The Tor Browser / file comparison

comparison: media/libvpx/vp8/encoder/mcomp.c

media/libvpx/vp8/encoder/mcomp.c

changeset 0
6474c204b198
equal deleted inserted replaced
-1:000000000000 0:420a17b96077
1 /*
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11
12 #include "onyx_int.h"
13 #include "mcomp.h"
14 #include "vpx_mem/vpx_mem.h"
15 #include "vpx_config.h"
16 #include <stdio.h>
17 #include <limits.h>
18 #include <math.h>
19 #include "vp8/common/findnearmv.h"
20
21 #ifdef VP8_ENTROPY_STATS
22 static int mv_ref_ct [31] [4] [2];
23 static int mv_mode_cts [4] [2];
24 #endif
25
26 int vp8_mv_bit_cost(int_mv *mv, int_mv *ref, int *mvcost[2], int Weight)
27 {
28 /* MV costing is based on the distribution of vectors in the previous
29 * frame and as such will tend to over state the cost of vectors. In
30 * addition coding a new vector can have a knock on effect on the cost
31 * of subsequent vectors and the quality of prediction from NEAR and
32 * NEAREST for subsequent blocks. The "Weight" parameter allows, to a
33 * limited extent, for some account to be taken of these factors.
34 */
35 return ((mvcost[0][(mv->as_mv.row - ref->as_mv.row) >> 1] + mvcost[1][(mv->as_mv.col - ref->as_mv.col) >> 1]) * Weight) >> 7;
36 }
37
38 static int mv_err_cost(int_mv *mv, int_mv *ref, int *mvcost[2], int error_per_bit)
39 {
40 /* Ignore mv costing if mvcost is NULL */
41 if (mvcost)
42 return ((mvcost[0][(mv->as_mv.row - ref->as_mv.row) >> 1] +
43 mvcost[1][(mv->as_mv.col - ref->as_mv.col) >> 1])
44 * error_per_bit + 128) >> 8;
45 return 0;
46 }
47
48 static int mvsad_err_cost(int_mv *mv, int_mv *ref, int *mvsadcost[2], int error_per_bit)
49 {
50 /* Calculate sad error cost on full pixel basis. */
51 /* Ignore mv costing if mvsadcost is NULL */
52 if (mvsadcost)
53 return ((mvsadcost[0][(mv->as_mv.row - ref->as_mv.row)] +
54 mvsadcost[1][(mv->as_mv.col - ref->as_mv.col)])
55 * error_per_bit + 128) >> 8;
56 return 0;
57 }
58
59 void vp8_init_dsmotion_compensation(MACROBLOCK *x, int stride)
60 {
61 int Len;
62 int search_site_count = 0;
63
64
65 /* Generate offsets for 4 search sites per step. */
66 Len = MAX_FIRST_STEP;
67 x->ss[search_site_count].mv.col = 0;
68 x->ss[search_site_count].mv.row = 0;
69 x->ss[search_site_count].offset = 0;
70 search_site_count++;
71
72 while (Len > 0)
73 {
74
75 /* Compute offsets for search sites. */
76 x->ss[search_site_count].mv.col = 0;
77 x->ss[search_site_count].mv.row = -Len;
78 x->ss[search_site_count].offset = -Len * stride;
79 search_site_count++;
80
81 /* Compute offsets for search sites. */
82 x->ss[search_site_count].mv.col = 0;
83 x->ss[search_site_count].mv.row = Len;
84 x->ss[search_site_count].offset = Len * stride;
85 search_site_count++;
86
87 /* Compute offsets for search sites. */
88 x->ss[search_site_count].mv.col = -Len;
89 x->ss[search_site_count].mv.row = 0;
90 x->ss[search_site_count].offset = -Len;
91 search_site_count++;
92
93 /* Compute offsets for search sites. */
94 x->ss[search_site_count].mv.col = Len;
95 x->ss[search_site_count].mv.row = 0;
96 x->ss[search_site_count].offset = Len;
97 search_site_count++;
98
99 /* Contract. */
100 Len /= 2;
101 }
102
103 x->ss_count = search_site_count;
104 x->searches_per_step = 4;
105 }
106
107 void vp8_init3smotion_compensation(MACROBLOCK *x, int stride)
108 {
109 int Len;
110 int search_site_count = 0;
111
112 /* Generate offsets for 8 search sites per step. */
113 Len = MAX_FIRST_STEP;
114 x->ss[search_site_count].mv.col = 0;
115 x->ss[search_site_count].mv.row = 0;
116 x->ss[search_site_count].offset = 0;
117 search_site_count++;
118
119 while (Len > 0)
120 {
121
122 /* Compute offsets for search sites. */
123 x->ss[search_site_count].mv.col = 0;
124 x->ss[search_site_count].mv.row = -Len;
125 x->ss[search_site_count].offset = -Len * stride;
126 search_site_count++;
127
128 /* Compute offsets for search sites. */
129 x->ss[search_site_count].mv.col = 0;
130 x->ss[search_site_count].mv.row = Len;
131 x->ss[search_site_count].offset = Len * stride;
132 search_site_count++;
133
134 /* Compute offsets for search sites. */
135 x->ss[search_site_count].mv.col = -Len;
136 x->ss[search_site_count].mv.row = 0;
137 x->ss[search_site_count].offset = -Len;
138 search_site_count++;
139
140 /* Compute offsets for search sites. */
141 x->ss[search_site_count].mv.col = Len;
142 x->ss[search_site_count].mv.row = 0;
143 x->ss[search_site_count].offset = Len;
144 search_site_count++;
145
146 /* Compute offsets for search sites. */
147 x->ss[search_site_count].mv.col = -Len;
148 x->ss[search_site_count].mv.row = -Len;
149 x->ss[search_site_count].offset = -Len * stride - Len;
150 search_site_count++;
151
152 /* Compute offsets for search sites. */
153 x->ss[search_site_count].mv.col = Len;
154 x->ss[search_site_count].mv.row = -Len;
155 x->ss[search_site_count].offset = -Len * stride + Len;
156 search_site_count++;
157
158 /* Compute offsets for search sites. */
159 x->ss[search_site_count].mv.col = -Len;
160 x->ss[search_site_count].mv.row = Len;
161 x->ss[search_site_count].offset = Len * stride - Len;
162 search_site_count++;
163
164 /* Compute offsets for search sites. */
165 x->ss[search_site_count].mv.col = Len;
166 x->ss[search_site_count].mv.row = Len;
167 x->ss[search_site_count].offset = Len * stride + Len;
168 search_site_count++;
169
170
171 /* Contract. */
172 Len /= 2;
173 }
174
175 x->ss_count = search_site_count;
176 x->searches_per_step = 8;
177 }
178
179 /*
180 * To avoid the penalty for crossing cache-line read, preload the reference
181 * area in a small buffer, which is aligned to make sure there won't be crossing
182 * cache-line read while reading from this buffer. This reduced the cpu
183 * cycles spent on reading ref data in sub-pixel filter functions.
184 * TODO: Currently, since sub-pixel search range here is -3 ~ 3, copy 22 rows x
185 * 32 cols area that is enough for 16x16 macroblock. Later, for SPLITMV, we
186 * could reduce the area.
