[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 851 - (view) (download)
Original Path: trunk/xvidcore/src/motion/motion_est.c

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 851 #define INITIAL_SKIP_THRESH (10)
50 :     #define FINAL_SKIP_THRESH (50)
51 :     #define MAX_SAD00_FOR_SKIP (20)
52 :     #define MAX_CHROMA_SAD_FOR_SKIP (22)
53 : Isibaar 3
54 : edgomez 851 #define CHECK_CANDIDATE(X,Y,D) { \
55 :     CheckCandidate((X),(Y), (D), &iDirection, data ); }
56 : Isibaar 3
57 : edgomez 851 static __inline uint32_t
58 :     d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
59 :     {
60 :     int xb, yb;
61 :     x = qpel ? x<<1 : x;
62 :     y = qpel ? y<<1 : y;
63 :     if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
64 : chl 141
65 : edgomez 851 x -= pred.x;
66 :     y -= pred.y;
67 : chl 141
68 : edgomez 851 if (x) {
69 :     x = ABS(x);
70 :     x += (1 << (iFcode - 1)) - 1;
71 :     x >>= (iFcode - 1);
72 :     if (x > 32) x = 32;
73 :     xb = mvtab[x] + iFcode;
74 :     } else xb = 1;
75 : chl 141
76 : edgomez 851 if (y) {
77 :     y = ABS(y);
78 :     y += (1 << (iFcode - 1)) - 1;
79 :     y >>= (iFcode - 1);
80 :     if (y > 32) y = 32;
81 :     yb = mvtab[y] + iFcode;
82 :     } else yb = 1;
83 :     return xb + yb;
84 :     }
85 : chl 141
86 : edgomez 851 static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data)
87 :     {
88 :     int sad;
89 :     const uint32_t stride = data->iEdgedWidth/2;
90 :     uint8_t * f_refu = data->RefQ,
91 :     * f_refv = data->RefQ + 8,
92 :     * b_refu = data->RefQ + 16,
93 :     * b_refv = data->RefQ + 24;
94 : Isibaar 3
95 : edgomez 851 switch (((fx & 1) << 1) | (fy & 1)) {
96 :     case 0:
97 :     fx = fx / 2; fy = fy / 2;
98 :     f_refu = (uint8_t*)data->RefCU + fy * stride + fx, stride;
99 :     f_refv = (uint8_t*)data->RefCV + fy * stride + fx, stride;
100 :     break;
101 :     case 1:
102 :     fx = fx / 2; fy = (fy - 1) / 2;
103 :     interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
104 :     interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
105 :     break;
106 :     case 2:
107 :     fx = (fx - 1) / 2; fy = fy / 2;
108 :     interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
109 :     interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
110 :     break;
111 :     default:
112 :     fx = (fx - 1) / 2; fy = (fy - 1) / 2;
113 :     interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
114 :     interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
115 :     break;
116 :     }
117 : Isibaar 3
118 : edgomez 851 switch (((bx & 1) << 1) | (by & 1)) {
119 :     case 0:
120 :     bx = bx / 2; by = by / 2;
121 :     b_refu = (uint8_t*)data->b_RefCU + by * stride + bx, stride;
122 :     b_refv = (uint8_t*)data->b_RefCV + by * stride + bx, stride;
123 :     break;
124 :     case 1:
125 :     bx = bx / 2; by = (by - 1) / 2;
126 :     interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
127 :     interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
128 :     break;
129 :     case 2:
130 :     bx = (bx - 1) / 2; by = by / 2;
131 :     interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
132 :     interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
133 :     break;
134 :     default:
135 :     bx = (bx - 1) / 2; by = (by - 1) / 2;
136 :     interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
137 :     interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
138 :     break;
139 :     }
140 : Isibaar 3
141 : edgomez 851 sad = sad8bi(data->CurU, b_refu, f_refu, stride);
142 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
143 : Isibaar 3
144 : edgomez 851 return sad;
145 :     }
146 : Isibaar 3
147 :    
148 : edgomez 851 static int32_t
149 :     ChromaSAD(int dx, int dy, const SearchData * const data)
150 :     {
151 :     int sad;
152 :     const uint32_t stride = data->iEdgedWidth/2;
153 : Isibaar 3
154 : edgomez 851 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently
155 :     data->temp[5] = dx; data->temp[6] = dy; // backup
156 : Isibaar 3
157 : edgomez 851 switch (((dx & 1) << 1) | (dy & 1)) {
158 :     case 0:
159 :     dx = dx / 2; dy = dy / 2;
160 :     sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride);
161 :     sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride);
162 :     break;
163 :     case 1:
164 :     dx = dx / 2; dy = (dy - 1) / 2;
165 :     sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride);
166 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride);
167 :     break;
168 :     case 2:
169 :     dx = (dx - 1) / 2; dy = dy / 2;
170 :     sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride);
171 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride);
172 :     break;
173 :     default:
174 :     dx = (dx - 1) / 2; dy = (dy - 1) / 2;
175 :     interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding);
176 :     sad = sad8(data->CurU, data->RefQ, stride);
177 :    
178 :     interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding);
179 :     sad += sad8(data->CurV, data->RefQ, stride);
180 :     break;
181 :     }
182 :     data->temp[7] = sad; //backup, part 2
183 :     return sad;
184 : Isibaar 3 }
185 :    
186 : edgomez 851 static __inline const uint8_t *
187 :     GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
188 : Isibaar 3 {
189 : edgomez 851 // dir : 0 = forward, 1 = backward
190 :     switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) {
191 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
192 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
193 :     case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
194 :     case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
195 :     case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth);
196 :     case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
197 :     case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
198 :     default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
199 :     }
200 : Isibaar 3 }
201 :    
202 : edgomez 851 // this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate
203 :     static __inline const uint8_t *
204 :     GetReference(const int x, const int y, const SearchData * const data)
205 : Isibaar 3 {
206 : edgomez 851 switch ( ((x&1)<<1) | (y&1) ) {
207 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
208 :     case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
209 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
210 :     default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2
211 :     }
212 : Isibaar 3 }
213 : edgomez 851
214 :     static uint8_t *
215 :     Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
216 : Isibaar 3 {
217 : edgomez 851 // create or find a qpel-precision reference picture; return pointer to it
218 :     uint8_t * Reference = data->RefQ + 16*dir;
219 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
220 :     const uint32_t rounding = data->rounding;
221 :     const int halfpel_x = x/2;
222 :     const int halfpel_y = y/2;
223 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
224 : suxen_drol 136
225 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
226 :     ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
227 :     switch( ((x&1)<<1) + (y&1) ) {
228 :     case 0: // pure halfpel position
229 :     return (uint8_t *) ref1;
230 :     break;
231 : chl 184
232 : edgomez 851 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
233 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
234 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
235 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
236 :     break;
237 : suxen_drol 136
238 : edgomez 851 case 2: // x qpel, y halfpel - left or right during qpel refinement
239 :     ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
240 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
241 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
242 :     break;
243 : edgomez 195
244 : edgomez 851 default: // x and y in qpel resolution - the "corners" (top left/right and
245 :     // bottom left/right) during qpel refinement
246 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
247 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
248 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
249 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
250 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
251 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
252 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
253 :     break;
254 :     }
255 :     return Reference;
256 :     }
257 : Isibaar 3
258 : edgomez 851 static uint8_t *
259 :     Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
260 :     {
261 :     // create or find a qpel-precision reference picture; return pointer to it
262 :     uint8_t * Reference = data->RefQ + 16*dir;
263 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
264 :     const uint32_t rounding = data->rounding;
265 :     const int halfpel_x = x/2;
266 :     const int halfpel_y = y/2;
267 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
268 : chl 326
269 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
270 :     switch( ((x&1)<<1) + (y&1) ) {
271 :     case 3: // x and y in qpel resolution - the "corners" (top left/right and
272 :     // bottom left/right) during qpel refinement
273 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
274 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
275 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
276 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
277 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
278 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
279 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
280 :     break;
281 : chl 184
282 : edgomez 851 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
283 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
284 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
285 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
286 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
287 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
288 :     break;
289 : chl 184
290 : edgomez 851 case 2: // x qpel, y halfpel - left or right during qpel refinement
291 :     ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
292 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
293 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
294 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
295 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
296 :     break;
297 : edgomez 195
298 : edgomez 851 case 0: // pure halfpel position
299 :     return (uint8_t *) ref1;
300 :     }
301 :     return Reference;
302 :     }
303 : edgomez 195
304 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS START */
305 : edgomez 195
306 : edgomez 851 static void
307 :     CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
308 :     {
309 :     int xc, yc;
310 :     const uint8_t * Reference;
311 :     VECTOR * current;
312 :     int32_t sad; uint32_t t;
313 : edgomez 195
314 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
315 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
316 : edgomez 195
317 : edgomez 851 if (!data->qpel_precision) {
318 :     Reference = GetReference(x, y, data);
319 :     current = data->currentMV;
320 :     xc = x; yc = y;
321 :     } else { // x and y are in 1/4 precision
322 :     Reference = Interpolate16x16qpel(x, y, 0, data);
323 :     xc = x/2; yc = y/2; //for chroma sad
324 :     current = data->currentQMV;
325 :     }
326 : edgomez 195
327 : edgomez 851 sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
328 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
329 : edgomez 195
330 : edgomez 851 sad += (data->lambda16 * t * sad)>>10;
331 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
332 : chl 175
333 : edgomez 851 if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
334 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
335 : edgomez 195
336 : edgomez 851 if (sad < data->iMinSAD[0]) {
337 :     data->iMinSAD[0] = sad;
338 :     current[0].x = x; current[0].y = y;
339 :     *dir = Direction;
340 :     }
341 : edgomez 195
342 : edgomez 851 if (data->temp[1] < data->iMinSAD[1]) {
343 :     data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
344 :     if (data->temp[2] < data->iMinSAD[2]) {
345 :     data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
346 :     if (data->temp[3] < data->iMinSAD[3]) {
347 :     data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
348 :     if (data->temp[4] < data->iMinSAD[4]) {
349 :     data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
350 :    
351 : Isibaar 3 }
352 :    
353 : edgomez 851 static void
354 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
355 :     {
356 :     int32_t sad; uint32_t t;
357 :     const uint8_t * Reference;
358 : chl 326
359 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
360 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
361 : Isibaar 3
362 : edgomez 851 if (!data->qpel_precision) Reference = GetReference(x, y, data);
363 :     else Reference = Interpolate8x8qpel(x, y, 0, 0, data);
364 : Isibaar 3
365 : edgomez 851 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
366 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
367 : Isibaar 3
368 : edgomez 851 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
369 : Isibaar 3
370 : edgomez 851 if (sad < *(data->iMinSAD)) {
371 :     *(data->iMinSAD) = sad;
372 :     data->currentMV->x = x; data->currentMV->y = y;
373 :     *dir = Direction;
374 :     }
375 : Isibaar 3 }
376 :    
377 :    
378 : edgomez 851 static void
379 :     CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
380 :     {
381 :     uint32_t t;
382 :     const uint8_t * Reference;
383 :    
384 :     if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value
385 :     (x > data->max_dx) || (x < data->min_dx)
386 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
387 :    
388 :     Reference = GetReference(x, y, data);
389 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
390 :    
391 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
392 :    
393 :     data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
394 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
395 :    
396 :     if (data->temp[0] < data->iMinSAD[0]) {
397 :     data->iMinSAD[0] = data->temp[0];
398 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
399 :     *dir = Direction; }
400 :    
401 :     if (data->temp[1] < data->iMinSAD[1]) {
402 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
403 :     if (data->temp[2] < data->iMinSAD[2]) {
404 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
405 :     if (data->temp[3] < data->iMinSAD[3]) {
