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

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

Parent Directory Parent Directory | Revision Log Revision Log


Revision 973 - (view) (download)

1 : edgomez 851 /**************************************************************************
2 : Isibaar 3 *
3 : edgomez 851 * XVID MPEG-4 VIDEO CODEC
4 :     * motion estimation
5 : chl 259 *
6 : edgomez 851 * This program is an implementation of a part of one or more MPEG-4
7 :     * Video tools as specified in ISO/IEC 14496-2 standard. Those intending
8 :     * to use this software module in hardware or software products are
9 :     * advised that its use may infringe existing patents or copyrights, and
10 :     * any such use would be at such party's own risk. The original
11 :     * developer of this software module and his/her company, and subsequent
12 :     * editors and their companies, will have no liability for use of this
13 :     * software or modifications or derivatives thereof.
14 : chl 259 *
15 : edgomez 851 * This program is free software; you can redistribute it and/or modify
16 :     * it under the terms of the GNU General Public License as published by
17 :     * the Free Software Foundation; either version 2 of the License, or
18 :     * (at your option) any later version.
19 : chl 259 *
20 : edgomez 851 * This program is distributed in the hope that it will be useful,
21 :     * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 :     * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 :     * GNU General Public License for more details.
24 : chl 259 *
25 : edgomez 851 * You should have received a copy of the GNU General Public License
26 :     * along with this program; if not, write to the Free Software
27 :     * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 : chl 259 *
29 :     *************************************************************************/
30 :    
31 : Isibaar 3 #include <assert.h>
32 :     #include <stdio.h>
33 : chl 96 #include <stdlib.h>
34 : edgomez 851 #include <string.h> // memcpy
35 :     #include <math.h> // lrint
36 : Isibaar 3
37 :     #include "../encoder.h"
38 :     #include "../utils/mbfunctions.h"
39 :     #include "../prediction/mbprediction.h"
40 :     #include "../global.h"
41 :     #include "../utils/timer.h"
42 : edgomez 851 #include "../image/interpolate8x8.h"
43 :     #include "motion_est.h"
44 : suxen_drol 118 #include "motion.h"
45 : Isibaar 3 #include "sad.h"
46 : edgomez 851 #include "../utils/emms.h"
47 :     #include "../dct/fdct.h"
48 : Isibaar 3
49 : edgomez 872 /*****************************************************************************
50 :     * Modified rounding tables -- declared in motion.h
51 :     * Original tables see ISO spec tables 7-6 -> 7-9
52 :     ****************************************************************************/
53 :    
54 :     const uint32_t roundtab[16] =
55 :     {0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2 };
56 :    
57 :     /* K = 4 */
58 :     const uint32_t roundtab_76[16] =
59 :     { 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1 };
60 :    
61 :     /* K = 2 */
62 :     const uint32_t roundtab_78[8] =
63 :     { 0, 0, 1, 1, 0, 0, 0, 1 };
64 :    
65 :     /* K = 1 */
66 :     const uint32_t roundtab_79[4] =
67 :     { 0, 1, 0, 0 };
68 :    
69 : edgomez 851 #define INITIAL_SKIP_THRESH (10)
70 :     #define FINAL_SKIP_THRESH (50)
71 :     #define MAX_SAD00_FOR_SKIP (20)
72 :     #define MAX_CHROMA_SAD_FOR_SKIP (22)
73 : Isibaar 3
74 : edgomez 851 #define CHECK_CANDIDATE(X,Y,D) { \
75 :     CheckCandidate((X),(Y), (D), &iDirection, data ); }
76 : Isibaar 3
77 : edgomez 872 /*****************************************************************************
78 :     * Code
79 :     ****************************************************************************/
80 :    
81 : edgomez 851 static __inline uint32_t
82 :     d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
83 :     {
84 : syskin 973 int bits;
85 :     const int q = (1 << (iFcode - 1)) - 1;
86 :    
87 :     x <<= qpel;
88 :     y <<= qpel;
89 : edgomez 851 if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
90 : chl 141
91 : edgomez 851 x -= pred.x;
92 : syskin 973 bits = (x != 0 ? iFcode:0);
93 :     x = abs(x);
94 :     x += q;
95 :     x >>= (iFcode - 1);
96 :     bits += mvtab[x];
97 :    
98 : edgomez 851 y -= pred.y;
99 : syskin 973 bits += (y != 0 ? iFcode:0);
100 :     y = abs(y);
101 :     y += q;
102 :     y >>= (iFcode - 1);
103 :     bits += mvtab[y];
104 : chl 141
105 : syskin 973 return bits;
106 : edgomez 851 }
107 : chl 141
108 : edgomez 851 static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data)
109 :     {
110 :     int sad;
111 :     const uint32_t stride = data->iEdgedWidth/2;
112 :     uint8_t * f_refu = data->RefQ,
113 :     * f_refv = data->RefQ + 8,
114 :     * b_refu = data->RefQ + 16,
115 :     * b_refv = data->RefQ + 24;
116 : Isibaar 3
117 : edgomez 851 switch (((fx & 1) << 1) | (fy & 1)) {
118 :     case 0:
119 :     fx = fx / 2; fy = fy / 2;
120 :     f_refu = (uint8_t*)data->RefCU + fy * stride + fx, stride;
121 :     f_refv = (uint8_t*)data->RefCV + fy * stride + fx, stride;
122 :     break;
123 :     case 1:
124 :     fx = fx / 2; fy = (fy - 1) / 2;
125 :     interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
126 :     interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
127 :     break;
128 :     case 2:
129 :     fx = (fx - 1) / 2; fy = fy / 2;
130 :     interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
131 :     interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
132 :     break;
133 :     default:
134 :     fx = (fx - 1) / 2; fy = (fy - 1) / 2;
135 :     interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
136 :     interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
137 :     break;
138 :     }
139 : Isibaar 3
140 : edgomez 851 switch (((bx & 1) << 1) | (by & 1)) {
141 :     case 0:
142 :     bx = bx / 2; by = by / 2;
143 :     b_refu = (uint8_t*)data->b_RefCU + by * stride + bx, stride;
144 :     b_refv = (uint8_t*)data->b_RefCV + by * stride + bx, stride;
145 :     break;
146 :     case 1:
147 :     bx = bx / 2; by = (by - 1) / 2;
148 :     interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
149 :     interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
150 :     break;
151 :     case 2:
152 :     bx = (bx - 1) / 2; by = by / 2;
153 :     interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
154 :     interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
155 :     break;
156 :     default:
157 :     bx = (bx - 1) / 2; by = (by - 1) / 2;
158 :     interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
159 :     interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
160 :     break;
161 :     }
162 : Isibaar 3
163 : edgomez 851 sad = sad8bi(data->CurU, b_refu, f_refu, stride);
164 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
165 : Isibaar 3
166 : edgomez 851 return sad;
167 :     }
168 : Isibaar 3
169 :    
170 : edgomez 851 static int32_t
171 :     ChromaSAD(int dx, int dy, const SearchData * const data)
172 :     {
173 :     int sad;
174 :     const uint32_t stride = data->iEdgedWidth/2;
175 : Isibaar 3
176 : edgomez 851 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently
177 :     data->temp[5] = dx; data->temp[6] = dy; // backup
178 : Isibaar 3
179 : edgomez 851 switch (((dx & 1) << 1) | (dy & 1)) {
180 :     case 0:
181 :     dx = dx / 2; dy = dy / 2;
182 :     sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride);
183 :     sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride);
184 :     break;
185 :     case 1:
186 :     dx = dx / 2; dy = (dy - 1) / 2;
187 :     sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride);
188 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride);
189 :     break;
190 :     case 2:
191 :     dx = (dx - 1) / 2; dy = dy / 2;
192 :     sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride);
193 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride);
194 :     break;
195 :     default:
196 :     dx = (dx - 1) / 2; dy = (dy - 1) / 2;
197 :     interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding);
198 :     sad = sad8(data->CurU, data->RefQ, stride);
199 :    
200 :     interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding);
201 :     sad += sad8(data->CurV, data->RefQ, stride);
202 :     break;
203 :     }
204 :     data->temp[7] = sad; //backup, part 2
205 :     return sad;
206 : Isibaar 3 }
207 :    
208 : edgomez 851 static __inline const uint8_t *
209 :     GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
210 : Isibaar 3 {
211 : edgomez 851 // dir : 0 = forward, 1 = backward
212 :     switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) {
213 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
214 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
215 :     case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
216 :     case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
217 :     case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth);
218 :     case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
219 :     case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
220 :     default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
221 :     }
222 : Isibaar 3 }
223 :    
224 : edgomez 851 // this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate
225 :     static __inline const uint8_t *
226 :     GetReference(const int x, const int y, const SearchData * const data)
227 : Isibaar 3 {
228 : edgomez 851 switch ( ((x&1)<<1) | (y&1) ) {
229 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
230 :     case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
231 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
232 :     default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2
233 :     }
234 : Isibaar 3 }
235 : edgomez 851
236 :     static uint8_t *
237 :     Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
238 : Isibaar 3 {
239 : edgomez 851 // create or find a qpel-precision reference picture; return pointer to it
240 :     uint8_t * Reference = data->RefQ + 16*dir;
241 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
242 :     const uint32_t rounding = data->rounding;
243 :     const int halfpel_x = x/2;
244 :     const int halfpel_y = y/2;
245 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
246 : suxen_drol 136
247 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
248 :     ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
249 :     switch( ((x&1)<<1) + (y&1) ) {
250 : edgomez 963 case 3: // x and y in qpel resolution - the "corners" (top left/right and
251 :     // bottom left/right) during qpel refinement
252 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
253 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
254 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
255 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
256 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
257 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
258 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
259 : edgomez 851 break;
260 : chl 184
261 : edgomez 851 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
262 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
263 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
264 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
265 :     break;
266 : suxen_drol 136
267 : edgomez 851 case 2: // x qpel, y halfpel - left or right during qpel refinement
268 :     ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
269 :     ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
270 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
271 :     break;
272 : edgomez 195
273 : edgomez 963 default: // pure halfpel position
274 :     return (uint8_t *) ref1;
275 :    
276 : edgomez 851 }
277 :     return Reference;
278 :     }
279 : Isibaar 3
280 : edgomez 851 static uint8_t *
281 :     Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
282 :     {
283 :     // create or find a qpel-precision reference picture; return pointer to it
284 :     uint8_t * Reference = data->RefQ + 16*dir;
285 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
286 :     const uint32_t rounding = data->rounding;
287 :     const int halfpel_x = x/2;
288 :     const int halfpel_y = y/2;
289 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
290 : chl 326
291 : edgomez 851 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
292 :     switch( ((x&1)<<1) + (y&1) ) {
293 :     case 3: // x and y in qpel resolution - the "corners" (top left/right and
294 :     // bottom left/right) during qpel refinement
295 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
296 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