187 */
188
189 /* estimated cost of a motion vector (r,c) */
190 #define MVC(r,c) (mvcost ? ((mvcost[0][(r)-rr] + mvcost[1][(c) - rc]) * error_per_bit + 128 )>>8 : 0)
191 /* pointer to predictor base of a motionvector */
192 #define PRE(r,c) (y + (((r)>>2) * y_stride + ((c)>>2) -(offset)))
193 /* convert motion vector component to offset for svf calc */
194 #define SP(x) (((x)&3)<<1)
195 /* returns subpixel variance error function. */
196 #define DIST(r,c) vfp->svf( PRE(r,c), y_stride, SP(c),SP(r), z,b->src_stride,&sse)
197 #define IFMVCV(r,c,s,e) if ( c >= minc && c <= maxc && r >= minr && r <= maxr) s else e;
198 /* returns distortion + motion vector cost */
199 #define ERR(r,c) (MVC(r,c)+DIST(r,c))
200 /* checks if (r,c) has better score than previous best */
201 #define CHECK_BETTER(v,r,c) IFMVCV(r,c,{thismse = DIST(r,c); if((v = (MVC(r,c)+thismse)) < besterr) { besterr = v; br=r; bc=c; *distortion = thismse; *sse1 = sse; }}, v=UINT_MAX;)
202
203 int vp8_find_best_sub_pixel_step_iteratively(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
204 int_mv *bestmv, int_mv *ref_mv,
205 int error_per_bit,
206 const vp8_variance_fn_ptr_t *vfp,
207 int *mvcost[2], int *distortion,
208 unsigned int *sse1)
209 {
210 unsigned char *z = (*(b->base_src) + b->src);
211
212 int rr = ref_mv->as_mv.row >> 1, rc = ref_mv->as_mv.col >> 1;
213 int br = bestmv->as_mv.row * 4, bc = bestmv->as_mv.col * 4;
214 int tr = br, tc = bc;
215 unsigned int besterr;
216 unsigned int left, right, up, down, diag;
217 unsigned int sse;
218 unsigned int whichdir;
219 unsigned int halfiters = 4;
220 unsigned int quarteriters = 4;
221 int thismse;
222
223 int minc = MAX(x->mv_col_min * 4,
224 (ref_mv->as_mv.col >> 1) - ((1 << mvlong_width) - 1));
225 int maxc = MIN(x->mv_col_max * 4,
226 (ref_mv->as_mv.col >> 1) + ((1 << mvlong_width) - 1));
227 int minr = MAX(x->mv_row_min * 4,
228 (ref_mv->as_mv.row >> 1) - ((1 << mvlong_width) - 1));
229 int maxr = MIN(x->mv_row_max * 4,
230 (ref_mv->as_mv.row >> 1) + ((1 << mvlong_width) - 1));
231
232 int y_stride;
233 int offset;
234 int pre_stride = x->e_mbd.pre.y_stride;
235 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
236
237
238 #if ARCH_X86 || ARCH_X86_64
239 MACROBLOCKD *xd = &x->e_mbd;
240 unsigned char *y_0 = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
241 unsigned char *y;
242 int buf_r1, buf_r2, buf_c1;
243
244 /* Clamping to avoid out-of-range data access */
245 buf_r1 = ((bestmv->as_mv.row - 3) < x->mv_row_min)?(bestmv->as_mv.row - x->mv_row_min):3;
246 buf_r2 = ((bestmv->as_mv.row + 3) > x->mv_row_max)?(x->mv_row_max - bestmv->as_mv.row):3;
247 buf_c1 = ((bestmv->as_mv.col - 3) < x->mv_col_min)?(bestmv->as_mv.col - x->mv_col_min):3;
248 y_stride = 32;
249
250 /* Copy to intermediate buffer before searching. */
251 vfp->copymem(y_0 - buf_c1 - pre_stride*buf_r1, pre_stride, xd->y_buf, y_stride, 16+buf_r1+buf_r2);
252 y = xd->y_buf + y_stride*buf_r1 +buf_c1;
253 #else
254 unsigned char *y = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
255 y_stride = pre_stride;
256 #endif
257
258 offset = (bestmv->as_mv.row) * y_stride + bestmv->as_mv.col;
259
260 /* central mv */
261 bestmv->as_mv.row *= 8;
262 bestmv->as_mv.col *= 8;
263
264 /* calculate central point error */
265 besterr = vfp->vf(y, y_stride, z, b->src_stride, sse1);
266 *distortion = besterr;
267 besterr += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
268
269 /* TODO: Each subsequent iteration checks at least one point in common
270 * with the last iteration could be 2 ( if diag selected)
271 */
272 while (--halfiters)
273 {
274 /* 1/2 pel */
275 CHECK_BETTER(left, tr, tc - 2);
276 CHECK_BETTER(right, tr, tc + 2);
277 CHECK_BETTER(up, tr - 2, tc);
278 CHECK_BETTER(down, tr + 2, tc);
279
280 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
281
282 switch (whichdir)
283 {
284 case 0:
285 CHECK_BETTER(diag, tr - 2, tc - 2);
286 break;
287 case 1:
288 CHECK_BETTER(diag, tr - 2, tc + 2);
289 break;
290 case 2:
291 CHECK_BETTER(diag, tr + 2, tc - 2);
292 break;
293 case 3:
294 CHECK_BETTER(diag, tr + 2, tc + 2);
295 break;
296 }
297
298 /* no reason to check the same one again. */
299 if (tr == br && tc == bc)
300 break;
301
302 tr = br;
303 tc = bc;
304 }
305
306 /* TODO: Each subsequent iteration checks at least one point in common
307 * with the last iteration could be 2 ( if diag selected)
308 */
309
310 /* 1/4 pel */
311 while (--quarteriters)
312 {
313 CHECK_BETTER(left, tr, tc - 1);
314 CHECK_BETTER(right, tr, tc + 1);
315 CHECK_BETTER(up, tr - 1, tc);
316 CHECK_BETTER(down, tr + 1, tc);
317
318 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
319
320 switch (whichdir)
321 {
322 case 0:
323 CHECK_BETTER(diag, tr - 1, tc - 1);
324 break;
325 case 1:
326 CHECK_BETTER(diag, tr - 1, tc + 1);
327 break;
328 case 2:
329 CHECK_BETTER(diag, tr + 1, tc - 1);
330 break;
331 case 3:
332 CHECK_BETTER(diag, tr + 1, tc + 1);
333 break;
334 }
335
336 /* no reason to check the same one again. */
337 if (tr == br && tc == bc)
338 break;
339
340 tr = br;
341 tc = bc;
342 }
343
344 bestmv->as_mv.row = br * 2;
345 bestmv->as_mv.col = bc * 2;
346
347 if ((abs(bestmv->as_mv.col - ref_mv->as_mv.col) > (MAX_FULL_PEL_VAL<<3)) ||
348 (abs(bestmv->as_mv.row - ref_mv->as_mv.row) > (MAX_FULL_PEL_VAL<<3)))
349 return INT_MAX;
350
351 return besterr;
352 }
353 #undef MVC
354 #undef PRE
355 #undef SP
356 #undef DIST
357 #undef IFMVCV
358 #undef ERR
359 #undef CHECK_BETTER
360
361 int vp8_find_best_sub_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
362 int_mv *bestmv, int_mv *ref_mv,
363 int error_per_bit,
364 const vp8_variance_fn_ptr_t *vfp,
365 int *mvcost[2], int *distortion,
366 unsigned int *sse1)
367 {
368 int bestmse = INT_MAX;
369 int_mv startmv;
370 int_mv this_mv;
371 unsigned char *z = (*(b->base_src) + b->src);
372 int left, right, up, down, diag;
373 unsigned int sse;
374 int whichdir ;
375 int thismse;
376 int y_stride;
377 int pre_stride = x->e_mbd.pre.y_stride;
378 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
379
380 #if ARCH_X86 || ARCH_X86_64
381 MACROBLOCKD *xd = &x->e_mbd;
382 unsigned char *y_0 = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
383 unsigned char *y;
384
385 y_stride = 32;
386 /* Copy 18 rows x 32 cols area to intermediate buffer before searching. */
387 vfp->copymem(y_0 - 1 - pre_stride, pre_stride, xd->y_buf, y_stride, 18);
388 y = xd->y_buf + y_stride + 1;
389 #else
390 unsigned char *y = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
391 y_stride = pre_stride;
392 #endif
393
394 /* central mv */
395 bestmv->as_mv.row <<= 3;
396 bestmv->as_mv.col <<= 3;
397 startmv = *bestmv;
398
399 /* calculate central point error */
400 bestmse = vfp->vf(y, y_stride, z, b->src_stride, sse1);
401 *distortion = bestmse;
402 bestmse += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
403
404 /* go left then right and check error */
405 this_mv.as_mv.row = startmv.as_mv.row;
406 this_mv.as_mv.col = ((startmv.as_mv.col - 8) | 4);
407 thismse = vfp->svf_halfpix_h(y - 1, y_stride, z, b->src_stride, &sse);
408 left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
409
410 if (left < bestmse)
411 {
412 *bestmv = this_mv;
413 bestmse = left;
414 *distortion = thismse;
415 *sse1 = sse;
416 }
417
418 this_mv.as_mv.col += 8;
419 thismse = vfp->svf_halfpix_h(y, y_stride, z, b->src_stride, &sse);
