ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambcomp.c
Revision: 2.41
Committed: Wed Apr 30 23:38:58 2014 UTC (10 years ago) by greg
Content type: text/plain
Branch: MAIN
Changes since 2.40: +156 -39 lines
Log Message:
Added support for ambient super-samples (-as) in new calculation (-DNEWAMB)

File Contents

# Content
1 #ifndef lint
2 static const char RCSid[] = "$Id: ambcomp.c,v 2.40 2014/04/30 18:27:14 greg Exp $";
3 #endif
4 /*
5 * Routines to compute "ambient" values using Monte Carlo
6 *
7 * Hessian calculations based on "Practical Hessian-Based Error Control
8 * for Irradiance Caching" by Schwarzhaupt, Wann Jensen, & Jarosz
9 * from ACM SIGGRAPH Asia 2012 conference proceedings.
10 *
11 * Declarations of external symbols in ambient.h
12 */
13
14 #include "copyright.h"
15
16 #include "ray.h"
17 #include "ambient.h"
18 #include "random.h"
19
20 #ifdef NEWAMB
21
22 extern void SDsquare2disk(double ds[2], double seedx, double seedy);
23
24 typedef struct {
25 RAY *rp; /* originating ray sample */
26 FVECT ux, uy; /* tangent axis unit vectors */
27 int ns; /* number of samples per axis */
28 COLOR acoef; /* division contribution coefficient */
29 struct s_ambsamp {
30 COLOR v; /* hemisphere sample value */
31 FVECT p; /* intersection point */
32 } sa[1]; /* sample array (extends struct) */
33 } AMBHEMI; /* ambient sample hemisphere */
34
35 typedef struct s_ambsamp AMBSAMP;
36
37 #define ambsamp(h,i,j) (h)->sa[(i)*(h)->ns + (j)]
38
39 typedef struct {
40 FVECT r_i, r_i1, e_i, rcp, rI2_eJ2;
41 double I1, I2;
42 } FFTRI; /* vectors and coefficients for Hessian calculation */
43
44
45 static AMBHEMI *
46 inithemi( /* initialize sampling hemisphere */
47 COLOR ac,
48 RAY *r,
49 double wt
50 )
51 {
52 AMBHEMI *hp;
53 double d;
54 int n, i;
55 /* set number of divisions */
56 if (ambacc <= FTINY &&
57 wt > (d = 0.8*intens(ac)*r->rweight/(ambdiv*minweight)))
58 wt = d; /* avoid ray termination */
59 n = sqrt(ambdiv * wt) + 0.5;
60 i = 1 + 5*(ambacc > FTINY); /* minimum number of samples */
61 if (n < i)
62 n = i;
63 /* allocate sampling array */
64 hp = (AMBHEMI *)malloc(sizeof(AMBHEMI) + sizeof(AMBSAMP)*(n*n - 1));
65 if (hp == NULL)
66 return(NULL);
67 hp->rp = r;
68 hp->ns = n;
69 /* assign coefficient */
70 copycolor(hp->acoef, ac);
71 d = 1.0/(n*n);
72 scalecolor(hp->acoef, d);
73 /* make tangent plane axes */
74 hp->uy[0] = 0.5 - frandom();
75 hp->uy[1] = 0.5 - frandom();
76 hp->uy[2] = 0.5 - frandom();
77 for (i = 3; i--; )
78 if ((-0.6 < r->ron[i]) & (r->ron[i] < 0.6))
79 break;
80 if (i < 0)
81 error(CONSISTENCY, "bad ray direction in inithemi");
82 hp->uy[i] = 1.0;
83 VCROSS(hp->ux, hp->uy, r->ron);
84 normalize(hp->ux);
85 VCROSS(hp->uy, r->ron, hp->ux);
86 /* we're ready to sample */
87 return(hp);
88 }
89
90
91 /* Prepare ambient division sample */
92 static int
93 prepambsamp(RAY *arp, AMBHEMI *hp, int i, int j, int n)
