[svn] / branches / dev-api-3 / xvidcore / src / motion / motion_est.c Repository:
ViewVC logotype

Annotation of /branches/dev-api-3/xvidcore/src/motion/motion_est.c

Parent Directory Parent Directory | Revision Log Revision Log


Revision 819 - (view) (download)

1 : Isibaar 3 /**************************************************************************
2 :     *
3 : chl 259 * XVID MPEG-4 VIDEO CODEC
4 : syskin 748 * motion estimation
5 : chl 259 *
6 :     * 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 :     *
15 :     * 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 :     *
20 :     * 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 :     *
25 :     * 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 :     *
29 :     *************************************************************************/
30 :    
31 : Isibaar 3 #include <assert.h>
32 :     #include <stdio.h>
33 : chl 96 #include <stdlib.h>
34 : suxen_drol 721 #include <string.h> // memcpy
35 : syskin 801 #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 : Isibaar 580 #include "../image/interpolate8x8.h"
43 : chl 530 #include "motion_est.h"
44 : suxen_drol 118 #include "motion.h"
45 : Isibaar 3 #include "sad.h"
46 : Isibaar 539 #include "../utils/emms.h"
47 : Isibaar 3
48 : chl 530 #define INITIAL_SKIP_THRESH (10)
49 :     #define FINAL_SKIP_THRESH (50)
50 :     #define MAX_SAD00_FOR_SKIP (20)
51 : Isibaar 539 #define MAX_CHROMA_SAD_FOR_SKIP (22)
52 : Isibaar 3
53 : suxen_drol 819 #define CHECK_CANDIDATE(X,Y,D) { CheckCandidate((X),(Y), (D), &iDirection, data ); }
54 : Isibaar 3
55 : syskin 768 static __inline uint32_t
56 : syskin 724 d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv)
57 : Isibaar 3 {
58 : chl 530 int xb, yb;
59 : syskin 817 x = qpel ? x<<1 : x;
60 :     y = qpel ? y<<1 : y;
61 : syskin 768 if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); }
62 : syskin 817
63 : syskin 748 x -= pred.x;
64 :     y -= pred.y;
65 : syskin 768
66 : syskin 748 if (x) {
67 : syskin 768 x = ABS(x);
68 : chl 530 x += (1 << (iFcode - 1)) - 1;
69 :     x >>= (iFcode - 1);
70 :     if (x > 32) x = 32;
71 :     xb = mvtab[x] + iFcode;
72 : syskin 748 } else xb = 1;
73 : Isibaar 3
74 : syskin 748 if (y) {
75 : syskin 768 y = ABS(y);
76 : chl 530 y += (1 << (iFcode - 1)) - 1;
77 :     y >>= (iFcode - 1);
78 :     if (y > 32) y = 32;
79 :     yb = mvtab[y] + iFcode;
80 : syskin 748 } else yb = 1;
81 : chl 530 return xb + yb;
82 : Isibaar 3 }
83 :    
84 : syskin 801 static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data)
85 :     {
86 :     int sad;
87 :     const uint32_t stride = data->iEdgedWidth/2;
88 :     uint8_t * f_refu = data->RefQ,
89 :     * f_refv = data->RefQ + 8,
90 :     * b_refu = data->RefQ + 16,
91 :     * b_refv = data->RefQ + 24;
92 :    
93 :     switch (((fx & 1) << 1) | (fy & 1)) {
94 :     case 0:
95 :     fx = fx / 2; fy = fy / 2;
96 :     f_refu = (uint8_t*)data->RefCU + fy * stride + fx, stride;
97 :     f_refv = (uint8_t*)data->RefCV + fy * stride + fx, stride;
98 :     break;
99 :     case 1:
100 :     fx = fx / 2; fy = (fy - 1) / 2;
101 :     interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
102 :     interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
103 :     break;
104 :     case 2:
105 :     fx = (fx - 1) / 2; fy = fy / 2;
106 :     interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
107 :     interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
108 :     break;
109 :     default:
110 :     fx = (fx - 1) / 2; fy = (fy - 1) / 2;
111 :     interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding);
112 :     interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding);
113 :     break;
114 :     }
115 :    
116 :     switch (((bx & 1) << 1) | (by & 1)) {
117 :     case 0:
118 :     bx = bx / 2; by = by / 2;
119 :     b_refu = (uint8_t*)data->b_RefCU + by * stride + bx, stride;
120 :     b_refv = (uint8_t*)data->b_RefCV + by * stride + bx, stride;
121 :     break;
122 :     case 1:
123 :     bx = bx / 2; by = (by - 1) / 2;
124 :     interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
125 :     interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
126 :     break;
127 :     case 2:
128 :     bx = (bx - 1) / 2; by = by / 2;
129 :     interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
130 :     interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
131 :     break;
132 :     default:
133 :     bx = (bx - 1) / 2; by = (by - 1) / 2;
134 :     interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding);
135 :     interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding);
136 :     break;
137 :     }
138 :    
139 :     sad = sad8bi(data->CurU, b_refu, f_refu, stride);
140 :     sad += sad8bi(data->CurV, b_refv, f_refv, stride);
141 :    
142 :     return sad;
143 :     }
144 :    
145 :    
146 : syskin 748 static int32_t
147 : syskin 628 ChromaSAD(int dx, int dy, const SearchData * const data)
148 : Isibaar 3 {
149 : syskin 628 int sad;
150 : syskin 748 const uint32_t stride = data->iEdgedWidth/2;
151 : Isibaar 3
152 : syskin 712 if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently
153 : syskin 768 data->temp[5] = dx; data->temp[6] = dy; // backup
154 : syskin 712
155 :     switch (((dx & 1) << 1) | (dy & 1)) {
156 : syskin 628 case 0:
157 : syskin 748 dx = dx / 2; dy = dy / 2;
158 :     sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride);
159 :     sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride);
160 : syskin 628 break;
161 :     case 1:
162 :     dx = dx / 2; dy = (dy - 1) / 2;
163 : syskin 748 sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride);
164 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride);
165 : syskin 628 break;
166 :     case 2:
167 :     dx = (dx - 1) / 2; dy = dy / 2;
168 : syskin 748 sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride);
169 :     sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride);
170 : syskin 628 break;
171 :     default:
172 :     dx = (dx - 1) / 2; dy = (dy - 1) / 2;
173 : syskin 748 interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding);
174 :     sad = sad8(data->CurU, data->RefQ, stride);
175 :    
176 :     interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding);
177 :     sad += sad8(data->CurV, data->RefQ, stride);
178 : syskin 628 break;
179 : chl 530 }
180 : syskin 748 data->temp[7] = sad; //backup, part 2
181 : syskin 628 return sad;
182 :     }
183 : Isibaar 3
184 : syskin 663 static __inline const uint8_t *
185 : syskin 768 GetReferenceB(const int x, const int y, const uint32_t dir, const SearchData * const data)
186 : Isibaar 606 {
187 : syskin 663 // dir : 0 = forward, 1 = backward
188 :     switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) {
189 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
190 :     case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
191 :     case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
192 :     case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
193 :     case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth);
194 :     case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
195 :     case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth);
196 :     default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
197 : Isibaar 606 }
198 :     }
199 :    
200 : syskin 748 // this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate
201 :     static __inline const uint8_t *
202 :     GetReference(const int x, const int y, const SearchData * const data)
203 :     {
204 :     switch ( ((x&1)<<1) | (y&1) ) {
205 :     case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth);
206 : syskin 817 case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth);
207 : syskin 748 case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth);
208 : syskin 817 default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2
209 : syskin 748 }
210 :     }
211 :    
212 :     static uint8_t *
213 : syskin 768 Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data)
214 : chl 530 {
215 : syskin 663 // create or find a qpel-precision reference picture; return pointer to it
216 : syskin 768 uint8_t * Reference = data->RefQ + 16*dir;
217 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
218 : syskin 663 const uint32_t rounding = data->rounding;
219 :     const int halfpel_x = x/2;
220 :     const int halfpel_y = y/2;
221 :     const uint8_t *ref1, *ref2, *ref3, *ref4;
222 : chl 326
223 : syskin 748 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
224 : syskin 663 ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
225 :     switch( ((x&1)<<1) + (y&1) ) {
226 :     case 0: // pure halfpel position
227 : syskin 740 return (uint8_t *) ref1;
228 : syskin 663 break;
229 : Isibaar 3
230 : syskin 663 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
231 : syskin 748 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
232 : syskin 663 ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
233 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
234 : syskin 663 break;
235 : Isibaar 3
236 : syskin 663 case 2: // x qpel, y halfpel - left or right during qpel refinement
237 : syskin 748 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
238 : syskin 663 ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
239 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
240 : syskin 663 break;
241 : Isibaar 3
242 : syskin 663 default: // x and y in qpel resolution - the "corners" (top left/right and
243 :     // bottom left/right) during qpel refinement
244 : syskin 748 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
245 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
246 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
247 : syskin 663 ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
248 :     ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
249 :     ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth;
250 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
251 :     break;
252 :     }
253 :     return Reference;
254 : Isibaar 3 }
255 :    
256 : syskin 748 static uint8_t *
257 : syskin 768 Interpolate16x16qpel(const int x, const int y, const uint32_t dir, const SearchData * const data)
258 : Isibaar 580 {
259 : syskin 663 // create or find a qpel-precision reference picture; return pointer to it
260 : syskin 768 uint8_t * Reference = data->RefQ + 16*dir;
261 :     const uint32_t iEdgedWidth = data->iEdgedWidth;
262 : syskin 663 const uint32_t rounding = data->rounding;
263 :     const int halfpel_x = x/2;
264 :     const int halfpel_y = y/2;
265 : Isibaar 580 const uint8_t *ref1, *ref2, *ref3, *ref4;
266 :    
267 : syskin 748 ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data);
268 : syskin 663 switch( ((x&1)<<1) + (y&1) ) {
269 : syskin 817 case 3: // x and y in qpel resolution - the "corners" (top left/right and
270 :     // bottom left/right) during qpel refinement
271 :     ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
272 :     ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
273 :     ref4 = GetReferenceB(x - halfpel_x, y - halfpel_y, dir, data);
274 :     interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding);
275 :     interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding);
276 :     interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding);
277 :     interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding);
278 :     break;
279 :    
280 : Isibaar 580 case 1: // x halfpel, y qpel - top or bottom during qpel refinement
281 : syskin 748 ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data);
282 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
283 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
284 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
285 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
286 : Isibaar 580 break;
287 :    
288 :     case 2: // x qpel, y halfpel - left or right during qpel refinement
289 : syskin 768 ref2 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data);
290 : Isibaar 665 interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8);
291 :     interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8);
292 :     interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8);
