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

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

Parent Directory Parent Directory | Revision Log Revision Log


Revision 962 - (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 : edgomez 953 (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 : edgomez 953 if ((xf > data->max_dx) || (xf < data->min_dx) ||
510 :     (yf > data->max_dy) || (yf < data->min_dy))
511 :     return;
512 : edgomez 851
513 :     if (!data->qpel_precision) {
514 :     ReferenceF = GetReference(xf, yf, data);
515 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
516 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
517 :     current = data->currentMV;
518 :     xcf = xf; ycf = yf;
519 :     xcb = xb; ycb = yb;
520 :     } else {
521 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
522 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
523 :     current = data->currentQMV;
524 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
525 :     xcf = xf/2; ycf = yf/2;
526 :     xcb = xb/2; ycb = yb/2;
527 :     }
528 :    
529 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
530 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
531 :    
532 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
533 :     sad += (data->lambda16 * t * sad)>>10;
534 :    
535 :     if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
536 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
537 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
538 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
539 :    
540 :     if (sad < *(data->iMinSAD)) {
541 :     *(data->iMinSAD) = sad;
542 :     current->x = xf; current->y = yf;
543 :     *dir = Direction;
544 :     }
545 : Isibaar 3 }
546 :    
547 : edgomez 851 static void
548 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
549 :     {
550 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
551 :     uint32_t k;
552 :     const uint8_t *ReferenceF;
553 :     const uint8_t *ReferenceB;
554 :     VECTOR mvs, b_mvs;
555 :    
556 : edgomez 953 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
557 : edgomez 851
558 :     for (k = 0; k < 4; k++) {
559 :     mvs.x = data->directmvF[k].x + x;
560 :     b_mvs.x = ((x == 0) ?
561 :     data->directmvB[k].x
562 :     : mvs.x - data->referencemv[k].x);
563 :    
564 :     mvs.y = data->directmvF[k].y + y;
565 :     b_mvs.y = ((y == 0) ?
566 :     data->directmvB[k].y
567 :     : mvs.y - data->referencemv[k].y);
568 :    
569 : edgomez 953 if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) ||
570 :     (mvs.y > data->max_dy) || (mvs.y < data->min_dy) ||
571 :     (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) ||
572 :     (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) )
573 :     return;
574 : edgomez 851
575 :     if (data->qpel) {
576 :     xcf += mvs.x/2; ycf += mvs.y/2;
577 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
578 :     } else {
579 :     xcf += mvs.x; ycf += mvs.y;
580 :     xcb += b_mvs.x; ycb += b_mvs.y;
581 :     mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway
582 :     b_mvs.x *= 2; b_mvs.y *= 2;
583 :     }
584 :    
585 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
586 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
587 :    
588 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
589 :     ReferenceF, ReferenceB, data->iEdgedWidth);
590 :     if (sad > *(data->iMinSAD)) return;
591 :     }
592 :    
593 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
594 :    
595 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
596 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
597 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
598 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
599 :    
600 :     if (sad < *(data->iMinSAD)) {
601 :     *(data->iMinSAD) = sad;
602 :     data->currentMV->x = x; data->currentMV->y = y;
603 :     *dir = Direction;
604 :     }
605 : Isibaar 3 }
606 :    
607 : edgomez 851 static void
608 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
609 : Isibaar 3 {
610 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
611 :     const uint8_t *ReferenceF;
612 :     const uint8_t *ReferenceB;
613 :     VECTOR mvs, b_mvs;
614 : Isibaar 3
615 : edgomez 851 if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return;
616 : Isibaar 3
617 : edgomez 851 mvs.x = data->directmvF[0].x + x;
618 :     b_mvs.x = ((x == 0) ?
619 :     data->directmvB[0].x
620 :     : mvs.x - data->referencemv[0].x);
621 : Isibaar 3
622 : edgomez 851 mvs.y = data->directmvF[0].y + y;
623 :     b_mvs.y = ((y == 0) ?
624 :     data->directmvB[0].y
625 :     : mvs.y - data->referencemv[0].y);
626 : Isibaar 3
627 : edgomez 851 if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx)
628 :     | (mvs.y > data->max_dy) | (mvs.y < data->min_dy)
629 :     | (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx)
630 :     | (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return;
631 : Isibaar 3
632 : edgomez 851 if (data->qpel) {
633 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
634 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
635 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
636 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
637 :     } else {
638 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
639 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
640 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
641 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
642 :     }
643 :    
644 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
645 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
646 :    
647 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
648 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
649 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
650 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
651 :    
652 :     if (sad < *(data->iMinSAD)) {
653 :     *(data->iMinSAD) = sad;
654 :     data->currentMV->x = x; data->currentMV->y = y;
655 :     *dir = Direction;
656 :     }
657 : Isibaar 3 }
658 :    
659 : edgomez 851
660 :     static void
661 :     CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
662 : Isibaar 3 {
663 :    
664 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
665 : edgomez 851 int32_t bits = 0, sum;
666 :     VECTOR * current;
667 :     const uint8_t * ptr;
668 :     int i, cbp = 0, t, xc, yc;
669 : edgomez 195
670 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
671 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
672 : edgomez 195
673 : edgomez 851 if (!data->qpel_precision) {
674 :     ptr = GetReference(x, y, data);
675 :     current = data->currentMV;
676 :     xc = x; yc = y;
677 :     } else { // x and y are in 1/4 precision
678 :     ptr = Interpolate16x16qpel(x, y, 0, data);
679 :     current = data->currentQMV;
680 :     xc = x/2; yc = y/2;
681 :     }
682 : Isibaar 3
683 : edgomez 851 for(i = 0; i < 4; i++) {
684 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
685 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
686 :     fdct(in);
687 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
688 :     else sum = quant4_inter(coeff, in, data->lambda16);
689 :     if (sum > 0) {
690 :     cbp |= 1 << (5 - i);
691 :     bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
692 :     } else data->temp[i] = 0;
693 :     }
694 : Isibaar 3
695 : edgomez 851 bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
696 : edgomez 195
697 : edgomez 851 if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma
698 :     xc = (xc >> 1) + roundtab_79[xc & 0x3];
699 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
700 :    
701 :     //chroma U
702 :     ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCU, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
703 :     transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
704 :     fdct(in);
705 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
706 :     else sum = quant4_inter(coeff, in, data->lambda16);
707 :     if (sum > 0) {
708 :     cbp |= 1 << (5 - 4);
709 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
710 : chl 344 }
711 : edgomez 851
712 :     if (bits < data->iMinSAD[0]) {
713 :     //chroma V
714 :     ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCV, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
715 :     transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2);
716 :     fdct(in);
717 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
718 :     else sum = quant4_inter(coeff, in, data->lambda16);
719 :     if (sum > 0) {
720 :     cbp |= 1 << (5 - 5);
721 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
722 :     }
723 :     }
724 : edgomez 78 }
725 : edgomez 851
726 : edgomez 958 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
727 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
728 :    
729 :     if (bits < data->iMinSAD[0]) {
730 :     data->iMinSAD[0] = bits;
731 :     current[0].x = x; current[0].y = y;
732 :     *dir = Direction;
733 :     }
734 :    
735 :     if (data->temp[0] + t < data->iMinSAD[1]) {
736 :     data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
737 :     if (data->temp[1] < data->iMinSAD[2]) {
738 :     data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
739 :     if (data->temp[2] < data->iMinSAD[3]) {
740 :     data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
741 :     if (data->temp[3] < data->iMinSAD[4]) {
742 :     data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
743 :    
744 : Isibaar 3 }
745 : edgomez 851 static void
746 :     CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
747 : chl 96 {
748 :    
749 : edgomez 962 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
750 : edgomez 851 int32_t sum, bits;
751 :     VECTOR * current;
752 :     const uint8_t * ptr;
753 :     int cbp;
754 : edgomez 195
755 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
756 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
757 : edgomez 195
758 : edgomez 851 if (!data->qpel_precision) {
759 :     ptr = GetReference(x, y, data);
760 :     current = data->currentMV;
761 :     } else { // x and y are in 1/4 precision
762 :     ptr = Interpolate8x8qpel(x, y, 0, 0, data);
763 :     current = data->currentQMV;
764 :     }
765 : chl 96
766 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
767 :     fdct(in);
768 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
769 :     else sum = quant4_inter(coeff, in, data->lambda16);
770 :     if (sum > 0) {
771 :     bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
772 :     cbp = 1;
773 :     } else cbp = bits = 0;
774 : chl 96
775 : edgomez 851 bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
776 : chl 96
777 : edgomez 851 if (bits < data->iMinSAD[0]) {
778 :     data->temp[0] = cbp;
779 :     data->iMinSAD[0] = bits;
780 :     current[0].x = x; current[0].y = y;
781 :     *dir = Direction;
782 :     }
783 :     }
784 : chl 96
785 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
786 : edgomez 195
787 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
788 : chl 96
789 : edgomez 851 static void
790 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
791 :     {
792 : edgomez 195
793 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
794 : edgomez 195
795 : edgomez 851 int iDirection;
796 : edgomez 195
797 : edgomez 851 for(;;) { //forever
798 :     iDirection = 0;
799 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
800 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
801 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
802 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
803 : edgomez 195
804 : edgomez 851 /* now we're doing diagonal checks near our candidate */
805 : edgomez 195
806 : edgomez 851 if (iDirection) { //if anything found
807 :     bDirection = iDirection;
808 :     iDirection = 0;
809 :     x = data->currentMV->x; y = data->currentMV->y;
810 :     if (bDirection & 3) { //our candidate is left or right
811 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
812 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
813 :     } else { // what remains here is up or down
814 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
815 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
816 :     }
817 : edgomez 195
818 : edgomez 851 if (iDirection) {
819 :     bDirection += iDirection;
820 :     x = data->currentMV->x; y = data->currentMV->y;
821 :     }
822 :     } else { //about to quit, eh? not so fast....