297 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
298 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
299 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
300 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
301 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
302 :     break;
303 : chl 184
304 : edgomez 851 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
305 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
306 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
307 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
308 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
309 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
310 :     break;
311 : chl 184
312 : edgomez 851 case 2: // x qpel, y halfpel - left or right during qpel refinement
313 :     ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
314 :     interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
315 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
316 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
317 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
318 :     break;
319 : edgomez 195
320 : edgomez 963 default: // pure halfpel position
321 : edgomez 851 return (uint8_t *) ref1;
322 :     }
323 :     return Reference;
324 :     }
325 : edgomez 195
326 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS START */
327 : edgomez 195
328 : edgomez 851 static void
329 :     CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
330 :     {
331 :     int xc, yc;
332 :     const uint8_t * Reference;
333 :     VECTOR * current;
334 :     int32_t sad; uint32_t t;
335 : edgomez 195
336 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
337 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
338 : edgomez 195
339 : edgomez 851 if (!data->qpel_precision) {
340 :     Reference = GetReference(x, y, data);
341 :     current = data->currentMV;
342 :     xc = x; yc = y;
343 :     } else { // x and y are in 1/4 precision
344 :     Reference = Interpolate16x16qpel(x, y, 0, data);
345 :     xc = x/2; yc = y/2; //for chroma sad
346 :     current = data->currentQMV;
347 :     }
348 : edgomez 195
349 : edgomez 851 sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
350 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
351 : edgomez 195
352 : edgomez 851 sad += (data->lambda16 * t * sad)>>10;
353 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
354 : chl 175
355 : edgomez 851 if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
356 : edgomez 963 (yc >> 1) + roundtab_79[yc & 0x3], data);
357 : edgomez 195
358 : edgomez 851 if (sad < data->iMinSAD[0]) {
359 :     data->iMinSAD[0] = sad;
360 :     current[0].x = x; current[0].y = y;
361 :     *dir = Direction;
362 :     }
363 : edgomez 195
364 : edgomez 851 if (data->temp[1] < data->iMinSAD[1]) {
365 :     data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
366 :     if (data->temp[2] < data->iMinSAD[2]) {
367 :     data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
368 :     if (data->temp[3] < data->iMinSAD[3]) {
369 :     data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
370 :     if (data->temp[4] < data->iMinSAD[4]) {
371 :     data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
372 :    
373 : Isibaar 3 }
374 :    
375 : edgomez 851 static void
376 :     CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
377 :     {
378 :     int32_t sad; uint32_t t;
379 :     const uint8_t * Reference;
380 : syskin 973 VECTOR * current;
381 : chl 326
382 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
383 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
384 : Isibaar 3
385 : syskin 973 if (!data->qpel_precision) {
386 :     Reference = GetReference(x, y, data);
387 :     current = data->currentMV;
388 :     } else { // x and y are in 1/4 precision
389 :     Reference = Interpolate8x8qpel(x, y, 0, 0, data);
390 :     current = data->currentQMV;
391 :     }
392 : Isibaar 3
393 : edgomez 851 sad = sad8(data->Cur, Reference, data->iEdgedWidth);
394 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
395 : Isibaar 3
396 : edgomez 851 sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
397 : Isibaar 3
398 : edgomez 851 if (sad < *(data->iMinSAD)) {
399 :     *(data->iMinSAD) = sad;
400 : syskin 973 current->x = x; current->y = y;
401 : edgomez 851 *dir = Direction;
402 :     }
403 : Isibaar 3 }
404 :    
405 :    
406 : edgomez 851 static void
407 :     CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
408 :     {
409 :     uint32_t t;
410 :     const uint8_t * Reference;
411 :    
412 : edgomez 963 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero even value
413 : edgomez 851 (x > data->max_dx) || (x < data->min_dx)
414 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
415 :    
416 :     Reference = GetReference(x, y, data);
417 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
418 :    
419 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
420 :    
421 :     data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
422 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
423 :    
424 :     if (data->temp[0] < data->iMinSAD[0]) {
425 :     data->iMinSAD[0] = data->temp[0];
426 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
427 :     *dir = Direction; }
428 :    
429 :     if (data->temp[1] < data->iMinSAD[1]) {
430 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
431 :     if (data->temp[2] < data->iMinSAD[2]) {
432 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
433 :     if (data->temp[3] < data->iMinSAD[3]) {
434 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
435 :     if (data->temp[4] < data->iMinSAD[4]) {
436 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
437 : Isibaar 3 }
438 : edgomez 195
439 : edgomez 851 static void
440 :     CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
441 :     {
442 :     int32_t sad, xc, yc;
443 :     const uint8_t * Reference;
444 :     uint32_t t;
445 :     VECTOR * current;
446 :    
447 : edgomez 963 if ( (x > data->max_dx) || ( x < data->min_dx)
448 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
449 : edgomez 851
450 :     if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value
451 :    
452 :     if (data->qpel_precision) { // x and y are in 1/4 precision
453 :     Reference = Interpolate16x16qpel(x, y, 0, data);
454 :     current = data->currentQMV;
455 :     xc = x/2; yc = y/2;
456 :     } else {
457 :     Reference = GetReference(x, y, data);
458 :     current = data->currentMV;
459 :     xc = x; yc = y;
460 :     }
461 :     t = d_mv_bits(x, y, data->predMV, data->iFcode,
462 :     data->qpel^data->qpel_precision, data->rrv);
463 :    
464 :     sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
465 :     sad += (data->lambda16 * t * sad)>>10;
466 :    
467 :     if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
468 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
469 :    
470 :     if (sad < *(data->iMinSAD)) {
471 :     *(data->iMinSAD) = sad;
472 :     current->x = x; current->y = y;
473 :     *dir = Direction;
474 :     }
475 : chl 96 }
476 : Isibaar 3
477 : edgomez 851 static void
478 :     CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
479 :     {
480 :     // maximum speed - for P/B/I decision
481 :     int32_t sad;
482 :    
483 :     if ( (x > data->max_dx) || (x < data->min_dx)
484 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
485 :    
486 :     sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
487 :     data->iEdgedWidth, data->temp+1);
488 :    
489 :     if (sad < *(data->iMinSAD)) {
490 :     *(data->iMinSAD) = sad;
491 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
492 :     *dir = Direction;
493 :     }
494 :     if (data->temp[1] < data->iMinSAD[1]) {
495 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
496 :     if (data->temp[2] < data->iMinSAD[2]) {
497 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
498 :     if (data->temp[3] < data->iMinSAD[3]) {
499 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
500 :     if (data->temp[4] < data->iMinSAD[4]) {
501 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
502 :    
503 : Isibaar 3 }
504 :    
505 : edgomez 851 static void
506 :     CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
507 :     {
508 :     int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
509 :     uint32_t t;
510 :     const uint8_t *ReferenceF, *ReferenceB;
511 :     VECTOR *current;
512 :    
513 : edgomez 963 if ((xf > data->max_dx) || (xf < data->min_dx) ||
514 :     (yf > data->max_dy) || (yf < data->min_dy))
515 :     return;
516 : edgomez 851
517 :     if (!data->qpel_precision) {
518 :     ReferenceF = GetReference(xf, yf, data);
519 :     xb = data->currentMV[1].x; yb = data->currentMV[1].y;
520 :     ReferenceB = GetReferenceB(xb, yb, 1, data);
521 :     current = data->currentMV;
522 :     xcf = xf; ycf = yf;
523 :     xcb = xb; ycb = yb;
524 :     } else {
525 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
526 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
527 :     current = data->currentQMV;
528 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
529 :     xcf = xf/2; ycf = yf/2;
530 :     xcb = xb/2; ycb = yb/2;
531 :     }
532 :    
533 :     t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
534 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
535 :    
536 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
537 :     sad += (data->lambda16 * t * sad)>>10;
538 :    
539 :     if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
540 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
541 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
542 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
543 :    
544 :     if (sad < *(data->iMinSAD)) {
545 :     *(data->iMinSAD) = sad;
546 :     current->x = xf; current->y = yf;
547 :     *dir = Direction;
548 :     }
549 : Isibaar 3 }
550 :    
551 : edgomez 851 static void
552 :     CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
553 :     {
554 :     int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
555 :     uint32_t k;
556 :     const uint8_t *ReferenceF;
557 :     const uint8_t *ReferenceB;
558 :     VECTOR mvs, b_mvs;
559 :    
560 : syskin 904 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
561 : edgomez 851
562 :     for (k = 0; k < 4; k++) {
563 :     mvs.x = data->directmvF[k].x + x;
564 :     b_mvs.x = ((x == 0) ?
565 :     data->directmvB[k].x
566 :     : mvs.x - data->referencemv[k].x);
567 :    
568 :     mvs.y = data->directmvF[k].y + y;
569 :     b_mvs.y = ((y == 0) ?
570 :     data->directmvB[k].y
571 :     : mvs.y - data->referencemv[k].y);
572 :    
573 : edgomez 963 if ((mvs.x > data->max_dx) || (mvs.x < data->min_dx) ||
574 :     (mvs.y > data->max_dy) || (mvs.y < data->min_dy) ||
575 :     (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx) ||
576 :     (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) )
577 :     return;
578 : edgomez 851
579 :     if (data->qpel) {
580 :     xcf += mvs.x/2; ycf += mvs.y/2;
581 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
582 :     } else {
583 :     xcf += mvs.x; ycf += mvs.y;
584 :     xcb += b_mvs.x; ycb += b_mvs.y;
585 :     mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway
586 :     b_mvs.x *= 2; b_mvs.y *= 2;
587 :     }
588 :    
589 :     ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
590 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
591 :    
592 :     sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
593 :     ReferenceF, ReferenceB, data->iEdgedWidth);
594 :     if (sad > *(data->iMinSAD)) return;
595 :     }
596 :    
597 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
598 :    
599 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
600 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
601 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
602 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
603 :    
604 :     if (sad < *(data->iMinSAD)) {
605 :     *(data->iMinSAD) = sad;
606 :     data->currentMV->x = x; data->currentMV->y = y;
607 :     *dir = Direction;
608 :     }
609 : Isibaar 3 }
610 :    
611 : edgomez 851 static void
612 :     CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
613 : Isibaar 3 {
614 : edgomez 851 int32_t sad, xcf, ycf, xcb, ycb;
615 :     const uint8_t *ReferenceF;
616 :     const uint8_t *ReferenceB;
617 :     VECTOR mvs, b_mvs;
618 : Isibaar 3
619 : edgomez 963 if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return;
620 : Isibaar 3
621 : edgomez 851 mvs.x = data->directmvF[0].x + x;
622 :     b_mvs.x = ((x == 0) ?
623 :     data->directmvB[0].x
624 :     : mvs.x - data->referencemv[0].x);
625 : Isibaar 3
626 : edgomez 851 mvs.y = data->directmvF[0].y + y;
627 :     b_mvs.y = ((y == 0) ?