293 :     interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8);
294 : Isibaar 580 break;
295 :    
296 : syskin 817 case 0: // pure halfpel position
297 :     return (uint8_t *) ref1;
298 : Isibaar 580 }
299 : syskin 663 return Reference;
300 :     }
301 :    
302 :     /* CHECK_CANDIATE FUNCTIONS START */
303 :    
304 : syskin 748 static void
305 : syskin 663 CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
306 :     {
307 : syskin 817 int xc, yc;
308 : syskin 663 const uint8_t * Reference;
309 :     VECTOR * current;
310 : syskin 817 int32_t sad; uint32_t t;
311 : syskin 663
312 : syskin 817 if ( (x > data->max_dx) || (x < data->min_dx)
313 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
314 : syskin 663
315 : syskin 817 if (!data->qpel_precision) {
316 :     Reference = GetReference(x, y, data);
317 :     current = data->currentMV;
318 :     xc = x; yc = y;
319 :     } else { // x and y are in 1/4 precision
320 : syskin 663 Reference = Interpolate16x16qpel(x, y, 0, data);
321 :     xc = x/2; yc = y/2; //for chroma sad
322 :     current = data->currentQMV;
323 :     }
324 : syskin 817
325 :     sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
326 : syskin 768 t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
327 :    
328 : syskin 817 sad += (data->lambda16 * t * sad)>>10;
329 : syskin 768 data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
330 : Isibaar 580
331 : syskin 817 if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
332 : syskin 753 (yc >> 1) + roundtab_79[yc & 0x3], data);
333 : syskin 628
334 : syskin 817 if (sad < data->iMinSAD[0]) {
335 :     data->iMinSAD[0] = sad;
336 : syskin 663 current[0].x = x; current[0].y = y;
337 : syskin 817 *dir = Direction;
338 :     }
339 : Isibaar 580
340 :     if (data->temp[1] < data->iMinSAD[1]) {
341 : syskin 817 data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y = y; }
342 : Isibaar 580 if (data->temp[2] < data->iMinSAD[2]) {
343 : syskin 663 data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; }
344 : Isibaar 580 if (data->temp[3] < data->iMinSAD[3]) {
345 : syskin 663 data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; }
346 : Isibaar 580 if (data->temp[4] < data->iMinSAD[4]) {
347 : syskin 663 data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; }
348 :    
349 : Isibaar 580 }
350 :    
351 : syskin 748 static void
352 : syskin 817 CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
353 :     {
354 :     int32_t sad; uint32_t t;
355 :     const uint8_t * Reference;
356 :    
357 :     if ( (x > data->max_dx) || (x < data->min_dx)
358 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
359 :    
360 :     if (!data->qpel_precision) Reference = GetReference(x, y, data);
361 : syskin 818 else Reference = Interpolate8x8qpel(x, y, 0, 0, data);
362 : syskin 817
363 :     sad = sad8(data->Cur, Reference, data->iEdgedWidth);
364 :     t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0);
365 :    
366 :     sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10;
367 :    
368 :     if (sad < *(data->iMinSAD)) {
369 :     *(data->iMinSAD) = sad;
370 :     data->currentMV->x = x; data->currentMV->y = y;
371 :     *dir = Direction;
372 :     }
373 :     }
374 :    
375 :    
376 :     static void
377 : syskin 704 CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
378 :     {
379 : syskin 768 uint32_t t;
380 : syskin 704 const uint8_t * Reference;
381 :    
382 : syskin 817 if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value
383 :     (x > data->max_dx) || (x < data->min_dx)
384 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
385 : syskin 704
386 : syskin 748 Reference = GetReference(x, y, data);
387 : syskin 724 t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1);
388 : syskin 704
389 :     data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1);
390 :    
391 : syskin 768 data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10;
392 :     data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10;
393 : syskin 704
394 :     if (data->temp[0] < data->iMinSAD[0]) {
395 :     data->iMinSAD[0] = data->temp[0];
396 :     data->currentMV[0].x = x; data->currentMV[0].y = y;
397 :     *dir = Direction; }
398 :    
399 :     if (data->temp[1] < data->iMinSAD[1]) {
400 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
401 :     if (data->temp[2] < data->iMinSAD[2]) {
402 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
403 :     if (data->temp[3] < data->iMinSAD[3]) {
404 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
405 :     if (data->temp[4] < data->iMinSAD[4]) {
406 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
407 :     }
408 :    
409 : syskin 748 static void
410 : syskin 663 CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
411 : syskin 658 {
412 : syskin 801 int32_t sad, xc, yc;
413 : syskin 663 const uint8_t * Reference;
414 : syskin 768 uint32_t t;
415 : syskin 663 VECTOR * current;
416 : syskin 658
417 : syskin 768 if ( (x > data->max_dx) | ( x < data->min_dx)
418 :     | (y > data->max_dy) | (y < data->min_dy) ) return;
419 : syskin 658
420 : syskin 768 if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value
421 : syskin 704
422 : syskin 663 if (data->qpel_precision) { // x and y are in 1/4 precision
423 :     Reference = Interpolate16x16qpel(x, y, 0, data);
424 :     current = data->currentQMV;
425 : syskin 801 xc = x/2; yc = y/2;
426 : syskin 663 } else {
427 : syskin 748 Reference = GetReference(x, y, data);
428 : syskin 663 current = data->currentMV;
429 : syskin 801 xc = x; yc = y;
430 : syskin 658 }
431 : syskin 724 t = d_mv_bits(x, y, data->predMV, data->iFcode,
432 : syskin 768 data->qpel^data->qpel_precision, data->rrv);
433 :    
434 : syskin 658 sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096);
435 : syskin 768 sad += (data->lambda16 * t * sad)>>10;
436 : syskin 658
437 : syskin 801 if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
438 :     (yc >> 1) + roundtab_79[yc & 0x3], data);
439 :    
440 :    
441 : syskin 663 if (sad < *(data->iMinSAD)) {
442 :     *(data->iMinSAD) = sad;
443 :     current->x = x; current->y = y;
444 : syskin 801 *dir = Direction;
445 :     }
446 : syskin 658 }
447 :    
448 : syskin 748 static void
449 : syskin 739 CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
450 : Isibaar 539 {
451 : syskin 658 // maximum speed - for P/B/I decision
452 : syskin 817 int32_t sad;
453 : Isibaar 539
454 : syskin 817 if ( (x > data->max_dx) || (x < data->min_dx)
455 :     || (y > data->max_dy) || (y < data->min_dy) ) return;
456 : Isibaar 539
457 : syskin 817 sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth),
458 : syskin 739 data->iEdgedWidth, data->temp+1);
459 : syskin 768
460 : syskin 817 if (sad < *(data->iMinSAD)) {
461 :     *(data->iMinSAD) = sad;
462 : Isibaar 539 data->currentMV[0].x = x; data->currentMV[0].y = y;
463 : syskin 817 *dir = Direction;
464 :     }
465 : syskin 739 if (data->temp[1] < data->iMinSAD[1]) {
466 :     data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; }
467 :     if (data->temp[2] < data->iMinSAD[2]) {
468 :     data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; }
469 :     if (data->temp[3] < data->iMinSAD[3]) {
470 :     data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; }
471 :     if (data->temp[4] < data->iMinSAD[4]) {
472 :     data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; }
473 :    
474 : Isibaar 539 }
475 :    
476 : syskin 748 static void
477 : chl 530 CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data)
478 :     {
479 : syskin 801 int32_t sad, xb, yb, xcf, ycf, xcb, ycb;
480 : syskin 768 uint32_t t;
481 : chl 530 const uint8_t *ReferenceF, *ReferenceB;
482 : syskin 663 VECTOR *current;
483 : Isibaar 3
484 : syskin 768 if ( (xf > data->max_dx) | (xf < data->min_dx)
485 :     | (yf > data->max_dy) | (yf < data->min_dy) ) return;
486 : Isibaar 3
487 : syskin 768 if (!data->qpel_precision) {
488 : syskin 748 ReferenceF = GetReference(xf, yf, data);
489 : syskin 663 xb = data->currentMV[1].x; yb = data->currentMV[1].y;
490 : syskin 748 ReferenceB = GetReferenceB(xb, yb, 1, data);
491 : syskin 663 current = data->currentMV;
492 : syskin 801 xcf = xf; ycf = yf;
493 :     xcb = xb; ycb = yb;
494 : syskin 768 } else {
495 :     ReferenceF = Interpolate16x16qpel(xf, yf, 0, data);
496 :     xb = data->currentQMV[1].x; yb = data->currentQMV[1].y;
497 :     current = data->currentQMV;
498 :     ReferenceB = Interpolate16x16qpel(xb, yb, 1, data);
499 : syskin 801 xcf = xf/2; ycf = yf/2;
500 :     xcb = xb/2; ycb = yb/2;
501 : chl 530 }
502 : edgomez 195
503 : syskin 768 t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0)
504 :     + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel^data->qpel_precision, 0);
505 :    
506 : syskin 628 sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
507 : syskin 768 sad += (data->lambda16 * t * sad)>>10;
508 : syskin 628
509 : syskin 801 if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3],
510 :     (ycf >> 1) + roundtab_79[ycf & 0x3],
511 :     (xcb >> 1) + roundtab_79[xcb & 0x3],
512 :     (ycb >> 1) + roundtab_79[ycb & 0x3], data);
513 :    
514 : chl 530 if (sad < *(data->iMinSAD)) {
515 :     *(data->iMinSAD) = sad;
516 : syskin 663 current->x = xf; current->y = yf;
517 : syskin 801 *dir = Direction;
518 :     }
519 : Isibaar 3 }
520 :    
521 : syskin 748 static void
522 : chl 530 CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
523 : Isibaar 3 {
524 : syskin 801 int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0;
525 : syskin 768 uint32_t k;
526 : chl 530 const uint8_t *ReferenceF;
527 :     const uint8_t *ReferenceB;
528 :     VECTOR mvs, b_mvs;
529 : Isibaar 3
530 : syskin 768 if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return;
531 : Isibaar 3
532 : chl 530 for (k = 0; k < 4; k++) {
533 :     mvs.x = data->directmvF[k].x + x;
534 :     b_mvs.x = ((x == 0) ?
535 :     data->directmvB[k].x
536 :     : mvs.x - data->referencemv[k].x);
537 : Isibaar 3
538 : chl 530 mvs.y = data->directmvF[k].y + y;
539 :     b_mvs.y = ((y == 0) ?
540 :     data->directmvB[k].y
541 :     : mvs.y - data->referencemv[k].y);
542 : syskin 768
543 : syskin 801 if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx)
544 :     | (mvs.y > data->max_dy) | (mvs.y < data->min_dy)
545 :     | (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx)
546 :     | (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return;
547 : Isibaar 3
548 : syskin 801 if (data->qpel) {
549 :     xcf += mvs.x/2; ycf += mvs.y/2;
550 :     xcb += b_mvs.x/2; ycb += b_mvs.y/2;
551 :     } else {
552 :     xcf += mvs.x; ycf += mvs.y;
553 :     xcb += b_mvs.x; ycb += b_mvs.y;
554 :     mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway
555 :     b_mvs.x *= 2; b_mvs.y *= 2;
556 :     }
557 :    
558 : syskin 663 ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data);
559 :     ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data);
560 : syskin 768
561 : chl 530 sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth),
562 : syskin 768 ReferenceF, ReferenceB, data->iEdgedWidth);
563 : chl 530 if (sad > *(data->iMinSAD)) return;
564 : edgomez 78 }
565 : chl 530
566 : syskin 768 sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
567 : syskin 628
568 : syskin 801 if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
569 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
570 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
571 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
572 :    
573 : chl 530 if (sad < *(data->iMinSAD)) {
574 :     *(data->iMinSAD) = sad;
575 :     data->currentMV->x = x; data->currentMV->y = y;
576 : syskin 801 *dir = Direction;
577 :     }
578 : Isibaar 3 }
579 :    
580 : syskin 748 static void
581 : chl 530 CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data)
582 : chl 96 {
583 : syskin 801 int32_t sad, xcf, ycf, xcb, ycb;
584 : chl 530 const uint8_t *ReferenceF;
585 :     const uint8_t *ReferenceB;
586 :     VECTOR mvs, b_mvs;
587 : chl 96
588 : syskin 768 if (( x > 31) | ( x < -32) | ( y > 31) | (y < -32)) return;
589 : edgomez 195
590 : chl 530 mvs.x = data->directmvF[0].x + x;
591 :     b_mvs.x = ((x == 0) ?
