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

Annotation of /trunk/xvidcore/src/motion/motion_est.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 897 - (view) (download)

1 : edgomez 851 /**************************************************************************
2 : Isibaar 3 *
3 : edgomez 851 * XVID MPEG-4 VIDEO CODEC
4 :     * motion estimation
5 : chl 259 *
6 : edgomez 851 * This program is an implementation of a part of one or more MPEG-4
7 :     * Video tools as specified in ISO/IEC 14496-2 standard. Those intending
8 :     * to use this software module in hardware or software products are
9 :     * advised that its use may infringe existing patents or copyrights, and
10 :     * any such use would be at such party's own risk. The original
11 :     * developer of this software module and his/her company, and subsequent
12 :     * editors and their companies, will have no liability for use of this
13 :     * software or modifications or derivatives thereof.
14 : chl 259 *
15 : edgomez 851 * This program is free software; you can redistribute it and/or modify
16 :     * it under the terms of the GNU General Public License as published by
17 :     * the Free Software Foundation; either version 2 of the License, or
18 :     * (at your option) any later version.
19 : chl 259 *
20 : edgomez 851 * This program is distributed in the hope that it will be useful,
21 :     * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 :     * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 :     * GNU General Public License for more details.
24 : chl 259 *
25 : edgomez 851 * You should have received a copy of the GNU General Public License
26 :     * along with this program; if not, write to the Free Software
27 :     * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 : chl 259 *
29 :     *************************************************************************/
30 :    
31 : Isibaar 3 #include <assert.h>
32 :     #include <stdio.h>
33 : chl 96 #include <stdlib.h>
34 : edgomez 851 #include <string.h> // memcpy
35 :     #include <math.h> // lrint
36 : Isibaar 3
37 :     #include "../encoder.h"
38 :     #include "../utils/mbfunctions.h"
39 :     #include "../prediction/mbprediction.h"
40 :     #include "../global.h"
41 :     #include "../utils/timer.h"
42 : edgomez 851 #include "../image/interpolate8x8.h"
43 :     #include "motion_est.h"
44 : suxen_drol 118 #include "motion.h"
45 : Isibaar 3 #include "sad.h"
46 : edgomez 851 #include "../utils/emms.h"
47 :     #include "../dct/fdct.h"
48 : Isibaar 3
49 : edgomez 872 /*****************************************************************************
50 :     * Modified rounding tables -- declared in motion.h
51 :     * Original tables see ISO spec tables 7-6 -> 7-9
52 :     ****************************************************************************/
53 :    
54 :     const uint32_t roundtab[16] =
55 :     {0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 };
56 :    
57 :     /* K = 4 */
58 :     const uint32_t roundtab_76[16] =
59 :     { 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 };
60 :    
61 :     /* K = 2 */
62 :     const uint32_t roundtab_78[8] =
63 :     { 0, 0, 1, 1, 0, 0, 0, 1 };
64 :    
65 :     /* K = 1 */
66 :     const uint32_t roundtab_79[4] =
67 :     { 0, 1, 0, 0 };
68 :    
69 : edgomez 851 #define INITIAL_SKIP_THRESH (10)
70 :     #define FINAL_SKIP_THRESH (50)
71 :     #define MAX_SAD00_FOR_SKIP (20)
72 :     #define MAX_CHROMA_SAD_FOR_SKIP (22)
73 : Isibaar 3
74 : edgomez 851 #define CHECK_CANDIDATE(X,Y,D) { \
75 :     CheckCandidate((X),(Y), (D), &iDirection, data ); }
76 : Isibaar 3
77 : edgomez 872 /*****************************************************************************
78 :     * Code
79 :     ****************************************************************************/
80 :    
81 : edgomez 851 static __inline uint32_t
82 :     d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
83 :     {
84 :     int xb, yb;
85 :     x = qpel ? x<<1 : x;
86 :     y = qpel ? y<<1 : y;
87 :     if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
88 : chl 141
89 : edgomez 851 x -= pred.x;
90 :     y -= pred.y;
91 : chl 141
92 : edgomez 851 if (x) {
93 :     x = ABS(x);
94 :     x += (1 << (iFcode - 1)) - 1;
95 :     x >>= (iFcode - 1);
96 :     if (x > 32) x = 32;
97 :     xb = mvtab[x] + iFcode;
98 :     } else xb = 1;
99 : chl 141
100 : edgomez 851 if (y) {
101 :     y = ABS(y);
102 :     y += (1 << (iFcode - 1)) - 1;
103 :     y >>= (iFcode - 1);
104 :     if (y > 32) y = 32;
105 :     yb = mvtab[y] + iFcode;
106 :     } else yb = 1;
107 :     return xb + yb;
108 :     }
109 : chl 141
110 : edgomez 851 static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data)
111 :     {
112 :     int sad;
113 :     const uint32_t stride = data->iEdgedWidth/2;
114 :     uint8_t * f_refu = data->RefQ,
115 :     * f_refv = data->RefQ + 8,
116 :     * b_refu = data->RefQ + 16,
117 :     * b_refv = data->RefQ + 24;
118 : Isibaar 3
119 : edgomez 851 switch (((fx & 1) << 1) | (fy & 1)) {
120 :     case 0:
121 :     fx = fx / 2; fy = fy / 2;
122 :     f_refu = (uint8_t*)data->RefCU + fy * stride + fx, stride;
123 :     f_refv = (uint8_t*)data->RefCV + fy * stride + fx, stride;
124 :     break;
125 :     case 1:
126 :     fx = fx / 2; fy = (fy - 1) / 2;
127 :     interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
128 :     interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
129 :     break;
130 :     case 2:
131 :     fx = (fx - 1) / 2; fy = fy / 2;
132 :     interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
133 :     interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
134 :     break;
135 :     default:
136 :     fx = (fx - 1) / 2; fy = (fy - 1) / 2;
137 :     interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
138 :     interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
139 :     break;
140 :     }
141 : Isibaar 3
142 : edgomez 851 switch (((bx & 1) << 1) | (by & 1)) {
143 :     case 0:
144 :     bx = bx / 2; by = by / 2;
145 :     b_refu = (uint8_t*)data->b_RefCU + by * stride + bx, stride;
146 :     b_refv = (uint8_t*)data->b_RefCV + by * stride + bx, stride;
147 :     break;
148 :     case 1:
149 :     bx = bx / 2; by = (by - 1) / 2;
150 :     interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
151 :     interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
152 :     break;
153 :     case 2:
154 :     bx = (bx - 1) / 2; by = by / 2;
155 :     interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
156 :     interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
157 :     break;
158 :     default:
159 :     bx = (bx - 1) / 2; by = (by - 1) / 2;
160 :     interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
161 :     interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
162 :     break;
163 :     }
164 : Isibaar 3
165 : edgomez 851 sad = sad8bi(data->CurU, b_refu, f_refu, stride);
166 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
167 : Isibaar 3
168 : edgomez 851 return sad;
169 :     }
170 : Isibaar 3
171 :    
172 : edgomez 851 static int32_t
173 :     ChromaSAD(int dx, int dy, const SearchData * const data)
174 :     {
175 :     int sad;
176 :     const uint32_t stride = data->iEdgedWidth/2;
177 : Isibaar 3
178 : edgomez 851 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently
179 :     data->temp[5] = dx; data->temp[6] = dy; // backup
180 : Isibaar 3
181 : edgomez 851 switch (((dx & 1) << 1) | (dy & 1)) {
182 :     case 0:
183 :     dx = dx / 2; dy = dy / 2;
184 :     sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride);
185 :     sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride);
186 :     break;
187 :     case 1:
188 :     dx = dx / 2; dy = (dy - 1) / 2;
189 :     sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride);
190 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride);
191 :     break;
192 :     case 2:
193 :     dx = (dx - 1) / 2; dy = dy / 2;
194 :     sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride);
195 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride);
196 :     break;
197 :     default:
198 :     dx = (dx - 1) / 2; dy = (dy - 1) / 2;
199 :     interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding);
200 :     sad = sad8(data->CurU, data->RefQ, stride);
201 :    
202 :     interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding);
203 :     sad += sad8(data->CurV, data->RefQ, stride);
204 :     break;
205 :     }
206 :     data->temp[7] = sad; //backup, part 2
207 :     return sad;
208 : Isibaar 3 }
209 :    
210 : edgomez 851 static __inline const uint8_t *
211 :     GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
212 : Isibaar 3 {
213 : edgomez 851 // dir : 0 = forward, 1 = backward
214 :     switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) {
215 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
216 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
217 :     case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
218 :     case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
219 :     case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth);
220 :     case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
221 :     case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
222 :     default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
223 :     }
224 : Isibaar 3 }
225 :    
226 : edgomez 851 // this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate
227 :     static __inline const uint8_t *
228 :     GetReference(const int x, const int y, const SearchData * const data)
229 : Isibaar 3 {
230 : edgomez 851 switch ( ((x&1)<<1) | (y&1) ) {
231 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
232 :     case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
233 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
234 :     default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2
235 :     }
236 : Isibaar 3 }
237 : edgomez 851
238 :     static uint8_t *
239 :     Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
240 : Isibaar 3 {
241 : edgomez 851 // create or find a qpel-precision reference picture; return pointer to it
242 :     uint8_t * Reference = data->RefQ + 16*dir;
243 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
244 :     const uint32_t rounding = data->rounding;
245 :     const int halfpel_x = x/2;
246 :     const int halfpel_y = y/2;
247 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
248 : suxen_drol 136
249 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
250 :     ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
251 :     switch( ((x&1)<<1) + (y&1) ) {
252 :     case 0: // pure halfpel position
253 :     return (uint8_t *) ref1;
254 :     break;
255 : chl 184
256 : edgomez 851 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
257 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
258 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
259 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
260 :     break;
261 : suxen_drol 136
262 : edgomez 851 case 2: // x qpel, y halfpel - left or right during qpel refinement
263 :     ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
264 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
265 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
266 :     break;
267 : edgomez 195
268 : edgomez 851 default: // x and y in qpel resolution - the "corners" (top left/right and
269 :     // bottom left/right) during qpel refinement
270 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
271 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
272 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
273 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
274 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
275 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
276 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
277 :     break;
278 :     }
279 :     return Reference;
280 :     }
281 : Isibaar 3
282 : edgomez 851 static uint8_t *
283 :     Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
284 :     {
285 :     // create or find a qpel-precision reference picture; return pointer to it
286 :     uint8_t * Reference = data->RefQ + 16*dir;
287 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
288 :     const uint32_t rounding = data->rounding;
289 :     const int halfpel_x = x/2;
290 :     const int halfpel_y = y/2;
291 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
292 : chl 326
293 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
294 :     switch( ((x&1)<<1) + (y&1) ) {
295 :     case 3: // x and y in qpel resolution - the "corners" (top left/right and
