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

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

Parent Directory Parent Directory | Revision Log Revision Log


Revision 982 - (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 : edgomez 974 int bits;
85 :     const int q = (1 << (iFcode - 1)) - 1;
86 :    
87 :     x <<= qpel;
88 :     y <<= qpel;
89 : edgomez 851 if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
90 : chl 141
91 : edgomez 851 x -= pred.x;
92 : edgomez 974 bits = (x != 0 ? iFcode:0);
93 :     x = abs(x);
94 :     x += q;
95 :     x >>= (iFcode - 1);
96 :     bits += mvtab[x];
97 :    
98 : edgomez 851 y -= pred.y;
99 : edgomez 974 bits += (y != 0 ? iFcode:0);
100 :     y = abs(y);
101 :     y += q;
102 :     y >>= (iFcode - 1);
103 :     bits += mvtab[y];
104 : chl 141
105 : edgomez 974 return bits;
106 : edgomez 851 }
107 : chl 141
108 : edgomez 978 static int32_t ChromaSAD2(const int fx, const int fy, const int bx, const int by,
109 :     const SearchData * const data)
110 : edgomez 851 {
111 :     int sad;
112 :     const uint32_t stride = data->iEdgedWidth/2;
113 :     uint8_t * f_refu = data->RefQ,
114 :     * f_refv = data->RefQ + 8,
115 :     * b_refu = data->RefQ + 16,
116 :     * b_refv = data->RefQ + 24;
117 : edgomez 978 int offset = (fx>>1) + (fy>>1)*stride;
118 : Isibaar 3
119 : edgomez 851 switch (((fx & 1) << 1) | (fy & 1)) {
120 :     case 0:
121 : edgomez 978 f_refu = (uint8_t*)data->RefP[4] + offset;
122 :     f_refv = (uint8_t*)data->RefP[5] + offset;
123 : edgomez 851 break;
124 :     case 1:
125 : edgomez 978 interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding);
126 :     interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding);
127 : edgomez 851 break;
128 :     case 2:
129 : edgomez 978 interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding);
130 :     interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding);
131 : edgomez 851 break;
132 :     default:
133 : edgomez 978 interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding);
134 :     interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding);
135 : edgomez 851 break;
136 :     }
137 : Isibaar 3
138 : edgomez 978 offset = (bx>>1) + (by>>1)*stride;
139 : edgomez 851 switch (((bx & 1) << 1) | (by & 1)) {
140 :     case 0:
141 : edgomez 978 b_refu = (uint8_t*)data->b_RefP[4] + offset;
142 :     b_refv = (uint8_t*)data->b_RefP[5] + offset;
143 : edgomez 851 break;
144 :     case 1:
145 : edgomez 978 interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
146 :     interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
147 : edgomez 851 break;
148 :     case 2:
149 : edgomez 978 interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
150 :     interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
151 : edgomez 851 break;
152 :     default:
153 : edgomez 978 interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding);
154 :     interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding);
155 : edgomez 851 break;
156 :     }
157 : Isibaar 3
158 : edgomez 851 sad = sad8bi(data->CurU, b_refu, f_refu, stride);
159 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
160 : Isibaar 3
161 : edgomez 851 return sad;
162 :     }
163 : Isibaar 3
164 : edgomez 851 static int32_t
165 : edgomez 978 ChromaSAD(const int dx, const int dy, const SearchData * const data)
166 : edgomez 851 {
167 :     int sad;
168 :     const uint32_t stride = data->iEdgedWidth/2;
169 : edgomez 978 int offset = (dx>>1) + (dy>>1)*stride;
170 : Isibaar 3
171 : edgomez 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 : edgomez 978 sad = sad8(data->CurU, data->RefP[4] + offset, stride);
177 :     sad += sad8(data->CurV, data->RefP[5] + offset, stride);
178 : edgomez 851 break;
179 :     case 1:
180 : edgomez 978 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride);
181 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride);
182 : edgomez 851 break;
183 :     case 2:
184 : edgomez 978 sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride);
185 :     sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride);
186 : edgomez 851 break;
187 :     default:
188 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding);
189 : edgomez 851 sad = sad8(data->CurU, data->RefQ, stride);
190 :    
191 : edgomez 978 interpolate8x8_halfpel_hv(data->RefQ, data->RefP[5] + offset, stride, data->rounding);
192 : edgomez 851 sad += sad8(data->CurV, data->RefQ, stride);
193 :     break;
194 :     }
195 :     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 : edgomez 978 const uint8_t* const *direction = ( dir == 0 ? data->RefP : data->b_RefP );
204 :     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 : edgomez 978 const int picture = ((x&1)<<1) | (y&1);
214 :     const int offset = (x>>1) + (y>>1)*data->iEdgedWidth;
215 :     return data->RefP[picture] + offset;
216 : Isibaar 3 }
217 : edgomez 851
218 :     static uint8_t *
219 :     Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
220 : Isibaar 3 {
221 : edgomez 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 : edgomez 978
258 : edgomez 851 }
259 :     return Reference;
260 :     }
261 : Isibaar 3
262 : edgomez 851 static uint8_t *
263 :     Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
264 :     {
265 :     // 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 : edgomez 978 // 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 : edgomez 851 if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
338 : edgomez 953 (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 :    
355 : Isibaar 3 }
356 :    
357 : edgomez 851 static void
358 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
359 :     {
360 :     int32_t sad; uint32_t t;
361 :     const uint8_t * Reference;
362 : edgomez 974 VECTOR * current;
363 : chl 326
364 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
365 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
366 : Isibaar 3
367 : edgomez 974 if (!data->qpel_precision) {
368 :     Reference = GetReference(x, y, data);
369 :     current = data->currentMV;
370 :     } else { // x and y are in 1/4 precision
371 :     Reference = Interpolate8x8qpel(x, y, 0, 0, data);
372 :     current = data->currentQMV;
373 :     }
374 : Isibaar 3
375 : edgomez 851 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
376 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
377 : Isibaar 3
378 : edgomez 851 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
379 : Isibaar 3
380 : edgomez 851 if (sad < *(data->iMinSAD)) {
381 :     *(data->iMinSAD) = sad;
382 : edgomez 974 current->x = x; current->y = y;
383 : edgomez 851 *dir = Direction;
384 :     }
385 : Isibaar 3 }
386 :    
387 :    
388 : edgomez 851 static void
389 :     CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
390 :     {
391 :     uint32_t t;
392 :     const uint8_t * Reference;
393 :    
394 : edgomez 963 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero even value
395 : edgomez 851 (x > data->max_dx) || (x < data->min_dx)
396 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
397 :    
398 :     Reference = GetReference(x, y, data);
399 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
400 :    
401 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
402 :    
403 :     data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
404 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
405 :    
406 :     if (data->temp[0] < data->iMinSAD[0]) {
407 :     data->iMinSAD[0] = data->temp[0];
408 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
409 :     *dir = Direction; }
410 :    
411 :     if (data->temp[1] < data->iMinSAD[1]) {
412 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
413 :     if (data->temp[2] < data->iMinSAD[2]) {
414 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
415 :     if (data->temp[3] < data->iMinSAD[3]) {
416 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
417 :     if (data->temp[4] < data->iMinSAD[4]) {
418 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
419 : Isibaar 3 }
420 : edgomez 195
421 : edgomez 851 static void
422 :     CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
423 :     {
424 :     int32_t sad, xc, yc;
425 :     const uint8_t * Reference;
426 :     uint32_t t;
427 :     VECTOR * current;
428 :    
429 : edgomez 963 if ( (x > data->max_dx) || ( x < data->min_dx)
430 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
431 : edgomez 851
432 :     if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value
433 :    
434 :     if (data->qpel_precision) { // x and y are in 1/4 precision
435 :     Reference = Interpolate16x16qpel(x, y, 0, data);
436 :     current = data->currentQMV;
437 :     xc = x/2; yc = y/2;
438 :     } else {
439 :     Reference = GetReference(x, y, data);
440 :     current = data->currentMV;
441 :     xc = x; yc = y;
442 :     }
443 :     t = d_mv_bits(x, y, data->predMV, data->iFcode,
444 :     data->qpel^data->qpel_precision, data->rrv);
445 :    
446 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
447 :     sad += (data->lambda16 * t * sad)>>10;
448 :    
449 :     if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
450 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
451 :    
452 :     if (sad < *(data->iMinSAD)) {
453 :     *(data->iMinSAD) = sad;
454 :     current->x = x; current->y = y;
455 :     *dir = Direction;
456 :     }
457 : chl 96 }
458 : Isibaar 3
459 : edgomez 851 static void
460 :     CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
461 :     {
462 :     // maximum speed - for P/B/I decision
463 :     int32_t sad;
464 :    
465 :     if ( (x > data->max_dx) || (x < data->min_dx)
466 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
467 :    
468 : edgomez 978 sad = sad32v_c(data->Cur, data->RefP[0] + x/2 + (y/2)*(data->iEdgedWidth),
469 : edgomez 851 data->iEdgedWidth, data->temp+1);
470 :    
471 :     if (sad < *(data->iMinSAD)) {
472 :     *(data->iMinSAD) = sad;
473 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
474 :     *dir = Direction;
475 :     }
476 :     if (data->temp[1] < data->iMinSAD[1]) {
477 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
478 :     if (data->temp[2] < data->iMinSAD[2]) {
479 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
480 :     if (data->temp[3] < data->iMinSAD[3]) {
481 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
482 :     if (data->temp[4] < data->iMinSAD[4]) {
483 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
484 :    
485 : Isibaar 3 }
486 :    
487 : edgomez 851 static void
488 :     CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
489 :     {
490 :     int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
491 :     uint32_t t;
492 :     const uint8_t *ReferenceF, *ReferenceB;
493 :     VECTOR *current;
494 :    
495 : edgomez 953 if ((xf > data->max_dx) || (xf < data->min_dx) ||
496 :     (yf > data->max_dy) || (yf < data->min_dy))
497 :     return;
498 : edgomez 851
499 :     if (!data->qpel_precision) {
500 :     ReferenceF = GetReference(xf, yf, data);
501 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
502 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
503 :     current = data->currentMV;
504 :     xcf = xf; ycf = yf;
505 :     xcb = xb; ycb = yb;
506 :     } else {
507 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
508 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
509 :     current = data->currentQMV;
510 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
511 :     xcf = xf/2; ycf = yf/2;
512 :     xcb = xb/2; ycb = yb/2;
513 :     }
514 :    
515 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
516 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
517 :    
518 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
519 :     sad += (data->lambda16 * t * sad)>>10;
520 :    
521 :     if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
522 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
523 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
524 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
525 :    
526 :     if (sad < *(data->iMinSAD)) {
527 :     *(data->iMinSAD) = sad;
528 :     current->x = xf; current->y = yf;
529 :     *dir = Direction;
530 :     }
531 : Isibaar 3 }
532 :    
533 : edgomez 851 static void
534 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
535 :     {
536 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
537 :     uint32_t k;
538 :     const uint8_t *ReferenceF;
539 :     const uint8_t *ReferenceB;
540 :     VECTOR mvs, b_mvs;
541 :    
542 : edgomez 953 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
543 : edgomez 851
544 :     for (k = 0; k < 4; k++) {
545 :     mvs.x = data->directmvF[k].x + x;
546 :     b_mvs.x = ((x == 0) ?