592 :     data->directmvB[0].x
593 :     : mvs.x - data->referencemv[0].x);
594 : edgomez 195
595 : chl 530 mvs.y = data->directmvF[0].y + y;
596 :     b_mvs.y = ((y == 0) ?
597 :     data->directmvB[0].y
598 :     : mvs.y - data->referencemv[0].y);
599 : syskin 768
600 : syskin 801 if ( (mvs.x > data->max_dx) | (mvs.x < data->min_dx)
601 :     | (mvs.y > data->max_dy) | (mvs.y < data->min_dy)
602 :     | (b_mvs.x > data->max_dx) | (b_mvs.x < data->min_dx)
603 :     | (b_mvs.y > data->max_dy) | (b_mvs.y < data->min_dy) ) return;
604 : chl 96
605 : syskin 801 if (data->qpel) {
606 :     xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2);
607 :     xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2);
608 :     ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data);
609 :     ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data);
610 :     } else {
611 :     xcf = 4*mvs.x; ycf = 4*mvs.y;
612 :     xcb = 4*b_mvs.x; ycb = 4*b_mvs.y;
613 :     ReferenceF = GetReference(mvs.x, mvs.y, data);
614 :     ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data);
615 :     }
616 : syskin 768
617 : syskin 628 sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth);
618 : syskin 768 sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10;
619 : chl 96
620 : syskin 801 if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf],
621 :     (ycf >> 3) + roundtab_76[ycf & 0xf],
622 :     (xcb >> 3) + roundtab_76[xcb & 0xf],
623 :     (ycb >> 3) + roundtab_76[ycb & 0xf], data);
624 :    
625 : chl 530 if (sad < *(data->iMinSAD)) {
626 :     *(data->iMinSAD) = sad;
627 :     data->currentMV->x = x; data->currentMV->y = y;
628 : syskin 801 *dir = Direction;
629 :     }
630 : chl 530 }
631 : chl 96
632 : Isibaar 539 /* CHECK_CANDIATE FUNCTIONS END */
633 : chl 96
634 : chl 530 /* MAINSEARCH FUNCTIONS START */
635 : edgomez 195
636 : chl 530 static void
637 :     AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection)
638 : chl 181 {
639 :    
640 :     /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
641 :    
642 : syskin 712 int iDirection;
643 : edgomez 195
644 : syskin 712 for(;;) { //forever
645 :     iDirection = 0;
646 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
647 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
648 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
649 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
650 : chl 181
651 : syskin 712 /* now we're doing diagonal checks near our candidate */
652 : chl 181
653 : syskin 768 if (iDirection) { //if anything found
654 : syskin 712 bDirection = iDirection;
655 :     iDirection = 0;
656 :     x = data->currentMV->x; y = data->currentMV->y;
657 :     if (bDirection & 3) { //our candidate is left or right
658 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
659 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
660 :     } else { // what remains here is up or down
661 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
662 : syskin 748 CHECK_CANDIDATE(x - iDiamondSize, y, 1);
663 : syskin 712 }
664 : chl 181
665 : syskin 712 if (iDirection) {
666 :     bDirection += iDirection;
667 : syskin 748 x = data->currentMV->x; y = data->currentMV->y;
668 : chl 181 }
669 : syskin 712 } else { //about to quit, eh? not so fast....
670 :     switch (bDirection) {
671 :     case 2:
672 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
673 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
674 :     break;
675 :     case 1:
676 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
677 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
678 :     break;
679 :     case 2 + 4:
680 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
681 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
682 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
683 :     break;
684 :     case 4:
685 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
686 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
687 :     break;
688 :     case 8:
689 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
690 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
691 :     break;
692 :     case 1 + 4:
693 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
694 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
695 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
696 :     break;
697 :     case 2 + 8:
698 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
699 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
700 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
701 :     break;
702 :     case 1 + 8:
703 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
704 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
705 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
706 :     break;
707 :     default: //1+2+4+8 == we didn't find anything at all
708 :     CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4);
709 :     CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8);
710 :     CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4);
711 :     CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8);
712 :     break;
713 :     }
714 :     if (!iDirection) break; //ok, the end. really
715 :     bDirection = iDirection;
716 :     x = data->currentMV->x; y = data->currentMV->y;
717 : chl 181 }
718 : syskin 712 }
719 : chl 181 }
720 :    
721 : chl 530 static void
722 :     SquareSearch(int x, int y, const SearchData * const data, int bDirection)
723 : chl 326 {
724 : chl 530 int iDirection;
725 : chl 326
726 : chl 530 do {
727 :     iDirection = 0;
728 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64);
729 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128);
730 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32);
731 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128);
732 :     if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64);
733 :     if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128);
734 :     if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128);
735 :     if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128);
736 : chl 326
737 : chl 530 bDirection = iDirection;
738 :     x = data->currentMV->x; y = data->currentMV->y;
739 :     } while (iDirection);
740 : chl 346 }
741 :    
742 : chl 530 static void
743 :     DiamondSearch(int x, int y, const SearchData * const data, int bDirection)
744 : chl 346 {
745 :    
746 : chl 181 /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */
747 :    
748 : syskin 712 int iDirection;
749 : edgomez 195
750 : syskin 712 do {
751 :     iDirection = 0;
752 :     if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1);
753 :     if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2);
754 :     if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4);
755 :     if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8);
756 : chl 181
757 : syskin 712 /* now we're doing diagonal checks near our candidate */
758 : chl 181
759 : syskin 712 if (iDirection) { //checking if anything found
760 :     bDirection = iDirection;
761 :     iDirection = 0;
762 :     x = data->currentMV->x; y = data->currentMV->y;
763 :     if (bDirection & 3) { //our candidate is left or right
764 :     CHECK_CANDIDATE(x, y + iDiamondSize, 8);
765 :     CHECK_CANDIDATE(x, y - iDiamondSize, 4);
766 :     } else { // what remains here is up or down
767 :     CHECK_CANDIDATE(x + iDiamondSize, y, 2);
768 : syskin 748 CHECK_CANDIDATE(x - iDiamondSize, y, 1);
769 : chl 181 }
770 : syskin 712 bDirection += iDirection;
771 :     x = data->currentMV->x; y = data->currentMV->y;
772 : chl 181 }
773 : syskin 712 }
774 :     while (iDirection);
775 : chl 181 }
776 :    
777 : chl 530 /* MAINSEARCH FUNCTIONS END */
778 : chl 181
779 : chl 530 static void
780 : syskin 663 SubpelRefine(const SearchData * const data)
781 : chl 96 {
782 : syskin 663 /* Do a half-pel or q-pel refinement */
783 : syskin 768 const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV;
784 :     int iDirection; //only needed because macro expects it
785 : chl 96
786 : syskin 768 CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0);
787 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0);
788 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0);
789 :     CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0);
790 :     CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0);
791 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0);
792 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0);
793 :     CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0);
794 : Isibaar 580 }
795 :    
796 : chl 530 static __inline int
797 :     SkipDecisionP(const IMAGE * current, const IMAGE * reference,
798 :     const int x, const int y,
799 : syskin 768 const uint32_t stride, const uint32_t iQuant, int rrv)
800 : chl 96
801 : Isibaar 3 {
802 : syskin 768 if(!rrv) {
803 :     uint32_t sadC = sad8(current->u + x*8 + y*stride*8,
804 :     reference->u + x*8 + y*stride*8, stride);
805 : syskin 704 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
806 : syskin 768 sadC += sad8(current->v + (x + y*stride)*8,
807 :     reference->v + (x + y*stride)*8, stride);
808 : syskin 704 if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
809 :     return 1;
810 : syskin 768
811 :     } else {
812 :     uint32_t sadC = sad16(current->u + x*16 + y*stride*16,
813 :     reference->u + x*16 + y*stride*16, stride, 256*4096);
814 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
815 :     sadC += sad16(current->v + (x + y*stride)*16,
816 :     reference->v + (x + y*stride)*16, stride, 256*4096);
817 :     if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0;
818 :     return 1;
819 : syskin 704 }
820 : chl 530 }
821 : edgomez 195
822 : chl 530 static __inline void
823 :     SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad)
824 :     {
825 :     pMB->mode = MODE_NOT_CODED;
826 : syskin 768 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = zeroMV;
827 :     pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = zeroMV;
828 : chl 530 pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;
829 : Isibaar 3 }
830 :    
831 : syskin 628 bool
832 : chl 530 MotionEstimation(MBParam * const pParam,
833 :     FRAMEINFO * const current,
834 :     FRAMEINFO * const reference,
835 :     const IMAGE * const pRefH,
836 :     const IMAGE * const pRefV,
837 :     const IMAGE * const pRefHV,
838 :     const uint32_t iLimit)
839 : Isibaar 3 {
840 : chl 530 MACROBLOCK *const pMBs = current->mbs;
841 :     const IMAGE *const pCurrent = &current->image;
842 :     const IMAGE *const pRef = &reference->image;
843 : Isibaar 3
844 : syskin 704 uint32_t mb_width = pParam->mb_width;
845 :     uint32_t mb_height = pParam->mb_height;
846 : syskin 768 const uint32_t iEdgedWidth = pParam->edged_width;
847 : suxen_drol 702
848 : chl 530 uint32_t x, y;
849 :     uint32_t iIntra = 0;
850 : syskin 628 int32_t InterBias, quant = current->quant, sad00;
851 : edgomez 195
852 : Isibaar 539 // some pre-initialized thingies for SearchP
853 : syskin 712 int32_t temp[8];
854 : Isibaar 539 VECTOR currentMV[5];
855 : Isibaar 580 VECTOR currentQMV[5];
856 : Isibaar 539 int32_t iMinSAD[5];
857 :     SearchData Data;
858 : syskin 704 memset(&Data, 0, sizeof(SearchData));
859 : syskin 768 Data.iEdgedWidth = iEdgedWidth;
860 : Isibaar 539 Data.currentMV = currentMV;
861 : Isibaar 580 Data.currentQMV = currentQMV;
862 : Isibaar 539 Data.iMinSAD = iMinSAD;
863 :     Data.temp = temp;
864 :     Data.iFcode = current->fcode;
865 : Isibaar 580 Data.rounding = pParam->m_rounding_type;
866 : syskin 628 Data.qpel = pParam->m_quarterpel;
867 : syskin 801 Data.chroma = current->motion_flags & PMV_CHROMA16;
868 : syskin 704 Data.rrv = current->global_flags & XVID_REDUCED;
869 : Isibaar 539
870 : syskin 704 if ((current->global_flags & XVID_REDUCED)) {