628 :     data->directmvB[0].y
629 :     : mvs.y - data->referencemv[0].y);
630 : Isibaar 3
631 : edgomez 963 if ( (mvs.x > data->max_dx) || (mvs.x < data->min_dx)
632 :     || (mvs.y > data->max_dy) || (mvs.y < data->min_dy)
633 :     || (b_mvs.x > data->max_dx) || (b_mvs.x < data->min_dx)
634 :     || (b_mvs.y > data->max_dy) || (b_mvs.y < data->min_dy) ) return;
635 : Isibaar 3
636 : edgomez 851 if (data->qpel) {
637 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
638 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
639 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
640 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
641 :     } else {
642 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
643 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
644 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
645 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
646 :     }
647 :    
648 :     sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
649 :     sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
650 :    
651 :     if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
652 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
653 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
654 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
655 :    
656 :     if (sad < *(data->iMinSAD)) {
657 :     *(data->iMinSAD) = sad;
658 :     data->currentMV->x = x; data->currentMV->y = y;
659 :     *dir = Direction;
660 :     }
661 : Isibaar 3 }
662 :    
663 : edgomez 851
664 :     static void
665 :     CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
666 : Isibaar 3 {
667 :    
668 : edgomez 961 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
669 : edgomez 851 int32_t bits = 0, sum;
670 :     VECTOR * current;
671 :     const uint8_t * ptr;
672 :     int i, cbp = 0, t, xc, yc;
673 : edgomez 195
674 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
675 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
676 : edgomez 195
677 : edgomez 851 if (!data->qpel_precision) {
678 :     ptr = GetReference(x, y, data);
679 :     current = data->currentMV;
680 :     xc = x; yc = y;
681 :     } else { // x and y are in 1/4 precision
682 :     ptr = Interpolate16x16qpel(x, y, 0, data);
683 :     current = data->currentQMV;
684 :     xc = x/2; yc = y/2;
685 :     }
686 : Isibaar 3
687 : edgomez 851 for(i = 0; i < 4; i++) {
688 :     int s = 8*((i&1) + (i>>1)*data->iEdgedWidth);
689 :     transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth);
690 :     fdct(in);
691 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
692 :     else sum = quant4_inter(coeff, in, data->lambda16);
693 :     if (sum > 0) {
694 :     cbp |= 1 << (5 - i);
695 :     bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
696 :     } else data->temp[i] = 0;
697 :     }
698 : Isibaar 3
699 : edgomez 851 bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
700 : edgomez 195
701 : edgomez 851 if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma
702 :     xc = (xc >> 1) + roundtab_79[xc & 0x3];
703 :     yc = (yc >> 1) + roundtab_79[yc & 0x3];
704 :    
705 :     //chroma U
706 :     ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCU, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
707 :     transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2);
708 :     fdct(in);
709 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
710 :     else sum = quant4_inter(coeff, in, data->lambda16);
711 :     if (sum > 0) {
712 :     cbp |= 1 << (5 - 4);
713 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
714 : chl 344 }
715 : edgomez 851
716 :     if (bits < data->iMinSAD[0]) {
717 :     //chroma V
718 :     ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCV, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding);
719 :     transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2);
720 :     fdct(in);
721 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
722 :     else sum = quant4_inter(coeff, in, data->lambda16);
723 :     if (sum > 0) {
724 :     cbp |= 1 << (5 - 5);
725 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
726 :     }
727 :     }
728 : edgomez 78 }
729 : edgomez 851
730 : chl 897 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
731 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len;
732 :    
733 :     if (bits < data->iMinSAD[0]) {
734 :     data->iMinSAD[0] = bits;
735 :     current[0].x = x; current[0].y = y;
736 :     *dir = Direction;
737 :     }
738 :    
739 :     if (data->temp[0] + t < data->iMinSAD[1]) {
740 :     data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; }
741 :     if (data->temp[1] < data->iMinSAD[2]) {
742 :     data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
743 :     if (data->temp[2] < data->iMinSAD[3]) {
744 :     data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
745 :     if (data->temp[3] < data->iMinSAD[4]) {
746 :     data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
747 :    
748 : Isibaar 3 }
749 : edgomez 851 static void
750 :     CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
751 : chl 96 {
752 :    
753 : edgomez 961 int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64;
754 : edgomez 851 int32_t sum, bits;
755 :     VECTOR * current;
756 :     const uint8_t * ptr;
757 :     int cbp;
758 : edgomez 195
759 : edgomez 851 if ( (x > data->max_dx) || (x < data->min_dx)
760 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
761 : edgomez 195
762 : edgomez 851 if (!data->qpel_precision) {
763 :     ptr = GetReference(x, y, data);
764 :     current = data->currentMV;
765 :     } else { // x and y are in 1/4 precision
766 :     ptr = Interpolate8x8qpel(x, y, 0, 0, data);
767 :     current = data->currentQMV;
768 :     }
769 : chl 96
770 : edgomez 851 transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth);
771 :     fdct(in);
772 :     if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16);
773 :     else sum = quant4_inter(coeff, in, data->lambda16);
774 :     if (sum > 0) {
775 :     bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
776 :     cbp = 1;
777 :     } else cbp = bits = 0;
778 : chl 96
779 : edgomez 851 bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
780 : chl 96
781 : edgomez 851 if (bits < data->iMinSAD[0]) {
782 :     data->temp[0] = cbp;
783 :     data->iMinSAD[0] = bits;
784 :     current[0].x = x; current[0].y = y;
785 :     *dir = Direction;
786 :     }
787 :     }
788 : chl 96
789 : edgomez 851 /* CHECK_CANDIATE FUNCTIONS END */
790 : edgomez 195
791 : edgomez 851 /* MAINSEARCH FUNCTIONS START */
792 : chl 96
793 : edgomez 851 static void
794 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
795 :     {
796 : edgomez 195
797 : edgomez 851 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
798 : edgomez 195
799 : edgomez 851 int iDirection;
800 : edgomez 195
801 : edgomez 851 for(;;) { //forever
802 :     iDirection = 0;
803 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
804 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
805 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
806 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
807 : edgomez 195
808 : edgomez 851 /* now we're doing diagonal checks near our candidate */
809 : edgomez 195
810 : edgomez 851 if (iDirection) { //if anything found
811 :     bDirection = iDirection;
812 :     iDirection = 0;
813 :     x = data->currentMV->x; y = data->currentMV->y;
814 :     if (bDirection & 3) { //our candidate is left or right
815 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
816 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
817 :     } else { // what remains here is up or down
818 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
819 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
820 :     }
821 : edgomez 195
822 : edgomez 851 if (iDirection) {
823 :     bDirection += iDirection;
824 :     x = data->currentMV->x; y = data->currentMV->y;
825 :     }
826 :     } else { //about to quit, eh? not so fast....
827 :     switch (bDirection) {
828 :     case 2:
829 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
830 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
831 : edgomez 195 break;
832 : edgomez 851 case 1:
833 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
834 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
835 : edgomez 195 break;
836 : edgomez 851 case 2 + 4:
837 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
838 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
839 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
840 :     break;
841 :     case 4:
842 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
843 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
844 :     break;
845 : edgomez 195 case 8:
846 : edgomez 851 CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
847 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
848 : edgomez 195 break;
849 : edgomez 851 case 1 + 4:
850 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
851 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
852 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
853 : edgomez 195 break;
854 : edgomez 851 case 2 + 8:
855 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
856 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
857 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
858 :     break;
859 :     case 1 + 8:
860 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
861 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
862 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
863 :     break;
864 :     default: //1+2+4+8 == we didn't find anything at all
865 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
866 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
867 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
868 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
869 :     break;
870 : chl 96 }
871 : edgomez 851 if (!iDirection) break; //ok, the end. really
872 :     bDirection = iDirection;
873 :     x = data->currentMV->x; y = data->currentMV->y;
874 : chl 345 }
875 : edgomez 195 }
876 : chl 96 }
877 :    
878 : edgomez 851 static void
879 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
880 : chl 96 {
881 : edgomez 851 int iDirection;
882 : edgomez 195
883 : edgomez 851 do {
884 :     iDirection = 0;
885 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
886 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
887 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
888 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
889 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
890 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
891 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
892 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
893 : chl 96
894 : edgomez 851 bDirection = iDirection;
895 :     x = data->currentMV->x; y = data->currentMV->y;
896 :     } while (iDirection);
897 : chl 96 }
898 :    
899 : edgomez 851 static void
900 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
901 : chl 181 {
902 :    
903 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
904 :    
905 : edgomez 851 int iDirection;
906 : edgomez 195
907 : edgomez 851 do {
908 :     iDirection = 0;
909 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
910 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
911 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
912 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
913 :    
914 :     /* now we're doing diagonal checks near our candidate */
915 :    
916 :     if (iDirection) { //checking if anything found
917 :     bDirection = iDirection;
918 : chl 181 iDirection = 0;
919 : edgomez 851 x = data->currentMV->x; y = data->currentMV->y;
920 :     if (bDirection & 3) { //our candidate is left or right
921 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
922 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
923 :     } else { // what remains here is up or down
924 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
925 :     CHECK_CANDIDATE(x - iDiamondSize, y, 1);
926 :     }
927 :     bDirection += iDirection;
928 :     x = data->currentMV->x; y = data->currentMV->y;
929 :     }
930 :     }
931 :     while (iDirection);
932 :     }
933 : chl 181
934 : edgomez 851 /* MAINSEARCH FUNCTIONS END */
935 : chl 181
936 : edgomez 851 static void
937 :     SubpelRefine(const SearchData * const data)
938 :     {
939 :     /* Do a half-pel or q-pel refinement */
940 :     const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
941 :     int iDirection; //only needed because macro expects it
942 : chl 181
943 : edgomez 851 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 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
948 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
949 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
950 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
951 :     }
952 : chl 181
953 : edgomez 851 static __inline int
954 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
955 :     const int x, const int y,
956 :     const uint32_t stride, const uint32_t iQuant, int rrv)
957 : chl 181
958 : edgomez 851 {
959 : edgomez 963 int offset = (x + y*stride)*8;
960 : edgomez 851 if(!rrv) {
961 : edgomez 963 uint32_t sadC = sad8(current->u + offset,
962 :     reference->u + offset, stride);
963 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
964 : edgomez 963 sadC += sad8(current->v + offset,
965 :     reference->v + offset, stride);
966 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
967 :     return 1;
968 : chl 181
969 : edgomez 851 } else {
970 : edgomez 963 uint32_t sadC = sad16(current->u + 2*offset,
971 :     reference->u + 2*offset, stride, 256*4096);
972 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
973 : edgomez 963 sadC += sad16(current->v + 2*offset,
974 :     reference->v + 2*offset, stride, 256*4096);
975 : edgomez 851 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
976 :     return 1;
977 : chl 181 }
978 : edgomez 851 }
979 : edgomez 504
980 : edgomez 851 static __inline void
981 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
982 :     {
983 :     pMB->mode = MODE_NOT_CODED;
984 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
985 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
986 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
987 : chl 181 }
988 :    
989 : edgomez 851 bool
990 :     MotionEstimation(MBParam * const pParam,
991 :     FRAMEINFO * const current,
992 :     FRAMEINFO * const reference,
993 :     const IMAGE * const pRefH,