420 right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
421
422 if (right < bestmse)
423 {
424 *bestmv = this_mv;
425 bestmse = right;
426 *distortion = thismse;
427 *sse1 = sse;
428 }
429
430 /* go up then down and check error */
431 this_mv.as_mv.col = startmv.as_mv.col;
432 this_mv.as_mv.row = ((startmv.as_mv.row - 8) | 4);
433 thismse = vfp->svf_halfpix_v(y - y_stride, y_stride, z, b->src_stride, &sse);
434 up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
435
436 if (up < bestmse)
437 {
438 *bestmv = this_mv;
439 bestmse = up;
440 *distortion = thismse;
441 *sse1 = sse;
442 }
443
444 this_mv.as_mv.row += 8;
445 thismse = vfp->svf_halfpix_v(y, y_stride, z, b->src_stride, &sse);
446 down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
447
448 if (down < bestmse)
449 {
450 *bestmv = this_mv;
451 bestmse = down;
452 *distortion = thismse;
453 *sse1 = sse;
454 }
455
456
457 /* now check 1 more diagonal */
458 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
459 this_mv = startmv;
460
461 switch (whichdir)
462 {
463 case 0:
464 this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
465 this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
466 thismse = vfp->svf_halfpix_hv(y - 1 - y_stride, y_stride, z, b->src_stride, &sse);
467 break;
468 case 1:
469 this_mv.as_mv.col += 4;
470 this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
471 thismse = vfp->svf_halfpix_hv(y - y_stride, y_stride, z, b->src_stride, &sse);
472 break;
473 case 2:
474 this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
475 this_mv.as_mv.row += 4;
476 thismse = vfp->svf_halfpix_hv(y - 1, y_stride, z, b->src_stride, &sse);
477 break;
478 case 3:
479 default:
480 this_mv.as_mv.col += 4;
481 this_mv.as_mv.row += 4;
482 thismse = vfp->svf_halfpix_hv(y, y_stride, z, b->src_stride, &sse);
483 break;
484 }
485
486 diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
487
488 if (diag < bestmse)
489 {
490 *bestmv = this_mv;
491 bestmse = diag;
492 *distortion = thismse;
493 *sse1 = sse;
494 }
495
496
497 /* time to check quarter pels. */
498 if (bestmv->as_mv.row < startmv.as_mv.row)
499 y -= y_stride;
500
501 if (bestmv->as_mv.col < startmv.as_mv.col)
502 y--;
503
504 startmv = *bestmv;
505
506
507
508 /* go left then right and check error */
509 this_mv.as_mv.row = startmv.as_mv.row;
510
511 if (startmv.as_mv.col & 7)
512 {
513 this_mv.as_mv.col = startmv.as_mv.col - 2;
514 thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
515 }
516 else
517 {
518 this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
519 thismse = vfp->svf(y - 1, y_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
520 }
521
522 left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
523
524 if (left < bestmse)
525 {
526 *bestmv = this_mv;
527 bestmse = left;
528 *distortion = thismse;
529 *sse1 = sse;
530 }
531
532 this_mv.as_mv.col += 4;
533 thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
534 right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
535
536 if (right < bestmse)
537 {
538 *bestmv = this_mv;
539 bestmse = right;
540 *distortion = thismse;
541 *sse1 = sse;
542 }
543
544 /* go up then down and check error */
545 this_mv.as_mv.col = startmv.as_mv.col;
546
547 if (startmv.as_mv.row & 7)
548 {
549 this_mv.as_mv.row = startmv.as_mv.row - 2;
550 thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
551 }
552 else
553 {
554 this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
555 thismse = vfp->svf(y - y_stride, y_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
556 }
557
558 up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
559
560 if (up < bestmse)
561 {
562 *bestmv = this_mv;
563 bestmse = up;
564 *distortion = thismse;
565 *sse1 = sse;
566 }
567
568 this_mv.as_mv.row += 4;
569 thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
570 down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
571
572 if (down < bestmse)
573 {
574 *bestmv = this_mv;
575 bestmse = down;
576 *distortion = thismse;
577 *sse1 = sse;
578 }
579
580
581 /* now check 1 more diagonal */
582 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
583
584 this_mv = startmv;
585
586 switch (whichdir)
587 {
588 case 0:
589
590 if (startmv.as_mv.row & 7)
591 {
592 this_mv.as_mv.row -= 2;
593
594 if (startmv.as_mv.col & 7)
595 {
596 this_mv.as_mv.col -= 2;
597 thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
598 }
599 else
600 {
601 this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
602 thismse = vfp->svf(y - 1, y_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);;
603 }
604 }
605 else
606 {
607 this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
608
609 if (startmv.as_mv.col & 7)
610 {
611 this_mv.as_mv.col -= 2;
612 thismse = vfp->svf(y - y_stride, y_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
613 }
614 else
615 {
616 this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
617 thismse = vfp->svf(y - y_stride - 1, y_stride, 6, 6, z, b->src_stride, &sse);
618 }
619 }
620
621 break;
622 case 1:
623 this_mv.as_mv.col += 2;
624
625 if (startmv.as_mv.row & 7)
626 {
627 this_mv.as_mv.row -= 2;
628 thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
629 }
630 else
631 {
632 this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
633 thismse = vfp->svf(y - y_stride, y_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
634 }
635
636 break;
637 case 2:
638 this_mv.as_mv.row += 2;
639
640 if (startmv.as_mv.col & 7)
641 {
642 this_mv.as_mv.col -= 2;
643 thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
644 }
645 else
646 {
647 this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
648 thismse = vfp->svf(y - 1, y_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
649 }
650
651 break;
652 case 3:
653 this_mv.as_mv.col += 2;
654 this_mv.as_mv.row += 2;
655 thismse = vfp->svf(y, y_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
656 break;
657 }
658
659 diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
660
661 if (diag < bestmse)
662 {
663 *bestmv = this_mv;
664 bestmse = diag;
665 *distortion = thismse;
666 *sse1 = sse;
667 }
668
669 return bestmse;
670 }
671
672 int vp8_find_best_half_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
673 int_mv *bestmv, int_mv *ref_mv,
674 int error_per_bit,
675 const vp8_variance_fn_ptr_t *vfp,
676 int *mvcost[2], int *distortion,
677 unsigned int *sse1)
678 {
679 int bestmse = INT_MAX;
680 int_mv startmv;
681 int_mv this_mv;
682 unsigned char *z = (*(b->base_src) + b->src);
683 int left, right, up, down, diag;
684 unsigned int sse;
685 int whichdir ;
686 int thismse;
687 int y_stride;
688 int pre_stride = x->e_mbd.pre.y_stride;
689 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
690
691 #if ARCH_X86 || ARCH_X86_64
692 MACROBLOCKD *xd = &x->e_mbd;
693 unsigned char *y_0 = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
694 unsigned char *y;
695
696 y_stride = 32;
697 /* Copy 18 rows x 32 cols area to intermediate buffer before searching. */
698 vfp->copymem(y_0 - 1 - pre_stride, pre_stride, xd->y_buf, y_stride, 18);
699 y = xd->y_buf + y_stride + 1;
700 #else
701 unsigned char *y = base_pre + d->offset + (bestmv->as_mv.row) * pre_stride + bestmv->as_mv.col;
702 y_stride = pre_stride;
703 #endif
704
705 /* central mv */
706 bestmv->as_mv.row *= 8;
707 bestmv->as_mv.col *= 8;
708 startmv = *bestmv;