871 : suxen_drol 702 mb_width = (pParam->width + 31) / 32;
872 :     mb_height = (pParam->height + 31) / 32;
873 : syskin 704 Data.qpel = Data.chroma = 0;
874 : suxen_drol 702 }
875 :    
876 : syskin 754 Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose)
877 : chl 530 if (sadInit) (*sadInit) ();
878 : edgomez 195
879 : suxen_drol 702 for (y = 0; y < mb_height; y++) {
880 :     for (x = 0; x < mb_width; x++) {
881 : chl 530 MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
882 : syskin 768
883 :     if (!Data.rrv) pMB->sad16 =
884 :     sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
885 :     pRef->y + (x + y * iEdgedWidth) * 16,
886 : syskin 704 pParam->edged_width, pMB->sad8 );
887 : syskin 768
888 : syskin 748 else pMB->sad16 =
889 : syskin 768 sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
890 :     pRef->y + (x + y * iEdgedWidth) * 32,
891 : chl 530 pParam->edged_width, pMB->sad8 );
892 : edgomez 195
893 : syskin 628 if (Data.chroma) {
894 : syskin 768 Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
895 :     pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
896 :     + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
897 :     pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
898 :     pMB->sad16 += Data.temp[7];
899 : syskin 628 }
900 :    
901 : syskin 768 sad00 = pMB->sad16;
902 : syskin 628
903 : chl 530 if (!(current->global_flags & XVID_LUMIMASKING)) {
904 :     pMB->dquant = NO_CHANGE;
905 : syskin 601 } else {
906 : syskin 574 if (pMB->dquant != NO_CHANGE) {
907 :     quant += DQtab[pMB->dquant];
908 :     if (quant > 31) quant = 31;
909 :     else if (quant < 1) quant = 1;
910 :     }
911 : syskin 601 }
912 : syskin 768 pMB->quant = current->quant;
913 : chl 181
914 : syskin 628 //initial skip decision
915 : chl 769 /* no early skip for GMC (global vector = skip vector is unknown!) */
916 :     if (!(current->global_flags & XVID_GMC)) { /* no fast SKIP for S(GMC)-VOPs */
917 : syskin 801 if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) )
918 : syskin 768 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
919 : chl 619 SkipMacroblockP(pMB, sad00);
920 :     continue;
921 :     }
922 :     }
923 : syskin 768
924 : syskin 628 SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
925 : chl 530 y, current->motion_flags, pMB->quant,
926 : Isibaar 539 &Data, pParam, pMBs, reference->mbs,
927 : chl 530 current->global_flags & XVID_INTER4V, pMB);
928 : edgomez 195
929 : chl 530 /* final skip decision, a.k.a. "the vector you found, really that good?" */
930 : chl 769 if (!(current->global_flags & XVID_GMC)) {
931 : chl 619 if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP)
932 : syskin 704 && ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) )
933 : syskin 768 if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
934 : chl 619 SkipMacroblockP(pMB, sad00);
935 :     continue;
936 :     }
937 :     }
938 : syskin 768
939 : chl 530 /* finally, intra decision */
940 : Isibaar 3
941 : chl 530 InterBias = MV16_INTER_BIAS;
942 : syskin 768 if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work
943 : chl 530 if (y != 0)
944 : syskin 639 if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
945 : chl 530 if (x != 0)
946 : syskin 639 if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;
947 : syskin 768
948 : syskin 628 if (Data.chroma) InterBias += 50; // to compensate bigger SAD
949 : syskin 768 if (Data.rrv) InterBias *= 4;
950 : Isibaar 3
951 : syskin 768 if (InterBias < pMB->sad16) {
952 : syskin 704 int32_t deviation;
953 : syskin 768 if (!Data.rrv)
954 :     deviation = dev16(pCurrent->y + (x + y * iEdgedWidth) * 16, iEdgedWidth);
955 :     else {
956 :     deviation = dev16(pCurrent->y + (x + y * iEdgedWidth) * 32, iEdgedWidth)
957 :     + dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16, iEdgedWidth)
958 :     + dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16 * iEdgedWidth, iEdgedWidth)
959 :     + dev16(pCurrent->y + (x + y * iEdgedWidth) * 32 + 16 * (iEdgedWidth+1), iEdgedWidth);
960 :     }
961 :     if (deviation < (pMB->sad16 - InterBias)) {
962 : syskin 754 if (++iIntra >= iLimit) return 1;
963 : syskin 768 SkipMacroblockP(pMB, 0); //same thing
964 : chl 530 pMB->mode = MODE_INTRA;
965 :     }
966 :     }
967 :     }
968 : Isibaar 3 }
969 : chl 619
970 : chl 770 if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */
971 : chl 769 {
972 :     current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV);
973 :     }
974 :    
975 : syskin 628 return 0;
976 : chl 530 }
977 : Isibaar 3
978 : edgomez 195
979 : chl 530 static __inline int
980 :     make_mask(const VECTOR * const pmv, const int i)
981 :     {
982 : Isibaar 539 int mask = 255, j;
983 : chl 530 for (j = 0; j < i; j++) {
984 :     if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already
985 :     if (pmv[i].x == pmv[j].x) {
986 : syskin 768 if (pmv[i].y == pmv[j].y + iDiamondSize) mask &= ~4;
987 :     else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8;
988 : syskin 748 } else
989 : chl 530 if (pmv[i].y == pmv[j].y) {
990 : syskin 768 if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1;
991 :     else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2;
992 : chl 530 }
993 : Isibaar 3 }
994 : chl 530 return mask;
995 :     }
996 : edgomez 195
997 : syskin 748 static __inline void
998 : syskin 704 PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
999 :     int iHcount, const MACROBLOCK * const prevMB, int rrv)
1000 : chl 530 {
1001 : edgomez 195
1002 : chl 530 //this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself
1003 : syskin 704 if (rrv) { iWcount /= 2; iHcount /= 2; }
1004 : edgomez 195
1005 : syskin 704 if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour
1006 : chl 530 pmv[5].x = EVEN(pmv[3].x);
1007 : syskin 748 pmv[5].y = EVEN(pmv[3].y);
1008 : Isibaar 539 } else pmv[5].x = pmv[5].y = 0;
1009 : chl 169
1010 : chl 530 if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour
1011 :     else pmv[3].x = pmv[3].y = 0;
1012 : Isibaar 3
1013 : chl 530 if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour
1014 : syskin 768 else pmv[4].x = pmv[4].y = 0;
1015 : chl 169
1016 : chl 530 // [1] median prediction
1017 : syskin 767 pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y);
1018 : chl 169
1019 : chl 530 pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask
1020 : chl 169
1021 : chl 530 pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame
1022 :     pmv[2].y = EVEN(prevMB->mvs[0].y);
1023 : chl 169
1024 : syskin 704 if ((x < iWcount-1) && (y < iHcount-1)) {
1025 : chl 530 pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame
1026 : syskin 748 pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y);
1027 : Isibaar 539 } else pmv[6].x = pmv[6].y = 0;
1028 : syskin 704
1029 :     if (rrv) {
1030 :     int i;
1031 :     for (i = 0; i < 7; i++) {
1032 : syskin 768 pmv[i].x = RRV_MV_SCALEUP(pmv[i].x);
1033 : syskin 767 pmv[i].y = RRV_MV_SCALEUP(pmv[i].y);
1034 : syskin 704 }
1035 :     }
1036 : chl 530 }
1037 : chl 169
1038 : chl 530 static void
1039 : syskin 628 SearchP(const IMAGE * const pRef,
1040 : chl 530 const uint8_t * const pRefH,
1041 :     const uint8_t * const pRefV,
1042 :     const uint8_t * const pRefHV,
1043 :     const IMAGE * const pCur,
1044 :     const int x,
1045 :     const int y,
1046 :     const uint32_t MotionFlags,
1047 :     const uint32_t iQuant,
1048 : Isibaar 539 SearchData * const Data,
1049 : chl 530 const MBParam * const pParam,
1050 :     const MACROBLOCK * const pMBs,
1051 :     const MACROBLOCK * const prevMBs,
1052 :     int inter4v,
1053 :     MACROBLOCK * const pMB)
1054 :     {
1055 : chl 169
1056 : chl 530 int i, iDirection = 255, mask, threshA;
1057 : Isibaar 539 VECTOR pmv[7];
1058 : Isibaar 3
1059 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1060 :     pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1061 : Isibaar 3
1062 : syskin 768 get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp);
1063 : syskin 628
1064 : syskin 768 Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache
1065 :     i = Data->rrv ? 2 : 1;
1066 : syskin 704 Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i;
1067 :     Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1068 :     Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1069 : Isibaar 3
1070 : syskin 704 Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
1071 :     Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
1072 :     Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
1073 :     Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
1074 :     Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1075 :     Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
1076 :    
1077 : syskin 601 Data->lambda16 = lambda_vec16[iQuant];
1078 : Isibaar 606 Data->lambda8 = lambda_vec8[iQuant];
1079 : syskin 663 Data->qpel_precision = 0;
1080 : chl 169
1081 : edgomez 568 if (pMB->dquant != NO_CHANGE) inter4v = 0;
1082 : edgomez 195
1083 : syskin 768 for(i = 0; i < 5; i++)
1084 : Isibaar 580 Data->currentMV[i].x = Data->currentMV[i].y = 0;
1085 : Isibaar 539
1086 : syskin 768 if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
1087 : syskin 663 else Data->predMV = pmv[0];
1088 : Isibaar 606
1089 : syskin 724 i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
1090 : syskin 768 Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
1091 :     Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
1092 : Isibaar 539 Data->iMinSAD[2] = pMB->sad8[1];
1093 :     Data->iMinSAD[3] = pMB->sad8[2];
1094 :     Data->iMinSAD[4] = pMB->sad8[3];
1095 :    
1096 : syskin 768 if (x | y) {
1097 : edgomez 568 threshA = Data->temp[0]; // that's when we keep this SAD atm
1098 : chl 530 if (threshA < 512) threshA = 512;
1099 : syskin 768 else if (threshA > 1024) threshA = 1024;
1100 :     } else threshA = 512;
1101 : Isibaar 3
1102 : chl 530 PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
1103 : syskin 704 prevMBs + x + y * pParam->mb_width, Data->rrv);
1104 : edgomez 195
1105 : syskin 768 if (!Data->rrv) {
1106 :     if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16;
1107 :     else CheckCandidate = CheckCandidate16no4v; //for extra speed
1108 :     } else CheckCandidate = CheckCandidate32;
1109 : edgomez 568
1110 : syskin 768 /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/
1111 : chl 530
1112 :     for (i = 1; i < 7; i++) {
1113 :     if (!(mask = make_mask(pmv, i)) ) continue;
1114 : suxen_drol 819 CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1115 : Isibaar 539 if (Data->iMinSAD[0] <= threshA) break;
1116 : chl 530 }
1117 :    
1118 : Isibaar 539 if ((Data->iMinSAD[0] <= threshA) ||
1119 :     (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
1120 : syskin 768 (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16)))
1121 : chl 530 inter4v = 0;
1122 : syskin 768 else {
1123 : Isibaar 3
1124 : syskin 748 MainSearchFunc * MainSearchPtr;
1125 : Isibaar 539 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1126 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1127 :     else MainSearchPtr = DiamondSearch;
1128 : chl 181
1129 : Isibaar 539 (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1130 : Isibaar 3
1131 : chl 530 /* extended search, diamond starting in 0,0 and in prediction.
1132 :     note that this search is/might be done in halfpel positions,
1133 :     which makes it more different than the diamond above */
1134 : chl 259
1135 : Isibaar 539 if (MotionFlags & PMV_EXTSEARCH16) {
1136 :     int32_t bSAD;
1137 :     VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
1138 : syskin 704 if (Data->rrv) {
1139 :     startMV.x = RRV_MV_SCALEUP(startMV.x);
1140 :     startMV.y = RRV_MV_SCALEUP(startMV.y);
1141 : syskin 748 } else
1142 : syskin 704 if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel?