823 :     switch (bDirection) {
824 :     case 2:
825 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
826 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
827 : edgomez 195 break;
828 : edgomez 851 case 1:
829 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
830 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
831 : edgomez 195 break;
832 : edgomez 851 case 2 + 4:
833 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
834 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
835 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
836 :     break;
837 :     case 4:
838 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
839 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
840 :     break;
841 : edgomez 195 case 8:
842 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
843 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
844 : edgomez 195 break;
845 : edgomez 851 case 1 + 4:
846 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
847 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
848 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
849 : edgomez 195 break;
850 : edgomez 851 case 2 + 8:
851 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
852 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
853 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
854 :     break;
855 :     case 1 + 8:
856 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
857 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
858 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
859 :     break;
860 :     default: //1+2+4+8 == we didn't find anything at all
861 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
862 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
863 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
864 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
865 :     break;
866 : chl 96 }
867 : edgomez 851 if (!iDirection) break; //ok, the end. really
868 :     bDirection = iDirection;
869 :     x = data->currentMV->x; y = data->currentMV->y;
870 : chl 345 }
871 : edgomez 195 }
872 : chl 96 }
873 :    
874 : edgomez 851 static void
875 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
876 : chl 96 {
877 : edgomez 851 int iDirection;
878 : edgomez 195
879 : edgomez 851 do {
880 :     iDirection = 0;
881 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
882 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
883 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
884 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
885 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
886 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
887 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
888 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
889 : chl 96
890 : edgomez 851 bDirection = iDirection;
891 :     x = data->currentMV->x; y = data->currentMV->y;
892 :     } while (iDirection);
893 : chl 96 }
894 :    
895 : edgomez 851 static void
896 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
897 : chl 181 {
898 :    
899 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
900 :    
901 : edgomez 851 int iDirection;
902 : edgomez 195
903 : edgomez 851 do {
904 :     iDirection = 0;
905 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
906 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
907 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
908 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
909 :    
910 :     /* now we're doing diagonal checks near our candidate */
911 :    
912 :     if (iDirection) { //checking if anything found
913 :     bDirection = iDirection;
914 : chl 181 iDirection = 0;
915 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
916 :     if (bDirection & 3) { //our candidate is left or right
917 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
918 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
919 :     } else { // what remains here is up or down
920 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
921 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
922 :     }
923 :     bDirection += iDirection;
924 :     x = data->currentMV->x; y = data->currentMV->y;
925 :     }
926 :     }
927 :     while (iDirection);
928 :     }
929 : chl 181
930 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
931 : chl 181
932 : edgomez 851 static void
933 :     SubpelRefine(const SearchData * const data)
934 :     {
935 :     /* Do a half-pel or q-pel refinement */
936 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
937 :     int iDirection; //only needed because macro expects it
938 : chl 181
939 : edgomez 851 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
940 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
941 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
942 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
943 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
944 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
945 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
946 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
947 :     }
948 : chl 181
949 : edgomez 851 static __inline int
950 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
951 :     const int x, const int y,
952 :     const uint32_t stride, const uint32_t iQuant, int rrv)
953 : chl 181
954 : edgomez 851 {
955 :     if(!rrv) {
956 :     uint32_t sadC = sad8(current->u + x*8 + y*stride*8,
957 :     reference->u + x*8 + y*stride*8, stride);
958 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
959 :     sadC += sad8(current->v + (x + y*stride)*8,
960 :     reference->v + (x + y*stride)*8, stride);
961 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
962 :     return 1;
963 : chl 181
964 : edgomez 851 } else {
965 :     uint32_t sadC = sad16(current->u + x*16 + y*stride*16,
966 :     reference->u + x*16 + y*stride*16, stride, 256*4096);
967 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
968 :     sadC += sad16(current->v + (x + y*stride)*16,
969 :     reference->v + (x + y*stride)*16, stride, 256*4096);
970 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
971 :     return 1;
972 : chl 181 }
973 : edgomez 851 }
974 : edgomez 504
975 : edgomez 851 static __inline void
976 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
977 :     {
978 :     pMB->mode = MODE_NOT_CODED;
979 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
980 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
981 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
982 : chl 181 }
983 :    
984 : edgomez 851 bool
985 :     MotionEstimation(MBParam * const pParam,
986 :     FRAMEINFO * const current,
987 :     FRAMEINFO * const reference,
988 :     const IMAGE * const pRefH,
989 :     const IMAGE * const pRefV,
990 :     const IMAGE * const pRefHV,
991 :     const uint32_t iLimit)
992 : chl 181 {
993 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
994 :     const IMAGE *const pCurrent = &current->image;
995 :     const IMAGE *const pRef = &reference->image;
996 : chl 181
997 : edgomez 851 uint32_t mb_width = pParam->mb_width;
998 :     uint32_t mb_height = pParam->mb_height;
999 :     const uint32_t iEdgedWidth = pParam->edged_width;
1000 : suxen_drol 890 const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
1001 : chl 181
1002 : edgomez 851 uint32_t x, y;
1003 :     uint32_t iIntra = 0;
1004 : edgomez 953 int32_t quant = current->quant, sad00;
1005 :     int skip_thresh = \
1006 :     INITIAL_SKIP_THRESH * \
1007 :     (current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \
1008 :     (current->vop_flags & XVID_VOP_MODEDECISION_BITS ? 2:1);
1009 : chl 181
1010 : edgomez 851 // some pre-initialized thingies for SearchP
1011 :     int32_t temp[8];
1012 :     VECTOR currentMV[5];
1013 :     VECTOR currentQMV[5];
1014 :     int32_t iMinSAD[5];
1015 : edgomez 962 DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE);
1016 : edgomez 851 SearchData Data;
1017 :     memset(&Data, 0, sizeof(SearchData));
1018 :     Data.iEdgedWidth = iEdgedWidth;
1019 :     Data.currentMV = currentMV;
1020 :     Data.currentQMV = currentQMV;
1021 :     Data.iMinSAD = iMinSAD;
1022 :     Data.temp = temp;
1023 :     Data.iFcode = current->fcode;
1024 :     Data.rounding = pParam->m_rounding_type;
1025 : edgomez 949 Data.qpel = current->vol_flags & XVID_VOL_QUARTERPEL;
1026 :     Data.chroma = MotionFlags & XVID_ME_CHROMA16;
1027 :     Data.rrv = current->vop_flags & XVID_VOP_REDUCED;
1028 : edgomez 962 Data.dctSpace = dct_space;
1029 : edgomez 195
1030 : edgomez 949 if ((current->vop_flags & XVID_VOP_REDUCED)) {
1031 : edgomez 851 mb_width = (pParam->width + 31) / 32;
1032 :     mb_height = (pParam->height + 31) / 32;
1033 :     Data.qpel = 0;
1034 :     }
1035 : chl 181
1036 : edgomez 851 Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose)
1037 :     if (sadInit) (*sadInit) ();
1038 : chl 181
1039 : edgomez 851 for (y = 0; y < mb_height; y++) {
1040 :     for (x = 0; x < mb_width; x++) {
1041 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1042 : chl 181
1043 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1044 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1045 :     pRef->y + (x + y * iEdgedWidth) * 16,
1046 :     pParam->edged_width, pMB->sad8 );
1047 : chl 181
1048 : edgomez 851 else pMB->sad16 =
1049 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1050 :     pRef->y + (x + y * iEdgedWidth) * 32,
1051 :     pParam->edged_width, pMB->sad8 );
1052 : chl 181
1053 : edgomez 851 if (Data.chroma) {
1054 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1055 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1056 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1057 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1058 :     pMB->sad16 += Data.temp[7];
1059 :     }
1060 :    
1061 :     sad00 = pMB->sad16;
1062 :    
1063 : edgomez 953 if (pMB->dquant != 0) {
1064 :     quant += DQtab[pMB->dquant];
1065 :     if (quant > 31) quant = 31;
1066 :     else if (quant < 1) quant = 1;
1067 :     }
1068 :    
1069 :     pMB->quant = current->quant;
1070 :    
1071 : edgomez 851 //initial skip decision
1072 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1073 : edgomez 949 if (!(current->vol_flags & XVID_VOL_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1074 : edgomez 953 if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh)
1075 : edgomez 851 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1076 :     SkipMacroblockP(pMB, sad00);
1077 :     continue;
1078 :     }
1079 :     }
1080 :    
1081 :     SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1082 : suxen_drol 890 y, MotionFlags, current->vol_flags, pMB->quant,
1083 : edgomez 851 &Data, pParam, pMBs, reference->mbs,
1084 : edgomez 949 current->vop_flags & XVID_VOP_INTER4V, pMB);
1085 : edgomez 851
1086 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
1087 : edgomez 953 if (!(current->vol_flags & XVID_VOL_GMC || current->vop_flags & XVID_VOP_MODEDECISION_BITS)) {
1088 : suxen_drol 926 if ( pMB->dquant == 0 && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) {
1089 : edgomez 953 if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) )
1090 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv))
1091 :     SkipMacroblockP(pMB, sad00);
1092 : chl 181 }
1093 :     }
1094 : edgomez 851 if (pMB->mode == MODE_INTRA)
1095 :     if (++iIntra > iLimit) return 1;
1096 : chl 181 }
1097 :     }
1098 : edgomez 851
1099 : edgomez 949 if (current->vol_flags & XVID_VOL_GMC ) /* GMC only for S(GMC)-VOPs */
1100 : edgomez 851 {
1101 :     current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1102 :     }
1103 :     return 0;
1104 : chl 181 }
1105 :    
1106 :    
1107 : edgomez 851 static __inline int
1108 :     make_mask(const VECTOR * const pmv, const int i)
1109 : chl 96 {
1110 : edgomez 851 int mask = 255, j;
1111 :     for (j = 0; j < i; j++) {
1112 :     if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
1113 :     if (pmv[i].x == pmv[j].x) {
