3 |
// 01.05.2002 updated MBMotionCompensationBVOP |
// 01.05.2002 updated MBMotionCompensationBVOP |
4 |
// 14.04.2002 bframe compensation |
// 14.04.2002 bframe compensation |
5 |
|
|
6 |
|
#include <stdio.h> |
7 |
|
|
8 |
#include "../encoder.h" |
#include "../encoder.h" |
9 |
#include "../utils/mbfunctions.h" |
#include "../utils/mbfunctions.h" |
10 |
#include "../image/interpolate8x8.h" |
#include "../image/interpolate8x8.h" |
12 |
#include "../utils/timer.h" |
#include "../utils/timer.h" |
13 |
#include "motion.h" |
#include "motion.h" |
14 |
|
|
15 |
|
#ifndef ABS |
16 |
|
#define ABS(X) (((X)>0)?(X):-(X)) |
17 |
|
#endif |
18 |
|
#ifndef SIGN |
19 |
|
#define SIGN(X) (((X)>0)?1:-1) |
20 |
|
#endif |
21 |
|
|
22 |
|
#ifndef RSHIFT |
23 |
|
#define RSHIFT(a,b) ((a) > 0 ? ((a) + (1<<((b)-1)))>>(b) : ((a) + (1<<((b)-1))-1)>>(b)) |
24 |
|
#endif |
25 |
|
|
26 |
|
/* assume b>0 */ |
27 |
|
#ifndef ROUNDED_DIV |
28 |
|
#define ROUNDED_DIV(a,b) (((a)>0 ? (a) + ((b)>>1) : (a) - ((b)>>1))/(b)) |
29 |
|
#endif |
30 |
|
|
31 |
|
|
32 |
|
/* This is borrowed from decoder.c */ |
33 |
|
static __inline int gmc_sanitize(int value, int quarterpel, int fcode) |
34 |
|
{ |
35 |
|
int length = 1 << (fcode+4); |
36 |
|
|
37 |
|
// if (quarterpel) value *= 2; |
38 |
|
|
39 |
|
if (value < -length) |
40 |
|
return -length; |
41 |
|
else if (value >= length) |
42 |
|
return length-1; |
43 |
|
else return value; |
44 |
|
} |
45 |
|
|
46 |
|
/* And this is borrowed from bitstream.c until we find a common solution */ |
47 |
|
|
48 |
|
static uint32_t __inline |
49 |
|
log2bin(uint32_t value) |
50 |
|
{ |
51 |
|
/* Changed by Chenm001 */ |
52 |
|
#if !defined(_MSC_VER) |
53 |
|
int n = 0; |
54 |
|
|
55 |
|
while (value) { |
56 |
|
value >>= 1; |
57 |
|
n++; |
58 |
|
} |
59 |
|
return n; |
60 |
|
#else |
61 |
|
__asm { |
62 |
|
bsr eax, value |
63 |
|
inc eax |
64 |
|
} |
65 |
|
#endif |
66 |
|
} |
67 |
|
|
68 |
|
|
69 |
static __inline void |
static __inline void |
70 |
compensate16x16_interpolate(int16_t * const dct_codes, |
compensate16x16_interpolate(int16_t * const dct_codes, |
71 |
uint8_t * const cur, |
uint8_t * const cur, |
78 |
uint32_t y, |
uint32_t y, |
79 |
const int32_t dx, |
const int32_t dx, |
80 |
const int32_t dy, |
const int32_t dy, |
81 |
const uint32_t stride, |
const int32_t stride, |
82 |
const int quarterpel, |
const int quarterpel, |
83 |
const int reduced_resolution, |
const int reduced_resolution, |
84 |
const uint32_t rounding) |
const int32_t rounding) |
85 |
{ |
{ |
86 |
const uint8_t * ptr; |
const uint8_t * ptr; |
87 |
|
|
88 |
if (!reduced_resolution) { |
if (!reduced_resolution) { |
89 |
|
|
90 |
if(quarterpel) { |
if(quarterpel) { |
91 |
if (dx&3 | dy&3) { |
if ((dx&3) | (dy&3)) { |
92 |
interpolate16x16_quarterpel(tmp - y * stride - x, |
interpolate16x16_quarterpel(tmp - y * stride - x, |
93 |
(uint8_t *) ref, tmp + 32, |
(uint8_t *) ref, tmp + 32, |
94 |
tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); |
tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); |
140 |
uint32_t y, |
uint32_t y, |
141 |
const int32_t dx, |
const int32_t dx, |
142 |
const int32_t dy, |
const int32_t dy, |
143 |
const uint32_t stride, |
const int32_t stride, |