709
710 /* calculate central point error */
711 bestmse = vfp->vf(y, y_stride, z, b->src_stride, sse1);
712 *distortion = bestmse;
713 bestmse += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
714
715 /* go left then right and check error */
716 this_mv.as_mv.row = startmv.as_mv.row;
717 this_mv.as_mv.col = ((startmv.as_mv.col - 8) | 4);
718 thismse = vfp->svf_halfpix_h(y - 1, y_stride, z, b->src_stride, &sse);
719 left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
720
721 if (left < bestmse)
722 {
723 *bestmv = this_mv;
724 bestmse = left;
725 *distortion = thismse;
726 *sse1 = sse;
727 }
728
729 this_mv.as_mv.col += 8;
730 thismse = vfp->svf_halfpix_h(y, y_stride, z, b->src_stride, &sse);
731 right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
732
733 if (right < bestmse)
734 {
735 *bestmv = this_mv;
736 bestmse = right;
737 *distortion = thismse;
738 *sse1 = sse;
739 }
740
741 /* go up then down and check error */
742 this_mv.as_mv.col = startmv.as_mv.col;
743 this_mv.as_mv.row = ((startmv.as_mv.row - 8) | 4);
744 thismse = vfp->svf_halfpix_v(y - y_stride, y_stride, z, b->src_stride, &sse);
745 up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
746
747 if (up < bestmse)
748 {
749 *bestmv = this_mv;
750 bestmse = up;
751 *distortion = thismse;
752 *sse1 = sse;
753 }
754
755 this_mv.as_mv.row += 8;
756 thismse = vfp->svf_halfpix_v(y, y_stride, z, b->src_stride, &sse);
757 down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
758
759 if (down < bestmse)
760 {
761 *bestmv = this_mv;
762 bestmse = down;
763 *distortion = thismse;
764 *sse1 = sse;
765 }
766
767 /* now check 1 more diagonal - */
768 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
769 this_mv = startmv;
770
771 switch (whichdir)
772 {
773 case 0:
774 this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
775 this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
776 thismse = vfp->svf_halfpix_hv(y - 1 - y_stride, y_stride, z, b->src_stride, &sse);
777 break;
778 case 1:
779 this_mv.as_mv.col += 4;
780 this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
781 thismse = vfp->svf_halfpix_hv(y - y_stride, y_stride, z, b->src_stride, &sse);
782 break;
783 case 2:
784 this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
785 this_mv.as_mv.row += 4;
786 thismse = vfp->svf_halfpix_hv(y - 1, y_stride, z, b->src_stride, &sse);
787 break;
788 case 3:
789 default:
790 this_mv.as_mv.col += 4;
791 this_mv.as_mv.row += 4;
792 thismse = vfp->svf_halfpix_hv(y, y_stride, z, b->src_stride, &sse);
793 break;
794 }
795
796 diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
797
798 if (diag < bestmse)
799 {
800 *bestmv = this_mv;
801 bestmse = diag;
802 *distortion = thismse;
803 *sse1 = sse;
804 }
805
806 return bestmse;
807 }
808
809 #define CHECK_BOUNDS(range) \
810 {\
811 all_in = 1;\
812 all_in &= ((br-range) >= x->mv_row_min);\
813 all_in &= ((br+range) <= x->mv_row_max);\
814 all_in &= ((bc-range) >= x->mv_col_min);\
815 all_in &= ((bc+range) <= x->mv_col_max);\
816 }
817
818 #define CHECK_POINT \
819 {\
820 if (this_mv.as_mv.col < x->mv_col_min) continue;\
821 if (this_mv.as_mv.col > x->mv_col_max) continue;\
822 if (this_mv.as_mv.row < x->mv_row_min) continue;\
823 if (this_mv.as_mv.row > x->mv_row_max) continue;\
824 }
825
826 #define CHECK_BETTER \
827 {\
828 if (thissad < bestsad)\
829 {\
830 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, sad_per_bit);\
831 if (thissad < bestsad)\
832 {\
833 bestsad = thissad;\
834 best_site = i;\
835 }\
836 }\
837 }
838
839 static const MV next_chkpts[6][3] =
840 {
841 {{ -2, 0}, { -1, -2}, {1, -2}},
842 {{ -1, -2}, {1, -2}, {2, 0}},
843 {{1, -2}, {2, 0}, {1, 2}},
844 {{2, 0}, {1, 2}, { -1, 2}},
845 {{1, 2}, { -1, 2}, { -2, 0}},
846 {{ -1, 2}, { -2, 0}, { -1, -2}}
847 };
848
849 int vp8_hex_search
850 (
851 MACROBLOCK *x,
852 BLOCK *b,
853 BLOCKD *d,
854 int_mv *ref_mv,
855 int_mv *best_mv,
856 int search_param,
857 int sad_per_bit,
858 const vp8_variance_fn_ptr_t *vfp,
859 int *mvsadcost[2],
860 int *mvcost[2],
861 int_mv *center_mv
862 )
863 {
864 MV hex[6] = { { -1, -2}, {1, -2}, {2, 0}, {1, 2}, { -1, 2}, { -2, 0} } ;
865 MV neighbors[4] = {{0, -1}, { -1, 0}, {1, 0}, {0, 1}} ;
866 int i, j;
867
868 unsigned char *what = (*(b->base_src) + b->src);
869 int what_stride = b->src_stride;
870 int pre_stride = x->e_mbd.pre.y_stride;
871 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
872
873 int in_what_stride = pre_stride;
874 int br, bc;
875 int_mv this_mv;
876 unsigned int bestsad;
877 unsigned int thissad;
878 unsigned char *base_offset;
879 unsigned char *this_offset;
880 int k = -1;
881 int all_in;
882 int best_site = -1;
883 int hex_range = 127;
884 int dia_range = 8;
885
886 int_mv fcenter_mv;
887 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
888 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
889
890 /* adjust ref_mv to make sure it is within MV range */
891 vp8_clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
892 br = ref_mv->as_mv.row;
893 bc = ref_mv->as_mv.col;
894
895 /* Work out the start point for the search */
896 base_offset = (unsigned char *)(base_pre + d->offset);
897 this_offset = base_offset + (br * (pre_stride)) + bc;
898 this_mv.as_mv.row = br;
899 this_mv.as_mv.col = bc;
900 bestsad = vfp->sdf(what, what_stride, this_offset, in_what_stride, UINT_MAX)
901 + mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, sad_per_bit);
902
903 #if CONFIG_MULTI_RES_ENCODING
904 /* Lower search range based on prediction info */
905 if (search_param >= 6) goto cal_neighbors;
906 else if (search_param >= 5) hex_range = 4;
907 else if (search_param >= 4) hex_range = 6;
908 else if (search_param >= 3) hex_range = 15;
909 else if (search_param >= 2) hex_range = 31;
910 else if (search_param >= 1) hex_range = 63;
911
912 dia_range = 8;
913 #endif
914
915 /* hex search */
916 CHECK_BOUNDS(2)
917
918 if(all_in)
919 {
920 for (i = 0; i < 6; i++)
921 {
922 this_mv.as_mv.row = br + hex[i].row;
923 this_mv.as_mv.col = bc + hex[i].col;
924 this_offset = base_offset + (this_mv.as_mv.row * in_what_stride) + this_mv.as_mv.col;
925 thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, bestsad);
926 CHECK_BETTER
927 }
928 }else
929 {
930 for (i = 0; i < 6; i++)
931 {
932 this_mv.as_mv.row = br + hex[i].row;
933 this_mv.as_mv.col = bc + hex[i].col;
934 CHECK_POINT
935 this_offset = base_offset + (this_mv.as_mv.row * in_what_stride) + this_mv.as_mv.col;
936 thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, bestsad);
937 CHECK_BETTER
938 }
939 }
940
941 if (best_site == -1)
942 goto cal_neighbors;
943 else
944 {
945 br += hex[best_site].row;
946 bc += hex[best_site].col;
947 k = best_site;
948 }
949
950 for (j = 1; j < hex_range; j++)
951 {
952 best_site = -1;
953 CHECK_BOUNDS(2)
954
955 if(all_in)
956 {
957 for (i = 0; i < 3; i++)
958 {
959 this_mv.as_mv.row = br + next_chkpts[k][i].row;
960 this_mv.as_mv.col = bc + next_chkpts[k][i].col;
961 this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
962 thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, bestsad);
963 CHECK_BETTER
964 }
965 }else
966 {
967 for (i = 0; i < 3; i++)
968 {
969 this_mv.as_mv.row = br + next_chkpts[k][i].row;
970 this_mv.as_mv.col = bc + next_chkpts[k][i].col;
971 CHECK_POINT
972 this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