1114 :     if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1115 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1116 :     } else
1117 :     if (pmv[i].y == pmv[j].y) {
1118 :     if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1119 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1120 :     }
1121 :     }
1122 :     return mask;
1123 :     }
1124 : edgomez 195
1125 : edgomez 851 static __inline void
1126 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1127 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1128 :     {
1129 : chl 96
1130 : edgomez 851 //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1131 :     if (rrv) { iWcount /= 2; iHcount /= 2; }
1132 : edgomez 195
1133 : edgomez 851 if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
1134 :     pmv[5].x = EVEN(pmv[3].x);
1135 :     pmv[5].y = EVEN(pmv[3].y);
1136 :     } else pmv[5].x = pmv[5].y = 0;
1137 : chl 96
1138 : edgomez 851 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1139 :     else pmv[3].x = pmv[3].y = 0;
1140 : chl 96
1141 : edgomez 851 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1142 :     else pmv[4].x = pmv[4].y = 0;
1143 : Isibaar 3
1144 : edgomez 851 // [1] median prediction
1145 :     pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1146 : edgomez 195
1147 : 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
1148 : edgomez 195
1149 : edgomez 851 pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1150 :     pmv[2].y = EVEN(prevMB->mvs[0].y);
1151 :    
1152 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1153 :     pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1154 :     pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1155 :     } else pmv[6].x = pmv[6].y = 0;
1156 :    
1157 :     if (rrv) {
1158 :     int i;
1159 :     for (i = 0; i < 7; i++) {
1160 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1161 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1162 :     }
1163 :     }
1164 : Isibaar 3 }
1165 :    
1166 : edgomez 851 static int
1167 :     ModeDecision(const uint32_t iQuant, SearchData * const Data,
1168 :     int inter4v,
1169 :     MACROBLOCK * const pMB,
1170 :     const MACROBLOCK * const pMBs,
1171 :     const int x, const int y,
1172 :     const MBParam * const pParam,
1173 :     const uint32_t MotionFlags,
1174 : suxen_drol 890 const uint32_t VopFlags)
1175 : edgomez 851 {
1176 : Isibaar 3
1177 : edgomez 851 int mode = MODE_INTER;
1178 : chl 96
1179 : edgomez 949 if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision
1180 : chl 871 // int intra = 0;
1181 : edgomez 851 int sad;
1182 :     int InterBias = MV16_INTER_BIAS;
1183 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1184 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1185 :     mode = 0; //inter
1186 :     sad = Data->iMinSAD[0];
1187 :     } else {
1188 :     mode = MODE_INTER4V;
1189 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1190 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1191 :     Data->iMinSAD[0] = sad;
1192 :     }
1193 : chl 326
1194 : edgomez 851 /* intra decision */
1195 : Isibaar 3
1196 : edgomez 851 if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work
1197 :     if (y != 0)
1198 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1199 :     if (x != 0)
1200 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1201 : Isibaar 3
1202 : edgomez 851 if (Data->chroma) InterBias += 50; // to compensate bigger SAD
1203 :     if (Data->rrv) InterBias *= 4;
1204 : edgomez 195
1205 : edgomez 851 if (InterBias < pMB->sad16) {
1206 :     int32_t deviation;
1207 :     if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
1208 :     else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
1209 :     dev16(Data->Cur+8, Data->iEdgedWidth) +
1210 :     dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) +
1211 :     dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth);
1212 : edgomez 195
1213 : edgomez 851 if (deviation < (sad - InterBias)) return MODE_INTRA;// intra
1214 :     }
1215 :     return mode;
1216 : Isibaar 3
1217 : edgomez 851 } else {
1218 : edgomez 195
1219 : edgomez 851 int bits, intra, i;
1220 :     VECTOR backup[5], *v;
1221 :     Data->lambda16 = iQuant;
1222 : edgomez 949 Data->lambda8 = (pParam->vol_flags & XVID_VOL_MPEGQUANT)?1:0;
1223 : chl 181
1224 : edgomez 851 v = Data->qpel ? Data->currentQMV : Data->currentMV;
1225 :     for (i = 0; i < 5; i++) {
1226 :     Data->iMinSAD[i] = 256*4096;
1227 :     backup[i] = v[i];
1228 :     }
1229 : Isibaar 3
1230 : edgomez 851 bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1231 :     if (bits == 0) return MODE_INTER; // quick stop
1232 : Isibaar 3
1233 : edgomez 851 if (inter4v) {
1234 :     int inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1235 :     if (inter4v < bits) { Data->iMinSAD[0] = bits = inter4v; mode = MODE_INTER4V; }
1236 :     }
1237 : edgomez 195
1238 : Isibaar 3
1239 : edgomez 851 intra = CountMBBitsIntra(Data);
1240 : Isibaar 3
1241 : edgomez 851 if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; }
1242 :    
1243 :     return mode;
1244 : edgomez 195 }
1245 : edgomez 851 }
1246 : Isibaar 3
1247 : edgomez 851 static void
1248 :     SearchP(const IMAGE * const pRef,
1249 :     const uint8_t * const pRefH,
1250 :     const uint8_t * const pRefV,
1251 :     const uint8_t * const pRefHV,
1252 :     const IMAGE * const pCur,
1253 :     const int x,
1254 :     const int y,
1255 :     const uint32_t MotionFlags,
1256 : suxen_drol 890 const uint32_t VopFlags,
1257 : edgomez 851 const uint32_t iQuant,
1258 :     SearchData * const Data,
1259 :     const MBParam * const pParam,
1260 :     const MACROBLOCK * const pMBs,
1261 :     const MACROBLOCK * const prevMBs,
1262 :     int inter4v,
1263 :     MACROBLOCK * const pMB)
1264 :     {
1265 : Isibaar 3
1266 : edgomez 851 int i, iDirection = 255, mask, threshA;
1267 :     VECTOR pmv[7];
1268 : Isibaar 3
1269 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1270 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1271 : edgomez 195
1272 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1273 : Isibaar 3
1274 : edgomez 851 Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache
1275 :     i = Data->rrv ? 2 : 1;
1276 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1277 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1278 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1279 : chl 326
1280 : edgomez 851 Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1281 :     Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1282 :     Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1283 :     Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1284 :     Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1285 :     Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1286 : edgomez 195
1287 : edgomez 851 Data->lambda16 = lambda_vec16[iQuant];
1288 :     Data->lambda8 = lambda_vec8[iQuant];
1289 :     Data->qpel_precision = 0;
1290 : edgomez 195
1291 : suxen_drol 926 if (pMB->dquant != 0) inter4v = 0;
1292 : edgomez 195
1293 : edgomez 851 for(i = 0; i < 5; i++)
1294 :     Data->currentMV[i].x = Data->currentMV[i].y = 0;
1295 : chl 169
1296 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1297 :     else Data->predMV = pmv[0];
1298 : Isibaar 3
1299 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1300 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1301 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1302 :     Data->iMinSAD[2] = pMB->sad8[1];
1303 :     Data->iMinSAD[3] = pMB->sad8[2];
1304 :     Data->iMinSAD[4] = pMB->sad8[3];
1305 : chl 169
1306 : edgomez 949 if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) || (x | y)) {
1307 : edgomez 851 threshA = Data->temp[0]; // that's where we keep this SAD atm
1308 :     if (threshA < 512) threshA = 512;
1309 :     else if (threshA > 1024) threshA = 1024;
1310 :     } else
1311 :     threshA = 512;
1312 : chl 169
1313 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1314 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1315 : chl 169
1316 : edgomez 851 if (!Data->rrv) {
1317 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1318 :     else CheckCandidate = CheckCandidate16no4v; //for extra speed
1319 :     } else CheckCandidate = CheckCandidate32;
1320 : chl 169
1321 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1322 : chl 169
1323 : edgomez 851 for (i = 1; i < 7; i++) {
1324 :     if (!(mask = make_mask(pmv, i)) ) continue;
1325 :     CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1326 :     if (Data->iMinSAD[0] <= threshA) break;
1327 :     }
1328 : chl 169
1329 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1330 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1331 :     (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1332 : edgomez 949 if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) inter4v = 0; }
1333 : edgomez 851 else {
1334 : Isibaar 3
1335 : edgomez 851 MainSearchFunc * MainSearchPtr;
1336 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1337 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1338 : edgomez 851 else MainSearchPtr = DiamondSearch;
1339 : Isibaar 3
1340 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1341 : Isibaar 3
1342 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1343 :     note that this search is/might be done in halfpel positions,
1344 :     which makes it more different than the diamond above */
1345 : chl 169
1346 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH16) {
1347 : edgomez 851 int32_t bSAD;
1348 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1349 :     if (Data->rrv) {
1350 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1351 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1352 :     }
1353 :     if (!(MVequal(startMV, backupMV))) {
1354 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1355 : edgomez 195
1356 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1357 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1358 :     if (bSAD < Data->iMinSAD[0]) {
1359 :     Data->currentMV[0] = backupMV;
1360 :     Data->iMinSAD[0] = bSAD; }
1361 :     }
1362 : chl 169
1363 : edgomez 851 backupMV = Data->currentMV[0];
1364 :     startMV.x = startMV.y = 1;
1365 :     if (!(MVequal(startMV, backupMV))) {
1366 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1367 : chl 169
1368 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1369 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1370 :     if (bSAD < Data->iMinSAD[0]) {
1371 :     Data->currentMV[0] = backupMV;
1372 :     Data->iMinSAD[0] = bSAD; }
1373 : edgomez 195 }
1374 : edgomez 851 }
1375 :     }
1376 : edgomez 195
1377 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE16)
1378 :     if ((!(MotionFlags & XVID_ME_HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant)
1379 : edgomez 851 SubpelRefine(Data);
1380 : edgomez 195
1381 : edgomez 851 for(i = 0; i < 5; i++) {
1382 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1383 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1384 :     }
1385 : Isibaar 3
1386 : edgomez 953 if (MotionFlags & XVID_ME_QUARTERPELREFINE16) {
1387 :    
1388 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1389 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1390 :    
1391 : edgomez 949 if ((!(MotionFlags & XVID_ME_QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) {
1392 : edgomez 851 Data->qpel_precision = 1;
1393 :     SubpelRefine(Data);
1394 :     }
1395 : edgomez 953 }
1396 : Isibaar 3
1397 : edgomez 949 if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0;
1398 : Isibaar 3
1399 : edgomez 949 if (inter4v && (!(VopFlags & XVID_VOP_MODEDECISION_BITS) ||
1400 :     (!(MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS)) || (!(MotionFlags & XVID_ME_HALFPELREFINE8_BITS)) ||
1401 :     ((!(MotionFlags & XVID_ME_EXTSEARCH_BITS)) && (!(MotionFlags&XVID_ME_EXTSEARCH8)) ))) {
1402 : edgomez 851 // if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop
1403 : Isibaar 3
1404 : edgomez 851 SearchData Data8;
1405 :     memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data
1406 : Isibaar 3
1407 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1408 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1409 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1410 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1411 : chl 181
1412 : edgomez 949 if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_BITS))) {
1413 : edgomez 851 // chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it
1414 :     int sumx = 0, sumy = 0;
1415 :     const int div = 1 + Data->qpel;
1416 :     const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs;
1417 : Isibaar 3
1418 : edgomez 851 for (i = 0; i < 4; i++) {
1419 :     sumx += mv[i].x / div;
1420 :     sumy += mv[i].y / div;
1421 :     }
1422 : chl 259
1423 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1424 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1425 :     }
1426 :     }
1427 : edgomez 195
1428 : suxen_drol 890 inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, VopFlags);
1429 : edgomez 851
1430 :     if (Data->rrv) {
1431 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1432 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1433 : Isibaar 3 }
1434 :    
1435 : edgomez 851 if (inter4v == MODE_INTER) {
1436 :     pMB->mode = MODE_INTER;
1437 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1438 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1439 : Isibaar 3
1440 : edgomez 851 if(Data->qpel) {
1441 :     pMB->qmvs[0] = pMB->qmvs[1]
1442 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1443 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1444 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1445 :     } else {
1446 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1447 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1448 : Isibaar 3 }
1449 :    
1450 : edgomez 851 } else if (inter4v == MODE_INTER4V) {
1451 :     pMB->mode = MODE_INTER4V;
1452 :     pMB->sad16 = Data->iMinSAD[0];
1453 :     } else { // INTRA mode
1454 :     SkipMacroblockP(pMB, 0); // not skip, but similar enough
1455 :     pMB->mode = MODE_INTRA;
1456 :     }
1457 : edgomez 195
1458 : edgomez 851 }
1459 :    
1460 :     static void
1461 :     Search8(const SearchData * const OldData,
1462 :     const int x, const int y,
1463 :     const uint32_t MotionFlags,
1464 :     const MBParam * const pParam,
1465 :     MACROBLOCK * const pMB,
1466 :     const MACROBLOCK * const pMBs,
1467 :     const int block,
1468 :     SearchData * const Data)
1469 :     {
1470 :     int i = 0;
1471 :     Data->iMinSAD = OldData->iMinSAD + 1 + block;
1472 :     Data->currentMV = OldData->currentMV + 1 + block;
1473 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1474 :    
1475 :     if(Data->qpel) {
1476 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1477 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1478 :     Data->predMV, Data->iFcode, 0, 0);
1479 :     } else {
1480 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1481 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1482 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1483 : Isibaar 3 }
1484 :    
1485 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1486 : Isibaar 3
1487 : edgomez 949 if (MotionFlags & (XVID_ME_EXTSEARCH8|XVID_ME_HALFPELREFINE8|XVID_ME_QUARTERPELREFINE8)) {
1488 : edgomez 851 if (Data->rrv) i = 2; else i = 1;
1489 : Isibaar 3
1490 : edgomez 851 Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1491 :     Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1492 :     Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1493 :     Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1494 : Isibaar 3
1495 : edgomez 851 Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1496 :     Data->qpel_precision = 0;
1497 : Isibaar 3
1498 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1499 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1500 : Isibaar 3
1501 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1502 :     else CheckCandidate = CheckCandidate16no4v;
1503 : Isibaar 3
1504 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_BITS))) {
1505 : edgomez 851 int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1506 : Isibaar 3
1507 : edgomez 851 MainSearchFunc *MainSearchPtr;
1508 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch;
1509 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1510 : edgomez 851 else MainSearchPtr = DiamondSearch;
1511 : Isibaar 3
1512 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1513 : Isibaar 3
1514 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1515 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1516 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1517 :     }
1518 :     }
1519 : edgomez 195
1520 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8) {
1521 : edgomez 851 int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1522 : edgomez 195
1523 : edgomez 851 SubpelRefine(Data); // perform halfpel refine of current best vector
1524 : Isibaar 3
1525 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1526 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1527 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1528 :     }
1529 :     }
1530 : Isibaar 3
1531 : edgomez 949 if (Data->qpel && MotionFlags & XVID_ME_QUARTERPELREFINE8) {
1532 : edgomez 851 Data->qpel_precision = 1;
1533 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1534 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1535 :     SubpelRefine(Data);
1536 :     }
1537 :     }
1538 : edgomez 195
1539 : edgomez 851 if (Data->rrv) {
1540 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1541 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1542 :     }
1543 :    
1544 :     if(Data->qpel) {
1545 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1546 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1547 :     pMB->qmvs[block] = *Data->currentQMV;
1548 : edgomez 195 } else {
1549 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1550 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1551 : edgomez 78 }
1552 : edgomez 851
1553 :     pMB->mvs[block] = *Data->currentMV;
1554 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1555 : Isibaar 3 }
1556 :    
1557 : edgomez 851 /* motion estimation for B-frames */
1558 : chl 345
1559 : edgomez 851 static __inline VECTOR
1560 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1561 :     {
1562 :     /* the stupidiest function ever */
1563 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1564 :     }
1565 : chl 345
1566 : edgomez 851 static void __inline
1567 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1568 :     const uint32_t iWcount,
1569 :     const MACROBLOCK * const pMB,
1570 :     const uint32_t mode_curr)
1571 : chl 345 {
1572 :    
1573 : edgomez 851 // [0] is prediction
1574 :     pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1575 : chl 345
1576 : edgomez 851 pmv[1].x = pmv[1].y = 0; // [1] is zero
1577 : chl 345
1578 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1579 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1580 : chl 345
1581 : edgomez 851 if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1582 :     pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1583 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1584 :     } else pmv[3].x = pmv[3].y = 0;
1585 : chl 345
1586 : edgomez 851 if (y != 0) {
1587 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1588 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1589 :     } else pmv[4].x = pmv[4].y = 0;
1590 : chl 345
1591 : edgomez 851 if (x != 0) {
1592 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1593 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1594 :     } else pmv[5].x = pmv[5].y = 0;
1595 : chl 345
1596 : edgomez 851 if (x != 0 && y != 0) {
1597 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1598 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1599 :     } else pmv[6].x = pmv[6].y = 0;
1600 :     }
1601 : chl 345
1602 :    
1603 : edgomez 851 /* search backward or forward */
1604 :     static void
1605 :     SearchBF( const IMAGE * const pRef,
1606 :     const uint8_t * const pRefH,
1607 :     const uint8_t * const pRefV,
1608 :     const uint8_t * const pRefHV,
1609 :     const IMAGE * const pCur,
1610 :     const int x, const int y,
1611 :     const uint32_t MotionFlags,
1612 :     const uint32_t iFcode,
1613 :     const MBParam * const pParam,
1614 :     MACROBLOCK * const pMB,
1615 :     const VECTOR * const predMV,
1616 :     int32_t * const best_sad,
1617 :     const int32_t mode_current,
1618 :     SearchData * const Data)
1619 :     {
1620 : chl 345
1621 : edgomez 851 int i, iDirection = 255, mask;
1622 :     VECTOR pmv[7];
1623 :     MainSearchFunc *MainSearchPtr;
1624 :     *Data->iMinSAD = MV_MAX_ERROR;
1625 :     Data->iFcode = iFcode;
1626 :     Data->qpel_precision = 0;
1627 :     Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache
1628 : chl 345
1629 : edgomez 851 Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16;
1630 :     Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16;
1631 :     Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16;
1632 :     Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16;
1633 :     Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8;
1634 :     Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8;
1635 : chl 345
1636 : edgomez 851 Data->predMV = *predMV;
1637 : chl 345
1638 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1639 :     pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1640 : chl 345
1641 : edgomez 851 pmv[0] = Data->predMV;
1642 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1643 : chl 345
1644 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1645 : chl 345
1646 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1647 :     CheckCandidate = CheckCandidate16no4v;
1648 : chl 345
1649 : edgomez 851 // main loop. checking all predictions
1650 :     for (i = 0; i < 7; i++) {
1651 :     if (!(mask = make_mask(pmv, i)) ) continue;
1652 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1653 : chl 345 }
1654 :    
1655 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1656 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1657 : edgomez 851 else MainSearchPtr = DiamondSearch;
1658 : chl 345
1659 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1660 : chl 345
1661 : edgomez 851 SubpelRefine(Data);
1662 : chl 345
1663 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1664 :     Data->currentQMV->x = 2*Data->currentMV->x;
1665 :     Data->currentQMV->y = 2*Data->currentMV->y;
1666 :     Data->qpel_precision = 1;
1667 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1668 :     pParam->width, pParam->height, iFcode, 1, 0);
1669 :     SubpelRefine(Data);
1670 :     }
1671 : chl 345
1672 : edgomez 851 // three bits are needed to code backward mode. four for forward
1673 : Isibaar 3
1674 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1675 :     else *Data->iMinSAD += 3 * Data->lambda16;
1676 : edgomez 195
1677 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1678 :     *best_sad = *Data->iMinSAD;
1679 :     pMB->mode = mode_current;
1680 :     if (Data->qpel) {
1681 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1682 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1683 :     if (mode_current == MODE_FORWARD)
1684 :     pMB->qmvs[0] = *Data->currentQMV;
1685 :     else
1686 :     pMB->b_qmvs[0] = *Data->currentQMV;
1687 :     } else {
1688 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1689 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1690 :     }
1691 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1692 :     else pMB->b_mvs[0] = *Data->currentMV;
1693 :     }
1694 : edgomez 195
1695 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1696 :     else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