547 :     data->directmvB[k].x
548 :     : mvs.x - data->referencemv[k].x);
549 :    
550 :     mvs.y = data->directmvF[k].y + y;
551 :     b_mvs.y = ((y == 0) ?
552 :     data->directmvB[k].y
553 :     : mvs.y - data->referencemv[k].y);
554 :    
555 : edgomez 953 if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) ||
556 :     (mvs.y > data->max_dy) || (mvs.y < data->min_dy) ||
557 :     (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) ||
558 :     (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) )
559 :     return;
560 : edgomez 851
561 :     if (data->qpel) {
562 :     xcf += mvs.x/2; ycf += mvs.y/2;
563 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
564 :     } else {
565 :     xcf += mvs.x; ycf += mvs.y;
566 :     xcb += b_mvs.x; ycb += b_mvs.y;
567 :     mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway
568 :     b_mvs.x *= 2; b_mvs.y *= 2;
569 :     }
570 :    
571 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
572 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
573 :    
574 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
575 :     ReferenceF, ReferenceB, data->iEdgedWidth);
576 :     if (sad > *(data->iMinSAD)) return;
577 :     }
578 :    
579 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
580 :    
581 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
582 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
583 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
584 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
585 :    
586 :     if (sad < *(data->iMinSAD)) {
587 :     *(data->iMinSAD) = sad;
588 :     data->currentMV->x = x; data->currentMV->y = y;
589 :     *dir = Direction;
590 :     }
591 : Isibaar 3 }
592 :    
593 : edgomez 851 static void
594 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
595 : Isibaar 3 {
596 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
597 :     const uint8_t *ReferenceF;
598 :     const uint8_t *ReferenceB;
599 :     VECTOR mvs, b_mvs;
600 : Isibaar 3
601 : edgomez 963 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
602 : Isibaar 3
603 : edgomez 851 mvs.x = data->directmvF[0].x + x;
604 :     b_mvs.x = ((x == 0) ?
605 :     data->directmvB[0].x
606 :     : mvs.x - data->referencemv[0].x);
607 : Isibaar 3
608 : edgomez 851 mvs.y = data->directmvF[0].y + y;
609 :     b_mvs.y = ((y == 0) ?
610 :     data->directmvB[0].y
611 :     : mvs.y - data->referencemv[0].y);
612 : Isibaar 3
613 : edgomez 963 if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx)
614 :     || (mvs.y > data->max_dy) || (mvs.y < data->min_dy)
615 :     || (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx)
616 :     || (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return;
617 : Isibaar 3
618 : edgomez 851 if (data->qpel) {
619 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
620 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
621 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
622 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
623 :     } else {
624 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
625 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
626 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
627 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
628 :     }
629 :    
630 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
631 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
632 :    
633 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
634 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
635 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
636 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
637 :    
638 :     if (sad < *(data->iMinSAD)) {
639 :     *(data->iMinSAD) = sad;
640 :     data->currentMV->x = x; data->currentMV->y = y;
641 :     *dir = Direction;
642 :     }
643 : Isibaar 3 }
644 :    
645 : edgomez 851
646 :     static void
647 :     CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
648 : Isibaar 3 {
649 :    
650 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
651 : edgomez 851 int32_t bits = 0, sum;
652 :     VECTOR * current;
653 :     const uint8_t * ptr;
654 :     int i, cbp = 0, t, xc, yc;
655 : edgomez 195
656 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
657 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
658 : edgomez 195
659 : edgomez 851 if (!data->qpel_precision) {
660 :     ptr = GetReference(x, y, data);
661 :     current = data->currentMV;
662 :     xc = x; yc = y;
663 :     } else { // x and y are in 1/4 precision
664 :     ptr = Interpolate16x16qpel(x, y, 0, data);
665 :     current = data->currentQMV;
666 :     xc = x/2; yc = y/2;
667 :     }
668 : Isibaar 3
669 : edgomez 851 for(i = 0; i < 4; i++) {
670 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
671 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
672 :     fdct(in);
673 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
674 :     else sum = quant4_inter(coeff, in, data->lambda16);
675 :     if (sum > 0) {
676 :     cbp |= 1 << (5 - i);
677 :     bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
678 :     } else data->temp[i] = 0;
679 :     }
680 : Isibaar 3
681 : edgomez 851 bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
682 : edgomez 195
683 : edgomez 851 if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma
684 :     xc = (xc >> 1) + roundtab_79[xc & 0x3];
685 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
686 :    
687 :     //chroma U
688 : edgomez 978 ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
689 : edgomez 851 transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
690 :     fdct(in);
691 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
692 :     else sum = quant4_inter(coeff, in, data->lambda16);
693 :     if (sum > 0) {
694 :     cbp |= 1 << (5 - 4);
695 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
696 : chl 344 }
697 : edgomez 851
698 :     if (bits < data->iMinSAD[0]) {
699 :     //chroma V
700 : edgomez 978 ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
701 : edgomez 851 transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2);
702 :     fdct(in);
703 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
704 :     else sum = quant4_inter(coeff, in, data->lambda16);
705 :     if (sum > 0) {
706 :     cbp |= 1 << (5 - 5);
707 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
708 :     }
709 :     }
710 : edgomez 78 }
711 : edgomez 851
712 : edgomez 958 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
713 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
714 :    
715 :     if (bits < data->iMinSAD[0]) {
716 :     data->iMinSAD[0] = bits;
717 :     current[0].x = x; current[0].y = y;
718 :     *dir = Direction;
719 :     }
720 :    
721 :     if (data->temp[0] + t < data->iMinSAD[1]) {
722 :     data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
723 :     if (data->temp[1] < data->iMinSAD[2]) {
724 :     data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
725 :     if (data->temp[2] < data->iMinSAD[3]) {
726 :     data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
727 :     if (data->temp[3] < data->iMinSAD[4]) {
728 :     data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
729 :    
730 : Isibaar 3 }
731 : edgomez 851 static void
732 :     CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
733 : chl 96 {
734 :    
735 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
736 : edgomez 851 int32_t sum, bits;
737 :     VECTOR * current;
738 :     const uint8_t * ptr;
739 :     int cbp;
740 : edgomez 195
741 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
742 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
743 : edgomez 195
744 : edgomez 851 if (!data->qpel_precision) {
745 :     ptr = GetReference(x, y, data);
746 :     current = data->currentMV;
747 :     } else { // x and y are in 1/4 precision
748 :     ptr = Interpolate8x8qpel(x, y, 0, 0, data);
749 :     current = data->currentQMV;
750 :     }
751 : chl 96
752 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
753 :     fdct(in);
754 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
755 :     else sum = quant4_inter(coeff, in, data->lambda16);
756 :     if (sum > 0) {
757 :     bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
758 :     cbp = 1;
759 :     } else cbp = bits = 0;
760 : chl 96
761 : edgomez 851 bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
762 : chl 96
763 : edgomez 851 if (bits < data->iMinSAD[0]) {
764 :     data->temp[0] = cbp;
765 :     data->iMinSAD[0] = bits;
766 :     current[0].x = x; current[0].y = y;
767 :     *dir = Direction;
768 :     }
769 :     }
770 : chl 96
771 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
772 : edgomez 195
773 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
774 : chl 96
775 : edgomez 851 static void
776 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
777 :     {
778 : edgomez 195
779 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
780 : edgomez 195
781 : edgomez 851 int iDirection;
782 : edgomez 195
783 : edgomez 851 for(;;) { //forever
784 :     iDirection = 0;
785 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
786 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
787 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
788 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
789 : edgomez 195
790 : edgomez 851 /* now we're doing diagonal checks near our candidate */
791 : edgomez 195
792 : edgomez 851 if (iDirection) { //if anything found
793 :     bDirection = iDirection;
794 :     iDirection = 0;
795 :     x = data->currentMV->x; y = data->currentMV->y;
796 :     if (bDirection & 3) { //our candidate is left or right
797 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
798 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
799 :     } else { // what remains here is up or down
800 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
801 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
802 :     }
803 : edgomez 195
804 : edgomez 851 if (iDirection) {
805 :     bDirection += iDirection;
806 :     x = data->currentMV->x; y = data->currentMV->y;
807 :     }
808 :     } else { //about to quit, eh? not so fast....