973 thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, bestsad);
974 CHECK_BETTER
975 }
976 }
977
978 if (best_site == -1)
979 break;
980 else
981 {
982 br += next_chkpts[k][best_site].row;
983 bc += next_chkpts[k][best_site].col;
984 k += 5 + best_site;
985 if (k >= 12) k -= 12;
986 else if (k >= 6) k -= 6;
987 }
988 }
989
990 /* check 4 1-away neighbors */
991 cal_neighbors:
992 for (j = 0; j < dia_range; j++)
993 {
994 best_site = -1;
995 CHECK_BOUNDS(1)
996
997 if(all_in)
998 {
999 for (i = 0; i < 4; i++)
1000 {
1001 this_mv.as_mv.row = br + neighbors[i].row;
1002 this_mv.as_mv.col = bc + neighbors[i].col;
1003 this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
1004 thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, bestsad);
1005 CHECK_BETTER
1006 }
1007 }else
1008 {
1009 for (i = 0; i < 4; i++)
1010 {
1011 this_mv.as_mv.row = br + neighbors[i].row;
1012 this_mv.as_mv.col = bc + neighbors[i].col;
1013 CHECK_POINT
1014 this_offset = base_offset + (this_mv.as_mv.row * (in_what_stride)) + this_mv.as_mv.col;
1015 thissad = vfp->sdf(what, what_stride, this_offset, in_what_stride, bestsad);
1016 CHECK_BETTER
1017 }
1018 }
1019
1020 if (best_site == -1)
1021 break;
1022 else
1023 {
1024 br += neighbors[best_site].row;
1025 bc += neighbors[best_site].col;
1026 }
1027 }
1028
1029 best_mv->as_mv.row = br;
1030 best_mv->as_mv.col = bc;
1031
1032 return bestsad;
1033 }
1034 #undef CHECK_BOUNDS
1035 #undef CHECK_POINT
1036 #undef CHECK_BETTER
1037
1038 int vp8_diamond_search_sad_c
1039 (
1040 MACROBLOCK *x,
1041 BLOCK *b,
1042 BLOCKD *d,
1043 int_mv *ref_mv,
1044 int_mv *best_mv,
1045 int search_param,
1046 int sad_per_bit,
1047 int *num00,
1048 vp8_variance_fn_ptr_t *fn_ptr,
1049 int *mvcost[2],
1050 int_mv *center_mv
1051 )
1052 {
1053 int i, j, step;
1054
1055 unsigned char *what = (*(b->base_src) + b->src);
1056 int what_stride = b->src_stride;
1057 unsigned char *in_what;
1058 int pre_stride = x->e_mbd.pre.y_stride;
1059 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
1060 int in_what_stride = pre_stride;
1061 unsigned char *best_address;
1062
1063 int tot_steps;
1064 int_mv this_mv;
1065
1066 unsigned int bestsad;
1067 unsigned int thissad;
1068 int best_site = 0;
1069 int last_site = 0;
1070
1071 int ref_row;
1072 int ref_col;
1073 int this_row_offset;
1074 int this_col_offset;
1075 search_site *ss;
1076
1077 unsigned char *check_here;
1078
1079 int *mvsadcost[2];
1080 int_mv fcenter_mv;
1081
1082 mvsadcost[0] = x->mvsadcost[0];
1083 mvsadcost[1] = x->mvsadcost[1];
1084 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1085 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1086
1087 vp8_clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
1088 ref_row = ref_mv->as_mv.row;
1089 ref_col = ref_mv->as_mv.col;
1090 *num00 = 0;
1091 best_mv->as_mv.row = ref_row;
1092 best_mv->as_mv.col = ref_col;
1093
1094 /* Work out the start point for the search */
1095 in_what = (unsigned char *)(base_pre + d->offset + (ref_row * pre_stride) + ref_col);
1096 best_address = in_what;
1097
1098 /* Check the starting position */
1099 bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, UINT_MAX)
1100 + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
1101
1102 /* search_param determines the length of the initial step and hence
1103 * the number of iterations 0 = initial step (MAX_FIRST_STEP) pel :
1104 * 1 = (MAX_FIRST_STEP/2) pel, 2 = (MAX_FIRST_STEP/4) pel... etc.
1105 */
1106 ss = &x->ss[search_param * x->searches_per_step];
1107 tot_steps = (x->ss_count / x->searches_per_step) - search_param;
1108
1109 i = 1;
1110
1111 for (step = 0; step < tot_steps ; step++)
1112 {
1113 for (j = 0 ; j < x->searches_per_step ; j++)
1114 {
1115 /* Trap illegal vectors */
1116 this_row_offset = best_mv->as_mv.row + ss[i].mv.row;
1117 this_col_offset = best_mv->as_mv.col + ss[i].mv.col;
1118
1119 if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
1120 (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
1121
1122 {
1123 check_here = ss[i].offset + best_address;
1124 thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, bestsad);
1125
1126 if (thissad < bestsad)
1127 {
1128 this_mv.as_mv.row = this_row_offset;
1129 this_mv.as_mv.col = this_col_offset;
1130 thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
1131 mvsadcost, sad_per_bit);
1132
1133 if (thissad < bestsad)
1134 {
1135 bestsad = thissad;
1136 best_site = i;
1137 }
1138 }
1139 }
1140
1141 i++;
1142 }
1143
1144 if (best_site != last_site)
1145 {
1146 best_mv->as_mv.row += ss[best_site].mv.row;
1147 best_mv->as_mv.col += ss[best_site].mv.col;
1148 best_address += ss[best_site].offset;
1149 last_site = best_site;
1150 }
1151 else if (best_address == in_what)
1152 (*num00)++;
1153 }
1154
1155 this_mv.as_mv.row = best_mv->as_mv.row << 3;
1156 this_mv.as_mv.col = best_mv->as_mv.col << 3;
1157
1158 return fn_ptr->vf(what, what_stride, best_address, in_what_stride, &thissad)
1159 + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
1160 }
1161
1162 int vp8_diamond_search_sadx4
1163 (
1164 MACROBLOCK *x,
1165 BLOCK *b,
1166 BLOCKD *d,
1167 int_mv *ref_mv,
1168 int_mv *best_mv,
1169 int search_param,
1170 int sad_per_bit,
1171 int *num00,
1172 vp8_variance_fn_ptr_t *fn_ptr,
1173 int *mvcost[2],
1174 int_mv *center_mv
1175 )
1176 {
1177 int i, j, step;
1178
1179 unsigned char *what = (*(b->base_src) + b->src);
1180 int what_stride = b->src_stride;
1181 unsigned char *in_what;
1182 int pre_stride = x->e_mbd.pre.y_stride;
1183 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
1184 int in_what_stride = pre_stride;
1185 unsigned char *best_address;
1186
1187 int tot_steps;
1188 int_mv this_mv;
1189
1190 unsigned int bestsad;
1191 unsigned int thissad;
1192 int best_site = 0;
1193 int last_site = 0;
1194
1195 int ref_row;
1196 int ref_col;
1197 int this_row_offset;
1198 int this_col_offset;
1199 search_site *ss;
1200
1201 unsigned char *check_here;
1202
1203 int *mvsadcost[2];
1204 int_mv fcenter_mv;
1205
1206 mvsadcost[0] = x->mvsadcost[0];
1207 mvsadcost[1] = x->mvsadcost[1];
1208 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1209 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1210
1211 vp8_clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
1212 ref_row = ref_mv->as_mv.row;
1213 ref_col = ref_mv->as_mv.col;
1214 *num00 = 0;
1215 best_mv->as_mv.row = ref_row;
1216 best_mv->as_mv.col = ref_col;
1217
1218 /* Work out the start point for the search */
1219 in_what = (unsigned char *)(base_pre + d->offset + (ref_row * pre_stride) + ref_col);
1220 best_address = in_what;
1221
1222 /* Check the starting position */
1223 bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, UINT_MAX)
1224 + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
1225
1226 /* search_param determines the length of the initial step and hence the
1227 * number of iterations 0 = initial step (MAX_FIRST_STEP) pel : 1 =
1228 * (MAX_FIRST_STEP/2) pel, 2 = (MAX_FIRST_STEP/4) pel... etc.
1229 */
1230 ss = &x->ss[search_param * x->searches_per_step];
1231 tot_steps = (x->ss_count / x->searches_per_step) - search_param;
1232
1233 i = 1;
1234
1235 for (step = 0; step < tot_steps ; step++)
1236 {
1237 int all_in = 1, t;
1238
1239 /* To know if all neighbor points are within the bounds, 4 bounds
1240 * checking are enough instead of checking 4 bounds for each
1241 * points.
