286 : edgomez 851 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
287 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
288 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
289 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
290 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
291 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
292 :     break;
293 : chl 184
294 : edgomez 851 case 2: // x qpel, y halfpel - left or right during qpel refinement
295 :     ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
296 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
297 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
298 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
299 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
300 :     break;
301 : edgomez 195
302 : edgomez 963 default: // pure halfpel position
303 : edgomez 851 return (uint8_t *) ref1;
304 :     }
305 :     return Reference;
306 :     }
307 : edgomez 195
308 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS START */
309 : edgomez 195
310 : edgomez 851 static void
311 :     CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
312 :     {
313 :     int xc, yc;
314 :     const uint8_t * Reference;
315 :     VECTOR * current;
316 :     int32_t sad; uint32_t t;
317 : edgomez 195
318 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
319 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
320 : edgomez 195
321 : edgomez 851 if (!data->qpel_precision) {
322 :     Reference = GetReference(x, y, data);
323 :     current = data->currentMV;
324 :     xc = x; yc = y;
325 :     } else { // x and y are in 1/4 precision
326 :     Reference = Interpolate16x16qpel(x, y, 0, data);
327 :     xc = x/2; yc = y/2; //for chroma sad
328 :     current = data->currentQMV;
329 :     }
330 : edgomez 195
331 : edgomez 851 sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
332 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
333 : edgomez 195
334 : edgomez 851 sad += (data->lambda16 * t * sad)>>10;
335 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
336 : chl 175
337 : syskin 976 if (data->chroma) sad += ChromaSAD( (xc >> 1) + roundtab_79[xc & 0x3],
338 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
339 : edgomez 195
340 : edgomez 851 if (sad < data->iMinSAD[0]) {
341 :     data->iMinSAD[0] = sad;
342 :     current[0].x = x; current[0].y = y;
343 :     *dir = Direction;
344 :     }
345 : edgomez 195
346 : edgomez 851 if (data->temp[1] < data->iMinSAD[1]) {
347 :     data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
348 :     if (data->temp[2] < data->iMinSAD[2]) {
349 :     data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
350 :     if (data->temp[3] < data->iMinSAD[3]) {
351 :     data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
352 :     if (data->temp[4] < data->iMinSAD[4]) {
353 :     data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
354 : Isibaar 3 }
355 :    
356 : edgomez 851 static void
357 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
358 :     {
359 :     int32_t sad; uint32_t t;
360 :     const uint8_t * Reference;
361 : syskin 973 VECTOR * current;
362 : chl 326
363 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
364 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
365 : Isibaar 3
366 : syskin 973 if (!data->qpel_precision) {
367 :     Reference = GetReference(x, y, data);
368 :     current = data->currentMV;
369 :     } else { // x and y are in 1/4 precision
370 :     Reference = Interpolate8x8qpel(x, y, 0, 0, data);
371 :     current = data->currentQMV;
372 :     }
373 : Isibaar 3
374 : edgomez 851 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
375 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
376 : Isibaar 3
377 : edgomez 851 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
378 : Isibaar 3
379 : edgomez 851 if (sad < *(data->iMinSAD)) {
380 :     *(data->iMinSAD) = sad;
381 : syskin 973 current->x = x; current->y = y;
382 : edgomez 851 *dir = Direction;
383 :     }
384 : Isibaar 3 }
385 :    
386 : edgomez 851 static void
387 :     CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
388 :     {
389 :     uint32_t t;
390 :     const uint8_t * Reference;
391 :    
392 : edgomez 963 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero even value
393 : edgomez 851 (x > data->max_dx) || (x < data->min_dx)
394 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
395 :    
396 :     Reference = GetReference(x, y, data);
397 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
398 :    
399 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
400 :    
401 :     data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
402 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
403 :    
404 :     if (data->temp[0] < data->iMinSAD[0]) {
405 :     data->iMinSAD[0] = data->temp[0];
406 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
407 :     *dir = Direction; }
408 :    
409 :     if (data->temp[1] < data->iMinSAD[1]) {
410 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
411 :     if (data->temp[2] < data->iMinSAD[2]) {
412 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
413 :     if (data->temp[3] < data->iMinSAD[3]) {
414 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
415 :     if (data->temp[4] < data->iMinSAD[4]) {
416 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
417 : Isibaar 3 }
418 : edgomez 195
419 : edgomez 851 static void
420 :     CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
421 :     {
422 :     int32_t sad, xc, yc;
423 :     const uint8_t * Reference;
424 :     uint32_t t;
425 :     VECTOR * current;
426 :    
427 : edgomez 963 if ( (x > data->max_dx) || ( x < data->min_dx)
428 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
429 : edgomez 851
430 :     if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value
431 :    
432 :     if (data->qpel_precision) { // x and y are in 1/4 precision
433 :     Reference = Interpolate16x16qpel(x, y, 0, data);
434 :     current = data->currentQMV;
435 :     xc = x/2; yc = y/2;
436 :     } else {
437 :     Reference = GetReference(x, y, data);
438 :     current = data->currentMV;
439 :     xc = x; yc = y;
440 :     }
441 :     t = d_mv_bits(x, y, data->predMV, data->iFcode,
442 :     data->qpel^data->qpel_precision, data->rrv);
443 :    
444 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
445 :     sad += (data->lambda16 * t * sad)>>10;
446 :    
447 :     if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
448 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
449 :    
450 :     if (sad < *(data->iMinSAD)) {
451 :     *(data->iMinSAD) = sad;
452 :     current->x = x; current->y = y;
453 :     *dir = Direction;
454 :     }
455 : chl 96 }
456 : Isibaar 3
457 : edgomez 851 static void
458 :     CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
459 :     {
460 :     // maximum speed - for P/B/I decision
461 :     int32_t sad;
462 :    
463 :     if ( (x > data->max_dx) || (x < data->min_dx)
464 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
465 :    
466 : syskin 985 sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*(data->iEdgedWidth),
467 :     data->iEdgedWidth, data->temp+1);
468 : edgomez 851
469 :     if (sad < *(data->iMinSAD)) {
470 :     *(data->iMinSAD) = sad;
471 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
472 :     *dir = Direction;
473 :     }
474 :     if (data->temp[1] < data->iMinSAD[1]) {
475 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
476 :     if (data->temp[2] < data->iMinSAD[2]) {
477 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
478 :     if (data->temp[3] < data->iMinSAD[3]) {
479 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
480 :     if (data->temp[4] < data->iMinSAD[4]) {
481 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
482 :    
483 : Isibaar 3 }
484 :    
485 : edgomez 851 static void
486 :     CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
487 :     {
488 :     int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
489 :     uint32_t t;
490 :     const uint8_t *ReferenceF, *ReferenceB;
491 :     VECTOR *current;
492 :    
493 : edgomez 963 if ((xf > data->max_dx) || (xf < data->min_dx) ||
494 :     (yf > data->max_dy) || (yf < data->min_dy))
495 :     return;
496 : edgomez 851
497 :     if (!data->qpel_precision) {
498 :     ReferenceF = GetReference(xf, yf, data);
499 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
500 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
501 :     current = data->currentMV;
502 :     xcf = xf; ycf = yf;
503 :     xcb = xb; ycb = yb;
504 :     } else {
505 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
506 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
507 :     current = data->currentQMV;
508 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
509 :     xcf = xf/2; ycf = yf/2;
510 :     xcb = xb/2; ycb = yb/2;
511 :     }
512 :    
513 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
514 : syskin 976 + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
515 : edgomez 851
516 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
517 :     sad += (data->lambda16 * t * sad)>>10;
518 :    
519 :     if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
520 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
521 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
522 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
523 :    
524 :     if (sad < *(data->iMinSAD)) {
525 :     *(data->iMinSAD) = sad;
526 :     current->x = xf; current->y = yf;
527 :     *dir = Direction;
528 :     }
529 : Isibaar 3 }
530 :    
531 : edgomez 851 static void
532 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
533 :     {
534 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
535 :     uint32_t k;
536 :     const uint8_t *ReferenceF;
537 :     const uint8_t *ReferenceB;
538 :     VECTOR mvs, b_mvs;
539 :    
540 : syskin 904 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
541 : edgomez 851
542 :     for (k = 0; k < 4; k++) {
543 :     mvs.x = data->directmvF[k].x + x;
544 :     b_mvs.x = ((x == 0) ?
545 :     data->directmvB[k].x
546 :     : mvs.x - data->referencemv[k].x);
547 :    
548 :     mvs.y = data->directmvF[k].y + y;
549 :     b_mvs.y = ((y == 0) ?
550 :     data->directmvB[k].y
551 :     : mvs.y - data->referencemv[k].y);
552 :    
553 : edgomez 963 if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) ||
554 :     (mvs.y > data->max_dy) || (mvs.y < data->min_dy) ||
555 :     (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) ||
556 :     (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) )
557 :     return;
558 : edgomez 851
559 :     if (data->qpel) {
560 :     xcf += mvs.x/2; ycf += mvs.y/2;
561 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
562 :     } else {
563 :     xcf += mvs.x; ycf += mvs.y;
564 :     xcb += b_mvs.x; ycb += b_mvs.y;
565 :     mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway
566 :     b_mvs.x *= 2; b_mvs.y *= 2;
567 :     }
568 :    
569 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
570 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
571 :    
572 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
573 :     ReferenceF, ReferenceB, data->iEdgedWidth);
574 :     if (sad > *(data->iMinSAD)) return;
575 :     }
576 :    
577 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
578 :    
579 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
580 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
581 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
582 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
583 :    
584 :     if (sad < *(data->iMinSAD)) {
585 :     *(data->iMinSAD) = sad;
586 :     data->currentMV->x = x; data->currentMV->y = y;
587 :     *dir = Direction;
588 :     }
589 : Isibaar 3 }
590 :    
591 : edgomez 851 static void
592 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
593 : Isibaar 3 {
594 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
595 :     const uint8_t *ReferenceF;
596 :     const uint8_t *ReferenceB;
597 :     VECTOR mvs, b_mvs;
598 : Isibaar 3
599 : edgomez 963 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
600 : Isibaar 3
601 : edgomez 851 mvs.x = data->directmvF[0].x + x;
602 :     b_mvs.x = ((x == 0) ?