296 :     // bottom left/right) during qpel refinement
297 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
298 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
299 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
300 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
301 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
302 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
303 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
304 :     break;
305 : chl 184
306 : edgomez 851 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
307 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
308 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
309 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
310 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
311 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
312 :     break;
313 : chl 184
314 : edgomez 851 case 2: // x qpel, y halfpel - left or right during qpel refinement
315 :     ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
316 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
317 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
318 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
319 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
320 :     break;
321 : edgomez 195
322 : edgomez 851 case 0: // pure halfpel position
323 :     return (uint8_t *) ref1;
324 :     }
325 :     return Reference;
326 :     }
327 : edgomez 195
328 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS START */
329 : edgomez 195
330 : edgomez 851 static void
331 :     CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
332 :     {
333 :     int xc, yc;
334 :     const uint8_t * Reference;
335 :     VECTOR * current;
336 :     int32_t sad; uint32_t t;
337 : edgomez 195
338 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
339 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
340 : edgomez 195
341 : edgomez 851 if (!data->qpel_precision) {
342 :     Reference = GetReference(x, y, data);
343 :     current = data->currentMV;
344 :     xc = x; yc = y;
345 :     } else { // x and y are in 1/4 precision
346 :     Reference = Interpolate16x16qpel(x, y, 0, data);
347 :     xc = x/2; yc = y/2; //for chroma sad
348 :     current = data->currentQMV;
349 :     }
350 : edgomez 195
351 : edgomez 851 sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
352 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
353 : edgomez 195
354 : edgomez 851 sad += (data->lambda16 * t * sad)>>10;
355 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
356 : chl 175
357 : edgomez 851 if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
358 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
359 : edgomez 195
360 : edgomez 851 if (sad < data->iMinSAD[0]) {
361 :     data->iMinSAD[0] = sad;
362 :     current[0].x = x; current[0].y = y;
363 :     *dir = Direction;
364 :     }
365 : edgomez 195
366 : edgomez 851 if (data->temp[1] < data->iMinSAD[1]) {
367 :     data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
368 :     if (data->temp[2] < data->iMinSAD[2]) {
369 :     data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
370 :     if (data->temp[3] < data->iMinSAD[3]) {
371 :     data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
372 :     if (data->temp[4] < data->iMinSAD[4]) {
373 :     data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
374 :    
375 : Isibaar 3 }
376 :    
377 : edgomez 851 static void
378 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
379 :     {
380 :     int32_t sad; uint32_t t;
381 :     const uint8_t * Reference;
382 : chl 326
383 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
384 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
385 : Isibaar 3
386 : edgomez 851 if (!data->qpel_precision) Reference = GetReference(x, y, data);
387 :     else Reference = Interpolate8x8qpel(x, y, 0, 0, data);
388 : Isibaar 3
389 : edgomez 851 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
390 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
391 : Isibaar 3
392 : edgomez 851 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
393 : Isibaar 3
394 : edgomez 851 if (sad < *(data->iMinSAD)) {
395 :     *(data->iMinSAD) = sad;
396 :     data->currentMV->x = x; data->currentMV->y = y;
397 :     *dir = Direction;
398 :     }
399 : Isibaar 3 }
400 :    
401 :    
402 : edgomez 851 static void
403 :     CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
404 :     {
405 :     uint32_t t;
406 :     const uint8_t * Reference;
407 :    
408 :     if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value
409 :     (x > data->max_dx) || (x < data->min_dx)
410 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
411 :    
412 :     Reference = GetReference(x, y, data);
413 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
414 :    
415 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
416 :    
417 :     data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
418 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
419 :    
420 :     if (data->temp[0] < data->iMinSAD[0]) {
421 :     data->iMinSAD[0] = data->temp[0];
422 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
423 :     *dir = Direction; }
424 :    
425 :     if (data->temp[1] < data->iMinSAD[1]) {
426 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
427 :     if (data->temp[2] < data->iMinSAD[2]) {
428 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
429 :     if (data->temp[3] < data->iMinSAD[3]) {
430 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
431 :     if (data->temp[4] < data->iMinSAD[4]) {
432 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
433 : Isibaar 3 }
434 : edgomez 195
435 : edgomez 851 static void
436 :     CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
437 :     {
438 :     int32_t sad, xc, yc;
439 :     const uint8_t * Reference;
440 :     uint32_t t;
441 :     VECTOR * current;
442 :    
443 :     if ( (x > data->max_dx) | ( x < data->min_dx)
444 :     | (y > data->max_dy) | (y < data->min_dy) ) return;
445 :    
446 :     if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value
447 :    
448 :     if (data->qpel_precision) { // x and y are in 1/4 precision
449 :     Reference = Interpolate16x16qpel(x, y, 0, data);
450 :     current = data->currentQMV;
451 :     xc = x/2; yc = y/2;
452 :     } else {
453 :     Reference = GetReference(x, y, data);
454 :     current = data->currentMV;
455 :     xc = x; yc = y;
456 :     }
457 :     t = d_mv_bits(x, y, data->predMV, data->iFcode,
458 :     data->qpel^data->qpel_precision, data->rrv);
459 :    
460 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
461 :     sad += (data->lambda16 * t * sad)>>10;
462 :    
463 :     if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
464 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
465 :    
466 :     if (sad < *(data->iMinSAD)) {
467 :     *(data->iMinSAD) = sad;
468 :     current->x = x; current->y = y;
469 :     *dir = Direction;
470 :     }
471 : chl 96 }
472 : Isibaar 3
473 : edgomez 851 static void
474 :     CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
475 :     {
476 :     // maximum speed - for P/B/I decision
477 :     int32_t sad;
478 :    
479 :     if ( (x > data->max_dx) || (x < data->min_dx)
480 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
481 :    
482 :     sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
483 :     data->iEdgedWidth, data->temp+1);
484 :    
485 :     if (sad < *(data->iMinSAD)) {
486 :     *(data->iMinSAD) = sad;
487 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
488 :     *dir = Direction;
489 :     }
490 :     if (data->temp[1] < data->iMinSAD[1]) {
491 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
492 :     if (data->temp[2] < data->iMinSAD[2]) {
493 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
494 :     if (data->temp[3] < data->iMinSAD[3]) {
495 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
496 :     if (data->temp[4] < data->iMinSAD[4]) {
497 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
498 :    
499 : Isibaar 3 }
500 :    
501 : edgomez 851 static void
502 :     CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
503 :     {
504 :     int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
505 :     uint32_t t;
506 :     const uint8_t *ReferenceF, *ReferenceB;
507 :     VECTOR *current;
508 :    
509 :     if ( (xf > data->max_dx) | (xf < data->min_dx)
510 :     | (yf > data->max_dy) | (yf < data->min_dy) ) return;
511 :    
512 :     if (!data->qpel_precision) {
513 :     ReferenceF = GetReference(xf, yf, data);
514 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
515 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
516 :     current = data->currentMV;
517 :     xcf = xf; ycf = yf;
518 :     xcb = xb; ycb = yb;
519 :     } else {
520 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
521 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
522 :     current = data->currentQMV;
523 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
524 :     xcf = xf/2; ycf = yf/2;
525 :     xcb = xb/2; ycb = yb/2;
526 :     }
527 :    
528 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
529 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
530 :    
531 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
532 :     sad += (data->lambda16 * t * sad)>>10;
533 :    
534 :     if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
535 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
536 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
537 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
538 :    
539 :     if (sad < *(data->iMinSAD)) {
540 :     *(data->iMinSAD) = sad;
541 :     current->x = xf; current->y = yf;
542 :     *dir = Direction;
543 :     }
544 : Isibaar 3 }
545 :    
546 : edgomez 851 static void
547 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
548 :     {
549 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
550 :     uint32_t k;
551 :     const uint8_t *ReferenceF;
552 :     const uint8_t *ReferenceB;
553 :     VECTOR mvs, b_mvs;
554 :    
555 :     if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return;
556 :    
557 :     for (k = 0; k < 4; k++) {
558 :     mvs.x = data->directmvF[k].x + x;
559 :     b_mvs.x = ((x == 0) ?
560 :     data->directmvB[k].x
561 :     : mvs.x - data->referencemv[k].x);
562 :    
563 :     mvs.y = data->directmvF[k].y + y;
564 :     b_mvs.y = ((y == 0) ?
565 :     data->directmvB[k].y
566 :     : mvs.y - data->referencemv[k].y);
567 :    
568 :     if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx)
569 :     | (mvs.y > data->max_dy) | (mvs.y < data->min_dy)
570 :     | (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx)
571 :     | (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return;
572 :    
573 :     if (data->qpel) {
574 :     xcf += mvs.x/2; ycf += mvs.y/2;
575 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
576 :     } else {
577 :     xcf += mvs.x; ycf += mvs.y;
578 :     xcb += b_mvs.x; ycb += b_mvs.y;
579 :     mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway
580 :     b_mvs.x *= 2; b_mvs.y *= 2;
581 :     }
582 :    
583 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
584 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
585 :    
586 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
587 :     ReferenceF, ReferenceB, data->iEdgedWidth);
588 :     if (sad > *(data->iMinSAD)) return;
589 :     }
590 :    
591 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
592 :    
593 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
594 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
595 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
596 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
597 :    
598 :     if (sad < *(data->iMinSAD)) {
599 :     *(data->iMinSAD) = sad;
600 :     data->currentMV->x = x; data->currentMV->y = y;
601 :     *dir = Direction;
602 :     }
603 : Isibaar 3 }
604 :    
605 : edgomez 851 static void
606 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
607 : Isibaar 3 {
608 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
609 :     const uint8_t *ReferenceF;
610 :     const uint8_t *ReferenceB;
611 :     VECTOR mvs, b_mvs;
612 : Isibaar 3
613 : edgomez 851 if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return;
614 : Isibaar 3
615 : edgomez 851 mvs.x = data->directmvF[0].x + x;
616 :     b_mvs.x = ((x == 0) ?
617 :     data->directmvB[0].x
618 :     : mvs.x - data->referencemv[0].x);
619 : Isibaar 3
620 : edgomez 851 mvs.y = data->directmvF[0].y + y;
621 :     b_mvs.y = ((y == 0) ?