809 :     switch (bDirection) {
810 :     case 2:
811 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
812 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
813 : edgomez 195 break;
814 : edgomez 851 case 1:
815 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
816 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
817 : edgomez 195 break;
818 : edgomez 851 case 2 + 4:
819 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
820 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
821 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
822 :     break;
823 :     case 4:
824 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
825 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
826 :     break;
827 : edgomez 195 case 8:
828 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
829 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
830 : edgomez 195 break;
831 : edgomez 851 case 1 + 4:
832 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
833 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
834 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
835 : edgomez 195 break;
836 : edgomez 851 case 2 + 8:
837 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
838 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
839 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
840 :     break;
841 :     case 1 + 8:
842 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
843 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
844 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
845 :     break;
846 :     default: //1+2+4+8 == we didn't find anything at all
847 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
848 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
849 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
850 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
851 :     break;
852 : chl 96 }
853 : edgomez 851 if (!iDirection) break; //ok, the end. really
854 :     bDirection = iDirection;
855 :     x = data->currentMV->x; y = data->currentMV->y;
856 : chl 345 }
857 : edgomez 195 }
858 : chl 96 }
859 :    
860 : edgomez 851 static void
861 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
862 : chl 96 {
863 : edgomez 851 int iDirection;
864 : edgomez 195
865 : edgomez 851 do {
866 :     iDirection = 0;
867 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
868 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
869 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
870 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
871 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
872 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
873 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
874 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
875 : chl 96
876 : edgomez 851 bDirection = iDirection;
877 :     x = data->currentMV->x; y = data->currentMV->y;
878 :     } while (iDirection);
879 : chl 96 }
880 :    
881 : edgomez 851 static void
882 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
883 : chl 181 {
884 :    
885 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
886 :    
887 : edgomez 851 int iDirection;
888 : edgomez 195
889 : edgomez 851 do {
890 :     iDirection = 0;
891 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
892 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
893 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
894 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
895 :    
896 :     /* now we're doing diagonal checks near our candidate */
897 :    
898 :     if (iDirection) { //checking if anything found
899 :     bDirection = iDirection;
900 : chl 181 iDirection = 0;
901 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
902 :     if (bDirection & 3) { //our candidate is left or right
903 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
904 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
905 :     } else { // what remains here is up or down
906 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
907 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
908 :     }
909 :     bDirection += iDirection;
910 :     x = data->currentMV->x; y = data->currentMV->y;
911 :     }
912 :     }
913 :     while (iDirection);
914 :     }
915 : chl 181
916 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
917 : chl 181
918 : edgomez 851 static void
919 :     SubpelRefine(const SearchData * const data)
920 :     {
921 :     /* Do a half-pel or q-pel refinement */
922 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
923 :     int iDirection; //only needed because macro expects it
924 : chl 181
925 : edgomez 851 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
926 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
927 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
928 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
929 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
930 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
931 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
932 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
933 :     }
934 : chl 181
935 : edgomez 851 static __inline int
936 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
937 :     const int x, const int y,
938 :     const uint32_t stride, const uint32_t iQuant, int rrv)
939 : chl 181
940 : edgomez 851 {
941 : edgomez 963 int offset = (x + y*stride)*8;
942 : edgomez 851 if(!rrv) {
943 : edgomez 963 uint32_t sadC = sad8(current->u + offset,
944 :     reference->u + offset, stride);
945 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
946 : edgomez 963 sadC += sad8(current->v + offset,
947 :     reference->v + offset, stride);
948 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
949 :     return 1;
950 : chl 181
951 : edgomez 851 } else {
952 : edgomez 963 uint32_t sadC = sad16(current->u + 2*offset,
953 :     reference->u + 2*offset, stride, 256*4096);
954 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
955 : edgomez 963 sadC += sad16(current->v + 2*offset,
956 :     reference->v + 2*offset, stride, 256*4096);
957 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
958 :     return 1;
959 : chl 181 }
960 : edgomez 851 }
961 : edgomez 504
962 : edgomez 851 static __inline void
963 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
964 :     {
965 :     pMB->mode = MODE_NOT_CODED;
966 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
967 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
968 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
969 : chl 181 }
970 :    
971 : edgomez 851 bool
972 :     MotionEstimation(MBParam * const pParam,
973 :     FRAMEINFO * const current,
974 :     FRAMEINFO * const reference,
975 :     const IMAGE * const pRefH,
976 :     const IMAGE * const pRefV,
977 :     const IMAGE * const pRefHV,
978 :     const uint32_t iLimit)
979 : chl 181 {
980 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
981 :     const IMAGE *const pCurrent = &current->image;
982 :     const IMAGE *const pRef = &reference->image;
983 : chl 181
984 : edgomez 851 uint32_t mb_width = pParam->mb_width;
985 :     uint32_t mb_height = pParam->mb_height;
986 :     const uint32_t iEdgedWidth = pParam->edged_width;
987 : suxen_drol 890 const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
988 : chl 181
989 : edgomez 851 uint32_t x, y;
990 :     uint32_t iIntra = 0;
991 : edgomez 953 int32_t quant = current->quant, sad00;
992 :     int skip_thresh = \
993 :     INITIAL_SKIP_THRESH * \
994 :     (current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \
995 :     (current->vop_flags & XVID_VOP_MODEDECISION_BITS ? 2:1);
996 : chl 181
997 : edgomez 851 // some pre-initialized thingies for SearchP
998 :     int32_t temp[8];
999 :     VECTOR currentMV[5];
1000 :     VECTOR currentQMV[5];
1001 :     int32_t iMinSAD[5];
1002 : edgomez 962 DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE);
1003 : edgomez 851 SearchData Data;
1004 :     memset(&Data, 0, sizeof(SearchData));
1005 :     Data.iEdgedWidth = iEdgedWidth;
1006 :     Data.currentMV = currentMV;
1007 :     Data.currentQMV = currentQMV;
1008 :     Data.iMinSAD = iMinSAD;
1009 :     Data.temp = temp;
1010 :     Data.iFcode = current->fcode;
1011 :     Data.rounding = pParam->m_rounding_type;
1012 : edgomez 963 Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0);
1013 : edgomez 949 Data.chroma = MotionFlags & XVID_ME_CHROMA16;
1014 : edgomez 963 Data.rrv = (current->vop_flags & XVID_VOP_REDUCED ? 1:0);
1015 : edgomez 962 Data.dctSpace = dct_space;
1016 : edgomez 195
1017 : edgomez 949 if ((current->vop_flags & XVID_VOP_REDUCED)) {
1018 : edgomez 851 mb_width = (pParam->width + 31) / 32;
1019 :     mb_height = (pParam->height + 31) / 32;
1020 :     Data.qpel = 0;
1021 :     }
1022 : chl 181
1023 : edgomez 851 Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose)
1024 :     if (sadInit) (*sadInit) ();
1025 : chl 181
1026 : edgomez 851 for (y = 0; y < mb_height; y++) {
1027 :     for (x = 0; x < mb_width; x++) {
1028 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1029 : chl 181
1030 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1031 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1032 :     pRef->y + (x + y * iEdgedWidth) * 16,
1033 :     pParam->edged_width, pMB->sad8 );
1034 : chl 181
1035 : edgomez 851 else pMB->sad16 =
1036 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1037 :     pRef->y + (x + y * iEdgedWidth) * 32,
1038 :     pParam->edged_width, pMB->sad8 );
1039 : chl 181
1040 : edgomez 851 if (Data.chroma) {
1041 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1042 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1043 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1044 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1045 :     pMB->sad16 += Data.temp[7];
1046 :     }
1047 :    
1048 :     sad00 = pMB->sad16;
1049 :    
1050 : edgomez 953 if (pMB->dquant != 0) {
1051 :     quant += DQtab[pMB->dquant];
1052 :     if (quant > 31) quant = 31;
1053 :     else if (quant < 1) quant = 1;
1054 :     }
1055 :    
1056 :     pMB->quant = current->quant;
1057 :    
1058 : edgomez 851 //initial skip decision
1059 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1060 : edgomez 949 if (!(current->vol_flags & XVID_VOL_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1061 : edgomez 953 if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh)
1062 : edgomez 851 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1063 :     SkipMacroblockP(pMB, sad00);
1064 :     continue;
1065 :     }
1066 :     }
1067 :    
1068 :     SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1069 : suxen_drol 890 y, MotionFlags, current->vol_flags, pMB->quant,
1070 : edgomez 851 &Data, pParam, pMBs, reference->mbs,
1071 : edgomez 949 current->vop_flags & XVID_VOP_INTER4V, pMB);
1072 : edgomez 851
1073 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
1074 : edgomez 953 if (!(current->vol_flags & XVID_VOL_GMC || current->vop_flags & XVID_VOP_MODEDECISION_BITS)) {
1075 : suxen_drol 926 if ( pMB->dquant == 0 && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) {
1076 : edgomez 953 if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) )
1077 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv))
1078 :     SkipMacroblockP(pMB, sad00);
1079 : chl 181 }
1080 :     }
1081 : edgomez 851 if (pMB->mode == MODE_INTRA)
1082 :     if (++iIntra > iLimit) return 1;
1083 : chl 181 }
1084 :     }
1085 : edgomez 851
1086 : edgomez 949 if (current->vol_flags & XVID_VOL_GMC ) /* GMC only for S(GMC)-VOPs */
1087 : edgomez 851 current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1088 : edgomez 963
1089 : edgomez 851 return 0;
1090 : chl 181 }
1091 :    
1092 :    
1093 : edgomez 851 static __inline int
1094 :     make_mask(const VECTOR * const pmv, const int i)
1095 : chl 96 {
1096 : edgomez 851 int mask = 255, j;
1097 :     for (j = 0; j < i; j++) {
1098 :     if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
1099 :     if (pmv[i].x == pmv[j].x) {
1100 :     if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1101 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1102 :     } else
1103 :     if (pmv[i].y == pmv[j].y) {
1104 :     if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1105 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1106 :     }
1107 :     }
1108 :     return mask;
1109 :     }
1110 : edgomez 195
1111 : edgomez 851 static __inline void
1112 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1113 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1114 :     {
1115 : chl 96
1116 : edgomez 851 //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1117 :     if (rrv) { iWcount /= 2; iHcount /= 2; }
1118 : edgomez 195
1119 : edgomez 851 if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
1120 :     pmv[5].x = EVEN(pmv[3].x);
1121 :     pmv[5].y = EVEN(pmv[3].y);
1122 :     } else pmv[5].x = pmv[5].y = 0;
1123 : chl 96
1124 : edgomez 851 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1125 :     else pmv[3].x = pmv[3].y = 0;
1126 : chl 96
1127 : edgomez 851 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1128 :     else pmv[4].x = pmv[4].y = 0;
1129 : Isibaar 3
1130 : edgomez 851 // [1] median prediction
1131 :     pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1132 : edgomez 195
1133 : 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
1134 : edgomez 195
1135 : edgomez 851 pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1136 :     pmv[2].y = EVEN(prevMB->mvs[0].y);
1137 :    
1138 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1139 :     pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1140 :     pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1141 :     } else pmv[6].x = pmv[6].y = 0;
1142 :    
1143 :     if (rrv) {
1144 :     int i;
1145 :     for (i = 0; i < 7; i++) {