144 |
const uint32_t quarterpel, |
const int32_t quarterpel, |
145 |
const int reduced_resolution, |
const int reduced_resolution, |
146 |
const uint32_t rounding) |
const int32_t rounding) |
147 |
{ |
{ |
148 |
const uint8_t * ptr; |
const uint8_t * ptr; |
149 |
|
|
150 |
if (!reduced_resolution) { |
if (!reduced_resolution) { |
151 |
|
|
152 |
if(quarterpel) { |
if(quarterpel) { |
153 |
if (dx&3 | dy&3) { |
if ((dx&3) | (dy&3)) { |
154 |
interpolate8x8_quarterpel(tmp - y*stride - x, |
interpolate8x8_quarterpel(tmp - y*stride - x, |
155 |
(uint8_t *) ref, tmp + 32, |
(uint8_t *) ref, tmp + 32, |
156 |
tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); |
tmp + 64, tmp + 96, x, y, dx, dy, stride, rounding); |
181 |
const uint32_t y, |
const uint32_t y, |
182 |
const int32_t dx, |
const int32_t dx, |
183 |
const int dy, |
const int dy, |
184 |
const uint32_t stride, |
const int32_t stride, |
185 |
const uint32_t rounding) |
const int32_t rounding) |
186 |
{ |
{ |
187 |
interpolate8x8_switch(cur, refn, x-1, y-1, dx, dy, stride, rounding); |
interpolate8x8_switch(cur, refn, x-1, y-1, dx, dy, stride, rounding); |
188 |
interpolate8x8_switch(cur, refn, x+7, y-1, dx, dy, stride, rounding); |
interpolate8x8_switch(cur, refn, x+7, y-1, dx, dy, stride, rounding); |
204 |
const IMAGE * const Ref, |
const IMAGE * const Ref, |
205 |
uint8_t * const temp, |
uint8_t * const temp, |
206 |
int16_t * const coeff, |
int16_t * const coeff, |
207 |
const uint32_t stride, |
const int32_t stride, |
208 |
const int rounding, |
const int rounding, |
209 |
const int rrv) |
const int rrv) |
210 |
{ /* uv-block-based compensation */ |
{ /* uv-block-based compensation */ |
214 |
interpolate8x8_switch2(temp, Ref->u, 8 * i, 8 * j, |
interpolate8x8_switch2(temp, Ref->u, 8 * i, 8 * j, |
215 |
dx, dy, stride, rounding), |
dx, dy, stride, rounding), |
216 |
stride); |
stride); |
|
|
|
217 |
transfer_8to16sub(coeff + 64, Cur->v + 8 * j * stride + 8 * i, |
transfer_8to16sub(coeff + 64, Cur->v + 8 * j * stride + 8 * i, |
218 |
interpolate8x8_switch2(temp, Ref->v, 8 * i, 8 * j, |
interpolate8x8_switch2(temp, Ref->v, 8 * i, 8 * j, |
219 |
dx, dy, stride, rounding), |
dx, dy, stride, rounding), |
244 |
const IMAGE * const refh, |
const IMAGE * const refh, |
245 |
const IMAGE * const refv, |
const IMAGE * const refv, |
246 |
const IMAGE * const refhv, |
const IMAGE * const refhv, |
247 |
|
const IMAGE * const refGMC, |
248 |
IMAGE * const cur, |
IMAGE * const cur, |
249 |
int16_t * dct_codes, |
int16_t * dct_codes, |
250 |
const uint32_t width, |
const uint32_t width, |
251 |
const uint32_t height, |
const uint32_t height, |
252 |
const uint32_t edged_width, |
const uint32_t edged_width, |
253 |
const int quarterpel, |
const int32_t quarterpel, |
254 |
const int reduced_resolution, |
const int reduced_resolution, |
255 |
const uint32_t rounding) |
const int32_t rounding) |
256 |
{ |
{ |
257 |
int32_t dx = (quarterpel ? mb->qmvs[0].x : mb->mvs[0].x); |
int32_t dx; |
258 |
int32_t dy = (quarterpel ? mb->qmvs[0].y : mb->mvs[0].y); |
int32_t dy; |
259 |
|
|
260 |
|
|
261 |
uint8_t * const tmp = refv->u; |
uint8_t * const tmp = refv->u; |
262 |
|
|
263 |
if ( mb->mode == MODE_NOT_CODED && dx==0 && dy==0 && !reduced_resolution) { /* quick copy */ |