1143 :     startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y);
1144 : Isibaar 539 if (!(MVequal(startMV, backupMV))) {
1145 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1146 : Isibaar 3
1147 : suxen_drol 819 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1148 : Isibaar 539 (*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
1149 :     if (bSAD < Data->iMinSAD[0]) {
1150 :     Data->currentMV[0] = backupMV;
1151 :     Data->iMinSAD[0] = bSAD; }
1152 :     }
1153 : Isibaar 3
1154 : Isibaar 539 backupMV = Data->currentMV[0];
1155 : syskin 768 if (MotionFlags & PMV_HALFPELREFINE16 && !Data->rrv) startMV.x = startMV.y = 1;
1156 :     else startMV.x = startMV.y = 0;
1157 : Isibaar 539 if (!(MVequal(startMV, backupMV))) {
1158 :     bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;
1159 : edgomez 195
1160 : suxen_drol 819 CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data);
1161 : Isibaar 539 (*MainSearchPtr)(startMV.x, startMV.y, Data, 255);
1162 :     if (bSAD < Data->iMinSAD[0]) {
1163 :     Data->currentMV[0] = backupMV;
1164 :     Data->iMinSAD[0] = bSAD; }
1165 :     }
1166 : Isibaar 3 }
1167 :     }
1168 :    
1169 : syskin 663 if (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(Data);
1170 : Isibaar 3
1171 : Isibaar 580 for(i = 0; i < 5; i++) {
1172 :     Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors
1173 :     Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
1174 :     }
1175 :    
1176 : syskin 768 if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE16) {
1177 : syskin 663 Data->qpel_precision = 1;
1178 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1179 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1180 : syskin 601
1181 : syskin 663 SubpelRefine(Data);
1182 : Isibaar 580 }
1183 :    
1184 : syskin 768 if (Data->iMinSAD[0] < (int32_t)iQuant * 30) inter4v = 0;
1185 : Isibaar 539 if (inter4v) {
1186 :     SearchData Data8;
1187 : syskin 768 memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data
1188 :    
1189 : Isibaar 539 Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
1190 :     Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
1191 :     Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
1192 :     Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);
1193 : syskin 768
1194 : syskin 628 if (Data->chroma) {
1195 : syskin 768 int sumx = 0, sumy = 0;
1196 :     const int div = 1 + Data->qpel;
1197 :     const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs;
1198 : syskin 628
1199 : syskin 768 for (i = 0; i < 4; i++) {
1200 :     sumx += mv[i].x / div;
1201 :     sumy += mv[i].y / div;
1202 : syskin 663 }
1203 : syskin 768
1204 : syskin 753 Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf],
1205 :     (sumy >> 3) + roundtab_76[sumy & 0xf], Data);
1206 : syskin 628 }
1207 : Isibaar 539 }
1208 : Isibaar 3
1209 : syskin 704 if (Data->rrv) {
1210 :     Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x);
1211 :     Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y);
1212 :     }
1213 : syskin 740
1214 : chl 530 if (!(inter4v) ||
1215 : syskin 748 (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
1216 : Isibaar 539 Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) {
1217 : chl 530 // INTER MODE
1218 :     pMB->mode = MODE_INTER;
1219 : syskin 768 pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];
1220 :     pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0];
1221 : Isibaar 3
1222 : syskin 768 if(Data->qpel) {
1223 : syskin 663 pMB->qmvs[0] = pMB->qmvs[1]
1224 :     = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
1225 :     pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
1226 :     pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
1227 : syskin 628 } else {
1228 : Isibaar 580 pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
1229 :     pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
1230 :     }
1231 : edgomez 195 } else {
1232 : syskin 748 // INTER4V MODE; all other things are already set in Search8
1233 : chl 530 pMB->mode = MODE_INTER4V;
1234 : syskin 748 pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] +
1235 : Isibaar 539 Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant;
1236 : edgomez 78 }
1237 : Isibaar 3 }
1238 :    
1239 : chl 530 static void
1240 :     Search8(const SearchData * const OldData,
1241 :     const int x, const int y,
1242 :     const uint32_t MotionFlags,
1243 :     const MBParam * const pParam,
1244 :     MACROBLOCK * const pMB,
1245 :     const MACROBLOCK * const pMBs,
1246 : Isibaar 539 const int block,
1247 :     SearchData * const Data)
1248 : chl 345 {
1249 : syskin 704 int i = 0;
1250 : Isibaar 539 Data->iMinSAD = OldData->iMinSAD + 1 + block;
1251 :     Data->currentMV = OldData->currentMV + 1 + block;
1252 : Isibaar 580 Data->currentQMV = OldData->currentQMV + 1 + block;
1253 : chl 345
1254 : syskin 768 if(Data->qpel) {
1255 : syskin 704 Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1256 : syskin 724 if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y,
1257 :     Data->predMV, Data->iFcode, 0, 0);
1258 : Isibaar 606 } else {
1259 : syskin 704 Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block);
1260 : syskin 768 if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y,
1261 :     Data->predMV, Data->iFcode, 0, Data->rrv);
1262 : Isibaar 606 }
1263 : chl 345
1264 : syskin 768 *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10;
1265 : syskin 704
1266 : syskin 768 if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) {
1267 : syskin 704 if (Data->rrv) i = 2; else i = 1;
1268 : chl 345
1269 : syskin 768 Data->Ref = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1270 :     Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1271 :     Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1272 :     Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1273 : chl 345
1274 : syskin 768 Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1));
1275 : syskin 663 Data->qpel_precision = 0;
1276 : chl 345
1277 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1278 : syskin 768 pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv);
1279 : syskin 712
1280 : syskin 768 if (!Data->rrv) CheckCandidate = CheckCandidate8;
1281 :     else CheckCandidate = CheckCandidate16no4v;
1282 : syskin 712
1283 : chl 530 if (MotionFlags & PMV_EXTSEARCH8) {
1284 : Isibaar 580 int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1285 : syskin 768
1286 : chl 530 MainSearchFunc *MainSearchPtr;
1287 :     if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch;
1288 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch;
1289 :     else MainSearchPtr = DiamondSearch;
1290 : chl 345
1291 : Isibaar 580 (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255);
1292 : chl 345
1293 : syskin 601 if(*(Data->iMinSAD) < temp_sad) {
1294 : Isibaar 580 Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1295 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1296 :     }
1297 :     }
1298 :    
1299 :     if (MotionFlags & PMV_HALFPELREFINE8) {
1300 :     int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD
1301 :    
1302 : syskin 663 SubpelRefine(Data); // perform halfpel refine of current best vector
1303 : Isibaar 580
1304 :     if(*(Data->iMinSAD) < temp_sad) { // we have found a better match
1305 :     Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector
1306 :     Data->currentQMV->y = 2 * Data->currentMV->y;
1307 :     }
1308 :     }
1309 :    
1310 : syskin 768 if (Data->qpel && MotionFlags & PMV_QUARTERPELREFINE8) {
1311 :     Data->qpel_precision = 1;
1312 :     get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8,
1313 :     pParam->width, pParam->height, Data->iFcode, 1, 0);
1314 :     SubpelRefine(Data);
1315 : Isibaar 580 }
1316 : chl 345 }
1317 : chl 530
1318 : syskin 704 if (Data->rrv) {
1319 :     Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x);
1320 :     Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y);
1321 :     }
1322 : syskin 768
1323 : syskin 712 if(Data->qpel) {
1324 : syskin 663 pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x;
1325 :     pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y;
1326 : syskin 768 pMB->qmvs[block] = *Data->currentQMV;
1327 : syskin 704 } else {
1328 : Isibaar 580 pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x;
1329 :     pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y;
1330 :     }
1331 :    
1332 : syskin 768 pMB->mvs[block] = *Data->currentMV;
1333 :     pMB->sad8[block] = 4 * *Data->iMinSAD;
1334 : chl 345 }
1335 :    
1336 : syskin 768 /* motion estimation for B-frames */
1337 : chl 345
1338 : chl 530 static __inline VECTOR
1339 :     ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode)
1340 : Isibaar 3 {
1341 : chl 530 /* the stupidiest function ever */
1342 : syskin 768 return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]);
1343 : Isibaar 3 }
1344 :    
1345 : chl 530 static void __inline
1346 :     PreparePredictionsBF(VECTOR * const pmv, const int x, const int y,
1347 :     const uint32_t iWcount,
1348 :     const MACROBLOCK * const pMB,
1349 :     const uint32_t mode_curr)
1350 : Isibaar 3 {
1351 :    
1352 : chl 530 // [0] is prediction
1353 :     pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y);
1354 : Isibaar 3
1355 : chl 530 pmv[1].x = pmv[1].y = 0; // [1] is zero
1356 : Isibaar 3
1357 : chl 530 pmv[2] = ChoosePred(pMB, mode_curr);
1358 :     pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y);
1359 : edgomez 195
1360 : chl 530 if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour
1361 :     pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr);
1362 : syskin 748 pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y);
1363 : Isibaar 539 } else pmv[3].x = pmv[3].y = 0;
1364 : Isibaar 3
1365 : chl 530 if (y != 0) {
1366 :     pmv[4] = ChoosePred(pMB-iWcount, mode_curr);
1367 :     pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y);
1368 :     } else pmv[4].x = pmv[4].y = 0;
1369 : Isibaar 3
1370 : chl 530 if (x != 0) {
1371 :     pmv[5] = ChoosePred(pMB-1, mode_curr);
1372 :     pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y);
1373 :     } else pmv[5].x = pmv[5].y = 0;
1374 : Isibaar 3
1375 : syskin 768 if (x != 0 && y != 0) {
1376 : chl 530 pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr);
1377 : syskin 767 pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y);
1378 : chl 530 } else pmv[6].x = pmv[6].y = 0;
1379 :     }
1380 : chl 181
1381 : edgomez 170
1382 : syskin 768 /* search backward or forward */
1383 : chl 530 static void
1384 : syskin 801 SearchBF( const IMAGE * const pRef,
1385 : chl 530 const uint8_t * const pRefH,
1386 :     const uint8_t * const pRefV,
1387 :     const uint8_t * const pRefHV,
1388 :     const IMAGE * const pCur,
1389 :     const int x, const int y,
1390 :     const uint32_t MotionFlags,
1391 :     const uint32_t iFcode,
1392 :     const MBParam * const pParam,
1393 :     MACROBLOCK * const pMB,
1394 :     const VECTOR * const predMV,
1395 :     int32_t * const best_sad,
1396 : h 545 const int32_t mode_current,
1397 :     SearchData * const Data)
1398 : chl 530 {
1399 : Isibaar 3
1400 : syskin 740 int i, iDirection = 255, mask;
1401 : h 545 VECTOR pmv[7];
1402 : chl 530 MainSearchFunc *MainSearchPtr;
1403 : h 545 *Data->iMinSAD = MV_MAX_ERROR;
1404 :     Data->iFcode = iFcode;
1405 : syskin 663 Data->qpel_precision = 0;
1406 : syskin 801 Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache
1407 : Isibaar 3