1146 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1147 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1148 :     }
1149 :     }
1150 : Isibaar 3 }
1151 :    
1152 : edgomez 851 static int
1153 :     ModeDecision(const uint32_t iQuant, SearchData * const Data,
1154 :     int inter4v,
1155 :     MACROBLOCK * const pMB,
1156 :     const MACROBLOCK * const pMBs,
1157 :     const int x, const int y,
1158 :     const MBParam * const pParam,
1159 :     const uint32_t MotionFlags,
1160 : suxen_drol 890 const uint32_t VopFlags)
1161 : edgomez 851 {
1162 : Isibaar 3
1163 : edgomez 851 int mode = MODE_INTER;
1164 : chl 96
1165 : edgomez 949 if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision
1166 : edgomez 851 int sad;
1167 :     int InterBias = MV16_INTER_BIAS;
1168 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1169 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1170 : edgomez 963 mode = MODE_INTER;
1171 :     sad = Data->iMinSAD[0];
1172 : edgomez 851 } else {
1173 :     mode = MODE_INTER4V;
1174 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1175 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1176 :     Data->iMinSAD[0] = sad;
1177 :     }
1178 : chl 326
1179 : edgomez 851 /* intra decision */
1180 : Isibaar 3
1181 : edgomez 851 if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work
1182 :     if (y != 0)
1183 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1184 :     if (x != 0)
1185 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1186 : Isibaar 3
1187 : edgomez 851 if (Data->chroma) InterBias += 50; // to compensate bigger SAD
1188 :     if (Data->rrv) InterBias *= 4;
1189 : edgomez 195
1190 : edgomez 851 if (InterBias < pMB->sad16) {
1191 :     int32_t deviation;
1192 :     if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
1193 :     else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
1194 :     dev16(Data->Cur+8, Data->iEdgedWidth) +
1195 :     dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) +
1196 :     dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth);
1197 : edgomez 195
1198 : edgomez 963 if (deviation < (sad - InterBias)) return MODE_INTRA;
1199 : edgomez 851 }
1200 :     return mode;
1201 : Isibaar 3
1202 : edgomez 851 } else {
1203 : edgomez 195
1204 : edgomez 851 int bits, intra, i;
1205 :     VECTOR backup[5], *v;
1206 :     Data->lambda16 = iQuant;
1207 : edgomez 949 Data->lambda8 = (pParam->vol_flags & XVID_VOL_MPEGQUANT)?1:0;
1208 : chl 181
1209 : edgomez 851 v = Data->qpel ? Data->currentQMV : Data->currentMV;
1210 :     for (i = 0; i < 5; i++) {
1211 :     Data->iMinSAD[i] = 256*4096;
1212 :     backup[i] = v[i];
1213 :     }
1214 : Isibaar 3
1215 : edgomez 851 bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1216 :     if (bits == 0) return MODE_INTER; // quick stop
1217 : Isibaar 3
1218 : edgomez 851 if (inter4v) {
1219 : edgomez 963 int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1220 :     if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; }
1221 : edgomez 851 }
1222 : edgomez 195
1223 : Isibaar 3
1224 : edgomez 851 intra = CountMBBitsIntra(Data);
1225 : Isibaar 3
1226 : edgomez 851 if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; }
1227 :    
1228 :     return mode;
1229 : edgomez 195 }
1230 : edgomez 851 }
1231 : Isibaar 3
1232 : edgomez 851 static void
1233 :     SearchP(const IMAGE * const pRef,
1234 :     const uint8_t * const pRefH,
1235 :     const uint8_t * const pRefV,
1236 :     const uint8_t * const pRefHV,
1237 :     const IMAGE * const pCur,
1238 :     const int x,
1239 :     const int y,
1240 :     const uint32_t MotionFlags,
1241 : suxen_drol 890 const uint32_t VopFlags,
1242 : edgomez 851 const uint32_t iQuant,
1243 :     SearchData * const Data,
1244 :     const MBParam * const pParam,
1245 :     const MACROBLOCK * const pMBs,
1246 :     const MACROBLOCK * const prevMBs,
1247 :     int inter4v,
1248 :     MACROBLOCK * const pMB)
1249 :     {
1250 : Isibaar 3
1251 : edgomez 851 int i, iDirection = 255, mask, threshA;
1252 :     VECTOR pmv[7];
1253 : Isibaar 3
1254 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1255 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1256 : edgomez 195
1257 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1258 : Isibaar 3
1259 : edgomez 851 Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache
1260 :     i = Data->rrv ? 2 : 1;
1261 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1262 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1263 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1264 : chl 326
1265 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1266 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1267 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1268 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1269 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1270 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1271 : edgomez 195
1272 : edgomez 851 Data->lambda16 = lambda_vec16[iQuant];
1273 :     Data->lambda8 = lambda_vec8[iQuant];
1274 :     Data->qpel_precision = 0;
1275 : edgomez 195
1276 : suxen_drol 926 if (pMB->dquant != 0) inter4v = 0;
1277 : edgomez 195
1278 : edgomez 963 memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1279 : chl 169
1280 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1281 :     else Data->predMV = pmv[0];
1282 : Isibaar 3
1283 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1284 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1285 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1286 :     Data->iMinSAD[2] = pMB->sad8[1];
1287 :     Data->iMinSAD[3] = pMB->sad8[2];
1288 :     Data->iMinSAD[4] = pMB->sad8[3];
1289 : chl 169
1290 : edgomez 949 if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) || (x | y)) {
1291 : edgomez 851 threshA = Data->temp[0]; // that's where we keep this SAD atm
1292 :     if (threshA < 512) threshA = 512;
1293 :     else if (threshA > 1024) threshA = 1024;
1294 :     } else
1295 :     threshA = 512;
1296 : chl 169
1297 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1298 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1299 : chl 169
1300 : edgomez 851 if (!Data->rrv) {
1301 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1302 :     else CheckCandidate = CheckCandidate16no4v; //for extra speed
1303 :     } else CheckCandidate = CheckCandidate32;
1304 : chl 169
1305 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1306 : chl 169
1307 : edgomez 851 for (i = 1; i < 7; i++) {
1308 :     if (!(mask = make_mask(pmv, i)) ) continue;
1309 :     CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1310 :     if (Data->iMinSAD[0] <= threshA) break;
1311 :     }
1312 : chl 169
1313 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1314 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1315 :     (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1316 : edgomez 949 if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) inter4v = 0; }
1317 : edgomez 851 else {
1318 : Isibaar 3
1319 : edgomez 851 MainSearchFunc * MainSearchPtr;
1320 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1321 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1322 : edgomez 851 else MainSearchPtr = DiamondSearch;
1323 : Isibaar 3
1324 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1325 : Isibaar 3
1326 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1327 :     note that this search is/might be done in halfpel positions,
1328 :     which makes it more different than the diamond above */
1329 : chl 169
1330 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH16) {
1331 : edgomez 851 int32_t bSAD;
1332 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1333 :     if (Data->rrv) {
1334 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1335 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1336 :     }
1337 :     if (!(MVequal(startMV, backupMV))) {
1338 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1339 : edgomez 195
1340 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1341 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1342 :     if (bSAD < Data->iMinSAD[0]) {
1343 :     Data->currentMV[0] = backupMV;
1344 :     Data->iMinSAD[0] = bSAD; }
1345 :     }
1346 : chl 169
1347 : edgomez 851 backupMV = Data->currentMV[0];
1348 :     startMV.x = startMV.y = 1;
1349 :     if (!(MVequal(startMV, backupMV))) {
1350 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1351 : chl 169
1352 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1353 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1354 :     if (bSAD < Data->iMinSAD[0]) {
1355 :     Data->currentMV[0] = backupMV;
1356 :     Data->iMinSAD[0] = bSAD; }
1357 : edgomez 195 }
1358 : edgomez 851 }
1359 :     }
1360 : edgomez 195
1361 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE16)
1362 :     if ((!(MotionFlags & XVID_ME_HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant)
1363 : edgomez 851 SubpelRefine(Data);
1364 : edgomez 195
1365 : edgomez 851 for(i = 0; i < 5; i++) {
1366 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1367 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1368 :     }
1369 : Isibaar 3
1370 : edgomez 953 if (MotionFlags & XVID_ME_QUARTERPELREFINE16) {
1371 :    
1372 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1373 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1374 :    
1375 : edgomez 949 if ((!(MotionFlags & XVID_ME_QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) {
1376 : edgomez 851 Data->qpel_precision = 1;
1377 :     SubpelRefine(Data);
1378 :     }
1379 : edgomez 953 }
1380 : Isibaar 3
1381 : edgomez 949 if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0;
1382 : Isibaar 3
1383 : edgomez 949 if (inter4v && (!(VopFlags & XVID_VOP_MODEDECISION_BITS) ||
1384 :     (!(MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS)) || (!(MotionFlags & XVID_ME_HALFPELREFINE8_BITS)) ||
1385 :     ((!(MotionFlags & XVID_ME_EXTSEARCH_BITS)) && (!(MotionFlags&XVID_ME_EXTSEARCH8)) ))) {
1386 : edgomez 851 // if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop
1387 : Isibaar 3
1388 : edgomez 851 SearchData Data8;
1389 :     memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data
1390 : Isibaar 3
1391 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1392 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1393 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1394 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1395 : chl 181
1396 : edgomez 949 if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_BITS))) {
1397 : edgomez 851 // chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it
1398 :     int sumx = 0, sumy = 0;
1399 :     const int div = 1 + Data->qpel;
1400 :     const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs;
1401 : Isibaar 3
1402 : edgomez 851 for (i = 0; i < 4; i++) {
1403 :     sumx += mv[i].x / div;
1404 :     sumy += mv[i].y / div;
1405 :     }
1406 : chl 259
1407 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1408 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1409 :     }
1410 :     }
1411 : edgomez 195
1412 : suxen_drol 890 inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, VopFlags);
1413 : edgomez 851
1414 :     if (Data->rrv) {
1415 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1416 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1417 : Isibaar 3 }
1418 :    
1419 : edgomez 851 if (inter4v == MODE_INTER) {
1420 :     pMB->mode = MODE_INTER;
1421 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1422 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1423 : Isibaar 3
1424 : edgomez 851 if(Data->qpel) {
1425 :     pMB->qmvs[0] = pMB->qmvs[1]
1426 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1427 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1428 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1429 :     } else {
1430 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1431 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1432 : Isibaar 3 }
1433 :    
1434 : edgomez 851 } else if (inter4v == MODE_INTER4V) {
1435 :     pMB->mode = MODE_INTER4V;
1436 :     pMB->sad16 = Data->iMinSAD[0];
1437 :     } else { // INTRA mode
1438 :     SkipMacroblockP(pMB, 0); // not skip, but similar enough
1439 :     pMB->mode = MODE_INTRA;
1440 :     }
1441 : edgomez 195
1442 : edgomez 851 }
1443 :    
1444 :     static void
1445 :     Search8(const SearchData * const OldData,
1446 :     const int x, const int y,
1447 :     const uint32_t MotionFlags,
1448 :     const MBParam * const pParam,
1449 :     MACROBLOCK * const pMB,
1450 :     const MACROBLOCK * const pMBs,
1451 :     const int block,
1452 :     SearchData * const Data)
1453 :     {
1454 :     int i = 0;
1455 :     Data->iMinSAD = OldData->iMinSAD + 1 + block;
1456 :     Data->currentMV = OldData->currentMV + 1 + block;
1457 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1458 :    
1459 :     if(Data->qpel) {
1460 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1461 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1462 :     Data->predMV, Data->iFcode, 0, 0);
1463 :     } else {
1464 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1465 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1466 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1467 : Isibaar 3 }
1468 :    
1469 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1470 : Isibaar 3
1471 : edgomez 949 if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) {
1472 : Isibaar 3
1473 : edgomez 978 if (Data->rrv) i = 16; else i = 8;
1474 : Isibaar 3
1475 : edgomez 978 Data->RefP[0] = OldData->RefP[0] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1476 :     Data->RefP[1] = OldData->RefP[1] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1477 :     Data->RefP[2] = OldData->RefP[2] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1478 :     Data->RefP[3] = OldData->RefP[3] + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1479 :    