if ( (!reduced_resolution) && (mb->mode == MODE_NOT_CODED) ) { /* quick copy for early SKIP */ |
264 |
|
/* early SKIP is only activated in P-VOPs, not in S-VOPs, so mcsel can never be 1 */ |
265 |
|
|
266 |
|
/* if (mb->mcsel) { |
267 |
|
transfer16x16_copy(cur->y + 16 * (i + j * edged_width), |
268 |
|
refGMC->y + 16 * (i + j * edged_width), |
269 |
|
edged_width); |
270 |
|
transfer8x8_copy(cur->u + 8 * (i + j * edged_width/2), |
271 |
|
refGMC->u + 8 * (i + j * edged_width/2), |
272 |
|
edged_width / 2); |
273 |
|
transfer8x8_copy(cur->v + 8 * (i + j * edged_width/2), |
274 |
|
refGMC->v + 8 * (i + j * edged_width/2), |
275 |
|
edged_width / 2); |
276 |
|
} else |
277 |
|
*/ |
278 |
|
{ |
279 |
transfer16x16_copy(cur->y + 16 * (i + j * edged_width), |
transfer16x16_copy(cur->y + 16 * (i + j * edged_width), |
280 |
ref->y + 16 * (i + j * edged_width), |
ref->y + 16 * (i + j * edged_width), |
281 |
edged_width); |
edged_width); |
286 |
transfer8x8_copy(cur->v + 8 * (i + j * edged_width/2), |
transfer8x8_copy(cur->v + 8 * (i + j * edged_width/2), |
287 |
ref->v + 8 * (i + j * edged_width/2), |
ref->v + 8 * (i + j * edged_width/2), |
288 |
edged_width / 2); |
edged_width / 2); |
289 |
|
} |
290 |
return; |
return; |
291 |
} |
} |
292 |
|
|
293 |
if ((mb->mode == MODE_NOT_CODED || mb->mode == MODE_INTER || mb->mode == MODE_INTER_Q) /*&& !quarterpel*/) { |
if ((mb->mode == MODE_NOT_CODED || mb->mode == MODE_INTER |
294 |
|
|| mb->mode == MODE_INTER_Q) /*&& !quarterpel*/) { |
295 |
|
|
296 |
|
/* reduced resolution + GMC: not possible */ |
297 |
|
|
298 |
/* quick MODE_NOT_CODED for GMC with MV!=(0,0) is still needed */ |
if (mb->mcsel) { |
299 |
|
|
300 |
|
/* call normal routine once, easier than "if (mcsel)"ing all the time */ |
301 |
|
|
302 |
|
transfer_8to16sub(&dct_codes[0*64], cur->y + 16*j*edged_width + 16*i, |
303 |
|
refGMC->y + 16*j*edged_width + 16*i, edged_width); |
304 |
|
transfer_8to16sub(&dct_codes[1*64], cur->y + 16*j*edged_width + 16*i+8, |
305 |
|
refGMC->y + 16*j*edged_width + 16*i+8, edged_width); |
306 |
|
transfer_8to16sub(&dct_codes[2*64], cur->y + (16*j+8)*edged_width + 16*i, |
307 |
|
refGMC->y + (16*j+8)*edged_width + 16*i, edged_width); |
308 |
|
transfer_8to16sub(&dct_codes[3*64], cur->y + (16*j+8)*edged_width + 16*i+8, |
309 |
|
refGMC->y + (16*j+8)*edged_width + 16*i+8, edged_width); |
310 |
|
|
311 |
|
/* lumi is needed earlier for mode decision, but chroma should be done block-based, but it isn't, yet. */ |
312 |
|
|
313 |
|
transfer_8to16sub(&dct_codes[4 * 64], cur->u + 8 *j*edged_width/2 + 8*i, |
314 |
|
refGMC->u + 8 *j*edged_width/2 + 8*i, edged_width/2); |
315 |
|
|
316 |
|
transfer_8to16sub(&dct_codes[5 * 64], cur->v + 8*j* edged_width/2 + 8*i, |
317 |
|
refGMC->v + 8*j* edged_width/2 + 8*i, edged_width/2); |
318 |
|
|
319 |
|
return; |
320 |
|
} |
321 |
|
|
322 |
|
/* ordinary compensation */ |
323 |
|
|
324 |
|
dx = (quarterpel ? mb->qmvs[0].x : mb->mvs[0].x); |
325 |
|
dy = (quarterpel ? mb->qmvs[0].y : mb->mvs[0].y); |
326 |
|
|
327 |
if (reduced_resolution) { |
if (reduced_resolution) { |
328 |
dx = RRV_MV_SCALEUP(dx); |
dx = RRV_MV_SCALEUP(dx); |
333 |