603 :     data->directmvB[0].x
604 :     : mvs.x - data->referencemv[0].x);
605 : Isibaar 3
606 : edgomez 851 mvs.y = data->directmvF[0].y + y;
607 :     b_mvs.y = ((y == 0) ?
608 :     data->directmvB[0].y
609 :     : mvs.y - data->referencemv[0].y);
610 : Isibaar 3
611 : edgomez 963 if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx)
612 :     || (mvs.y > data->max_dy) || (mvs.y < data->min_dy)
613 :     || (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx)
614 :     || (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return;
615 : Isibaar 3
616 : edgomez 851 if (data->qpel) {
617 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
618 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
619 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
620 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
621 :     } else {
622 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
623 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
624 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
625 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
626 :     }
627 :    
628 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
629 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
630 :    
631 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
632 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
633 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
634 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
635 :    
636 :     if (sad < *(data->iMinSAD)) {
637 :     *(data->iMinSAD) = sad;
638 :     data->currentMV->x = x; data->currentMV->y = y;
639 :     *dir = Direction;
640 :     }
641 : Isibaar 3 }
642 :    
643 : edgomez 851
644 :     static void
645 :     CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
646 : Isibaar 3 {
647 :    
648 : edgomez 961 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
649 : edgomez 851 int32_t bits = 0, sum;
650 :     VECTOR * current;
651 :     const uint8_t * ptr;
652 :     int i, cbp = 0, t, xc, yc;
653 : edgomez 195
654 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
655 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
656 : edgomez 195
657 : edgomez 851 if (!data->qpel_precision) {
658 :     ptr = GetReference(x, y, data);
659 :     current = data->currentMV;
660 :     xc = x; yc = y;
661 :     } else { // x and y are in 1/4 precision
662 :     ptr = Interpolate16x16qpel(x, y, 0, data);
663 :     current = data->currentQMV;
664 :     xc = x/2; yc = y/2;
665 :     }
666 : Isibaar 3
667 : edgomez 851 for(i = 0; i < 4; i++) {
668 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
669 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
670 :     fdct(in);
671 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
672 :     else sum = quant4_inter(coeff, in, data->lambda16);
673 :     if (sum > 0) {
674 :     cbp |= 1 << (5 - i);
675 :     bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
676 :     } else data->temp[i] = 0;
677 :     }
678 : Isibaar 3
679 : edgomez 851 bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
680 : edgomez 195
681 : edgomez 851 if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma
682 :     xc = (xc >> 1) + roundtab_79[xc & 0x3];
683 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
684 :    
685 :     //chroma U
686 : syskin 976 ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
687 : edgomez 851 transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
688 :     fdct(in);
689 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
690 :     else sum = quant4_inter(coeff, in, data->lambda16);
691 :     if (sum > 0) {
692 :     cbp |= 1 << (5 - 4);
693 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
694 : chl 344 }
695 : edgomez 851
696 :     if (bits < data->iMinSAD[0]) {
697 :     //chroma V
698 : syskin 976 ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
699 : edgomez 851 transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2);
700 :     fdct(in);
701 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
702 :     else sum = quant4_inter(coeff, in, data->lambda16);
703 :     if (sum > 0) {
704 :     cbp |= 1 << (5 - 5);
705 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
706 :     }
707 :     }
708 : edgomez 78 }
709 : edgomez 851
710 : chl 897 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
711 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
712 :    
713 :     if (bits < data->iMinSAD[0]) {
714 :     data->iMinSAD[0] = bits;
715 :     current[0].x = x; current[0].y = y;
716 :     *dir = Direction;
717 :     }
718 :    
719 :     if (data->temp[0] + t < data->iMinSAD[1]) {
720 :     data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
721 :     if (data->temp[1] < data->iMinSAD[2]) {
722 :     data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
723 :     if (data->temp[2] < data->iMinSAD[3]) {
724 :     data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
725 :     if (data->temp[3] < data->iMinSAD[4]) {
726 :     data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
727 :    
728 : Isibaar 3 }
729 : edgomez 851 static void
730 :     CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
731 : chl 96 {
732 :    
733 : edgomez 961 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
734 : edgomez 851 int32_t sum, bits;
735 :     VECTOR * current;
736 :     const uint8_t * ptr;
737 :     int cbp;
738 : edgomez 195
739 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
740 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
741 : edgomez 195
742 : edgomez 851 if (!data->qpel_precision) {
743 :     ptr = GetReference(x, y, data);
744 :     current = data->currentMV;
745 :     } else { // x and y are in 1/4 precision
746 :     ptr = Interpolate8x8qpel(x, y, 0, 0, data);
747 :     current = data->currentQMV;
748 :     }
749 : chl 96
750 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
751 :     fdct(in);
752 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
753 :     else sum = quant4_inter(coeff, in, data->lambda16);
754 :     if (sum > 0) {
755 :     bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
756 :     cbp = 1;
757 :     } else cbp = bits = 0;
758 : chl 96
759 : edgomez 851 bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
760 : chl 96
761 : edgomez 851 if (bits < data->iMinSAD[0]) {
762 :     data->temp[0] = cbp;
763 :     data->iMinSAD[0] = bits;
764 :     current[0].x = x; current[0].y = y;
765 :     *dir = Direction;
766 :     }
767 :     }
768 : chl 96
769 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
770 : edgomez 195
771 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
772 : chl 96
773 : edgomez 851 static void
774 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
775 :     {
776 : edgomez 195
777 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
778 : edgomez 195
779 : edgomez 851 int iDirection;
780 : edgomez 195
781 : edgomez 851 for(;;) { //forever
782 :     iDirection = 0;
783 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
784 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
785 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
786 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
787 : edgomez 195
788 : edgomez 851 /* now we're doing diagonal checks near our candidate */
789 : edgomez 195
790 : edgomez 851 if (iDirection) { //if anything found
791 :     bDirection = iDirection;
792 :     iDirection = 0;
793 :     x = data->currentMV->x; y = data->currentMV->y;
794 :     if (bDirection & 3) { //our candidate is left or right
795 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
796 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
797 :     } else { // what remains here is up or down
798 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
799 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
800 :     }
801 : edgomez 195
802 : edgomez 851 if (iDirection) {
803 :     bDirection += iDirection;
804 :     x = data->currentMV->x; y = data->currentMV->y;
805 :     }
806 :     } else { //about to quit, eh? not so fast....
807 :     switch (bDirection) {
808 :     case 2:
809 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
810 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
811 : edgomez 195 break;
812 : edgomez 851 case 1:
813 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
814 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
815 : edgomez 195 break;
816 : edgomez 851 case 2 + 4:
817 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
818 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
819 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
820 :     break;
821 :     case 4:
822 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
823 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
824 :     break;
825 : edgomez 195 case 8:
826 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
827 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
828 : edgomez 195 break;
829 : edgomez 851 case 1 + 4:
830 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
831 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
832 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
833 : edgomez 195 break;
834 : edgomez 851 case 2 + 8:
835 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
836 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
837 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
838 :     break;
839 :     case 1 + 8:
840 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
841 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
842 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
843 :     break;
844 :     default: //1+2+4+8 == we didn't find anything at all
845 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
846 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
847 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
848 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
849 :     break;
850 : chl 96 }
851 : edgomez 851 if (!iDirection) break; //ok, the end. really
852 :     bDirection = iDirection;
853 :     x = data->currentMV->x; y = data->currentMV->y;
854 : chl 345 }
855 : edgomez 195 }
856 : chl 96 }
857 :    
858 : edgomez 851 static void
859 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
860 : chl 96 {
861 : edgomez 851 int iDirection;
862 : edgomez 195
863 : edgomez 851 do {
864 :     iDirection = 0;
865 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
866 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
867 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
868 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
869 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
870 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
871 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
872 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
873 : chl 96
874 : edgomez 851 bDirection = iDirection;
875 :     x = data->currentMV->x; y = data->currentMV->y;
876 :     } while (iDirection);
877 : chl 96 }
878 :    
879 : edgomez 851 static void
880 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
881 : chl 181 {
882 :    
883 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
884 :    
885 : edgomez 851 int iDirection;
886 : edgomez 195
887 : edgomez 851 do {
888 :     iDirection = 0;
889 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
890 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
891 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
892 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
893 :    
894 :     /* now we're doing diagonal checks near our candidate */
895 :    
896 :     if (iDirection) { //checking if anything found
897 :     bDirection = iDirection;
898 : chl 181 iDirection = 0;
899 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
900 :     if (bDirection & 3) { //our candidate is left or right
901 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
902 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
903 :     } else { // what remains here is up or down
904 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
905 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
906 :     }
907 :     bDirection += iDirection;
908 :     x = data->currentMV->x; y = data->currentMV->y;
909 :     }
910 :     }
911 :     while (iDirection);
912 :     }
913 : chl 181
914 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
915 : chl 181