1480 :     Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1));
1481 : edgomez 851 Data->qpel_precision = 0;
1482 : Isibaar 3
1483 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1484 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1485 : Isibaar 3
1486 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1487 :     else CheckCandidate = CheckCandidate16no4v;
1488 : Isibaar 3
1489 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_BITS))) {
1490 : edgomez 851 int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1491 : Isibaar 3
1492 : edgomez 851 MainSearchFunc *MainSearchPtr;
1493 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch;
1494 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1495 : edgomez 851 else MainSearchPtr = DiamondSearch;
1496 : Isibaar 3
1497 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1498 : Isibaar 3
1499 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1500 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1501 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1502 :     }
1503 :     }
1504 : edgomez 195
1505 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8) {
1506 : edgomez 851 int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1507 : edgomez 195
1508 : edgomez 851 SubpelRefine(Data); // perform halfpel refine of current best vector
1509 : Isibaar 3
1510 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1511 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1512 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1513 :     }
1514 :     }
1515 : Isibaar 3
1516 : edgomez 949 if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) {
1517 : edgomez 851 Data->qpel_precision = 1;
1518 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1519 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1520 :     SubpelRefine(Data);
1521 :     }
1522 :     }
1523 : edgomez 195
1524 : edgomez 851 if (Data->rrv) {
1525 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1526 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1527 :     }
1528 :    
1529 :     if(Data->qpel) {
1530 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1531 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1532 :     pMB->qmvs[block] = *Data->currentQMV;
1533 : edgomez 195 } else {
1534 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1535 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1536 : edgomez 78 }
1537 : edgomez 851
1538 :     pMB->mvs[block] = *Data->currentMV;
1539 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1540 : Isibaar 3 }
1541 :    
1542 : edgomez 851 /* motion estimation for B-frames */
1543 : chl 345
1544 : edgomez 851 static __inline VECTOR
1545 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1546 :     {
1547 :     /* the stupidiest function ever */
1548 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1549 :     }
1550 : chl 345
1551 : edgomez 851 static void __inline
1552 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1553 :     const uint32_t iWcount,
1554 :     const MACROBLOCK * const pMB,
1555 :     const uint32_t mode_curr)
1556 : chl 345 {
1557 :    
1558 : edgomez 851 // [0] is prediction
1559 :     pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1560 : chl 345
1561 : edgomez 851 pmv[1].x = pmv[1].y = 0; // [1] is zero
1562 : chl 345
1563 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1564 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1565 : chl 345
1566 : edgomez 851 if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1567 :     pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1568 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1569 :     } else pmv[3].x = pmv[3].y = 0;
1570 : chl 345
1571 : edgomez 851 if (y != 0) {
1572 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1573 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1574 :     } else pmv[4].x = pmv[4].y = 0;
1575 : chl 345
1576 : edgomez 851 if (x != 0) {
1577 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1578 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1579 :     } else pmv[5].x = pmv[5].y = 0;
1580 : chl 345
1581 : edgomez 851 if (x != 0 && y != 0) {
1582 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1583 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1584 :     } else pmv[6].x = pmv[6].y = 0;
1585 :     }
1586 : chl 345
1587 :    
1588 : edgomez 851 /* search backward or forward */
1589 :     static void
1590 :     SearchBF( const IMAGE * const pRef,
1591 :     const uint8_t * const pRefH,
1592 :     const uint8_t * const pRefV,
1593 :     const uint8_t * const pRefHV,
1594 :     const IMAGE * const pCur,
1595 :     const int x, const int y,
1596 :     const uint32_t MotionFlags,
1597 :     const uint32_t iFcode,
1598 :     const MBParam * const pParam,
1599 :     MACROBLOCK * const pMB,
1600 :     const VECTOR * const predMV,
1601 :     int32_t * const best_sad,
1602 :     const int32_t mode_current,
1603 :     SearchData * const Data)
1604 :     {
1605 : chl 345
1606 : edgomez 851 int i, iDirection = 255, mask;
1607 :     VECTOR pmv[7];
1608 :     MainSearchFunc *MainSearchPtr;
1609 :     *Data->iMinSAD = MV_MAX_ERROR;
1610 :     Data->iFcode = iFcode;
1611 :     Data->qpel_precision = 0;
1612 :     Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache
1613 : chl 345
1614 : edgomez 978 Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16;
1615 :     Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16;
1616 :     Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16;
1617 :     Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16;
1618 :     Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8;
1619 :     Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8;
1620 : chl 345
1621 : edgomez 851 Data->predMV = *predMV;
1622 : chl 345
1623 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1624 :     pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1625 : chl 345
1626 : edgomez 851 pmv[0] = Data->predMV;
1627 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1628 : chl 345
1629 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1630 : chl 345
1631 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1632 :     CheckCandidate = CheckCandidate16no4v;
1633 : chl 345
1634 : edgomez 851 // main loop. checking all predictions
1635 :     for (i = 0; i < 7; i++) {
1636 :     if (!(mask = make_mask(pmv, i)) ) continue;
1637 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1638 : chl 345 }
1639 :    
1640 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1641 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1642 : edgomez 851 else MainSearchPtr = DiamondSearch;
1643 : chl 345
1644 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1645 : chl 345
1646 : edgomez 851 SubpelRefine(Data);
1647 : chl 345
1648 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1649 :     Data->currentQMV->x = 2*Data->currentMV->x;
1650 :     Data->currentQMV->y = 2*Data->currentMV->y;
1651 :     Data->qpel_precision = 1;
1652 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1653 :     pParam->width, pParam->height, iFcode, 1, 0);
1654 :     SubpelRefine(Data);
1655 :     }
1656 : chl 345
1657 : edgomez 851 // three bits are needed to code backward mode. four for forward
1658 : Isibaar 3
1659 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1660 :     else *Data->iMinSAD += 3 * Data->lambda16;
1661 : edgomez 195
1662 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1663 :     *best_sad = *Data->iMinSAD;
1664 :     pMB->mode = mode_current;
1665 :     if (Data->qpel) {
1666 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1667 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1668 :     if (mode_current == MODE_FORWARD)
1669 :     pMB->qmvs[0] = *Data->currentQMV;
1670 :     else
1671 :     pMB->b_qmvs[0] = *Data->currentQMV;
1672 :     } else {
1673 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1674 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1675 :     }
1676 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1677 :     else pMB->b_mvs[0] = *Data->currentMV;
1678 :     }
1679 : edgomez 195
1680 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1681 :     else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
1682 : Isibaar 3 }
1683 :    
1684 : edgomez 851 static void
1685 :     SkipDecisionB(const IMAGE * const pCur,
1686 :     const IMAGE * const f_Ref,
1687 :     const IMAGE * const b_Ref,
1688 :     MACROBLOCK * const pMB,
1689 :     const uint32_t x, const uint32_t y,
1690 :     const SearchData * const Data)
1691 :     {
1692 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1693 :     int32_t sum;
1694 :     const int div = 1 + Data->qpel;
1695 :     int k;
1696 :     const uint32_t stride = Data->iEdgedWidth/2;
1697 :     //this is not full chroma compensation, only it's fullpel approximation. should work though
1698 : Isibaar 3
1699 : edgomez 851 for (k = 0; k < 4; k++) {
1700 :     dy += Data->directmvF[k].y / div;
1701 : edgomez 963 dx += Data->directmvF[k].x / div;
1702 :     b_dy += Data->directmvB[k].y / div;
1703 :     b_dx += Data->directmvB[k].x / div;
1704 : edgomez 851 }
1705 : Isibaar 3
1706 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1707 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1708 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1709 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1710 : Isibaar 3
1711 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1712 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1713 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1714 :     stride);
1715 : Isibaar 3
1716 : edgomez 851 if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip
1717 : Isibaar 3
1718 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1719 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1720 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1721 :     stride);
1722 : edgomez 195
1723 : edgomez 963 if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1724 :     pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1725 : edgomez 978 for (k = 0; k < 4; k++) {
1726 : edgomez 963 pMB->qmvs[k] = pMB->mvs[k];
1727 :     pMB->b_qmvs[k] = pMB->b_mvs[k];
1728 :     }
1729 :     }
1730 : edgomez 851 }
1731 : Isibaar 3
1732 : edgomez 851 static __inline uint32_t
1733 :     SearchDirect(const IMAGE * const f_Ref,
1734 :     const uint8_t * const f_RefH,
1735 :     const uint8_t * const f_RefV,
1736 :     const uint8_t * const f_RefHV,
1737 :     const IMAGE * const b_Ref,
1738 :     const uint8_t * const b_RefH,
1739 :     const uint8_t * const b_RefV,
1740 :     const uint8_t * const b_RefHV,
1741 :     const IMAGE * const pCur,
1742 :     const int x, const int y,
1743 :     const uint32_t MotionFlags,
1744 :     const int32_t TRB, const int32_t TRD,
1745 :     const MBParam * const pParam,
1746 :     MACROBLOCK * const pMB,
1747 :     const MACROBLOCK * const b_mb,
1748 :     int32_t * const best_sad,
1749 :     SearchData * const Data)
1750 : Isibaar 3
1751 : edgomez 851 {
1752 :     int32_t skip_sad;
1753 :     int k = (x + Data->iEdgedWidth*y) * 16;
1754 :     MainSearchFunc *MainSearchPtr;
1755 : Isibaar 3
1756 : edgomez 851 *Data->iMinSAD = 256*4096;
1757 : edgomez 978 Data->RefP[0] = f_Ref->y + k;
1758 :     Data->RefP[2] = f_RefH + k;
1759 :     Data->RefP[1] = f_RefV + k;
1760 :     Data->RefP[3] = f_RefHV + k;
1761 :     Data->b_RefP[0] = b_Ref->y + k;
1762 :     Data->b_RefP[2] = b_RefH + k;
1763 :     Data->b_RefP[1] = b_RefV + k;
1764 :     Data->b_RefP[3] = b_RefHV + k;
1765 :     Data->RefP[4] = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1766 :     Data->RefP[5] = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1767 :     Data->b_RefP[4] = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1768 :     Data->b_RefP[5] = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1769 : edgomez 195
1770 : edgomez 851 k = Data->qpel ? 4 : 2;
1771 :     Data->max_dx = k * (pParam->width - x * 16);
1772 :     Data->max_dy = k * (pParam->height - y * 16);
1773 :     Data->min_dx = -k * (16 + x * 16);
1774 :     Data->min_dy = -k * (16 + y * 16);
1775 : chl 181
1776 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1777 :     Data->qpel_precision = 0;
1778 : edgomez 170
1779 : edgomez 851 for (k = 0; k < 4; k++) {
1780 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1781 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1782 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1783 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1784 : Isibaar 3
1785 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1786 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1787 :    
1788 :     *best_sad = 256*4096; // in that case, we won't use direct mode
1789 :     pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1790 :     pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1791 :     return 256*4096;
1792 :     }
1793 :     if (b_mb->mode != MODE_INTER4V) {
1794 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1795 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1796 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1797 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1798 :     break;
1799 :     }
1800 : edgomez 195 }
1801 : Isibaar 3
1802 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1803 : Isibaar 3
1804 : edgomez 851 CheckCandidate(0, 0, 255, &k, Data);
1805 : edgomez 195
1806 : edgomez 851 // initial (fast) skip decision
1807 :     if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) {
1808 :     //possible skip
1809 :     if (Data->chroma) {
1810 :     pMB->mode = MODE_DIRECT_NONE_MV;
1811 :     return *Data->iMinSAD; // skip.