1408 : syskin 801 Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16;
1409 : syskin 768 Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16;
1410 :     Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16;
1411 :     Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16;
1412 : syskin 801 Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8;
1413 :     Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8;
1414 : Isibaar 3
1415 : h 545 Data->predMV = *predMV;
1416 : edgomez 195
1417 : h 545 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1418 : syskin 712 pParam->width, pParam->height, iFcode - Data->qpel, 0, 0);
1419 : Isibaar 3
1420 : h 545 pmv[0] = Data->predMV;
1421 : syskin 658 if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; }
1422 : syskin 768
1423 : h 545 PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current);
1424 : edgomez 195
1425 : h 545 Data->currentMV->x = Data->currentMV->y = 0;
1426 : chl 530 CheckCandidate = CheckCandidate16no4v;
1427 : chl 181
1428 : chl 530 // main loop. checking all predictions
1429 : syskin 740 for (i = 0; i < 7; i++) {
1430 : chl 530 if (!(mask = make_mask(pmv, i)) ) continue;
1431 : h 545 CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data);
1432 : edgomez 78 }
1433 : Isibaar 3
1434 : syskin 768 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1435 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1436 : chl 530 else MainSearchPtr = DiamondSearch;
1437 : chl 96
1438 : syskin 740 (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection);
1439 : chl 169
1440 : syskin 663 SubpelRefine(Data);
1441 : syskin 768
1442 : syskin 740 if (Data->qpel && *Data->iMinSAD < *best_sad + 300) {
1443 : syskin 658 Data->currentQMV->x = 2*Data->currentMV->x;
1444 :     Data->currentQMV->y = 2*Data->currentMV->y;
1445 : syskin 663 Data->qpel_precision = 1;
1446 : syskin 712 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1447 :     pParam->width, pParam->height, iFcode, 1, 0);
1448 : syskin 663 SubpelRefine(Data);
1449 : syskin 658 }
1450 : chl 169
1451 : chl 530 // three bits are needed to code backward mode. four for forward
1452 : syskin 740
1453 : syskin 768 if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16;
1454 :     else *Data->iMinSAD += 3 * Data->lambda16;
1455 : chl 169
1456 : h 545 if (*Data->iMinSAD < *best_sad) {
1457 :     *best_sad = *Data->iMinSAD;
1458 : chl 530 pMB->mode = mode_current;
1459 : syskin 658 if (Data->qpel) {
1460 :     pMB->pmvs[0].x = Data->currentQMV->x - predMV->x;
1461 :     pMB->pmvs[0].y = Data->currentQMV->y - predMV->y;
1462 : syskin 748 if (mode_current == MODE_FORWARD)
1463 : syskin 658 pMB->qmvs[0] = *Data->currentQMV;
1464 : syskin 748 else
1465 : syskin 658 pMB->b_qmvs[0] = *Data->currentQMV;
1466 :     } else {
1467 :     pMB->pmvs[0].x = Data->currentMV->x - predMV->x;
1468 :     pMB->pmvs[0].y = Data->currentMV->y - predMV->y;
1469 :     }
1470 : syskin 739 if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV;
1471 :     else pMB->b_mvs[0] = *Data->currentMV;
1472 :     }
1473 : syskin 658
1474 : syskin 768 if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV;
1475 :     else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search
1476 : chl 530 }
1477 : chl 169
1478 : syskin 748 static void
1479 : syskin 712 SkipDecisionB(const IMAGE * const pCur,
1480 : syskin 767 const IMAGE * const f_Ref,
1481 :     const IMAGE * const b_Ref,
1482 :     MACROBLOCK * const pMB,
1483 :     const uint32_t x, const uint32_t y,
1484 :     const SearchData * const Data)
1485 : syskin 712 {
1486 : syskin 767 int dx = 0, dy = 0, b_dx = 0, b_dy = 0;
1487 : syskin 768 int32_t sum;
1488 : syskin 767 const int div = 1 + Data->qpel;
1489 :     int k;
1490 : syskin 768 const uint32_t stride = Data->iEdgedWidth/2;
1491 : syskin 712 //this is not full chroma compensation, only it's fullpel approximation. should work though
1492 :    
1493 : syskin 767 for (k = 0; k < 4; k++) {
1494 :     dy += Data->directmvF[k].y / div;
1495 :     dx += Data->directmvF[0].x / div;
1496 :     b_dy += Data->directmvB[0].y / div;
1497 :     b_dx += Data->directmvB[0].x / div;
1498 : syskin 712 }
1499 :    
1500 :     dy = (dy >> 3) + roundtab_76[dy & 0xf];
1501 :     dx = (dx >> 3) + roundtab_76[dx & 0xf];
1502 :     b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf];
1503 :     b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf];
1504 :    
1505 : syskin 768 sum = sad8bi(pCur->u + 8 * x + 8 * y * stride,
1506 :     f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2,
1507 :     b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1508 :     stride);
1509 : syskin 767
1510 : syskin 768 if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip
1511 : syskin 767
1512 : syskin 768 sum += sad8bi(pCur->v + 8*x + 8 * y * stride,
1513 :     f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2,
1514 :     b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2,
1515 :     stride);
1516 : syskin 712
1517 : syskin 768 if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped
1518 : syskin 712 }
1519 :    
1520 :     static __inline uint32_t
1521 : Isibaar 539 SearchDirect(const IMAGE * const f_Ref,
1522 : chl 530 const uint8_t * const f_RefH,
1523 :     const uint8_t * const f_RefV,
1524 :     const uint8_t * const f_RefHV,
1525 : Isibaar 539 const IMAGE * const b_Ref,
1526 : chl 530 const uint8_t * const b_RefH,
1527 :     const uint8_t * const b_RefV,
1528 :     const uint8_t * const b_RefHV,
1529 :     const IMAGE * const pCur,
1530 :     const int x, const int y,
1531 :     const uint32_t MotionFlags,
1532 :     const int32_t TRB, const int32_t TRD,
1533 :     const MBParam * const pParam,
1534 :     MACROBLOCK * const pMB,
1535 :     const MACROBLOCK * const b_mb,
1536 : h 545 int32_t * const best_sad,
1537 :     SearchData * const Data)
1538 : edgomez 195
1539 : chl 530 {
1540 : h 545 int32_t skip_sad;
1541 : syskin 768 int k = (x + Data->iEdgedWidth*y) * 16;
1542 : chl 530 MainSearchFunc *MainSearchPtr;
1543 : Isibaar 3
1544 : h 545 *Data->iMinSAD = 256*4096;
1545 : syskin 768 Data->Ref = f_Ref->y + k;
1546 :     Data->RefH = f_RefH + k;
1547 :     Data->RefV = f_RefV + k;
1548 :     Data->RefHV = f_RefHV + k;
1549 :     Data->bRef = b_Ref->y + k;
1550 :     Data->bRefH = b_RefH + k;
1551 :     Data->bRefV = b_RefV + k;
1552 :     Data->bRefHV = b_RefHV + k;
1553 : syskin 801 Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1554 :     Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1555 :     Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8;
1556 :     Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8;
1557 : chl 140
1558 : syskin 768 k = Data->qpel ? 4 : 2;
1559 :     Data->max_dx = k * (pParam->width - x * 16);
1560 :     Data->max_dy = k * (pParam->height - y * 16);
1561 :     Data->min_dx = -k * (16 + x * 16);
1562 :     Data->min_dy = -k * (16 + y * 16);
1563 : chl 140
1564 : syskin 768 Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs;
1565 :     Data->qpel_precision = 0;
1566 : Isibaar 3
1567 : chl 530 for (k = 0; k < 4; k++) {
1568 : h 545 pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD);
1569 :     pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD;
1570 :     pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD);
1571 :     pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD;
1572 : Isibaar 3
1573 : syskin 768 if ( (pMB->b_mvs[k].x > Data->max_dx) | (pMB->b_mvs[k].x < Data->min_dx)
1574 :     | (pMB->b_mvs[k].y > Data->max_dy) | (pMB->b_mvs[k].y < Data->min_dy) ) {
1575 : Isibaar 3
1576 : Isibaar 539 *best_sad = 256*4096; // in that case, we won't use direct mode
1577 :     pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV"
1578 :     pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0;
1579 : syskin 712 return 256*4096;
1580 : Isibaar 539 }
1581 :     if (b_mb->mode != MODE_INTER4V) {
1582 :     pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0];
1583 :     pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0];
1584 : h 545 Data->directmvF[1] = Data->directmvF[2] = Data->directmvF[3] = Data->directmvF[0];
1585 :     Data->directmvB[1] = Data->directmvB[2] = Data->directmvB[3] = Data->directmvB[0];
1586 : Isibaar 539 break;
1587 :     }
1588 :     }
1589 : Isibaar 3
1590 : syskin 768 CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v;
1591 :    
1592 : suxen_drol 819 CheckCandidate(0, 0, 255, &k, Data);
1593 : Isibaar 3
1594 : syskin 712 // initial (fast) skip decision
1595 : syskin 801 if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) {
1596 :     //possible skip
1597 :     if (Data->chroma) {
1598 :     pMB->mode = MODE_DIRECT_NONE_MV;
1599 :     return *Data->iMinSAD; // skip.
1600 :     } else {
1601 :     SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data);
1602 :     if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip.
1603 :     }
1604 : Isibaar 539 }
1605 :    
1606 : h 545 skip_sad = *Data->iMinSAD;
1607 : edgomez 195
1608 : syskin 768 // DIRECT MODE DELTA VECTOR SEARCH.
1609 : chl 530 // This has to be made more effective, but at the moment I'm happy it's running at all
1610 : Isibaar 3
1611 : chl 530 if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch;
1612 :     else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch;
1613 :     else MainSearchPtr = DiamondSearch;
1614 : edgomez 195
1615 : h 545 (*MainSearchPtr)(0, 0, Data, 255);
1616 : Isibaar 3
1617 : syskin 663 SubpelRefine(Data);
1618 : Isibaar 3
1619 : h 545 *best_sad = *Data->iMinSAD;
1620 : Isibaar 3
1621 : syskin 768 if (Data->qpel || b_mb->mode == MODE_INTER4V) pMB->mode = MODE_DIRECT;
1622 : syskin 672 else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation
1623 : edgomez 195
1624 : h 545 pMB->pmvs[3] = *Data->currentMV;
1625 : edgomez 195
1626 : chl 530 for (k = 0; k < 4; k++) {
1627 : h 545 pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x;
1628 : syskin 658 pMB->b_mvs[k].x = ( (Data->currentMV->x == 0)
1629 : h 545 ? Data->directmvB[k].x
1630 : syskin 658 :pMB->mvs[k].x - Data->referencemv[k].x);
1631 : h 545 pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y);
1632 :     pMB->b_mvs[k].y = ((Data->currentMV->y == 0)
1633 :     ? Data->directmvB[k].y
1634 :     : pMB->mvs[k].y - Data->referencemv[k].y);
1635 : syskin 658 if (Data->qpel) {
1636 :     pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2;
1637 :     pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2;
1638 :     pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2;
1639 :     pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2;
1640 :     }
1641 :    
1642 : chl 530 if (b_mb->mode != MODE_INTER4V) {
1643 :     pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0];
1644 :     pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0];
1645 : syskin 658 pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0];
1646 :     pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0];
1647 : chl 530 break;
1648 :     }
1649 :     }
1650 : h 545 return skip_sad;
1651 : Isibaar 3 }
1652 : chl 96
1653 : syskin 712 static void
1654 : syskin 801 SearchInterpolate(const IMAGE * const f_Ref,
1655 : chl 530 const uint8_t * const f_RefH,
1656 :     const uint8_t * const f_RefV,
1657 :     const uint8_t * const f_RefHV,
1658 : syskin 801 const IMAGE * const b_Ref,
1659 : chl 530 const uint8_t * const b_RefH,
1660 :     const uint8_t * const b_RefV,
1661 :     const uint8_t * const b_RefHV,
1662 :     const IMAGE * const pCur,
1663 :     const int x, const int y,
1664 :     const uint32_t fcode,