406 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
407 :     if (data->temp[4] < data->iMinSAD[4]) {
408 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
409 : Isibaar 3 }
410 : edgomez 195
411 : edgomez 851 static void
412 :     CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
413 :     {
414 :     int32_t sad, xc, yc;
415 :     const uint8_t * Reference;
416 :     uint32_t t;
417 :     VECTOR * current;
418 :    
419 :     if ( (x > data->max_dx) | ( x < data->min_dx)
420 :     | (y > data->max_dy) | (y < data->min_dy) ) return;
421 :    
422 :     if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value
423 :    
424 :     if (data->qpel_precision) { // x and y are in 1/4 precision
425 :     Reference = Interpolate16x16qpel(x, y, 0, data);
426 :     current = data->currentQMV;
427 :     xc = x/2; yc = y/2;
428 :     } else {
429 :     Reference = GetReference(x, y, data);
430 :     current = data->currentMV;
431 :     xc = x; yc = y;
432 :     }
433 :     t = d_mv_bits(x, y, data->predMV, data->iFcode,
434 :     data->qpel^data->qpel_precision, data->rrv);
435 :    
436 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
437 :     sad += (data->lambda16 * t * sad)>>10;
438 :    
439 :     if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
440 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
441 :    
442 :     if (sad < *(data->iMinSAD)) {
443 :     *(data->iMinSAD) = sad;
444 :     current->x = x; current->y = y;
445 :     *dir = Direction;
446 :     }
447 : chl 96 }
448 : Isibaar 3
449 : edgomez 851 static void
450 :     CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
451 :     {
452 :     // maximum speed - for P/B/I decision
453 :     int32_t sad;
454 :    
455 :     if ( (x > data->max_dx) || (x < data->min_dx)
456 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
457 :    
458 :     sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
459 :     data->iEdgedWidth, data->temp+1);
460 :    
461 :     if (sad < *(data->iMinSAD)) {
462 :     *(data->iMinSAD) = sad;
463 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
464 :     *dir = Direction;
465 :     }
466 :     if (data->temp[1] < data->iMinSAD[1]) {
467 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
468 :     if (data->temp[2] < data->iMinSAD[2]) {
469 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
470 :     if (data->temp[3] < data->iMinSAD[3]) {
471 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
472 :     if (data->temp[4] < data->iMinSAD[4]) {
473 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
474 :    
475 : Isibaar 3 }
476 :    
477 : edgomez 851 static void
478 :     CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
479 :     {
480 :     int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
481 :     uint32_t t;
482 :     const uint8_t *ReferenceF, *ReferenceB;
483 :     VECTOR *current;
484 :    
485 :     if ( (xf > data->max_dx) | (xf < data->min_dx)
486 :     | (yf > data->max_dy) | (yf < data->min_dy) ) return;
487 :    
488 :     if (!data->qpel_precision) {
489 :     ReferenceF = GetReference(xf, yf, data);
490 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
491 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
492 :     current = data->currentMV;
493 :     xcf = xf; ycf = yf;
494 :     xcb = xb; ycb = yb;
495 :     } else {
496 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
497 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
498 :     current = data->currentQMV;
499 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
500 :     xcf = xf/2; ycf = yf/2;
501 :     xcb = xb/2; ycb = yb/2;
502 :     }
503 :    
504 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
505 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
506 :    
507 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
508 :     sad += (data->lambda16 * t * sad)>>10;
509 :    
510 :     if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
511 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
512 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
513 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
514 :    
515 :     if (sad < *(data->iMinSAD)) {
516 :     *(data->iMinSAD) = sad;
517 :     current->x = xf; current->y = yf;
518 :     *dir = Direction;
519 :     }
520 : Isibaar 3 }
521 :    
522 : edgomez 851 static void
523 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
524 :     {
525 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
526 :     uint32_t k;
527 :     const uint8_t *ReferenceF;
528 :     const uint8_t *ReferenceB;
529 :     VECTOR mvs, b_mvs;
530 :    
531 :     if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return;
532 :    
533 :     for (k = 0; k < 4; k++) {
534 :     mvs.x = data->directmvF[k].x + x;
535 :     b_mvs.x = ((x == 0) ?
536 :     data->directmvB[k].x
537 :     : mvs.x - data->referencemv[k].x);
538 :    
539 :     mvs.y = data->directmvF[k].y + y;
540 :     b_mvs.y = ((y == 0) ?
541 :     data->directmvB[k].y
542 :     : mvs.y - data->referencemv[k].y);
543 :    
544 :     if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx)
545 :     | (mvs.y > data->max_dy) | (mvs.y < data->min_dy)
546 :     | (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx)
547 :     | (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return;
548 :    
549 :     if (data->qpel) {
550 :     xcf += mvs.x/2; ycf += mvs.y/2;
551 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
552 :     } else {
553 :     xcf += mvs.x; ycf += mvs.y;
554 :     xcb += b_mvs.x; ycb += b_mvs.y;
555 :     mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway
556 :     b_mvs.x *= 2; b_mvs.y *= 2;
557 :     }
558 :    
559 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
560 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
561 :    
562 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
563 :     ReferenceF, ReferenceB, data->iEdgedWidth);
564 :     if (sad > *(data->iMinSAD)) return;
565 :     }
566 :    
567 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
568 :    
569 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
570 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
571 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
572 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
573 :    
574 :     if (sad < *(data->iMinSAD)) {
575 :     *(data->iMinSAD) = sad;
576 :     data->currentMV->x = x; data->currentMV->y = y;
577 :     *dir = Direction;
578 :     }
579 : Isibaar 3 }
580 :    
581 : edgomez 851 static void
582 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
583 : Isibaar 3 {
584 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
585 :     const uint8_t *ReferenceF;
586 :     const uint8_t *ReferenceB;
587 :     VECTOR mvs, b_mvs;
588 : Isibaar 3
589 : edgomez 851 if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return;
590 : Isibaar 3
591 : edgomez 851 mvs.x = data->directmvF[0].x + x;
592 :     b_mvs.x = ((x == 0) ?
593 :     data->directmvB[0].x
594 :     : mvs.x - data->referencemv[0].x);
595 : Isibaar 3
596 : edgomez 851 mvs.y = data->directmvF[0].y + y;
597 :     b_mvs.y = ((y == 0) ?
598 :     data->directmvB[0].y
599 :     : mvs.y - data->referencemv[0].y);
600 : Isibaar 3
601 : edgomez 851 if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx)
602 :     | (mvs.y > data->max_dy) | (mvs.y < data->min_dy)
603 :     | (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx)
604 :     | (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return;
605 : Isibaar 3
606 : edgomez 851 if (data->qpel) {
607 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
608 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
609 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
610 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
611 :     } else {
612 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
613 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
614 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
615 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
616 :     }
617 :    
618 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
619 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
620 :    
621 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
622 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
623 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
624 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
625 :    
626 :     if (sad < *(data->iMinSAD)) {
627 :     *(data->iMinSAD) = sad;
628 :     data->currentMV->x = x; data->currentMV->y = y;
629 :     *dir = Direction;
630 :     }
631 : Isibaar 3 }
632 :    
633 : edgomez 851
634 :     static void
635 :     CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
636 : Isibaar 3 {
637 :    
638 : edgomez 851 static int16_t in[64], coeff[64];
639 :     int32_t bits = 0, sum;
640 :     VECTOR * current;
641 :     const uint8_t * ptr;
642 :     int i, cbp = 0, t, xc, yc;
643 : edgomez 195
644 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
645 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
646 : edgomez 195
647 : edgomez 851 if (!data->qpel_precision) {
648 :     ptr = GetReference(x, y, data);
649 :     current = data->currentMV;
650 :     xc = x; yc = y;
651 :     } else { // x and y are in 1/4 precision
652 :     ptr = Interpolate16x16qpel(x, y, 0, data);
653 :     current = data->currentQMV;
654 :     xc = x/2; yc = y/2;
655 :     }
656 : Isibaar 3
657 : edgomez 851 for(i = 0; i < 4; i++) {
658 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
659 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
660 :     fdct(in);
661 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
662 :     else sum = quant4_inter(coeff, in, data->lambda16);
663 :     if (sum > 0) {
664 :     cbp |= 1 << (5 - i);
665 :     bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
666 :     } else data->temp[i] = 0;
667 :     }
668 : Isibaar 3
669 : edgomez 851 bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
670 : edgomez 195
671 : edgomez 851 if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma
672 :     xc = (xc >> 1) + roundtab_79[xc & 0x3];
673 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
674 :    
675 :     //chroma U
676 :     ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCU, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
677 :     transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
678 :     fdct(in);
679 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
680 :     else sum = quant4_inter(coeff, in, data->lambda16);
681 :     if (sum > 0) {
682 :     cbp |= 1 << (5 - 4);
683 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
684 : chl 344 }
685 : edgomez 851
686 :     if (bits < data->iMinSAD[0]) {
687 :     //chroma V
688 :     ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCV, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
689 :     transfer_8to16subro(in, ptr, data->CurV, 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 - 5);
695 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
696 :     }
697 :     }
698 : edgomez 78 }
699 : edgomez 851
700 :     bits += cbpy_tab[15-(cbp>>2)].len;
701 :     bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
702 :    
703 :     if (bits < data->iMinSAD[0]) {
704 :     data->iMinSAD[0] = bits;
705 :     current[0].x = x; current[0].y = y;
706 :     *dir = Direction;
707 :     }
708 :    
709 :     if (data->temp[0] + t < data->iMinSAD[1]) {
710 :     data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
711 :     if (data->temp[1] < data->iMinSAD[2]) {
712 :     data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
713 :     if (data->temp[2] < data->iMinSAD[3]) {
714 :     data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
715 :     if (data->temp[3] < data->iMinSAD[4]) {
716 :     data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
717 :    
718 : Isibaar 3 }
719 : edgomez 851 static void
720 :     CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
721 : chl 96 {
722 :    
723 : edgomez 851 static int16_t in[64], coeff[64];
724 :     int32_t sum, bits;
725 :     VECTOR * current;
726 :     const uint8_t * ptr;
727 :     int cbp;
728 : edgomez 195
729 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
730 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
731 : edgomez 195
732 : edgomez 851 if (!data->qpel_precision) {
733 :     ptr = GetReference(x, y, data);
734 :     current = data->currentMV;
735 :     } else { // x and y are in 1/4 precision
736 :     ptr = Interpolate8x8qpel(x, y, 0, 0, data);
737 :     current = data->currentQMV;
738 :     }
739 : chl 96
740 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
741 :     fdct(in);
742 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
743 :     else sum = quant4_inter(coeff, in, data->lambda16);
744 :     if (sum > 0) {
745 :     bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
746 :     cbp = 1;
747 :     } else cbp = bits = 0;
748 : chl 96
749 : edgomez 851 bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
750 : chl 96
751 : edgomez 851 if (bits < data->iMinSAD[0]) {
752 :     data->temp[0] = cbp;
753 :     data->iMinSAD[0] = bits;
754 :     current[0].x = x; current[0].y = y;
755 :     *dir = Direction;
756 :     }
757 :     }
758 : chl 96
759 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
760 : edgomez 195
761 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
762 : chl 96
763 : edgomez 851 static void
764 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
765 :     {
766 : edgomez 195
767 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
768 : edgomez 195
769 : edgomez 851 int iDirection;
770 : edgomez 195
771 : edgomez 851 for(;;) { //forever
772 :     iDirection = 0;
773 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
774 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
775 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
776 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
777 : edgomez 195
778 : edgomez 851 /* now we're doing diagonal checks near our candidate */
779 : edgomez 195
780 : edgomez 851 if (iDirection) { //if anything found
781 :     bDirection = iDirection;
782 :     iDirection = 0;
783 :     x = data->currentMV->x; y = data->currentMV->y;
784 :     if (bDirection & 3) { //our candidate is left or right
785 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
786 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
787 :     } else { // what remains here is up or down
788 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
789 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
790 :     }
791 : edgomez 195
792 : edgomez 851 if (iDirection) {
793 :     bDirection += iDirection;
794 :     x = data->currentMV->x; y = data->currentMV->y;
795 :     }
796 :     } else { //about to quit, eh? not so fast....