622 :     data->directmvB[0].y
623 :     : mvs.y - data->referencemv[0].y);
624 : Isibaar 3
625 : edgomez 851 if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx)
626 :     | (mvs.y > data->max_dy) | (mvs.y < data->min_dy)
627 :     | (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx)
628 :     | (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return;
629 : Isibaar 3
630 : edgomez 851 if (data->qpel) {
631 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
632 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
633 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
634 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
635 :     } else {
636 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
637 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
638 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
639 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
640 :     }
641 :    
642 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
643 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
644 :    
645 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
646 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
647 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
648 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
649 :    
650 :     if (sad < *(data->iMinSAD)) {
651 :     *(data->iMinSAD) = sad;
652 :     data->currentMV->x = x; data->currentMV->y = y;
653 :     *dir = Direction;
654 :     }
655 : Isibaar 3 }
656 :    
657 : edgomez 851
658 :     static void
659 :     CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
660 : Isibaar 3 {
661 :    
662 : edgomez 851 static int16_t in[64], coeff[64];
663 :     int32_t bits = 0, sum;
664 :     VECTOR * current;
665 :     const uint8_t * ptr;
666 :     int i, cbp = 0, t, xc, yc;
667 : edgomez 195
668 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
669 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
670 : edgomez 195
671 : edgomez 851 if (!data->qpel_precision) {
672 :     ptr = GetReference(x, y, data);
673 :     current = data->currentMV;
674 :     xc = x; yc = y;
675 :     } else { // x and y are in 1/4 precision
676 :     ptr = Interpolate16x16qpel(x, y, 0, data);
677 :     current = data->currentQMV;
678 :     xc = x/2; yc = y/2;
679 :     }
680 : Isibaar 3
681 : edgomez 851 for(i = 0; i < 4; i++) {
682 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
683 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
684 :     fdct(in);
685 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
686 :     else sum = quant4_inter(coeff, in, data->lambda16);
687 :     if (sum > 0) {
688 :     cbp |= 1 << (5 - i);
689 :     bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
690 :     } else data->temp[i] = 0;
691 :     }
692 : Isibaar 3
693 : edgomez 851 bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
694 : edgomez 195
695 : edgomez 851 if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma
696 :     xc = (xc >> 1) + roundtab_79[xc & 0x3];
697 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
698 :    
699 :     //chroma U
700 :     ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCU, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
701 :     transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
702 :     fdct(in);
703 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
704 :     else sum = quant4_inter(coeff, in, data->lambda16);
705 :     if (sum > 0) {
706 :     cbp |= 1 << (5 - 4);
707 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
708 : chl 344 }
709 : edgomez 851
710 :     if (bits < data->iMinSAD[0]) {
711 :     //chroma V
712 :     ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCV, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
713 :     transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2);
714 :     fdct(in);
715 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
716 :     else sum = quant4_inter(coeff, in, data->lambda16);
717 :     if (sum > 0) {
718 :     cbp |= 1 << (5 - 5);
719 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
720 :     }
721 :     }
722 : edgomez 78 }
723 : edgomez 851
724 : chl 897 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
725 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
726 :    
727 :     if (bits < data->iMinSAD[0]) {
728 :     data->iMinSAD[0] = bits;
729 :     current[0].x = x; current[0].y = y;
730 :     *dir = Direction;
731 :     }
732 :    
733 :     if (data->temp[0] + t < data->iMinSAD[1]) {
734 :     data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
735 :     if (data->temp[1] < data->iMinSAD[2]) {
736 :     data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
737 :     if (data->temp[2] < data->iMinSAD[3]) {
738 :     data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
739 :     if (data->temp[3] < data->iMinSAD[4]) {
740 :     data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
741 :    
742 : Isibaar 3 }
743 : edgomez 851 static void
744 :     CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
745 : chl 96 {
746 :    
747 : edgomez 851 static int16_t in[64], coeff[64];
748 :     int32_t sum, bits;
749 :     VECTOR * current;
750 :     const uint8_t * ptr;
751 :     int cbp;
752 : edgomez 195
753 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
754 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
755 : edgomez 195
756 : edgomez 851 if (!data->qpel_precision) {
757 :     ptr = GetReference(x, y, data);
758 :     current = data->currentMV;
759 :     } else { // x and y are in 1/4 precision
760 :     ptr = Interpolate8x8qpel(x, y, 0, 0, data);
761 :     current = data->currentQMV;
762 :     }
763 : chl 96
764 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
765 :     fdct(in);
766 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
767 :     else sum = quant4_inter(coeff, in, data->lambda16);
768 :     if (sum > 0) {
769 :     bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
770 :     cbp = 1;
771 :     } else cbp = bits = 0;
772 : chl 96
773 : edgomez 851 bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
774 : chl 96
775 : edgomez 851 if (bits < data->iMinSAD[0]) {
776 :     data->temp[0] = cbp;
777 :     data->iMinSAD[0] = bits;
778 :     current[0].x = x; current[0].y = y;
779 :     *dir = Direction;
780 :     }
781 :     }
782 : chl 96
783 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
784 : edgomez 195
785 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
786 : chl 96
787 : edgomez 851 static void
788 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
789 :     {
790 : edgomez 195
791 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
792 : edgomez 195
793 : edgomez 851 int iDirection;
794 : edgomez 195
795 : edgomez 851 for(;;) { //forever
796 :     iDirection = 0;
797 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
798 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
799 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
800 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
801 : edgomez 195
802 : edgomez 851 /* now we're doing diagonal checks near our candidate */
803 : edgomez 195
804 : edgomez 851 if (iDirection) { //if anything found
805 :     bDirection = iDirection;
806 :     iDirection = 0;
807 :     x = data->currentMV->x; y = data->currentMV->y;
808 :     if (bDirection & 3) { //our candidate is left or right
809 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
810 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
811 :     } else { // what remains here is up or down
812 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
813 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
814 :     }
815 : edgomez 195
816 : edgomez 851 if (iDirection) {
817 :     bDirection += iDirection;
818 :     x = data->currentMV->x; y = data->currentMV->y;
819 :     }
820 :     } else { //about to quit, eh? not so fast....
821 :     switch (bDirection) {
822 :     case 2:
823 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
824 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
825 : edgomez 195 break;
826 : edgomez 851 case 1:
827 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
828 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
829 : edgomez 195 break;
830 : edgomez 851 case 2 + 4:
831 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
832 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
833 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
834 :     break;
835 :     case 4:
836 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
837 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
838 :     break;
839 : edgomez 195 case 8:
840 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
841 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
842 : edgomez 195 break;
843 : edgomez 851 case 1 + 4:
844 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
845 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
846 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
847 : edgomez 195 break;
848 : edgomez 851 case 2 + 8:
849 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
850 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
851 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
852 :     break;
853 :     case 1 + 8:
854 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
855 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
856 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
857 :     break;
858 :     default: //1+2+4+8 == we didn't find anything at all
859 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
860 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
861 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
862 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
863 :     break;
864 : chl 96 }
865 : edgomez 851 if (!iDirection) break; //ok, the end. really
866 :     bDirection = iDirection;
867 :     x = data->currentMV->x; y = data->currentMV->y;
868 : chl 345 }
869 : edgomez 195 }
870 : chl 96 }
871 :    
872 : edgomez 851 static void
873 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
874 : chl 96 {
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+16+64);
880 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
881 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
882 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
883 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
884 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
885 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
886 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
887 : chl 96
888 : edgomez 851 bDirection = iDirection;
889 :     x = data->currentMV->x; y = data->currentMV->y;
890 :     } while (iDirection);
891 : chl 96 }
892 :    
893 : edgomez 851 static void
894 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
895 : chl 181 {
896 :    
897 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
898 :    
899 : edgomez 851 int iDirection;
900 : edgomez 195
901 : edgomez 851 do {
902 :     iDirection = 0;
903 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
904 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
905 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
906 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
907 :    
908 :     /* now we're doing diagonal checks near our candidate */
909 :    
910 :     if (iDirection) { //checking if anything found
911 :     bDirection = iDirection;
912 : chl 181 iDirection = 0;
913 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
914 :     if (bDirection & 3) { //our candidate is left or right
915 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
916 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
917 :     } else { // what remains here is up or down
918 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
919 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
920 :     }
921 :     bDirection += iDirection;
922 :     x = data->currentMV->x; y = data->currentMV->y;
923 :     }
924 :     }
925 :     while (iDirection);
926 :     }
927 : chl 181
928 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
929 : chl 181
930 : edgomez 851 static void
931 :     SubpelRefine(const SearchData * const data)
932 :     {
933 :     /* Do a half-pel or q-pel refinement */
934 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
935 :     int iDirection; //only needed because macro expects it
936 : chl 181
937 : edgomez 851 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
938 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
939 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
940 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
941 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
942 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
943 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
944 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
945 :     }
946 : chl 181
947 : edgomez 851 static __inline int
948 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
949 :     const int x, const int y,
950 :     const uint32_t stride, const uint32_t iQuant, int rrv)
951 : chl 181
952 : edgomez 851 {
953 :     if(!rrv) {
954 :     uint32_t sadC = sad8(current->u + x*8 + y*stride*8,
955 :     reference->u + x*8 + y*stride*8, stride);
956 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
957 :     sadC += sad8(current->v + (x + y*stride)*8,
958 :     reference->v + (x + y*stride)*8, stride);
959 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
960 :     return 1;
961 : chl 181
962 : edgomez 851 } else {
963 :     uint32_t sadC = sad16(current->u + x*16 + y*stride*16,
964 :     reference->u + x*16 + y*stride*16, stride, 256*4096);
965 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
966 :     sadC += sad16(current->v + (x + y*stride)*16,
967 :     reference->v + (x + y*stride)*16, stride, 256*4096);
968 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
969 :     return 1;
970 : chl 181 }
971 : edgomez 851 }
972 : edgomez 504
973 : edgomez 851 static __inline void
974 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
975 :     {
976 :     pMB->mode = MODE_NOT_CODED;
977 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
978 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
979 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
980 : chl 181 }
981 :    
982 : edgomez 851 bool
983 :     MotionEstimation(MBParam * const pParam,
984 :     FRAMEINFO * const current,
985 :     FRAMEINFO * const reference,
986 :     const IMAGE * const pRefH,
987 :     const IMAGE * const pRefV,
988 :     const IMAGE * const pRefHV,
989 :     const uint32_t iLimit)
990 : chl 181 {
991 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
992 :     const IMAGE *const pCurrent = &current->image;
993 :     const IMAGE *const pRef = &reference->image;
994 : chl 181
995 : edgomez 851 uint32_t mb_width = pParam->mb_width;
996 :     uint32_t mb_height = pParam->mb_height;
997 :     const uint32_t iEdgedWidth = pParam->edged_width;
998 :     const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags);
999 : chl 181
1000 : edgomez 851 uint32_t x, y;
1001 :     uint32_t iIntra = 0;
1002 :     int32_t quant = current->quant, sad00;
1003 : chl 181
1004 : edgomez 851 // some pre-initialized thingies for SearchP
1005 :     int32_t temp[8];
1006 :     VECTOR currentMV[5];
1007 :     VECTOR currentQMV[5];
1008 :     int32_t iMinSAD[5];
1009 :     SearchData Data;
1010 :     memset(&Data, 0, sizeof(SearchData));
1011 :     Data.iEdgedWidth = iEdgedWidth;
1012 :     Data.currentMV = currentMV;
1013 :     Data.currentQMV = currentQMV;
1014 :     Data.iMinSAD = iMinSAD;
1015 :     Data.temp = temp;
1016 :     Data.iFcode = current->fcode;
1017 :     Data.rounding = pParam->m_rounding_type;
1018 :     Data.qpel = pParam->m_quarterpel;
1019 :     Data.chroma = MotionFlags & PMV_CHROMA16;
1020 :     Data.rrv = current->global_flags & XVID_REDUCED;
1021 : edgomez 195
1022 : edgomez 851 if ((current->global_flags & XVID_REDUCED)) {
1023 :     mb_width = (pParam->width + 31) / 32;
1024 :     mb_height = (pParam->height + 31) / 32;
1025 :     Data.qpel = 0;