1665 :     const uint32_t bcode,
1666 :     const uint32_t MotionFlags,
1667 :     const MBParam * const pParam,
1668 :     const VECTOR * const f_predMV,
1669 :     const VECTOR * const b_predMV,
1670 :     MACROBLOCK * const pMB,
1671 : h 545 int32_t * const best_sad,
1672 :     SearchData * const fData)
1673 : chl 530
1674 : chl 96 {
1675 :    
1676 : chl 530 int iDirection, i, j;
1677 : h 545 SearchData bData;
1678 : chl 96
1679 : syskin 712 fData->qpel_precision = 0;
1680 :     memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data
1681 :     *fData->iMinSAD = 4096*256;
1682 : syskin 739 bData.currentMV++; bData.currentQMV++;
1683 : h 545 fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode;
1684 : edgomez 195
1685 : syskin 712 i = (x + y * fData->iEdgedWidth) * 16;
1686 : syskin 801 bData.bRef = fData->Ref = f_Ref->y + i;
1687 : syskin 712 bData.bRefH = fData->RefH = f_RefH + i;
1688 :     bData.bRefV = fData->RefV = f_RefV + i;
1689 :     bData.bRefHV = fData->RefHV = f_RefHV + i;
1690 : syskin 801 bData.Ref = fData->bRef = b_Ref->y + i;
1691 : syskin 712 bData.RefH = fData->bRefH = b_RefH + i;
1692 :     bData.RefV = fData->bRefV = b_RefV + i;
1693 :     bData.RefHV = fData->bRefHV = b_RefHV + i;
1694 : syskin 801 bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1695 :     bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1696 :     bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8;
1697 :     bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8;
1698 : chl 96
1699 : syskin 801
1700 : h 545 bData.bpredMV = fData->predMV = *f_predMV;
1701 :     fData->bpredMV = bData.predMV = *b_predMV;
1702 : syskin 661 fData->currentMV[0] = fData->currentMV[2];
1703 : chl 96
1704 : syskin 712 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);
1705 :     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);
1706 :    
1707 : h 545 if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx;
1708 : syskin 675 if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dx;
1709 :     if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy;
1710 :     if (fData->currentMV[0].y < fData->min_dy) fData->currentMV[0].y = fData->min_dy;
1711 : Isibaar 539
1712 : h 545 if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx;
1713 : syskin 675 if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dx;
1714 :     if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy;
1715 :     if (fData->currentMV[1].y < bData.min_dy) fData->currentMV[1].y = bData.min_dy;
1716 : Isibaar 539
1717 : h 545 CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData);
1718 : chl 96
1719 : syskin 768 //diamond
1720 : chl 530 do {
1721 :     iDirection = 255;
1722 :     // forward MV moves
1723 : h 545 i = fData->currentMV[0].x; j = fData->currentMV[0].y;
1724 : chl 96
1725 : h 545 CheckCandidateInt(i + 1, j, 0, &iDirection, fData);
1726 :     CheckCandidateInt(i, j + 1, 0, &iDirection, fData);
1727 :     CheckCandidateInt(i - 1, j, 0, &iDirection, fData);
1728 :     CheckCandidateInt(i, j - 1, 0, &iDirection, fData);
1729 : chl 96
1730 : chl 530 // backward MV moves
1731 : h 545 i = fData->currentMV[1].x; j = fData->currentMV[1].y;
1732 :     fData->currentMV[2] = fData->currentMV[0];
1733 : Isibaar 539 CheckCandidateInt(i + 1, j, 0, &iDirection, &bData);
1734 :     CheckCandidateInt(i, j + 1, 0, &iDirection, &bData);
1735 :     CheckCandidateInt(i - 1, j, 0, &iDirection, &bData);
1736 :     CheckCandidateInt(i, j - 1, 0, &iDirection, &bData);
1737 : chl 96
1738 : chl 530 } while (!(iDirection));
1739 : edgomez 195
1740 : syskin 768 //qpel refinement
1741 : syskin 658 if (fData->qpel) {
1742 : syskin 740 if (*fData->iMinSAD > *best_sad + 500) return;
1743 : syskin 675 CheckCandidate = CheckCandidateInt;
1744 : syskin 663 fData->qpel_precision = bData.qpel_precision = 1;
1745 : syskin 712 get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0);
1746 :     get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0);
1747 : syskin 658 fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x;
1748 :     fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y;
1749 :     fData->currentQMV[1].x = 2 * fData->currentMV[1].x;
1750 :     fData->currentQMV[1].y = 2 * fData->currentMV[1].y;
1751 : syskin 663 SubpelRefine(fData);
1752 : syskin 740 if (*fData->iMinSAD > *best_sad + 300) return;
1753 : syskin 658 fData->currentQMV[2] = fData->currentQMV[0];
1754 : syskin 663 SubpelRefine(&bData);
1755 : syskin 658 }
1756 :    
1757 : syskin 768 *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode.
1758 :    
1759 : h 545 if (*fData->iMinSAD < *best_sad) {
1760 :     *best_sad = *fData->iMinSAD;
1761 :     pMB->mvs[0] = fData->currentMV[0];
1762 :     pMB->b_mvs[0] = fData->currentMV[1];
1763 : chl 530 pMB->mode = MODE_INTERPOLATE;
1764 : syskin 658 if (fData->qpel) {
1765 :     pMB->qmvs[0] = fData->currentQMV[0];
1766 :     pMB->b_qmvs[0] = fData->currentQMV[1];
1767 :     pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x;
1768 :     pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y;
1769 :     pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x;
1770 :     pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y;
1771 :     } else {
1772 :     pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x;
1773 :     pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y;
1774 :     pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x;
1775 :     pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y;
1776 :     }
1777 : chl 530 }
1778 :     }
1779 : chl 96
1780 : chl 530 void
1781 :     MotionEstimationBVOP(MBParam * const pParam,
1782 :     FRAMEINFO * const frame,
1783 :     const int32_t time_bp,
1784 :     const int32_t time_pp,
1785 :     // forward (past) reference
1786 :     const MACROBLOCK * const f_mbs,
1787 :     const IMAGE * const f_ref,
1788 :     const IMAGE * const f_refH,
1789 :     const IMAGE * const f_refV,
1790 :     const IMAGE * const f_refHV,
1791 :     // backward (future) reference
1792 : syskin 644 const FRAMEINFO * const b_reference,
1793 : chl 530 const IMAGE * const b_ref,
1794 :     const IMAGE * const b_refH,
1795 :     const IMAGE * const b_refV,
1796 :     const IMAGE * const b_refHV)
1797 :     {
1798 :     uint32_t i, j;
1799 : syskin 712 int32_t best_sad;
1800 :     uint32_t skip_sad;
1801 : chl 530 int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0;
1802 : syskin 644 const MACROBLOCK * const b_mbs = b_reference->mbs;
1803 : chl 96
1804 : chl 530 VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/
1805 : chl 96
1806 : chl 530 const int32_t TRB = time_pp - time_bp;
1807 :     const int32_t TRD = time_pp;
1808 : chl 96
1809 : h 545 // some pre-inintialized data for the rest of the search
1810 :    
1811 :     SearchData Data;
1812 :     int32_t iMinSAD;
1813 :     VECTOR currentMV[3];
1814 : syskin 658 VECTOR currentQMV[3];
1815 : syskin 801 int32_t temp[8];
1816 : syskin 704 memset(&Data, 0, sizeof(SearchData));
1817 : h 545 Data.iEdgedWidth = pParam->edged_width;
1818 : syskin 658 Data.currentMV = currentMV; Data.currentQMV = currentQMV;
1819 : h 545 Data.iMinSAD = &iMinSAD;
1820 : syskin 712 Data.lambda16 = lambda_vec16[frame->quant];
1821 : syskin 658 Data.qpel = pParam->m_quarterpel;
1822 : syskin 663 Data.rounding = 0;
1823 : syskin 801 Data.chroma = frame->motion_flags & PMV_CHROMA8;
1824 :     Data.temp = temp;
1825 : h 545
1826 : syskin 754 Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose)
1827 : chl 530 // note: i==horizontal, j==vertical
1828 :     for (j = 0; j < pParam->mb_height; j++) {
1829 : chl 96
1830 : chl 530 f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */
1831 : edgomez 195
1832 : chl 530 for (i = 0; i < pParam->mb_width; i++) {
1833 :     MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width;
1834 :     const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width;
1835 : edgomez 195
1836 : chl 619 /* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */
1837 : syskin 644 if (b_reference->coding_type != S_VOP)
1838 :     if (b_mb->mode == MODE_NOT_CODED) {
1839 :     pMB->mode = MODE_NOT_CODED;
1840 :     continue;
1841 :     }
1842 : chl 96
1843 : h 545 Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16;
1844 : syskin 801 Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8;
1845 :     Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8;
1846 : syskin 601 pMB->quant = frame->quant;
1847 : syskin 628
1848 : chl 530 /* direct search comes first, because it (1) checks for SKIP-mode
1849 :     and (2) sets very good predictions for forward and backward search */
1850 : Isibaar 539 skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y,
1851 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
1852 : chl 530 &frame->image,
1853 :     i, j,
1854 :     frame->motion_flags,
1855 :     TRB, TRD,
1856 :     pParam,
1857 :     pMB, b_mb,
1858 : h 545 &best_sad,
1859 :     &Data);
1860 : chl 96
1861 : Isibaar 539 if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; }
1862 : syskin 739
1863 : chl 530 // forward search
1864 : syskin 801 SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y,
1865 : chl 530 &frame->image, i, j,
1866 :     frame->motion_flags,
1867 : h 545 frame->fcode, pParam,
1868 : chl 530 pMB, &f_predMV, &best_sad,
1869 : h 545 MODE_FORWARD, &Data);
1870 : chl 96
1871 : chl 530 // backward search
1872 : syskin 801 SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y,
1873 : chl 530 &frame->image, i, j,
1874 :     frame->motion_flags,
1875 : h 545 frame->bcode, pParam,
1876 : chl 530 pMB, &b_predMV, &best_sad,
1877 : h 545 MODE_BACKWARD, &Data);
1878 : chl 96
1879 : chl 530 // interpolate search comes last, because it uses data from forward and backward as prediction
1880 : syskin 801 SearchInterpolate(f_ref, f_refH->y, f_refV->y, f_refHV->y,
1881 :     b_ref, b_refH->y, b_refV->y, b_refHV->y,
1882 : chl 530 &frame->image,
1883 :     i, j,
1884 :     frame->fcode, frame->bcode,
1885 :     frame->motion_flags,
1886 : h 545 pParam,
1887 : chl 530 &f_predMV, &b_predMV,
1888 : h 545 pMB, &best_sad,
1889 :     &Data);
1890 : syskin 662
1891 : syskin 712 // final skip decision
1892 : syskin 801 if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP * 2)
1893 : syskin 712 && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) )
1894 : syskin 767 SkipDecisionB(&frame->image, f_ref, b_ref, pMB, i, j, &Data);
1895 : syskin 712
1896 : chl 530 switch (pMB->mode) {
1897 :     case MODE_FORWARD:
1898 :     f_count++;
1899 : syskin 768 f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
1900 : chl 530 break;
1901 :     case MODE_BACKWARD:
1902 :     b_count++;
1903 : syskin 768 b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
1904 : chl 530 break;
1905 :     case MODE_INTERPOLATE:
1906 :     i_count++;
1907 : syskin 768 f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0];
1908 :     b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0];
1909 : chl 530 break;
1910 :     case MODE_DIRECT:
1911 :     case MODE_DIRECT_NO4V:
1912 :     d_count++;
1913 :     default:
1914 :     break;
1915 : chl 96 }
1916 :     }
1917 :     }
1918 :     }
1919 :    
1920 : syskin 739 static __inline void
1921 : Isibaar 539 MEanalyzeMB ( const uint8_t * const pRef,
1922 :     const uint8_t * const pCur,
1923 :     const int x,
1924 :     const int y,
1925 :     const MBParam * const pParam,
1926 : syskin 739 MACROBLOCK * const pMBs,
1927 : h 545 SearchData * const Data)
1928 : Isibaar 539 {
1929 : chl 317
1930 : syskin 739 int i, mask;
1931 : h 545 VECTOR pmv[3];