916 : edgomez 851 static void
917 :     SubpelRefine(const SearchData * const data)
918 :     {
919 :     /* Do a half-pel or q-pel refinement */
920 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
921 :     int iDirection; //only needed because macro expects it
922 : chl 181
923 : edgomez 851 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
924 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
925 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
926 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
927 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
928 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
929 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
930 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
931 :     }
932 : chl 181
933 : edgomez 851 static __inline int
934 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
935 :     const int x, const int y,
936 :     const uint32_t stride, const uint32_t iQuant, int rrv)
937 : chl 181
938 : edgomez 851 {
939 : edgomez 963 int offset = (x + y*stride)*8;
940 : edgomez 851 if(!rrv) {
941 : edgomez 963 uint32_t sadC = sad8(current->u + offset,
942 :     reference->u + offset, stride);
943 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
944 : edgomez 963 sadC += sad8(current->v + offset,
945 :     reference->v + offset, stride);
946 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
947 :     return 1;
948 : chl 181
949 : edgomez 851 } else {
950 : edgomez 963 uint32_t sadC = sad16(current->u + 2*offset,
951 :     reference->u + 2*offset, stride, 256*4096);
952 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
953 : edgomez 963 sadC += sad16(current->v + 2*offset,
954 :     reference->v + 2*offset, stride, 256*4096);
955 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
956 :     return 1;
957 : chl 181 }
958 : edgomez 851 }
959 : edgomez 504
960 : edgomez 851 static __inline void
961 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
962 :     {
963 :     pMB->mode = MODE_NOT_CODED;
964 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
965 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
966 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
967 : chl 181 }
968 :    
969 : Isibaar 1004 static __inline void
970 :     ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad)
971 :     {
972 :     pMB->mode = MODE_INTER;
973 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
974 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
975 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
976 :     }
977 :    
978 : edgomez 851 bool
979 :     MotionEstimation(MBParam * const pParam,
980 : syskin 976 FRAMEINFO * const current,
981 :     FRAMEINFO * const reference,
982 :     const IMAGE * const pRefH,
983 :     const IMAGE * const pRefV,
984 :     const IMAGE * const pRefHV,
985 :     const uint32_t iLimit)
986 : chl 181 {
987 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
988 :     const IMAGE *const pCurrent = &current->image;
989 :     const IMAGE *const pRef = &reference->image;
990 : chl 181
991 : edgomez 851 uint32_t mb_width = pParam->mb_width;
992 :     uint32_t mb_height = pParam->mb_height;
993 :     const uint32_t iEdgedWidth = pParam->edged_width;
994 :     const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags);
995 : chl 181
996 : edgomez 851 uint32_t x, y;
997 :     uint32_t iIntra = 0;
998 :     int32_t quant = current->quant, sad00;
999 : syskin 976 int skip_thresh = INITIAL_SKIP_THRESH *
1000 :     (current->global_flags & XVID_REDUCED ? 4:1) *
1001 : syskin 904 (current->global_flags & XVID_MODEDECISION_BITS ? 2:1);
1002 : chl 181
1003 : edgomez 851 // some pre-initialized thingies for SearchP
1004 :     int32_t temp[8];
1005 :     VECTOR currentMV[5];
1006 :     VECTOR currentQMV[5];
1007 :     int32_t iMinSAD[5];
1008 : edgomez 961 DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE);
1009 : edgomez 851 SearchData Data;
1010 :     memset(&Data, 0, sizeof(SearchData));
1011 :     Data.iEdgedWidth = iEdgedWidth;
1012 :     Data.currentMV = currentMV;
1013 :     Data.currentQMV = currentQMV;
1014 :     Data.iMinSAD = iMinSAD;
1015 :     Data.temp = temp;
1016 :     Data.iFcode = current->fcode;
1017 :     Data.rounding = pParam->m_rounding_type;
1018 :     Data.qpel = pParam->m_quarterpel;
1019 :     Data.chroma = MotionFlags & PMV_CHROMA16;
1020 :     Data.rrv = current->global_flags & XVID_REDUCED;
1021 : edgomez 961 Data.dctSpace = dct_space;
1022 : edgomez 195
1023 : edgomez 851 if ((current->global_flags & XVID_REDUCED)) {
1024 :     mb_width = (pParam->width + 31) / 32;
1025 :     mb_height = (pParam->height + 31) / 32;
1026 :     Data.qpel = 0;
1027 :     }
1028 : chl 181
1029 : edgomez 851 Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose)
1030 :     if (sadInit) (*sadInit) ();
1031 : chl 181
1032 : edgomez 851 for (y = 0; y < mb_height; y++) {
1033 :     for (x = 0; x < mb_width; x++) {
1034 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1035 : chl 181
1036 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1037 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1038 :     pRef->y + (x + y * iEdgedWidth) * 16,
1039 :     pParam->edged_width, pMB->sad8 );
1040 : chl 181
1041 : edgomez 851 else pMB->sad16 =
1042 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1043 :     pRef->y + (x + y * iEdgedWidth) * 32,
1044 :     pParam->edged_width, pMB->sad8 );
1045 : chl 181
1046 : edgomez 851 if (Data.chroma) {
1047 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1048 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1049 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1050 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1051 :     pMB->sad16 += Data.temp[7];
1052 :     }
1053 :    
1054 :     sad00 = pMB->sad16;
1055 :    
1056 :     if (!(current->global_flags & XVID_LUMIMASKING)) {
1057 :     pMB->dquant = NO_CHANGE;
1058 :     } else {
1059 :     if (pMB->dquant != NO_CHANGE) {
1060 :     quant += DQtab[pMB->dquant];
1061 :     if (quant > 31) quant = 31;
1062 :     else if (quant < 1) quant = 1;
1063 : chl 181 }
1064 : edgomez 851 }
1065 :     pMB->quant = current->quant;
1066 : chl 181
1067 : edgomez 851 //initial skip decision
1068 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1069 :     if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1070 : syskin 904 if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * skip_thresh)
1071 : edgomez 851 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1072 :     SkipMacroblockP(pMB, sad00);
1073 :     continue;
1074 :     }
1075 :     }
1076 :    
1077 : Isibaar 1004 if ((current->global_flags & XVID_CARTOON_MODE) &&
1078 :     (sad00 < pMB->quant * 4 * skip_thresh)) { /* favorize (0,0) vector for cartoons */
1079 :     ZeroMacroblockP(pMB, sad00);
1080 :     continue;
1081 :     }
1082 :    
1083 : edgomez 851 SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1084 :     y, MotionFlags, current->global_flags, pMB->quant,
1085 :     &Data, pParam, pMBs, reference->mbs,
1086 :     current->global_flags & XVID_INTER4V, pMB);
1087 :    
1088 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
1089 : syskin 904 if (!(current->global_flags & XVID_GMC || current->global_flags & XVID_MODEDECISION_BITS)) {
1090 : edgomez 851 if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) {
1091 : syskin 904 if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) )
1092 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv))
1093 :     SkipMacroblockP(pMB, sad00);
1094 : chl 181 }
1095 :     }
1096 : edgomez 851 if (pMB->mode == MODE_INTRA)
1097 :     if (++iIntra > iLimit) return 1;
1098 : chl 181 }
1099 :     }
1100 : edgomez 851
1101 :     if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */
1102 :     {
1103 :     current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1104 :     }
1105 :     return 0;
1106 : chl 181 }
1107 :    
1108 :    
1109 : edgomez 851 static __inline int
1110 :     make_mask(const VECTOR * const pmv, const int i)
1111 : chl 96 {
1112 : edgomez 851 int mask = 255, j;
1113 :     for (j = 0; j < i; j++) {
1114 :     if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
1115 :     if (pmv[i].x == pmv[j].x) {
1116 :     if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1117 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1118 :     } else
1119 :     if (pmv[i].y == pmv[j].y) {
1120 :     if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1121 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1122 :     }
1123 :     }
1124 :     return mask;
1125 :     }
1126 : edgomez 195
1127 : edgomez 851 static __inline void
1128 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1129 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1130 :     {
1131 : chl 96
1132 : edgomez 851 //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1133 :     if (rrv) { iWcount /= 2; iHcount /= 2; }
1134 : edgomez 195
1135 : edgomez 851 if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
1136 :     pmv[5].x = EVEN(pmv[3].x);
1137 :     pmv[5].y = EVEN(pmv[3].y);
1138 :     } else pmv[5].x = pmv[5].y = 0;
1139 : chl 96
1140 : edgomez 851 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1141 :     else pmv[3].x = pmv[3].y = 0;
1142 : chl 96
1143 : edgomez 851 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1144 :     else pmv[4].x = pmv[4].y = 0;
1145 : Isibaar 3
1146 : edgomez 851 // [1] median prediction
1147 :     pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1148 : edgomez 195
1149 : edgomez 851 pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask
1150 : edgomez 195
1151 : edgomez 851 pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1152 :     pmv[2].y = EVEN(prevMB->mvs[0].y);
1153 :    
1154 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1155 :     pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1156 :     pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1157 :     } else pmv[6].x = pmv[6].y = 0;
1158 :    
1159 :     if (rrv) {
1160 :     int i;
1161 :     for (i = 0; i < 7; i++) {
1162 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1163 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1164 :     }
1165 :     }
1166 : Isibaar 3 }
1167 :    
1168 : edgomez 851 static int
1169 :     ModeDecision(const uint32_t iQuant, SearchData * const Data,
1170 :     int inter4v,
1171 :     MACROBLOCK * const pMB,
1172 :     const MACROBLOCK * const pMBs,
1173 :     const int x, const int y,
1174 :     const MBParam * const pParam,
1175 :     const uint32_t MotionFlags,
1176 :     const uint32_t GlobalFlags)
1177 :     {
1178 : Isibaar 3
1179 : edgomez 851 int mode = MODE_INTER;
1180 : chl 96
1181 : edgomez 851 if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision
1182 :     int sad;
1183 :     int InterBias = MV16_INTER_BIAS;
1184 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1185 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1186 : edgomez 963 mode = MODE_INTER;
1187 :     sad = Data->iMinSAD[0];
1188 : edgomez 851 } else {
1189 :     mode = MODE_INTER4V;
1190 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1191 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1192 :     Data->iMinSAD[0] = sad;
1193 :     }
1194 : chl 326
1195 : edgomez 851 /* intra decision */
1196 : Isibaar 3
1197 : edgomez 851 if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work
1198 :     if (y != 0)
1199 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1200 :     if (x != 0)
1201 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1202 : Isibaar 3
1203 : edgomez 851 if (Data->chroma) InterBias += 50; // to compensate bigger SAD
1204 :     if (Data->rrv) InterBias *= 4;
1205 : edgomez 195
1206 : edgomez 851 if (InterBias < pMB->sad16) {
1207 :     int32_t deviation;
1208 :     if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
1209 :     else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
1210 :     dev16(Data->Cur+8, Data->iEdgedWidth) +
1211 :     dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) +
1212 :     dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth);
1213 : edgomez 195
1214 : edgomez 963 if (deviation < (sad - InterBias)) return MODE_INTRA;
1215 : edgomez 851 }
1216 :     return mode;
1217 : Isibaar 3
1218 : edgomez 851 } else {
1219 : edgomez 195
1220 : edgomez 851 int bits, intra, i;