refv->y, refhv->y, tmp, 16 * i, 16 * j, dx, dy, |
refv->y, refhv->y, tmp, 16 * i, 16 * j, dx, dy, |
334 |
edged_width, quarterpel, reduced_resolution, rounding); |
edged_width, quarterpel, reduced_resolution, rounding); |
335 |
|
|
336 |
dx /= 1 + quarterpel; |
dx /= (int)(1 + quarterpel); |
337 |
dy /= 1 + quarterpel; |
dy /= (int)(1 + quarterpel); |
338 |
|
|
339 |
dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
dx = (dx >> 1) + roundtab_79[dx & 0x3]; |
340 |
dy = (dy >> 1) + roundtab_79[dy & 0x3]; |
dy = (dy >> 1) + roundtab_79[dy & 0x3]; |
341 |
|
|
364 |
|
|
365 |
CompensateChroma(dx, dy, i, j, cur, ref, tmp, |
CompensateChroma(dx, dy, i, j, cur, ref, tmp, |
366 |
&dct_codes[4 * 64], edged_width / 2, rounding, reduced_resolution); |
&dct_codes[4 * 64], edged_width / 2, rounding, reduced_resolution); |
|
|
|
367 |
} |
} |
368 |
|
|
369 |
|
|
432 |
|
|
433 |
if (quarterpel) { |
if (quarterpel) { |
434 |
|
|
435 |
if (dx&3 | dy&3) { |
if ((dx&3) | (dy&3)) { |
436 |
interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width, |
interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width, |
437 |
(uint8_t *) f_ref->y, tmp + 32, |
(uint8_t *) f_ref->y, tmp + 32, |
438 |
tmp + 64, tmp + 96, 16*i, 16*j, dx, dy, edged_width, 0); |
tmp + 64, tmp + 96, 16*i, 16*j, dx, dy, edged_width, 0); |
439 |
ptr1 = tmp; |
ptr1 = tmp; |
440 |
} else ptr1 = f_ref->y + (16*j + dy/4)*edged_width + 16*i + dx/4; // fullpixel position |
} else ptr1 = f_ref->y + (16*j + dy/4)*edged_width + 16*i + dx/4; // fullpixel position |
441 |
|
|
442 |
if (b_dx&3 | b_dy&3) { |
if ((b_dx&3) | (b_dy&3)) { |
443 |
interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width + 16, |
interpolate16x16_quarterpel(tmp - i * 16 - j * 16 * edged_width + 16, |
444 |
(uint8_t *) b_ref->y, tmp + 32, |
(uint8_t *) b_ref->y, tmp + 32, |
445 |
tmp + 64, tmp + 96, 16*i, 16*j, b_dx, b_dy, edged_width, 0); |
tmp + 64, tmp + 96, 16*i, 16*j, b_dx, b_dy, edged_width, 0); |
485 |
sumx += dx/2; sumy += dy/2; |
sumx += dx/2; sumy += dy/2; |
486 |
b_sumx += b_dx/2; b_sumy += b_dy/2; |
b_sumx += b_dx/2; b_sumy += b_dy/2; |
487 |
|
|
488 |
if (dx&3 | dy&3) { |
if ((dx&3) | (dy&3)) { |
489 |
interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width, |
interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width, |
490 |
(uint8_t *) f_ref->y, |
(uint8_t *) f_ref->y, |
491 |
tmp + 32, tmp + 64, tmp + 96, |
tmp + 32, tmp + 64, tmp + 96, |
493 |
ptr1 = tmp; |
ptr1 = tmp; |
494 |
} else ptr1 = f_ref->y + (16*j + (k>>1)*8 + dy/4)*edged_width + 16*i + (k&1)*8 + dx/4; |
} else ptr1 = f_ref->y + (16*j + (k>>1)*8 + dy/4)*edged_width + 16*i + (k&1)*8 + dx/4; |
495 |
|
|
496 |
if (b_dx&3 | b_dy&3) { |
if ((b_dx&3) | (b_dy&3)) { |
497 |
interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width + 16, |
interpolate8x8_quarterpel(tmp - (i * 16+(k&1)*8) - (j * 16+((k>>1)*8)) * edged_width + 16, |
498 |
(uint8_t *) b_ref->y, |
(uint8_t *) b_ref->y, |
499 |
tmp + 16, tmp + 32, tmp + 48, |
tmp + 16, tmp + 32, tmp + 48, |
540 |
dx, dy, edged_width / 2, 0), |
dx, dy, edged_width / 2, 0), |
541 |
edged_width / 2); |
edged_width / 2); |
542 |
} |
} |
543 |
|
|
544 |
|
|
545 |
|
|
546 |
|
void |
547 |
|