994 :     const IMAGE * const pRefV,
995 :     const IMAGE * const pRefHV,
996 :     const uint32_t iLimit)
997 : chl 181 {
998 : edgomez 851 MACROBLOCK *const pMBs = current->mbs;
999 :     const IMAGE *const pCurrent = &current->image;
1000 :     const IMAGE *const pRef = &reference->image;
1001 : chl 181
1002 : edgomez 851 uint32_t mb_width = pParam->mb_width;
1003 :     uint32_t mb_height = pParam->mb_height;
1004 :     const uint32_t iEdgedWidth = pParam->edged_width;
1005 :     const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags);
1006 : chl 181
1007 : edgomez 851 uint32_t x, y;
1008 :     uint32_t iIntra = 0;
1009 :     int32_t quant = current->quant, sad00;
1010 : syskin 904 int skip_thresh = INITIAL_SKIP_THRESH *
1011 :     (current->global_flags & XVID_REDUCED ? 4:1) *
1012 :     (current->global_flags & XVID_MODEDECISION_BITS ? 2:1);
1013 : chl 181
1014 : edgomez 851 // some pre-initialized thingies for SearchP
1015 :     int32_t temp[8];
1016 :     VECTOR currentMV[5];
1017 :     VECTOR currentQMV[5];
1018 :     int32_t iMinSAD[5];
1019 : edgomez 961 DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE);
1020 : edgomez 851 SearchData Data;
1021 :     memset(&Data, 0, sizeof(SearchData));
1022 :     Data.iEdgedWidth = iEdgedWidth;
1023 :     Data.currentMV = currentMV;
1024 :     Data.currentQMV = currentQMV;
1025 :     Data.iMinSAD = iMinSAD;
1026 :     Data.temp = temp;
1027 :     Data.iFcode = current->fcode;
1028 :     Data.rounding = pParam->m_rounding_type;
1029 :     Data.qpel = pParam->m_quarterpel;
1030 :     Data.chroma = MotionFlags & PMV_CHROMA16;
1031 :     Data.rrv = current->global_flags & XVID_REDUCED;
1032 : edgomez 961 Data.dctSpace = dct_space;
1033 : edgomez 195
1034 : edgomez 851 if ((current->global_flags & XVID_REDUCED)) {
1035 :     mb_width = (pParam->width + 31) / 32;
1036 :     mb_height = (pParam->height + 31) / 32;
1037 :     Data.qpel = 0;
1038 :     }
1039 : chl 181
1040 : edgomez 851 Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose)
1041 :     if (sadInit) (*sadInit) ();
1042 : chl 181
1043 : edgomez 851 for (y = 0; y < mb_height; y++) {
1044 :     for (x = 0; x < mb_width; x++) {
1045 :     MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
1046 : chl 181
1047 : edgomez 851 if (!Data.rrv) pMB->sad16 =
1048 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
1049 :     pRef->y + (x + y * iEdgedWidth) * 16,
1050 :     pParam->edged_width, pMB->sad8 );
1051 : chl 181
1052 : edgomez 851 else pMB->sad16 =
1053 :     sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
1054 :     pRef->y + (x + y * iEdgedWidth) * 32,
1055 :     pParam->edged_width, pMB->sad8 );
1056 : chl 181
1057 : edgomez 851 if (Data.chroma) {
1058 :     Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
1059 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
1060 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
1061 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
1062 :     pMB->sad16 += Data.temp[7];
1063 :     }
1064 :    
1065 :     sad00 = pMB->sad16;
1066 :    
1067 :     if (!(current->global_flags & XVID_LUMIMASKING)) {
1068 :     pMB->dquant = NO_CHANGE;
1069 :     } else {
1070 :     if (pMB->dquant != NO_CHANGE) {
1071 :     quant += DQtab[pMB->dquant];
1072 :     if (quant > 31) quant = 31;
1073 :     else if (quant < 1) quant = 1;
1074 : chl 181 }
1075 : edgomez 851 }
1076 :     pMB->quant = current->quant;
1077 : chl 181
1078 : edgomez 851 //initial skip decision
1079 :     /* no early skip for GMC (global vector = skip vector is unknown!) */
1080 :     if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
1081 : syskin 904 if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * skip_thresh)
1082 : edgomez 851 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
1083 :     SkipMacroblockP(pMB, sad00);
1084 :     continue;
1085 :     }
1086 :     }
1087 :    
1088 :     SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
1089 :     y, MotionFlags, current->global_flags, pMB->quant,
1090 :     &Data, pParam, pMBs, reference->mbs,
1091 :     current->global_flags & XVID_INTER4V, pMB);
1092 :    
1093 :     /* final skip decision, a.k.a. "the vector you found, really that good?" */
1094 : syskin 904 if (!(current->global_flags & XVID_GMC || current->global_flags & XVID_MODEDECISION_BITS)) {
1095 : edgomez 851 if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) {
1096 : syskin 904 if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) )
1097 :     if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv))
1098 :     SkipMacroblockP(pMB, sad00);
1099 : chl 181 }
1100 :     }
1101 : edgomez 851 if (pMB->mode == MODE_INTRA)
1102 :     if (++iIntra > iLimit) return 1;
1103 : chl 181 }
1104 :     }
1105 : edgomez 851
1106 :     if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */
1107 :     {
1108 :     current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
1109 :     }
1110 :     return 0;
1111 : chl 181 }
1112 :    
1113 :    
1114 : edgomez 851 static __inline int
1115 :     make_mask(const VECTOR * const pmv, const int i)
1116 : chl 96 {
1117 : edgomez 851 int mask = 255, j;
1118 :     for (j = 0; j < i; j++) {
1119 :     if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
1120 :     if (pmv[i].x == pmv[j].x) {
1121 :     if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
1122 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
1123 :     } else
1124 :     if (pmv[i].y == pmv[j].y) {
1125 :     if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
1126 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
1127 :     }
1128 :     }
1129 :     return mask;
1130 :     }
1131 : edgomez 195
1132 : edgomez 851 static __inline void
1133 :     PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
1134 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1135 :     {
1136 : chl 96
1137 : edgomez 851 //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1138 :     if (rrv) { iWcount /= 2; iHcount /= 2; }
1139 : edgomez 195
1140 : edgomez 851 if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
1141 :     pmv[5].x = EVEN(pmv[3].x);
1142 :     pmv[5].y = EVEN(pmv[3].y);
1143 :     } else pmv[5].x = pmv[5].y = 0;
1144 : chl 96
1145 : edgomez 851 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1146 :     else pmv[3].x = pmv[3].y = 0;
1147 : chl 96
1148 : edgomez 851 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1149 :     else pmv[4].x = pmv[4].y = 0;
1150 : Isibaar 3
1151 : edgomez 851 // [1] median prediction
1152 :     pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1153 : edgomez 195
1154 : 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
1155 : edgomez 195
1156 : edgomez 851 pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1157 :     pmv[2].y = EVEN(prevMB->mvs[0].y);
1158 :    
1159 :     if ((x < iWcount-1) && (y < iHcount-1)) {
1160 :     pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1161 :     pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1162 :     } else pmv[6].x = pmv[6].y = 0;
1163 :    
1164 :     if (rrv) {
1165 :     int i;
1166 :     for (i = 0; i < 7; i++) {
1167 :     pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1168 :     pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1169 :     }
1170 :     }
1171 : Isibaar 3 }
1172 :    
1173 : edgomez 851 static int
1174 :     ModeDecision(const uint32_t iQuant, SearchData * const Data,
1175 :     int inter4v,
1176 :     MACROBLOCK * const pMB,
1177 :     const MACROBLOCK * const pMBs,
1178 :     const int x, const int y,
1179 :     const MBParam * const pParam,
1180 :     const uint32_t MotionFlags,
1181 :     const uint32_t GlobalFlags)
1182 :     {
1183 : Isibaar 3
1184 : edgomez 851 int mode = MODE_INTER;
1185 : chl 96
1186 : edgomez 851 if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision
1187 :     int sad;
1188 :     int InterBias = MV16_INTER_BIAS;
1189 :     if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1190 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
1191 : edgomez 963 mode = MODE_INTER;
1192 :     sad = Data->iMinSAD[0];
1193 : edgomez 851 } else {
1194 :     mode = MODE_INTER4V;
1195 :     sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
1196 :     Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
1197 :     Data->iMinSAD[0] = sad;
1198 :     }
1199 : chl 326
1200 : edgomez 851 /* intra decision */
1201 : Isibaar 3
1202 : edgomez 851 if (iQuant > 8) InterBias += 100 * (iQuant - 8); // to make high quants work
1203 :     if (y != 0)
1204 :     if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
1205 :     if (x != 0)
1206 :     if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
1207 : Isibaar 3
1208 : edgomez 851 if (Data->chroma) InterBias += 50; // to compensate bigger SAD
1209 :     if (Data->rrv) InterBias *= 4;
1210 : edgomez 195
1211 : edgomez 851 if (InterBias < pMB->sad16) {
1212 :     int32_t deviation;
1213 :     if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth);
1214 :     else deviation = dev16(Data->Cur, Data->iEdgedWidth) +
1215 :     dev16(Data->Cur+8, Data->iEdgedWidth) +
1216 :     dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) +
1217 :     dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth);
1218 : edgomez 195
1219 : edgomez 963 if (deviation < (sad - InterBias)) return MODE_INTRA;
1220 : edgomez 851 }
1221 :     return mode;
1222 : Isibaar 3
1223 : edgomez 851 } else {
1224 : edgomez 195
1225 : edgomez 851 int bits, intra, i;
1226 :     VECTOR backup[5], *v;
1227 :     Data->lambda16 = iQuant;
1228 :     Data->lambda8 = pParam->m_quant_type;
1229 : chl 181
1230 : edgomez 851 v = Data->qpel ? Data->currentQMV : Data->currentMV;
1231 :     for (i = 0; i < 5; i++) {
1232 :     Data->iMinSAD[i] = 256*4096;
1233 :     backup[i] = v[i];
1234 :     }
1235 : Isibaar 3
1236 : edgomez 851 bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags);
1237 :     if (bits == 0) return MODE_INTER; // quick stop
1238 : Isibaar 3
1239 : edgomez 851 if (inter4v) {
1240 : edgomez 963 int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup);
1241 :     if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; }
1242 : edgomez 851 }
1243 : edgomez 195
1244 : Isibaar 3
1245 : edgomez 851 intra = CountMBBitsIntra(Data);
1246 : Isibaar 3
1247 : edgomez 851 if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; }
1248 :    
1249 :     return mode;
1250 : edgomez 195 }
1251 : edgomez 851 }
1252 : Isibaar 3
1253 : edgomez 851 static void
1254 :     SearchP(const IMAGE * const pRef,
1255 :     const uint8_t * const pRefH,
1256 :     const uint8_t * const pRefV,
1257 :     const uint8_t * const pRefHV,
1258 :     const IMAGE * const pCur,
1259 :     const int x,
1260 :     const int y,
1261 :     const uint32_t MotionFlags,
1262 :     const uint32_t GlobalFlags,
1263 :     const uint32_t iQuant,
1264 :     SearchData * const Data,
1265 :     const MBParam * const pParam,
1266 :     const MACROBLOCK * const pMBs,
1267 :     const MACROBLOCK * const prevMBs,
1268 :     int inter4v,
1269 :     MACROBLOCK * const pMB)
1270 :     {
1271 : Isibaar 3
1272 : edgomez 851 int i, iDirection = 255, mask, threshA;
1273 :     VECTOR pmv[7];
1274 : Isibaar 3
1275 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1276 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1277 : edgomez 195
1278 : edgomez 851 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1279 : Isibaar 3
1280 : edgomez 851 Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache
1281 :     i = Data->rrv ? 2 : 1;
1282 :     Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1283 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1284 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1285 : chl 326
1286 : edgomez 851 Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1287 :     Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1288 :     Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1289 :     Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1290 :     Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1291 :     Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1292 : edgomez 195
1293 : edgomez 851 Data->lambda16 = lambda_vec16[iQuant];
1294 :     Data->lambda8 = lambda_vec8[iQuant];
1295 :     Data->qpel_precision = 0;
1296 : edgomez 195
1297 : edgomez 851 if (pMB->dquant != NO_CHANGE) inter4v = 0;
1298 : edgomez 195
1299 : edgomez 963 memset(Data->currentMV, 0, 5*sizeof(VECTOR));
1300 : chl 169
1301 : edgomez 851 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1302 :     else Data->predMV = pmv[0];
1303 : Isibaar 3
1304 : edgomez 851 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1305 :     Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1306 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1307 :     Data->iMinSAD[2] = pMB->sad8[1];
1308 :     Data->iMinSAD[3] = pMB->sad8[2];
1309 :     Data->iMinSAD[4] = pMB->sad8[3];
1310 : chl 169
1311 : edgomez 851 if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) {
1312 :     threshA = Data->temp[0]; // that's where we keep this SAD atm
1313 :     if (threshA < 512) threshA = 512;
1314 :     else if (threshA > 1024) threshA = 1024;
1315 :     } else
1316 :     threshA = 512;
1317 : chl 169
1318 : edgomez 851 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1319 :     prevMBs + x + y * pParam->mb_width, Data->rrv);
1320 : chl 169
1321 : edgomez 851 if (!Data->rrv) {
1322 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1323 :     else CheckCandidate = CheckCandidate16no4v; //for extra speed
1324 :     } else CheckCandidate = CheckCandidate32;
1325 : chl 169
1326 : edgomez 851 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1327 : chl 169
1328 : edgomez 851 for (i = 1; i < 7; i++) {
1329 :     if (!(mask = make_mask(pmv, i)) ) continue;
1330 :     CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1331 :     if (Data->iMinSAD[0] <= threshA) break;
1332 :     }
1333 : chl 169
1334 : edgomez 851 if ((Data->iMinSAD[0] <= threshA) ||
1335 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1336 :     (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) {
1337 :     if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; }
1338 :     else {
1339 : Isibaar 3
1340 : edgomez 851 MainSearchFunc * MainSearchPtr;
1341 :     if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1342 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1343 :     else MainSearchPtr = DiamondSearch;
1344 : Isibaar 3
1345 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1346 : Isibaar 3
1347 : edgomez 851 /* extended search, diamond starting in 0,0 and in prediction.