1221 :     VECTOR backup[5], *v;
1222 :     Data->lambda16 = iQuant;
1223 :     Data->lambda8 = pParam->m_quant_type;
1224 : chl 181
1225 : edgomez 851 v = Data->qpel ? Data->currentQMV : Data->currentMV;
1226 :     for (i = 0; i < 5; i++) {
1227 :     Data->iMinSAD[i] = 256*4096;
1228 :     backup[i] = v[i];
1229 :     }
1230 : Isibaar 3
1231 : edgomez 851 bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1232 :     if (bits == 0) return MODE_INTER; // quick stop
1233 : Isibaar 3
1234 : edgomez 851 if (inter4v) {
1235 : edgomez 963 int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1236 :     if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; }
1237 : edgomez 851 }
1238 : edgomez 195
1239 : edgomez 851 intra = CountMBBitsIntra(Data);
1240 : Isibaar 3
1241 : edgomez 851 if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; }
1242 :    
1243 :     return mode;
1244 : edgomez 195 }
1245 : edgomez 851 }
1246 : Isibaar 3
1247 : edgomez 851 static void
1248 :     SearchP(const IMAGE * const pRef,
1249 :     const uint8_t * const pRefH,
1250 :     const uint8_t * const pRefV,
1251 :     const uint8_t * const pRefHV,
1252 :     const IMAGE * const pCur,
1253 :     const int x,
1254 :     const int y,
1255 :     const uint32_t MotionFlags,
1256 :     const uint32_t GlobalFlags,
1257 :     const uint32_t iQuant,
1258 :     SearchData * const Data,
1259 :     const MBParam * const pParam,
1260 :     const MACROBLOCK * const pMBs,
1261 :     const MACROBLOCK * const prevMBs,
1262 :     int inter4v,
1263 :     MACROBLOCK * const pMB)
1264 :     {
1265 : Isibaar 3
1266 : edgomez 851 int i, iDirection = 255, mask, threshA;
1267 :     VECTOR pmv[7];
1268 : Isibaar 3
1269 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1270 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1271 : edgomez 195
1272 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1273 : Isibaar 3
1274 : edgomez 851 Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache
1275 :     i = Data->rrv ? 2 : 1;
1276 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1277 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1278 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1279 : chl 326
1280 : syskin 976 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1281 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1282 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1283 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1284 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1285 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1286 : edgomez 195
1287 : edgomez 851 Data->lambda16 = lambda_vec16[iQuant];
1288 :     Data->lambda8 = lambda_vec8[iQuant];
1289 :     Data->qpel_precision = 0;
1290 : edgomez 195
1291 : edgomez 851 if (pMB->dquant != NO_CHANGE) inter4v = 0;
1292 : edgomez 195
1293 : edgomez 963 memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1294 : chl 169
1295 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1296 :     else Data->predMV = pmv[0];
1297 : Isibaar 3
1298 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1299 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1300 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1301 :     Data->iMinSAD[2] = pMB->sad8[1];
1302 :     Data->iMinSAD[3] = pMB->sad8[2];
1303 :     Data->iMinSAD[4] = pMB->sad8[3];
1304 : chl 169
1305 : edgomez 851 if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) {
1306 :     threshA = Data->temp[0]; // that's where we keep this SAD atm
1307 :     if (threshA < 512) threshA = 512;
1308 :     else if (threshA > 1024) threshA = 1024;
1309 :     } else
1310 :     threshA = 512;
1311 : chl 169
1312 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1313 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1314 : chl 169
1315 : edgomez 851 if (!Data->rrv) {
1316 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1317 :     else CheckCandidate = CheckCandidate16no4v; //for extra speed
1318 :     } else CheckCandidate = CheckCandidate32;
1319 : chl 169
1320 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1321 : chl 169
1322 : edgomez 851 for (i = 1; i < 7; i++) {
1323 :     if (!(mask = make_mask(pmv, i)) ) continue;
1324 :     CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1325 :     if (Data->iMinSAD[0] <= threshA) break;
1326 :     }
1327 : chl 169
1328 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1329 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1330 :     (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1331 :     if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; }
1332 :     else {
1333 : Isibaar 3
1334 : edgomez 851 MainSearchFunc * MainSearchPtr;
1335 :     if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1336 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1337 :     else MainSearchPtr = DiamondSearch;
1338 : Isibaar 3
1339 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1340 : Isibaar 3
1341 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1342 :     note that this search is/might be done in halfpel positions,
1343 :     which makes it more different than the diamond above */
1344 : chl 169
1345 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH16) {
1346 :     int32_t bSAD;
1347 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1348 :     if (Data->rrv) {
1349 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1350 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1351 :     }
1352 :     if (!(MVequal(startMV, backupMV))) {
1353 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1354 : edgomez 195
1355 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1356 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1357 :     if (bSAD < Data->iMinSAD[0]) {
1358 :     Data->currentMV[0] = backupMV;
1359 :     Data->iMinSAD[0] = bSAD; }
1360 :     }
1361 : chl 169
1362 : edgomez 851 backupMV = Data->currentMV[0];
1363 :     startMV.x = startMV.y = 1;
1364 :     if (!(MVequal(startMV, backupMV))) {
1365 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1366 : chl 169
1367 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1368 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1369 :     if (bSAD < Data->iMinSAD[0]) {
1370 :     Data->currentMV[0] = backupMV;
1371 :     Data->iMinSAD[0] = bSAD; }
1372 : edgomez 195 }
1373 : edgomez 851 }
1374 :     }
1375 : edgomez 195
1376 : edgomez 851 if (MotionFlags & PMV_HALFPELREFINE16)
1377 : syskin 985 SubpelRefine(Data);
1378 : edgomez 195
1379 : edgomez 851 for(i = 0; i < 5; i++) {
1380 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1381 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1382 :     }
1383 : Isibaar 3
1384 : syskin 904 if (MotionFlags & PMV_QUARTERPELREFINE16) {
1385 : syskin 976
1386 : syskin 904 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1387 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1388 : syskin 985 Data->qpel_precision = 1;
1389 :     SubpelRefine(Data);
1390 : syskin 904 }
1391 : Isibaar 3
1392 : edgomez 851 if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0;
1393 : Isibaar 3
1394 : syskin 985 if (inter4v) {
1395 : edgomez 851 SearchData Data8;
1396 :     memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data
1397 : Isibaar 3
1398 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1399 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1400 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1401 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1402 : chl 181
1403 : edgomez 851 if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) {
1404 :     // chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it
1405 :     int sumx = 0, sumy = 0;
1406 : syskin 986 const int div = Data->qpel ? 2 : 1;
1407 : edgomez 851 const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs;
1408 : Isibaar 3
1409 : edgomez 851 for (i = 0; i < 4; i++) {
1410 :     sumx += mv[i].x / div;
1411 :     sumy += mv[i].y / div;
1412 :     }
1413 : chl 259
1414 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1415 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1416 :     }
1417 :     }
1418 : edgomez 195
1419 : edgomez 851 inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags);
1420 :    
1421 :     if (Data->rrv) {
1422 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1423 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1424 : Isibaar 3 }
1425 :    
1426 : edgomez 851 if (inter4v == MODE_INTER) {
1427 :     pMB->mode = MODE_INTER;
1428 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1429 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1430 : Isibaar 3
1431 : edgomez 851 if(Data->qpel) {
1432 :     pMB->qmvs[0] = pMB->qmvs[1]
1433 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1434 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1435 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1436 :     } else {
1437 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1438 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1439 : Isibaar 3 }
1440 :    
1441 : edgomez 851 } else if (inter4v == MODE_INTER4V) {
1442 :     pMB->mode = MODE_INTER4V;
1443 :     pMB->sad16 = Data->iMinSAD[0];
1444 :     } else { // INTRA mode
1445 :     SkipMacroblockP(pMB, 0); // not skip, but similar enough
1446 :     pMB->mode = MODE_INTRA;
1447 :     }
1448 : edgomez 195
1449 : edgomez 851 }
1450 :    
1451 :     static void
1452 :     Search8(const SearchData * const OldData,
1453 :     const int x, const int y,
1454 :     const uint32_t MotionFlags,
1455 :     const MBParam * const pParam,
1456 :     MACROBLOCK * const pMB,
1457 :     const MACROBLOCK * const pMBs,
1458 :     const int block,
1459 :     SearchData * const Data)
1460 :     {
1461 :     int i = 0;
1462 :     Data->iMinSAD = OldData->iMinSAD + 1 + block;
1463 :     Data->currentMV = OldData->currentMV + 1 + block;
1464 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1465 :    
1466 :     if(Data->qpel) {
1467 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1468 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1469 :     Data->predMV, Data->iFcode, 0, 0);
1470 :     } else {
1471 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1472 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1473 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1474 : Isibaar 3 }
1475 :    
1476 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1477 : Isibaar 3
1478 : edgomez 851 if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) {
1479 : Isibaar 3
1480 : syskin 976 if (Data->rrv) i = 16; else i = 8;
1481 : Isibaar 3
1482 : syskin 976 Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1483 :     Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1484 :     Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1485 :     Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1486 :    
1487 :     Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1488 : edgomez 851 Data->qpel_precision = 0;
1489 : Isibaar 3
1490 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1491 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1492 : Isibaar 3
1493 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1494 :     else CheckCandidate = CheckCandidate16no4v;
1495 : Isibaar 3
1496 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) {
1497 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1498 : Isibaar 3
1499 : edgomez 851 MainSearchFunc *MainSearchPtr;
1500 :     if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1501 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1502 :     else MainSearchPtr = DiamondSearch;
1503 : Isibaar 3
1504 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1505 : Isibaar 3
1506 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1507 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1508 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1509 :     }
1510 :     }
1511 : edgomez 195
1512 : edgomez 851 if (MotionFlags & PMV_HALFPELREFINE8) {
1513 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1514 : edgomez 195
1515 : edgomez 851 SubpelRefine(Data); // perform halfpel refine of current best vector
1516 : Isibaar 3
1517 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1518 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1519 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1520 :     }
1521 :     }
1522 : Isibaar 3