generate_GMCparameters( const int num_wp, // [input]: number of warppoints |
548 |
|
const int res, // [input]: resolution |
549 |
|
const WARPPOINTS *const warp, // [input]: warp points |
550 |
|
const int width, const int height, |
551 |
|
GMC_DATA *const gmc) // [output] precalculated parameters |
552 |
|
{ |
553 |
|
|
554 |
|
/* We follow mainly two sources: The original standard, which is ugly, and the |
555 |
|
thesis from Andreas Dehnhardt, which is much nicer. |
556 |
|
|
557 |
|
Notation is: indices are written next to the variable, |
558 |
|
primes in the standard are denoted by a suffix 'p'. |
559 |
|
types are "c"=constant, "i"=input parameter, "f"=calculated, then fixed, |
560 |
|
"o"=output data, " "=other, "u" = unused, "p"=calc for every pixel |
561 |
|
|
562 |
|
type | variable name | ISO name (TeX-style) | value or range | usage |
563 |
|
------------------------------------------------------------------------------------- |
564 |
|
c | H | H | [16 , ?] | image width (w/o edges) |
565 |
|
c | W | W | [16 , ?] | image height (w/o edges) |
566 |
|
|
567 |
|
c | i0 | i_0 | 0 | ref. point #1, X |
568 |
|
c | j0 | j_0 | 0 | ref. point #1, Y |
569 |
|
c | i1 | i_1 | W | ref. point #2, X |
570 |
|
c | j1 | j_1 | 0 | ref. point #2, Y |
571 |
|
cu | i2 | i_2 | 0 | ref. point #3, X |
572 |
|
cu | i2 | j_2 | H | ref. point #3, Y |
573 |
|
|
574 |
|
i | du0 | du[0] | [-16863,16863] | warp vector #1, Y |
575 |
|
i | dv0 | dv[0] | [-16863,16863] | warp vector #1, Y |
576 |
|
i | du1 | du[1] | [-16863,16863] | warp vector #2, Y |
577 |
|
i | dv1 | dv[1] | [-16863,16863] | warp vector #2, Y |
578 |
|
iu | du2 | du[2] | [-16863,16863] | warp vector #3, Y |
579 |
|
iu | dv2 | dv[2] | [-16863,16863] | warp vector #3, Y |
580 |
|
|
581 |
|
i | s | s | {2,4,8,16} | interpol. resolution |
582 |
|
f | sigma | - | log2(s) | X / s == X >> sigma |
583 |
|
f | r | r | =16/s | complementary res. |
584 |
|
f | rho | \rho | log2(r) | X / r == X >> rho |
585 |
|
|
586 |
|
f | i0s | i'_0 | | |
587 |
|
f | j0s | j'_0 | | |
588 |
|
f | i1s | i'_1 | | |
589 |
|
f | j1s | j'_1 | | |
590 |
|
f | i2s | i'_2 | | |
591 |
|
f | j2s | j'_2 | | |
592 |
|
|
593 |
|
f | alpha | \alpha | | 2^{alpha-1} < W <= 2^alpha |
594 |
|
f | beta | \beta | | 2^{beta-1} < H <= 2^beta |
595 |
|
|
596 |
|
f | Ws | W' | W = 2^{alpha} | scaled width |
597 |
|
f | Hs | H' | W = 2^{beta} | scaled height |
598 |
|
|
599 |
|
f | i1ss | i''_1 | "virtual sprite stuff" |
600 |
|
f | j1ss | j''_1 | "virtual sprite stuff" |
601 |
|
f | i2ss | i''_2 | "virtual sprite stuff" |
602 |
|
f | j2ss | j''_2 | "virtual sprite stuff" |
603 |
|
*/ |
604 |
|
|
605 |
|
/* Some calculations are disabled because we only use 2 warppoints at the moment */ |
606 |
|
|
607 |
|
int du0 = warp->duv[0].x; |
608 |
|
int dv0 = warp->duv[0].y; |
609 |
|
int du1 = warp->duv[1].x; |
610 |
|
int dv1 = warp->duv[1].y; |
611 |
|
// int du2 = warp->duv[2].x; |
612 |
|
// int dv2 = warp->duv[2].y; |
613 |
|
|
614 |
|
gmc->num_wp = num_wp; |
615 |
|
|
616 |
|
gmc->s = res; /* scaling parameters 2,4,8 or 16 */ |
617 |
|
gmc->sigma = log2bin(res-1); /* log2bin(15)=4, log2bin(16)=5, log2bin(17)=5 */ |
618 |
|
gmc->r = 16/res; |
619 |
|
gmc->rho = 4 - gmc->sigma; /* = log2bin(r-1) */ |