1026 :     }
1027 : chl 181
1028 : edgomez 851 Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose)
1029 :     if (sadInit) (*sadInit) ();
1030 : chl 181
1031 : edgomez 851 for (y = 0; y < mb_height; y++) {
1032 :     for (x = 0; x < mb_width; x++) {
1033 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1034 : chl 181
1035 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1036 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1037 :     pRef->y + (x + y * iEdgedWidth) * 16,
1038 :     pParam->edged_width, pMB->sad8 );
1039 : chl 181
1040 : edgomez 851 else pMB->sad16 =
1041 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1042 :     pRef->y + (x + y * iEdgedWidth) * 32,
1043 :     pParam->edged_width, pMB->sad8 );
1044 : chl 181
1045 : edgomez 851 if (Data.chroma) {
1046 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1047 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1048 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1049 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1050 :     pMB->sad16 += Data.temp[7];
1051 :     }
1052 :    
1053 :     sad00 = pMB->sad16;
1054 :    
1055 :     if (!(current->global_flags & XVID_LUMIMASKING)) {
1056 :     pMB->dquant = NO_CHANGE;
1057 :     } else {
1058 :     if (pMB->dquant != NO_CHANGE) {
1059 :     quant += DQtab[pMB->dquant];
1060 :     if (quant > 31) quant = 31;
1061 :     else if (quant < 1) quant = 1;
1062 : chl 181 }
1063 : edgomez 851 }
1064 :     pMB->quant = current->quant;
1065 : chl 181
1066 : edgomez 851 //initial skip decision
1067 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1068 :     if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1069 :     if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) )
1070 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1071 :     SkipMacroblockP(pMB, sad00);
1072 :     continue;
1073 :     }
1074 :     }
1075 :    
1076 :     SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1077 :     y, MotionFlags, current->global_flags, pMB->quant,
1078 :     &Data, pParam, pMBs, reference->mbs,
1079 :     current->global_flags & XVID_INTER4V, pMB);
1080 :    
1081 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
1082 :     if (!(current->global_flags & XVID_GMC)) {
1083 :     if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) {
1084 :     if (!(current->global_flags & XVID_MODEDECISION_BITS)) {
1085 :     if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) )
1086 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv))
1087 :     SkipMacroblockP(pMB, sad00);
1088 :     } else { // BITS mode decision
1089 :     if (pMB->sad16 > 10)
1090 :     SkipMacroblockP(pMB, sad00); // more than 10 bits would be used for this MB - skip
1091 :    
1092 :     }
1093 : chl 181 }
1094 :     }
1095 : edgomez 851 if (pMB->mode == MODE_INTRA)
1096 :     if (++iIntra > iLimit) return 1;
1097 : chl 181 }
1098 :     }
1099 : edgomez 851
1100 :     if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */
1101 :     {
1102 :     current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1103 :     }
1104 :     return 0;
1105 : chl 181 }
1106 :    
1107 :    
1108 : edgomez 851 static __inline int
1109 :     make_mask(const VECTOR * const pmv, const int i)
1110 : chl 96 {
1111 : edgomez 851 int mask = 255, j;
1112 :     for (j = 0; j < i; j++) {
1113 :     if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
1114 :     if (pmv[i].x == pmv[j].x) {
1115 :     if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1116 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1117 :     } else
1118 :     if (pmv[i].y == pmv[j].y) {
1119 :     if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1120 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1121 :     }
1122 :     }
1123 :     return mask;
1124 :     }
1125 : edgomez 195
1126 : edgomez 851 static __inline void
1127 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1128 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1129 :     {
1130 : chl 96
1131 : edgomez 851 //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1132 :     if (rrv) { iWcount /= 2; iHcount /= 2; }
1133 : edgomez 195
1134 : edgomez 851 if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
1135 :     pmv[5].x = EVEN(pmv[3].x);
1136 :     pmv[5].y = EVEN(pmv[3].y);
1137 :     } else pmv[5].x = pmv[5].y = 0;
1138 : chl 96
1139 : edgomez 851 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1140 :     else pmv[3].x = pmv[3].y = 0;
1141 : chl 96
1142 : edgomez 851 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1143 :     else pmv[4].x = pmv[4].y = 0;
1144 : Isibaar 3
1145 : edgomez 851 // [1] median prediction
1146 :     pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1147 : edgomez 195
1148 : 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
1149 : edgomez 195
1150 : edgomez 851 pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1151 :     pmv[2].y = EVEN(prevMB->mvs[0].y);
1152 :    
1153 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1154 :     pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1155 :     pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1156 :     } else pmv[6].x = pmv[6].y = 0;
1157 :    
1158 :     if (rrv) {
1159 :     int i;
1160 :     for (i = 0; i < 7; i++) {
1161 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1162 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1163 :     }
1164 :     }
1165 : Isibaar 3 }
1166 :    
1167 : edgomez 851 static int
1168 :     ModeDecision(const uint32_t iQuant, SearchData * const Data,
1169 :     int inter4v,
1170 :     MACROBLOCK * const pMB,
1171 :     const MACROBLOCK * const pMBs,
1172 :     const int x, const int y,
1173 :     const MBParam * const pParam,
1174 :     const uint32_t MotionFlags,
1175 :     const uint32_t GlobalFlags)
1176 :     {
1177 : Isibaar 3
1178 : edgomez 851 int mode = MODE_INTER;
1179 : chl 96
1180 : edgomez 851 if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision
1181 : chl 871 // int intra = 0;
1182 : edgomez 851 int sad;
1183 :     int InterBias = MV16_INTER_BIAS;
1184 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1185 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1186 :     mode = 0; //inter
1187 :     sad = Data->iMinSAD[0];
1188 :     } else {
1189 :     mode = MODE_INTER4V;
1190 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1191 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1192 :     Data->iMinSAD[0] = sad;
1193 :     }
1194 : chl 326
1195 : edgomez 851 /* intra decision */
1196 : Isibaar 3
1197 : edgomez 851 if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work
1198 :     if (y != 0)
1199 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1200 :     if (x != 0)
1201 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1202 : Isibaar 3
1203 : edgomez 851 if (Data->chroma) InterBias += 50; // to compensate bigger SAD
1204 :     if (Data->rrv) InterBias *= 4;
1205 : edgomez 195
1206 : edgomez 851 if (InterBias < pMB->sad16) {
1207 :     int32_t deviation;
1208 :     if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
1209 :     else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
1210 :     dev16(Data->Cur+8, Data->iEdgedWidth) +
1211 :     dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) +
1212 :     dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth);
1213 : edgomez 195
1214 : edgomez 851 if (deviation < (sad - InterBias)) return MODE_INTRA;// intra
1215 :     }
1216 :     return mode;
1217 : Isibaar 3
1218 : edgomez 851 } else {
1219 : edgomez 195
1220 : edgomez 851 int bits, intra, i;
1221 :     VECTOR backup[5], *v;
1222 :     Data->lambda16 = iQuant;
1223 :     Data->lambda8 = pParam->m_quant_type;
1224 : chl 181
1225 : edgomez 851 v = Data->qpel ? Data->currentQMV : Data->currentMV;
1226 :     for (i = 0; i < 5; i++) {
1227 :     Data->iMinSAD[i] = 256*4096;
1228 :     backup[i] = v[i];
1229 :     }
1230 : Isibaar 3
1231 : edgomez 851 bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1232 :     if (bits == 0) return MODE_INTER; // quick stop
1233 : Isibaar 3
1234 : edgomez 851 if (inter4v) {
1235 :     int inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1236 :     if (inter4v < bits) { Data->iMinSAD[0] = bits = inter4v; mode = MODE_INTER4V; }
1237 :     }
1238 : edgomez 195
1239 : Isibaar 3
1240 : edgomez 851 intra = CountMBBitsIntra(Data);
1241 : Isibaar 3
1242 : edgomez 851 if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; }
1243 :    
1244 :     return mode;
1245 : edgomez 195 }
1246 : edgomez 851 }
1247 : Isibaar 3
1248 : edgomez 851 static void
1249 :     SearchP(const IMAGE * const pRef,
1250 :     const uint8_t * const pRefH,
1251 :     const uint8_t * const pRefV,
1252 :     const uint8_t * const pRefHV,
1253 :     const IMAGE * const pCur,
1254 :     const int x,
1255 :     const int y,
1256 :     const uint32_t MotionFlags,
1257 :     const uint32_t GlobalFlags,
1258 :     const uint32_t iQuant,
1259 :     SearchData * const Data,
1260 :     const MBParam * const pParam,
1261 :     const MACROBLOCK * const pMBs,
1262 :     const MACROBLOCK * const prevMBs,
1263 :     int inter4v,
1264 :     MACROBLOCK * const pMB)
1265 :     {
1266 : Isibaar 3
1267 : edgomez 851 int i, iDirection = 255, mask, threshA;
1268 :     VECTOR pmv[7];
1269 : Isibaar 3
1270 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1271 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1272 : edgomez 195
1273 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1274 : Isibaar 3
1275 : edgomez 851 Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache
1276 :     i = Data->rrv ? 2 : 1;
1277 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1278 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1279 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1280 : chl 326
1281 : edgomez 851 Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1282 :     Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1283 :     Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1284 :     Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1285 :     Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1286 :     Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1287 : edgomez 195
1288 : edgomez 851 Data->lambda16 = lambda_vec16[iQuant];
1289 :     Data->lambda8 = lambda_vec8[iQuant];
1290 :     Data->qpel_precision = 0;
1291 : edgomez 195
1292 : edgomez 851 if (pMB->dquant != NO_CHANGE) inter4v = 0;
1293 : edgomez 195
1294 : edgomez 851 for(i = 0; i < 5; i++)
1295 :     Data->currentMV[i].x = Data->currentMV[i].y = 0;
1296 : chl 169
1297 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1298 :     else Data->predMV = pmv[0];
1299 : Isibaar 3
1300 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1301 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1302 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1303 :     Data->iMinSAD[2] = pMB->sad8[1];
1304 :     Data->iMinSAD[3] = pMB->sad8[2];
1305 :     Data->iMinSAD[4] = pMB->sad8[3];
1306 : chl 169
1307 : edgomez 851 if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) {
1308 :     threshA = Data->temp[0]; // that's where we keep this SAD atm
1309 :     if (threshA < 512) threshA = 512;
1310 :     else if (threshA > 1024) threshA = 1024;
1311 :     } else
1312 :     threshA = 512;
1313 : chl 169
1314 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1315 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1316 : chl 169
1317 : edgomez 851 if (!Data->rrv) {
1318 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1319 :     else CheckCandidate = CheckCandidate16no4v; //for extra speed
1320 :     } else CheckCandidate = CheckCandidate32;
1321 : chl 169
1322 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1323 : chl 169
1324 : edgomez 851 for (i = 1; i < 7; i++) {
1325 :     if (!(mask = make_mask(pmv, i)) ) continue;
1326 :     CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1327 :     if (Data->iMinSAD[0] <= threshA) break;
1328 :     }
1329 : chl 169
1330 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1331 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1332 :     (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1333 :     if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; }
1334 :     else {
1335 : Isibaar 3
1336 : edgomez 851 MainSearchFunc * MainSearchPtr;
1337 :     if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1338 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1339 :     else MainSearchPtr = DiamondSearch;
1340 : Isibaar 3
1341 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1342 : Isibaar 3
1343 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1344 :     note that this search is/might be done in halfpel positions,
1345 :     which makes it more different than the diamond above */
1346 : chl 169
1347 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH16) {
1348 :     int32_t bSAD;
1349 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1350 :     if (Data->rrv) {
1351 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1352 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1353 :     }
1354 :     if (!(MVequal(startMV, backupMV))) {
1355 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1356 : edgomez 195
1357 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1358 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1359 :     if (bSAD < Data->iMinSAD[0]) {
1360 :     Data->currentMV[0] = backupMV;
1361 :     Data->iMinSAD[0] = bSAD; }
1362 :     }
1363 : chl 169
1364 : edgomez 851 backupMV = Data->currentMV[0];
1365 :     startMV.x = startMV.y = 1;
1366 :     if (!(MVequal(startMV, backupMV))) {
1367 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1368 : chl 169
1369 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1370 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1371 :     if (bSAD < Data->iMinSAD[0]) {
1372 :     Data->currentMV[0] = backupMV;
1373 :     Data->iMinSAD[0] = bSAD; }
1374 : edgomez 195 }
1375 : edgomez 851 }
1376 :     }
1377 : edgomez 195
1378 : edgomez 851 if (MotionFlags & PMV_HALFPELREFINE16)
1379 :     if ((!(MotionFlags & HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant)
1380 :     SubpelRefine(Data);
1381 : edgomez 195
1382 : edgomez 851 for(i = 0; i < 5; i++) {
1383 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1384 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1385 :     }
1386 : Isibaar 3
1387 : edgomez 851 if (MotionFlags & PMV_QUARTERPELREFINE16)
1388 :     if ((!(MotionFlags & QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) {
1389 :     Data->qpel_precision = 1;
1390 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1391 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1392 : edgomez 195
1393 : edgomez 851 SubpelRefine(Data);
1394 :     }
1395 : Isibaar 3
1396 : edgomez 851 if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0;
1397 : Isibaar 3
1398 : edgomez 851 if (inter4v && (!(GlobalFlags & XVID_MODEDECISION_BITS) ||
1399 :     (!(MotionFlags & QUARTERPELREFINE8_BITS)) || (!(MotionFlags & HALFPELREFINE8_BITS)) ||
1400 :     ((!(MotionFlags & EXTSEARCH_BITS)) && (!(MotionFlags&PMV_EXTSEARCH8)) ))) {
1401 :     // if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop
1402 : Isibaar 3
1403 : edgomez 851 SearchData Data8;
1404 :     memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data
1405 : Isibaar 3
1406 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1407 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1408 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1409 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1410 : chl 181
1411 : edgomez 851 if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) {
1412 :     // chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it
1413 :     int sumx = 0, sumy = 0;
1414 :     const int div = 1 + Data->qpel;
1415 :     const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs;
1416 : Isibaar 3
1417 : edgomez 851 for (i = 0; i < 4; i++) {
1418 :     sumx += mv[i].x / div;
1419 :     sumy += mv[i].y / div;
1420 :     }
1421 : chl 259