1523 : edgomez 851 if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) {
1524 :     Data->qpel_precision = 1;
1525 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1526 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1527 :     SubpelRefine(Data);
1528 :     }
1529 :     }
1530 : edgomez 195
1531 : edgomez 851 if (Data->rrv) {
1532 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1533 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1534 :     }
1535 :    
1536 :     if(Data->qpel) {
1537 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1538 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1539 :     pMB->qmvs[block] = *Data->currentQMV;
1540 : edgomez 195 } else {
1541 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1542 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1543 : edgomez 78 }
1544 : edgomez 851
1545 :     pMB->mvs[block] = *Data->currentMV;
1546 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1547 : Isibaar 3 }
1548 :    
1549 : edgomez 851 /* motion estimation for B-frames */
1550 : chl 345
1551 : edgomez 851 static __inline VECTOR
1552 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1553 :     {
1554 :     /* the stupidiest function ever */
1555 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1556 :     }
1557 : chl 345
1558 : edgomez 851 static void __inline
1559 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1560 :     const uint32_t iWcount,
1561 :     const MACROBLOCK * const pMB,
1562 :     const uint32_t mode_curr)
1563 : chl 345 {
1564 :    
1565 : edgomez 851 // [0] is prediction
1566 :     pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1567 : chl 345
1568 : edgomez 851 pmv[1].x = pmv[1].y = 0; // [1] is zero
1569 : chl 345
1570 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1571 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1572 : chl 345
1573 : edgomez 851 if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1574 :     pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1575 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1576 :     } else pmv[3].x = pmv[3].y = 0;
1577 : chl 345
1578 : edgomez 851 if (y != 0) {
1579 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1580 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1581 :     } else pmv[4].x = pmv[4].y = 0;
1582 : chl 345
1583 : edgomez 851 if (x != 0) {
1584 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1585 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1586 :     } else pmv[5].x = pmv[5].y = 0;
1587 : chl 345
1588 : edgomez 851 if (x != 0 && y != 0) {
1589 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1590 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1591 :     } else pmv[6].x = pmv[6].y = 0;
1592 :     }
1593 : chl 345
1594 :    
1595 : edgomez 851 /* search backward or forward */
1596 :     static void
1597 :     SearchBF( const IMAGE * const pRef,
1598 :     const uint8_t * const pRefH,
1599 :     const uint8_t * const pRefV,
1600 :     const uint8_t * const pRefHV,
1601 :     const IMAGE * const pCur,
1602 :     const int x, const int y,
1603 :     const uint32_t MotionFlags,
1604 :     const uint32_t iFcode,
1605 :     const MBParam * const pParam,
1606 :     MACROBLOCK * const pMB,
1607 :     const VECTOR * const predMV,
1608 :     int32_t * const best_sad,
1609 :     const int32_t mode_current,
1610 :     SearchData * const Data)
1611 :     {
1612 : chl 345
1613 : edgomez 851 int i, iDirection = 255, mask;
1614 :     VECTOR pmv[7];
1615 :     MainSearchFunc *MainSearchPtr;
1616 :     *Data->iMinSAD = MV_MAX_ERROR;
1617 :     Data->iFcode = iFcode;
1618 :     Data->qpel_precision = 0;
1619 :     Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache
1620 : chl 345
1621 : syskin 976 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16;
1622 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16;
1623 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16;
1624 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16;
1625 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1626 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1627 : chl 345
1628 : edgomez 851 Data->predMV = *predMV;
1629 : chl 345
1630 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1631 :     pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1632 : chl 345
1633 : edgomez 851 pmv[0] = Data->predMV;
1634 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1635 : chl 345
1636 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1637 : chl 345
1638 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1639 :     CheckCandidate = CheckCandidate16no4v;
1640 : chl 345
1641 : edgomez 851 // main loop. checking all predictions
1642 :     for (i = 0; i < 7; i++) {
1643 :     if (!(mask = make_mask(pmv, i)) ) continue;
1644 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1645 : chl 345 }
1646 :    
1647 : edgomez 851 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1648 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1649 :     else MainSearchPtr = DiamondSearch;
1650 : chl 345
1651 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1652 : chl 345
1653 : edgomez 851 SubpelRefine(Data);
1654 : chl 345
1655 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1656 :     Data->currentQMV->x = 2*Data->currentMV->x;
1657 :     Data->currentQMV->y = 2*Data->currentMV->y;
1658 :     Data->qpel_precision = 1;
1659 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1660 :     pParam->width, pParam->height, iFcode, 1, 0);
1661 :     SubpelRefine(Data);
1662 :     }
1663 : chl 345
1664 : edgomez 851 // three bits are needed to code backward mode. four for forward
1665 : Isibaar 3
1666 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1667 :     else *Data->iMinSAD += 3 * Data->lambda16;
1668 : edgomez 195
1669 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1670 :     *best_sad = *Data->iMinSAD;
1671 :     pMB->mode = mode_current;
1672 :     if (Data->qpel) {
1673 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1674 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1675 :     if (mode_current == MODE_FORWARD)
1676 :     pMB->qmvs[0] = *Data->currentQMV;
1677 :     else
1678 :     pMB->b_qmvs[0] = *Data->currentQMV;
1679 :     } else {
1680 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1681 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1682 :     }
1683 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1684 :     else pMB->b_mvs[0] = *Data->currentMV;
1685 :     }
1686 : edgomez 195
1687 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1688 :     else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
1689 : Isibaar 3 }
1690 :    
1691 : edgomez 851 static void
1692 :     SkipDecisionB(const IMAGE * const pCur,
1693 :     const IMAGE * const f_Ref,
1694 :     const IMAGE * const b_Ref,
1695 :     MACROBLOCK * const pMB,
1696 :     const uint32_t x, const uint32_t y,
1697 :     const SearchData * const Data)
1698 :     {
1699 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1700 :     int32_t sum;
1701 :     const int div = 1 + Data->qpel;
1702 :     int k;
1703 :     const uint32_t stride = Data->iEdgedWidth/2;
1704 :     //this is not full chroma compensation, only it's fullpel approximation. should work though
1705 : Isibaar 3
1706 : edgomez 851 for (k = 0; k < 4; k++) {
1707 :     dy += Data->directmvF[k].y / div;
1708 : edgomez 963 dx += Data->directmvF[k].x / div;
1709 :     b_dy += Data->directmvB[k].y / div;
1710 :     b_dx += Data->directmvB[k].x / div;
1711 : edgomez 851 }
1712 : Isibaar 3
1713 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1714 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1715 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1716 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1717 : Isibaar 3
1718 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1719 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1720 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1721 :     stride);
1722 : Isibaar 3
1723 : edgomez 851 if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip
1724 : Isibaar 3
1725 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1726 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1727 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1728 :     stride);
1729 : edgomez 195
1730 : edgomez 963 if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1731 :     pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1732 : syskin 976 for (k = 0; k < 4; k++) {
1733 : edgomez 963 pMB->qmvs[k] = pMB->mvs[k];
1734 :     pMB->b_qmvs[k] = pMB->b_mvs[k];
1735 :     }
1736 :     }
1737 : edgomez 851 }
1738 : Isibaar 3
1739 : edgomez 851 static __inline uint32_t
1740 :     SearchDirect(const IMAGE * const f_Ref,
1741 :     const uint8_t * const f_RefH,
1742 :     const uint8_t * const f_RefV,
1743 :     const uint8_t * const f_RefHV,
1744 :     const IMAGE * const b_Ref,
1745 :     const uint8_t * const b_RefH,
1746 :     const uint8_t * const b_RefV,
1747 :     const uint8_t * const b_RefHV,
1748 :     const IMAGE * const pCur,
1749 :     const int x, const int y,
1750 :     const uint32_t MotionFlags,
1751 :     const int32_t TRB, const int32_t TRD,
1752 :     const MBParam * const pParam,
1753 :     MACROBLOCK * const pMB,
1754 :     const MACROBLOCK * const b_mb,
1755 :     int32_t * const best_sad,
1756 :     SearchData * const Data)
1757 : Isibaar 3
1758 : edgomez 851 {
1759 :     int32_t skip_sad;
1760 :     int k = (x + Data->iEdgedWidth*y) * 16;
1761 :     MainSearchFunc *MainSearchPtr;
1762 : Isibaar 3
1763 : edgomez 851 *Data->iMinSAD = 256*4096;
1764 : syskin 976 Data->RefP[0] = f_Ref->y + k;
1765 :     Data->RefP[2] = f_RefH + k;
1766 :     Data->RefP[1] = f_RefV + k;
1767 :     Data->RefP[3] = f_RefHV + k;
1768 :     Data->b_RefP[0] = b_Ref->y + k;
1769 :     Data->b_RefP[2] = b_RefH + k;
1770 :     Data->b_RefP[1] = b_RefV + k;
1771 :     Data->b_RefP[3] = b_RefHV + k;
1772 :     Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1773 :     Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1774 :     Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1775 :     Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1776 : edgomez 195
1777 : edgomez 851 k = Data->qpel ? 4 : 2;
1778 :     Data->max_dx = k * (pParam->width - x * 16);
1779 :     Data->max_dy = k * (pParam->height - y * 16);
1780 :     Data->min_dx = -k * (16 + x * 16);
1781 :     Data->min_dy = -k * (16 + y * 16);
1782 : chl 181
1783 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1784 :     Data->qpel_precision = 0;
1785 : edgomez 170
1786 : edgomez 851 for (k = 0; k < 4; k++) {
1787 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1788 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1789 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1790 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1791 : Isibaar 3
1792 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1793 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1794 :    
1795 :     *best_sad = 256*4096; // in that case, we won't use direct mode
1796 :     pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1797 :     pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1798 :     return 256*4096;
1799 :     }
1800 :     if (b_mb->mode != MODE_INTER4V) {
1801 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1802 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1803 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1804 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1805 :     break;
1806 :     }
1807 : edgomez 195 }
1808 : Isibaar 3
1809 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1810 : Isibaar 3
1811 : edgomez 851 CheckCandidate(0, 0, 255, &k, Data);
1812 : edgomez 195
1813 : edgomez 851 // initial (fast) skip decision
1814 :     if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) {
1815 :     //possible skip
1816 :     if (Data->chroma) {
1817 :     pMB->mode = MODE_DIRECT_NONE_MV;
1818 :     return *Data->iMinSAD; // skip.
1819 :     } else {
1820 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1821 :     if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1822 :     }
1823 : Isibaar 3 }
1824 :    
1825 : syskin 904 *Data->iMinSAD += Data->lambda16;
1826 : edgomez 851 skip_sad = *Data->iMinSAD;
1827 : edgomez 195
1828 : edgomez 851 // DIRECT MODE DELTA VECTOR SEARCH.