1697 : Isibaar 3 }
1698 :    
1699 : edgomez 851 static void
1700 :     SkipDecisionB(const IMAGE * const pCur,
1701 :     const IMAGE * const f_Ref,
1702 :     const IMAGE * const b_Ref,
1703 :     MACROBLOCK * const pMB,
1704 :     const uint32_t x, const uint32_t y,
1705 :     const SearchData * const Data)
1706 :     {
1707 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1708 :     int32_t sum;
1709 :     const int div = 1 + Data->qpel;
1710 :     int k;
1711 :     const uint32_t stride = Data->iEdgedWidth/2;
1712 :     //this is not full chroma compensation, only it's fullpel approximation. should work though
1713 : Isibaar 3
1714 : edgomez 851 for (k = 0; k < 4; k++) {
1715 :     dy += Data->directmvF[k].y / div;
1716 :     dx += Data->directmvF[0].x / div;
1717 :     b_dy += Data->directmvB[0].y / div;
1718 :     b_dx += Data->directmvB[0].x / div;
1719 :     }
1720 : Isibaar 3
1721 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1722 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1723 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1724 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1725 : Isibaar 3
1726 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1727 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1728 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1729 :     stride);
1730 : Isibaar 3
1731 : edgomez 851 if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip
1732 : Isibaar 3
1733 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1734 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1735 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1736 :     stride);
1737 : edgomez 195
1738 : edgomez 851 if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1739 :     }
1740 : Isibaar 3
1741 : edgomez 851 static __inline uint32_t
1742 :     SearchDirect(const IMAGE * const f_Ref,
1743 :     const uint8_t * const f_RefH,
1744 :     const uint8_t * const f_RefV,
1745 :     const uint8_t * const f_RefHV,
1746 :     const IMAGE * const b_Ref,
1747 :     const uint8_t * const b_RefH,
1748 :     const uint8_t * const b_RefV,
1749 :     const uint8_t * const b_RefHV,
1750 :     const IMAGE * const pCur,
1751 :     const int x, const int y,
1752 :     const uint32_t MotionFlags,
1753 :     const int32_t TRB, const int32_t TRD,
1754 :     const MBParam * const pParam,
1755 :     MACROBLOCK * const pMB,
1756 :     const MACROBLOCK * const b_mb,
1757 :     int32_t * const best_sad,
1758 :     SearchData * const Data)
1759 : Isibaar 3
1760 : edgomez 851 {
1761 :     int32_t skip_sad;
1762 :     int k = (x + Data->iEdgedWidth*y) * 16;
1763 :     MainSearchFunc *MainSearchPtr;
1764 : Isibaar 3
1765 : edgomez 851 *Data->iMinSAD = 256*4096;
1766 :     Data->Ref = f_Ref->y + k;
1767 :     Data->RefH = f_RefH + k;
1768 :     Data->RefV = f_RefV + k;
1769 :     Data->RefHV = f_RefHV + k;
1770 :     Data->bRef = b_Ref->y + k;
1771 :     Data->bRefH = b_RefH + k;
1772 :     Data->bRefV = b_RefV + k;
1773 :     Data->bRefHV = b_RefHV + k;
1774 :     Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1775 :     Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1776 :     Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1777 :     Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1778 : edgomez 195
1779 : edgomez 851 k = Data->qpel ? 4 : 2;
1780 :     Data->max_dx = k * (pParam->width - x * 16);
1781 :     Data->max_dy = k * (pParam->height - y * 16);
1782 :     Data->min_dx = -k * (16 + x * 16);
1783 :     Data->min_dy = -k * (16 + y * 16);
1784 : chl 181
1785 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1786 :     Data->qpel_precision = 0;
1787 : edgomez 170
1788 : edgomez 851 for (k = 0; k < 4; k++) {
1789 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1790 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1791 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1792 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1793 : Isibaar 3
1794 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1795 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1796 :    
1797 :     *best_sad = 256*4096; // in that case, we won't use direct mode
1798 :     pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1799 :     pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1800 :     return 256*4096;
1801 :     }
1802 :     if (b_mb->mode != MODE_INTER4V) {
1803 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1804 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1805 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1806 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1807 :     break;
1808 :     }
1809 : edgomez 195 }
1810 : Isibaar 3
1811 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1812 : Isibaar 3
1813 : edgomez 851 CheckCandidate(0, 0, 255, &k, Data);
1814 : edgomez 195
1815 : edgomez 851 // initial (fast) skip decision
1816 :     if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) {
1817 :     //possible skip
1818 :     if (Data->chroma) {
1819 :     pMB->mode = MODE_DIRECT_NONE_MV;
1820 :     return *Data->iMinSAD; // skip.
1821 :     } else {
1822 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1823 :     if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1824 :     }
1825 : Isibaar 3 }
1826 :    
1827 : edgomez 953 *Data->iMinSAD += Data->lambda16;
1828 : edgomez 851 skip_sad = *Data->iMinSAD;
1829 : edgomez 195
1830 : edgomez 851 // DIRECT MODE DELTA VECTOR SEARCH.
1831 :     // This has to be made more effective, but at the moment I'm happy it's running at all
1832 : Isibaar 3
1833 : edgomez 949 if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch;
1834 :     else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1835 : edgomez 851 else MainSearchPtr = DiamondSearch;
1836 : Isibaar 3
1837 : edgomez 851 MainSearchPtr(0, 0, Data, 255);
1838 : Isibaar 3
1839 : edgomez 851 SubpelRefine(Data);
1840 : chl 181
1841 : edgomez 851 *best_sad = *Data->iMinSAD;
1842 : chl 181
1843 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1844 :     else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1845 : chl 181
1846 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1847 : edgomez 195
1848 : edgomez 851 for (k = 0; k < 4; k++) {
1849 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1850 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1851 :     ? Data->directmvB[k].x
1852 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1853 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1854 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1855 :     ? Data->directmvB[k].y
1856 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1857 :     if (Data->qpel) {
1858 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1859 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1860 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1861 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1862 :     }
1863 : edgomez 195
1864 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1865 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1866 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1867 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1868 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1869 :     break;
1870 :     }
1871 : edgomez 78 }
1872 : edgomez 851 return skip_sad;
1873 :     }
1874 : Isibaar 3
1875 : edgomez 851 static void
1876 :     SearchInterpolate(const IMAGE * const f_Ref,
1877 :     const uint8_t * const f_RefH,
1878 :     const uint8_t * const f_RefV,
1879 :     const uint8_t * const f_RefHV,
1880 :     const IMAGE * const b_Ref,
1881 :     const uint8_t * const b_RefH,
1882 :     const uint8_t * const b_RefV,
1883 :     const uint8_t * const b_RefHV,
1884 :     const IMAGE * const pCur,
1885 :     const int x, const int y,
1886 :     const uint32_t fcode,
1887 :     const uint32_t bcode,
1888 :     const uint32_t MotionFlags,
1889 :     const MBParam * const pParam,
1890 :     const VECTOR * const f_predMV,
1891 :     const VECTOR * const b_predMV,
1892 :     MACROBLOCK * const pMB,
1893 :     int32_t * const best_sad,
1894 :     SearchData * const fData)
1895 : chl 96
1896 : edgomez 851 {
1897 : chl 169
1898 : edgomez 851 int iDirection, i, j;
1899 :     SearchData bData;
1900 : chl 169
1901 : edgomez 851 fData->qpel_precision = 0;
1902 :     memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1903 :     *fData->iMinSAD = 4096*256;
1904 :     bData.currentMV++; bData.currentQMV++;
1905 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1906 : chl 169
1907 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1908 :     bData.bRef = fData->Ref = f_Ref->y + i;
1909 :     bData.bRefH = fData->RefH = f_RefH + i;
1910 :     bData.bRefV = fData->RefV = f_RefV + i;
1911 :     bData.bRefHV = fData->RefHV = f_RefHV + i;
1912 :     bData.Ref = fData->bRef = b_Ref->y + i;
1913 :     bData.RefH = fData->bRefH = b_RefH + i;
1914 :     bData.RefV = fData->bRefV = b_RefV + i;
1915 :     bData.RefHV = fData->bRefHV = b_RefHV + i;
1916 :     bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1917 :     bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1918 :     bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1919 :     bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1920 : chl 169
1921 :    
1922 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1923 :     fData->bpredMV = bData.predMV = *b_predMV;
1924 :     fData->currentMV[0] = fData->currentMV[2];
1925 : Isibaar 3
1926 : 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);
1927 :     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);
1928 : Isibaar 3
1929 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1930 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1931 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1932 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1933 : chl 169
1934 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1935 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1936 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1937 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1938 : Isibaar 3
1939 : edgomez 851 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1940 : chl 169
1941 : edgomez 851 //diamond
1942 :     do {
1943 :     iDirection = 255;
1944 :     // forward MV moves
1945 :     i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1946 : chl 169
1947 : edgomez 851 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1948 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1949 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1950 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1951 : edgomez 195
1952 : edgomez 851 // backward MV moves
1953 :     i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1954 :     fData->currentMV[2] = fData->currentMV[0];
1955 :     CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1956 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1957 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1958 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1959 : Isibaar 3
1960 : edgomez 851 } while (!(iDirection));
1961 : chl 140
1962 : edgomez 851 //qpel refinement
1963 :     if (fData->qpel) {
1964 :     if (*fData->iMinSAD > *best_sad + 500) return;
1965 :     CheckCandidate = CheckCandidateInt;
1966 :     fData->qpel_precision = bData.qpel_precision = 1;
1967 :     get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1968 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1969 :     fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1970 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1971 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1972 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1973 :     SubpelRefine(fData);
1974 :     if (*fData->iMinSAD > *best_sad + 300) return;
1975 :     fData->currentQMV[2] = fData->currentQMV[0];
1976 :     SubpelRefine(&bData);
1977 :     }
1978 : chl 140
1979 : edgomez 851 *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode.