1812 :     } else {
1813 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1814 :     if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1815 :     }
1816 : Isibaar 3 }
1817 :    
1818 : edgomez 953 *Data->iMinSAD += Data->lambda16;
1819 : edgomez 851 skip_sad = *Data->iMinSAD;
1820 : edgomez 195
1821 : edgomez 851 // DIRECT MODE DELTA VECTOR SEARCH.
1822 :     // This has to be made more effective, but at the moment I'm happy it's running at all
1823 : Isibaar 3
1824 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1825 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1826 : edgomez 851 else MainSearchPtr = DiamondSearch;
1827 : Isibaar 3
1828 : edgomez 851 MainSearchPtr(0, 0, Data, 255);
1829 : Isibaar 3
1830 : edgomez 851 SubpelRefine(Data);
1831 : chl 181
1832 : edgomez 851 *best_sad = *Data->iMinSAD;
1833 : chl 181
1834 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1835 :     else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1836 : chl 181
1837 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1838 : edgomez 195
1839 : edgomez 851 for (k = 0; k < 4; k++) {
1840 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1841 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1842 :     ? Data->directmvB[k].x
1843 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1844 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1845 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1846 :     ? Data->directmvB[k].y
1847 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1848 :     if (Data->qpel) {
1849 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1850 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1851 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1852 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1853 :     }
1854 : edgomez 195
1855 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1856 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1857 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1858 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1859 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1860 :     break;
1861 :     }
1862 : edgomez 78 }
1863 : edgomez 851 return skip_sad;
1864 :     }
1865 : Isibaar 3
1866 : edgomez 851 static void
1867 :     SearchInterpolate(const IMAGE * const f_Ref,
1868 :     const uint8_t * const f_RefH,
1869 :     const uint8_t * const f_RefV,
1870 :     const uint8_t * const f_RefHV,
1871 :     const IMAGE * const b_Ref,
1872 :     const uint8_t * const b_RefH,
1873 :     const uint8_t * const b_RefV,
1874 :     const uint8_t * const b_RefHV,
1875 :     const IMAGE * const pCur,
1876 :     const int x, const int y,
1877 :     const uint32_t fcode,
1878 :     const uint32_t bcode,
1879 :     const uint32_t MotionFlags,
1880 :     const MBParam * const pParam,
1881 :     const VECTOR * const f_predMV,
1882 :     const VECTOR * const b_predMV,
1883 :     MACROBLOCK * const pMB,
1884 :     int32_t * const best_sad,
1885 :     SearchData * const fData)
1886 : chl 96
1887 : edgomez 851 {
1888 : chl 169
1889 : edgomez 851 int iDirection, i, j;
1890 :     SearchData bData;
1891 : chl 169
1892 : edgomez 851 fData->qpel_precision = 0;
1893 :     memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1894 :     *fData->iMinSAD = 4096*256;
1895 :     bData.currentMV++; bData.currentQMV++;
1896 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1897 : chl 169
1898 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1899 : chl 169
1900 : edgomez 978 bData.b_RefP[0] = fData->RefP[0] = f_Ref->y + i;
1901 :     bData.b_RefP[2] = fData->RefP[2] = f_RefH + i;
1902 :     bData.b_RefP[1] = fData->RefP[1] = f_RefV + i;
1903 :     bData.b_RefP[3] = fData->RefP[3] = f_RefHV + i;
1904 :     bData.RefP[0] = fData->b_RefP[0] = b_Ref->y + i;
1905 :     bData.RefP[2] = fData->b_RefP[2] = b_RefH + i;
1906 :     bData.RefP[1] = fData->b_RefP[1] = b_RefV + i;
1907 :     bData.RefP[3] = fData->b_RefP[3] = b_RefHV + i;
1908 :     bData.b_RefP[4] = fData->RefP[4] = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1909 :     bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1910 :     bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1911 :     bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1912 : chl 169
1913 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1914 :     fData->bpredMV = bData.predMV = *b_predMV;
1915 :     fData->currentMV[0] = fData->currentMV[2];
1916 : Isibaar 3
1917 : edgomez 851 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0);
1918 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0);
1919 : Isibaar 3
1920 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1921 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1922 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1923 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1924 : chl 169
1925 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1926 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1927 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1928 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1929 : Isibaar 3
1930 : edgomez 851 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1931 : chl 169
1932 : edgomez 851 //diamond
1933 :     do {
1934 :     iDirection = 255;
1935 :     // forward MV moves
1936 :     i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1937 : chl 169
1938 : edgomez 851 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1939 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1940 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1941 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1942 : edgomez 195
1943 : edgomez 851 // backward MV moves
1944 :     i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1945 :     fData->currentMV[2] = fData->currentMV[0];
1946 :     CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1947 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1948 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1949 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1950 : Isibaar 3
1951 : edgomez 851 } while (!(iDirection));
1952 : chl 140
1953 : edgomez 851 //qpel refinement
1954 :     if (fData->qpel) {
1955 :     if (*fData->iMinSAD > *best_sad + 500) return;
1956 :     CheckCandidate = CheckCandidateInt;
1957 :     fData->qpel_precision = bData.qpel_precision = 1;
1958 :     get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1959 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1960 :     fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1961 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1962 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1963 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1964 :     SubpelRefine(fData);
1965 :     if (*fData->iMinSAD > *best_sad + 300) return;
1966 :     fData->currentQMV[2] = fData->currentQMV[0];
1967 :     SubpelRefine(&bData);
1968 :     }
1969 : chl 140
1970 : edgomez 851 *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode.
1971 : Isibaar 3
1972 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
1973 :     *best_sad = *fData->iMinSAD;
1974 :     pMB->mvs[0] = fData->currentMV[0];
1975 :     pMB->b_mvs[0] = fData->currentMV[1];
1976 :     pMB->mode = MODE_INTERPOLATE;
1977 :     if (fData->qpel) {
1978 :     pMB->qmvs[0] = fData->currentQMV[0];
1979 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1980 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1981 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1982 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1983 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1984 :     } else {
1985 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1986 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1987 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1988 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1989 :     }
1990 : edgomez 78 }
1991 : edgomez 851 }
1992 : Isibaar 3
1993 : edgomez 851 void
1994 :     MotionEstimationBVOP(MBParam * const pParam,
1995 :     FRAMEINFO * const frame,
1996 :     const int32_t time_bp,
1997 :     const int32_t time_pp,
1998 :     // forward (past) reference
1999 :     const MACROBLOCK * const f_mbs,
2000 :     const IMAGE * const f_ref,
2001 :     const IMAGE * const f_refH,
2002 :     const IMAGE * const f_refV,
2003 :     const IMAGE * const f_refHV,
2004 :     // backward (future) reference
2005 :     const FRAMEINFO * const b_reference,
2006 :     const IMAGE * const b_ref,
2007 :     const IMAGE * const b_refH,
2008 :     const IMAGE * const b_refV,
2009 :     const IMAGE * const b_refHV)
2010 :     {
2011 :     uint32_t i, j;
2012 :     int32_t best_sad;
2013 :     uint32_t skip_sad;
2014 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2015 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
2016 : Isibaar 3
2017 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2018 : Isibaar 3
2019 : edgomez 851 const int32_t TRB = time_pp - time_bp;
2020 :     const int32_t TRD = time_pp;
2021 : edgomez 195
2022 : edgomez 851 // some pre-inintialized data for the rest of the search
2023 : Isibaar 3
2024 : edgomez 851 SearchData Data;
2025 :     int32_t iMinSAD;
2026 :     VECTOR currentMV[3];
2027 :     VECTOR currentQMV[3];
2028 :     int32_t temp[8];
2029 :     memset(&Data, 0, sizeof(SearchData));
2030 :     Data.iEdgedWidth = pParam->edged_width;
2031 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2032 :     Data.iMinSAD = &iMinSAD;
2033 :     Data.lambda16 = lambda_vec16[frame->quant];
2034 : edgomez 949 Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL;
2035 : edgomez 851 Data.rounding = 0;
2036 : edgomez 949 Data.chroma = frame->motion_flags & XVID_ME_CHROMA8;
2037 : edgomez 851 Data.temp = temp;
2038 : Isibaar 3
2039 : edgomez 851 Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose)
2040 :     // note: i==horizontal, j==vertical
2041 :     for (j = 0; j < pParam->mb_height; j++) {
2042 : edgomez 195
2043 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2044 : Isibaar 3
2045 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2046 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2047 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2048 : edgomez 195
2049 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2050 :     if (b_reference->coding_type != S_VOP)
2051 :     if (b_mb->mode == MODE_NOT_CODED) {
2052 :     pMB->mode = MODE_NOT_CODED;
2053 :     continue;
2054 :     }
2055 : Isibaar 3
2056 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2057 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2058 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2059 : edgomez 953 pMB->quant = frame->quant;
2060 : Isibaar 3
2061 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2062 :     and (2) sets very good predictions for forward and backward search */
2063 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2064 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2065 :     &frame->image,
2066 :     i, j,
2067 :     frame->motion_flags,
2068 :     TRB, TRD,
2069 :     pParam,
2070 :     pMB, b_mb,
2071 :     &best_sad,
2072 :     &Data);
2073 : Isibaar 3
2074 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2075 : edgomez 195
2076 : edgomez 851 // forward search
2077 :     SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2078 :     &frame->image, i, j,
2079 :     frame->motion_flags,
2080 :     frame->fcode, pParam,
2081 :     pMB, &f_predMV, &best_sad,
2082 :     MODE_FORWARD, &Data);
2083 : edgomez 195
2084 : edgomez 851 // backward search
2085 :     SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2086 :     &frame->image, i, j,
2087 :     frame->motion_flags,
2088 :     frame->bcode, pParam,
2089 :     pMB, &b_predMV, &best_sad,
2090 :     MODE_BACKWARD, &Data);
2091 :    
2092 :     // interpolate search comes last, because it uses data from forward and backward as prediction
2093 :     SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2094 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2095 :     &frame->image,
2096 :     i, j,
2097 :     frame->fcode, frame->bcode,
2098 :     frame->motion_flags,
2099 :     pParam,
2100 :     &f_predMV, &b_predMV,
2101 :     pMB, &best_sad,
2102 :     &Data);
2103 :    
2104 :     // final skip decision
2105 :     if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2106 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2107 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2108 :    
2109 :     switch (pMB->mode) {
2110 :     case MODE_FORWARD:
2111 :     f_count++;
2112 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2113 :     break;
2114 :     case MODE_BACKWARD:
2115 :     b_count++;
2116 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2117 :     break;
2118 :     case MODE_INTERPOLATE:
2119 :     i_count++;
2120 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2121 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2122 :     break;
2123 :     case MODE_DIRECT:
2124 :     case MODE_DIRECT_NO4V:
2125 :     d_count++;
2126 :     default:
2127 :     break;
2128 :     }
2129 :     }
2130 :     }
2131 : Isibaar 3 }
2132 : chl 96
2133 : edgomez 851 static __inline void
2134 :     MEanalyzeMB ( const uint8_t * const pRef,
2135 :     const uint8_t * const pCur,
2136 :     const int x,
2137 :     const int y,