1829 :     // This has to be made more effective, but at the moment I'm happy it's running at all
1830 : Isibaar 3
1831 : edgomez 851 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1832 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1833 :     else MainSearchPtr = DiamondSearch;
1834 : Isibaar 3
1835 : edgomez 851 MainSearchPtr(0, 0, Data, 255);
1836 : Isibaar 3
1837 : edgomez 851 SubpelRefine(Data);
1838 : chl 181
1839 : edgomez 851 *best_sad = *Data->iMinSAD;
1840 : chl 181
1841 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1842 :     else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1843 : chl 181
1844 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1845 : edgomez 195
1846 : edgomez 851 for (k = 0; k < 4; k++) {
1847 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1848 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1849 :     ? Data->directmvB[k].x
1850 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1851 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1852 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1853 :     ? Data->directmvB[k].y
1854 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1855 :     if (Data->qpel) {
1856 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1857 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1858 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1859 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1860 :     }
1861 : edgomez 195
1862 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1863 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1864 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1865 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1866 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1867 :     break;
1868 :     }
1869 : edgomez 78 }
1870 : edgomez 851 return skip_sad;
1871 :     }
1872 : Isibaar 3
1873 : edgomez 851 static void
1874 :     SearchInterpolate(const IMAGE * const f_Ref,
1875 :     const uint8_t * const f_RefH,
1876 :     const uint8_t * const f_RefV,
1877 :     const uint8_t * const f_RefHV,
1878 :     const IMAGE * const b_Ref,
1879 :     const uint8_t * const b_RefH,
1880 :     const uint8_t * const b_RefV,
1881 :     const uint8_t * const b_RefHV,
1882 :     const IMAGE * const pCur,
1883 :     const int x, const int y,
1884 :     const uint32_t fcode,
1885 :     const uint32_t bcode,
1886 :     const uint32_t MotionFlags,
1887 :     const MBParam * const pParam,
1888 :     const VECTOR * const f_predMV,
1889 :     const VECTOR * const b_predMV,
1890 :     MACROBLOCK * const pMB,
1891 :     int32_t * const best_sad,
1892 :     SearchData * const fData)
1893 : chl 96
1894 : edgomez 851 {
1895 : chl 169
1896 : edgomez 851 int iDirection, i, j;
1897 :     SearchData bData;
1898 : chl 169
1899 : edgomez 851 fData->qpel_precision = 0;
1900 :     memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1901 :     *fData->iMinSAD = 4096*256;
1902 :     bData.currentMV++; bData.currentQMV++;
1903 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1904 : chl 169
1905 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1906 : chl 169
1907 : syskin 976 bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i;
1908 :     bData.b_RefP[2] = fData->RefP[2] = f_RefH + i;
1909 :     bData.b_RefP[1] = fData->RefP[1] = f_RefV + i;
1910 :     bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i;
1911 :     bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i;
1912 :     bData.RefP[2] = fData->b_RefP[2] = b_RefH + i;
1913 :     bData.RefP[1] = fData->b_RefP[1] = b_RefV + i;
1914 :     bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i;
1915 :     bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1916 :     bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1917 :     bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1918 :     bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1919 : chl 169
1920 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1921 :     fData->bpredMV = bData.predMV = *b_predMV;
1922 :     fData->currentMV[0] = fData->currentMV[2];
1923 : Isibaar 3
1924 : 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);
1925 :     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);
1926 : Isibaar 3
1927 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1928 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1929 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1930 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1931 : chl 169
1932 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1933 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1934 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1935 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1936 : Isibaar 3
1937 : edgomez 851 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1938 : chl 169
1939 : edgomez 851 //diamond
1940 :     do {
1941 :     iDirection = 255;
1942 :     // forward MV moves
1943 :     i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1944 : chl 169
1945 : edgomez 851 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1946 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1947 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1948 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1949 : edgomez 195
1950 : edgomez 851 // backward MV moves
1951 :     i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1952 :     fData->currentMV[2] = fData->currentMV[0];
1953 :     CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1954 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1955 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1956 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1957 : Isibaar 3
1958 : edgomez 851 } while (!(iDirection));
1959 : chl 140
1960 : edgomez 851 //qpel refinement
1961 :     if (fData->qpel) {
1962 :     if (*fData->iMinSAD > *best_sad + 500) return;
1963 :     CheckCandidate = CheckCandidateInt;
1964 :     fData->qpel_precision = bData.qpel_precision = 1;
1965 :     get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1966 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1967 :     fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1968 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1969 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1970 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1971 :     SubpelRefine(fData);
1972 :     if (*fData->iMinSAD > *best_sad + 300) return;
1973 :     fData->currentQMV[2] = fData->currentQMV[0];
1974 :     SubpelRefine(&bData);
1975 :     }
1976 : chl 140
1977 : edgomez 851 *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode.
1978 : Isibaar 3
1979 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
1980 :     *best_sad = *fData->iMinSAD;
1981 :     pMB->mvs[0] = fData->currentMV[0];
1982 :     pMB->b_mvs[0] = fData->currentMV[1];
1983 :     pMB->mode = MODE_INTERPOLATE;
1984 :     if (fData->qpel) {
1985 :     pMB->qmvs[0] = fData->currentQMV[0];
1986 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1987 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1988 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1989 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1990 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1991 :     } else {
1992 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1993 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1994 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1995 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1996 :     }
1997 : edgomez 78 }
1998 : edgomez 851 }
1999 : Isibaar 3
2000 : edgomez 851 void
2001 :     MotionEstimationBVOP(MBParam * const pParam,
2002 : syskin 976 FRAMEINFO * const frame,
2003 :     const int32_t time_bp,
2004 :     const int32_t time_pp,
2005 :     // forward (past) reference
2006 :     const MACROBLOCK * const f_mbs,
2007 :     const IMAGE * const f_ref,
2008 :     const IMAGE * const f_refH,
2009 :     const IMAGE * const f_refV,
2010 :     const IMAGE * const f_refHV,
2011 :     // backward (future) reference
2012 :     const FRAMEINFO * const b_reference,
2013 :     const IMAGE * const b_ref,
2014 :     const IMAGE * const b_refH,
2015 :     const IMAGE * const b_refV,
2016 :     const IMAGE * const b_refHV)
2017 : edgomez 851 {
2018 :     uint32_t i, j;
2019 :     int32_t best_sad;
2020 :     uint32_t skip_sad;
2021 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2022 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
2023 : Isibaar 3
2024 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2025 : Isibaar 3
2026 : edgomez 851 const int32_t TRB = time_pp - time_bp;
2027 :     const int32_t TRD = time_pp;
2028 : edgomez 195
2029 : edgomez 851 // some pre-inintialized data for the rest of the search
2030 : Isibaar 3
2031 : edgomez 851 SearchData Data;
2032 :     int32_t iMinSAD;
2033 :     VECTOR currentMV[3];
2034 :     VECTOR currentQMV[3];
2035 :     int32_t temp[8];
2036 :     memset(&Data, 0, sizeof(SearchData));
2037 :     Data.iEdgedWidth = pParam->edged_width;
2038 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2039 :     Data.iMinSAD = &iMinSAD;
2040 :     Data.lambda16 = lambda_vec16[frame->quant];
2041 :     Data.qpel = pParam->m_quarterpel;
2042 :     Data.rounding = 0;
2043 :     Data.chroma = frame->motion_flags & PMV_CHROMA8;
2044 :     Data.temp = temp;
2045 : Isibaar 3
2046 : edgomez 851 Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose)
2047 :     // note: i==horizontal, j==vertical
2048 :     for (j = 0; j < pParam->mb_height; j++) {
2049 : edgomez 195
2050 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2051 : Isibaar 3
2052 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2053 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2054 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2055 : edgomez 195
2056 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2057 :     if (b_reference->coding_type != S_VOP)
2058 :     if (b_mb->mode == MODE_NOT_CODED) {
2059 :     pMB->mode = MODE_NOT_CODED;
2060 :     continue;
2061 :     }
2062 : Isibaar 3
2063 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2064 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2065 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2066 :     pMB->quant = frame->quant;
2067 : Isibaar 3
2068 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2069 :     and (2) sets very good predictions for forward and backward search */
2070 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2071 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2072 :     &frame->image,
2073 :     i, j,
2074 :     frame->motion_flags,
2075 :     TRB, TRD,
2076 :     pParam,
2077 :     pMB, b_mb,
2078 :     &best_sad,
2079 :     &Data);
2080 : Isibaar 3
2081 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2082 : edgomez 195
2083 : edgomez 851 // forward search
2084 :     SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2085 :     &frame->image, i, j,
2086 :     frame->motion_flags,
2087 :     frame->fcode, pParam,
2088 :     pMB, &f_predMV, &best_sad,
2089 :     MODE_FORWARD, &Data);
2090 : edgomez 195
2091 : edgomez 851 // backward search
2092 :     SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2093 :     &frame->image, i, j,
2094 :     frame->motion_flags,
2095 :     frame->bcode, pParam,
2096 :     pMB, &b_predMV, &best_sad,
2097 :     MODE_BACKWARD, &Data);
2098 :    
2099 :     // interpolate search comes last, because it uses data from forward and backward as prediction
2100 :     SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2101 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2102 :     &frame->image,
2103 :     i, j,
2104 :     frame->fcode, frame->bcode,
2105 :     frame->motion_flags,
2106 :     pParam,
2107 :     &f_predMV, &b_predMV,
2108 :     pMB, &best_sad,
2109 :     &Data);
2110 :    
2111 :     // final skip decision
2112 :     if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2113 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2114 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2115 :    
2116 :     switch (pMB->mode) {
2117 :     case MODE_FORWARD:
2118 :     f_count++;
2119 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2120 :     break;
2121 :     case MODE_BACKWARD:
2122 :     b_count++;
2123 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2124 :     break;
2125 :     case MODE_INTERPOLATE:
2126 :     i_count++;
2127 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2128 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2129 :     break;
2130 :     case MODE_DIRECT:
2131 :     case MODE_DIRECT_NO4V:
2132 :     d_count++;
2133 :     default:
2134 :     break;
2135 :     }
2136 :     }
2137 :     }
2138 : Isibaar 3 }
2139 : chl 96
2140 : edgomez 851 static __inline void
2141 :     MEanalyzeMB ( const uint8_t * const pRef,
2142 :     const uint8_t * const pCur,
2143 :     const int x,
2144 :     const int y,
2145 :     const MBParam * const pParam,
2146 :     MACROBLOCK * const pMBs,
2147 :     SearchData * const Data)
2148 : chl 96 {
2149 :    
2150 : edgomez 851 int i, mask;