620 |
|
|
621 |
|
gmc->W = width; |
622 |
|
gmc->H = height; /* fixed reference coordinates */ |
623 |
|
|
624 |
|
gmc->alpha = log2bin(gmc->W-1); |
625 |
|
gmc->Ws= 1<<gmc->alpha; |
626 |
|
|
627 |
|
// gmc->beta = log2bin(gmc->H-1); |
628 |
|
// gmc->Hs= 1<<gmc->beta; |
629 |
|
|
630 |
|
// printf("du0=%d dv0=%d du1=%d dv1=%d s=%d sigma=%d W=%d alpha=%d, Ws=%d, rho=%d\n",du0,dv0,du1,dv1,gmc->s,gmc->sigma,gmc->W,gmc->alpha,gmc->Ws,gmc->rho); |
631 |
|
|
632 |
|
/* i2s is only needed for num_wp >= 3, etc. */ |
633 |
|
/* the 's' values are in 1/s pel resolution */ |
634 |
|
gmc->i0s = res/2 * ( du0 ); |
635 |
|
gmc->j0s = res/2 * ( dv0 ); |
636 |
|
gmc->i1s = res/2 * (2*width + du1 + du0 ); |
637 |
|
gmc->j1s = res/2 * ( dv1 + dv0 ); |
638 |
|
// gmc->i2s = res/2 * ( du2 + du0 ); |
639 |
|
// gmc->j2s = res/2 * (2*height + dv2 + dv0 ); |
640 |
|
|
641 |
|
/* i2s and i2ss are only needed for num_wp == 3, etc. */ |
642 |
|
|
643 |
|
/* the 'ss' values are in 1/16 pel resolution */ |
644 |
|
gmc->i1ss = 16*gmc->Ws + ROUNDED_DIV(((gmc->W-gmc->Ws)*(gmc->r*gmc->i0s) + gmc->Ws*(gmc->r*gmc->i1s - 16*gmc->W)),gmc->W); |
645 |
|
gmc->j1ss = ROUNDED_DIV( ((gmc->W - gmc->Ws)*(gmc->r*gmc->j0s) + gmc->Ws*gmc->r*gmc->j1s) ,gmc->W ); |
646 |
|
|
647 |
|
// gmc->i2ss = ROUNDED_DIV( ((gmc->H - gmc->Hs)*(gmc->r*gmc->i0s) + gmc->Hs*(gmc->r*gmc->i2s)), gmc->H); |
648 |
|
// gmc->j2ss = 16*gmc->Hs + ROUNDED_DIV( ((gmc->H-gmc->Hs)*(gmc->r*gmc->j0s) + gmc->Ws*(gmc->r*gmc->j2s - 16*gmc->H)), gmc->H); |
649 |
|
|
650 |
|
return; |
651 |
|
} |
652 |
|
|
653 |
|
|
654 |
|
|
655 |
|
void |
656 |
|
generate_GMCimage( const GMC_DATA *const gmc_data, // [input] precalculated data |
657 |
|
const IMAGE *const pRef, // [input] |
658 |
|
const int mb_width, |
659 |
|
const int mb_height, |
660 |
|
const int stride, |
661 |
|
const int stride2, |
662 |
|
const int fcode, // [input] some parameters... |
663 |
|
const int32_t quarterpel, // [input] for rounding avgMV |
664 |
|
const int reduced_resolution, // [input] ignored |
665 |
|
const int32_t rounding, // [input] for rounding image data |
666 |
|
MACROBLOCK *const pMBs, // [output] average motion vectors |
667 |
|
IMAGE *const pGMC) // [output] full warped image |
668 |
|
{ |
669 |
|
|
670 |
|
unsigned int mj,mi; |
671 |
|
VECTOR avgMV; |
672 |
|
|
673 |
|
for (mj=0;mj<mb_height;mj++) |
674 |
|
for (mi=0;mi<mb_width; mi++) |
675 |
|
{ |
676 |
|
avgMV = generate_GMCimageMB(gmc_data, pRef, mi, mj, |
677 |
|
stride, stride2, quarterpel, rounding, pGMC); |
678 |
|
|
679 |
|
pMBs[mj*mb_width+mi].amv.x = gmc_sanitize(avgMV.x, quarterpel, fcode); |
680 |
|
pMBs[mj*mb_width+mi].amv.y = gmc_sanitize(avgMV.y, quarterpel, fcode); |
681 |
|
pMBs[mj*mb_width+mi].mcsel = 0; /* until mode decision */ |
682 |
|
} |
683 |
|
} |
684 |
|
|
685 |
|
|
686 |
|
VECTOR generate_GMCimageMB( const GMC_DATA *const gmc_data, /* [input] all precalc data */ |
687 |
|
const IMAGE *const pRef, /* [input] */ |
688 |
|
const int mi, const int mj, /* [input] MB position */ |
689 |
|
const int stride, /* [input] Lumi stride */ |
690 |
|
const int stride2, /* [input] chroma stride */ |
691 |
|
const int quarterpel, /* [input] for rounding of avgMV */ |
692 |
|