1348 :     note that this search is/might be done in halfpel positions,
1349 :     which makes it more different than the diamond above */
1350 : chl 169
1351 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH16) {
1352 :     int32_t bSAD;
1353 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1354 :     if (Data->rrv) {
1355 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1356 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1357 :     }
1358 :     if (!(MVequal(startMV, backupMV))) {
1359 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1360 : edgomez 195
1361 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1362 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1363 :     if (bSAD < Data->iMinSAD[0]) {
1364 :     Data->currentMV[0] = backupMV;
1365 :     Data->iMinSAD[0] = bSAD; }
1366 :     }
1367 : chl 169
1368 : edgomez 851 backupMV = Data->currentMV[0];
1369 :     startMV.x = startMV.y = 1;
1370 :     if (!(MVequal(startMV, backupMV))) {
1371 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1372 : chl 169
1373 : edgomez 851 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1374 :     MainSearchPtr(startMV.x, startMV.y, Data, 255);
1375 :     if (bSAD < Data->iMinSAD[0]) {
1376 :     Data->currentMV[0] = backupMV;
1377 :     Data->iMinSAD[0] = bSAD; }
1378 : edgomez 195 }
1379 : edgomez 851 }
1380 :     }
1381 : edgomez 195
1382 : edgomez 851 if (MotionFlags & PMV_HALFPELREFINE16)
1383 :     if ((!(MotionFlags & HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant)
1384 :     SubpelRefine(Data);
1385 : edgomez 195
1386 : edgomez 851 for(i = 0; i < 5; i++) {
1387 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1388 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1389 :     }
1390 : Isibaar 3
1391 : syskin 904 if (MotionFlags & PMV_QUARTERPELREFINE16) {
1392 :    
1393 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1394 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1395 :    
1396 : edgomez 851 if ((!(MotionFlags & QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) {
1397 :     Data->qpel_precision = 1;
1398 :     SubpelRefine(Data);
1399 :     }
1400 : syskin 904 }
1401 : Isibaar 3
1402 : edgomez 851 if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0;
1403 : Isibaar 3
1404 : edgomez 851 if (inter4v && (!(GlobalFlags & XVID_MODEDECISION_BITS) ||
1405 :     (!(MotionFlags & QUARTERPELREFINE8_BITS)) || (!(MotionFlags & HALFPELREFINE8_BITS)) ||
1406 :     ((!(MotionFlags & EXTSEARCH_BITS)) && (!(MotionFlags&PMV_EXTSEARCH8)) ))) {
1407 :     // if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop
1408 : Isibaar 3
1409 : edgomez 851 SearchData Data8;
1410 :     memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data
1411 : Isibaar 3
1412 : edgomez 851 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1413 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1414 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1415 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1416 : chl 181
1417 : edgomez 851 if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) {
1418 :     // chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it
1419 :     int sumx = 0, sumy = 0;
1420 :     const int div = 1 + Data->qpel;
1421 :     const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs;
1422 : Isibaar 3
1423 : edgomez 851 for (i = 0; i < 4; i++) {
1424 :     sumx += mv[i].x / div;
1425 :     sumy += mv[i].y / div;
1426 :     }
1427 : chl 259
1428 : edgomez 851 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1429 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1430 :     }
1431 :     }
1432 : edgomez 195
1433 : edgomez 851 inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags);
1434 :    
1435 :     if (Data->rrv) {
1436 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1437 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1438 : Isibaar 3 }
1439 :    
1440 : edgomez 851 if (inter4v == MODE_INTER) {
1441 :     pMB->mode = MODE_INTER;
1442 :     pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1443 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1444 : Isibaar 3
1445 : edgomez 851 if(Data->qpel) {
1446 :     pMB->qmvs[0] = pMB->qmvs[1]
1447 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1448 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1449 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1450 :     } else {
1451 :     pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1452 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1453 : Isibaar 3 }
1454 :    
1455 : edgomez 851 } else if (inter4v == MODE_INTER4V) {
1456 :     pMB->mode = MODE_INTER4V;
1457 :     pMB->sad16 = Data->iMinSAD[0];
1458 :     } else { // INTRA mode
1459 :     SkipMacroblockP(pMB, 0); // not skip, but similar enough
1460 :     pMB->mode = MODE_INTRA;
1461 :     }
1462 : edgomez 195
1463 : edgomez 851 }
1464 :    
1465 :     static void
1466 :     Search8(const SearchData * const OldData,
1467 :     const int x, const int y,
1468 :     const uint32_t MotionFlags,
1469 :     const MBParam * const pParam,
1470 :     MACROBLOCK * const pMB,
1471 :     const MACROBLOCK * const pMBs,
1472 :     const int block,
1473 :     SearchData * const Data)
1474 :     {
1475 :     int i = 0;
1476 :     Data->iMinSAD = OldData->iMinSAD + 1 + block;
1477 :     Data->currentMV = OldData->currentMV + 1 + block;
1478 :     Data->currentQMV = OldData->currentQMV + 1 + block;
1479 :    
1480 :     if(Data->qpel) {
1481 :     Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1482 :     if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1483 :     Data->predMV, Data->iFcode, 0, 0);
1484 :     } else {
1485 :     Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1486 :     if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1487 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1488 : Isibaar 3 }
1489 :    
1490 : edgomez 851 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1491 : Isibaar 3
1492 : edgomez 851 if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) {
1493 :     if (Data->rrv) i = 2; else i = 1;
1494 : Isibaar 3
1495 : edgomez 851 Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1496 :     Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1497 :     Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1498 :     Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1499 : Isibaar 3
1500 : edgomez 851 Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1501 :     Data->qpel_precision = 0;
1502 : Isibaar 3
1503 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1504 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1505 : Isibaar 3
1506 : edgomez 851 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1507 :     else CheckCandidate = CheckCandidate16no4v;
1508 : Isibaar 3
1509 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) {
1510 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1511 : Isibaar 3
1512 : edgomez 851 MainSearchFunc *MainSearchPtr;
1513 :     if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1514 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1515 :     else MainSearchPtr = DiamondSearch;
1516 : Isibaar 3
1517 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255);
1518 : Isibaar 3
1519 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) {
1520 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1521 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1522 :     }
1523 :     }
1524 : edgomez 195
1525 : edgomez 851 if (MotionFlags & PMV_HALFPELREFINE8) {
1526 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1527 : edgomez 195
1528 : edgomez 851 SubpelRefine(Data); // perform halfpel refine of current best vector
1529 : Isibaar 3
1530 : edgomez 851 if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1531 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1532 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1533 :     }
1534 :     }
1535 : Isibaar 3
1536 : edgomez 851 if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) {
1537 :     Data->qpel_precision = 1;
1538 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1539 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1540 :     SubpelRefine(Data);
1541 :     }
1542 :     }
1543 : edgomez 195
1544 : edgomez 851 if (Data->rrv) {
1545 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1546 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1547 :     }
1548 :    
1549 :     if(Data->qpel) {
1550 :     pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1551 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1552 :     pMB->qmvs[block] = *Data->currentQMV;
1553 : edgomez 195 } else {
1554 : edgomez 851 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1555 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1556 : edgomez 78 }
1557 : edgomez 851
1558 :     pMB->mvs[block] = *Data->currentMV;
1559 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1560 : Isibaar 3 }
1561 :    
1562 : edgomez 851 /* motion estimation for B-frames */
1563 : chl 345
1564 : edgomez 851 static __inline VECTOR
1565 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1566 :     {
1567 :     /* the stupidiest function ever */
1568 :     return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1569 :     }
1570 : chl 345
1571 : edgomez 851 static void __inline
1572 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1573 :     const uint32_t iWcount,
1574 :     const MACROBLOCK * const pMB,
1575 :     const uint32_t mode_curr)
1576 : chl 345 {
1577 :    
1578 : edgomez 851 // [0] is prediction
1579 :     pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1580 : chl 345
1581 : edgomez 851 pmv[1].x = pmv[1].y = 0; // [1] is zero
1582 : chl 345
1583 : edgomez 851 pmv[2] = ChoosePred(pMB, mode_curr);
1584 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1585 : chl 345
1586 : edgomez 851 if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1587 :     pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1588 :     pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1589 :     } else pmv[3].x = pmv[3].y = 0;
1590 : chl 345
1591 : edgomez 851 if (y != 0) {
1592 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1593 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1594 :     } else pmv[4].x = pmv[4].y = 0;
1595 : chl 345
1596 : edgomez 851 if (x != 0) {
1597 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1598 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1599 :     } else pmv[5].x = pmv[5].y = 0;
1600 : chl 345
1601 : edgomez 851 if (x != 0 && y != 0) {
1602 :     pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1603 :     pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1604 :     } else pmv[6].x = pmv[6].y = 0;
1605 :     }
1606 : chl 345
1607 :    
1608 : edgomez 851 /* search backward or forward */
1609 :     static void
1610 :     SearchBF( const IMAGE * const pRef,
1611 :     const uint8_t * const pRefH,
1612 :     const uint8_t * const pRefV,
1613 :     const uint8_t * const pRefHV,
1614 :     const IMAGE * const pCur,
1615 :     const int x, const int y,
1616 :     const uint32_t MotionFlags,
1617 :     const uint32_t iFcode,
1618 :     const MBParam * const pParam,
1619 :     MACROBLOCK * const pMB,
1620 :     const VECTOR * const predMV,
1621 :     int32_t * const best_sad,
1622 :     const int32_t mode_current,
1623 :     SearchData * const Data)
1624 :     {
1625 : chl 345
1626 : edgomez 851 int i, iDirection = 255, mask;
1627 :     VECTOR pmv[7];
1628 :     MainSearchFunc *MainSearchPtr;
1629 :     *Data->iMinSAD = MV_MAX_ERROR;
1630 :     Data->iFcode = iFcode;
1631 :     Data->qpel_precision = 0;
1632 :     Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache
1633 : chl 345
1634 : edgomez 851 Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16;
1635 :     Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16;
1636 :     Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16;
1637 :     Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16;
1638 :     Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8;
1639 :     Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8;
1640 : chl 345
1641 : edgomez 851 Data->predMV = *predMV;
1642 : chl 345
1643 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1644 :     pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1645 : chl 345
1646 : edgomez 851 pmv[0] = Data->predMV;
1647 :     if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1648 : chl 345
1649 : edgomez 851 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1650 : chl 345
1651 : edgomez 851 Data->currentMV->x = Data->currentMV->y = 0;
1652 :     CheckCandidate = CheckCandidate16no4v;
1653 : chl 345
1654 : edgomez 851 // main loop. checking all predictions
1655 :     for (i = 0; i < 7; i++) {
1656 :     if (!(mask = make_mask(pmv, i)) ) continue;
1657 :     CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1658 : chl 345 }
1659 :    
1660 : edgomez 851 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1661 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1662 :     else MainSearchPtr = DiamondSearch;
1663 : chl 345
1664 : edgomez 851 MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1665 : chl 345
1666 : edgomez 851 SubpelRefine(Data);
1667 : chl 345
1668 : edgomez 851 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1669 :     Data->currentQMV->x = 2*Data->currentMV->x;
1670 :     Data->currentQMV->y = 2*Data->currentMV->y;
1671 :     Data->qpel_precision = 1;
1672 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1673 :     pParam->width, pParam->height, iFcode, 1, 0);
1674 :     SubpelRefine(Data);
1675 :     }
1676 : chl 345
1677 : edgomez 851 // three bits are needed to code backward mode. four for forward
1678 : Isibaar 3
1679 : edgomez 851 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1680 :     else *Data->iMinSAD += 3 * Data->lambda16;
1681 : edgomez 195
1682 : edgomez 851 if (*Data->iMinSAD < *best_sad) {
1683 :     *best_sad = *Data->iMinSAD;
1684 :     pMB->mode = mode_current;
1685 :     if (Data->qpel) {
1686 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1687 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1688 :     if (mode_current == MODE_FORWARD)
1689 :     pMB->qmvs[0] = *Data->currentQMV;
1690 :     else
1691 :     pMB->b_qmvs[0] = *Data->currentQMV;
1692 :     } else {
1693 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1694 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1695 :     }
1696 :     if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1697 :     else pMB->b_mvs[0] = *Data->currentMV;
1698 :     }
1699 : edgomez 195
1700 : edgomez 851 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1701 :     else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
1702 : Isibaar 3 }
1703 :    
1704 : edgomez 851 static void
1705 :     SkipDecisionB(const IMAGE * const pCur,
1706 :     const IMAGE * const f_Ref,
1707 :     const IMAGE * const b_Ref,
1708 :     MACROBLOCK * const pMB,
1709 :     const uint32_t x, const uint32_t y,
1710 :     const SearchData * const Data)
1711 :     {
1712 :     int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1713 :     int32_t sum;
1714 :     const int div = 1 + Data->qpel;
1715 :     int k;
1716 :     const uint32_t stride = Data->iEdgedWidth/2;
1717 :     //this is not full chroma compensation, only it's fullpel approximation. should work though
1718 : Isibaar 3
1719 : edgomez 851 for (k = 0; k < 4; k++) {
1720 :     dy += Data->directmvF[k].y / div;
1721 : edgomez 963 dx += Data->directmvF[k].x / div;
1722 :     b_dy += Data->directmvB[k].y / div;
1723 :     b_dx += Data->directmvB[k].x / div;
1724 : edgomez 851 }
1725 : Isibaar 3
1726 : edgomez 851 dy = (dy >> 3) + roundtab_76[dy & 0xf];
1727 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1728 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1729 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1730 : Isibaar 3
1731 : edgomez 851 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1732 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1733 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1734 :     stride);
1735 : Isibaar 3
1736 : edgomez 851 if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip
1737 : Isibaar 3
1738 : edgomez 851 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1739 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1740 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1741 :     stride);
1742 : edgomez 195
1743 : edgomez 963 if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) {
1744 :     pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1745 :     for (k = 0; k < 4; k++) {
1746 :     pMB->qmvs[k] = pMB->mvs[k];
1747 :     pMB->b_qmvs[k] = pMB->b_mvs[k];
1748 :     }
1749 :     }
1750 : edgomez 851 }
1751 : Isibaar 3
1752 : edgomez 851 static __inline uint32_t
1753 :     SearchDirect(const IMAGE * const f_Ref,
1754 :     const uint8_t * const f_RefH,
1755 :     const uint8_t * const f_RefV,
1756 :     const uint8_t * const f_RefHV,
1757 :     const IMAGE * const b_Ref,
1758 :     const uint8_t * const b_RefH,
1759 :     const uint8_t * const b_RefV,
1760 :     const uint8_t * const b_RefHV,
1761 :     const IMAGE * const pCur,
1762 :     const int x, const int y,
1763 :     const uint32_t MotionFlags,
1764 :     const int32_t TRB, const int32_t TRD,
1765 :     const MBParam * const pParam,
1766 :     MACROBLOCK * const pMB,
1767 :     const MACROBLOCK * const b_mb,
1768 :     int32_t * const best_sad,
1769 :     SearchData * const Data)
1770 : Isibaar 3
1771 : edgomez 851 {
1772 :     int32_t skip_sad;
1773 :     int k = (x + Data->iEdgedWidth*y) * 16;
1774 :     MainSearchFunc *MainSearchPtr;
1775 : Isibaar 3
1776 : edgomez 851 *Data->iMinSAD = 256*4096;
1777 :     Data->Ref = f_Ref->y + k;
1778 :     Data->RefH = f_RefH + k;
1779 :     Data->RefV = f_RefV + k;
1780 :     Data->RefHV = f_RefHV + k;
1781 :     Data->bRef = b_Ref->y + k;
1782 :     Data->bRefH = b_RefH + k;
1783 :     Data->bRefV = b_RefV + k;
1784 :     Data->bRefHV = b_RefHV + k;
1785 :     Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1786 :     Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1787 :     Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1788 :     Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1789 : edgomez 195
1790 : edgomez 851 k = Data->qpel ? 4 : 2;
1791 :     Data->max_dx = k * (pParam->width - x * 16);
1792 :     Data->max_dy = k * (pParam->height - y * 16);
1793 :     Data->min_dx = -k * (16 + x * 16);
1794 :     Data->min_dy = -k * (16 + y * 16);
1795 : chl 181
1796 : edgomez 851 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1797 :     Data->qpel_precision = 0;
1798 : edgomez 170
1799 : edgomez 851 for (k = 0; k < 4; k++) {
1800 :     pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1801 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1802 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1803 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1804 : Isibaar 3
1805 : edgomez 851 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1806 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1807 :    
1808 :     *best_sad = 256*4096; // in that case, we won't use direct mode
1809 :     pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1810 :     pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1811 :     return 256*4096;
1812 :     }
1813 :     if (b_mb->mode != MODE_INTER4V) {
1814 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1815 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1816 :     Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1817 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1818 :     break;
1819 :     }
1820 : edgomez 195 }
1821 : Isibaar 3
1822 : edgomez 851 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1823 : Isibaar 3
1824 : edgomez 851 CheckCandidate(0, 0, 255, &k, Data);
1825 : edgomez 195
1826 : edgomez 851 // initial (fast) skip decision
1827 :     if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) {
1828 :     //possible skip
1829 :     if (Data->chroma) {
1830 :     pMB->mode = MODE_DIRECT_NONE_MV;
1831 :     return *Data->iMinSAD; // skip.