2151 :     VECTOR pmv[3];
2152 : syskin 904 MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2153 : chl 96
2154 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2155 : chl 96
2156 : edgomez 851 //median is only used as prediction. it doesn't have to be real
2157 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2158 :     else
2159 :     if (x == 1) //left macroblock does not have any vector now
2160 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2161 :     else if (y == 1) // top macroblock doesn't have it's vector
2162 :     Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
2163 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2164 : edgomez 195
2165 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2166 : syskin 904 pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0);
2167 : edgomez 195
2168 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2169 : syskin 976 Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16;
2170 : chl 96
2171 : edgomez 851 pmv[1].x = EVEN(pMB->mvs[0].x);
2172 :     pmv[1].y = EVEN(pMB->mvs[0].y);
2173 :     pmv[2].x = EVEN(Data->predMV.x);
2174 :     pmv[2].y = EVEN(Data->predMV.y);
2175 :     pmv[0].x = pmv[0].y = 0;
2176 : edgomez 195
2177 : edgomez 851 CheckCandidate32I(0, 0, 255, &i, Data);
2178 : edgomez 195
2179 : syskin 904 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) {
2180 : chl 96
2181 : edgomez 851 if (!(mask = make_mask(pmv, 1)))
2182 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2183 :     if (!(mask = make_mask(pmv, 2)))
2184 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2185 : chl 96
2186 : syskin 904 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) // diamond only if needed
2187 : edgomez 851 DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2188 : syskin 904 }
2189 : edgomez 851
2190 : syskin 904 for (i = 0; i < 4; i++) {
2191 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2192 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2193 :     MB->mode = MODE_INTER;
2194 :     MB->sad16 = Data->iMinSAD[i+1];
2195 : chl 96 }
2196 : edgomez 851 }
2197 : chl 96
2198 : syskin 1001 #define INTRA_THRESH 1800
2199 : syskin 985 #define INTER_THRESH 1200
2200 : chl 96
2201 : edgomez 851 int
2202 :     MEanalysis( const IMAGE * const pRef,
2203 : syskin 904 const FRAMEINFO * const Current,
2204 :     const MBParam * const pParam,
2205 :     const int maxIntra, //maximum number if non-I frames
2206 :     const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2207 :     const int bCount, // number of B frames in a row
2208 :     const int b_thresh)
2209 : edgomez 851 {
2210 :     uint32_t x, y, intra = 0;
2211 :     int sSAD = 0;
2212 :     MACROBLOCK * const pMBs = Current->mbs;
2213 :     const IMAGE * const pCurrent = &Current->image;
2214 : syskin 904 int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh;
2215 :     int s = 0, blocks = 0;
2216 : chl 96
2217 : edgomez 851 int32_t iMinSAD[5], temp[5];
2218 :     VECTOR currentMV[5];
2219 :     SearchData Data;
2220 :     Data.iEdgedWidth = pParam->edged_width;
2221 :     Data.currentMV = currentMV;
2222 :     Data.iMinSAD = iMinSAD;
2223 :     Data.iFcode = Current->fcode;
2224 :     Data.temp = temp;
2225 :     CheckCandidate = CheckCandidate32I;
2226 : chl 96
2227 : edgomez 851 if (intraCount != 0 && intraCount < 10) // we're right after an I frame
2228 : syskin 1001 IntraThresh += 15 * (intraCount - 10) * (intraCount - 10);
2229 : edgomez 851 else
2230 :     if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2231 : syskin 1001 IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra;
2232 : chl 96
2233 : syskin 904 InterThresh -= (350 - 8*b_thresh) * bCount;
2234 :     if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh;
2235 : chl 326
2236 : edgomez 851 if (sadInit) (*sadInit) ();
2237 : chl 96
2238 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2239 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2240 :     int i;
2241 : syskin 904 blocks += 4;
2242 : edgomez 195
2243 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2244 : syskin 904 else { //extrapolation of the vector found for last frame
2245 : syskin 976 pMBs[x + y * pParam->mb_width].mvs[0].x =
2246 : syskin 904 (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2247 : syskin 976 pMBs[x + y * pParam->mb_width].mvs[0].y =
2248 : syskin 904 (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2249 :     }
2250 : edgomez 195
2251 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2252 : edgomez 195
2253 : edgomez 851 for (i = 0; i < 4; i++) {
2254 :     int dev;
2255 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2256 :     if (pMB->sad16 > IntraThresh) {
2257 :     dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2258 :     pParam->edged_width);
2259 :     if (dev + IntraThresh < pMB->sad16) {
2260 :     pMB->mode = MODE_INTRA;
2261 : syskin 904 if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2262 : edgomez 851 }
2263 :     }
2264 : syskin 904 if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++;
2265 :    
2266 : edgomez 851 sSAD += pMB->sad16;
2267 :     }
2268 :     }
2269 : edgomez 195 }
2270 : syskin 904
2271 :     sSAD /= blocks;
2272 :     s = (10*s) / blocks;
2273 :    
2274 : syskin 992 if (s > 4) sSAD += (s - 2) * (60 - 2*b_thresh); //static block - looks bad when in bframe...
2275 : syskin 904
2276 : edgomez 851 if (sSAD > InterThresh ) return P_VOP;
2277 :     emms();
2278 :     return B_VOP;
2279 :     }
2280 : chl 96
2281 :    
2282 : edgomez 851 static WARPPOINTS
2283 :     GlobalMotionEst(const MACROBLOCK * const pMBs,
2284 :     const MBParam * const pParam,
2285 :     const FRAMEINFO * const current,
2286 :     const FRAMEINFO * const reference,
2287 :     const IMAGE * const pRefH,
2288 :     const IMAGE * const pRefV,
2289 :     const IMAGE * const pRefHV )
2290 :     {
2291 : chl 96
2292 : edgomez 851 const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2293 :     const int deltay=8;
2294 :     const int grad=512; // lower bound for deviation in MB
2295 : chl 96
2296 : edgomez 851 WARPPOINTS gmc;
2297 : chl 96
2298 : edgomez 851 uint32_t mx, my;
2299 : chl 96
2300 : edgomez 851 int MBh = pParam->mb_height;
2301 :     int MBw = pParam->mb_width;
2302 : chl 96
2303 : edgomez 851 int *MBmask= calloc(MBh*MBw,sizeof(int));
2304 :     double DtimesF[4] = { 0.,0., 0., 0. };
2305 :     double sol[4] = { 0., 0., 0., 0. };
2306 :     double a,b,c,n,denom;
2307 :     double meanx,meany;
2308 :     int num,oldnum;
2309 : edgomez 195
2310 : syskin 976 if (!MBmask) { fprintf(stderr,"Mem error\n");
2311 :     gmc.duv[0].x= gmc.duv[0].y =
2312 :     gmc.duv[1].x= gmc.duv[1].y =
2313 : chl 871 gmc.duv[2].x= gmc.duv[2].y = 0;
2314 :     return gmc; }
2315 : edgomez 195
2316 : edgomez 851 // filter mask of all blocks
2317 :    
2318 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2319 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2320 : edgomez 851 {
2321 :     const int mbnum = mx + my * MBw;
2322 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2323 :     const VECTOR mv = pMB->mvs[0];
2324 :    
2325 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2326 :     continue;
2327 :    
2328 : syskin 985 if ( ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2329 :     && ( (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2330 :     && ( (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2331 :     && ( (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2332 : edgomez 851 MBmask[mbnum]=1;
2333 : chl 96 }
2334 :    
2335 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2336 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2337 : edgomez 851 {
2338 :     const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx;
2339 : chl 96
2340 : edgomez 851 const int mbnum = mx + my * MBw;
2341 :     if (!MBmask[mbnum])
2342 :     continue;
2343 :    
2344 : edgomez 876 if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad )
2345 : edgomez 851 MBmask[mbnum] = 0;
2346 : edgomez 876 if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad )
2347 : edgomez 851 MBmask[mbnum] = 0;
2348 :    
2349 : edgomez 195 }
2350 : chl 96
2351 : edgomez 851 emms();
2352 : chl 96
2353 : edgomez 851 do { /* until convergence */
2354 : chl 96
2355 : edgomez 851 a = b = c = n = 0;
2356 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2357 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2358 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2359 : edgomez 851 {
2360 :     const int mbnum = mx + my * MBw;
2361 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2362 :     const VECTOR mv = pMB->mvs[0];
2363 : chl 96
2364 : edgomez 851 if (!MBmask[mbnum])
2365 :     continue;
2366 : chl 96
2367 : edgomez 851 n++;
2368 :     a += 16*mx+8;
2369 :     b += 16*my+8;
2370 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2371 : chl 96
2372 : edgomez 851 DtimesF[0] += (double)mv.x;
2373 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2374 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2375 :     DtimesF[3] += (double)mv.y;
2376 :     }
2377 : chl 96
2378 : edgomez 851 denom = a*a+b*b-c*n;
2379 : chl 96
2380 : syskin 976 /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2381 : edgomez 851 /* D'*E*F has been calculated in the same loop as matrix */
2382 : chl 96
2383 : edgomez 851 sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2384 : syskin 976 sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2385 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2386 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2387 : chl 96
2388 : edgomez 851 sol[0] /= denom;
2389 :     sol[1] /= denom;
2390 :     sol[2] /= denom;
2391 :     sol[3] /= denom;
2392 : chl 96
2393 : edgomez 851 meanx = meany = 0.;
2394 :     oldnum = 0;
2395 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2396 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2397 : edgomez 851 {
2398 :     const int mbnum = mx + my * MBw;
2399 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2400 :     const VECTOR mv = pMB->mvs[0];
2401 : chl 96
2402 : edgomez 851 if (!MBmask[mbnum])
2403 :     continue;
2404 : chl 184
2405 : edgomez 851 oldnum++;
2406 : syskin 985 meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x );
2407 :     meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y );
2408 : edgomez 851 }
2409 : chl 96
2410 : edgomez 851 if (4*meanx > oldnum) /* better fit than 0.25 is useless */
2411 :     meanx /= oldnum;
2412 :     else
2413 :     meanx = 0.25;
2414 : edgomez 195
2415 : edgomez 851 if (4*meany > oldnum)
2416 :     meany /= oldnum;
2417 :     else
2418 :     meany = 0.25;
2419 : chl 96
2420 : edgomez 851 /* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2421 :     fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2422 :     */
2423 :     num = 0;
2424 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2425 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2426 : edgomez 851 {
2427 :     const int mbnum = mx + my * MBw;
2428 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2429 :     const VECTOR mv = pMB->mvs[0];
2430 : chl 96
2431 : edgomez 851 if (!MBmask[mbnum])
2432 :     continue;
2433 : chl 96
2434 : syskin 985 if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx )
2435 :     || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) )
2436 : edgomez 851 MBmask[mbnum]=0;
2437 :     else
2438 :     num++;
2439 : chl 96 }
2440 : edgomez 195
2441 : edgomez 851 } while ( (oldnum != num) && (num>=4) );
2442 : edgomez 195
2443 : edgomez 851 if (num < 4)
2444 :     {
2445 :     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;
2446 :     } else {
2447 : edgomez 195
2448 : edgomez 851 gmc.duv[0].x=(int)(sol[0]+0.5);
2449 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2450 : chl 96
2451 : edgomez 851 gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2452 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2453 : chl 96
2454 : edgomez 851 gmc.duv[2].x=0;
2455 :     gmc.duv[2].y=0;
2456 :     }
2457 :     // fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y);
2458 : chl 96
2459 : edgomez 851 free(MBmask);
2460 : chl 96
2461 : edgomez 851 return gmc;
2462 : chl 96 }
2463 :    
2464 : edgomez 851 // functions which perform BITS-based search/bitcount
2465 : chl 96
2466 : edgomez 851 static int
2467 :     CountMBBitsInter(SearchData * const Data,
2468 :     const MACROBLOCK * const pMBs, const int x, const int y,
2469 :     const MBParam * const pParam,
2470 :     const uint32_t MotionFlags)
2471 : chl 96 {
2472 : edgomez 851 int i, iDirection;
2473 :     int32_t bsad[5];
2474 : chl 141
2475 : edgomez 851 CheckCandidate = CheckCandidateBits16;
2476 : chl 96
2477 : edgomez 851 if (Data->qpel) {
2478 :     for(i = 0; i < 5; i++) {
2479 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2480 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2481 :     }
2482 :     Data->qpel_precision = 1;
2483 :     CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2484 : chl 96
2485 : edgomez 851 //checking if this vector is perfect. if it is, we stop.