797 :     switch (bDirection) {
798 :     case 2:
799 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
800 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
801 : edgomez 195 break;
802 : edgomez 851 case 1:
803 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
804 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
805 : edgomez 195 break;
806 : edgomez 851 case 2 + 4:
807 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
808 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
809 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
810 :     break;
811 :     case 4:
812 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
813 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
814 :     break;
815 : edgomez 195 case 8:
816 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
817 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
818 : edgomez 195 break;
819 : edgomez 851 case 1 + 4:
820 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
821 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
822 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
823 : edgomez 195 break;
824 : edgomez 851 case 2 + 8:
825 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
826 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
827 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
828 :     break;
829 :     case 1 + 8:
830 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
831 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
832 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
833 :     break;
834 :     default: //1+2+4+8 == we didn't find anything at all
835 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
836 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
837 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
838 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
839 :     break;
840 : chl 96 }
841 : edgomez 851 if (!iDirection) break; //ok, the end. really
842 :     bDirection = iDirection;
843 :     x = data->currentMV->x; y = data->currentMV->y;
844 : chl 345 }
845 : edgomez 195 }
846 : chl 96 }
847 :    
848 : edgomez 851 static void
849 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
850 : chl 96 {
851 : edgomez 851 int iDirection;
852 : edgomez 195
853 : edgomez 851 do {
854 :     iDirection = 0;
855 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
856 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
857 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
858 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
859 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
860 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
861 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
862 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
863 : chl 96
864 : edgomez 851 bDirection = iDirection;
865 :     x = data->currentMV->x; y = data->currentMV->y;
866 :     } while (iDirection);
867 : chl 96 }
868 :    
869 : edgomez 851 static void
870 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
871 : chl 181 {
872 :    
873 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
874 :    
875 : edgomez 851 int iDirection;
876 : edgomez 195
877 : edgomez 851 do {
878 :     iDirection = 0;
879 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
880 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
881 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
882 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
883 :    
884 :     /* now we're doing diagonal checks near our candidate */
885 :    
886 :     if (iDirection) { //checking if anything found
887 :     bDirection = iDirection;
888 : chl 181 iDirection = 0;
889 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
890 :     if (bDirection & 3) { //our candidate is left or right
891 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
892 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
893 :     } else { // what remains here is up or down
894 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
895 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
896 :     }
897 :     bDirection += iDirection;
898 :     x = data->currentMV->x; y = data->currentMV->y;
899 :     }
900 :     }
901 :     while (iDirection);
902 :     }
903 : chl 181
904 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
905 : chl 181
906 : edgomez 851 static void
907 :     SubpelRefine(const SearchData * const data)
908 :     {
909 :     /* Do a half-pel or q-pel refinement */
910 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
911 :     int iDirection; //only needed because macro expects it
912 : chl 181
913 : edgomez 851 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
914 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
915 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
916 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
917 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
918 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
919 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
920 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
921 :     }
922 : chl 181
923 : edgomez 851 static __inline int
924 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
925 :     const int x, const int y,
926 :     const uint32_t stride, const uint32_t iQuant, int rrv)
927 : chl 181
928 : edgomez 851 {
929 :     if(!rrv) {
930 :     uint32_t sadC = sad8(current->u + x*8 + y*stride*8,
931 :     reference->u + x*8 + y*stride*8, stride);
932 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
933 :     sadC += sad8(current->v + (x + y*stride)*8,
934 :     reference->v + (x + y*stride)*8, stride);
935 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
936 :     return 1;
937 : chl 181
938 : edgomez 851 } else {
939 :     uint32_t sadC = sad16(current->u + x*16 + y*stride*16,
940 :     reference->u + x*16 + y*stride*16, stride, 256*4096);
941 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
942 :     sadC += sad16(current->v + (x + y*stride)*16,
943 :     reference->v + (x + y*stride)*16, stride, 256*4096);
944 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
945 :     return 1;
946 : chl 181 }
947 : edgomez 851 }
948 : edgomez 504
949 : edgomez 851 static __inline void
950 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
951 :     {
952 :     pMB->mode = MODE_NOT_CODED;
953 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
954 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
955 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
956 : chl 181 }
957 :    
958 : edgomez 851 bool
959 :     MotionEstimation(MBParam * const pParam,
960 :     FRAMEINFO * const current,
961 :     FRAMEINFO * const reference,
962 :     const IMAGE * const pRefH,
963 :     const IMAGE * const pRefV,
964 :     const IMAGE * const pRefHV,
965 :     const uint32_t iLimit)
966 : chl 181 {
967 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
968 :     const IMAGE *const pCurrent = &current->image;
969 :     const IMAGE *const pRef = &reference->image;
970 : chl 181
971 : edgomez 851 uint32_t mb_width = pParam->mb_width;
972 :     uint32_t mb_height = pParam->mb_height;
973 :     const uint32_t iEdgedWidth = pParam->edged_width;
974 :     const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags);
975 : chl 181
976 : edgomez 851 uint32_t x, y;
977 :     uint32_t iIntra = 0;
978 :     int32_t quant = current->quant, sad00;
979 : chl 181
980 : edgomez 851 // some pre-initialized thingies for SearchP
981 :     int32_t temp[8];
982 :     VECTOR currentMV[5];
983 :     VECTOR currentQMV[5];
984 :     int32_t iMinSAD[5];
985 :     SearchData Data;
986 :     memset(&Data, 0, sizeof(SearchData));
987 :     Data.iEdgedWidth = iEdgedWidth;
988 :     Data.currentMV = currentMV;
989 :     Data.currentQMV = currentQMV;
990 :     Data.iMinSAD = iMinSAD;
991 :     Data.temp = temp;
992 :     Data.iFcode = current->fcode;
993 :     Data.rounding = pParam->m_rounding_type;
994 :     Data.qpel = pParam->m_quarterpel;
995 :     Data.chroma = MotionFlags & PMV_CHROMA16;
996 :     Data.rrv = current->global_flags & XVID_REDUCED;
997 : edgomez 195
998 : edgomez 851 if ((current->global_flags & XVID_REDUCED)) {
999 :     mb_width = (pParam->width + 31) / 32;
1000 :     mb_height = (pParam->height + 31) / 32;
1001 :     Data.qpel = 0;
1002 :     }
1003 : chl 181
1004 : edgomez 851 Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose)
1005 :     if (sadInit) (*sadInit) ();
1006 : chl 181
1007 : edgomez 851 for (y = 0; y < mb_height; y++) {
1008 :     for (x = 0; x < mb_width; x++) {
1009 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1010 : chl 181
1011 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1012 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1013 :     pRef->y + (x + y * iEdgedWidth) * 16,
1014 :     pParam->edged_width, pMB->sad8 );
1015 : chl 181
1016 : edgomez 851 else pMB->sad16 =
1017 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1018 :     pRef->y + (x + y * iEdgedWidth) * 32,
1019 :     pParam->edged_width, pMB->sad8 );
1020 : chl 181
1021 : edgomez 851 if (Data.chroma) {
1022 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1023 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1024 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1025 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1026 :     pMB->sad16 += Data.temp[7];
1027 :     }
1028 :    
1029 :     sad00 = pMB->sad16;
1030 :    
1031 :     if (!(current->global_flags & XVID_LUMIMASKING)) {
1032 :     pMB->dquant = NO_CHANGE;
1033 :     } else {
1034 :     if (pMB->dquant != NO_CHANGE) {
1035 :     quant += DQtab[pMB->dquant];
1036 :     if (quant > 31) quant = 31;
1037 :     else if (quant < 1) quant = 1;
1038 : chl 181 }
1039 : edgomez 851 }
1040 :     pMB->quant = current->quant;
1041 : chl 181
1042 : edgomez 851 //initial skip decision
1043 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1044 :     if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1045 :     if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) )
1046 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1047 :     SkipMacroblockP(pMB, sad00);
1048 :     continue;
1049 :     }
1050 :     }
1051 :    
1052 :     SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1053 :     y, MotionFlags, current->global_flags, pMB->quant,
1054 :     &Data, pParam, pMBs, reference->mbs,
1055 :     current->global_flags & XVID_INTER4V, pMB);
1056 :    
1057 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
1058 :     if (!(current->global_flags & XVID_GMC)) {
1059 :     if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) {
1060 :     if (!(current->global_flags & XVID_MODEDECISION_BITS)) {
1061 :     if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) )
1062 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv))
1063 :     SkipMacroblockP(pMB, sad00);
1064 :     } else { // BITS mode decision
1065 :     if (pMB->sad16 > 10)
1066 :     SkipMacroblockP(pMB, sad00); // more than 10 bits would be used for this MB - skip
1067 :    
1068 :     }
1069 : chl 181 }
1070 :     }
1071 : edgomez 851 if (pMB->mode == MODE_INTRA)
1072 :     if (++iIntra > iLimit) return 1;
1073 : chl 181 }
1074 :     }
1075 : edgomez 851
1076 :     if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */
1077 :     {
1078 :     current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1079 :     }
1080 :     return 0;
1081 : chl 181 }
1082 :    
1083 :    
1084 : edgomez 851 static __inline int
1085 :     make_mask(const VECTOR * const pmv, const int i)
1086 : chl 96 {
1087 : edgomez 851 int mask = 255, j;
1088 :     for (j = 0; j < i; j++) {
1089 :     if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
1090 :     if (pmv[i].x == pmv[j].x) {
1091 :     if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1092 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1093 :     } else
1094 :     if (pmv[i].y == pmv[j].y) {
1095 :     if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1096 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1097 :     }
1098 :     }
1099 :     return mask;
1100 :     }
1101 : edgomez 195
1102 : edgomez 851 static __inline void
1103 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1104 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1105 :     {
1106 : chl 96
1107 : edgomez 851 //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1108 :     if (rrv) { iWcount /= 2; iHcount /= 2; }
1109 : edgomez 195
1110 : edgomez 851 if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
1111 :     pmv[5].x = EVEN(pmv[3].x);
1112 :     pmv[5].y = EVEN(pmv[3].y);
1113 :     } else pmv[5].x = pmv[5].y = 0;
1114 : chl 96
1115 : edgomez 851 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1116 :     else pmv[3].x = pmv[3].y = 0;
1117 : chl 96
1118 : edgomez 851 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1119 :     else pmv[4].x = pmv[4].y = 0;
1120 : Isibaar 3
1121 : edgomez 851 // [1] median prediction
1122 :     pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1123 : edgomez 195
1124 : 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
1125 : edgomez 195
1126 : edgomez 851 pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1127 :     pmv[2].y = EVEN(prevMB->mvs[0].y);
1128 :    
1129 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1130 :     pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1131 :     pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1132 :     } else pmv[6].x = pmv[6].y = 0;
1133 :    