1980 : Isibaar 3
1981 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
1982 :     *best_sad = *fData->iMinSAD;
1983 :     pMB->mvs[0] = fData->currentMV[0];
1984 :     pMB->b_mvs[0] = fData->currentMV[1];
1985 :     pMB->mode = MODE_INTERPOLATE;
1986 :     if (fData->qpel) {
1987 :     pMB->qmvs[0] = fData->currentQMV[0];
1988 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1989 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1990 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1991 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1992 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1993 :     } else {
1994 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1995 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1996 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1997 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1998 :     }
1999 : edgomez 78 }
2000 : edgomez 851 }
2001 : Isibaar 3
2002 : edgomez 851 void
2003 :     MotionEstimationBVOP(MBParam * const pParam,
2004 :     FRAMEINFO * const frame,
2005 :     const int32_t time_bp,
2006 :     const int32_t time_pp,
2007 :     // forward (past) reference
2008 :     const MACROBLOCK * const f_mbs,
2009 :     const IMAGE * const f_ref,
2010 :     const IMAGE * const f_refH,
2011 :     const IMAGE * const f_refV,
2012 :     const IMAGE * const f_refHV,
2013 :     // backward (future) reference
2014 :     const FRAMEINFO * const b_reference,
2015 :     const IMAGE * const b_ref,
2016 :     const IMAGE * const b_refH,
2017 :     const IMAGE * const b_refV,
2018 :     const IMAGE * const b_refHV)
2019 :     {
2020 :     uint32_t i, j;
2021 :     int32_t best_sad;
2022 :     uint32_t skip_sad;
2023 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2024 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
2025 : Isibaar 3
2026 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2027 : Isibaar 3
2028 : edgomez 851 const int32_t TRB = time_pp - time_bp;
2029 :     const int32_t TRD = time_pp;
2030 : edgomez 195
2031 : edgomez 851 // some pre-inintialized data for the rest of the search
2032 : Isibaar 3
2033 : edgomez 851 SearchData Data;
2034 :     int32_t iMinSAD;
2035 :     VECTOR currentMV[3];
2036 :     VECTOR currentQMV[3];
2037 :     int32_t temp[8];
2038 :     memset(&Data, 0, sizeof(SearchData));
2039 :     Data.iEdgedWidth = pParam->edged_width;
2040 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2041 :     Data.iMinSAD = &iMinSAD;
2042 :     Data.lambda16 = lambda_vec16[frame->quant];
2043 : edgomez 949 Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL;
2044 : edgomez 851 Data.rounding = 0;
2045 : edgomez 949 Data.chroma = frame->motion_flags & XVID_ME_CHROMA8;
2046 : edgomez 851 Data.temp = temp;
2047 : Isibaar 3
2048 : edgomez 851 Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose)
2049 :     // note: i==horizontal, j==vertical
2050 :     for (j = 0; j < pParam->mb_height; j++) {
2051 : edgomez 195
2052 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2053 : Isibaar 3
2054 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2055 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2056 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2057 : edgomez 195
2058 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2059 :     if (b_reference->coding_type != S_VOP)
2060 :     if (b_mb->mode == MODE_NOT_CODED) {
2061 :     pMB->mode = MODE_NOT_CODED;
2062 :     continue;
2063 :     }
2064 : Isibaar 3
2065 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2066 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2067 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2068 : edgomez 953 pMB->quant = frame->quant;
2069 : Isibaar 3
2070 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2071 :     and (2) sets very good predictions for forward and backward search */
2072 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2073 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2074 :     &frame->image,
2075 :     i, j,
2076 :     frame->motion_flags,
2077 :     TRB, TRD,
2078 :     pParam,
2079 :     pMB, b_mb,
2080 :     &best_sad,
2081 :     &Data);
2082 : Isibaar 3
2083 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2084 : edgomez 195
2085 : edgomez 851 // forward search
2086 :     SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2087 :     &frame->image, i, j,
2088 :     frame->motion_flags,
2089 :     frame->fcode, pParam,
2090 :     pMB, &f_predMV, &best_sad,
2091 :     MODE_FORWARD, &Data);
2092 : edgomez 195
2093 : edgomez 851 // backward search
2094 :     SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2095 :     &frame->image, i, j,
2096 :     frame->motion_flags,
2097 :     frame->bcode, pParam,
2098 :     pMB, &b_predMV, &best_sad,
2099 :     MODE_BACKWARD, &Data);
2100 :    
2101 :     // interpolate search comes last, because it uses data from forward and backward as prediction
2102 :     SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2103 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2104 :     &frame->image,
2105 :     i, j,
2106 :     frame->fcode, frame->bcode,
2107 :     frame->motion_flags,
2108 :     pParam,
2109 :     &f_predMV, &b_predMV,
2110 :     pMB, &best_sad,
2111 :     &Data);
2112 :    
2113 :     // final skip decision
2114 :     if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2115 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2116 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2117 :    
2118 :     switch (pMB->mode) {
2119 :     case MODE_FORWARD:
2120 :     f_count++;
2121 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2122 :     break;
2123 :     case MODE_BACKWARD:
2124 :     b_count++;
2125 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2126 :     break;
2127 :     case MODE_INTERPOLATE:
2128 :     i_count++;
2129 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2130 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2131 :     break;
2132 :     case MODE_DIRECT:
2133 :     case MODE_DIRECT_NO4V:
2134 :     d_count++;
2135 :     default:
2136 :     break;
2137 :     }
2138 :     }
2139 :     }
2140 : Isibaar 3 }
2141 : chl 96
2142 : edgomez 851 static __inline void
2143 :     MEanalyzeMB ( const uint8_t * const pRef,
2144 :     const uint8_t * const pCur,
2145 :     const int x,
2146 :     const int y,
2147 :     const MBParam * const pParam,
2148 :     MACROBLOCK * const pMBs,
2149 :     SearchData * const Data)
2150 : chl 96 {
2151 :    
2152 : edgomez 851 int i, mask;
2153 : edgomez 953 int quarterpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0;
2154 : edgomez 851 VECTOR pmv[3];
2155 : edgomez 953 MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2156 : chl 96
2157 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2158 : chl 96
2159 : edgomez 851 //median is only used as prediction. it doesn't have to be real
2160 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2161 :     else
2162 :     if (x == 1) //left macroblock does not have any vector now
2163 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2164 :     else if (y == 1) // top macroblock doesn't have it's vector
2165 :     Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
2166 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2167 : edgomez 195
2168 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2169 : edgomez 953 pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0);
2170 : edgomez 195
2171 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2172 :     Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
2173 : chl 96
2174 : edgomez 851 pmv[1].x = EVEN(pMB->mvs[0].x);
2175 :     pmv[1].y = EVEN(pMB->mvs[0].y);
2176 :     pmv[2].x = EVEN(Data->predMV.x);
2177 :     pmv[2].y = EVEN(Data->predMV.y);
2178 :     pmv[0].x = pmv[0].y = 0;
2179 : edgomez 195
2180 : edgomez 851 CheckCandidate32I(0, 0, 255, &i, Data);
2181 : edgomez 953 Data->iMinSAD[1] -= 50;
2182 :     Data->iMinSAD[2] -= 50;
2183 :     Data->iMinSAD[3] -= 50;
2184 :     Data->iMinSAD[4] -= 50;
2185 : edgomez 195
2186 : edgomez 953 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) {
2187 : chl 96
2188 : edgomez 851 if (!(mask = make_mask(pmv, 1)))
2189 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2190 :     if (!(mask = make_mask(pmv, 2)))
2191 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2192 : chl 96
2193 : edgomez 953 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) // diamond only if needed
2194 : edgomez 851 DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2195 : edgomez 953 }
2196 : edgomez 851
2197 : edgomez 953 for (i = 0; i < 4; i++) {
2198 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2199 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2200 :     MB->mode = MODE_INTER;
2201 :     MB->sad16 = Data->iMinSAD[i+1];
2202 : chl 96 }
2203 : edgomez 851 }
2204 : chl 96
2205 : edgomez 953 #define INTRA_THRESH 2400
2206 :     #define INTER_THRESH 1100
2207 : chl 96
2208 : edgomez 851 int
2209 :     MEanalysis( const IMAGE * const pRef,
2210 : edgomez 953 const FRAMEINFO * const Current,
2211 :     const MBParam * const pParam,
2212 :     const int maxIntra, //maximum number if non-I frames
2213 :     const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2214 :     const int bCount, // number of B frames in a row
2215 :     const int b_thresh)
2216 : edgomez 851 {
2217 :     uint32_t x, y, intra = 0;
2218 :     int sSAD = 0;
2219 :     MACROBLOCK * const pMBs = Current->mbs;
2220 :     const IMAGE * const pCurrent = &Current->image;
2221 : edgomez 953 int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh;
2222 :     int s = 0, blocks = 0;
2223 : chl 96
2224 : edgomez 851 int32_t iMinSAD[5], temp[5];
2225 :     VECTOR currentMV[5];
2226 :     SearchData Data;
2227 :     Data.iEdgedWidth = pParam->edged_width;
2228 :     Data.currentMV = currentMV;
2229 :     Data.iMinSAD = iMinSAD;
2230 :     Data.iFcode = Current->fcode;
2231 :     Data.temp = temp;
2232 :     CheckCandidate = CheckCandidate32I;
2233 : chl 96
2234 : edgomez 851 if (intraCount != 0 && intraCount < 10) // we're right after an I frame
2235 : edgomez 953 IntraThresh += 8 * (intraCount - 10) * (intraCount - 10);
2236 : edgomez 851 else
2237 :     if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2238 :     IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
2239 : chl 96
2240 : edgomez 953 InterThresh -= (350 - 8*b_thresh) * bCount;
2241 :     if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh;
2242 : chl 326
2243 : edgomez 851 if (sadInit) (*sadInit) ();
2244 : chl 96
2245 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2246 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2247 :     int i;
2248 : edgomez 953 blocks += 4;
2249 : edgomez 195
2250 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2251 : edgomez 953 else { //extrapolation of the vector found for last frame
2252 :     pMBs[x + y * pParam->mb_width].mvs[0].x =
2253 :     (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2254 :     pMBs[x + y * pParam->mb_width].mvs[0].y =
2255 :     (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2256 :     }
2257 : edgomez 195
2258 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2259 : edgomez 195
2260 : edgomez 851 for (i = 0; i < 4; i++) {
2261 :     int dev;
2262 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2263 :     if (pMB->sad16 > IntraThresh) {
2264 :     dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2265 :     pParam->edged_width);
2266 :     if (dev + IntraThresh < pMB->sad16) {
2267 :     pMB->mode = MODE_INTRA;
2268 : edgomez 953 if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2269 : edgomez 851 }
2270 :     }
2271 : edgomez 953 if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++;
2272 :    
2273 : edgomez 851 sSAD += pMB->sad16;
2274 :     }
2275 :     }
2276 : edgomez 195 }
2277 : edgomez 953
2278 :     sSAD /= blocks;
2279 :     s = (10*s) / blocks;
2280 :    
2281 :     if (s > 5) sSAD += (s - 4) * (180 - 2*b_thresh); //static block - looks bad when in bframe...