1422 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1423 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1424 :     }
1425 :     }
1426 : edgomez 195
1427 : edgomez 851 inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags);
1428 :    
1429 :     if (Data->rrv) {
1430 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1431 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1432 : Isibaar 3 }
1433 :    
1434 : edgomez 851 if (inter4v == MODE_INTER) {
1435 :     pMB->mode = MODE_INTER;
1436 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1437 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1438 : Isibaar 3
1439 : edgomez 851 if(Data->qpel) {
1440 :     pMB->qmvs[0] = pMB->qmvs[1]
1441 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1442 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1443 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1444 :     } else {
1445 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1446 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1447 : Isibaar 3 }
1448 :    
1449 : edgomez 851 } else if (inter4v == MODE_INTER4V) {
1450 :     pMB->mode = MODE_INTER4V;
1451 :     pMB->sad16 = Data->iMinSAD[0];
1452 :     } else { // INTRA mode
1453 :     SkipMacroblockP(pMB, 0); // not skip, but similar enough
1454 :     pMB->mode = MODE_INTRA;
1455 :     }
1456 : edgomez 195
1457 : edgomez 851 }
1458 :    
1459 :     static void
1460 :     Search8(const SearchData * const OldData,
1461 :     const int x, const int y,
1462 :     const uint32_t MotionFlags,
1463 :     const MBParam * const pParam,
1464 :     MACROBLOCK * const pMB,
1465 :     const MACROBLOCK * const pMBs,
1466 :     const int block,
1467 :     SearchData * const Data)
1468 :     {
1469 :     int i = 0;
1470 :     Data->iMinSAD = OldData->iMinSAD + 1 + block;
1471 :     Data->currentMV = OldData->currentMV + 1 + block;
1472 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1473 :    
1474 :     if(Data->qpel) {
1475 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1476 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1477 :     Data->predMV, Data->iFcode, 0, 0);
1478 :     } else {
1479 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1480 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1481 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1482 : Isibaar 3 }
1483 :    
1484 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1485 : Isibaar 3
1486 : edgomez 851 if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) {
1487 :     if (Data->rrv) i = 2; else i = 1;
1488 : Isibaar 3
1489 : edgomez 851 Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1490 :     Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1491 :     Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1492 :     Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1493 : Isibaar 3
1494 : edgomez 851 Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1495 :     Data->qpel_precision = 0;
1496 : Isibaar 3
1497 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1498 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1499 : Isibaar 3
1500 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1501 :     else CheckCandidate = CheckCandidate16no4v;
1502 : Isibaar 3
1503 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) {
1504 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1505 : Isibaar 3
1506 : edgomez 851 MainSearchFunc *MainSearchPtr;
1507 :     if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1508 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1509 :     else MainSearchPtr = DiamondSearch;
1510 : Isibaar 3
1511 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1512 : Isibaar 3
1513 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1514 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1515 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1516 :     }
1517 :     }
1518 : edgomez 195
1519 : edgomez 851 if (MotionFlags & PMV_HALFPELREFINE8) {
1520 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1521 : edgomez 195
1522 : edgomez 851 SubpelRefine(Data); // perform halfpel refine of current best vector
1523 : Isibaar 3
1524 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1525 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1526 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1527 :     }
1528 :     }
1529 : Isibaar 3
1530 : edgomez 851 if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) {
1531 :     Data->qpel_precision = 1;
1532 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1533 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1534 :     SubpelRefine(Data);
1535 :     }
1536 :     }
1537 : edgomez 195
1538 : edgomez 851 if (Data->rrv) {
1539 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1540 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1541 :     }
1542 :    
1543 :     if(Data->qpel) {
1544 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1545 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1546 :     pMB->qmvs[block] = *Data->currentQMV;
1547 : edgomez 195 } else {
1548 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1549 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1550 : edgomez 78 }
1551 : edgomez 851
1552 :     pMB->mvs[block] = *Data->currentMV;
1553 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1554 : Isibaar 3 }
1555 :    
1556 : edgomez 851 /* motion estimation for B-frames */
1557 : chl 345
1558 : edgomez 851 static __inline VECTOR
1559 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1560 :     {
1561 :     /* the stupidiest function ever */
1562 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1563 :     }
1564 : chl 345
1565 : edgomez 851 static void __inline
1566 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1567 :     const uint32_t iWcount,
1568 :     const MACROBLOCK * const pMB,
1569 :     const uint32_t mode_curr)
1570 : chl 345 {
1571 :    
1572 : edgomez 851 // [0] is prediction
1573 :     pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1574 : chl 345
1575 : edgomez 851 pmv[1].x = pmv[1].y = 0; // [1] is zero
1576 : chl 345
1577 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1578 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1579 : chl 345
1580 : edgomez 851 if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1581 :     pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1582 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1583 :     } else pmv[3].x = pmv[3].y = 0;
1584 : chl 345
1585 : edgomez 851 if (y != 0) {
1586 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1587 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1588 :     } else pmv[4].x = pmv[4].y = 0;
1589 : chl 345
1590 : edgomez 851 if (x != 0) {
1591 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1592 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1593 :     } else pmv[5].x = pmv[5].y = 0;
1594 : chl 345
1595 : edgomez 851 if (x != 0 && y != 0) {
1596 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1597 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1598 :     } else pmv[6].x = pmv[6].y = 0;
1599 :     }
1600 : chl 345
1601 :    
1602 : edgomez 851 /* search backward or forward */
1603 :     static void
1604 :     SearchBF( const IMAGE * const pRef,
1605 :     const uint8_t * const pRefH,
1606 :     const uint8_t * const pRefV,
1607 :     const uint8_t * const pRefHV,
1608 :     const IMAGE * const pCur,
1609 :     const int x, const int y,
1610 :     const uint32_t MotionFlags,
1611 :     const uint32_t iFcode,
1612 :     const MBParam * const pParam,
1613 :     MACROBLOCK * const pMB,
1614 :     const VECTOR * const predMV,
1615 :     int32_t * const best_sad,
1616 :     const int32_t mode_current,
1617 :     SearchData * const Data)
1618 :     {
1619 : chl 345
1620 : edgomez 851 int i, iDirection = 255, mask;
1621 :     VECTOR pmv[7];
1622 :     MainSearchFunc *MainSearchPtr;
1623 :     *Data->iMinSAD = MV_MAX_ERROR;
1624 :     Data->iFcode = iFcode;
1625 :     Data->qpel_precision = 0;
1626 :     Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache
1627 : chl 345
1628 : edgomez 851 Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16;
1629 :     Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16;
1630 :     Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16;
1631 :     Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16;
1632 :     Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8;
1633 :     Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8;
1634 : chl 345
1635 : edgomez 851 Data->predMV = *predMV;
1636 : chl 345
1637 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1638 :     pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1639 : chl 345
1640 : edgomez 851 pmv[0] = Data->predMV;
1641 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1642 : chl 345
1643 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1644 : chl 345
1645 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1646 :     CheckCandidate = CheckCandidate16no4v;
1647 : chl 345
1648 : edgomez 851 // main loop. checking all predictions
1649 :     for (i = 0; i < 7; i++) {
1650 :     if (!(mask = make_mask(pmv, i)) ) continue;
1651 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1652 : chl 345 }
1653 :    
1654 : edgomez 851 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1655 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1656 :     else MainSearchPtr = DiamondSearch;
1657 : chl 345
1658 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1659 : chl 345
1660 : edgomez 851 SubpelRefine(Data);
1661 : chl 345
1662 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1663 :     Data->currentQMV->x = 2*Data->currentMV->x;
1664 :     Data->currentQMV->y = 2*Data->currentMV->y;
1665 :     Data->qpel_precision = 1;
1666 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1667 :     pParam->width, pParam->height, iFcode, 1, 0);
1668 :     SubpelRefine(Data);
1669 :     }
1670 : chl 345
1671 : edgomez 851 // three bits are needed to code backward mode. four for forward
1672 : Isibaar 3
1673 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1674 :     else *Data->iMinSAD += 3 * Data->lambda16;
1675 : edgomez 195
1676 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1677 :     *best_sad = *Data->iMinSAD;
1678 :     pMB->mode = mode_current;
1679 :     if (Data->qpel) {
1680 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1681 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1682 :     if (mode_current == MODE_FORWARD)
1683 :     pMB->qmvs[0] = *Data->currentQMV;
1684 :     else
1685 :     pMB->b_qmvs[0] = *Data->currentQMV;
1686 :     } else {
1687 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1688 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1689 :     }
1690 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1691 :     else pMB->b_mvs[0] = *Data->currentMV;
1692 :     }
1693 : edgomez 195
1694 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1695 :     else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
1696 : Isibaar 3 }
1697 :    
1698 : edgomez 851 static void
1699 :     SkipDecisionB(const IMAGE * const pCur,
1700 :     const IMAGE * const f_Ref,
1701 :     const IMAGE * const b_Ref,
1702 :     MACROBLOCK * const pMB,
1703 :     const uint32_t x, const uint32_t y,
1704 :     const SearchData * const Data)
1705 :     {
1706 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1707 :     int32_t sum;
1708 :     const int div = 1 + Data->qpel;
1709 :     int k;
1710 :     const uint32_t stride = Data->iEdgedWidth/2;
1711 :     //this is not full chroma compensation, only it's fullpel approximation. should work though
1712 : Isibaar 3
1713 : edgomez 851 for (k = 0; k < 4; k++) {
1714 :     dy += Data->directmvF[k].y / div;
1715 :     dx += Data->directmvF[0].x / div;
1716 :     b_dy += Data->directmvB[0].y / div;
1717 :     b_dx += Data->directmvB[0].x / div;
1718 :     }
1719 : Isibaar 3
1720 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1721 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1722 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1723 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1724 : Isibaar 3
1725 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1726 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1727 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1728 :     stride);
1729 : Isibaar 3
1730 : edgomez 851 if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip
1731 : Isibaar 3
1732 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1733 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1734 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1735 :     stride);
1736 : edgomez 195
1737 : edgomez 851 if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1738 :     }
1739 : Isibaar 3
1740 : edgomez 851 static __inline uint32_t
1741 :     SearchDirect(const IMAGE * const f_Ref,
1742 :     const uint8_t * const f_RefH,
1743 :     const uint8_t * const f_RefV,
1744 :     const uint8_t * const f_RefHV,
1745 :     const IMAGE * const b_Ref,
1746 :     const uint8_t * const b_RefH,
1747 :     const uint8_t * const b_RefV,
1748 :     const uint8_t * const b_RefHV,
1749 :     const IMAGE * const pCur,
1750 :     const int x, const int y,
1751 :     const uint32_t MotionFlags,
1752 :     const int32_t TRB, const int32_t TRD,
1753 :     const MBParam * const pParam,
1754 :     MACROBLOCK * const pMB,
1755 :     const MACROBLOCK * const b_mb,
1756 :     int32_t * const best_sad,
1757 :     SearchData * const Data)
1758 : Isibaar 3
1759 : edgomez 851 {
1760 :     int32_t skip_sad;
1761 :     int k = (x + Data->iEdgedWidth*y) * 16;
1762 :     MainSearchFunc *MainSearchPtr;
1763 : Isibaar 3
1764 : edgomez 851 *Data->iMinSAD = 256*4096;
1765 :     Data->Ref = f_Ref->y + k;
1766 :     Data->RefH = f_RefH + k;
1767 :     Data->RefV = f_RefV + k;
1768 :     Data->RefHV = f_RefHV + k;
1769 :     Data->bRef = b_Ref->y + k;
1770 :     Data->bRefH = b_RefH + k;
1771 :     Data->bRefV = b_RefV + k;
1772 :     Data->bRefHV = b_RefHV + k;
1773 :     Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1774 :     Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1775 :     Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1776 :     Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1777 : edgomez 195
1778 : edgomez 851 k = Data->qpel ? 4 : 2;
1779 :     Data->max_dx = k * (pParam->width - x * 16);
1780 :     Data->max_dy = k * (pParam->height - y * 16);
1781 :     Data->min_dx = -k * (16 + x * 16);
1782 :     Data->min_dy = -k * (16 + y * 16);
1783 : chl 181
1784 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1785 :     Data->qpel_precision = 0;
1786 : edgomez 170
1787 : edgomez 851 for (k = 0; k < 4; k++) {
1788 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1789 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1790 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1791 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1792 : Isibaar 3
1793 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1794 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1795 :    
1796 :     *best_sad = 256*4096; // in that case, we won't use direct mode
1797 :     pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1798 :     pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1799 :     return 256*4096;
1800 :     }
1801 :     if (b_mb->mode != MODE_INTER4V) {
1802 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1803 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1804 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1805 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1806 :     break;
1807 :     }
1808 : edgomez 195 }
1809 : Isibaar 3
1810 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1811 : Isibaar 3
1812 : edgomez 851 CheckCandidate(0, 0, 255, &k, Data);
1813 : edgomez 195
1814 : edgomez 851 // initial (fast) skip decision
1815 :     if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) {
1816 :     //possible skip
1817 :     if (Data->chroma) {
1818 :     pMB->mode = MODE_DIRECT_NONE_MV;
1819 :     return *Data->iMinSAD; // skip.