1134 :     if (rrv) {
1135 :     int i;
1136 :     for (i = 0; i < 7; i++) {
1137 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1138 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1139 :     }
1140 :     }
1141 : Isibaar 3 }
1142 :    
1143 : edgomez 851 static int
1144 :     ModeDecision(const uint32_t iQuant, SearchData * const Data,
1145 :     int inter4v,
1146 :     MACROBLOCK * const pMB,
1147 :     const MACROBLOCK * const pMBs,
1148 :     const int x, const int y,
1149 :     const MBParam * const pParam,
1150 :     const uint32_t MotionFlags,
1151 :     const uint32_t GlobalFlags)
1152 :     {
1153 : Isibaar 3
1154 : edgomez 851 int mode = MODE_INTER;
1155 : chl 96
1156 : edgomez 851 if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision
1157 :     int intra = 0;
1158 :     int sad;
1159 :     int InterBias = MV16_INTER_BIAS;
1160 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1161 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1162 :     mode = 0; //inter
1163 :     sad = Data->iMinSAD[0];
1164 :     } else {
1165 :     mode = MODE_INTER4V;
1166 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1167 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1168 :     Data->iMinSAD[0] = sad;
1169 :     }
1170 : chl 326
1171 : edgomez 851 /* intra decision */
1172 : Isibaar 3
1173 : edgomez 851 if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work
1174 :     if (y != 0)
1175 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1176 :     if (x != 0)
1177 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1178 : Isibaar 3
1179 : edgomez 851 if (Data->chroma) InterBias += 50; // to compensate bigger SAD
1180 :     if (Data->rrv) InterBias *= 4;
1181 : edgomez 195
1182 : edgomez 851 if (InterBias < pMB->sad16) {
1183 :     int32_t deviation;
1184 :     if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
1185 :     else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
1186 :     dev16(Data->Cur+8, Data->iEdgedWidth) +
1187 :     dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) +
1188 :     dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth);
1189 : edgomez 195
1190 : edgomez 851 if (deviation < (sad - InterBias)) return MODE_INTRA;// intra
1191 :     }
1192 :     return mode;
1193 : Isibaar 3
1194 : edgomez 851 } else {
1195 : edgomez 195
1196 : edgomez 851 int bits, intra, i;
1197 :     VECTOR backup[5], *v;
1198 :     Data->lambda16 = iQuant;
1199 :     Data->lambda8 = pParam->m_quant_type;
1200 : chl 181
1201 : edgomez 851 v = Data->qpel ? Data->currentQMV : Data->currentMV;
1202 :     for (i = 0; i < 5; i++) {
1203 :     Data->iMinSAD[i] = 256*4096;
1204 :     backup[i] = v[i];
1205 :     }
1206 : Isibaar 3
1207 : edgomez 851 bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1208 :     if (bits == 0) return MODE_INTER; // quick stop
1209 : Isibaar 3
1210 : edgomez 851 if (inter4v) {
1211 :     int inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1212 :     if (inter4v < bits) { Data->iMinSAD[0] = bits = inter4v; mode = MODE_INTER4V; }
1213 :     }
1214 : edgomez 195
1215 : Isibaar 3
1216 : edgomez 851 intra = CountMBBitsIntra(Data);
1217 : Isibaar 3
1218 : edgomez 851 if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; }
1219 :    
1220 :     return mode;
1221 : edgomez 195 }
1222 : edgomez 851 }
1223 : Isibaar 3
1224 : edgomez 851 static void
1225 :     SearchP(const IMAGE * const pRef,
1226 :     const uint8_t * const pRefH,
1227 :     const uint8_t * const pRefV,
1228 :     const uint8_t * const pRefHV,
1229 :     const IMAGE * const pCur,
1230 :     const int x,
1231 :     const int y,
1232 :     const uint32_t MotionFlags,
1233 :     const uint32_t GlobalFlags,
1234 :     const uint32_t iQuant,
1235 :     SearchData * const Data,
1236 :     const MBParam * const pParam,
1237 :     const MACROBLOCK * const pMBs,
1238 :     const MACROBLOCK * const prevMBs,
1239 :     int inter4v,
1240 :     MACROBLOCK * const pMB)
1241 :     {
1242 : Isibaar 3
1243 : edgomez 851 int i, iDirection = 255, mask, threshA;
1244 :     VECTOR pmv[7];
1245 : Isibaar 3
1246 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1247 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1248 : edgomez 195
1249 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1250 : Isibaar 3
1251 : edgomez 851 Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache
1252 :     i = Data->rrv ? 2 : 1;
1253 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1254 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1255 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1256 : chl 326
1257 : edgomez 851 Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1258 :     Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1259 :     Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1260 :     Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1261 :     Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1262 :     Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1263 : edgomez 195
1264 : edgomez 851 Data->lambda16 = lambda_vec16[iQuant];
1265 :     Data->lambda8 = lambda_vec8[iQuant];
1266 :     Data->qpel_precision = 0;
1267 : edgomez 195
1268 : edgomez 851 if (pMB->dquant != NO_CHANGE) inter4v = 0;
1269 : edgomez 195
1270 : edgomez 851 for(i = 0; i < 5; i++)
1271 :     Data->currentMV[i].x = Data->currentMV[i].y = 0;
1272 : chl 169
1273 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1274 :     else Data->predMV = pmv[0];
1275 : Isibaar 3
1276 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1277 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1278 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1279 :     Data->iMinSAD[2] = pMB->sad8[1];
1280 :     Data->iMinSAD[3] = pMB->sad8[2];
1281 :     Data->iMinSAD[4] = pMB->sad8[3];
1282 : chl 169
1283 : edgomez 851 if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) {
1284 :     threshA = Data->temp[0]; // that's where we keep this SAD atm
1285 :     if (threshA < 512) threshA = 512;
1286 :     else if (threshA > 1024) threshA = 1024;
1287 :     } else
1288 :     threshA = 512;
1289 : chl 169
1290 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1291 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1292 : chl 169
1293 : edgomez 851 if (!Data->rrv) {
1294 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1295 :     else CheckCandidate = CheckCandidate16no4v; //for extra speed
1296 :     } else CheckCandidate = CheckCandidate32;
1297 : chl 169
1298 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1299 : chl 169
1300 : edgomez 851 for (i = 1; i < 7; i++) {
1301 :     if (!(mask = make_mask(pmv, i)) ) continue;
1302 :     CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1303 :     if (Data->iMinSAD[0] <= threshA) break;
1304 :     }
1305 : chl 169
1306 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1307 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1308 :     (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1309 :     if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; }
1310 :     else {
1311 : Isibaar 3
1312 : edgomez 851 MainSearchFunc * MainSearchPtr;
1313 :     if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1314 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1315 :     else MainSearchPtr = DiamondSearch;
1316 : Isibaar 3
1317 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1318 : Isibaar 3
1319 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1320 :     note that this search is/might be done in halfpel positions,
1321 :     which makes it more different than the diamond above */
1322 : chl 169
1323 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH16) {
1324 :     int32_t bSAD;
1325 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1326 :     if (Data->rrv) {
1327 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1328 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1329 :     }
1330 :     if (!(MVequal(startMV, backupMV))) {
1331 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1332 : edgomez 195
1333 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1334 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1335 :     if (bSAD < Data->iMinSAD[0]) {
1336 :     Data->currentMV[0] = backupMV;
1337 :     Data->iMinSAD[0] = bSAD; }
1338 :     }
1339 : chl 169
1340 : edgomez 851 backupMV = Data->currentMV[0];
1341 :     startMV.x = startMV.y = 1;
1342 :     if (!(MVequal(startMV, backupMV))) {
1343 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1344 : chl 169
1345 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1346 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1347 :     if (bSAD < Data->iMinSAD[0]) {
1348 :     Data->currentMV[0] = backupMV;
1349 :     Data->iMinSAD[0] = bSAD; }
1350 : edgomez 195 }
1351 : edgomez 851 }
1352 :     }
1353 : edgomez 195
1354 : edgomez 851 if (MotionFlags & PMV_HALFPELREFINE16)
1355 :     if ((!(MotionFlags & HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant)
1356 :     SubpelRefine(Data);
1357 : edgomez 195
1358 : edgomez 851 for(i = 0; i < 5; i++) {
1359 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1360 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1361 :     }
1362 : Isibaar 3
1363 : edgomez 851 if (MotionFlags & PMV_QUARTERPELREFINE16)
1364 :     if ((!(MotionFlags & QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) {
1365 :     Data->qpel_precision = 1;
1366 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1367 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1368 : edgomez 195
1369 : edgomez 851 SubpelRefine(Data);
1370 :     }
1371 : Isibaar 3
1372 : edgomez 851 if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0;
1373 : Isibaar 3
1374 : edgomez 851 if (inter4v && (!(GlobalFlags & XVID_MODEDECISION_BITS) ||
1375 :     (!(MotionFlags & QUARTERPELREFINE8_BITS)) || (!(MotionFlags & HALFPELREFINE8_BITS)) ||
1376 :     ((!(MotionFlags & EXTSEARCH_BITS)) && (!(MotionFlags&PMV_EXTSEARCH8)) ))) {
1377 :     // if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop
1378 : Isibaar 3
1379 : edgomez 851 SearchData Data8;
1380 :     memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data
1381 : Isibaar 3
1382 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1383 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1384 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1385 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1386 : chl 181
1387 : edgomez 851 if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) {
1388 :     // chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it
1389 :     int sumx = 0, sumy = 0;
1390 :     const int div = 1 + Data->qpel;
1391 :     const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs;
1392 : Isibaar 3
1393 : edgomez 851 for (i = 0; i < 4; i++) {
1394 :     sumx += mv[i].x / div;
1395 :     sumy += mv[i].y / div;
1396 :     }
1397 : chl 259
1398 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1399 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1400 :     }
1401 :     }
1402 : edgomez 195
1403 : edgomez 851 inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags);
1404 :    
1405 :     if (Data->rrv) {
1406 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1407 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1408 : Isibaar 3 }
1409 :    
1410 : edgomez 851 if (inter4v == MODE_INTER) {
1411 :     pMB->mode = MODE_INTER;
1412 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1413 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1414 : Isibaar 3
1415 : edgomez 851 if(Data->qpel) {
1416 :     pMB->qmvs[0] = pMB->qmvs[1]
1417 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1418 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1419 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1420 :     } else {
1421 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1422 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1423 : Isibaar 3 }
1424 :    
1425 : edgomez 851 } else if (inter4v == MODE_INTER4V) {
1426 :     pMB->mode = MODE_INTER4V;
1427 :     pMB->sad16 = Data->iMinSAD[0];
1428 :     } else { // INTRA mode
1429 :     SkipMacroblockP(pMB, 0); // not skip, but similar enough
1430 :     pMB->mode = MODE_INTRA;
1431 :     }
1432 : edgomez 195
1433 : edgomez 851 }
1434 :    
1435 :     static void
1436 :     Search8(const SearchData * const OldData,
1437 :     const int x, const int y,
1438 :     const uint32_t MotionFlags,
1439 :     const MBParam * const pParam,
1440 :     MACROBLOCK * const pMB,
1441 :     const MACROBLOCK * const pMBs,
1442 :     const int block,
1443 :     SearchData * const Data)
1444 :     {
1445 :     int i = 0;
1446 :     Data->iMinSAD = OldData->iMinSAD + 1 + block;
1447 :     Data->currentMV = OldData->currentMV + 1 + block;
1448 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1449 :    
1450 :     if(Data->qpel) {
1451 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1452 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1453 :     Data->predMV, Data->iFcode, 0, 0);
1454 :     } else {
1455 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1456 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1457 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1458 : Isibaar 3 }
1459 :    
1460 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1461 : Isibaar 3
1462 : edgomez 851 if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) {
1463 :     if (Data->rrv) i = 2; else i = 1;
1464 : Isibaar 3