2282 :    
2283 : edgomez 851 if (sSAD > InterThresh ) return P_VOP;
2284 :     emms();
2285 :     return B_VOP;
2286 :     }
2287 : chl 96
2288 :    
2289 : edgomez 851 static WARPPOINTS
2290 :     GlobalMotionEst(const MACROBLOCK * const pMBs,
2291 :     const MBParam * const pParam,
2292 :     const FRAMEINFO * const current,
2293 :     const FRAMEINFO * const reference,
2294 :     const IMAGE * const pRefH,
2295 :     const IMAGE * const pRefV,
2296 :     const IMAGE * const pRefHV )
2297 :     {
2298 : chl 96
2299 : edgomez 851 const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2300 :     const int deltay=8;
2301 :     const int grad=512; // lower bound for deviation in MB
2302 : chl 96
2303 : edgomez 851 WARPPOINTS gmc;
2304 : chl 96
2305 : edgomez 851 uint32_t mx, my;
2306 : chl 96
2307 : edgomez 851 int MBh = pParam->mb_height;
2308 :     int MBw = pParam->mb_width;
2309 : chl 96
2310 : edgomez 851 int *MBmask= calloc(MBh*MBw,sizeof(int));
2311 :     double DtimesF[4] = { 0.,0., 0., 0. };
2312 :     double sol[4] = { 0., 0., 0., 0. };
2313 :     double a,b,c,n,denom;
2314 :     double meanx,meany;
2315 :     int num,oldnum;
2316 : edgomez 195
2317 : chl 871 if (!MBmask) { fprintf(stderr,"Mem error\n");
2318 :     gmc.duv[0].x= gmc.duv[0].y =
2319 :     gmc.duv[1].x= gmc.duv[1].y =
2320 :     gmc.duv[2].x= gmc.duv[2].y = 0;
2321 :     return gmc; }
2322 : edgomez 195
2323 : edgomez 851 // filter mask of all blocks
2324 :    
2325 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2326 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2327 : edgomez 851 {
2328 :     const int mbnum = mx + my * MBw;
2329 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2330 :     const VECTOR mv = pMB->mvs[0];
2331 :    
2332 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2333 :     continue;
2334 :    
2335 :     if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2336 :     && ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2337 :     && ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2338 :     && ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2339 :     MBmask[mbnum]=1;
2340 : chl 96 }
2341 :    
2342 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2343 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2344 : edgomez 851 {
2345 :     const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx;
2346 : chl 96
2347 : edgomez 851 const int mbnum = mx + my * MBw;
2348 :     if (!MBmask[mbnum])
2349 :     continue;
2350 :    
2351 : edgomez 876 if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad )
2352 : edgomez 851 MBmask[mbnum] = 0;
2353 : edgomez 876 if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad )
2354 : edgomez 851 MBmask[mbnum] = 0;
2355 :    
2356 : edgomez 195 }
2357 : chl 96
2358 : edgomez 851 emms();
2359 : chl 96
2360 : edgomez 851 do { /* until convergence */
2361 : chl 96
2362 : edgomez 851 a = b = c = n = 0;
2363 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2364 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2365 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2366 : edgomez 851 {
2367 :     const int mbnum = mx + my * MBw;
2368 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2369 :     const VECTOR mv = pMB->mvs[0];
2370 : chl 96
2371 : edgomez 851 if (!MBmask[mbnum])
2372 :     continue;
2373 : chl 96
2374 : edgomez 851 n++;
2375 :     a += 16*mx+8;
2376 :     b += 16*my+8;
2377 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2378 : chl 96
2379 : edgomez 851 DtimesF[0] += (double)mv.x;
2380 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2381 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2382 :     DtimesF[3] += (double)mv.y;
2383 :     }
2384 : chl 96
2385 : edgomez 851 denom = a*a+b*b-c*n;
2386 : chl 96
2387 : edgomez 851 /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2388 :     /* D'*E*F has been calculated in the same loop as matrix */
2389 : chl 96
2390 : edgomez 851 sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2391 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2392 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2393 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2394 : chl 96
2395 : edgomez 851 sol[0] /= denom;
2396 :     sol[1] /= denom;
2397 :     sol[2] /= denom;
2398 :     sol[3] /= denom;
2399 : chl 96
2400 : edgomez 851 meanx = meany = 0.;
2401 :     oldnum = 0;
2402 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2403 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2404 : edgomez 851 {
2405 :     const int mbnum = mx + my * MBw;
2406 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2407 :     const VECTOR mv = pMB->mvs[0];
2408 : chl 96
2409 : edgomez 851 if (!MBmask[mbnum])
2410 :     continue;
2411 : chl 184
2412 : edgomez 851 oldnum++;
2413 :     meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x );
2414 :     meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y );
2415 :     }
2416 : chl 96
2417 : edgomez 851 if (4*meanx > oldnum) /* better fit than 0.25 is useless */
2418 :     meanx /= oldnum;
2419 :     else
2420 :     meanx = 0.25;
2421 : edgomez 195
2422 : edgomez 851 if (4*meany > oldnum)
2423 :     meany /= oldnum;
2424 :     else
2425 :     meany = 0.25;
2426 : chl 96
2427 : edgomez 851 /* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2428 :     fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2429 :     */
2430 :     num = 0;
2431 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2432 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2433 : edgomez 851 {
2434 :     const int mbnum = mx + my * MBw;
2435 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2436 :     const VECTOR mv = pMB->mvs[0];
2437 : chl 96
2438 : edgomez 851 if (!MBmask[mbnum])
2439 :     continue;
2440 : chl 96
2441 : edgomez 851 if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx )
2442 :     || ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) )
2443 :     MBmask[mbnum]=0;
2444 :     else
2445 :     num++;
2446 : chl 96 }
2447 : edgomez 195
2448 : edgomez 851 } while ( (oldnum != num) && (num>=4) );
2449 : edgomez 195
2450 : edgomez 851 if (num < 4)
2451 :     {
2452 :     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;
2453 :     } else {
2454 : edgomez 195
2455 : edgomez 851 gmc.duv[0].x=(int)(sol[0]+0.5);
2456 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2457 : chl 96
2458 : edgomez 851 gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2459 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2460 : chl 96
2461 : edgomez 851 gmc.duv[2].x=0;
2462 :     gmc.duv[2].y=0;
2463 :     }
2464 :     // 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);
2465 : chl 96
2466 : edgomez 851 free(MBmask);
2467 : chl 96
2468 : edgomez 851 return gmc;
2469 : chl 96 }
2470 :    
2471 : edgomez 851 // functions which perform BITS-based search/bitcount
2472 : chl 96
2473 : edgomez 851 static int
2474 :     CountMBBitsInter(SearchData * const Data,
2475 :     const MACROBLOCK * const pMBs, const int x, const int y,
2476 :     const MBParam * const pParam,
2477 :     const uint32_t MotionFlags)
2478 : chl 96 {
2479 : edgomez 851 int i, iDirection;
2480 :     int32_t bsad[5];
2481 : chl 141
2482 : edgomez 851 CheckCandidate = CheckCandidateBits16;
2483 : chl 96
2484 : edgomez 851 if (Data->qpel) {
2485 :     for(i = 0; i < 5; i++) {
2486 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2487 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2488 :     }
2489 :     Data->qpel_precision = 1;
2490 :     CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2491 : chl 96
2492 : edgomez 851 //checking if this vector is perfect. if it is, we stop.