1832 :     } else {
1833 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1834 :     if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1835 :     }
1836 : Isibaar 3 }
1837 :    
1838 : syskin 904 *Data->iMinSAD += Data->lambda16;
1839 : edgomez 851 skip_sad = *Data->iMinSAD;
1840 : edgomez 195
1841 : edgomez 851 // DIRECT MODE DELTA VECTOR SEARCH.
1842 :     // This has to be made more effective, but at the moment I'm happy it's running at all
1843 : Isibaar 3
1844 : edgomez 851 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1845 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1846 :     else MainSearchPtr = DiamondSearch;
1847 : Isibaar 3
1848 : edgomez 851 MainSearchPtr(0, 0, Data, 255);
1849 : Isibaar 3
1850 : edgomez 851 SubpelRefine(Data);
1851 : chl 181
1852 : edgomez 851 *best_sad = *Data->iMinSAD;
1853 : chl 181
1854 : edgomez 851 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1855 :     else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1856 : chl 181
1857 : edgomez 851 pMB->pmvs[3] = *Data->currentMV;
1858 : edgomez 195
1859 : edgomez 851 for (k = 0; k < 4; k++) {
1860 :     pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1861 :     pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1862 :     ? Data->directmvB[k].x
1863 :     :pMB->mvs[k].x - Data->referencemv[k].x);
1864 :     pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1865 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1866 :     ? Data->directmvB[k].y
1867 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1868 :     if (Data->qpel) {
1869 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1870 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1871 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1872 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1873 :     }
1874 : edgomez 195
1875 : edgomez 851 if (b_mb->mode != MODE_INTER4V) {
1876 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1877 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1878 :     pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1879 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1880 :     break;
1881 :     }
1882 : edgomez 78 }
1883 : edgomez 851 return skip_sad;
1884 :     }
1885 : Isibaar 3
1886 : edgomez 851 static void
1887 :     SearchInterpolate(const IMAGE * const f_Ref,
1888 :     const uint8_t * const f_RefH,
1889 :     const uint8_t * const f_RefV,
1890 :     const uint8_t * const f_RefHV,
1891 :     const IMAGE * const b_Ref,
1892 :     const uint8_t * const b_RefH,
1893 :     const uint8_t * const b_RefV,
1894 :     const uint8_t * const b_RefHV,
1895 :     const IMAGE * const pCur,
1896 :     const int x, const int y,
1897 :     const uint32_t fcode,
1898 :     const uint32_t bcode,
1899 :     const uint32_t MotionFlags,
1900 :     const MBParam * const pParam,
1901 :     const VECTOR * const f_predMV,
1902 :     const VECTOR * const b_predMV,
1903 :     MACROBLOCK * const pMB,
1904 :     int32_t * const best_sad,
1905 :     SearchData * const fData)
1906 : chl 96
1907 : edgomez 851 {
1908 : chl 169
1909 : edgomez 851 int iDirection, i, j;
1910 :     SearchData bData;
1911 : chl 169
1912 : edgomez 851 fData->qpel_precision = 0;
1913 :     memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1914 :     *fData->iMinSAD = 4096*256;
1915 :     bData.currentMV++; bData.currentQMV++;
1916 :     fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1917 : chl 169
1918 : edgomez 851 i = (x + y * fData->iEdgedWidth) * 16;
1919 :     bData.bRef = fData->Ref = f_Ref->y + i;
1920 :     bData.bRefH = fData->RefH = f_RefH + i;
1921 :     bData.bRefV = fData->RefV = f_RefV + i;
1922 :     bData.bRefHV = fData->RefHV = f_RefHV + i;
1923 :     bData.Ref = fData->bRef = b_Ref->y + i;
1924 :     bData.RefH = fData->bRefH = b_RefH + i;
1925 :     bData.RefV = fData->bRefV = b_RefV + i;
1926 :     bData.RefHV = fData->bRefHV = b_RefHV + i;
1927 :     bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1928 :     bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1929 :     bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1930 :     bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1931 : chl 169
1932 :    
1933 : edgomez 851 bData.bpredMV = fData->predMV = *f_predMV;
1934 :     fData->bpredMV = bData.predMV = *b_predMV;
1935 :     fData->currentMV[0] = fData->currentMV[2];
1936 : Isibaar 3
1937 : 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);
1938 :     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);
1939 : Isibaar 3
1940 : edgomez 851 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1941 :     if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1942 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1943 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1944 : chl 169
1945 : edgomez 851 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1946 :     if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1947 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1948 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1949 : Isibaar 3
1950 : edgomez 851 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1951 : chl 169
1952 : edgomez 851 //diamond
1953 :     do {
1954 :     iDirection = 255;
1955 :     // forward MV moves
1956 :     i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1957 : chl 169
1958 : edgomez 851 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1959 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1960 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1961 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1962 : edgomez 195
1963 : edgomez 851 // backward MV moves
1964 :     i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1965 :     fData->currentMV[2] = fData->currentMV[0];
1966 :     CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1967 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1968 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1969 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1970 : Isibaar 3
1971 : edgomez 851 } while (!(iDirection));
1972 : chl 140
1973 : edgomez 851 //qpel refinement
1974 :     if (fData->qpel) {
1975 :     if (*fData->iMinSAD > *best_sad + 500) return;
1976 :     CheckCandidate = CheckCandidateInt;
1977 :     fData->qpel_precision = bData.qpel_precision = 1;
1978 :     get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1979 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1980 :     fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1981 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1982 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1983 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1984 :     SubpelRefine(fData);
1985 :     if (*fData->iMinSAD > *best_sad + 300) return;
1986 :     fData->currentQMV[2] = fData->currentQMV[0];
1987 :     SubpelRefine(&bData);
1988 :     }
1989 : chl 140
1990 : edgomez 851 *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode.
1991 : Isibaar 3
1992 : edgomez 851 if (*fData->iMinSAD < *best_sad) {
1993 :     *best_sad = *fData->iMinSAD;
1994 :     pMB->mvs[0] = fData->currentMV[0];
1995 :     pMB->b_mvs[0] = fData->currentMV[1];
1996 :     pMB->mode = MODE_INTERPOLATE;
1997 :     if (fData->qpel) {
1998 :     pMB->qmvs[0] = fData->currentQMV[0];
1999 :     pMB->b_qmvs[0] = fData->currentQMV[1];
2000 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
2001 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
2002 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
2003 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
2004 :     } else {
2005 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
2006 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
2007 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
2008 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
2009 :     }
2010 : edgomez 78 }
2011 : edgomez 851 }
2012 : Isibaar 3
2013 : edgomez 851 void
2014 :     MotionEstimationBVOP(MBParam * const pParam,
2015 :     FRAMEINFO * const frame,
2016 :     const int32_t time_bp,
2017 :     const int32_t time_pp,
2018 :     // forward (past) reference
2019 :     const MACROBLOCK * const f_mbs,
2020 :     const IMAGE * const f_ref,
2021 :     const IMAGE * const f_refH,
2022 :     const IMAGE * const f_refV,
2023 :     const IMAGE * const f_refHV,
2024 :     // backward (future) reference
2025 :     const FRAMEINFO * const b_reference,
2026 :     const IMAGE * const b_ref,
2027 :     const IMAGE * const b_refH,
2028 :     const IMAGE * const b_refV,
2029 :     const IMAGE * const b_refHV)
2030 :     {
2031 :     uint32_t i, j;
2032 :     int32_t best_sad;
2033 :     uint32_t skip_sad;
2034 :     int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
2035 :     const MACROBLOCK * const b_mbs = b_reference->mbs;
2036 : Isibaar 3
2037 : edgomez 851 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
2038 : Isibaar 3
2039 : edgomez 851 const int32_t TRB = time_pp - time_bp;
2040 :     const int32_t TRD = time_pp;
2041 : edgomez 195
2042 : edgomez 851 // some pre-inintialized data for the rest of the search
2043 : Isibaar 3
2044 : edgomez 851 SearchData Data;
2045 :     int32_t iMinSAD;
2046 :     VECTOR currentMV[3];
2047 :     VECTOR currentQMV[3];
2048 :     int32_t temp[8];
2049 :     memset(&Data, 0, sizeof(SearchData));
2050 :     Data.iEdgedWidth = pParam->edged_width;
2051 :     Data.currentMV = currentMV; Data.currentQMV = currentQMV;
2052 :     Data.iMinSAD = &iMinSAD;
2053 :     Data.lambda16 = lambda_vec16[frame->quant];
2054 :     Data.qpel = pParam->m_quarterpel;
2055 :     Data.rounding = 0;
2056 :     Data.chroma = frame->motion_flags & PMV_CHROMA8;
2057 :     Data.temp = temp;
2058 : Isibaar 3
2059 : edgomez 851 Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose)
2060 :     // note: i==horizontal, j==vertical
2061 :     for (j = 0; j < pParam->mb_height; j++) {
2062 : edgomez 195
2063 : edgomez 851 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
2064 : Isibaar 3
2065 : edgomez 851 for (i = 0; i < pParam->mb_width; i++) {
2066 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
2067 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
2068 : edgomez 195
2069 : edgomez 851 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
2070 :     if (b_reference->coding_type != S_VOP)
2071 :     if (b_mb->mode == MODE_NOT_CODED) {
2072 :     pMB->mode = MODE_NOT_CODED;
2073 :     continue;
2074 :     }
2075 : Isibaar 3
2076 : edgomez 851 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
2077 :     Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
2078 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
2079 :     pMB->quant = frame->quant;
2080 : Isibaar 3
2081 : edgomez 851 /* direct search comes first, because it (1) checks for SKIP-mode
2082 :     and (2) sets very good predictions for forward and backward search */
2083 :     skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2084 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2085 :     &frame->image,
2086 :     i, j,
2087 :     frame->motion_flags,
2088 :     TRB, TRD,
2089 :     pParam,
2090 :     pMB, b_mb,
2091 :     &best_sad,
2092 :     &Data);
2093 : Isibaar 3
2094 : edgomez 851 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
2095 : edgomez 195
2096 : edgomez 851 // forward search
2097 :     SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2098 :     &frame->image, i, j,
2099 :     frame->motion_flags,
2100 :     frame->fcode, pParam,
2101 :     pMB, &f_predMV, &best_sad,
2102 :     MODE_FORWARD, &Data);
2103 : edgomez 195
2104 : edgomez 851 // backward search
2105 :     SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
2106 :     &frame->image, i, j,
2107 :     frame->motion_flags,
2108 :     frame->bcode, pParam,
2109 :     pMB, &b_predMV, &best_sad,
2110 :     MODE_BACKWARD, &Data);
2111 :    
2112 :     // interpolate search comes last, because it uses data from forward and backward as prediction
2113 :     SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
2114 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
2115 :     &frame->image,
2116 :     i, j,
2117 :     frame->fcode, frame->bcode,
2118 :     frame->motion_flags,
2119 :     pParam,
2120 :     &f_predMV, &b_predMV,
2121 :     pMB, &best_sad,
2122 :     &Data);
2123 :    
2124 :     // final skip decision
2125 :     if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
2126 :     && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
2127 :     SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
2128 :    
2129 :     switch (pMB->mode) {
2130 :     case MODE_FORWARD:
2131 :     f_count++;
2132 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2133 :     break;
2134 :     case MODE_BACKWARD:
2135 :     b_count++;
2136 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2137 :     break;
2138 :     case MODE_INTERPOLATE:
2139 :     i_count++;
2140 :     f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
2141 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
2142 :     break;
2143 :     case MODE_DIRECT:
2144 :     case MODE_DIRECT_NO4V:
2145 :     d_count++;
2146 :     default:
2147 :     break;
2148 :     }
2149 :     }
2150 :     }
2151 : Isibaar 3 }
2152 : chl 96
2153 : edgomez 851 static __inline void
2154 :     MEanalyzeMB ( const uint8_t * const pRef,
2155 :     const uint8_t * const pCur,
2156 :     const int x,
2157 :     const int y,