2138 :     const MBParam * const pParam,
2139 :     MACROBLOCK * const pMBs,
2140 :     SearchData * const Data)
2141 : chl 96 {
2142 :    
2143 : edgomez 851 int i, mask;
2144 : edgomez 953 int quarterpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0;
2145 : edgomez 851 VECTOR pmv[3];
2146 : edgomez 953 MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2147 : chl 96
2148 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2149 : chl 96
2150 : edgomez 851 //median is only used as prediction. it doesn't have to be real
2151 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2152 :     else
2153 :     if (x == 1) //left macroblock does not have any vector now
2154 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2155 :     else if (y == 1) // top macroblock doesn't have it's vector
2156 :     Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
2157 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2158 : edgomez 195
2159 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2160 : edgomez 953 pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0);
2161 : edgomez 195
2162 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2163 : edgomez 978 Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16;
2164 : chl 96
2165 : edgomez 851 pmv[1].x = EVEN(pMB->mvs[0].x);
2166 :     pmv[1].y = EVEN(pMB->mvs[0].y);
2167 :     pmv[2].x = EVEN(Data->predMV.x);
2168 :     pmv[2].y = EVEN(Data->predMV.y);
2169 :     pmv[0].x = pmv[0].y = 0;
2170 : edgomez 195
2171 : edgomez 851 CheckCandidate32I(0, 0, 255, &i, Data);
2172 : edgomez 195
2173 : edgomez 953 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) {
2174 : chl 96
2175 : edgomez 851 if (!(mask = make_mask(pmv, 1)))
2176 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2177 :     if (!(mask = make_mask(pmv, 2)))
2178 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2179 : chl 96
2180 : edgomez 953 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) // diamond only if needed
2181 : edgomez 851 DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2182 : edgomez 953 }
2183 : edgomez 851
2184 : edgomez 953 for (i = 0; i < 4; i++) {
2185 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2186 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2187 :     MB->mode = MODE_INTER;
2188 :     MB->sad16 = Data->iMinSAD[i+1];
2189 : chl 96 }
2190 : edgomez 851 }
2191 : chl 96
2192 : edgomez 953 #define INTRA_THRESH 2400
2193 : edgomez 978 #define INTER_THRESH 1300
2194 : chl 96
2195 : edgomez 851 int
2196 :     MEanalysis( const IMAGE * const pRef,
2197 : edgomez 953 const FRAMEINFO * const Current,
2198 :     const MBParam * const pParam,
2199 :     const int maxIntra, //maximum number if non-I frames
2200 :     const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2201 :     const int bCount, // number of B frames in a row
2202 :     const int b_thresh)
2203 : edgomez 851 {
2204 :     uint32_t x, y, intra = 0;
2205 :     int sSAD = 0;
2206 :     MACROBLOCK * const pMBs = Current->mbs;
2207 :     const IMAGE * const pCurrent = &Current->image;
2208 : edgomez 953 int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh;
2209 :     int s = 0, blocks = 0;
2210 : chl 96
2211 : edgomez 851 int32_t iMinSAD[5], temp[5];
2212 :     VECTOR currentMV[5];
2213 :     SearchData Data;
2214 :     Data.iEdgedWidth = pParam->edged_width;
2215 :     Data.currentMV = currentMV;
2216 :     Data.iMinSAD = iMinSAD;
2217 :     Data.iFcode = Current->fcode;
2218 :     Data.temp = temp;
2219 :     CheckCandidate = CheckCandidate32I;
2220 : chl 96
2221 : edgomez 851 if (intraCount != 0 && intraCount < 10) // we're right after an I frame
2222 : edgomez 953 IntraThresh += 8 * (intraCount - 10) * (intraCount - 10);
2223 : edgomez 851 else
2224 :     if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2225 :     IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
2226 : chl 96
2227 : edgomez 953 InterThresh -= (350 - 8*b_thresh) * bCount;
2228 :     if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh;
2229 : chl 326
2230 : edgomez 851 if (sadInit) (*sadInit) ();
2231 : chl 96
2232 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2233 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2234 :     int i;
2235 : edgomez 953 blocks += 4;
2236 : edgomez 195
2237 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2238 : edgomez 953 else { //extrapolation of the vector found for last frame
2239 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].x =
2240 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2241 : edgomez 978 pMBs[x + y * pParam->mb_width].mvs[0].y =
2242 : edgomez 953 (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2243 :     }
2244 : edgomez 195
2245 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2246 : edgomez 195
2247 : edgomez 851 for (i = 0; i < 4; i++) {
2248 :     int dev;
2249 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2250 :     if (pMB->sad16 > IntraThresh) {
2251 :     dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2252 :     pParam->edged_width);
2253 :     if (dev + IntraThresh < pMB->sad16) {
2254 :     pMB->mode = MODE_INTRA;
2255 : edgomez 953 if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2256 : edgomez 851 }
2257 :     }
2258 : edgomez 953 if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++;
2259 :    
2260 : edgomez 851 sSAD += pMB->sad16;
2261 :     }
2262 :     }
2263 : edgomez 195 }
2264 : edgomez 953
2265 :     sSAD /= blocks;
2266 :     s = (10*s) / blocks;
2267 :    
2268 : edgomez 978 if (s > 4) sSAD += (s - 3) * (300 - 2*b_thresh); //static block - looks bad when in bframe...
2269 : edgomez 953
2270 : edgomez 851 if (sSAD > InterThresh ) return P_VOP;
2271 :     emms();
2272 :     return B_VOP;
2273 :     }
2274 : chl 96
2275 :    
2276 : edgomez 851 static WARPPOINTS
2277 :     GlobalMotionEst(const MACROBLOCK * const pMBs,
2278 :     const MBParam * const pParam,
2279 :     const FRAMEINFO * const current,
2280 :     const FRAMEINFO * const reference,
2281 :     const IMAGE * const pRefH,
2282 :     const IMAGE * const pRefV,
2283 :     const IMAGE * const pRefHV )
2284 :     {
2285 : chl 96
2286 : edgomez 851 const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2287 :     const int deltay=8;
2288 :     const int grad=512; // lower bound for deviation in MB
2289 : chl 96
2290 : edgomez 851 WARPPOINTS gmc;
2291 : chl 96
2292 : edgomez 851 uint32_t mx, my;
2293 : chl 96
2294 : edgomez 851 int MBh = pParam->mb_height;
2295 :     int MBw = pParam->mb_width;
2296 : chl 96
2297 : edgomez 851 int *MBmask= calloc(MBh*MBw,sizeof(int));
2298 :     double DtimesF[4] = { 0.,0., 0., 0. };
2299 :     double sol[4] = { 0., 0., 0., 0. };
2300 :     double a,b,c,n,denom;
2301 :     double meanx,meany;
2302 :     int num,oldnum;
2303 : edgomez 195
2304 : edgomez 978 if (!MBmask) { fprintf(stderr,"Mem error\n");
2305 :     gmc.duv[0].x= gmc.duv[0].y =
2306 :     gmc.duv[1].x= gmc.duv[1].y =
2307 : chl 871 gmc.duv[2].x= gmc.duv[2].y = 0;
2308 :     return gmc; }
2309 : edgomez 195
2310 : edgomez 851 // filter mask of all blocks
2311 :    
2312 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2313 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2314 : edgomez 851 {
2315 :     const int mbnum = mx + my * MBw;
2316 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2317 :     const VECTOR mv = pMB->mvs[0];
2318 :    
2319 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2320 :     continue;
2321 :    
2322 : edgomez 982 if ( ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2323 :     && ( (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2324 :     && ( (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2325 :     && ( (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2326 : edgomez 851 MBmask[mbnum]=1;
2327 : chl 96 }
2328 :    
2329 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2330 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2331 : edgomez 851 {
2332 :     const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx;
2333 : chl 96
2334 : edgomez 851 const int mbnum = mx + my * MBw;
2335 :     if (!MBmask[mbnum])
2336 :     continue;
2337 :    
2338 : edgomez 876 if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad )
2339 : edgomez 851 MBmask[mbnum] = 0;
2340 : edgomez 876 if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad )
2341 : edgomez 851 MBmask[mbnum] = 0;
2342 :    
2343 : edgomez 195 }
2344 : chl 96
2345 : edgomez 851 emms();
2346 : chl 96
2347 : edgomez 851 do { /* until convergence */
2348 : chl 96
2349 : edgomez 851 a = b = c = n = 0;
2350 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2351 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2352 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2353 : edgomez 851 {
2354 :     const int mbnum = mx + my * MBw;
2355 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2356 :     const VECTOR mv = pMB->mvs[0];
2357 : chl 96
2358 : edgomez 851 if (!MBmask[mbnum])
2359 :     continue;
2360 : chl 96
2361 : edgomez 851 n++;
2362 :     a += 16*mx+8;
2363 :     b += 16*my+8;
2364 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2365 : chl 96
2366 : edgomez 851 DtimesF[0] += (double)mv.x;
2367 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2368 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2369 :     DtimesF[3] += (double)mv.y;
2370 :     }
2371 : chl 96
2372 : edgomez 851 denom = a*a+b*b-c*n;
2373 : chl 96
2374 : edgomez 978 /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2375 : edgomez 851 /* D'*E*F has been calculated in the same loop as matrix */
2376 : chl 96
2377 : edgomez 851 sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2378 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2379 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2380 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2381 : chl 96
2382 : edgomez 851 sol[0] /= denom;
2383 :     sol[1] /= denom;
2384 :     sol[2] /= denom;
2385 :     sol[3] /= denom;
2386 : chl 96
2387 : edgomez 851 meanx = meany = 0.;
2388 :     oldnum = 0;
2389 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2390 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2391 : edgomez 851 {
2392 :     const int mbnum = mx + my * MBw;
2393 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2394 :     const VECTOR mv = pMB->mvs[0];
2395 : chl 96
2396 : edgomez 851 if (!MBmask[mbnum])
2397 :     continue;
2398 : chl 184
2399 : edgomez 851 oldnum++;
2400 : edgomez 982 meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x );
2401 :     meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y );
2402 : edgomez 851 }
2403 : chl 96
2404 : edgomez 851 if (4*meanx > oldnum) /* better fit than 0.25 is useless */
2405 :     meanx /= oldnum;
2406 :     else
2407 :     meanx = 0.25;
2408 : edgomez 195
2409 : edgomez 851 if (4*meany > oldnum)
2410 :     meany /= oldnum;
2411 :     else
2412 :     meany = 0.25;
2413 : chl 96
2414 : edgomez 851 /* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2415 :     fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2416 :     */
2417 :     num = 0;
2418 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2419 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2420 : edgomez 851 {
2421 :     const int mbnum = mx + my * MBw;
2422 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2423 :     const VECTOR mv = pMB->mvs[0];
2424 : chl 96
2425 : edgomez 851 if (!MBmask[mbnum])
2426 :     continue;
2427 : chl 96
2428 : edgomez 982 if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx )
2429 :     || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) )
2430 : edgomez 851 MBmask[mbnum]=0;
2431 :     else
2432 :     num++;
2433 : chl 96 }
2434 : edgomez 195
2435 : edgomez 851 } while ( (oldnum != num) && (num>=4) );
2436 : edgomez 195
2437 : edgomez 851 if (num < 4)
2438 :     {
2439 :     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;
2440 :     } else {
2441 : edgomez 195
2442 : edgomez 851 gmc.duv[0].x=(int)(sol[0]+0.5);
2443 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2444 : chl 96
2445 : edgomez 851 gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2446 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2447 : chl 96
2448 : edgomez 851 gmc.duv[2].x=0;
2449 :     gmc.duv[2].y=0;
2450 :     }
2451 :     // 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);
2452 : chl 96
2453 : edgomez 851 free(MBmask);
2454 : chl 96
2455 : edgomez 851 return gmc;
2456 : chl 96 }
2457 :    
2458 : edgomez 851 // functions which perform BITS-based search/bitcount
2459 : chl 96
2460 : edgomez 851 static int
2461 :     CountMBBitsInter(SearchData * const Data,
2462 :     const MACROBLOCK * const pMBs, const int x, const int y,
2463 :     const MBParam * const pParam,
2464 :     const uint32_t MotionFlags)
2465 : chl 96 {
2466 : edgomez 851 int i, iDirection;
2467 :     int32_t bsad[5];
2468 : chl 141
2469 : edgomez 851 CheckCandidate = CheckCandidateBits16;
2470 : chl 96
2471 : edgomez 851 if (Data->qpel) {
2472 :     for(i = 0; i < 5; i++) {
2473 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2474 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2475 :     }
2476 :     Data->qpel_precision = 1;
2477 :     CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2478 : chl 96
2479 : edgomez 851 //checking if this vector is perfect. if it is, we stop.