2493 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0)
2494 :     return 0; //quick stop
2495 : edgomez 195
2496 : edgomez 949 if (MotionFlags & (XVID_ME_HALFPELREFINE16_BITS | XVID_ME_EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search
2497 : edgomez 851 for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2498 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2499 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2500 :     Data->qpel_precision = 0;
2501 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2502 :     CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2503 :     }
2504 : edgomez 195
2505 : edgomez 851 } else { // not qpel
2506 : edgomez 195
2507 : edgomez 851 CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2508 :     //checking if this vector is perfect. if it is, we stop.
2509 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) {
2510 :     return 0; //inter
2511 :     }
2512 :     }
2513 : chl 96
2514 : edgomez 949 if (MotionFlags&XVID_ME_EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2515 : chl 96
2516 : edgomez 949 if (MotionFlags&XVID_ME_HALFPELREFINE16_BITS) SubpelRefine(Data);
2517 : chl 96
2518 : edgomez 851 if (Data->qpel) {
2519 : edgomez 949 if (MotionFlags&(XVID_ME_EXTSEARCH_BITS | XVID_ME_HALFPELREFINE16_BITS)) { // there was halfpel-precision search
2520 : edgomez 851 for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2521 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match
2522 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2523 :     }
2524 : edgomez 195
2525 : edgomez 851 // preparing for qpel-precision search
2526 :     Data->qpel_precision = 1;
2527 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2528 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
2529 :     }
2530 : edgomez 949 if (MotionFlags&XVID_ME_QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2531 : edgomez 851 }
2532 : chl 96
2533 : edgomez 949 if (MotionFlags&XVID_ME_CHECKPREDICTION_BITS) { //let's check vector equal to prediction
2534 : edgomez 851 VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2535 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2536 :     CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2537 :     }
2538 :     return Data->iMinSAD[0];
2539 :     }
2540 : chl 96
2541 :    
2542 : edgomez 851 static int
2543 :     CountMBBitsInter4v(const SearchData * const Data,
2544 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2545 :     const int x, const int y,
2546 :     const MBParam * const pParam, const uint32_t MotionFlags,
2547 :     const VECTOR * const backup)
2548 :     {
2549 : chl 96
2550 : edgomez 851 int cbp = 0, bits = 0, t = 0, i, iDirection;
2551 :     SearchData Data2, *Data8 = &Data2;
2552 :     int sumx = 0, sumy = 0;
2553 : edgomez 962 int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64;
2554 : chl 96
2555 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2556 :     CheckCandidate = CheckCandidateBits8;
2557 : chl 96
2558 : edgomez 851 for (i = 0; i < 4; i++) {
2559 :     Data8->iMinSAD = Data->iMinSAD + i + 1;
2560 :     Data8->currentMV = Data->currentMV + i + 1;
2561 :     Data8->currentQMV = Data->currentQMV + i + 1;
2562 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2563 :     Data8->Ref = Data->Ref + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2564 :     Data8->RefH = Data->RefH + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2565 :     Data8->RefV = Data->RefV + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2566 :     Data8->RefHV = Data->RefHV + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2567 : edgomez 195
2568 : edgomez 851 if(Data->qpel) {
2569 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2570 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2571 :     Data8->predMV, Data8->iFcode, 0, 0);
2572 :     } else {
2573 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2574 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2575 :     Data8->predMV, Data8->iFcode, 0, 0);
2576 :     }
2577 : chl 96
2578 : 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,
2579 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2580 : chl 96
2581 : edgomez 851 *Data8->iMinSAD += t;
2582 : chl 96
2583 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2584 :     // checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far)
2585 :     if (Data8->qpel) {
2586 :     if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y))
2587 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2588 :     } else {
2589 :     if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y))
2590 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2591 :     }
2592 : edgomez 195
2593 : edgomez 851 if (Data8->qpel) {
2594 : edgomez 949 if (MotionFlags&XVID_ME_HALFPELREFINE8_BITS || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_BITS)) { // halfpixel motion search follows
2595 : edgomez 851 int32_t s = *Data8->iMinSAD;
2596 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2597 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2598 :     Data8->qpel_precision = 0;
2599 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2600 :     pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2601 : edgomez 195
2602 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2603 :     CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2604 : edgomez 195
2605 : edgomez 949 if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS)
2606 : edgomez 851 SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2607 : chl 96
2608 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8);
2609 : chl 96
2610 : edgomez 851 if(s > *Data8->iMinSAD) { //we have found a better match
2611 :     Data8->currentQMV->x = 2*Data8->currentMV->x;
2612 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2613 :     }
2614 : chl 96
2615 : edgomez 851 Data8->qpel_precision = 1;
2616 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2617 :     pParam->width, pParam->height, Data8->iFcode, 1, 0);
2618 : chl 96
2619 : edgomez 851 }
2620 : edgomez 949 if (MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2621 : edgomez 195
2622 : edgomez 851 } else // not qpel
2623 : edgomez 949 if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement
2624 : edgomez 851
2625 :     //checking vector equal to predicion
2626 : edgomez 949 if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_BITS) {
2627 : edgomez 851 const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2628 :     if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y))
2629 :     CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2630 : chl 141 }
2631 : edgomez 195
2632 : edgomez 851 bits += *Data8->iMinSAD;
2633 :     if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V
2634 :    
2635 :     // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else
2636 :     if(Data->qpel) {
2637 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2638 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2639 :     pMB->qmvs[i] = *Data8->currentQMV;
2640 :     sumx += Data8->currentQMV->x/2;
2641 :     sumy += Data8->currentQMV->y/2;
2642 :     } else {
2643 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2644 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2645 :     sumx += Data8->currentMV->x;
2646 :     sumy += Data8->currentMV->y;
2647 : chl 141 }
2648 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2649 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2650 :     if (Data8->temp[0]) cbp |= 1 << (5 - i);
2651 : chl 141 }
2652 :    
2653 : edgomez 851 if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma
2654 :     const uint8_t * ptr;
2655 :     sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2656 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2657 : chl 141
2658 : edgomez 851 //chroma U
2659 :     ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCU, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2660 :     transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2661 :     fdct(in);
2662 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2663 :     else i = quant4_inter(coeff, in, Data->lambda16);
2664 :     if (i > 0) {
2665 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2666 :     cbp |= 1 << (5 - 4);
2667 :     }
2668 : chl 141
2669 : edgomez 851 if (bits < *Data->iMinSAD) { // still possible
2670 :     //chroma V
2671 :     ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCV, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2672 :     transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2673 :     fdct(in);
2674 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2675 :     else i = quant4_inter(coeff, in, Data->lambda16);
2676 :     if (i > 0) {
2677 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2678 :     cbp |= 1 << (5 - 5);
2679 :     }
2680 : edgomez 958 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
2681 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2682 :     }
2683 : edgomez 195 }
2684 : chl 96
2685 : edgomez 851 return bits;
2686 :     }
2687 : chl 96
2688 :    
2689 : edgomez 851 static int
2690 :     CountMBBitsIntra(const SearchData * const Data)
2691 :     {
2692 :     int bits = 1; //this one is ac/dc prediction flag. always 1.
2693 :     int cbp = 0, i, t, dc = 0, b_dc = 1024;
2694 :     const uint32_t iQuant = Data->lambda16;
2695 : edgomez 962 int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2696 : edgomez 195
2697 : edgomez 851 for(i = 0; i < 4; i++) {
2698 :     uint32_t iDcScaler = get_dc_scaler(iQuant, 1);
2699 : chl 96
2700 : edgomez 851 int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2701 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2702 :     fdct(in);
2703 :     b_dc = dc;
2704 :     dc = in[0];
2705 :     in[0] -= b_dc;
2706 :     if (Data->lambda8 == 0) quant_intra_c(coeff, in, iQuant, iDcScaler);
2707 :     else quant4_intra_c(coeff, in, iQuant, iDcScaler);
2708 : chl 141
2709 : edgomez 851 b_dc = dc;
2710 :     dc = coeff[0];
2711 :     if (i != 0) coeff[0] -= b_dc;
2712 : chl 181
2713 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;;
2714 :     Data->temp[i] = t;
2715 :     if (t != 0) cbp |= 1 << (5 - i);
2716 :     if (bits >= Data->iMinSAD[0]) break;
2717 :     }
2718 : chl 181
2719 : edgomez 851 if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma
2720 :     uint32_t iDcScaler = get_dc_scaler(iQuant, 0);
2721 :     //chroma U
2722 :     transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2723 :     fdct(in);
2724 :     in[0] -= 1024;
2725 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2726 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2727 : chl 96
2728 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2729 :     if (t != 0) cbp |= 1 << (5 - 4);
2730 : chl 141
2731 : edgomez 851 if (bits < Data->iMinSAD[0]) {
2732 : edgomez 953 iDcScaler = get_dc_scaler(iQuant, 1);
2733 : edgomez 851 //chroma V
2734 :     transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2735 :     fdct(in);
2736 :     in[0] -= 1024;
2737 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2738 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2739 : chl 96
2740 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2741 :     if (t != 0) cbp |= 1 << (5 - 5);
2742 : chl 96
2743 : edgomez 958 bits += xvid_cbpy_tab[cbp>>2].len;
2744 : edgomez 953 bits += mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2745 : chl 96 }
2746 :     }
2747 : edgomez 851 return bits;
2748 : chl 96 }

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