1465 : edgomez 851 Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1466 :     Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1467 :     Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1468 :     Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1469 : Isibaar 3
1470 : edgomez 851 Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1471 :     Data->qpel_precision = 0;
1472 : Isibaar 3
1473 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1474 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1475 : Isibaar 3
1476 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1477 :     else CheckCandidate = CheckCandidate16no4v;
1478 : Isibaar 3
1479 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) {
1480 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1481 : Isibaar 3
1482 : edgomez 851 MainSearchFunc *MainSearchPtr;
1483 :     if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1484 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1485 :     else MainSearchPtr = DiamondSearch;
1486 : Isibaar 3
1487 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1488 : Isibaar 3
1489 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1490 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1491 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1492 :     }
1493 :     }
1494 : edgomez 195
1495 : edgomez 851 if (MotionFlags & PMV_HALFPELREFINE8) {
1496 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1497 : edgomez 195
1498 : edgomez 851 SubpelRefine(Data); // perform halfpel refine of current best vector
1499 : Isibaar 3
1500 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1501 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1502 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1503 :     }
1504 :     }
1505 : Isibaar 3
1506 : edgomez 851 if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) {
1507 :     Data->qpel_precision = 1;
1508 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1509 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1510 :     SubpelRefine(Data);
1511 :     }
1512 :     }
1513 : edgomez 195
1514 : edgomez 851 if (Data->rrv) {
1515 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1516 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1517 :     }
1518 :    
1519 :     if(Data->qpel) {
1520 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1521 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1522 :     pMB->qmvs[block] = *Data->currentQMV;
1523 : edgomez 195 } else {
1524 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1525 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1526 : edgomez 78 }
1527 : edgomez 851
1528 :     pMB->mvs[block] = *Data->currentMV;
1529 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1530 : Isibaar 3 }
1531 :    
1532 : edgomez 851 /* motion estimation for B-frames */
1533 : chl 345
1534 : edgomez 851 static __inline VECTOR
1535 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1536 :     {
1537 :     /* the stupidiest function ever */
1538 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1539 :     }
1540 : chl 345
1541 : edgomez 851 static void __inline
1542 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1543 :     const uint32_t iWcount,
1544 :     const MACROBLOCK * const pMB,
1545 :     const uint32_t mode_curr)
1546 : chl 345 {
1547 :    
1548 : edgomez 851 // [0] is prediction
1549 :     pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1550 : chl 345
1551 : edgomez 851 pmv[1].x = pmv[1].y = 0; // [1] is zero
1552 : chl 345
1553 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1554 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1555 : chl 345
1556 : edgomez 851 if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1557 :     pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1558 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1559 :     } else pmv[3].x = pmv[3].y = 0;
1560 : chl 345
1561 : edgomez 851 if (y != 0) {
1562 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1563 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1564 :     } else pmv[4].x = pmv[4].y = 0;
1565 : chl 345
1566 : edgomez 851 if (x != 0) {
1567 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1568 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1569 :     } else pmv[5].x = pmv[5].y = 0;
1570 : chl 345
1571 : edgomez 851 if (x != 0 && y != 0) {
1572 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1573 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1574 :     } else pmv[6].x = pmv[6].y = 0;
1575 :     }
1576 : chl 345
1577 :    
1578 : edgomez 851 /* search backward or forward */
1579 :     static void
1580 :     SearchBF( const IMAGE * const pRef,
1581 :     const uint8_t * const pRefH,
1582 :     const uint8_t * const pRefV,
1583 :     const uint8_t * const pRefHV,
1584 :     const IMAGE * const pCur,
1585 :     const int x, const int y,
1586 :     const uint32_t MotionFlags,
1587 :     const uint32_t iFcode,
1588 :     const MBParam * const pParam,
1589 :     MACROBLOCK * const pMB,
1590 :     const VECTOR * const predMV,
1591 :     int32_t * const best_sad,
1592 :     const int32_t mode_current,
1593 :     SearchData * const Data)
1594 :     {
1595 : chl 345
1596 : edgomez 851 int i, iDirection = 255, mask;
1597 :     VECTOR pmv[7];
1598 :     MainSearchFunc *MainSearchPtr;
1599 :     *Data->iMinSAD = MV_MAX_ERROR;
1600 :     Data->iFcode = iFcode;
1601 :     Data->qpel_precision = 0;
1602 :     Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache
1603 : chl 345
1604 : edgomez 851 Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16;
1605 :     Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16;
1606 :     Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16;
1607 :     Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16;
1608 :     Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8;
1609 :     Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8;
1610 : chl 345
1611 : edgomez 851 Data->predMV = *predMV;
1612 : chl 345
1613 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1614 :     pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1615 : chl 345
1616 : edgomez 851 pmv[0] = Data->predMV;
1617 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1618 : chl 345
1619 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1620 : chl 345
1621 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1622 :     CheckCandidate = CheckCandidate16no4v;
1623 : chl 345
1624 : edgomez 851 // main loop. checking all predictions
1625 :     for (i = 0; i < 7; i++) {
1626 :     if (!(mask = make_mask(pmv, i)) ) continue;
1627 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1628 : chl 345 }
1629 :    
1630 : edgomez 851 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1631 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1632 :     else MainSearchPtr = DiamondSearch;
1633 : chl 345
1634 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1635 : chl 345
1636 : edgomez 851 SubpelRefine(Data);
1637 : chl 345
1638 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1639 :     Data->currentQMV->x = 2*Data->currentMV->x;
1640 :     Data->currentQMV->y = 2*Data->currentMV->y;
1641 :     Data->qpel_precision = 1;
1642 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1643 :     pParam->width, pParam->height, iFcode, 1, 0);
1644 :     SubpelRefine(Data);
1645 :     }
1646 : chl 345
1647 : edgomez 851 // three bits are needed to code backward mode. four for forward
1648 : Isibaar 3
1649 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1650 :     else *Data->iMinSAD += 3 * Data->lambda16;
1651 : edgomez 195
1652 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1653 :     *best_sad = *Data->iMinSAD;
1654 :     pMB->mode = mode_current;
1655 :     if (Data->qpel) {
1656 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1657 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1658 :     if (mode_current == MODE_FORWARD)
1659 :     pMB->qmvs[0] = *Data->currentQMV;
1660 :     else
1661 :     pMB->b_qmvs[0] = *Data->currentQMV;
1662 :     } else {
1663 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1664 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1665 :     }
1666 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1667 :     else pMB->b_mvs[0] = *Data->currentMV;
1668 :     }
1669 : edgomez 195
1670 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1671 :     else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
1672 : Isibaar 3 }
1673 :    
1674 : edgomez 851 static void
1675 :     SkipDecisionB(const IMAGE * const pCur,
1676 :     const IMAGE * const f_Ref,
1677 :     const IMAGE * const b_Ref,
1678 :     MACROBLOCK * const pMB,
1679 :     const uint32_t x, const uint32_t y,
1680 :     const SearchData * const Data)
1681 :     {
1682 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1683 :     int32_t sum;
1684 :     const int div = 1 + Data->qpel;
1685 :     int k;
1686 :     const uint32_t stride = Data->iEdgedWidth/2;
1687 :     //this is not full chroma compensation, only it's fullpel approximation. should work though
1688 : Isibaar 3
1689 : edgomez 851 for (k = 0; k < 4; k++) {
1690 :     dy += Data->directmvF[k].y / div;
1691 :     dx += Data->directmvF[0].x / div;
1692 :     b_dy += Data->directmvB[0].y / div;
1693 :     b_dx += Data->directmvB[0].x / div;
1694 :     }
1695 : Isibaar 3
1696 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1697 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1698 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1699 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1700 : Isibaar 3
1701 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1702 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1703 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1704 :     stride);
1705 : Isibaar 3
1706 : edgomez 851 if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip
1707 : Isibaar 3
1708 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1709 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1710 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1711 :     stride);
1712 : edgomez 195
1713 : edgomez 851 if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1714 :     }
1715 : Isibaar 3
1716 : edgomez 851 static __inline uint32_t
1717 :     SearchDirect(const IMAGE * const f_Ref,
1718 :     const uint8_t * const f_RefH,
1719 :     const uint8_t * const f_RefV,
1720 :     const uint8_t * const f_RefHV,
1721 :     const IMAGE * const b_Ref,
1722 :     const uint8_t * const b_RefH,
1723 :     const uint8_t * const b_RefV,
1724 :     const uint8_t * const b_RefHV,
1725 :     const IMAGE * const pCur,
1726 :     const int x, const int y,
1727 :     const uint32_t MotionFlags,
1728 :     const int32_t TRB, const int32_t TRD,
1729 :     const MBParam * const pParam,
1730 :     MACROBLOCK * const pMB,
1731 :     const MACROBLOCK * const b_mb,
1732 :     int32_t * const best_sad,
1733 :     SearchData * const Data)
1734 : Isibaar 3
1735 : edgomez 851 {
1736 :     int32_t skip_sad;
1737 :     int k = (x + Data->iEdgedWidth*y) * 16;
1738 :     MainSearchFunc *MainSearchPtr;
1739 : Isibaar 3
1740 : edgomez 851 *Data->iMinSAD = 256*4096;
1741 :     Data->Ref = f_Ref->y + k;
1742 :     Data->RefH = f_RefH + k;
1743 :     Data->RefV = f_RefV + k;
1744 :     Data->RefHV = f_RefHV + k;
1745 :     Data->bRef = b_Ref->y + k;
1746 :     Data->bRefH = b_RefH + k;
1747 :     Data->bRefV = b_RefV + k;
1748 :     Data->bRefHV = b_RefHV + k;
1749 :     Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1750 :     Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1751 :     Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1752 :     Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1753 : edgomez 195
1754 : edgomez 851 k = Data->qpel ? 4 : 2;
1755 :     Data->max_dx = k * (pParam->width - x * 16);
1756 :     Data->max_dy = k * (pParam->height - y * 16);
1757 :     Data->min_dx = -k * (16 + x * 16);
1758 :     Data->min_dy = -k * (16 + y * 16);
1759 : chl 181
1760 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1761 :     Data->qpel_precision = 0;
1762 : edgomez 170
1763 : edgomez 851 for (k = 0; k < 4; k++) {
1764 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1765 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1766 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1767 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1768 : Isibaar 3
1769 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1770 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1771 :    
1772 :     *best_sad = 256*4096; // in that case, we won't use direct mode
1773 :     pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1774 :     pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1775 :     return 256*4096;
1776 :     }
1777 :     if (b_mb->mode != MODE_INTER4V) {
1778 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1779 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1780 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1781 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1782 :     break;
1783 :     }
1784 : edgomez 195 }
1785 : Isibaar 3
1786 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1787 : Isibaar 3
1788 : edgomez 851 CheckCandidate(0, 0, 255, &k, Data);
1789 : edgomez 195
1790 : edgomez 851 // initial (fast) skip decision
1791 :     if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) {
1792 :     //possible skip
1793 :     if (Data->chroma) {
1794 :     pMB->mode = MODE_DIRECT_NONE_MV;
1795 :     return *Data->iMinSAD; // skip.