1242 */
1243 all_in &= ((best_mv->as_mv.row + ss[i].mv.row)> x->mv_row_min);
1244 all_in &= ((best_mv->as_mv.row + ss[i+1].mv.row) < x->mv_row_max);
1245 all_in &= ((best_mv->as_mv.col + ss[i+2].mv.col) > x->mv_col_min);
1246 all_in &= ((best_mv->as_mv.col + ss[i+3].mv.col) < x->mv_col_max);
1247
1248 if (all_in)
1249 {
1250 unsigned int sad_array[4];
1251
1252 for (j = 0 ; j < x->searches_per_step ; j += 4)
1253 {
1254 const unsigned char *block_offset[4];
1255
1256 for (t = 0; t < 4; t++)
1257 block_offset[t] = ss[i+t].offset + best_address;
1258
1259 fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, sad_array);
1260
1261 for (t = 0; t < 4; t++, i++)
1262 {
1263 if (sad_array[t] < bestsad)
1264 {
1265 this_mv.as_mv.row = best_mv->as_mv.row + ss[i].mv.row;
1266 this_mv.as_mv.col = best_mv->as_mv.col + ss[i].mv.col;
1267 sad_array[t] += mvsad_err_cost(&this_mv, &fcenter_mv,
1268 mvsadcost, sad_per_bit);
1269
1270 if (sad_array[t] < bestsad)
1271 {
1272 bestsad = sad_array[t];
1273 best_site = i;
1274 }
1275 }
1276 }
1277 }
1278 }
1279 else
1280 {
1281 for (j = 0 ; j < x->searches_per_step ; j++)
1282 {
1283 /* Trap illegal vectors */
1284 this_row_offset = best_mv->as_mv.row + ss[i].mv.row;
1285 this_col_offset = best_mv->as_mv.col + ss[i].mv.col;
1286
1287 if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
1288 (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
1289 {
1290 check_here = ss[i].offset + best_address;
1291 thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, bestsad);
1292
1293 if (thissad < bestsad)
1294 {
1295 this_mv.as_mv.row = this_row_offset;
1296 this_mv.as_mv.col = this_col_offset;
1297 thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
1298 mvsadcost, sad_per_bit);
1299
1300 if (thissad < bestsad)
1301 {
1302 bestsad = thissad;
1303 best_site = i;
1304 }
1305 }
1306 }
1307 i++;
1308 }
1309 }
1310
1311 if (best_site != last_site)
1312 {
1313 best_mv->as_mv.row += ss[best_site].mv.row;
1314 best_mv->as_mv.col += ss[best_site].mv.col;
1315 best_address += ss[best_site].offset;
1316 last_site = best_site;
1317 }
1318 else if (best_address == in_what)
1319 (*num00)++;
1320 }
1321
1322 this_mv.as_mv.row = best_mv->as_mv.row * 8;
1323 this_mv.as_mv.col = best_mv->as_mv.col * 8;
1324
1325 return fn_ptr->vf(what, what_stride, best_address, in_what_stride, &thissad)
1326 + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
1327 }
1328
1329 int vp8_full_search_sad_c(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
1330 int sad_per_bit, int distance,
1331 vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
1332 int_mv *center_mv)
1333 {
1334 unsigned char *what = (*(b->base_src) + b->src);
1335 int what_stride = b->src_stride;
1336 unsigned char *in_what;
1337 int pre_stride = x->e_mbd.pre.y_stride;
1338 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
1339 int in_what_stride = pre_stride;
1340 int mv_stride = pre_stride;
1341 unsigned char *bestaddress;
1342 int_mv *best_mv = &d->bmi.mv;
1343 int_mv this_mv;
1344 unsigned int bestsad;
1345 unsigned int thissad;
1346 int r, c;
1347
1348 unsigned char *check_here;
1349
1350 int ref_row = ref_mv->as_mv.row;
1351 int ref_col = ref_mv->as_mv.col;
1352
1353 int row_min = ref_row - distance;
1354 int row_max = ref_row + distance;
1355 int col_min = ref_col - distance;
1356 int col_max = ref_col + distance;
1357
1358 int *mvsadcost[2];
1359 int_mv fcenter_mv;
1360
1361 mvsadcost[0] = x->mvsadcost[0];
1362 mvsadcost[1] = x->mvsadcost[1];
1363 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1364 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1365
1366 /* Work out the mid point for the search */
1367 in_what = base_pre + d->offset;
1368 bestaddress = in_what + (ref_row * pre_stride) + ref_col;
1369
1370 best_mv->as_mv.row = ref_row;
1371 best_mv->as_mv.col = ref_col;
1372
1373 /* Baseline value at the centre */
1374 bestsad = fn_ptr->sdf(what, what_stride, bestaddress,
1375 in_what_stride, UINT_MAX)
1376 + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
1377
1378 /* Apply further limits to prevent us looking using vectors that
1379 * stretch beyiond the UMV border
1380 */
1381 if (col_min < x->mv_col_min)
1382 col_min = x->mv_col_min;
1383
1384 if (col_max > x->mv_col_max)
1385 col_max = x->mv_col_max;
1386
1387 if (row_min < x->mv_row_min)
1388 row_min = x->mv_row_min;
1389
1390 if (row_max > x->mv_row_max)
1391 row_max = x->mv_row_max;
1392
1393 for (r = row_min; r < row_max ; r++)
1394 {
1395 this_mv.as_mv.row = r;
1396 check_here = r * mv_stride + in_what + col_min;
1397
1398 for (c = col_min; c < col_max; c++)
1399 {
1400 thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, bestsad);
1401
1402 this_mv.as_mv.col = c;
1403 thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
1404 mvsadcost, sad_per_bit);
1405
1406 if (thissad < bestsad)
1407 {
1408 bestsad = thissad;
1409 best_mv->as_mv.row = r;
1410 best_mv->as_mv.col = c;
1411 bestaddress = check_here;
1412 }
1413
1414 check_here++;
1415 }
1416 }
1417
1418 this_mv.as_mv.row = best_mv->as_mv.row << 3;
1419 this_mv.as_mv.col = best_mv->as_mv.col << 3;
1420
1421 return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, &thissad)
1422 + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
1423 }
1424
1425 int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
1426 int sad_per_bit, int distance,
1427 vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
1428 int_mv *center_mv)
1429 {
1430 unsigned char *what = (*(b->base_src) + b->src);
1431 int what_stride = b->src_stride;
1432 unsigned char *in_what;
1433 int pre_stride = x->e_mbd.pre.y_stride;
1434 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
1435 int in_what_stride = pre_stride;
1436 int mv_stride = pre_stride;
1437 unsigned char *bestaddress;
1438 int_mv *best_mv = &d->bmi.mv;
1439 int_mv this_mv;
1440 unsigned int bestsad;
1441 unsigned int thissad;
1442 int r, c;
1443
1444 unsigned char *check_here;
1445
1446 int ref_row = ref_mv->as_mv.row;
1447 int ref_col = ref_mv->as_mv.col;
1448
1449 int row_min = ref_row - distance;
1450 int row_max = ref_row + distance;
1451 int col_min = ref_col - distance;
1452 int col_max = ref_col + distance;
1453
1454 unsigned int sad_array[3];
1455
1456 int *mvsadcost[2];
1457 int_mv fcenter_mv;
1458
1459 mvsadcost[0] = x->mvsadcost[0];
1460 mvsadcost[1] = x->mvsadcost[1];
1461 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1462 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1463
1464 /* Work out the mid point for the search */
1465 in_what = base_pre + d->offset;
1466 bestaddress = in_what + (ref_row * pre_stride) + ref_col;
1467
1468 best_mv->as_mv.row = ref_row;
1469 best_mv->as_mv.col = ref_col;
1470
1471 /* Baseline value at the centre */
1472 bestsad = fn_ptr->sdf(what, what_stride, bestaddress,
1473 in_what_stride, UINT_MAX)
1474 + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
1475
1476 /* Apply further limits to prevent us looking using vectors that stretch
1477 * beyond the UMV border
1478 */
1479 if (col_min < x->mv_col_min)
1480 col_min = x->mv_col_min;
1481
1482 if (col_max > x->mv_col_max)
1483 col_max = x->mv_col_max;