1820 :     } else {
1821 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1822 :     if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1823 :     }
1824 : Isibaar 3 }
1825 :    
1826 : edgomez 851 skip_sad = *Data->iMinSAD;
1827 : edgomez 195
1828 : edgomez 851 // DIRECT MODE DELTA VECTOR SEARCH.
1829 :     // This has to be made more effective, but at the moment I'm happy it's running at all
1830 : Isibaar 3
1831 : edgomez 851 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1832 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1833 :     else MainSearchPtr = DiamondSearch;
1834 : Isibaar 3
1835 : edgomez 851 MainSearchPtr(0, 0, Data, 255);
1836 : Isibaar 3
1837 : edgomez 851 SubpelRefine(Data);
1838 : chl 181
1839 : edgomez 851 *best_sad = *Data->iMinSAD;
1840 : chl 181
1841 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1842 :     else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1843 : chl 181
1844 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1845 : edgomez 195
1846 : edgomez 851 for (k = 0; k < 4; k++) {
1847 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1848 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1849 :     ? Data->directmvB[k].x
1850 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1851 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1852 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1853 :     ? Data->directmvB[k].y
1854 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1855 :     if (Data->qpel) {
1856 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1857 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1858 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1859 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1860 :     }
1861 : edgomez 195
1862 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1863 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1864 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1865 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1866 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1867 :     break;
1868 :     }
1869 : edgomez 78 }
1870 : edgomez 851 return skip_sad;
1871 :     }
1872 : Isibaar 3
1873 : edgomez 851 static void
1874 :     SearchInterpolate(const IMAGE * const f_Ref,
1875 :     const uint8_t * const f_RefH,
1876 :     const uint8_t * const f_RefV,
1877 :     const uint8_t * const f_RefHV,
1878 :     const IMAGE * const b_Ref,
1879 :     const uint8_t * const b_RefH,
1880 :     const uint8_t * const b_RefV,
1881 :     const uint8_t * const b_RefHV,
1882 :     const IMAGE * const pCur,
1883 :     const int x, const int y,
1884 :     const uint32_t fcode,
1885 :     const uint32_t bcode,
1886 :     const uint32_t MotionFlags,
1887 :     const MBParam * const pParam,
1888 :     const VECTOR * const f_predMV,
1889 :     const VECTOR * const b_predMV,
1890 :     MACROBLOCK * const pMB,
1891 :     int32_t * const best_sad,
1892 :     SearchData * const fData)
1893 : chl 96
1894 : edgomez 851 {
1895 : chl 169
1896 : edgomez 851 int iDirection, i, j;
1897 :     SearchData bData;
1898 : chl 169
1899 : edgomez 851 fData->qpel_precision = 0;
1900 :     memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1901 :     *fData->iMinSAD = 4096*256;
1902 :     bData.currentMV++; bData.currentQMV++;
1903 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1904 : chl 169
1905 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1906 :     bData.bRef = fData->Ref = f_Ref->y + i;
1907 :     bData.bRefH = fData->RefH = f_RefH + i;
1908 :     bData.bRefV = fData->RefV = f_RefV + i;
1909 :     bData.bRefHV = fData->RefHV = f_RefHV + i;
1910 :     bData.Ref = fData->bRef = b_Ref->y + i;
1911 :     bData.RefH = fData->bRefH = b_RefH + i;
1912 :     bData.RefV = fData->bRefV = b_RefV + i;
1913 :     bData.RefHV = fData->bRefHV = b_RefHV + i;
1914 :     bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1915 :     bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1916 :     bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1917 :     bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1918 : chl 169
1919 :    
1920 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1921 :     fData->bpredMV = bData.predMV = *b_predMV;
1922 :     fData->currentMV[0] = fData->currentMV[2];
1923 : Isibaar 3
1924 : edgomez 851 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0);
1925 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0);
1926 : Isibaar 3
1927 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1928 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1929 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1930 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1931 : chl 169
1932 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1933 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1934 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1935 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1936 : Isibaar 3
1937 : edgomez 851 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1938 : chl 169
1939 : edgomez 851 //diamond
1940 :     do {
1941 :     iDirection = 255;
1942 :     // forward MV moves
1943 :     i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1944 : chl 169
1945 : edgomez 851 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1946 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1947 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1948 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1949 : edgomez 195
1950 : edgomez 851 // backward MV moves
1951 :     i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1952 :     fData->currentMV[2] = fData->currentMV[0];
1953 :     CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1954 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1955 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1956 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1957 : Isibaar 3
1958 : edgomez 851 } while (!(iDirection));
1959 : chl 140
1960 : edgomez 851 //qpel refinement
1961 :     if (fData->qpel) {
1962 :     if (*fData->iMinSAD > *best_sad + 500) return;
1963 :     CheckCandidate = CheckCandidateInt;
1964 :     fData->qpel_precision = bData.qpel_precision = 1;
1965 :     get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1966 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1967 :     fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1968 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1969 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1970 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1971 :     SubpelRefine(fData);
1972 :     if (*fData->iMinSAD > *best_sad + 300) return;
1973 :     fData->currentQMV[2] = fData->currentQMV[0];
1974 :     SubpelRefine(&bData);
1975 :     }
1976 : chl 140
1977 : edgomez 851 *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode.