2158 :     const MBParam * const pParam,
2159 :     MACROBLOCK * const pMBs,
2160 :     SearchData * const Data)
2161 : chl 96 {
2162 :    
2163 : edgomez 851 int i, mask;
2164 :     VECTOR pmv[3];
2165 : syskin 904 MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width];
2166 : chl 96
2167 : edgomez 851 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
2168 : chl 96
2169 : edgomez 851 //median is only used as prediction. it doesn't have to be real
2170 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
2171 :     else
2172 :     if (x == 1) //left macroblock does not have any vector now
2173 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
2174 :     else if (y == 1) // top macroblock doesn't have it's vector
2175 :     Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
2176 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
2177 : edgomez 195
2178 : edgomez 851 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2179 : syskin 904 pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, 0);
2180 : edgomez 195
2181 : edgomez 851 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
2182 :     Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
2183 : chl 96
2184 : edgomez 851 pmv[1].x = EVEN(pMB->mvs[0].x);
2185 :     pmv[1].y = EVEN(pMB->mvs[0].y);
2186 :     pmv[2].x = EVEN(Data->predMV.x);
2187 :     pmv[2].y = EVEN(Data->predMV.y);
2188 :     pmv[0].x = pmv[0].y = 0;
2189 : edgomez 195
2190 : edgomez 851 CheckCandidate32I(0, 0, 255, &i, Data);
2191 : syskin 904 Data->iMinSAD[1] -= 50;
2192 :     Data->iMinSAD[2] -= 50;
2193 :     Data->iMinSAD[3] -= 50;
2194 :     Data->iMinSAD[4] -= 50;
2195 : edgomez 195
2196 : syskin 904 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) {
2197 : chl 96
2198 : edgomez 851 if (!(mask = make_mask(pmv, 1)))
2199 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
2200 :     if (!(mask = make_mask(pmv, 2)))
2201 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
2202 : chl 96
2203 : syskin 904 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) // diamond only if needed
2204 : edgomez 851 DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
2205 : syskin 904 }
2206 : edgomez 851
2207 : syskin 904 for (i = 0; i < 4; i++) {
2208 :     MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
2209 :     MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
2210 :     MB->mode = MODE_INTER;
2211 :     MB->sad16 = Data->iMinSAD[i+1];
2212 : chl 96 }
2213 : edgomez 851 }
2214 : chl 96
2215 : syskin 904 #define INTRA_THRESH 2400
2216 :     #define INTER_THRESH 1100
2217 : chl 96
2218 : edgomez 851 int
2219 :     MEanalysis( const IMAGE * const pRef,
2220 : syskin 904 const FRAMEINFO * const Current,
2221 :     const MBParam * const pParam,
2222 :     const int maxIntra, //maximum number if non-I frames
2223 :     const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
2224 :     const int bCount, // number of B frames in a row
2225 :     const int b_thresh)
2226 : edgomez 851 {
2227 :     uint32_t x, y, intra = 0;
2228 :     int sSAD = 0;
2229 :     MACROBLOCK * const pMBs = Current->mbs;
2230 :     const IMAGE * const pCurrent = &Current->image;
2231 : syskin 904 int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + 10*b_thresh;
2232 :     int s = 0, blocks = 0;
2233 : chl 96
2234 : edgomez 851 int32_t iMinSAD[5], temp[5];
2235 :     VECTOR currentMV[5];
2236 :     SearchData Data;
2237 :     Data.iEdgedWidth = pParam->edged_width;
2238 :     Data.currentMV = currentMV;
2239 :     Data.iMinSAD = iMinSAD;
2240 :     Data.iFcode = Current->fcode;
2241 :     Data.temp = temp;
2242 :     CheckCandidate = CheckCandidate32I;
2243 : chl 96
2244 : edgomez 851 if (intraCount != 0 && intraCount < 10) // we're right after an I frame
2245 : syskin 904 IntraThresh += 8 * (intraCount - 10) * (intraCount - 10);
2246 : edgomez 851 else
2247 :     if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2248 :     IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
2249 : chl 96
2250 : syskin 904 InterThresh -= (350 - 8*b_thresh) * bCount;
2251 :     if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh;
2252 : chl 326
2253 : edgomez 851 if (sadInit) (*sadInit) ();
2254 : chl 96
2255 : edgomez 851 for (y = 1; y < pParam->mb_height-1; y += 2) {
2256 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2257 :     int i;
2258 : syskin 904 blocks += 4;
2259 : edgomez 195
2260 : edgomez 851 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2261 : syskin 904 else { //extrapolation of the vector found for last frame
2262 :     pMBs[x + y * pParam->mb_width].mvs[0].x =
2263 :     (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount;
2264 :     pMBs[x + y * pParam->mb_width].mvs[0].y =
2265 :     (pMBs[x + y * pParam->mb_width].mvs[0].y * (bCount+1) ) / bCount;
2266 :     }
2267 : edgomez 195
2268 : edgomez 851 MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2269 : edgomez 195
2270 : edgomez 851 for (i = 0; i < 4; i++) {
2271 :     int dev;
2272 :     MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2273 :     if (pMB->sad16 > IntraThresh) {
2274 :     dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2275 :     pParam->edged_width);
2276 :     if (dev + IntraThresh < pMB->sad16) {
2277 :     pMB->mode = MODE_INTRA;
2278 : syskin 904 if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP;
2279 : edgomez 851 }
2280 :     }
2281 : syskin 904 if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++;
2282 :    
2283 : edgomez 851 sSAD += pMB->sad16;
2284 :     }
2285 :     }
2286 : edgomez 195 }
2287 : syskin 904
2288 :     sSAD /= blocks;
2289 :     s = (10*s) / blocks;
2290 :    
2291 :     if (s > 5) sSAD += (s - 4) * (180 - 2*b_thresh); //static block - looks bad when in bframe...
2292 :    
2293 : edgomez 851 if (sSAD > InterThresh ) return P_VOP;
2294 :     emms();
2295 :     return B_VOP;
2296 :     }
2297 : chl 96
2298 :    
2299 : edgomez 851 static WARPPOINTS
2300 :     GlobalMotionEst(const MACROBLOCK * const pMBs,
2301 :     const MBParam * const pParam,
2302 :     const FRAMEINFO * const current,
2303 :     const FRAMEINFO * const reference,
2304 :     const IMAGE * const pRefH,
2305 :     const IMAGE * const pRefV,
2306 :     const IMAGE * const pRefHV )
2307 :     {
2308 : chl 96
2309 : edgomez 851 const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2310 :     const int deltay=8;
2311 :     const int grad=512; // lower bound for deviation in MB
2312 : chl 96
2313 : edgomez 851 WARPPOINTS gmc;
2314 : chl 96
2315 : edgomez 851 uint32_t mx, my;
2316 : chl 96
2317 : edgomez 851 int MBh = pParam->mb_height;
2318 :     int MBw = pParam->mb_width;
2319 : chl 96
2320 : edgomez 851 int *MBmask= calloc(MBh*MBw,sizeof(int));
2321 :     double DtimesF[4] = { 0.,0., 0., 0. };
2322 :     double sol[4] = { 0., 0., 0., 0. };
2323 :     double a,b,c,n,denom;
2324 :     double meanx,meany;
2325 :     int num,oldnum;
2326 : edgomez 195
2327 : chl 871 if (!MBmask) { fprintf(stderr,"Mem error\n");
2328 :     gmc.duv[0].x= gmc.duv[0].y =
2329 :     gmc.duv[1].x= gmc.duv[1].y =
2330 :     gmc.duv[2].x= gmc.duv[2].y = 0;
2331 :     return gmc; }
2332 : edgomez 195
2333 : edgomez 851 // filter mask of all blocks
2334 :    
2335 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2336 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2337 : edgomez 851 {
2338 :     const int mbnum = mx + my * MBw;
2339 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2340 :     const VECTOR mv = pMB->mvs[0];
2341 :    
2342 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2343 :     continue;
2344 :    
2345 :     if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2346 :     && ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2347 :     && ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2348 :     && ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2349 :     MBmask[mbnum]=1;
2350 : chl 96 }
2351 :    
2352 : edgomez 876 for (my = 1; my < (uint32_t)MBh-1; my++)
2353 :     for (mx = 1; mx < (uint32_t)MBw-1; mx++)
2354 : edgomez 851 {
2355 :     const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx;
2356 : chl 96
2357 : edgomez 851 const int mbnum = mx + my * MBw;
2358 :     if (!MBmask[mbnum])
2359 :     continue;
2360 :    
2361 : edgomez 876 if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad )
2362 : edgomez 851 MBmask[mbnum] = 0;
2363 : edgomez 876 if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad )
2364 : edgomez 851 MBmask[mbnum] = 0;
2365 :    
2366 : edgomez 195 }
2367 : chl 96
2368 : edgomez 851 emms();
2369 : chl 96
2370 : edgomez 851 do { /* until convergence */
2371 : chl 96
2372 : edgomez 851 a = b = c = n = 0;
2373 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2374 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2375 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2376 : edgomez 851 {
2377 :     const int mbnum = mx + my * MBw;
2378 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2379 :     const VECTOR mv = pMB->mvs[0];
2380 : chl 96
2381 : edgomez 851 if (!MBmask[mbnum])
2382 :     continue;
2383 : chl 96
2384 : edgomez 851 n++;
2385 :     a += 16*mx+8;
2386 :     b += 16*my+8;
2387 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2388 : chl 96
2389 : edgomez 851 DtimesF[0] += (double)mv.x;
2390 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2391 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2392 :     DtimesF[3] += (double)mv.y;
2393 :     }
2394 : chl 96
2395 : edgomez 851 denom = a*a+b*b-c*n;
2396 : chl 96
2397 : edgomez 851 /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2398 :     /* D'*E*F has been calculated in the same loop as matrix */
2399 : chl 96
2400 : edgomez 851 sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2401 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2402 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2403 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2404 : chl 96
2405 : edgomez 851 sol[0] /= denom;
2406 :     sol[1] /= denom;
2407 :     sol[2] /= denom;
2408 :     sol[3] /= denom;
2409 : chl 96
2410 : edgomez 851 meanx = meany = 0.;
2411 :     oldnum = 0;
2412 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2413 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2414 : edgomez 851 {
2415 :     const int mbnum = mx + my * MBw;
2416 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2417 :     const VECTOR mv = pMB->mvs[0];
2418 : chl 96
2419 : edgomez 851 if (!MBmask[mbnum])
2420 :     continue;
2421 : chl 184
2422 : edgomez 851 oldnum++;
2423 :     meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x );
2424 :     meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y );
2425 :     }
2426 : chl 96
2427 : edgomez 851 if (4*meanx > oldnum) /* better fit than 0.25 is useless */
2428 :     meanx /= oldnum;
2429 :     else
2430 :     meanx = 0.25;
2431 : edgomez 195
2432 : edgomez 851 if (4*meany > oldnum)
2433 :     meany /= oldnum;
2434 :     else
2435 :     meany = 0.25;
2436 : chl 96
2437 : edgomez 851 /* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2438 :     fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2439 :     */
2440 :     num = 0;
2441 : edgomez 876 for (my = 0; my < (uint32_t)MBh; my++)
2442 :     for (mx = 0; mx < (uint32_t)MBw; mx++)
2443 : edgomez 851 {
2444 :     const int mbnum = mx + my * MBw;
2445 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2446 :     const VECTOR mv = pMB->mvs[0];
2447 : chl 96
2448 : edgomez 851 if (!MBmask[mbnum])
2449 :     continue;
2450 : chl 96
2451 : edgomez 851 if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx )
2452 :     || ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) )
2453 :     MBmask[mbnum]=0;
2454 :     else
2455 :     num++;
2456 : chl 96 }
2457 : edgomez 195
2458 : edgomez 851 } while ( (oldnum != num) && (num>=4) );
2459 : edgomez 195
2460 : edgomez 851 if (num < 4)
2461 :     {
2462 :     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;
2463 :     } else {
2464 : edgomez 195
2465 : edgomez 851 gmc.duv[0].x=(int)(sol[0]+0.5);
2466 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2467 : chl 96
2468 : edgomez 851 gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2469 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2470 : chl 96
2471 : edgomez 851 gmc.duv[2].x=0;
2472 :     gmc.duv[2].y=0;
2473 :     }
2474 :     // 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);
2475 : chl 96
2476 : edgomez 851 free(MBmask);
2477 : chl 96
2478 : edgomez 851 return gmc;
2479 : chl 96 }
2480 :    
2481 : edgomez 851 // functions which perform BITS-based search/bitcount
2482 : chl 96
2483 : edgomez 851 static int
2484 :     CountMBBitsInter(SearchData * const Data,
2485 :     const MACROBLOCK * const pMBs, const int x, const int y,
2486 :     const MBParam * const pParam,
2487 :     const uint32_t MotionFlags)
2488 : chl 96 {
2489 : edgomez 851 int i, iDirection;
2490 :     int32_t bsad[5];
2491 : chl 141
2492 : edgomez 851 CheckCandidate = CheckCandidateBits16;
2493 : chl 96
2494 : edgomez 851 if (Data->qpel) {
2495 :     for(i = 0; i < 5; i++) {
2496 :     Data->currentMV[i].x = Data->currentQMV[i].x/2;
2497 :     Data->currentMV[i].y = Data->currentQMV[i].y/2;
2498 :     }
2499 :     Data->qpel_precision = 1;
2500 :     CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data);
2501 : chl 96
2502 : edgomez 851 //checking if this vector is perfect. if it is, we stop.