1484
1485 if (row_min < x->mv_row_min)
1486 row_min = x->mv_row_min;
1487
1488 if (row_max > x->mv_row_max)
1489 row_max = x->mv_row_max;
1490
1491 for (r = row_min; r < row_max ; r++)
1492 {
1493 this_mv.as_mv.row = r;
1494 check_here = r * mv_stride + in_what + col_min;
1495 c = col_min;
1496
1497 while ((c + 2) < col_max)
1498 {
1499 int i;
1500
1501 fn_ptr->sdx3f(what, what_stride, check_here, in_what_stride, sad_array);
1502
1503 for (i = 0; i < 3; i++)
1504 {
1505 thissad = sad_array[i];
1506
1507 if (thissad < bestsad)
1508 {
1509 this_mv.as_mv.col = c;
1510 thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
1511 mvsadcost, sad_per_bit);
1512
1513 if (thissad < bestsad)
1514 {
1515 bestsad = thissad;
1516 best_mv->as_mv.row = r;
1517 best_mv->as_mv.col = c;
1518 bestaddress = check_here;
1519 }
1520 }
1521
1522 check_here++;
1523 c++;
1524 }
1525 }
1526
1527 while (c < col_max)
1528 {
1529 thissad = fn_ptr->sdf(what, what_stride, check_here, in_what_stride, bestsad);
1530
1531 if (thissad < bestsad)
1532 {
1533 this_mv.as_mv.col = c;
1534 thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
1535 mvsadcost, sad_per_bit);
1536
1537 if (thissad < bestsad)
1538 {
1539 bestsad = thissad;
1540 best_mv->as_mv.row = r;
1541 best_mv->as_mv.col = c;
1542 bestaddress = check_here;
1543 }
1544 }
1545
1546 check_here ++;
1547 c ++;
1548 }
1549
1550 }
1551
1552 this_mv.as_mv.row = best_mv->as_mv.row << 3;
1553 this_mv.as_mv.col = best_mv->as_mv.col << 3;
1554
1555 return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, &thissad)
1556 + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
1557 }
1558
1559 int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
1560 int sad_per_bit, int distance,
1561 vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
1562 int_mv *center_mv)
1563 {
1564 unsigned char *what = (*(b->base_src) + b->src);
1565 int what_stride = b->src_stride;
1566 int pre_stride = x->e_mbd.pre.y_stride;
1567 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
1568 unsigned char *in_what;
1569 int in_what_stride = pre_stride;
1570 int mv_stride = pre_stride;
1571 unsigned char *bestaddress;
1572 int_mv *best_mv = &d->bmi.mv;
1573 int_mv this_mv;
1574 unsigned int bestsad;
1575 unsigned int thissad;
1576 int r, c;
1577
1578 unsigned char *check_here;
1579
1580 int ref_row = ref_mv->as_mv.row;
1581 int ref_col = ref_mv->as_mv.col;
1582
1583 int row_min = ref_row - distance;
1584 int row_max = ref_row + distance;
1585 int col_min = ref_col - distance;
1586 int col_max = ref_col + distance;
1587
1588 DECLARE_ALIGNED_ARRAY(16, unsigned short, sad_array8, 8);
1589 unsigned int sad_array[3];
1590
1591 int *mvsadcost[2];
1592 int_mv fcenter_mv;
1593
1594 mvsadcost[0] = x->mvsadcost[0];
1595 mvsadcost[1] = x->mvsadcost[1];
1596 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1597 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1598
1599 /* Work out the mid point for the search */
1600 in_what = base_pre + d->offset;
1601 bestaddress = in_what + (ref_row * pre_stride) + ref_col;
1602
1603 best_mv->as_mv.row = ref_row;
1604 best_mv->as_mv.col = ref_col;
1605
1606 /* Baseline value at the centre */
1607 bestsad = fn_ptr->sdf(what, what_stride,
1608 bestaddress, in_what_stride, UINT_MAX)
1609 + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, sad_per_bit);
1610
1611 /* Apply further limits to prevent us looking using vectors that stretch
1612 * beyond the UMV border
1613 */
1614 if (col_min < x->mv_col_min)
1615 col_min = x->mv_col_min;
1616
1617 if (col_max > x->mv_col_max)
1618 col_max = x->mv_col_max;
1619
1620 if (row_min < x->mv_row_min)
1621 row_min = x->mv_row_min;
1622
1623 if (row_max > x->mv_row_max)
1624 row_max = x->mv_row_max;
1625
1626 for (r = row_min; r < row_max ; r++)
1627 {
1628 this_mv.as_mv.row = r;
1629 check_here = r * mv_stride + in_what + col_min;
1630 c = col_min;
1631
1632 while ((c + 7) < col_max)
1633 {
1634 int i;
1635
1636 fn_ptr->sdx8f(what, what_stride, check_here, in_what_stride, sad_array8);
1637
1638 for (i = 0; i < 8; i++)
1639 {
1640 thissad = sad_array8[i];
1641
1642 if (thissad < bestsad)
1643 {
1644 this_mv.as_mv.col = c;
1645 thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
1646 mvsadcost, sad_per_bit);
1647
1648 if (thissad < bestsad)
1649 {
1650 bestsad = thissad;
1651 best_mv->as_mv.row = r;
1652 best_mv->as_mv.col = c;
1653 bestaddress = check_here;
1654 }
1655 }
1656
1657 check_here++;
1658 c++;
1659 }
1660 }
1661
1662 while ((c + 2) < col_max)
1663 {
1664 int i;
1665
1666 fn_ptr->sdx3f(what, what_stride, check_here , in_what_stride, sad_array);
1667
1668 for (i = 0; i < 3; i++)
1669 {
1670 thissad = sad_array[i];
1671
1672 if (thissad < bestsad)
1673 {
1674 this_mv.as_mv.col = c;
1675 thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
1676 mvsadcost, sad_per_bit);
1677
1678 if (thissad < bestsad)
1679 {
1680 bestsad = thissad;
1681 best_mv->as_mv.row = r;
1682 best_mv->as_mv.col = c;
1683 bestaddress = check_here;
1684 }
1685 }
1686
1687 check_here++;
1688 c++;
1689 }
1690 }
1691
1692 while (c < col_max)
1693 {
1694 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1695
1696 if (thissad < bestsad)
1697 {
1698 this_mv.as_mv.col = c;
1699 thissad += mvsad_err_cost(&this_mv, &fcenter_mv,
1700 mvsadcost, sad_per_bit);
1701
1702 if (thissad < bestsad)
1703 {
1704 bestsad = thissad;
1705 best_mv->as_mv.row = r;
1706 best_mv->as_mv.col = c;
1707 bestaddress = check_here;
1708 }
1709 }
1710
1711 check_here ++;
1712 c ++;
1713 }
1714 }
1715
1716 this_mv.as_mv.row = best_mv->as_mv.row * 8;
1717 this_mv.as_mv.col = best_mv->as_mv.col * 8;
1718
1719 return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, &thissad)
1720 + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
1721 }
1722
1723 int vp8_refining_search_sad_c(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
1724 int error_per_bit, int search_range,
1725 vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
1726 int_mv *center_mv)
1727 {
1728 MV neighbors[4] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0}};
1729 int i, j;
1730 short this_row_offset, this_col_offset;
1731
1732 int what_stride = b->src_stride;
1733 int pre_stride = x->e_mbd.pre.y_stride;
1734 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
1735 int in_what_stride = pre_stride;
1736 unsigned char *what = (*(b->base_src) + b->src);
1737 unsigned char *best_address = (unsigned char *)(base_pre + d->offset +
1738 (ref_mv->as_mv.row * pre_stride) + ref_mv->as_mv.col);
1739 unsigned char *check_here;
1740 int_mv this_mv;
1741 unsigned int bestsad;
1742 unsigned int thissad;
1743
1744 int *mvsadcost[2];
1745 int_mv fcenter_mv;
1746
1747 mvsadcost[0] = x->mvsadcost[0];
1748 mvsadcost[1] = x->mvsadcost[1];
1749 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1750 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1751
1752 bestsad = fn_ptr->sdf(what, what_stride, best_address,
1753 in_what_stride, UINT_MAX)
1754 + mvsad_err_cost(ref_mv, &fcenter_mv, mvsadcost, error_per_bit);
1755
1756 for (i=0; i<search_range; i++)
1757 {
1758 int best_site = -1;
1759
1760 for (j = 0 ; j < 4 ; j++)
1761 {