1796 :     } else {
1797 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1798 :     if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1799 :     }
1800 : Isibaar 3 }
1801 :    
1802 : edgomez 851 skip_sad = *Data->iMinSAD;
1803 : edgomez 195
1804 : edgomez 851 // DIRECT MODE DELTA VECTOR SEARCH.
1805 :     // This has to be made more effective, but at the moment I'm happy it's running at all
1806 : Isibaar 3
1807 : edgomez 851 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1808 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1809 :     else MainSearchPtr = DiamondSearch;
1810 : Isibaar 3
1811 : edgomez 851 MainSearchPtr(0, 0, Data, 255);
1812 : Isibaar 3
1813 : edgomez 851 SubpelRefine(Data);
1814 : chl 181
1815 : edgomez 851 *best_sad = *Data->iMinSAD;
1816 : chl 181
1817 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1818 :     else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1819 : chl 181
1820 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1821 : edgomez 195
1822 : edgomez 851 for (k = 0; k < 4; k++) {
1823 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1824 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1825 :     ? Data->directmvB[k].x
1826 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1827 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1828 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1829 :     ? Data->directmvB[k].y
1830 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1831 :     if (Data->qpel) {
1832 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1833 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1834 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1835 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1836 :     }
1837 : edgomez 195
1838 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1839 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1840 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1841 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1842 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1843 :     break;
1844 :     }
1845 : edgomez 78 }
1846 : edgomez 851 return skip_sad;
1847 :     }
1848 : Isibaar 3
1849 : edgomez 851 static void
1850 :     SearchInterpolate(const IMAGE * const f_Ref,
1851 :     const uint8_t * const f_RefH,
1852 :     const uint8_t * const f_RefV,
1853 :     const uint8_t * const f_RefHV,
1854 :     const IMAGE * const b_Ref,
1855 :     const uint8_t * const b_RefH,
1856 :     const uint8_t * const b_RefV,
1857 :     const uint8_t * const b_RefHV,
1858 :     const IMAGE * const pCur,
1859 :     const int x, const int y,
1860 :     const uint32_t fcode,
1861 :     const uint32_t bcode,
1862 :     const uint32_t MotionFlags,
1863 :     const MBParam * const pParam,
1864 :     const VECTOR * const f_predMV,
1865 :     const VECTOR * const b_predMV,
1866 :     MACROBLOCK * const pMB,
1867 :     int32_t * const best_sad,
1868 :     SearchData * const fData)
1869 : chl 96
1870 : edgomez 851 {
1871 : chl 169
1872 : edgomez 851 int iDirection, i, j;
1873 :     SearchData bData;
1874 : chl 169
1875 : edgomez 851 fData->qpel_precision = 0;
1876 :     memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1877 :     *fData->iMinSAD = 4096*256;
1878 :     bData.currentMV++; bData.currentQMV++;
1879 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1880 : chl 169
1881 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1882 :     bData.bRef = fData->Ref = f_Ref->y + i;
1883 :     bData.bRefH = fData->RefH = f_RefH + i;
1884 :     bData.bRefV = fData->RefV = f_RefV + i;
1885 :     bData.bRefHV = fData->RefHV = f_RefHV + i;
1886 :     bData.Ref = fData->bRef = b_Ref->y + i;
1887 :     bData.RefH = fData->bRefH = b_RefH + i;
1888 :     bData.RefV = fData->bRefV = b_RefV + i;
1889 :     bData.RefHV = fData->bRefHV = b_RefHV + i;
1890 :     bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1891 :     bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1892 :     bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1893 :     bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1894 : chl 169
1895 :    
1896 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1897 :     fData->bpredMV = bData.predMV = *b_predMV;
1898 :     fData->currentMV[0] = fData->currentMV[2];
1899 : Isibaar 3
1900 : 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);
1901 :     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);
1902 : Isibaar 3
1903 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1904 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1905 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1906 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1907 : chl 169
1908 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1909 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1910 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1911 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1912 : Isibaar 3
1913 : edgomez 851 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1914 : chl 169
1915 : edgomez 851 //diamond
1916 :     do {
1917 :     iDirection = 255;
1918 :     // forward MV moves
1919 :     i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1920 : chl 169
1921 : edgomez 851 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1922 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1923 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1924 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1925 : edgomez 195
1926 : edgomez 851 // backward MV moves
1927 :     i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1928 :     fData->currentMV[2] = fData->currentMV[0];
1929 :     CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1930 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1931 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1932 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1933 : Isibaar 3
1934 : edgomez 851 } while (!(iDirection));
1935 : chl 140
1936 : edgomez 851 //qpel refinement
1937 :     if (fData->qpel) {
1938 :     if (*fData->iMinSAD > *best_sad + 500) return;
1939 :     CheckCandidate = CheckCandidateInt;
1940 :     fData->qpel_precision = bData.qpel_precision = 1;
1941 :     get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1942 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1943 :     fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1944 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1945 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1946 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1947 :     SubpelRefine(fData);
1948 :     if (*fData->iMinSAD > *best_sad + 300) return;
1949 :     fData->currentQMV[2] = fData->currentQMV[0];
1950 :     SubpelRefine(&bData);
1951 :     }
1952 : chl 140
1953 : edgomez 851 *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode.
1954 : Isibaar 3
1955 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
1956 :     *best_sad = *fData->iMinSAD;
1957 :     pMB->mvs[0] = fData->currentMV[0];
1958 :     pMB->b_mvs[0] = fData->currentMV[1];
1959 :     pMB->mode = MODE_INTERPOLATE;
1960 :     if (fData->qpel) {
1961 :     pMB->qmvs[0] = fData->currentQMV[0];
1962 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1963 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1964 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1965 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1966 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1967 :     } else {
1968 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1969 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1970 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1971 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1972 :     }
1973 : edgomez 78 }
1974 : edgomez 851 }
1975 : Isibaar 3
1976 : edgomez 851 void
1977 :     MotionEstimationBVOP(MBParam * const pParam,
1978 :     FRAMEINFO * const frame,
1979 :     const int32_t time_bp,
1980 :     const int32_t time_pp,
1981 :     // forward (past) reference
1982 :     const MACROBLOCK * const f_mbs,
1983 :     const IMAGE * const f_ref,
1984 :     const IMAGE * const f_refH,
1985 :     const IMAGE * const f_refV,
1986 :     const IMAGE * const f_refHV,
1987 :     // backward (future) reference
1988 :     const FRAMEINFO * const b_reference,
1989 :     const IMAGE * const b_ref,
1990 :     const IMAGE * const b_refH,
1991 :     const IMAGE * const b_refV,
1992 :     const IMAGE * const b_refHV)
1993 :     {
1994 :     uint32_t i, j;
1995 :     int32_t best_sad;
1996 :     uint32_t skip_sad;
1997 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
1998 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
1999 : Isibaar 3
2000 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2001 : Isibaar 3
2002 : edgomez 851 const int32_t TRB = time_pp - time_bp;
2003 :     const int32_t TRD = time_pp;
2004 : edgomez 195
2005 : edgomez 851 // some pre-inintialized data for the rest of the search
2006 : Isibaar 3
2007 : edgomez 851 SearchData Data;
2008 :     int32_t iMinSAD;
2009 :     VECTOR currentMV[3];
2010 :     VECTOR currentQMV[3];
2011 :     int32_t temp[8];
2012 :     memset(&Data, 0, sizeof(SearchData));
2013 :     Data.iEdgedWidth = pParam->edged_width;
2014 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2015 :     Data.iMinSAD = &iMinSAD;
2016 :     Data.lambda16 = lambda_vec16[frame->quant];
2017 :     Data.qpel = pParam->m_quarterpel;
2018 :     Data.rounding = 0;
2019 :     Data.chroma = frame->motion_flags & PMV_CHROMA8;
2020 :     Data.temp = temp;
2021 : Isibaar 3
2022 : edgomez 851 Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose)
2023 :     // note: i==horizontal, j==vertical
2024 :     for (j = 0; j < pParam->mb_height; j++) {
2025 : edgomez 195
2026 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2027 : Isibaar 3
2028 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2029 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2030 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2031 : edgomez 195
2032 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2033 :     if (b_reference->coding_type != S_VOP)
2034 :     if (b_mb->mode == MODE_NOT_CODED) {
2035 :     pMB->mode = MODE_NOT_CODED;
2036 :     continue;
2037 :     }
2038 : Isibaar 3
2039 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2040 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2041 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2042 :     pMB->quant = frame->quant;
2043 : Isibaar 3
2044 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2045 :     and (2) sets very good predictions for forward and backward search */
2046 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2047 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2048 :     &frame->image,
2049 :     i, j,
2050 :     frame->motion_flags,
2051 :     TRB, TRD,
2052 :     pParam,
2053 :     pMB, b_mb,
2054 :     &best_sad,
2055 :     &Data);
2056 : Isibaar 3
2057 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2058 : edgomez 195
2059 : edgomez 851 // forward search
2060 :     SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2061 :     &frame->image, i, j,
2062 :     frame->motion_flags,
2063 :     frame->fcode, pParam,
2064 :     pMB, &f_predMV, &best_sad,
2065 :     MODE_FORWARD, &Data);
2066 : edgomez 195
2067 : edgomez 851 // backward search
2068 :     SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2069 :     &frame->image, i, j,
2070 :     frame->motion_flags,
2071 :     frame->bcode, pParam,
2072 :     pMB, &b_predMV, &best_sad,
2073 :     MODE_BACKWARD, &Data);
2074 :    
2075 :     // interpolate search comes last, because it uses data from forward and backward as prediction
2076 :     SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2077 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2078 :     &frame->image,
2079 :     i, j,
2080 :     frame->fcode, frame->bcode,
2081 :     frame->motion_flags,
2082 :     pParam,
2083 :     &f_predMV, &b_predMV,
2084 :     pMB, &best_sad,
2085 :     &Data);
2086 :    
2087 :     // final skip decision
2088 :     if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2089 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2090 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2091 :    
2092 :     switch (pMB->mode) {
2093 :     case MODE_FORWARD:
2094 :     f_count++;
2095 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2096 :     break;
2097 :     case MODE_BACKWARD:
2098 :     b_count++;
2099 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2100 :     break;
2101 :     case MODE_INTERPOLATE:
2102 :     i_count++;