1932 : syskin 739 MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width];
1933 : syskin 768
1934 : syskin 739 for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR;
1935 : suxen_drol 118
1936 : syskin 640 //median is only used as prediction. it doesn't have to be real
1937 :     if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0;
1938 :     else
1939 :     if (x == 1) //left macroblock does not have any vector now
1940 :     Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median
1941 : syskin 748 else if (y == 1) // top macroblock doesn't have it's vector
1942 : syskin 640 Data->predMV = (pMB - 1)->mvs[0]; // left instead of median
1943 :     else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median
1944 :    
1945 : h 545 get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16,
1946 : syskin 739 pParam->width, pParam->height, Data->iFcode - pParam->m_quarterpel, 0, Data->rrv);
1947 : chl 326
1948 : h 545 Data->Cur = pCur + (x + y * pParam->edged_width) * 16;
1949 :     Data->Ref = pRef + (x + y * pParam->edged_width) * 16;
1950 : syskin 739
1951 : Isibaar 539 pmv[1].x = EVEN(pMB->mvs[0].x);
1952 :     pmv[1].y = EVEN(pMB->mvs[0].y);
1953 : syskin 640 pmv[2].x = EVEN(Data->predMV.x);
1954 :     pmv[2].y = EVEN(Data->predMV.y);
1955 :     pmv[0].x = pmv[0].y = 0;
1956 : chl 326
1957 : syskin 739 CheckCandidate32I(0, 0, 255, &i, Data);
1958 : syskin 640
1959 : syskin 739 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) {
1960 : syskin 640
1961 : syskin 739 if (!(mask = make_mask(pmv, 1)))
1962 :     CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data);
1963 :     if (!(mask = make_mask(pmv, 2)))
1964 :     CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data);
1965 : Isibaar 539
1966 : syskin 739 if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed
1967 :     DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i);
1968 : Isibaar 539
1969 : syskin 739 for (i = 0; i < 4; i++) {
1970 : syskin 768 MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1)) * pParam->mb_width];
1971 : syskin 739 MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i];
1972 :     MB->mode = MODE_INTER;
1973 :     MB->sad16 = Data->iMinSAD[i+1];
1974 :     }
1975 :     }
1976 : Isibaar 539 }
1977 :    
1978 : syskin 739 #define INTRA_BIAS 2500
1979 :     #define INTRA_THRESH 1500
1980 :     #define INTER_THRESH 1400
1981 : Isibaar 539
1982 :     int
1983 : syskin 768 MEanalysis( const IMAGE * const pRef,
1984 : syskin 644 FRAMEINFO * const Current,
1985 : Isibaar 539 MBParam * const pParam,
1986 : syskin 644 int maxIntra, //maximum number if non-I frames
1987 :     int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame
1988 : syskin 768 int bCount) // number of B frames in a row
1989 : Isibaar 539 {
1990 :     uint32_t x, y, intra = 0;
1991 :     int sSAD = 0;
1992 : syskin 644 MACROBLOCK * const pMBs = Current->mbs;
1993 :     const IMAGE * const pCurrent = &Current->image;
1994 :     int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH;
1995 : h 545
1996 : syskin 739 int32_t iMinSAD[5], temp[5];
1997 :     VECTOR currentMV[5];
1998 : h 545 SearchData Data;
1999 :     Data.iEdgedWidth = pParam->edged_width;
2000 : syskin 739 Data.currentMV = currentMV;
2001 :     Data.iMinSAD = iMinSAD;
2002 : syskin 644 Data.iFcode = Current->fcode;
2003 : syskin 739 Data.rrv = Current->global_flags & XVID_REDUCED;
2004 :     Data.temp = temp;
2005 :     CheckCandidate = CheckCandidate32I;
2006 : h 545
2007 : syskin 768 if (intraCount != 0 && intraCount < 10) // we're right after an I frame
2008 : syskin 658 IntraThresh += 4 * (intraCount - 10) * (intraCount - 10);
2009 : syskin 644 else
2010 :     if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec
2011 :     IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra;
2012 :    
2013 : syskin 672 InterThresh += 400 * (1 - bCount);
2014 : syskin 739 if (InterThresh < 300) InterThresh = 300;
2015 : syskin 644
2016 : Isibaar 539 if (sadInit) (*sadInit) ();
2017 :    
2018 : syskin 768 for (y = 1; y < pParam->mb_height-1; y += 2) {
2019 :     for (x = 1; x < pParam->mb_width-1; x += 2) {
2020 : syskin 739 int i;
2021 : Isibaar 539
2022 : syskin 739 if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV;
2023 :    
2024 :     MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data);
2025 : syskin 768
2026 : syskin 739 for (i = 0; i < 4; i++) {
2027 :     int dev;
2028 : syskin 768 MACROBLOCK *pMB = &pMBs[x+(i&1) + (y+(i>>1)) * pParam->mb_width];
2029 : syskin 739 if (pMB->sad16 > IntraThresh) {
2030 : syskin 768 dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16,
2031 :     pParam->edged_width);
2032 : syskin 739 if (dev + IntraThresh < pMB->sad16) {
2033 :     pMB->mode = MODE_INTRA;
2034 :     if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP;
2035 :     }
2036 : Isibaar 539 }
2037 : syskin 739 sSAD += pMB->sad16;
2038 : syskin 644 }
2039 : suxen_drol 118 }
2040 :     }
2041 : syskin 704 sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2);
2042 : syskin 793 // if (sSAD > IntraThresh + INTRA_BIAS) return I_VOP;
2043 : suxen_drol 709 if (sSAD > InterThresh ) return P_VOP;
2044 : Isibaar 539 emms();
2045 : suxen_drol 709 return B_VOP;
2046 : syskin 712
2047 : suxen_drol 118 }
2048 : syskin 576
2049 : chl 769
2050 :     static WARPPOINTS
2051 : syskin 801 GlobalMotionEst(const MACROBLOCK * const pMBs,
2052 : chl 769 const MBParam * const pParam,
2053 :     const FRAMEINFO * const current,
2054 :     const FRAMEINFO * const reference,
2055 :     const IMAGE * const pRefH,
2056 :     const IMAGE * const pRefV,
2057 :     const IMAGE * const pRefHV )
2058 : syskin 628 {
2059 : chl 619
2060 : chl 769 const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs
2061 :     const int deltay=8;
2062 : syskin 801 const int grad=512; // lower bound for deviation in MB
2063 : chl 769
2064 :     WARPPOINTS gmc;
2065 :    
2066 :     uint32_t mx, my;
2067 : chl 619
2068 : chl 769 int MBh = pParam->mb_height;
2069 :     int MBw = pParam->mb_width;
2070 :    
2071 :     int *MBmask= calloc(MBh*MBw,sizeof(int));
2072 :     double DtimesF[4] = { 0.,0., 0., 0. };
2073 :     double sol[4] = { 0., 0., 0., 0. };
2074 :     double a,b,c,n,denom;
2075 :     double meanx,meany;
2076 :     int num,oldnum;
2077 :    
2078 :     if (!MBmask) { fprintf(stderr,"Mem error\n"); return gmc;}
2079 :    
2080 : syskin 801 // filter mask of all blocks
2081 : chl 769
2082 :     for (my = 1; my < MBh-1; my++)
2083 :     for (mx = 1; mx < MBw-1; mx++)
2084 :     {
2085 :     const int mbnum = mx + my * MBw;
2086 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2087 :     const VECTOR mv = pMB->mvs[0];
2088 :    
2089 :     if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED)
2090 :     continue;
2091 :    
2092 :     if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) )
2093 :     && ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) )
2094 :     && ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) )
2095 :     && ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) )
2096 :     MBmask[mbnum]=1;
2097 : syskin 628 }
2098 : chl 769
2099 :     for (my = 1; my < MBh-1; my++)
2100 :     for (mx = 1; mx < MBw-1; mx++)
2101 :     {
2102 :     const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx;
2103 :    
2104 :     const int mbnum = mx + my * MBw;
2105 :     if (!MBmask[mbnum])
2106 :     continue;
2107 : syskin 628
2108 : chl 769 if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= grad )
2109 :     MBmask[mbnum] = 0;
2110 :     if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= grad )
2111 :     MBmask[mbnum] = 0;
2112 :    
2113 :     }
2114 : syskin 628
2115 : chl 769 emms();
2116 : chl 619
2117 : chl 769 do { /* until convergence */
2118 : chl 619
2119 : chl 769 a = b = c = n = 0;
2120 :     DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.;
2121 :     for (my = 0; my < MBh; my++)
2122 :     for (mx = 0; mx < MBw; mx++)
2123 :     {
2124 :     const int mbnum = mx + my * MBw;
2125 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2126 :     const VECTOR mv = pMB->mvs[0];
2127 : chl 619
2128 : chl 769 if (!MBmask[mbnum])
2129 :     continue;
2130 :    
2131 :     n++;
2132 :     a += 16*mx+8;
2133 :     b += 16*my+8;
2134 :     c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8);
2135 : syskin 768
2136 : chl 769 DtimesF[0] += (double)mv.x;
2137 :     DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8);
2138 :     DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8);
2139 :     DtimesF[3] += (double)mv.y;
2140 :     }
2141 : syskin 768
2142 : chl 769 denom = a*a+b*b-c*n;
2143 :    
2144 :     /* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */
2145 :     /* D'*E*F has been calculated in the same loop as matrix */
2146 : syskin 768
2147 : chl 769 sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2];
2148 :     sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3];
2149 :     sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3];
2150 :     sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3];
2151 : syskin 628
2152 : chl 769 sol[0] /= denom;
2153 :     sol[1] /= denom;
2154 :     sol[2] /= denom;
2155 :     sol[3] /= denom;
2156 :    
2157 :     meanx = meany = 0.;
2158 :     oldnum = 0;
2159 :     for (my = 0; my < MBh; my++)
2160 :     for (mx = 0; mx < MBw; mx++)
2161 :     {
2162 :     const int mbnum = mx + my * MBw;
2163 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2164 :     const VECTOR mv = pMB->mvs[0];
2165 : syskin 628
2166 : chl 769 if (!MBmask[mbnum])
2167 :     continue;
2168 : syskin 628
2169 : chl 769 oldnum++;
2170 :     meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x );
2171 : chl 770 meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y );
2172 :     }
2173 : chl 769
2174 : chl 770 if (4*meanx > oldnum) /* better fit than 0.25 is useless */
2175 : chl 769 meanx /= oldnum;
2176 : syskin 801 else
2177 : chl 770 meanx = 0.25;
2178 :    
2179 :     if (4*meany > oldnum)
2180 : chl 769 meany /= oldnum;
2181 : syskin 801 else
2182 : chl 770 meany = 0.25;
2183 : chl 769
2184 : chl 770 /* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]);
2185 :     fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum);
2186 :     */
2187 : chl 769 num = 0;
2188 :     for (my = 0; my < MBh; my++)
2189 :     for (mx = 0; mx < MBw; mx++)
2190 :     {
2191 :     const int mbnum = mx + my * MBw;
2192 :     const MACROBLOCK *pMB = &pMBs[mbnum];
2193 :     const VECTOR mv = pMB->mvs[0];
2194 :    
2195 :     if (!MBmask[mbnum])
2196 :     continue;
2197 :    
2198 :     if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx )
2199 : chl 770 || ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) )
2200 : chl 769 MBmask[mbnum]=0;
2201 :     else
2202 :     num++;
2203 :     }
2204 :    
2205 :     } while ( (oldnum != num) && (num>=4) );
2206 :    
2207 :     if (num < 4)
2208 :     {
2209 :     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;
2210 :     } else {
2211 :    
2212 :     gmc.duv[0].x=(int)(sol[0]+0.5);
2213 :     gmc.duv[0].y=(int)(sol[3]+0.5);
2214 :    
2215 :     gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5);
2216 :     gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5);
2217 :    
2218 :     gmc.duv[2].x=0;
2219 :     gmc.duv[2].y=0;
2220 : chl 619 }
2221 : chl 769 // 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);
2222 : chl 619
2223 : chl 769 free(MBmask);
2224 :    
2225 : chl 619 return gmc;
2226 :     }

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