const int rounding, /* [input] for rounding of imgae data */ |
693 |
|
IMAGE *const pGMC) /* [outut] generate image */ |
694 |
|
|
695 |
|
/* |
696 |
|
type | variable name | ISO name (TeX-style) | value or range | usage |
697 |
|
------------------------------------------------------------------------------------- |
698 |
|
p | F | F(i,j) | | pelwise motion vector X in s-th pel |
699 |
|
p | G | G(i,j) | | pelwise motion vector Y in s-th pel |
700 |
|
p | Fc | F_c(i,j) | | |
701 |
|
p | Gc | G_c(i,j) | | same for chroma |
702 |
|
|
703 |
|
p | Y00 | Y_{00} | [0,255*s*s] | first: 4 neighbouring Y-values |
704 |
|
p | Y01 | Y_{01} | [0,255] | at fullpel position, around the |
705 |
|
p | Y10 | Y_{10} | [0,255*s] | position where pelweise MV points to |
706 |
|
p | Y11 | Y_{11} | [0,255] | later: bilinear interpol Y-values in Y00 |
707 |
|
|
708 |
|
p | C00 | C_{00} | [0,255*s*s] | same for chroma Cb and Cr |
709 |
|
p | C01 | C_{01} | [0,255] | |
710 |
|
p | C10 | C_{10} | [0,255*s] | |
711 |
|
p | C11 | C_{11} | [0,255] | |
712 |
|
|
713 |
|
*/ |
714 |
|
{ |
715 |
|
const int W = gmc_data->W; |
716 |
|
const int H = gmc_data->H; |
717 |
|
|
718 |
|
const int s = gmc_data->s; |
719 |
|
const int sigma = gmc_data->sigma; |
720 |
|
|
721 |
|
const int r = gmc_data->r; |
722 |
|
const int rho = gmc_data->rho; |
723 |
|
|
724 |
|
const int i0s = gmc_data->i0s; |
725 |
|
const int j0s = gmc_data->j0s; |
726 |
|
|
727 |
|
const int i1ss = gmc_data->i1ss; |
728 |
|
const int j1ss = gmc_data->j1ss; |
729 |
|
// const int i2ss = gmc_data->i2ss; |
730 |
|
// const int j2ss = gmc_data->j2ss; |
731 |
|
|
732 |
|
const int alpha = gmc_data->alpha; |
733 |
|
const int Ws = gmc_data->Ws; |
734 |
|
|
735 |
|
// const int beta = gmc_data->beta; |
736 |
|
// const int Hs = gmc_data->Hs; |
737 |
|
|
738 |
|
int I,J; |
739 |
|
VECTOR avgMV = {0,0}; |
740 |
|
|
741 |
|
for (J=16*mj;J<16*(mj+1);J++) |
742 |
|
for (I=16*mi;I<16*(mi+1);I++) |
743 |
|
{ |
744 |
|
int F= i0s + ( ((-r*i0s+i1ss)*I + (r*j0s-j1ss)*J + (1<<(alpha+rho-1))) >> (alpha+rho) ); |
745 |
|
int G= j0s + ( ((-r*j0s+j1ss)*I + (-r*i0s+i1ss)*J + (1<<(alpha+rho-1))) >> (alpha+rho) ); |
746 |
|
|
747 |
|
/* this naive implementation (with lots of multiplications) isn't slower (rather faster) than |
748 |
|
working incremental. Don't ask me why... maybe the whole this is memory bound? */ |
749 |
|
|
750 |
|
const int ri= F & (s-1); // fractional part of pelwise MV X |
751 |
|
const int rj= G & (s-1); // fractional part of pelwise MV Y |
752 |
|
|
753 |
|
int Y00,Y01,Y10,Y11; |
754 |
|
|
755 |
|
/* unclipped values are used for avgMV */ |
756 |
|
avgMV.x += F-(I<<sigma); /* shift position to 1/s-pel, as the MV is */ |
757 |
|
avgMV.y += G-(J<<sigma); /* TODO: don't do this (of course) */ |
758 |
|
|
759 |
|
F >>= sigma; |
760 |
|
G >>= sigma; |
761 |
|
|
762 |
|
/* clip values to be in range. Since we have edges, clip to 1 less than lower boundary |
763 |
|
this way positions F+1/G+1 are still right */ |
764 |
|
|
765 |
|
if (F< -1) |
766 |
|
F=-1; |
767 |
|
else if (F>W) |
768 |
|
F=W; /* W or W-1 doesn't matter, so save 1 subtract ;-) */ |
769 |
|
if (G< -1) |
770 |
|
G=-1; |
771 |
|
else if (G>H) |
772 |
|
G=H; /* dito */ |
773 |
|
|
774 |
|
Y00 = pRef->y[ G*stride + F ]; // Lumi values |
775 |
|
Y01 = pRef->y[ G*stride + F+1 ]; |
776 |
|
Y10 = pRef->y[ G*stride + F+stride ]; |
777 |
|
Y11 = pRef->y[ G*stride + F+stride+1 ]; |
778 |
|
|
779 |
|
/* bilinear interpolation */ |
780 |
|
Y00 = ((s-ri)*Y00 + ri*Y01); |
781 |
|
Y10 = ((s-ri)*Y10 + ri*Y11); |
782 |
|
Y00 = ((s-rj)*Y00 + rj*Y10 + s*s/2 - rounding ) >> (sigma+sigma); |
783 |
|
|
784 |
|
pGMC->y[J*stride+I] = (uint8_t)Y00; /* output 1 Y-pixel */ |
785 |
|
} |
786 |
|
|
787 |
|
|
788 |
|
/* doing chroma _here_ is even more stupid and slow, because won't be used until Compensation and |
789 |
|
most likely not even then (only if the block really _is_ GMC) |
790 |
|
*/ |
791 |
|
|
792 |
|
for (J=8*mj;J<8*(mj+1);J++) /* this plays the role of j_c,i_c in the standard */ |
793 |
|
for (I=8*mi;I<8*(mi+1);I++) /* For I_c we have to use I_c = 4*i_c+1 ! */ |
794 |
|
{ |
795 |
|
/* same positions for both chroma components, U=Cb and V=Cr */ |
796 |
|
int Fc=((-r*i0s+i1ss)*(4*I+1) + (r*j0s-j1ss)*(4*J+1) +2*Ws*r*i0s |
797 |
|
-16*Ws +(1<<(alpha+rho+1)))>>(alpha+rho+2); |
798 |
|
int Gc=((-r*j0s+j1ss)*(4*I+1) +(-r*i0s+i1ss)*(4*J+1) +2*Ws*r*j0s |
799 |
|
-16*Ws +(1<<(alpha+rho+1))) >>(alpha+rho+2); |
800 |
|
|
801 |
|
const int ri= Fc & (s-1); // fractional part of pelwise MV X |
802 |
|
const int rj= Gc & (s-1); // fractional part of pelwise MV Y |
803 |
|
|
804 |
|
int C00,C01,C10,C11; |
805 |
|
|
806 |
|
Fc >>= sigma; |
807 |
|
Gc >>= sigma; |
808 |
|
|
809 |
|
if (Fc< -1) |
810 |
|
Fc=-1; |
811 |
|
else if (Fc>=W/2) |
812 |
|
Fc=W/2; /* W or W-1 doesn't matter, so save 1 subtraction ;-) */ |
813 |
|
if (Gc< -1) |
814 |
|
Gc=-1; |
815 |
|
else if (Gc>=H/2) |
816 |
|
Gc=H/2; /* dito */ |
817 |
|
|
818 |
|
/* now calculate U data */ |
819 |
|
C00 = pRef->u[ Gc*stride2 + Fc ]; // chroma-value Cb |
820 |
|
C01 = pRef->u[ Gc*stride2 + Fc+1 ]; |
821 |
|
C10 = pRef->u[ (Gc+1)*stride2 + Fc ]; |
822 |
|
C11 = pRef->u[ (Gc+1)*stride2 + Fc+1 ]; |
823 |
|
|
824 |
|
/* bilinear interpolation */ |
825 |
|
C00 = ((s-ri)*C00 + ri*C01); |
826 |
|
C10 = ((s-ri)*C10 + ri*C11); |
827 |
|
C00 = ((s-rj)*C00 + rj*C10 + s*s/2 - rounding ) >> (sigma+sigma); |
828 |
|
|
829 |
|
pGMC->u[J*stride2+I] = (uint8_t)C00; /* output 1 U-pixel */ |
830 |
|
|
831 |
|
/* now calculate V data */ |
832 |
|
C00 = pRef->v[ Gc*stride2 + Fc ]; // chroma-value Cr |
833 |
|
C01 = pRef->v[ Gc*stride2 + Fc+1 ]; |
834 |
|
C10 = pRef->v[ (Gc+1)*stride2 + Fc ]; |
835 |
|
C11 = pRef->v[ (Gc+1)*stride2 + Fc+1 ]; |
836 |
|
|
837 |
|
/* bilinear interpolation */ |
838 |
|
C00 = ((s-ri)*C00 + ri*C01); |
839 |
|
C10 = ((s-ri)*C10 + ri*C11); |
840 |
|
C00 = ((s-rj)*C00 + rj*C10 + s*s/2 - rounding ) >> (sigma+sigma); |
841 |
|
|
842 |
|
pGMC->v[J*stride2+I] = (uint8_t)C00; /* output 1 V-pixel */ |
843 |
|
} |
844 |
|
|
845 |
|
/* The average vector is rounded from 1/s-pel to 1/2 or 1/4 using the '//' operator*/ |
846 |
|
|
847 |
|
avgMV.x = RSHIFT( avgMV.x, (sigma+7-quarterpel) ); |
848 |
|
avgMV.y = RSHIFT( avgMV.y, (sigma+7-quarterpel) ); |
849 |
|
|
850 |
|
/* ^^^^ this is the way MS Reference Software does it */ |
851 |
|
|
852 |
|
return avgMV; /* clipping to fcode area is done outside! */ |
853 |
|
} |
854 |
|
|