2103 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2104 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2105 :     break;
2106 :     case MODE_DIRECT:
2107 :     case MODE_DIRECT_NO4V:
2108 :     d_count++;
2109 :     default:
2110 :     break;
2111 :     }
2112 :     }
2113 :     }
2114 : Isibaar 3 }
2115 : chl 96
2116 : edgomez 851 static __inline void
2117 :     MEanalyzeMB ( const uint8_t * const pRef,
2118 :     const uint8_t * const pCur,
2119 :     const int x,
2120 :     const int y,
2121 :     const MBParam * const pParam,
2122 :     MACROBLOCK * const pMBs,
2123 :     SearchData * const Data)
2124 : chl 96 {
2125 :    
2126 : edgomez 851 int i, mask;
2127 :     VECTOR pmv[3];
2128 :     MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width];
2129 : chl 96
2130 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2131 : chl 96
2132 : edgomez 851 //median is only used as prediction. it doesn't have to be real
2133 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2134 :     else
2135 :     if (x == 1) //left macroblock does not have any vector now
2136 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2137 :     else if (y == 1) // top macroblock doesn't have it's vector
2138 :     Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
2139 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2140 : edgomez 195
2141 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2142 :     pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv);
2143 : edgomez 195
2144 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2145 :     Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
2146 : chl 96
2147 : edgomez 851 pmv[1].x = EVEN(pMB->mvs[0].x);
2148 :     pmv[1].y = EVEN(pMB->mvs[0].y);
2149 :     pmv[2].x = EVEN(Data->predMV.x);
2150 :     pmv[2].y = EVEN(Data->predMV.y);
2151 :     pmv[0].x = pmv[0].y = 0;
2152 : edgomez 195
2153 : edgomez 851 CheckCandidate32I(0, 0, 255, &i, Data);
2154 : edgomez 195
2155 : edgomez 851 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) {
2156 : chl 96
2157 : edgomez 851 if (!(mask = make_mask(pmv, 1)))
2158 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2159 :     if (!(mask = make_mask(pmv, 2)))
2160 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2161 : chl 96
2162 : edgomez 851 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed
2163 :     DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2164 :    
2165 :     for (i = 0; i < 4; i++) {
2166 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2167 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2168 :     MB->mode = MODE_INTER;
2169 :     MB->sad16 = Data->iMinSAD[i+1];
2170 :     }
2171 : chl 96 }
2172 : edgomez 851 }
2173 : chl 96
2174 : edgomez 851 #define INTRA_BIAS 2500
2175 :     #define INTRA_THRESH 1500
2176 :     #define INTER_THRESH 1400
2177 : chl 96
2178 : edgomez 851 int
2179 :     MEanalysis( const IMAGE * const pRef,
2180 :     FRAMEINFO * const Current,
2181 :     MBParam * const pParam,
2182 :     int maxIntra, //maximum number if non-I frames
2183 :     int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2184 :     int bCount) // number of B frames in a row
2185 :     {
2186 :     uint32_t x, y, intra = 0;
2187 :     int sSAD = 0;
2188 :     MACROBLOCK * const pMBs = Current->mbs;
2189 :     const IMAGE * const pCurrent = &Current->image;
2190 :     int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH;
2191 : chl 96
2192 : edgomez 851 int32_t iMinSAD[5], temp[5];
2193 :     VECTOR currentMV[5];
2194 :     SearchData Data;
2195 :     Data.iEdgedWidth = pParam->edged_width;
2196 :     Data.currentMV = currentMV;
2197 :     Data.iMinSAD = iMinSAD;
2198 :     Data.iFcode = Current->fcode;
2199 :     Data.rrv = Current->global_flags & XVID_REDUCED;
2200 :     Data.temp = temp;
2201 :     CheckCandidate = CheckCandidate32I;
2202 : chl 96
2203 : edgomez 851 if (intraCount != 0 && intraCount < 10) // we're right after an I frame
2204 :     IntraThresh += 4 * (intraCount - 10) * (intraCount - 10);
2205 :     else
2206 :     if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2207 :     IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
2208 : chl 96
2209 : edgomez 851 InterThresh += 400 * (1 - bCount);
2210 :     if (InterThresh < 300) InterThresh = 300;
2211 : chl 326
2212 : edgomez 851 if (sadInit) (*sadInit) ();
2213 : chl 96
2214 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2215 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2216 :     int i;
2217 : edgomez 195
2218 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2219 : edgomez 195
2220 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2221 : edgomez 195
2222 : edgomez 851 for (i = 0; i < 4; i++) {
2223 :     int dev;
2224 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2225 :     if (pMB->sad16 > IntraThresh) {
2226 :     dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2227 :     pParam->edged_width);
2228 :     if (dev + IntraThresh < pMB->sad16) {
2229 :     pMB->mode = MODE_INTRA;
2230 :     if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP;
2231 :     }
2232 :     }
2233 :     sSAD += pMB->sad16;
2234 :     }
2235 :     }
2236 : edgomez 195 }
2237 : edgomez 851 sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2);
2238 :     // if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP;
2239 :     if (sSAD > InterThresh ) return P_VOP;
2240 :     emms();
2241 :     return B_VOP;
2242 : chl 96
2243 : edgomez 851 }
2244 : chl 96
2245 :    
2246 : edgomez 851 static WARPPOINTS
2247 :     GlobalMotionEst(const MACROBLOCK * const pMBs,
2248 :     const MBParam * const pParam,
2249 :     const FRAMEINFO * const current,
2250 :     const FRAMEINFO * const reference,
2251 :     const IMAGE * const pRefH,
2252 :     const IMAGE * const pRefV,
2253 :     const IMAGE * const pRefHV )
2254 :     {
2255 : chl 96
2256 : edgomez 851 const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2257 :     const int deltay=8;
2258 :     const int grad=512; // lower bound for deviation in MB
2259 : chl 96
2260 : edgomez 851 WARPPOINTS gmc;
2261 : chl 96
2262 : edgomez 851 uint32_t mx, my;
2263 : chl 96
2264 : edgomez 851 int MBh = pParam->mb_height;
2265 :     int MBw = pParam->mb_width;
2266 : chl 96
2267 : edgomez 851 int *MBmask= calloc(MBh*MBw,sizeof(int));
2268 :     double DtimesF[4] = { 0.,0., 0., 0. };
2269 :     double sol[4] = { 0., 0., 0., 0. };
2270 :     double a,b,c,n,denom;
2271 :     double meanx,meany;
2272 :     int num,oldnum;
2273 : edgomez 195
2274 : edgomez 851 if (!MBmask) { fprintf(stderr,"Mem error\n"); return gmc;}
2275 : edgomez 195
2276 : edgomez 851 // filter mask of all blocks
2277 :    
2278 :     for (my = 1; my < MBh-1; my++)
2279 :     for (mx = 1; mx < MBw-1; mx++)
2280 :     {
2281 :     const int mbnum = mx + my * MBw;
2282 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2283 :     const VECTOR mv = pMB->mvs[0];
2284 :    
2285 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2286 :     continue;
2287 :    
2288 :     if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2289 :     && ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2290 :     && ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2291 :     && ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2292 :     MBmask[mbnum]=1;
2293 : chl 96 }
2294 :    
2295 : edgomez 851 for (my = 1; my < MBh-1; my++)
2296 :     for (mx = 1; mx < MBw-1; mx++)
2297 :     {
2298 :     const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx;
2299 : chl 96
2300 : edgomez 851 const int mbnum = mx + my * MBw;
2301 :     if (!MBmask[mbnum])
2302 :     continue;
2303 :    
2304 :     if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= grad )
2305 :     MBmask[mbnum] = 0;
2306 :     if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= grad )
2307 :     MBmask[mbnum] = 0;
2308 :    
2309 : edgomez 195 }
2310 : chl 96
2311 : edgomez 851 emms();
2312 : chl 96
2313 : edgomez 851 do { /* until convergence */
2314 : chl 96
2315 : edgomez 851 a = b = c = n = 0;
2316 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2317 :     for (my = 0; my < MBh; my++)
2318 :     for (mx = 0; mx < MBw; mx++)
2319 :     {
2320 :     const int mbnum = mx + my * MBw;
2321 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2322 :     const VECTOR mv = pMB->mvs[0];
2323 : chl 96
2324 : edgomez 851 if (!MBmask[mbnum])
2325 :     continue;
2326 : chl 96
2327 : edgomez 851 n++;
2328 :     a += 16*mx+8;
2329 :     b += 16*my+8;
2330 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2331 : chl 96
2332 : edgomez 851 DtimesF[0] += (double)mv.x;
2333 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2334 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2335 :     DtimesF[3] += (double)mv.y;
2336 :     }
2337 : chl 96
2338 : edgomez 851 denom = a*a+b*b-c*n;
2339 : chl 96
2340 : edgomez 851 /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2341 :     /* D'*E*F has been calculated in the same loop as matrix */
2342 : chl 96
2343 : edgomez 851 sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2344 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2345 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2346 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2347 : chl 96
2348 : edgomez 851 sol[0] /= denom;
2349 :     sol[1] /= denom;
2350 :     sol[2] /= denom;
2351 :     sol[3] /= denom;
2352 : chl 96
2353 : edgomez 851 meanx = meany = 0.;
2354 :     oldnum = 0;
2355 :     for (my = 0; my < MBh; my++)
2356 :     for (mx = 0; mx < MBw; mx++)
2357 :     {
2358 :     const int mbnum = mx + my * MBw;
2359 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2360 :     const VECTOR mv = pMB->mvs[0];
2361 : chl 96
2362 : edgomez 851 if (!MBmask[mbnum])
2363 :     continue;
2364 : chl 184
2365 : edgomez 851 oldnum++;
2366 :     meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x );
2367 :     meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y );
2368 :     }
2369 : chl 96
2370 : edgomez 851 if (4*meanx > oldnum) /* better fit than 0.25 is useless */
2371 :     meanx /= oldnum;
2372 :     else
2373 :     meanx = 0.25;
2374 : edgomez 195
2375 : edgomez 851 if (4*meany > oldnum)
2376 :     meany /= oldnum;
2377 :     else
2378 :     meany = 0.25;
2379 : chl 96
2380 : edgomez 851 /* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2381 :     fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2382 :     */
2383 :     num = 0;
2384 :     for (my = 0; my < MBh; my++)
2385 :     for (mx = 0; mx < MBw; mx++)
2386 :     {
2387 :     const int mbnum = mx + my * MBw;
2388 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2389 :     const VECTOR mv = pMB->mvs[0];
2390 : chl 96
2391 : edgomez 851 if (!MBmask[mbnum])
2392 :     continue;
2393 : chl 96
2394 : edgomez 851 if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx )
2395 :     || ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) )
2396 :     MBmask[mbnum]=0;
2397 :     else
2398 :     num++;
2399 : chl 96 }
2400 : edgomez 195
2401 : edgomez 851 } while ( (oldnum != num) && (num>=4) );
2402 : edgomez 195
2403 : edgomez 851 if (num < 4)
2404 :     {
2405 :     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;
2406 :     } else {
2407 : edgomez 195
2408 : edgomez 851 gmc.duv[0].x=(int)(sol[0]+0.5);
2409 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2410 : chl 96
2411 : edgomez 851 gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2412 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2413 : chl 96
2414 : edgomez 851 gmc.duv[2].x=0;
2415 :     gmc.duv[2].y=0;
2416 :     }
2417 :     // 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);
2418 : chl 96
2419 : edgomez 851 free(MBmask);
2420 : chl 96
2421 : edgomez 851 return gmc;
2422 : chl 96 }
2423 :    
2424 : edgomez 851 // functions which perform BITS-based search/bitcount
2425 : chl 96
2426 : edgomez 851 static int
2427 :     CountMBBitsInter(SearchData * const Data,
2428 :     const MACROBLOCK * const pMBs, const int x, const int y,
2429 :     const MBParam * const pParam,
2430 :     const uint32_t MotionFlags)
2431 : chl 96 {
2432 : edgomez 851 int i, iDirection;
2433 :     int32_t bsad[5];
2434 : chl 141
2435 : edgomez 851 CheckCandidate = CheckCandidateBits16;
2436 : chl 96
2437 : edgomez 851 if (Data->qpel) {
2438 :     for(i = 0; i < 5; i++) {
2439 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2440 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2441 :     }
2442 :     Data->qpel_precision = 1;
2443 :     CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2444 : chl 96
2445 : edgomez 851 //checking if this vector is perfect. if it is, we stop.