1978 : Isibaar 3
1979 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
1980 :     *best_sad = *fData->iMinSAD;
1981 :     pMB->mvs[0] = fData->currentMV[0];
1982 :     pMB->b_mvs[0] = fData->currentMV[1];
1983 :     pMB->mode = MODE_INTERPOLATE;
1984 :     if (fData->qpel) {
1985 :     pMB->qmvs[0] = fData->currentQMV[0];
1986 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1987 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1988 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1989 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1990 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1991 :     } else {
1992 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1993 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1994 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1995 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1996 :     }
1997 : edgomez 78 }
1998 : edgomez 851 }
1999 : Isibaar 3
2000 : edgomez 851 void
2001 :     MotionEstimationBVOP(MBParam * const pParam,
2002 :     FRAMEINFO * const frame,
2003 :     const int32_t time_bp,
2004 :     const int32_t time_pp,
2005 :     // forward (past) reference
2006 :     const MACROBLOCK * const f_mbs,
2007 :     const IMAGE * const f_ref,
2008 :     const IMAGE * const f_refH,
2009 :     const IMAGE * const f_refV,
2010 :     const IMAGE * const f_refHV,
2011 :     // backward (future) reference
2012 :     const FRAMEINFO * const b_reference,
2013 :     const IMAGE * const b_ref,
2014 :     const IMAGE * const b_refH,
2015 :     const IMAGE * const b_refV,
2016 :     const IMAGE * const b_refHV)
2017 :     {
2018 :     uint32_t i, j;
2019 :     int32_t best_sad;
2020 :     uint32_t skip_sad;
2021 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2022 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
2023 : Isibaar 3
2024 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2025 : Isibaar 3
2026 : edgomez 851 const int32_t TRB = time_pp - time_bp;
2027 :     const int32_t TRD = time_pp;
2028 : edgomez 195
2029 : edgomez 851 // some pre-inintialized data for the rest of the search
2030 : Isibaar 3
2031 : edgomez 851 SearchData Data;
2032 :     int32_t iMinSAD;
2033 :     VECTOR currentMV[3];
2034 :     VECTOR currentQMV[3];
2035 :     int32_t temp[8];
2036 :     memset(&Data, 0, sizeof(SearchData));
2037 :     Data.iEdgedWidth = pParam->edged_width;
2038 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2039 :     Data.iMinSAD = &iMinSAD;
2040 :     Data.lambda16 = lambda_vec16[frame->quant];
2041 :     Data.qpel = pParam->m_quarterpel;
2042 :     Data.rounding = 0;
2043 :     Data.chroma = frame->motion_flags & PMV_CHROMA8;
2044 :     Data.temp = temp;
2045 : Isibaar 3
2046 : edgomez 851 Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose)
2047 :     // note: i==horizontal, j==vertical
2048 :     for (j = 0; j < pParam->mb_height; j++) {
2049 : edgomez 195
2050 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2051 : Isibaar 3
2052 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2053 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2054 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2055 : edgomez 195
2056 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2057 :     if (b_reference->coding_type != S_VOP)
2058 :     if (b_mb->mode == MODE_NOT_CODED) {
2059 :     pMB->mode = MODE_NOT_CODED;
2060 :     continue;
2061 :     }
2062 : Isibaar 3
2063 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2064 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2065 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2066 :     pMB->quant = frame->quant;
2067 : Isibaar 3
2068 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2069 :     and (2) sets very good predictions for forward and backward search */
2070 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2071 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2072 :     &frame->image,
2073 :     i, j,
2074 :     frame->motion_flags,
2075 :     TRB, TRD,
2076 :     pParam,
2077 :     pMB, b_mb,
2078 :     &best_sad,
2079 :     &Data);
2080 : Isibaar 3
2081 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2082 : edgomez 195
2083 : edgomez 851 // forward search
2084 :     SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2085 :     &frame->image, i, j,
2086 :     frame->motion_flags,
2087 :     frame->fcode, pParam,
2088 :     pMB, &f_predMV, &best_sad,
2089 :     MODE_FORWARD, &Data);
2090 : edgomez 195
2091 : edgomez 851 // backward search
2092 :     SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2093 :     &frame->image, i, j,
2094 :     frame->motion_flags,
2095 :     frame->bcode, pParam,
2096 :     pMB, &b_predMV, &best_sad,
2097 :     MODE_BACKWARD, &Data);
2098 :    
2099 :     // interpolate search comes last, because it uses data from forward and backward as prediction
2100 :     SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2101 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2102 :     &frame->image,
2103 :     i, j,
2104 :     frame->fcode, frame->bcode,
2105 :     frame->motion_flags,
2106 :     pParam,
2107 :     &f_predMV, &b_predMV,
2108 :     pMB, &best_sad,
2109 :     &Data);
2110 :    
2111 :     // final skip decision
2112 :     if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2113 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2114 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2115 :    
2116 :     switch (pMB->mode) {
2117 :     case MODE_FORWARD:
2118 :     f_count++;
2119 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2120 :     break;
2121 :     case MODE_BACKWARD:
2122 :     b_count++;
2123 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2124 :     break;
2125 :     case MODE_INTERPOLATE:
2126 :     i_count++;
2127 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2128 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2129 :     break;
2130 :     case MODE_DIRECT:
2131 :     case MODE_DIRECT_NO4V:
2132 :     d_count++;
2133 :     default:
2134 :     break;
2135 :     }
2136 :     }
2137 :     }
2138 : Isibaar 3 }
2139 : chl 96
2140 : edgomez 851 static __inline void
2141 :     MEanalyzeMB ( const uint8_t * const pRef,
2142 :     const uint8_t * const pCur,
2143 :     const int x,
2144 :     const int y,
2145 :     const MBParam * const pParam,
2146 :     MACROBLOCK * const pMBs,
2147 :     SearchData * const Data)
2148 : chl 96 {
2149 :    
2150 : edgomez 851 int i, mask;
2151 :     VECTOR pmv[3];
2152 :     MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width];
2153 : chl 96
2154 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2155 : chl 96
2156 : edgomez 851 //median is only used as prediction. it doesn't have to be real
2157 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2158 :     else
2159 :     if (x == 1) //left macroblock does not have any vector now
2160 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2161 :     else if (y == 1) // top macroblock doesn't have it's vector
2162 :     Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
2163 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2164 : edgomez 195
2165 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2166 :     pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv);
2167 : edgomez 195
2168 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2169 :     Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
2170 : chl 96
2171 : edgomez 851 pmv[1].x = EVEN(pMB->mvs[0].x);
2172 :     pmv[1].y = EVEN(pMB->mvs[0].y);
2173 :     pmv[2].x = EVEN(Data->predMV.x);
2174 :     pmv[2].y = EVEN(Data->predMV.y);
2175 :     pmv[0].x = pmv[0].y = 0;
2176 : edgomez 195
2177 : edgomez 851 CheckCandidate32I(0, 0, 255, &i, Data);
2178 : edgomez 195
2179 : edgomez 851 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) {
2180 : chl 96
2181 : edgomez 851 if (!(mask = make_mask(pmv, 1)))
2182 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2183 :     if (!(mask = make_mask(pmv, 2)))
2184 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2185 : chl 96
2186 : edgomez 851 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed
2187 :     DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2188 :    
2189 :     for (i = 0; i < 4; i++) {
2190 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2191 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2192 :     MB->mode = MODE_INTER;
2193 :     MB->sad16 = Data->iMinSAD[i+1];
2194 :     }
2195 : chl 96 }
2196 : edgomez 851 }
2197 : chl 96
2198 : edgomez 851 #define INTRA_BIAS 2500
2199 :     #define INTRA_THRESH 1500
2200 :     #define INTER_THRESH 1400
2201 : chl 96
2202 : edgomez 851 int
2203 :     MEanalysis( const IMAGE * const pRef,
2204 :     FRAMEINFO * const Current,
2205 :     MBParam * const pParam,
2206 :     int maxIntra, //maximum number if non-I frames
2207 :     int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2208 :     int bCount) // number of B frames in a row
2209 :     {
2210 :     uint32_t x, y, intra = 0;
2211 :     int sSAD = 0;
2212 :     MACROBLOCK * const pMBs = Current->mbs;
2213 :     const IMAGE * const pCurrent = &Current->image;
2214 :     int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH;
2215 : chl 96
2216 : edgomez 851 int32_t iMinSAD[5], temp[5];
2217 :     VECTOR currentMV[5];
2218 :     SearchData Data;
2219 :     Data.iEdgedWidth = pParam->edged_width;
2220 :     Data.currentMV = currentMV;
2221 :     Data.iMinSAD = iMinSAD;
2222 :     Data.iFcode = Current->fcode;
2223 :     Data.rrv = Current->global_flags & XVID_REDUCED;
2224 :     Data.temp = temp;
2225 :     CheckCandidate = CheckCandidate32I;
2226 : chl 96
2227 : edgomez 851 if (intraCount != 0 && intraCount < 10) // we're right after an I frame
2228 :     IntraThresh += 4 * (intraCount - 10) * (intraCount - 10);
2229 :     else
2230 :     if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2231 :     IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
2232 : chl 96
2233 : edgomez 851 InterThresh += 400 * (1 - bCount);
2234 :     if (InterThresh < 300) InterThresh = 300;
2235 : chl 326
2236 : edgomez 851 if (sadInit) (*sadInit) ();
2237 : chl 96
2238 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2239 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2240 :     int i;
2241 : edgomez 195
2242 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2243 : edgomez 195
2244 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2245 : edgomez 195
2246 : edgomez 851 for (i = 0; i < 4; i++) {
2247 :     int dev;
2248 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2249 :     if (pMB->sad16 > IntraThresh) {
2250 :     dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2251 :     pParam->edged_width);
2252 :     if (dev + IntraThresh < pMB->sad16) {
2253 :     pMB->mode = MODE_INTRA;
2254 :     if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP;
2255 :     }
2256 :     }
2257 :     sSAD += pMB->sad16;
2258 :     }
2259 :     }
2260 : edgomez 195 }
2261 : edgomez 851 sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2);
2262 :     // if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP;
2263 :     if (sSAD > InterThresh ) return P_VOP;
2264 :     emms();
2265 :     return B_VOP;
2266 : chl 96
2267 : edgomez 851 }
2268 : chl 96
2269 :    
2270 : edgomez 851 static WARPPOINTS
2271 :     GlobalMotionEst(const MACROBLOCK * const pMBs,
2272 :     const MBParam * const pParam,
2273 :     const FRAMEINFO * const current,
2274 :     const FRAMEINFO * const reference,
2275 :     const IMAGE * const pRefH,
2276 :     const IMAGE * const pRefV,
2277 :     const IMAGE * const pRefHV )
2278 :     {
2279 : chl 96
2280 : edgomez 851 const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2281 :     const int deltay=8;
2282 :     const int grad=512; // lower bound for deviation in MB
2283 : chl 96
2284 : edgomez 851 WARPPOINTS gmc;
2285 : chl 96
2286 : edgomez 851 uint32_t mx, my;
2287 : chl 96
2288 : edgomez 851 int MBh = pParam->mb_height;
2289 :     int MBw = pParam->mb_width;
2290 : chl 96
2291 : edgomez 851 int *MBmask= calloc(MBh*MBw,sizeof(int));
2292 :     double DtimesF[4] = { 0.,0., 0., 0. };
2293 :     double sol[4] = { 0., 0., 0., 0. };
2294 :     double a,b,c,n,denom;
2295 :     double meanx,meany;
2296 :     int num,oldnum;
2297 : edgomez 195
2298 : chl 871 if (!MBmask) { fprintf(stderr,"Mem error\n");
2299 :     gmc.duv[0].x= gmc.duv[0].y =
2300 :     gmc.duv[1].x= gmc.duv[1].y =
2301 :     gmc.duv[2].x= gmc.duv[2].y = 0;
2302 :     return gmc; }
2303 : edgomez 195
2304 : edgomez 851 // filter mask of all blocks
2305 :    
2306 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2307 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2308 : edgomez 851 {
2309 :     const int mbnum = mx + my * MBw;
2310 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2311 :     const VECTOR mv = pMB->mvs[0];
2312 :    
2313 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2314 :     continue;
2315 :    
2316 :     if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2317 :     && ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2318 :     && ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2319 :     && ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2320 :     MBmask[mbnum]=1;
2321 : chl 96 }
2322 :    
2323 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2324 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2325 : edgomez 851 {
2326 :     const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx;
2327 : chl 96
2328 : edgomez 851 const int mbnum = mx + my * MBw;
2329 :     if (!MBmask[mbnum])
2330 :     continue;
2331 :    
2332 : edgomez 876 if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad )
2333 : edgomez 851 MBmask[mbnum] = 0;
2334 : edgomez 876 if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad )
2335 : edgomez 851 MBmask[mbnum] = 0;
2336 :    
2337 : edgomez 195 }
2338 : chl 96
2339 : edgomez 851 emms();
2340 : chl 96
2341 : edgomez 851 do { /* until convergence */
2342 : chl 96
2343 : edgomez 851 a = b = c = n = 0;
2344 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2345 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2346 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2347 : edgomez 851 {
2348 :     const int mbnum = mx + my * MBw;
2349 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2350 :     const VECTOR mv = pMB->mvs[0];
2351 : chl 96
2352 : edgomez 851 if (!MBmask[mbnum])
2353 :     continue;
2354 : chl 96
2355 : edgomez 851 n++;
2356 :     a += 16*mx+8;
2357 :     b += 16*my+8;
2358 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2359 : chl 96
2360 : edgomez 851 DtimesF[0] += (double)mv.x;
2361 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2362 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2363 :     DtimesF[3] += (double)mv.y;
2364 :     }
2365 : chl 96
2366 : edgomez 851 denom = a*a+b*b-c*n;
2367 : chl 96
2368 : edgomez 851 /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2369 :     /* D'*E*F has been calculated in the same loop as matrix */
2370 : chl 96
2371 : edgomez 851 sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2372 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2373 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2374 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2375 : chl 96
2376 : edgomez 851 sol[0] /= denom;
2377 :     sol[1] /= denom;
2378 :     sol[2] /= denom;
2379 :     sol[3] /= denom;
2380 : chl 96
2381 : edgomez 851 meanx = meany = 0.;
2382 :     oldnum = 0;
2383 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2384 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2385 : edgomez 851 {
2386 :     const int mbnum = mx + my * MBw;
2387 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2388 :     const VECTOR mv = pMB->mvs[0];
2389 : chl 96
2390 : edgomez 851 if (!MBmask[mbnum])
2391 :     continue;
2392 : chl 184
2393 : edgomez 851 oldnum++;
2394 :     meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x );
2395 :     meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y );
2396 :     }
2397 : chl 96
2398 : edgomez 851 if (4*meanx > oldnum) /* better fit than 0.25 is useless */
2399 :     meanx /= oldnum;
2400 :     else
2401 :     meanx = 0.25;
2402 : edgomez 195
2403 : edgomez 851 if (4*meany > oldnum)
2404 :     meany /= oldnum;
2405 :     else
2406 :     meany = 0.25;
2407 : chl 96
2408 : edgomez 851 /* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2409 :     fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2410 :     */
2411 :     num = 0;
2412 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2413 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2414 : edgomez 851 {
2415 :     const int mbnum = mx + my * MBw;
2416 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2417 :     const VECTOR mv = pMB->mvs[0];
2418 : chl 96
2419 : edgomez 851 if (!MBmask[mbnum])
2420 :     continue;
2421 : chl 96
2422 : edgomez 851 if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx )
2423 :     || ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) )
2424 :     MBmask[mbnum]=0;
2425 :     else
2426 :     num++;
2427 : chl 96 }
2428 : edgomez 195
2429 : edgomez 851 } while ( (oldnum != num) && (num>=4) );
2430 : edgomez 195
2431 : edgomez 851 if (num < 4)
2432 :     {
2433 :     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;
2434 :     } else {
2435 : edgomez 195
2436 : edgomez 851 gmc.duv[0].x=(int)(sol[0]+0.5);
2437 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2438 : chl 96
2439 : edgomez 851 gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2440 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2441 : chl 96
2442 : edgomez 851 gmc.duv[2].x=0;
2443 :     gmc.duv[2].y=0;
2444 :     }
2445 :     // 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);
2446 : chl 96
2447 : edgomez 851 free(MBmask);
2448 : chl 96
2449 : edgomez 851 return gmc;
2450 : chl 96 }
2451 :    
2452 : edgomez 851 // functions which perform BITS-based search/bitcount
2453 : chl 96
2454 : edgomez 851 static int
2455 :     CountMBBitsInter(SearchData * const Data,
2456 :     const MACROBLOCK * const pMBs, const int x, const int y,
2457 :     const MBParam * const pParam,
2458 :     const uint32_t MotionFlags)
2459 : chl 96 {
2460 : edgomez 851 int i, iDirection;
2461 :     int32_t bsad[5];
2462 : chl 141
2463 : edgomez 851 CheckCandidate = CheckCandidateBits16;
2464 : chl 96
2465 : edgomez 851 if (Data->qpel) {
2466 :     for(i = 0; i < 5; i++) {
2467 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2468 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2469 :     }
2470 :     Data->qpel_precision = 1;
2471 :     CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2472 : chl 96
2473 : edgomez 851 //checking if this vector is perfect. if it is, we stop.