2503 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0)
2504 :     return 0; //quick stop
2505 : edgomez 195
2506 : edgomez 851 if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search
2507 :     for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i];
2508 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2509 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
2510 :     Data->qpel_precision = 0;
2511 :     if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1)
2512 :     CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2513 :     }
2514 : edgomez 195
2515 : edgomez 851 } else { // not qpel
2516 : edgomez 195
2517 : edgomez 851 CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data);
2518 :     //checking if this vector is perfect. if it is, we stop.
2519 :     if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) {
2520 :     return 0; //inter
2521 :     }
2522 :     }
2523 : chl 96
2524 : edgomez 851 if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
2525 : chl 96
2526 : edgomez 851 if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data);
2527 : chl 96
2528 : edgomez 851 if (Data->qpel) {
2529 :     if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search
2530 :     for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) {
2531 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match
2532 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
2533 :     }
2534 : edgomez 195
2535 : edgomez 851 // preparing for qpel-precision search
2536 :     Data->qpel_precision = 1;
2537 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
2538 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
2539 :     }
2540 :     if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data);
2541 :     }
2542 : chl 96
2543 : edgomez 851 if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction
2544 :     VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV;
2545 :     if (!(Data->predMV.x == v->x && Data->predMV.y == v->y))
2546 :     CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data);
2547 :     }
2548 :     return Data->iMinSAD[0];
2549 :     }
2550 : chl 96
2551 :    
2552 : edgomez 851 static int
2553 :     CountMBBitsInter4v(const SearchData * const Data,
2554 :     MACROBLOCK * const pMB, const MACROBLOCK * const pMBs,
2555 :     const int x, const int y,
2556 :     const MBParam * const pParam, const uint32_t MotionFlags,
2557 :     const VECTOR * const backup)
2558 :     {
2559 : chl 96
2560 : edgomez 851 int cbp = 0, bits = 0, t = 0, i, iDirection;
2561 :     SearchData Data2, *Data8 = &Data2;
2562 :     int sumx = 0, sumy = 0;
2563 : edgomez 961 int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64;
2564 : chl 96
2565 : edgomez 851 memcpy(Data8, Data, sizeof(SearchData));
2566 :     CheckCandidate = CheckCandidateBits8;
2567 : chl 96
2568 : edgomez 851 for (i = 0; i < 4; i++) {
2569 :     Data8->iMinSAD = Data->iMinSAD + i + 1;
2570 :     Data8->currentMV = Data->currentMV + i + 1;
2571 :     Data8->currentQMV = Data->currentQMV + i + 1;
2572 :     Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2573 :     Data8->Ref = Data->Ref + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2574 :     Data8->RefH = Data->RefH + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2575 :     Data8->RefV = Data->RefV + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2576 :     Data8->RefHV = Data->RefHV + 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2577 : edgomez 195
2578 : edgomez 851 if(Data->qpel) {
2579 :     Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i);
2580 :     if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y,
2581 :     Data8->predMV, Data8->iFcode, 0, 0);
2582 :     } else {
2583 :     Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i);
2584 :     if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y,
2585 :     Data8->predMV, Data8->iFcode, 0, 0);
2586 :     }
2587 : chl 96
2588 : 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,
2589 :     pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0);
2590 : chl 96
2591 : edgomez 851 *Data8->iMinSAD += t;
2592 : chl 96
2593 : edgomez 851 Data8->qpel_precision = Data8->qpel;
2594 :     // checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far)
2595 :     if (Data8->qpel) {
2596 :     if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y))
2597 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2598 :     } else {
2599 :     if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y))
2600 :     CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8);
2601 :     }
2602 : edgomez 195
2603 : edgomez 851 if (Data8->qpel) {
2604 :     if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows
2605 :     int32_t s = *Data8->iMinSAD;
2606 :     Data8->currentMV->x = Data8->currentQMV->x/2;
2607 :     Data8->currentMV->y = Data8->currentQMV->y/2;
2608 :     Data8->qpel_precision = 0;
2609 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2610 :     pParam->width, pParam->height, Data8->iFcode - 1, 0, 0);
2611 : edgomez 195
2612 : edgomez 851 if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1)
2613 :     CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8);
2614 : edgomez 195
2615 : edgomez 851 if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS)
2616 :     SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255);
2617 : chl 96
2618 : edgomez 851 if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8);
2619 : chl 96
2620 : edgomez 851 if(s > *Data8->iMinSAD) { //we have found a better match
2621 :     Data8->currentQMV->x = 2*Data8->currentMV->x;
2622 :     Data8->currentQMV->y = 2*Data8->currentMV->y;
2623 :     }
2624 : chl 96
2625 : edgomez 851 Data8->qpel_precision = 1;
2626 :     get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8,
2627 :     pParam->width, pParam->height, Data8->iFcode, 1, 0);
2628 : chl 96
2629 : edgomez 851 }
2630 :     if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8);
2631 : edgomez 195
2632 : edgomez 851 } else // not qpel
2633 :     if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement
2634 :    
2635 :     //checking vector equal to predicion
2636 :     if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) {
2637 :     const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV;
2638 :     if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y))
2639 :     CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8);
2640 : chl 141 }
2641 : edgomez 195
2642 : edgomez 851 bits += *Data8->iMinSAD;
2643 :     if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V
2644 :    
2645 :     // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else
2646 :     if(Data->qpel) {
2647 :     pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x;
2648 :     pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y;
2649 :     pMB->qmvs[i] = *Data8->currentQMV;
2650 :     sumx += Data8->currentQMV->x/2;
2651 :     sumy += Data8->currentQMV->y/2;
2652 :     } else {
2653 :     pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x;
2654 :     pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y;
2655 :     sumx += Data8->currentMV->x;
2656 :     sumy += Data8->currentMV->y;
2657 : chl 141 }
2658 : edgomez 851 pMB->mvs[i] = *Data8->currentMV;
2659 :     pMB->sad8[i] = 4 * *Data8->iMinSAD;
2660 :     if (Data8->temp[0]) cbp |= 1 << (5 - i);
2661 : chl 141 }
2662 :    
2663 : edgomez 851 if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma
2664 :     const uint8_t * ptr;
2665 :     sumx = (sumx >> 3) + roundtab_76[sumx & 0xf];
2666 :     sumy = (sumy >> 3) + roundtab_76[sumy & 0xf];
2667 : chl 141
2668 : edgomez 851 //chroma U
2669 :     ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCU, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2670 :     transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2);
2671 :     fdct(in);
2672 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2673 :     else i = quant4_inter(coeff, in, Data->lambda16);
2674 :     if (i > 0) {
2675 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2676 :     cbp |= 1 << (5 - 4);
2677 :     }
2678 : chl 141
2679 : edgomez 851 if (bits < *Data->iMinSAD) { // still possible
2680 :     //chroma V
2681 :     ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCV, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding);
2682 :     transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2);
2683 :     fdct(in);
2684 :     if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16);
2685 :     else i = quant4_inter(coeff, in, Data->lambda16);
2686 :     if (i > 0) {
2687 :     bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]);
2688 :     cbp |= 1 << (5 - 5);
2689 :     }
2690 : chl 897 bits += xvid_cbpy_tab[15-(cbp>>2)].len;
2691 : edgomez 851 bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len;
2692 :     }
2693 : edgomez 195 }
2694 : chl 96
2695 : edgomez 851 return bits;
2696 :     }
2697 : chl 96
2698 :    
2699 : edgomez 851 static int
2700 :     CountMBBitsIntra(const SearchData * const Data)
2701 :     {
2702 :     int bits = 1; //this one is ac/dc prediction flag. always 1.
2703 :     int cbp = 0, i, t, dc = 0, b_dc = 1024;
2704 :     const uint32_t iQuant = Data->lambda16;
2705 : edgomez 961 int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64;
2706 : edgomez 195
2707 : edgomez 851 for(i = 0; i < 4; i++) {
2708 :     uint32_t iDcScaler = get_dc_scaler(iQuant, 1);
2709 : chl 96
2710 : edgomez 851 int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth);
2711 :     transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth);
2712 :     fdct(in);
2713 :     b_dc = dc;
2714 :     dc = in[0];
2715 :     in[0] -= b_dc;
2716 :     if (Data->lambda8 == 0) quant_intra_c(coeff, in, iQuant, iDcScaler);
2717 :     else quant4_intra_c(coeff, in, iQuant, iDcScaler);
2718 : chl 141
2719 : edgomez 851 b_dc = dc;
2720 :     dc = coeff[0];
2721 :     if (i != 0) coeff[0] -= b_dc;
2722 : chl 181
2723 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;;
2724 :     Data->temp[i] = t;
2725 :     if (t != 0) cbp |= 1 << (5 - i);
2726 :     if (bits >= Data->iMinSAD[0]) break;
2727 :     }
2728 : chl 181
2729 : edgomez 851 if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma
2730 :     uint32_t iDcScaler = get_dc_scaler(iQuant, 0);
2731 :     //chroma U
2732 :     transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2);
2733 :     fdct(in);
2734 :     in[0] -= 1024;
2735 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2736 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2737 : chl 96
2738 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2739 :     if (t != 0) cbp |= 1 << (5 - 4);
2740 : chl 141
2741 : edgomez 851 if (bits < Data->iMinSAD[0]) {
2742 : syskin 904 iDcScaler = get_dc_scaler(iQuant, 1);
2743 : edgomez 851 //chroma V
2744 :     transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2);
2745 :     fdct(in);
2746 :     in[0] -= 1024;
2747 :     if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler);
2748 :     else quant4_intra(coeff, in, iQuant, iDcScaler);
2749 : chl 96
2750 : edgomez 851 bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len;
2751 :     if (t != 0) cbp |= 1 << (5 - 5);
2752 : chl 96
2753 : syskin 904 bits += xvid_cbpy_tab[cbp>>2].len;
2754 :     bits += mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len;
2755 : chl 96 }
2756 :     }
2757 : edgomez 851 return bits;
2758 : chl 96 }

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