1762 this_row_offset = ref_mv->as_mv.row + neighbors[j].row;
1763 this_col_offset = ref_mv->as_mv.col + neighbors[j].col;
1764
1765 if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
1766 (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
1767 {
1768 check_here = (neighbors[j].row)*in_what_stride + neighbors[j].col + best_address;
1769 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1770
1771 if (thissad < bestsad)
1772 {
1773 this_mv.as_mv.row = this_row_offset;
1774 this_mv.as_mv.col = this_col_offset;
1775 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1776
1777 if (thissad < bestsad)
1778 {
1779 bestsad = thissad;
1780 best_site = j;
1781 }
1782 }
1783 }
1784 }
1785
1786 if (best_site == -1)
1787 break;
1788 else
1789 {
1790 ref_mv->as_mv.row += neighbors[best_site].row;
1791 ref_mv->as_mv.col += neighbors[best_site].col;
1792 best_address += (neighbors[best_site].row)*in_what_stride + neighbors[best_site].col;
1793 }
1794 }
1795
1796 this_mv.as_mv.row = ref_mv->as_mv.row << 3;
1797 this_mv.as_mv.col = ref_mv->as_mv.col << 3;
1798
1799 return fn_ptr->vf(what, what_stride, best_address, in_what_stride, &thissad)
1800 + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
1801 }
1802
1803 int vp8_refining_search_sadx4(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
1804 int_mv *ref_mv, int error_per_bit,
1805 int search_range, vp8_variance_fn_ptr_t *fn_ptr,
1806 int *mvcost[2], int_mv *center_mv)
1807 {
1808 MV neighbors[4] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0}};
1809 int i, j;
1810 short this_row_offset, this_col_offset;
1811
1812 int what_stride = b->src_stride;
1813 int pre_stride = x->e_mbd.pre.y_stride;
1814 unsigned char *base_pre = x->e_mbd.pre.y_buffer;
1815 int in_what_stride = pre_stride;
1816 unsigned char *what = (*(b->base_src) + b->src);
1817 unsigned char *best_address = (unsigned char *)(base_pre + d->offset +
1818 (ref_mv->as_mv.row * pre_stride) + ref_mv->as_mv.col);
1819 unsigned char *check_here;
1820 int_mv this_mv;
1821 unsigned int bestsad;
1822 unsigned int thissad;
1823
1824 int *mvsadcost[2];
1825 int_mv fcenter_mv;
1826
1827 mvsadcost[0] = x->mvsadcost[0];
1828 mvsadcost[1] = x->mvsadcost[1];
1829 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1830 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1831
1832 bestsad = fn_ptr->sdf(what, what_stride, best_address,
1833 in_what_stride, UINT_MAX)
1834 + mvsad_err_cost(ref_mv, &fcenter_mv, mvsadcost, error_per_bit);
1835
1836 for (i=0; i<search_range; i++)
1837 {
1838 int best_site = -1;
1839 int all_in = 1;
1840
1841 all_in &= ((ref_mv->as_mv.row - 1) > x->mv_row_min);
1842 all_in &= ((ref_mv->as_mv.row + 1) < x->mv_row_max);
1843 all_in &= ((ref_mv->as_mv.col - 1) > x->mv_col_min);
1844 all_in &= ((ref_mv->as_mv.col + 1) < x->mv_col_max);
1845
1846 if(all_in)
1847 {
1848 unsigned int sad_array[4];
1849 const unsigned char *block_offset[4];
1850 block_offset[0] = best_address - in_what_stride;
1851 block_offset[1] = best_address - 1;
1852 block_offset[2] = best_address + 1;
1853 block_offset[3] = best_address + in_what_stride;
1854
1855 fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, sad_array);
1856
1857 for (j = 0; j < 4; j++)
1858 {
1859 if (sad_array[j] < bestsad)
1860 {
1861 this_mv.as_mv.row = ref_mv->as_mv.row + neighbors[j].row;
1862 this_mv.as_mv.col = ref_mv->as_mv.col + neighbors[j].col;
1863 sad_array[j] += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1864
1865 if (sad_array[j] < bestsad)
1866 {
1867 bestsad = sad_array[j];
1868 best_site = j;
1869 }
1870 }
1871 }
1872 }
1873 else
1874 {
1875 for (j = 0 ; j < 4 ; j++)
1876 {
1877 this_row_offset = ref_mv->as_mv.row + neighbors[j].row;
1878 this_col_offset = ref_mv->as_mv.col + neighbors[j].col;
1879
1880 if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
1881 (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
1882 {
1883 check_here = (neighbors[j].row)*in_what_stride + neighbors[j].col + best_address;
1884 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1885
1886 if (thissad < bestsad)
1887 {
1888 this_mv.as_mv.row = this_row_offset;
1889 this_mv.as_mv.col = this_col_offset;
1890 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1891
1892 if (thissad < bestsad)
1893 {
1894 bestsad = thissad;
1895 best_site = j;
1896 }
1897 }
1898 }
1899 }
1900 }
1901
1902 if (best_site == -1)
1903 break;
1904 else
1905 {
1906 ref_mv->as_mv.row += neighbors[best_site].row;
1907 ref_mv->as_mv.col += neighbors[best_site].col;
1908 best_address += (neighbors[best_site].row)*in_what_stride + neighbors[best_site].col;
1909 }
1910 }
1911
1912 this_mv.as_mv.row = ref_mv->as_mv.row * 8;
1913 this_mv.as_mv.col = ref_mv->as_mv.col * 8;
1914
1915 return fn_ptr->vf(what, what_stride, best_address, in_what_stride, &thissad)
1916 + mv_err_cost(&this_mv, center_mv, mvcost, x->errorperbit);
1917 }
1918
1919 #ifdef VP8_ENTROPY_STATS
1920 void print_mode_context(void)
1921 {
1922 FILE *f = fopen("modecont.c", "w");
1923 int i, j;
1924
1925 fprintf(f, "#include \"entropy.h\"\n");
1926 fprintf(f, "const int vp8_mode_contexts[6][4] =\n");
1927 fprintf(f, "{\n");
1928
1929 for (j = 0; j < 6; j++)
1930 {
1931 fprintf(f, " { /* %d */\n", j);
1932 fprintf(f, " ");
1933
1934 for (i = 0; i < 4; i++)
1935 {
1936 int overal_prob;
1937 int this_prob;
1938 int count;
1939
1940 /* Overall probs */
1941 count = mv_mode_cts[i][0] + mv_mode_cts[i][1];
1942
1943 if (count)
1944 overal_prob = 256 * mv_mode_cts[i][0] / count;
1945 else
1946 overal_prob = 128;
1947
1948 if (overal_prob == 0)
1949 overal_prob = 1;
1950
1951 /* context probs */
1952 count = mv_ref_ct[j][i][0] + mv_ref_ct[j][i][1];
1953
1954 if (count)
1955 this_prob = 256 * mv_ref_ct[j][i][0] / count;
1956 else
1957 this_prob = 128;
1958
1959 if (this_prob == 0)
1960 this_prob = 1;
1961
1962 fprintf(f, "%5d, ", this_prob);
1963 }
1964
1965 fprintf(f, " },\n");
1966 }
1967
1968 fprintf(f, "};\n");
1969 fclose(f);
1970 }
1971
1972 /* MV ref count VP8_ENTROPY_STATS stats code */
1973 #ifdef VP8_ENTROPY_STATS
1974 void init_mv_ref_counts()
1975 {
1976 vpx_memset(mv_ref_ct, 0, sizeof(mv_ref_ct));
1977 vpx_memset(mv_mode_cts, 0, sizeof(mv_mode_cts));
1978 }
1979
1980 void accum_mv_refs(MB_PREDICTION_MODE m, const int ct[4])
1981 {
1982 if (m == ZEROMV)
1983 {
1984 ++mv_ref_ct [ct[0]] [0] [0];
1985 ++mv_mode_cts[0][0];
1986 }
1987 else
1988 {
1989 ++mv_ref_ct [ct[0]] [0] [1];
1990 ++mv_mode_cts[0][1];
1991
1992 if (m == NEARESTMV)
1993 {
1994 ++mv_ref_ct [ct[1]] [1] [0];
1995 ++mv_mode_cts[1][0];
1996 }
1997 else
1998 {
1999 ++mv_ref_ct [ct[1]] [1] [1];
2000 ++mv_mode_cts[1][1];
2001
2002 if (m == NEARMV)
2003 {
2004 ++mv_ref_ct [ct[2]] [2] [0];
2005 ++mv_mode_cts[2][0];
2006 }
2007 else
2008 {
2009 ++mv_ref_ct [ct[2]] [2] [1];
2010 ++mv_mode_cts[2][1];
2011
2012 if (m == NEWMV)
2013 {
2014 ++mv_ref_ct [ct[3]] [3] [0];
2015 ++mv_mode_cts[3][0];
2016 }
2017 else
2018 {
2019 ++mv_ref_ct [ct[3]] [3] [1];
2020 ++mv_mode_cts[3][1];
2021 }
2022 }
2023 }
2024 }
2025 }
2026
2027 #endif/* END MV ref count VP8_ENTROPY_STATS stats code */
2028
2029 #endif

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial