2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
13 #include "vpx_mem/vpx_mem.h"
20 static int mv_ref_ct [31] [4] [2];
21 static int mv_mode_cts [4] [2];
24 int vp8_mv_bit_cost(int_mv *mv, int_mv *ref, int *mvcost[2], int Weight)
26 // MV costing is based on the distribution of vectors in the previous frame and as such will tend to
27 // over state the cost of vectors. In addition coding a new vector can have a knock on effect on the
28 // cost of subsequent vectors and the quality of prediction from NEAR and NEAREST for subsequent blocks.
29 // The "Weight" parameter allows, to a limited extent, for some account to be taken of these factors.
30 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;
33 static int mv_err_cost(int_mv *mv, int_mv *ref, int *mvcost[2], int error_per_bit)
35 return ((mvcost[0][(mv->as_mv.row - ref->as_mv.row) >> 1] +
36 mvcost[1][(mv->as_mv.col - ref->as_mv.col) >> 1])
37 * error_per_bit + 128) >> 8;
40 static int mvsad_err_cost(int_mv *mv, int_mv *ref, int *mvsadcost[2], int error_per_bit)
42 /* Calculate sad error cost on full pixel basis. */
43 return ((mvsadcost[0][(mv->as_mv.row - ref->as_mv.row)] +
44 mvsadcost[1][(mv->as_mv.col - ref->as_mv.col)])
45 * error_per_bit + 128) >> 8;
48 void vp8_init_dsmotion_compensation(MACROBLOCK *x, int stride)
51 int search_site_count = 0;
54 // Generate offsets for 4 search sites per step.
56 x->ss[search_site_count].mv.col = 0;
57 x->ss[search_site_count].mv.row = 0;
58 x->ss[search_site_count].offset = 0;
64 // Compute offsets for search sites.
65 x->ss[search_site_count].mv.col = 0;
66 x->ss[search_site_count].mv.row = -Len;
67 x->ss[search_site_count].offset = -Len * stride;
70 // Compute offsets for search sites.
71 x->ss[search_site_count].mv.col = 0;
72 x->ss[search_site_count].mv.row = Len;
73 x->ss[search_site_count].offset = Len * stride;
76 // Compute offsets for search sites.
77 x->ss[search_site_count].mv.col = -Len;
78 x->ss[search_site_count].mv.row = 0;
79 x->ss[search_site_count].offset = -Len;
82 // Compute offsets for search sites.
83 x->ss[search_site_count].mv.col = Len;
84 x->ss[search_site_count].mv.row = 0;
85 x->ss[search_site_count].offset = Len;
92 x->ss_count = search_site_count;
93 x->searches_per_step = 4;
96 void vp8_init3smotion_compensation(MACROBLOCK *x, int stride)
99 int search_site_count = 0;
101 // Generate offsets for 8 search sites per step.
102 Len = MAX_FIRST_STEP;
103 x->ss[search_site_count].mv.col = 0;
104 x->ss[search_site_count].mv.row = 0;
105 x->ss[search_site_count].offset = 0;
111 // Compute offsets for search sites.
112 x->ss[search_site_count].mv.col = 0;
113 x->ss[search_site_count].mv.row = -Len;
114 x->ss[search_site_count].offset = -Len * stride;
117 // Compute offsets for search sites.
118 x->ss[search_site_count].mv.col = 0;
119 x->ss[search_site_count].mv.row = Len;
120 x->ss[search_site_count].offset = Len * stride;
123 // Compute offsets for search sites.
124 x->ss[search_site_count].mv.col = -Len;
125 x->ss[search_site_count].mv.row = 0;
126 x->ss[search_site_count].offset = -Len;
129 // Compute offsets for search sites.
130 x->ss[search_site_count].mv.col = Len;
131 x->ss[search_site_count].mv.row = 0;
132 x->ss[search_site_count].offset = Len;
135 // Compute offsets for search sites.
136 x->ss[search_site_count].mv.col = -Len;
137 x->ss[search_site_count].mv.row = -Len;
138 x->ss[search_site_count].offset = -Len * stride - Len;
141 // Compute offsets for search sites.
142 x->ss[search_site_count].mv.col = Len;
143 x->ss[search_site_count].mv.row = -Len;
144 x->ss[search_site_count].offset = -Len * stride + Len;
147 // Compute offsets for search sites.
148 x->ss[search_site_count].mv.col = -Len;
149 x->ss[search_site_count].mv.row = Len;
150 x->ss[search_site_count].offset = Len * stride - Len;
153 // Compute offsets for search sites.
154 x->ss[search_site_count].mv.col = Len;
155 x->ss[search_site_count].mv.row = Len;
156 x->ss[search_site_count].offset = Len * stride + Len;
164 x->ss_count = search_site_count;
165 x->searches_per_step = 8;
169 #define MVC(r,c) (((mvcost[0][(r)-rr] + mvcost[1][(c) - rc]) * error_per_bit + 128 )>>8 ) // estimated cost of a motion vector (r,c)
170 #define PRE(r,c) (*(d->base_pre) + d->pre + ((r)>>2) * d->pre_stride + ((c)>>2)) // pointer to predictor base of a motionvector
171 #define SP(x) (((x)&3)<<1) // convert motion vector component to offset for svf calc
172 #define DIST(r,c) vfp->svf( PRE(r,c), d->pre_stride, SP(c),SP(r), z,b->src_stride,&sse) // returns subpixel variance error function.
173 #define IFMVCV(r,c,s,e) if ( c >= minc && c <= maxc && r >= minr && r <= maxr) s else e;
174 #define ERR(r,c) (MVC(r,c)+DIST(r,c)) // returns distortion + motion vector cost
175 #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=INT_MAX;)// checks if (r,c) has better score than previous best
176 #define MIN(x,y) (((x)<(y))?(x):(y))
177 #define MAX(x,y) (((x)>(y))?(x):(y))
179 //#define CHECK_BETTER(v,r,c) if((v = ERR(r,c)) < besterr) { besterr = v; br=r; bc=c; }
181 int vp8_find_best_sub_pixel_step_iteratively(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
182 int_mv *bestmv, int_mv *ref_mv,
184 const vp8_variance_fn_ptr_t *vfp,
185 int *mvcost[2], int *distortion,
188 unsigned char *y = *(d->base_pre) + d->pre + (bestmv->as_mv.row) * d->pre_stride + bestmv->as_mv.col;
189 unsigned char *z = (*(b->base_src) + b->src);
191 int rr = ref_mv->as_mv.row >> 1, rc = ref_mv->as_mv.col >> 1;
192 int br = bestmv->as_mv.row << 2, bc = bestmv->as_mv.col << 2;
193 int tr = br, tc = bc;
194 unsigned int besterr = INT_MAX;
195 unsigned int left, right, up, down, diag;
197 unsigned int whichdir;
198 unsigned int halfiters = 4;
199 unsigned int quarteriters = 4;
202 int minc = MAX(x->mv_col_min << 2, (ref_mv->as_mv.col >> 1) - ((1 << mvlong_width) - 1));
203 int maxc = MIN(x->mv_col_max << 2, (ref_mv->as_mv.col >> 1) + ((1 << mvlong_width) - 1));
204 int minr = MAX(x->mv_row_min << 2, (ref_mv->as_mv.row >> 1) - ((1 << mvlong_width) - 1));
205 int maxr = MIN(x->mv_row_max << 2, (ref_mv->as_mv.row >> 1) + ((1 << mvlong_width) - 1));
208 bestmv->as_mv.row <<= 3;
209 bestmv->as_mv.col <<= 3;
211 // calculate central point error
212 besterr = vfp->vf(y, d->pre_stride, z, b->src_stride, sse1);
213 *distortion = besterr;
214 besterr += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
216 // TODO: Each subsequent iteration checks at least one point in common with the last iteration could be 2 ( if diag selected)
220 CHECK_BETTER(left, tr, tc - 2);
221 CHECK_BETTER(right, tr, tc + 2);
222 CHECK_BETTER(up, tr - 2, tc);
223 CHECK_BETTER(down, tr + 2, tc);
225 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
230 CHECK_BETTER(diag, tr - 2, tc - 2);
233 CHECK_BETTER(diag, tr - 2, tc + 2);
236 CHECK_BETTER(diag, tr + 2, tc - 2);
239 CHECK_BETTER(diag, tr + 2, tc + 2);
243 // no reason to check the same one again.
244 if (tr == br && tc == bc)
251 // TODO: Each subsequent iteration checks at least one point in common with the last iteration could be 2 ( if diag selected)
253 while (--quarteriters)
255 CHECK_BETTER(left, tr, tc - 1);
256 CHECK_BETTER(right, tr, tc + 1);
257 CHECK_BETTER(up, tr - 1, tc);
258 CHECK_BETTER(down, tr + 1, tc);
260 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
265 CHECK_BETTER(diag, tr - 1, tc - 1);
268 CHECK_BETTER(diag, tr - 1, tc + 1);
271 CHECK_BETTER(diag, tr + 1, tc - 1);
274 CHECK_BETTER(diag, tr + 1, tc + 1);
278 // no reason to check the same one again.
279 if (tr == br && tc == bc)
286 bestmv->as_mv.row = br << 1;
287 bestmv->as_mv.col = bc << 1;
289 if ((abs(bestmv->as_mv.col - ref_mv->as_mv.col) > MAX_FULL_PEL_VAL) ||
290 (abs(bestmv->as_mv.row - ref_mv->as_mv.row) > MAX_FULL_PEL_VAL))
303 int vp8_find_best_sub_pixel_step(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
304 int_mv *bestmv, int_mv *ref_mv,
306 const vp8_variance_fn_ptr_t *vfp,
307 int *mvcost[2], int *distortion,
310 int bestmse = INT_MAX;
313 unsigned char *y = *(d->base_pre) + d->pre + (bestmv->as_mv.row) * d->pre_stride + bestmv->as_mv.col;
314 unsigned char *z = (*(b->base_src) + b->src);
315 int left, right, up, down, diag;
321 // Trap uncodable vectors
322 if ((abs((bestmv->as_mv.col << 3) - ref_mv->as_mv.col) > MAX_FULL_PEL_VAL)
323 || (abs((bestmv->as_mv.row << 3) - ref_mv->as_mv.row) > MAX_FULL_PEL_VAL))
325 bestmv->as_mv.row <<= 3;
326 bestmv->as_mv.col <<= 3;
327 *distortion = INT_MAX;
332 bestmv->as_mv.row <<= 3;
333 bestmv->as_mv.col <<= 3;
336 // calculate central point error
337 bestmse = vfp->vf(y, d->pre_stride, z, b->src_stride, sse1);
338 *distortion = bestmse;
339 bestmse += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
341 // go left then right and check error
342 this_mv.as_mv.row = startmv.as_mv.row;
343 this_mv.as_mv.col = ((startmv.as_mv.col - 8) | 4);
344 thismse = vfp->svf_halfpix_h(y - 1, d->pre_stride, z, b->src_stride, &sse);
345 left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
351 *distortion = thismse;
355 this_mv.as_mv.col += 8;
356 thismse = vfp->svf_halfpix_h(y, d->pre_stride, z, b->src_stride, &sse);
357 right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
363 *distortion = thismse;
367 // go up then down and check error
368 this_mv.as_mv.col = startmv.as_mv.col;
369 this_mv.as_mv.row = ((startmv.as_mv.row - 8) | 4);
370 thismse = vfp->svf_halfpix_v(y - d->pre_stride, d->pre_stride, z, b->src_stride, &sse);
371 up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
377 *distortion = thismse;
381 this_mv.as_mv.row += 8;
382 thismse = vfp->svf_halfpix_v(y, d->pre_stride, z, b->src_stride, &sse);
383 down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
389 *distortion = thismse;
394 // now check 1 more diagonal
395 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
396 //for(whichdir =0;whichdir<4;whichdir++)
403 this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
404 this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
405 thismse = vfp->svf_halfpix_hv(y - 1 - d->pre_stride, d->pre_stride, z, b->src_stride, &sse);
408 this_mv.as_mv.col += 4;
409 this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
410 thismse = vfp->svf_halfpix_hv(y - d->pre_stride, d->pre_stride, z, b->src_stride, &sse);
413 this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
414 this_mv.as_mv.row += 4;
415 thismse = vfp->svf_halfpix_hv(y - 1, d->pre_stride, z, b->src_stride, &sse);
419 this_mv.as_mv.col += 4;
420 this_mv.as_mv.row += 4;
421 thismse = vfp->svf_halfpix_hv(y, d->pre_stride, z, b->src_stride, &sse);
425 diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
431 *distortion = thismse;
438 // time to check quarter pels.
439 if (bestmv->as_mv.row < startmv.as_mv.row)
442 if (bestmv->as_mv.col < startmv.as_mv.col)
449 // go left then right and check error
450 this_mv.as_mv.row = startmv.as_mv.row;
452 if (startmv.as_mv.col & 7)
454 this_mv.as_mv.col = startmv.as_mv.col - 2;
455 thismse = vfp->svf(y, d->pre_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
459 this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
460 thismse = vfp->svf(y - 1, d->pre_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
463 left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
469 *distortion = thismse;
473 this_mv.as_mv.col += 4;
474 thismse = vfp->svf(y, d->pre_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
475 right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
481 *distortion = thismse;
485 // go up then down and check error
486 this_mv.as_mv.col = startmv.as_mv.col;
488 if (startmv.as_mv.row & 7)
490 this_mv.as_mv.row = startmv.as_mv.row - 2;
491 thismse = vfp->svf(y, d->pre_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
495 this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
496 thismse = vfp->svf(y - d->pre_stride, d->pre_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
499 up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
505 *distortion = thismse;
509 this_mv.as_mv.row += 4;
510 thismse = vfp->svf(y, d->pre_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
511 down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
517 *distortion = thismse;
522 // now check 1 more diagonal
523 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
525 // for(whichdir=0;whichdir<4;whichdir++)
533 if (startmv.as_mv.row & 7)
535 this_mv.as_mv.row -= 2;
537 if (startmv.as_mv.col & 7)
539 this_mv.as_mv.col -= 2;
540 thismse = vfp->svf(y, d->pre_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
544 this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
545 thismse = vfp->svf(y - 1, d->pre_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);;
550 this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
552 if (startmv.as_mv.col & 7)
554 this_mv.as_mv.col -= 2;
555 thismse = vfp->svf(y - d->pre_stride, d->pre_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
559 this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
560 thismse = vfp->svf(y - d->pre_stride - 1, d->pre_stride, 6, 6, z, b->src_stride, &sse);
566 this_mv.as_mv.col += 2;
568 if (startmv.as_mv.row & 7)
570 this_mv.as_mv.row -= 2;
571 thismse = vfp->svf(y, d->pre_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
575 this_mv.as_mv.row = (startmv.as_mv.row - 8) | 6;
576 thismse = vfp->svf(y - d->pre_stride, d->pre_stride, this_mv.as_mv.col & 7, 6, z, b->src_stride, &sse);
581 this_mv.as_mv.row += 2;
583 if (startmv.as_mv.col & 7)
585 this_mv.as_mv.col -= 2;
586 thismse = vfp->svf(y, d->pre_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
590 this_mv.as_mv.col = (startmv.as_mv.col - 8) | 6;
591 thismse = vfp->svf(y - 1, d->pre_stride, 6, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
596 this_mv.as_mv.col += 2;
597 this_mv.as_mv.row += 2;
598 thismse = vfp->svf(y, d->pre_stride, this_mv.as_mv.col & 7, this_mv.as_mv.row & 7, z, b->src_stride, &sse);
602 diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
608 *distortion = thismse;
615 int vp8_find_best_half_pixel_step(MACROBLOCK *mb, BLOCK *b, BLOCKD *d,
616 int_mv *bestmv, int_mv *ref_mv,
618 const vp8_variance_fn_ptr_t *vfp,
619 int *mvcost[2], int *distortion,
622 int bestmse = INT_MAX;
625 unsigned char *y = *(d->base_pre) + d->pre + (bestmv->as_mv.row) * d->pre_stride + bestmv->as_mv.col;
626 unsigned char *z = (*(b->base_src) + b->src);
627 int left, right, up, down, diag;
631 // Trap uncodable vectors
632 if ((abs((bestmv->as_mv.col << 3) - ref_mv->as_mv.col) > MAX_FULL_PEL_VAL)
633 || (abs((bestmv->as_mv.row << 3) - ref_mv->as_mv.row) > MAX_FULL_PEL_VAL))
635 bestmv->as_mv.row <<= 3;
636 bestmv->as_mv.col <<= 3;
637 *distortion = INT_MAX;
642 bestmv->as_mv.row <<= 3;
643 bestmv->as_mv.col <<= 3;
646 // calculate central point error
647 bestmse = vfp->vf(y, d->pre_stride, z, b->src_stride, sse1);
648 *distortion = bestmse;
649 bestmse += mv_err_cost(bestmv, ref_mv, mvcost, error_per_bit);
651 // go left then right and check error
652 this_mv.as_mv.row = startmv.as_mv.row;
653 this_mv.as_mv.col = ((startmv.as_mv.col - 8) | 4);
654 thismse = vfp->svf_halfpix_h(y - 1, d->pre_stride, z, b->src_stride, &sse);
655 left = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
661 *distortion = thismse;
665 this_mv.as_mv.col += 8;
666 thismse = vfp->svf_halfpix_h(y, d->pre_stride, z, b->src_stride, &sse);
667 right = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
673 *distortion = thismse;
677 // go up then down and check error
678 this_mv.as_mv.col = startmv.as_mv.col;
679 this_mv.as_mv.row = ((startmv.as_mv.row - 8) | 4);
680 thismse = vfp->svf_halfpix_v(y - d->pre_stride, d->pre_stride, z, b->src_stride, &sse);
681 up = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
687 *distortion = thismse;
691 this_mv.as_mv.row += 8;
692 thismse = vfp->svf_halfpix_v(y, d->pre_stride, z, b->src_stride, &sse);
693 down = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
699 *distortion = thismse;
703 // somewhat strangely not doing all the diagonals for half pel is slower than doing them.
705 // now check 1 more diagonal -
706 whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);
712 this_mv.col = (this_mv.col - 8) | 4;
713 this_mv.row = (this_mv.row - 8) | 4;
714 diag = vfp->svf(y - 1 - d->pre_stride, d->pre_stride, 4, 4, z, b->src_stride, &sse);
718 this_mv.row = (this_mv.row - 8) | 4;
719 diag = vfp->svf(y - d->pre_stride, d->pre_stride, 4, 4, z, b->src_stride, &sse);
722 this_mv.col = (this_mv.col - 8) | 4;
724 diag = vfp->svf(y - 1, d->pre_stride, 4, 4, z, b->src_stride, &sse);
729 diag = vfp->svf(y, d->pre_stride, 4, 4, z, b->src_stride, &sse);
733 diag += mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
742 this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
743 this_mv.as_mv.row = (this_mv.as_mv.row - 8) | 4;
744 thismse = vfp->svf_halfpix_hv(y - 1 - d->pre_stride, d->pre_stride, z, b->src_stride, &sse);
745 diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
751 *distortion = thismse;
755 this_mv.as_mv.col += 8;
756 thismse = vfp->svf_halfpix_hv(y - d->pre_stride, d->pre_stride, z, b->src_stride, &sse);
757 diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
763 *distortion = thismse;
767 this_mv.as_mv.col = (this_mv.as_mv.col - 8) | 4;
768 this_mv.as_mv.row = startmv.as_mv.row + 4;
769 thismse = vfp->svf_halfpix_hv(y - 1, d->pre_stride, z, b->src_stride, &sse);
770 diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
776 *distortion = thismse;
780 this_mv.as_mv.col += 8;
781 thismse = vfp->svf_halfpix_hv(y, d->pre_stride, z, b->src_stride, &sse);
782 diag = thismse + mv_err_cost(&this_mv, ref_mv, mvcost, error_per_bit);
788 *distortion = thismse;
797 #define MVC(r,c) (((mvsadcost[0][r-rr] + mvsadcost[1][c-rc]) * error_per_bit + 128 )>>8 ) // estimated cost of a motion vector (r,c)
798 #define PRE(r,c) (*(d->base_pre) + d->pre + (r) * d->pre_stride + (c)) // pointer to predictor base of a motionvector
799 #define DIST(r,c,v) vfp->sdf( src,src_stride,PRE(r,c),d->pre_stride, v) // returns sad error score.
800 #define ERR(r,c,v) (MVC(r,c)+DIST(r,c,v)) // returns distortion + motion vector cost
801 #define CHECK_BETTER(v,r,c) if ((v = ERR(r,c,besterr)) < besterr) { besterr = v; br=r; bc=c; } // checks if (r,c) has better score than previous best
802 static const MV next_chkpts[6][3] =
804 {{ -2, 0}, { -1, -2}, {1, -2}},
805 {{ -1, -2}, {1, -2}, {2, 0}},
806 {{1, -2}, {2, 0}, {1, 2}},
807 {{2, 0}, {1, 2}, { -1, 2}},
808 {{1, 2}, { -1, 2}, { -2, 0}},
809 {{ -1, 2}, { -2, 0}, { -1, -2}}
821 const vp8_variance_fn_ptr_t *vfp,
827 MV hex[6] = { { -1, -2}, {1, -2}, {2, 0}, {1, 2}, { -1, 2}, { -2, 0} } ;
828 //MV neighbors[8] = { { -1, -1}, {0, -1}, {1, -1}, { -1, 0}, {1, 0}, { -1, 1}, {0, 1}, {1, 1} } ;
829 MV neighbors[4] = {{0, -1}, { -1, 0}, {1, 0}, {0, 1}} ;
832 unsigned char *src = (*(b->base_src) + b->src);
833 int src_stride = b->src_stride;
834 int rr = center_mv->as_mv.row, rc = center_mv->as_mv.col;
835 int br = ref_mv->as_mv.row >> 3, bc = ref_mv->as_mv.col >> 3, tr, tc;
836 unsigned int besterr, thiserr = 0x7fffffff;
839 if (bc < x->mv_col_min) bc = x->mv_col_min;
841 if (bc > x->mv_col_max) bc = x->mv_col_max;
843 if (br < x->mv_row_min) br = x->mv_row_min;
845 if (br > x->mv_row_max) br = x->mv_row_max;
850 besterr = ERR(br, bc, thiserr);
857 for (i = 0; i < 6; i++)
859 int nr = tr + hex[i].row, nc = tc + hex[i].col;
861 if (nc < x->mv_col_min) continue;
863 if (nc > x->mv_col_max) continue;
865 if (nr < x->mv_row_min) continue;
867 if (nr > x->mv_row_max) continue;
869 //CHECK_BETTER(thiserr,nr,nc);
870 if ((thiserr = ERR(nr, nc, besterr)) < besterr)
879 if (tr == br && tc == bc)
882 for (j = 1; j < 127; j++)
888 for (i = 0; i < 3; i++)
890 int nr = tr + next_chkpts[tk][i].row, nc = tc + next_chkpts[tk][i].col;
892 if (nc < x->mv_col_min) continue;
894 if (nc > x->mv_col_max) continue;
896 if (nr < x->mv_row_min) continue;
898 if (nr > x->mv_row_max) continue;
900 //CHECK_BETTER(thiserr,nr,nc);
901 if ((thiserr = ERR(nr, nc, besterr)) < besterr)
905 bc = nc; //k=(tk+5+i)%6;}
908 if (k >= 12) k -= 12;
909 else if (k >= 6) k -= 6;
913 if (tr == br && tc == bc)
917 // check 4 1-away neighbors
920 for (j = 0; j < 32; j++)
925 for (i = 0; i < 4; i++)
927 int nr = tr + neighbors[i].row, nc = tc + neighbors[i].col;
929 if (nc < x->mv_col_min) continue;
931 if (nc > x->mv_col_max) continue;
933 if (nr < x->mv_row_min) continue;
935 if (nr > x->mv_row_max) continue;
937 CHECK_BETTER(thiserr, nr, nc);
940 if (tr == br && tc == bc)
944 best_mv->as_mv.row = br;
945 best_mv->as_mv.col = bc;
947 return vfp->vf(src, src_stride, PRE(br, bc), d->pre_stride, &thiserr) + mv_err_cost(best_mv, center_mv, mvcost, error_per_bit) ;
957 int vp8_diamond_search_sad
967 vp8_variance_fn_ptr_t *fn_ptr,
974 unsigned char *what = (*(b->base_src) + b->src);
975 int what_stride = b->src_stride;
976 unsigned char *in_what;
977 int in_what_stride = d->pre_stride;
978 unsigned char *best_address;
983 int bestsad = INT_MAX;
987 int ref_row = ref_mv->as_mv.row >> 3;
988 int ref_col = ref_mv->as_mv.col >> 3;
993 unsigned char *check_here;
996 int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
998 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
999 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1003 best_mv->as_mv.row = ref_row;
1004 best_mv->as_mv.col = ref_col;
1006 // Work out the start point for the search
1007 in_what = (unsigned char *)(*(d->base_pre) + d->pre + (ref_row * (d->pre_stride)) + ref_col);
1008 best_address = in_what;
1010 // We need to check that the starting point for the search (as indicated by ref_mv) is within the buffer limits
1011 if ((ref_col > x->mv_col_min) && (ref_col < x->mv_col_max) &&
1012 (ref_row > x->mv_row_min) && (ref_row < x->mv_row_max))
1014 // Check the starting position
1015 bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, error_per_bit);
1018 // search_param determines the length of the initial step and hence the number of iterations
1019 // 0 = initial step (MAX_FIRST_STEP) pel : 1 = (MAX_FIRST_STEP/2) pel, 2 = (MAX_FIRST_STEP/4) pel... etc.
1020 ss = &x->ss[search_param * x->searches_per_step];
1021 tot_steps = (x->ss_count / x->searches_per_step) - search_param;
1025 for (step = 0; step < tot_steps ; step++)
1027 for (j = 0 ; j < x->searches_per_step ; j++)
1029 // Trap illegal vectors
1030 this_row_offset = best_mv->as_mv.row + ss[i].mv.row;
1031 this_col_offset = best_mv->as_mv.col + ss[i].mv.col;
1033 if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
1034 (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
1037 check_here = ss[i].offset + best_address;
1038 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1040 if (thissad < bestsad)
1042 this_mv.as_mv.row = this_row_offset;
1043 this_mv.as_mv.col = this_col_offset;
1044 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1046 if (thissad < bestsad)
1057 if (best_site != last_site)
1059 best_mv->as_mv.row += ss[best_site].mv.row;
1060 best_mv->as_mv.col += ss[best_site].mv.col;
1061 best_address += ss[best_site].offset;
1062 last_site = best_site;
1064 else if (best_address == in_what)
1068 this_mv.as_mv.row = best_mv->as_mv.row << 3;
1069 this_mv.as_mv.col = best_mv->as_mv.col << 3;
1071 if (bestsad == INT_MAX)
1074 return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
1075 + mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
1078 int vp8_diamond_search_sadx4
1088 vp8_variance_fn_ptr_t *fn_ptr,
1095 unsigned char *what = (*(b->base_src) + b->src);
1096 int what_stride = b->src_stride;
1097 unsigned char *in_what;
1098 int in_what_stride = d->pre_stride;
1099 unsigned char *best_address;
1104 int bestsad = INT_MAX;
1108 int ref_row = ref_mv->as_mv.row >> 3;
1109 int ref_col = ref_mv->as_mv.col >> 3;
1110 int this_row_offset;
1111 int this_col_offset;
1114 unsigned char *check_here;
1115 unsigned int thissad;
1117 int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
1119 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1120 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1123 best_mv->as_mv.row = ref_row;
1124 best_mv->as_mv.col = ref_col;
1126 // Work out the start point for the search
1127 in_what = (unsigned char *)(*(d->base_pre) + d->pre + (ref_row * (d->pre_stride)) + ref_col);
1128 best_address = in_what;
1130 // We need to check that the starting point for the search (as indicated by ref_mv) is within the buffer limits
1131 if ((ref_col > x->mv_col_min) && (ref_col < x->mv_col_max) &&
1132 (ref_row > x->mv_row_min) && (ref_row < x->mv_row_max))
1134 // Check the starting position
1135 bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride, 0x7fffffff) + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, error_per_bit);
1138 // search_param determines the length of the initial step and hence the number of iterations
1139 // 0 = initial step (MAX_FIRST_STEP) pel : 1 = (MAX_FIRST_STEP/2) pel, 2 = (MAX_FIRST_STEP/4) pel... etc.
1140 ss = &x->ss[search_param * x->searches_per_step];
1141 tot_steps = (x->ss_count / x->searches_per_step) - search_param;
1145 for (step = 0; step < tot_steps ; step++)
1149 // To know if all neighbor points are within the bounds, 4 bounds checking are enough instead of
1150 // checking 4 bounds for each points.
1151 all_in &= ((best_mv->as_mv.row + ss[i].mv.row)> x->mv_row_min);
1152 all_in &= ((best_mv->as_mv.row + ss[i+1].mv.row) < x->mv_row_max);
1153 all_in &= ((best_mv->as_mv.col + ss[i+2].mv.col) > x->mv_col_min);
1154 all_in &= ((best_mv->as_mv.col + ss[i+3].mv.col) < x->mv_col_max);
1158 unsigned int sad_array[4];
1160 for (j = 0 ; j < x->searches_per_step ; j += 4)
1162 unsigned char *block_offset[4];
1164 for (t = 0; t < 4; t++)
1165 block_offset[t] = ss[i+t].offset + best_address;
1167 fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, sad_array);
1169 for (t = 0; t < 4; t++, i++)
1171 if (sad_array[t] < bestsad)
1173 this_mv.as_mv.row = best_mv->as_mv.row + ss[i].mv.row;
1174 this_mv.as_mv.col = best_mv->as_mv.col + ss[i].mv.col;
1175 sad_array[t] += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1177 if (sad_array[t] < bestsad)
1179 bestsad = sad_array[t];
1188 for (j = 0 ; j < x->searches_per_step ; j++)
1190 // Trap illegal vectors
1191 this_row_offset = best_mv->as_mv.row + ss[i].mv.row;
1192 this_col_offset = best_mv->as_mv.col + ss[i].mv.col;
1194 if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
1195 (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
1197 check_here = ss[i].offset + best_address;
1198 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1200 if (thissad < bestsad)
1202 this_mv.as_mv.row = this_row_offset;
1203 this_mv.as_mv.col = this_col_offset;
1204 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1206 if (thissad < bestsad)
1217 if (best_site != last_site)
1219 best_mv->as_mv.row += ss[best_site].mv.row;
1220 best_mv->as_mv.col += ss[best_site].mv.col;
1221 best_address += ss[best_site].offset;
1222 last_site = best_site;
1224 else if (best_address == in_what)
1228 this_mv.as_mv.row = best_mv->as_mv.row << 3;
1229 this_mv.as_mv.col = best_mv->as_mv.col << 3;
1231 if (bestsad == INT_MAX)
1234 return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
1235 + mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
1238 int vp8_full_search_sad(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
1239 int error_per_bit, int distance,
1240 vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
1243 unsigned char *what = (*(b->base_src) + b->src);
1244 int what_stride = b->src_stride;
1245 unsigned char *in_what;
1246 int in_what_stride = d->pre_stride;
1247 int mv_stride = d->pre_stride;
1248 unsigned char *bestaddress;
1249 int_mv *best_mv = &d->bmi.mv;
1251 int bestsad = INT_MAX;
1254 unsigned char *check_here;
1257 int ref_row = ref_mv->as_mv.row;
1258 int ref_col = ref_mv->as_mv.col;
1260 int row_min = ref_row - distance;
1261 int row_max = ref_row + distance;
1262 int col_min = ref_col - distance;
1263 int col_max = ref_col + distance;
1265 int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
1267 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1268 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1270 // Work out the mid point for the search
1271 in_what = *(d->base_pre) + d->pre;
1272 bestaddress = in_what + (ref_row * d->pre_stride) + ref_col;
1274 best_mv->as_mv.row = ref_row;
1275 best_mv->as_mv.col = ref_col;
1277 // We need to check that the starting point for the search (as indicated by ref_mv) is within the buffer limits
1278 if ((ref_col > x->mv_col_min) && (ref_col < x->mv_col_max) &&
1279 (ref_row > x->mv_row_min) && (ref_row < x->mv_row_max))
1281 // Baseline value at the centre
1283 //bestsad = fn_ptr->sf( what,what_stride,bestaddress,in_what_stride) + (int)sqrt(mv_err_cost(ref_mv,ref_mv, mvcost,error_per_bit*14));
1284 bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, error_per_bit);
1287 // Apply further limits to prevent us looking using vectors that stretch beyiond the UMV border
1288 if (col_min < x->mv_col_min)
1289 col_min = x->mv_col_min;
1291 if (col_max > x->mv_col_max)
1292 col_max = x->mv_col_max;
1294 if (row_min < x->mv_row_min)
1295 row_min = x->mv_row_min;
1297 if (row_max > x->mv_row_max)
1298 row_max = x->mv_row_max;
1300 for (r = row_min; r < row_max ; r++)
1302 this_mv.as_mv.row = r;
1303 check_here = r * mv_stride + in_what + col_min;
1305 for (c = col_min; c < col_max; c++)
1307 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1309 this_mv.as_mv.col = c;
1310 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1312 if (thissad < bestsad)
1315 best_mv->as_mv.row = r;
1316 best_mv->as_mv.col = c;
1317 bestaddress = check_here;
1324 this_mv.as_mv.row = best_mv->as_mv.row << 3;
1325 this_mv.as_mv.col = best_mv->as_mv.col << 3;
1327 if (bestsad < INT_MAX)
1328 return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, (unsigned int *)(&thissad))
1329 + mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
1334 int vp8_full_search_sadx3(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
1335 int error_per_bit, int distance,
1336 vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
1339 unsigned char *what = (*(b->base_src) + b->src);
1340 int what_stride = b->src_stride;
1341 unsigned char *in_what;
1342 int in_what_stride = d->pre_stride;
1343 int mv_stride = d->pre_stride;
1344 unsigned char *bestaddress;
1345 int_mv *best_mv = &d->bmi.mv;
1347 int bestsad = INT_MAX;
1350 unsigned char *check_here;
1351 unsigned int thissad;
1353 int ref_row = ref_mv->as_mv.row;
1354 int ref_col = ref_mv->as_mv.col;
1356 int row_min = ref_row - distance;
1357 int row_max = ref_row + distance;
1358 int col_min = ref_col - distance;
1359 int col_max = ref_col + distance;
1361 unsigned int sad_array[3];
1363 int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
1365 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1366 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1368 // Work out the mid point for the search
1369 in_what = *(d->base_pre) + d->pre;
1370 bestaddress = in_what + (ref_row * d->pre_stride) + ref_col;
1372 best_mv->as_mv.row = ref_row;
1373 best_mv->as_mv.col = ref_col;
1375 // We need to check that the starting point for the search (as indicated by ref_mv) is within the buffer limits
1376 if ((ref_col > x->mv_col_min) && (ref_col < x->mv_col_max) &&
1377 (ref_row > x->mv_row_min) && (ref_row < x->mv_row_max))
1379 // Baseline value at the centre
1380 bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, error_per_bit);
1383 // Apply further limits to prevent us looking using vectors that stretch beyiond the UMV border
1384 if (col_min < x->mv_col_min)
1385 col_min = x->mv_col_min;
1387 if (col_max > x->mv_col_max)
1388 col_max = x->mv_col_max;
1390 if (row_min < x->mv_row_min)
1391 row_min = x->mv_row_min;
1393 if (row_max > x->mv_row_max)
1394 row_max = x->mv_row_max;
1396 for (r = row_min; r < row_max ; r++)
1398 this_mv.as_mv.row = r;
1399 check_here = r * mv_stride + in_what + col_min;
1402 while ((c + 2) < col_max)
1406 fn_ptr->sdx3f(what, what_stride, check_here , in_what_stride, sad_array);
1408 for (i = 0; i < 3; i++)
1410 thissad = sad_array[i];
1412 if (thissad < bestsad)
1414 this_mv.as_mv.col = c;
1415 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1417 if (thissad < bestsad)
1420 best_mv->as_mv.row = r;
1421 best_mv->as_mv.col = c;
1422 bestaddress = check_here;
1433 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1435 if (thissad < bestsad)
1437 this_mv.as_mv.col = c;
1438 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1440 if (thissad < bestsad)
1443 best_mv->as_mv.row = r;
1444 best_mv->as_mv.col = c;
1445 bestaddress = check_here;
1455 this_mv.as_mv.row = best_mv->as_mv.row << 3;
1456 this_mv.as_mv.col = best_mv->as_mv.col << 3;
1458 if (bestsad < INT_MAX)
1459 return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, (unsigned int *)(&thissad))
1460 + mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
1465 int vp8_full_search_sadx8(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
1466 int error_per_bit, int distance,
1467 vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
1470 unsigned char *what = (*(b->base_src) + b->src);
1471 int what_stride = b->src_stride;
1472 unsigned char *in_what;
1473 int in_what_stride = d->pre_stride;
1474 int mv_stride = d->pre_stride;
1475 unsigned char *bestaddress;
1476 int_mv *best_mv = &d->bmi.mv;
1478 int bestsad = INT_MAX;
1481 unsigned char *check_here;
1482 unsigned int thissad;
1484 int ref_row = ref_mv->as_mv.row;
1485 int ref_col = ref_mv->as_mv.col;
1487 int row_min = ref_row - distance;
1488 int row_max = ref_row + distance;
1489 int col_min = ref_col - distance;
1490 int col_max = ref_col + distance;
1492 DECLARE_ALIGNED_ARRAY(16, unsigned short, sad_array8, 8);
1493 unsigned int sad_array[3];
1495 int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
1497 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1498 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1500 // Work out the mid point for the search
1501 in_what = *(d->base_pre) + d->pre;
1502 bestaddress = in_what + (ref_row * d->pre_stride) + ref_col;
1504 best_mv->as_mv.row = ref_row;
1505 best_mv->as_mv.col = ref_col;
1507 // We need to check that the starting point for the search (as indicated by ref_mv) is within the buffer limits
1508 if ((ref_col > x->mv_col_min) && (ref_col < x->mv_col_max) &&
1509 (ref_row > x->mv_row_min) && (ref_row < x->mv_row_max))
1511 // Baseline value at the centre
1512 bestsad = fn_ptr->sdf(what, what_stride, bestaddress, in_what_stride, 0x7fffffff) + mvsad_err_cost(best_mv, &fcenter_mv, mvsadcost, error_per_bit);
1515 // Apply further limits to prevent us looking using vectors that stretch beyiond the UMV border
1516 if (col_min < x->mv_col_min)
1517 col_min = x->mv_col_min;
1519 if (col_max > x->mv_col_max)
1520 col_max = x->mv_col_max;
1522 if (row_min < x->mv_row_min)
1523 row_min = x->mv_row_min;
1525 if (row_max > x->mv_row_max)
1526 row_max = x->mv_row_max;
1528 for (r = row_min; r < row_max ; r++)
1530 this_mv.as_mv.row = r;
1531 check_here = r * mv_stride + in_what + col_min;
1534 while ((c + 7) < col_max)
1538 fn_ptr->sdx8f(what, what_stride, check_here , in_what_stride, sad_array8);
1540 for (i = 0; i < 8; i++)
1542 thissad = (unsigned int)sad_array8[i];
1544 if (thissad < bestsad)
1546 this_mv.as_mv.col = c;
1547 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1549 if (thissad < bestsad)
1552 best_mv->as_mv.row = r;
1553 best_mv->as_mv.col = c;
1554 bestaddress = check_here;
1563 while ((c + 2) < col_max)
1567 fn_ptr->sdx3f(what, what_stride, check_here , in_what_stride, sad_array);
1569 for (i = 0; i < 3; i++)
1571 thissad = sad_array[i];
1573 if (thissad < bestsad)
1575 this_mv.as_mv.col = c;
1576 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1578 if (thissad < bestsad)
1581 best_mv->as_mv.row = r;
1582 best_mv->as_mv.col = c;
1583 bestaddress = check_here;
1594 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1596 if (thissad < bestsad)
1598 this_mv.as_mv.col = c;
1599 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1601 if (thissad < bestsad)
1604 best_mv->as_mv.row = r;
1605 best_mv->as_mv.col = c;
1606 bestaddress = check_here;
1615 this_mv.as_mv.row = best_mv->as_mv.row << 3;
1616 this_mv.as_mv.col = best_mv->as_mv.col << 3;
1618 if (bestsad < INT_MAX)
1619 return fn_ptr->vf(what, what_stride, bestaddress, in_what_stride, (unsigned int *)(&thissad))
1620 + mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
1625 int vp8_refining_search_sad(MACROBLOCK *x, BLOCK *b, BLOCKD *d, int_mv *ref_mv,
1626 int error_per_bit, int search_range,
1627 vp8_variance_fn_ptr_t *fn_ptr, int *mvcost[2],
1630 MV neighbors[4] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0}};
1632 short this_row_offset, this_col_offset;
1634 int what_stride = b->src_stride;
1635 int in_what_stride = d->pre_stride;
1636 unsigned char *what = (*(b->base_src) + b->src);
1637 unsigned char *best_address = (unsigned char *)(*(d->base_pre) + d->pre +
1638 (ref_mv->as_mv.row * (d->pre_stride)) + ref_mv->as_mv.col);
1639 unsigned char *check_here;
1640 unsigned int thissad;
1642 unsigned int bestsad = INT_MAX;
1644 int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
1647 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1648 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1650 bestsad = fn_ptr->sdf(what, what_stride, best_address, in_what_stride, 0x7fffffff) + mvsad_err_cost(ref_mv, &fcenter_mv, mvsadcost, error_per_bit);
1652 for (i=0; i<search_range; i++)
1656 for (j = 0 ; j < 4 ; j++)
1658 this_row_offset = ref_mv->as_mv.row + neighbors[j].row;
1659 this_col_offset = ref_mv->as_mv.col + neighbors[j].col;
1661 if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
1662 (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
1664 check_here = (neighbors[j].row)*in_what_stride + neighbors[j].col + best_address;
1665 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1667 if (thissad < bestsad)
1669 this_mv.as_mv.row = this_row_offset;
1670 this_mv.as_mv.col = this_col_offset;
1671 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1673 if (thissad < bestsad)
1682 if (best_site == -1)
1686 ref_mv->as_mv.row += neighbors[best_site].row;
1687 ref_mv->as_mv.col += neighbors[best_site].col;
1688 best_address += (neighbors[best_site].row)*in_what_stride + neighbors[best_site].col;
1692 this_mv.as_mv.row = ref_mv->as_mv.row << 3;
1693 this_mv.as_mv.col = ref_mv->as_mv.col << 3;
1695 if (bestsad < INT_MAX)
1696 return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
1697 + mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
1702 int vp8_refining_search_sadx4(MACROBLOCK *x, BLOCK *b, BLOCKD *d,
1703 int_mv *ref_mv, int error_per_bit,
1704 int search_range, vp8_variance_fn_ptr_t *fn_ptr,
1705 int *mvcost[2], int_mv *center_mv)
1707 MV neighbors[4] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0}};
1709 short this_row_offset, this_col_offset;
1711 int what_stride = b->src_stride;
1712 int in_what_stride = d->pre_stride;
1713 unsigned char *what = (*(b->base_src) + b->src);
1714 unsigned char *best_address = (unsigned char *)(*(d->base_pre) + d->pre +
1715 (ref_mv->as_mv.row * (d->pre_stride)) + ref_mv->as_mv.col);
1716 unsigned char *check_here;
1717 unsigned int thissad;
1719 unsigned int bestsad = INT_MAX;
1721 int *mvsadcost[2] = {x->mvsadcost[0], x->mvsadcost[1]};
1724 fcenter_mv.as_mv.row = center_mv->as_mv.row >> 3;
1725 fcenter_mv.as_mv.col = center_mv->as_mv.col >> 3;
1727 bestsad = fn_ptr->sdf(what, what_stride, best_address, in_what_stride, 0x7fffffff) + mvsad_err_cost(ref_mv, &fcenter_mv, mvsadcost, error_per_bit);
1729 for (i=0; i<search_range; i++)
1734 all_in &= ((ref_mv->as_mv.row - 1) > x->mv_row_min);
1735 all_in &= ((ref_mv->as_mv.row + 1) < x->mv_row_max);
1736 all_in &= ((ref_mv->as_mv.col - 1) > x->mv_col_min);
1737 all_in &= ((ref_mv->as_mv.col + 1) < x->mv_col_max);
1741 unsigned int sad_array[4];
1742 unsigned char *block_offset[4];
1743 block_offset[0] = best_address - in_what_stride;
1744 block_offset[1] = best_address - 1;
1745 block_offset[2] = best_address + 1;
1746 block_offset[3] = best_address + in_what_stride;
1748 fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride, sad_array);
1750 for (j = 0; j < 4; j++)
1752 if (sad_array[j] < bestsad)
1754 this_mv.as_mv.row = ref_mv->as_mv.row + neighbors[j].row;
1755 this_mv.as_mv.col = ref_mv->as_mv.col + neighbors[j].col;
1756 sad_array[j] += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1758 if (sad_array[j] < bestsad)
1760 bestsad = sad_array[j];
1768 for (j = 0 ; j < 4 ; j++)
1770 this_row_offset = ref_mv->as_mv.row + neighbors[j].row;
1771 this_col_offset = ref_mv->as_mv.col + neighbors[j].col;
1773 if ((this_col_offset > x->mv_col_min) && (this_col_offset < x->mv_col_max) &&
1774 (this_row_offset > x->mv_row_min) && (this_row_offset < x->mv_row_max))
1776 check_here = (neighbors[j].row)*in_what_stride + neighbors[j].col + best_address;
1777 thissad = fn_ptr->sdf(what, what_stride, check_here , in_what_stride, bestsad);
1779 if (thissad < bestsad)
1781 this_mv.as_mv.row = this_row_offset;
1782 this_mv.as_mv.col = this_col_offset;
1783 thissad += mvsad_err_cost(&this_mv, &fcenter_mv, mvsadcost, error_per_bit);
1785 if (thissad < bestsad)
1795 if (best_site == -1)
1799 ref_mv->as_mv.row += neighbors[best_site].row;
1800 ref_mv->as_mv.col += neighbors[best_site].col;
1801 best_address += (neighbors[best_site].row)*in_what_stride + neighbors[best_site].col;
1805 this_mv.as_mv.row = ref_mv->as_mv.row << 3;
1806 this_mv.as_mv.col = ref_mv->as_mv.col << 3;
1808 if (bestsad < INT_MAX)
1809 return fn_ptr->vf(what, what_stride, best_address, in_what_stride, (unsigned int *)(&thissad))
1810 + mv_err_cost(&this_mv, center_mv, mvcost, error_per_bit);
1815 #ifdef ENTROPY_STATS
1816 void print_mode_context(void)
1818 FILE *f = fopen("modecont.c", "w");
1821 fprintf(f, "#include \"entropy.h\"\n");
1822 fprintf(f, "const int vp8_mode_contexts[6][4] =\n");
1825 for (j = 0; j < 6; j++)
1827 fprintf(f, " { // %d \n", j);
1830 for (i = 0; i < 4; i++)
1834 int count; // = mv_ref_ct[j][i][0]+mv_ref_ct[j][i][1];
1837 count = mv_mode_cts[i][0] + mv_mode_cts[i][1];
1840 overal_prob = 256 * mv_mode_cts[i][0] / count;
1844 if (overal_prob == 0)
1848 count = mv_ref_ct[j][i][0] + mv_ref_ct[j][i][1];
1851 this_prob = 256 * mv_ref_ct[j][i][0] / count;
1858 fprintf(f, "%5d, ", this_prob);
1859 //fprintf(f,"%5d, %5d, %8d,", this_prob, overal_prob, (this_prob << 10)/overal_prob);
1860 //fprintf(f,"%8d, ", (this_prob << 10)/overal_prob);
1863 fprintf(f, " },\n");
1870 /* MV ref count ENTROPY_STATS stats code */
1871 #ifdef ENTROPY_STATS
1872 void init_mv_ref_counts()
1874 vpx_memset(mv_ref_ct, 0, sizeof(mv_ref_ct));
1875 vpx_memset(mv_mode_cts, 0, sizeof(mv_mode_cts));
1878 void accum_mv_refs(MB_PREDICTION_MODE m, const int ct[4])
1882 ++mv_ref_ct [ct[0]] [0] [0];
1883 ++mv_mode_cts[0][0];
1887 ++mv_ref_ct [ct[0]] [0] [1];
1888 ++mv_mode_cts[0][1];
1892 ++mv_ref_ct [ct[1]] [1] [0];
1893 ++mv_mode_cts[1][0];
1897 ++mv_ref_ct [ct[1]] [1] [1];
1898 ++mv_mode_cts[1][1];
1902 ++mv_ref_ct [ct[2]] [2] [0];
1903 ++mv_mode_cts[2][0];
1907 ++mv_ref_ct [ct[2]] [2] [1];
1908 ++mv_mode_cts[2][1];
1912 ++mv_ref_ct [ct[3]] [3] [0];
1913 ++mv_mode_cts[3][0];
1917 ++mv_ref_ct [ct[3]] [3] [1];
1918 ++mv_mode_cts[3][1];
1925 #endif/* END MV ref count ENTROPY_STATS stats code */