2474 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0)
2475 :     return 0; //quick stop
2476 : edgomez 195
2477 : edgomez 851 if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search
2478 :     for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2479 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2480 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2481 :     Data->qpel_precision = 0;
2482 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2483 :     CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2484 :     }
2485 : edgomez 195
2486 : edgomez 851 } else { // not qpel
2487 : edgomez 195
2488 : edgomez 851 CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2489 :     //checking if this vector is perfect. if it is, we stop.
2490 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) {
2491 :     return 0; //inter
2492 :     }
2493 :     }
2494 : chl 96
2495 : edgomez 851 if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2496 : chl 96
2497 : edgomez 851 if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data);
2498 : chl 96
2499 : edgomez 851 if (Data->qpel) {
2500 :     if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search
2501 :     for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2502 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match
2503 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2504 :     }
2505 : edgomez 195
2506 : edgomez 851 // preparing for qpel-precision search
2507 :     Data->qpel_precision = 1;
2508 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2509 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
2510 :     }
2511 :     if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2512 :     }
2513 : chl 96
2514 : edgomez 851 if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction
2515 :     VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2516 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2517 :     CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2518 :     }
2519 :     return Data->iMinSAD[0];
2520 :     }
2521 : chl 96
2522 :    
2523 : edgomez 851 static int
2524 :     CountMBBitsInter4v(const SearchData * const Data,
2525 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2526 :     const int x, const int y,
2527 :     const MBParam * const pParam, const uint32_t MotionFlags,
2528 :     const VECTOR * const backup)
2529 :     {
2530 : chl 96
2531 : edgomez 851 int cbp = 0, bits = 0, t = 0, i, iDirection;
2532 :     SearchData Data2, *Data8 = &Data2;
2533 :     int sumx = 0, sumy = 0;
2534 :     int16_t in[64], coeff[64];
2535 : chl 96
2536 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2537 :     CheckCandidate = CheckCandidateBits8;
2538 : chl 96
2539 : edgomez 851 for (i = 0; i < 4; i++) {
2540 :     Data8->iMinSAD = Data->iMinSAD + i + 1;
2541 :     Data8->currentMV = Data->currentMV + i + 1;
2542 :     Data8->currentQMV = Data->currentQMV + i + 1;
2543 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2544 :     Data8->Ref = Data->Ref + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2545 :     Data8->RefH = Data->RefH + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2546 :     Data8->RefV = Data->RefV + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2547 :     Data8->RefHV = Data->RefHV + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2548 : edgomez 195
2549 : edgomez 851 if(Data->qpel) {
2550 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2551 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2552 :     Data8->predMV, Data8->iFcode, 0, 0);
2553 :     } else {
2554 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2555 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2556 :     Data8->predMV, Data8->iFcode, 0, 0);
2557 :     }
2558 : chl 96
2559 : 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,
2560 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2561 : chl 96
2562 : edgomez 851 *Data8->iMinSAD += t;
2563 : chl 96
2564 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2565 :     // checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far)
2566 :     if (Data8->qpel) {
2567 :     if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y))
2568 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2569 :     } else {
2570 :     if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y))
2571 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2572 :     }
2573 : edgomez 195
2574 : edgomez 851 if (Data8->qpel) {
2575 :     if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows
2576 :     int32_t s = *Data8->iMinSAD;
2577 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2578 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2579 :     Data8->qpel_precision = 0;
2580 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2581 :     pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2582 : edgomez 195
2583 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2584 :     CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2585 : edgomez 195
2586 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS)
2587 :     SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2588 : chl 96
2589 : edgomez 851 if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8);
2590 : chl 96
2591 : edgomez 851 if(s > *Data8->iMinSAD) { //we have found a better match
2592 :     Data8->currentQMV->x = 2*Data8->currentMV->x;
2593 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2594 :     }
2595 : chl 96
2596 : edgomez 851 Data8->qpel_precision = 1;
2597 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2598 :     pParam->width, pParam->height, Data8->iFcode, 1, 0);
2599 : chl 96
2600 : edgomez 851 }
2601 :     if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2602 : edgomez 195
2603 : edgomez 851 } else // not qpel
2604 :     if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement
2605 :    
2606 :     //checking vector equal to predicion
2607 :     if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) {
2608 :     const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2609 :     if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y))
2610 :     CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2611 : chl 141 }
2612 : edgomez 195
2613 : edgomez 851 bits += *Data8->iMinSAD;
2614 :     if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V
2615 :    
2616 :     // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else
2617 :     if(Data->qpel) {
2618 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2619 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2620 :     pMB->qmvs[i] = *Data8->currentQMV;
2621 :     sumx += Data8->currentQMV->x/2;
2622 :     sumy += Data8->currentQMV->y/2;
2623 :     } else {
2624 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2625 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2626 :     sumx += Data8->currentMV->x;
2627 :     sumy += Data8->currentMV->y;
2628 : chl 141 }
2629 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2630 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2631 :     if (Data8->temp[0]) cbp |= 1 << (5 - i);
2632 : chl 141 }
2633 :    
2634 : edgomez 851 if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma
2635 :     const uint8_t * ptr;
2636 :     sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2637 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2638 : chl 141
2639 : edgomez 851 //chroma U
2640 :     ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCU, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2641 :     transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2642 :     fdct(in);
2643 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2644 :     else i = quant4_inter(coeff, in, Data->lambda16);
2645 :     if (i > 0) {
2646 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2647 :     cbp |= 1 << (5 - 4);
2648 :     }
2649 : chl 141
2650 : edgomez 851 if (bits < *Data->iMinSAD) { // still possible
2651 :     //chroma V
2652 :     ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCV, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2653 :     transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2654 :     fdct(in);
2655 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2656 :     else i = quant4_inter(coeff, in, Data->lambda16);
2657 :     if (i > 0) {
2658 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2659 :     cbp |= 1 << (5 - 5);
2660 :     }
2661 : chl 897 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
2662 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2663 :     }
2664 : edgomez 195 }
2665 : chl 96
2666 : edgomez 851 return bits;
2667 :     }
2668 : chl 96
2669 :    
2670 : edgomez 851 static int
2671 :     CountMBBitsIntra(const SearchData * const Data)
2672 :     {
2673 :     int bits = 1; //this one is ac/dc prediction flag. always 1.
2674 :     int cbp = 0, i, t, dc = 0, b_dc = 1024;
2675 :     const uint32_t iQuant = Data->lambda16;
2676 :     int16_t in[64], coeff[64];
2677 : edgomez 195
2678 : edgomez 851 for(i = 0; i < 4; i++) {
2679 :     uint32_t iDcScaler = get_dc_scaler(iQuant, 1);
2680 : chl 96
2681 : edgomez 851 int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2682 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2683 :     fdct(in);
2684 :     b_dc = dc;
2685 :     dc = in[0];
2686 :     in[0] -= b_dc;
2687 :     if (Data->lambda8 == 0) quant_intra_c(coeff, in, iQuant, iDcScaler);
2688 :     else quant4_intra_c(coeff, in, iQuant, iDcScaler);
2689 : chl 141
2690 : edgomez 851 b_dc = dc;
2691 :     dc = coeff[0];
2692 :     if (i != 0) coeff[0] -= b_dc;
2693 : chl 181
2694 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;;
2695 :     Data->temp[i] = t;
2696 :     if (t != 0) cbp |= 1 << (5 - i);
2697 :     if (bits >= Data->iMinSAD[0]) break;
2698 :     }
2699 : chl 181
2700 : edgomez 851 if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma
2701 :     uint32_t iDcScaler = get_dc_scaler(iQuant, 0);
2702 :     //chroma U
2703 :     transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2704 :     fdct(in);
2705 :     in[0] -= 1024;
2706 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2707 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2708 : chl 96
2709 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2710 :     if (t != 0) cbp |= 1 << (5 - 4);
2711 :     Data->temp[4] = t;
2712 : chl 141
2713 : edgomez 851 if (bits < Data->iMinSAD[0]) {
2714 :     //chroma V
2715 :     transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2716 :     fdct(in);
2717 :     in[0] -= 1024;
2718 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2719 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2720 : chl 96
2721 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2722 :     if (t != 0) cbp |= 1 << (5 - 5);
2723 : chl 96
2724 : edgomez 851 Data->temp[5] = t;
2725 : edgomez 195
2726 : chl 897 bits += t = xvid_cbpy_tab[cbp>>2].len;
2727 : edgomez 851 Data->temp[6] = t;
2728 : chl 96
2729 : edgomez 851 bits += t = mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2730 :     Data->temp[7] = t;
2731 : edgomez 195
2732 : chl 96 }
2733 :     }
2734 :    
2735 : edgomez 851 return bits;
2736 : chl 96 }

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