2480 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0)
2481 :     return 0; //quick stop
2482 : edgomez 195
2483 : edgomez 949 if (MotionFlags & (XVID_ME_HALFPELREFINE16_BITS | XVID_ME_EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search
2484 : edgomez 851 for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2485 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2486 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2487 :     Data->qpel_precision = 0;
2488 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2489 :     CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2490 :     }
2491 : edgomez 195
2492 : edgomez 851 } else { // not qpel
2493 : edgomez 195
2494 : edgomez 851 CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2495 :     //checking if this vector is perfect. if it is, we stop.
2496 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) {
2497 :     return 0; //inter
2498 :     }
2499 :     }
2500 : chl 96
2501 : edgomez 949 if (MotionFlags&XVID_ME_EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2502 : chl 96
2503 : edgomez 949 if (MotionFlags&XVID_ME_HALFPELREFINE16_BITS) SubpelRefine(Data);
2504 : chl 96
2505 : edgomez 851 if (Data->qpel) {
2506 : edgomez 949 if (MotionFlags&(XVID_ME_EXTSEARCH_BITS | XVID_ME_HALFPELREFINE16_BITS)) { // there was halfpel-precision search
2507 : edgomez 851 for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2508 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match
2509 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2510 :     }
2511 : edgomez 195
2512 : edgomez 851 // preparing for qpel-precision search
2513 :     Data->qpel_precision = 1;
2514 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2515 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
2516 :     }
2517 : edgomez 949 if (MotionFlags&XVID_ME_QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2518 : edgomez 851 }
2519 : chl 96
2520 : edgomez 949 if (MotionFlags&XVID_ME_CHECKPREDICTION_BITS) { //let's check vector equal to prediction
2521 : edgomez 851 VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2522 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2523 :     CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2524 :     }
2525 :     return Data->iMinSAD[0];
2526 :     }
2527 : chl 96
2528 :    
2529 : edgomez 851 static int
2530 :     CountMBBitsInter4v(const SearchData * const Data,
2531 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2532 :     const int x, const int y,
2533 :     const MBParam * const pParam, const uint32_t MotionFlags,
2534 :     const VECTOR * const backup)
2535 :     {
2536 : chl 96
2537 : edgomez 851 int cbp = 0, bits = 0, t = 0, i, iDirection;
2538 :     SearchData Data2, *Data8 = &Data2;
2539 :     int sumx = 0, sumy = 0;
2540 : edgomez 962 int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64;
2541 : chl 96
2542 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2543 :     CheckCandidate = CheckCandidateBits8;
2544 : chl 96
2545 : edgomez 851 for (i = 0; i < 4; i++) {
2546 :     Data8->iMinSAD = Data->iMinSAD + i + 1;
2547 :     Data8->currentMV = Data->currentMV + i + 1;
2548 :     Data8->currentQMV = Data->currentQMV + i + 1;
2549 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2550 : edgomez 978 Data8->RefP[0] = Data->RefP[0] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2551 :     Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2552 :     Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2553 :     Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2554 : edgomez 195
2555 : edgomez 851 if(Data->qpel) {
2556 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2557 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2558 :     Data8->predMV, Data8->iFcode, 0, 0);
2559 :     } else {
2560 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2561 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2562 :     Data8->predMV, Data8->iFcode, 0, 0);
2563 :     }
2564 : chl 96
2565 : 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,
2566 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2567 : chl 96
2568 : edgomez 851 *Data8->iMinSAD += t;
2569 : chl 96
2570 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2571 :     // checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far)
2572 :     if (Data8->qpel) {
2573 :     if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y))
2574 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2575 :     } else {
2576 :     if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y))
2577 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2578 :     }
2579 : edgomez 195
2580 : edgomez 851 if (Data8->qpel) {
2581 : edgomez 949 if (MotionFlags&XVID_ME_HALFPELREFINE8_BITS || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_BITS)) { // halfpixel motion search follows
2582 : edgomez 851 int32_t s = *Data8->iMinSAD;
2583 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2584 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2585 :     Data8->qpel_precision = 0;
2586 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2587 :     pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2588 : edgomez 195
2589 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2590 :     CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2591 : edgomez 195
2592 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS)
2593 : edgomez 851 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2594 : chl 96
2595 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8);
2596 : chl 96
2597 : edgomez 851 if(s > *Data8->iMinSAD) { //we have found a better match
2598 :     Data8->currentQMV->x = 2*Data8->currentMV->x;
2599 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2600 :     }
2601 : chl 96
2602 : edgomez 851 Data8->qpel_precision = 1;
2603 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2604 :     pParam->width, pParam->height, Data8->iFcode, 1, 0);
2605 : chl 96
2606 : edgomez 851 }
2607 : edgomez 949 if (MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2608 : edgomez 195
2609 : edgomez 851 } else // not qpel
2610 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement
2611 : edgomez 851
2612 :     //checking vector equal to predicion
2613 : edgomez 949 if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_BITS) {
2614 : edgomez 851 const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2615 :     if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y))
2616 :     CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2617 : chl 141 }
2618 : edgomez 195
2619 : edgomez 851 bits += *Data8->iMinSAD;
2620 :     if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V
2621 :    
2622 :     // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else
2623 :     if(Data->qpel) {
2624 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2625 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2626 :     pMB->qmvs[i] = *Data8->currentQMV;
2627 :     sumx += Data8->currentQMV->x/2;
2628 :     sumy += Data8->currentQMV->y/2;
2629 :     } else {
2630 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2631 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2632 :     sumx += Data8->currentMV->x;
2633 :     sumy += Data8->currentMV->y;
2634 : chl 141 }
2635 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2636 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2637 :     if (Data8->temp[0]) cbp |= 1 << (5 - i);
2638 : chl 141 }
2639 :    
2640 : edgomez 851 if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma
2641 :     const uint8_t * ptr;
2642 :     sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2643 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2644 : chl 141
2645 : edgomez 851 //chroma U
2646 : edgomez 978 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2647 : edgomez 851 transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2648 :     fdct(in);
2649 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2650 :     else i = quant4_inter(coeff, in, Data->lambda16);
2651 :     if (i > 0) {
2652 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2653 :     cbp |= 1 << (5 - 4);
2654 :     }
2655 : chl 141
2656 : edgomez 851 if (bits < *Data->iMinSAD) { // still possible
2657 :     //chroma V
2658 : edgomez 978 ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2659 : edgomez 851 transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2660 :     fdct(in);
2661 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2662 :     else i = quant4_inter(coeff, in, Data->lambda16);
2663 :     if (i > 0) {
2664 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2665 :     cbp |= 1 << (5 - 5);
2666 :     }
2667 : edgomez 958 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
2668 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2669 :     }
2670 : edgomez 195 }
2671 : chl 96
2672 : edgomez 851 return bits;
2673 :     }
2674 : chl 96
2675 :    
2676 : edgomez 851 static int
2677 :     CountMBBitsIntra(const SearchData * const Data)
2678 :     {
2679 :     int bits = 1; //this one is ac/dc prediction flag. always 1.
2680 : edgomez 978 int cbp = 0, i, t, dc = 1024, b_dc;
2681 : edgomez 851 const uint32_t iQuant = Data->lambda16;
2682 : edgomez 962 int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2683 : edgomez 978 uint32_t iDcScaler = get_dc_scaler(iQuant, 1);;
2684 : edgomez 195
2685 : edgomez 851 for(i = 0; i < 4; i++) {
2686 :     int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2687 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2688 :     fdct(in);
2689 : edgomez 978 b_dc = in[0];
2690 :     in[0] -= dc;
2691 :     dc = b_dc;
2692 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2693 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2694 : chl 141
2695 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;;
2696 :     Data->temp[i] = t;
2697 :     if (t != 0) cbp |= 1 << (5 - i);
2698 :     if (bits >= Data->iMinSAD[0]) break;
2699 :     }
2700 : chl 181
2701 : edgomez 851 if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma
2702 : edgomez 978 iDcScaler = get_dc_scaler(iQuant, 0);
2703 : edgomez 851 //chroma U
2704 :     transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2705 :     fdct(in);
2706 :     in[0] -= 1024;
2707 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2708 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2709 : chl 96
2710 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2711 :     if (t != 0) cbp |= 1 << (5 - 4);
2712 : chl 141
2713 : edgomez 851 if (bits < Data->iMinSAD[0]) {
2714 :     //chroma V
2715 :     transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2716 :     fdct(in);
2717 :     in[0] -= 1024;
2718 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2719 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2720 : chl 96
2721 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2722 :     if (t != 0) cbp |= 1 << (5 - 5);
2723 : chl 96
2724 : edgomez 958 bits += xvid_cbpy_tab[cbp>>2].len;
2725 : edgomez 953 bits += mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2726 : chl 96 }
2727 :     }
2728 : edgomez 851 return bits;
2729 : chl 96 }

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