2486 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0)
2487 :     return 0; //quick stop
2488 : edgomez 195
2489 : edgomez 851 if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search
2490 :     for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2491 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2492 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2493 :     Data->qpel_precision = 0;
2494 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2495 :     CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2496 :     }
2497 : edgomez 195
2498 : edgomez 851 } else { // not qpel
2499 : edgomez 195
2500 : edgomez 851 CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2501 :     //checking if this vector is perfect. if it is, we stop.
2502 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) {
2503 :     return 0; //inter
2504 :     }
2505 :     }
2506 : chl 96
2507 : edgomez 851 if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2508 : syskin 985
2509 : edgomez 851 if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data);
2510 : chl 96
2511 : edgomez 851 if (Data->qpel) {
2512 :     if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search
2513 :     for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2514 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match
2515 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2516 :     }
2517 : edgomez 195
2518 : edgomez 851 // preparing for qpel-precision search
2519 :     Data->qpel_precision = 1;
2520 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2521 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
2522 :     }
2523 :     if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2524 :     }
2525 : chl 96
2526 : edgomez 851 if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction
2527 :     VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2528 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2529 :     CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2530 :     }
2531 :     return Data->iMinSAD[0];
2532 :     }
2533 : chl 96
2534 :    
2535 : edgomez 851 static int
2536 :     CountMBBitsInter4v(const SearchData * const Data,
2537 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2538 :     const int x, const int y,
2539 :     const MBParam * const pParam, const uint32_t MotionFlags,
2540 :     const VECTOR * const backup)
2541 :     {
2542 : chl 96
2543 : edgomez 851 int cbp = 0, bits = 0, t = 0, i, iDirection;
2544 :     SearchData Data2, *Data8 = &Data2;
2545 :     int sumx = 0, sumy = 0;
2546 : edgomez 961 int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64;
2547 : chl 96
2548 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2549 :     CheckCandidate = CheckCandidateBits8;
2550 : chl 96
2551 : edgomez 851 for (i = 0; i < 4; i++) {
2552 :     Data8->iMinSAD = Data->iMinSAD + i + 1;
2553 :     Data8->currentMV = Data->currentMV + i + 1;
2554 :     Data8->currentQMV = Data->currentQMV + i + 1;
2555 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2556 : syskin 976 Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2557 :     Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2558 :     Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2559 :     Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2560 : edgomez 195
2561 : edgomez 851 if(Data->qpel) {
2562 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2563 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2564 :     Data8->predMV, Data8->iFcode, 0, 0);
2565 :     } else {
2566 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2567 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2568 :     Data8->predMV, Data8->iFcode, 0, 0);
2569 :     }
2570 : chl 96
2571 : 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,
2572 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2573 : chl 96
2574 : edgomez 851 *Data8->iMinSAD += t;
2575 : chl 96
2576 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2577 :     // checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far)
2578 : syskin 985 {
2579 :     VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV;
2580 :     if (!( (v->x == backup[i+1].x) && (v->y == backup[i+1].y) ))
2581 : edgomez 851 CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2582 :     }
2583 : edgomez 195
2584 : edgomez 851 if (Data8->qpel) {
2585 :     if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows
2586 :     int32_t s = *Data8->iMinSAD;
2587 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2588 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2589 :     Data8->qpel_precision = 0;
2590 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2591 :     pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2592 : edgomez 195
2593 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2594 :     CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2595 : edgomez 195
2596 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS)
2597 :     SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2598 : chl 96
2599 : edgomez 851 if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8);
2600 : chl 96
2601 : edgomez 851 if(s > *Data8->iMinSAD) { //we have found a better match
2602 :     Data8->currentQMV->x = 2*Data8->currentMV->x;
2603 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2604 :     }
2605 : chl 96
2606 : edgomez 851 Data8->qpel_precision = 1;
2607 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2608 :     pParam->width, pParam->height, Data8->iFcode, 1, 0);
2609 : chl 96
2610 : edgomez 851 }
2611 :     if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2612 : edgomez 195
2613 : edgomez 851 } else // not qpel
2614 :     if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement
2615 :    
2616 :     //checking vector equal to predicion
2617 :     if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) {
2618 :     const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2619 :     if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y))
2620 :     CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2621 : chl 141 }
2622 : edgomez 195
2623 : edgomez 851 bits += *Data8->iMinSAD;
2624 :     if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V
2625 :    
2626 :     // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else
2627 :     if(Data->qpel) {
2628 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2629 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2630 :     pMB->qmvs[i] = *Data8->currentQMV;
2631 :     sumx += Data8->currentQMV->x/2;
2632 :     sumy += Data8->currentQMV->y/2;
2633 :     } else {
2634 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2635 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2636 :     sumx += Data8->currentMV->x;
2637 :     sumy += Data8->currentMV->y;
2638 : chl 141 }
2639 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2640 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2641 :     if (Data8->temp[0]) cbp |= 1 << (5 - i);
2642 : chl 141 }
2643 :    
2644 : edgomez 851 if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma
2645 :     const uint8_t * ptr;
2646 :     sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2647 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2648 : chl 141
2649 : edgomez 851 //chroma U
2650 : syskin 976 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2651 : edgomez 851 transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2652 :     fdct(in);
2653 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2654 :     else i = quant4_inter(coeff, in, Data->lambda16);
2655 :     if (i > 0) {
2656 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2657 :     cbp |= 1 << (5 - 4);
2658 :     }
2659 : chl 141
2660 : edgomez 851 if (bits < *Data->iMinSAD) { // still possible
2661 :     //chroma V
2662 : syskin 976 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2663 : edgomez 851 transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2664 :     fdct(in);
2665 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2666 :     else i = quant4_inter(coeff, in, Data->lambda16);
2667 :     if (i > 0) {
2668 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2669 :     cbp |= 1 << (5 - 5);
2670 :     }
2671 : chl 897 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
2672 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2673 :     }
2674 : edgomez 195 }
2675 : chl 96
2676 : edgomez 851 return bits;
2677 :     }
2678 : chl 96
2679 :    
2680 : edgomez 851 static int
2681 :     CountMBBitsIntra(const SearchData * const Data)
2682 :     {
2683 :     int bits = 1; //this one is ac/dc prediction flag. always 1.
2684 : syskin 976 int cbp = 0, i, t, dc = 1024, b_dc;
2685 : edgomez 851 const uint32_t iQuant = Data->lambda16;
2686 : edgomez 961 int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2687 : syskin 976 uint32_t iDcScaler = get_dc_scaler(iQuant, 1);;
2688 : edgomez 195
2689 : edgomez 851 for(i = 0; i < 4; i++) {
2690 :     int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2691 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2692 :     fdct(in);
2693 : syskin 976 b_dc = in[0];
2694 :     in[0] -= dc;
2695 :     dc = b_dc;
2696 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2697 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2698 : chl 141
2699 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;;
2700 :     Data->temp[i] = t;
2701 :     if (t != 0) cbp |= 1 << (5 - i);
2702 :     if (bits >= Data->iMinSAD[0]) break;
2703 :     }
2704 : chl 181
2705 : edgomez 851 if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma
2706 : syskin 976 iDcScaler = get_dc_scaler(iQuant, 0);
2707 : edgomez 851 //chroma U
2708 :     transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2709 :     fdct(in);
2710 :     in[0] -= 1024;
2711 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2712 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2713 : chl 96
2714 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2715 :     if (t != 0) cbp |= 1 << (5 - 4);
2716 : chl 141
2717 : edgomez 851 if (bits < Data->iMinSAD[0]) {
2718 :     //chroma V
2719 :     transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2720 :     fdct(in);
2721 :     in[0] -= 1024;
2722 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2723 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2724 : chl 96
2725 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2726 :     if (t != 0) cbp |= 1 << (5 - 5);
2727 : chl 96
2728 : syskin 904 bits += xvid_cbpy_tab[cbp>>2].len;
2729 :     bits += mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2730 : chl 96 }
2731 :     }
2732 : edgomez 851 return bits;
2733 : chl 96 }

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