2446 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0)
2447 :     return 0; //quick stop
2448 : edgomez 195
2449 : edgomez 851 if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search
2450 :     for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2451 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2452 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2453 :     Data->qpel_precision = 0;
2454 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2455 :     CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2456 :     }
2457 : edgomez 195
2458 : edgomez 851 } else { // not qpel
2459 : edgomez 195
2460 : edgomez 851 CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2461 :     //checking if this vector is perfect. if it is, we stop.
2462 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) {
2463 :     return 0; //inter
2464 :     }
2465 :     }
2466 : chl 96
2467 : edgomez 851 if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2468 : chl 96
2469 : edgomez 851 if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data);
2470 : chl 96
2471 : edgomez 851 if (Data->qpel) {
2472 :     if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search
2473 :     for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2474 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match
2475 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2476 :     }
2477 : edgomez 195
2478 : edgomez 851 // preparing for qpel-precision search
2479 :     Data->qpel_precision = 1;
2480 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2481 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
2482 :     }
2483 :     if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2484 :     }
2485 : chl 96
2486 : edgomez 851 if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction
2487 :     VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2488 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2489 :     CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2490 :     }
2491 :     return Data->iMinSAD[0];
2492 :     }
2493 : chl 96
2494 :    
2495 : edgomez 851 static int
2496 :     CountMBBitsInter4v(const SearchData * const Data,
2497 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2498 :     const int x, const int y,
2499 :     const MBParam * const pParam, const uint32_t MotionFlags,
2500 :     const VECTOR * const backup)
2501 :     {
2502 : chl 96
2503 : edgomez 851 int cbp = 0, bits = 0, t = 0, i, iDirection;
2504 :     SearchData Data2, *Data8 = &Data2;
2505 :     int sumx = 0, sumy = 0;
2506 :     int16_t in[64], coeff[64];
2507 : chl 96
2508 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2509 :     CheckCandidate = CheckCandidateBits8;
2510 : chl 96
2511 : edgomez 851 for (i = 0; i < 4; i++) {
2512 :     Data8->iMinSAD = Data->iMinSAD + i + 1;
2513 :     Data8->currentMV = Data->currentMV + i + 1;
2514 :     Data8->currentQMV = Data->currentQMV + i + 1;
2515 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2516 :     Data8->Ref = Data->Ref + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2517 :     Data8->RefH = Data->RefH + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2518 :     Data8->RefV = Data->RefV + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2519 :     Data8->RefHV = Data->RefHV + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2520 : edgomez 195
2521 : edgomez 851 if(Data->qpel) {
2522 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2523 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2524 :     Data8->predMV, Data8->iFcode, 0, 0);
2525 :     } else {
2526 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2527 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2528 :     Data8->predMV, Data8->iFcode, 0, 0);
2529 :     }
2530 : chl 96
2531 : 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,
2532 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2533 : chl 96
2534 : edgomez 851 *Data8->iMinSAD += t;
2535 : chl 96
2536 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2537 :     // checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far)
2538 :     if (Data8->qpel) {
2539 :     if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y))
2540 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2541 :     } else {
2542 :     if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y))
2543 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2544 :     }
2545 : edgomez 195
2546 : edgomez 851 if (Data8->qpel) {
2547 :     if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows
2548 :     int32_t s = *Data8->iMinSAD;
2549 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2550 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2551 :     Data8->qpel_precision = 0;
2552 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2553 :     pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2554 : edgomez 195
2555 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2556 :     CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2557 : edgomez 195
2558 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS)
2559 :     SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2560 : chl 96
2561 : edgomez 851 if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8);
2562 : chl 96
2563 : edgomez 851 if(s > *Data8->iMinSAD) { //we have found a better match
2564 :     Data8->currentQMV->x = 2*Data8->currentMV->x;
2565 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2566 :     }
2567 : chl 96
2568 : edgomez 851 Data8->qpel_precision = 1;
2569 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2570 :     pParam->width, pParam->height, Data8->iFcode, 1, 0);
2571 : chl 96
2572 : edgomez 851 }
2573 :     if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2574 : edgomez 195
2575 : edgomez 851 } else // not qpel
2576 :     if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement
2577 :    
2578 :     //checking vector equal to predicion
2579 :     if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) {
2580 :     const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2581 :     if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y))
2582 :     CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2583 : chl 141 }
2584 : edgomez 195
2585 : edgomez 851 bits += *Data8->iMinSAD;
2586 :     if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V
2587 :    
2588 :     // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else
2589 :     if(Data->qpel) {
2590 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2591 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2592 :     pMB->qmvs[i] = *Data8->currentQMV;
2593 :     sumx += Data8->currentQMV->x/2;
2594 :     sumy += Data8->currentQMV->y/2;
2595 :     } else {
2596 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2597 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2598 :     sumx += Data8->currentMV->x;
2599 :     sumy += Data8->currentMV->y;
2600 : chl 141 }
2601 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2602 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2603 :     if (Data8->temp[0]) cbp |= 1 << (5 - i);
2604 : chl 141 }
2605 :    
2606 : edgomez 851 if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma
2607 :     const uint8_t * ptr;
2608 :     sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2609 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2610 : chl 141
2611 : edgomez 851 //chroma U
2612 :     ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCU, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2613 :     transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2614 :     fdct(in);
2615 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2616 :     else i = quant4_inter(coeff, in, Data->lambda16);
2617 :     if (i > 0) {
2618 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2619 :     cbp |= 1 << (5 - 4);
2620 :     }
2621 : chl 141
2622 : edgomez 851 if (bits < *Data->iMinSAD) { // still possible
2623 :     //chroma V
2624 :     ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCV, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2625 :     transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2626 :     fdct(in);
2627 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2628 :     else i = quant4_inter(coeff, in, Data->lambda16);
2629 :     if (i > 0) {
2630 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2631 :     cbp |= 1 << (5 - 5);
2632 :     }
2633 :     bits += cbpy_tab[15-(cbp>>2)].len;
2634 :     bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2635 :     }
2636 : edgomez 195 }
2637 : chl 96
2638 : edgomez 851 return bits;
2639 :     }
2640 : chl 96
2641 :    
2642 : edgomez 851 static int
2643 :     CountMBBitsIntra(const SearchData * const Data)
2644 :     {
2645 :     int bits = 1; //this one is ac/dc prediction flag. always 1.
2646 :     int cbp = 0, i, t, dc = 0, b_dc = 1024;
2647 :     const uint32_t iQuant = Data->lambda16;
2648 :     int16_t in[64], coeff[64];
2649 : edgomez 195
2650 : edgomez 851 for(i = 0; i < 4; i++) {
2651 :     uint32_t iDcScaler = get_dc_scaler(iQuant, 1);
2652 : chl 96
2653 : edgomez 851 int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2654 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2655 :     fdct(in);
2656 :     b_dc = dc;
2657 :     dc = in[0];
2658 :     in[0] -= b_dc;
2659 :     if (Data->lambda8 == 0) quant_intra_c(coeff, in, iQuant, iDcScaler);
2660 :     else quant4_intra_c(coeff, in, iQuant, iDcScaler);
2661 : chl 141
2662 : edgomez 851 b_dc = dc;
2663 :     dc = coeff[0];
2664 :     if (i != 0) coeff[0] -= b_dc;
2665 : chl 181
2666 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;;
2667 :     Data->temp[i] = t;
2668 :     if (t != 0) cbp |= 1 << (5 - i);
2669 :     if (bits >= Data->iMinSAD[0]) break;
2670 :     }
2671 : chl 181
2672 : edgomez 851 if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma
2673 :     uint32_t iDcScaler = get_dc_scaler(iQuant, 0);
2674 :     //chroma U
2675 :     transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2676 :     fdct(in);
2677 :     in[0] -= 1024;
2678 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2679 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2680 : chl 96
2681 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2682 :     if (t != 0) cbp |= 1 << (5 - 4);
2683 :     Data->temp[4] = t;
2684 : chl 141
2685 : edgomez 851 if (bits < Data->iMinSAD[0]) {
2686 :     //chroma V
2687 :     transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2688 :     fdct(in);
2689 :     in[0] -= 1024;
2690 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2691 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2692 : chl 96
2693 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2694 :     if (t != 0) cbp |= 1 << (5 - 5);
2695 : chl 96
2696 : edgomez 851 Data->temp[5] = t;
2697 : edgomez 195
2698 : edgomez 851 bits += t = cbpy_tab[cbp>>2].len;
2699 :     Data->temp[6] = t;
2700 : chl 96
2701 : edgomez 851 bits += t = mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2702 :     Data->temp[7] = t;
2703 : edgomez 195
2704 : chl 96 }
2705 :     }
2706 :    
2707 : edgomez 851 return bits;
2708 : chl 96 }

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