94 {
95 int hlist[3], ii;
96 double spt[2], zd;
97 /* ambient coefficient for weight */
98 if (ambacc > FTINY)
99 setcolor(arp->rcoef, AVGREFL, AVGREFL, AVGREFL);
100 else
101 copycolor(arp->rcoef, hp->acoef);
102 if (rayorigin(arp, AMBIENT, hp->rp, arp->rcoef) < 0)
103 return(0);
104 if (ambacc > FTINY) {
105 multcolor(arp->rcoef, hp->acoef);
106 scalecolor(arp->rcoef, 1./AVGREFL);
107 }
108 hlist[0] = hp->rp->rno;
109 hlist[1] = i;
110 hlist[2] = j;
111 multisamp(spt, 2, urand(ilhash(hlist,3)+n));
112 if (!n) { /* avoid border samples for n==0 */
113 if ((spt[0] < 0.1) | (spt[0] > 0.9))
114 spt[0] = 0.1 + 0.8*frandom();
115 if ((spt[1] < 0.1) | (spt[1] > 0.9))
116 spt[1] = 0.1 + 0.8*frandom();
117 }
118 SDsquare2disk(spt, (i+spt[0])/hp->ns, (j+spt[1])/hp->ns);
119 zd = sqrt(1. - spt[0]*spt[0] - spt[1]*spt[1]);
120 for (ii = 3; ii--; )
121 arp->rdir[ii] = spt[0]*hp->ux[ii] +
122 spt[1]*hp->uy[ii] +
123 zd*hp->rp->ron[ii];
124 checknorm(arp->rdir);
125 return(1);
126 }
127
128
129 static AMBSAMP *
130 ambsample( /* sample an ambient direction */
131 AMBHEMI *hp,
132 int i,
133 int j
134 )
135 {
136 AMBSAMP *ap = &ambsamp(hp,i,j);
137 RAY ar;
138 /* generate hemispherical sample */
139 if (!prepambsamp(&ar, hp, i, j, 0))
140 goto badsample;
141 dimlist[ndims++] = i*hp->ns + j + 90171;
142 rayvalue(&ar); /* evaluate ray */
143 ndims--;
144 /* limit vertex distance */
145 if (ar.rt > 10.0*thescene.cusize)
146 ar.rt = 10.0*thescene.cusize;
147 else if (ar.rt <= FTINY) /* should never happen! */
148 goto badsample;
149 VSUM(ap->p, ar.rorg, ar.rdir, ar.rt);
150 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
151 copycolor(ap->v, ar.rcol);
152 return(ap);
153 badsample:
154 setcolor(ap->v, 0., 0., 0.);
155 VCOPY(ap->p, hp->rp->rop);
156 return(NULL);
157 }
158
159
160 /* Estimate errors based on ambient division differences */
161 static float *
162 getambdiffs(AMBHEMI *hp)
163 {
164 float *earr = calloc(hp->ns*hp->ns, sizeof(float));
165 float *ep;
166 double b, d2;
167 int i, j;
168
169 if (earr == NULL) /* out of memory? */
170 return(NULL);
171 /* compute squared neighbor diffs */
172 for (ep = earr, i = 0; i < hp->ns; i++)
173 for (j = 0; j < hp->ns; j++, ep++) {
174 b = bright(ambsamp(hp,i,j).v);
175 if (i) { /* from above */
176 d2 = b - bright(ambsamp(hp,i-1,j).v);
177 d2 *= d2;
178 ep[0] += d2;
179 ep[-hp->ns] += d2;
180 }
181 if (j) { /* from behind */
182 d2 = b - bright(ambsamp(hp,i,j-1).v);
183 d2 *= d2;
184 ep[0] += d2;
185 ep[-1] += d2;
186 }
187 }
188 /* correct for number of neighbors */
189 earr[0] *= 2.f;
190 earr[hp->ns-1] *= 2.f;
191 earr[(hp->ns-1)*hp->ns] *= 2.f;
192 earr[(hp->ns-1)*hp->ns + hp->ns-1] *= 2.f;
193 for (i = 1; i < hp->ns-1; i++) {
194 earr[i*hp->ns] *= 4./3.;
195 earr[i*hp->ns + hp->ns-1] *= 4./3.;
196 }
197 for (j = 1; j < hp->ns-1; j++) {
198 earr[j] *= 4./3.;
199 earr[(hp->ns-1)*hp->ns + j] *= 4./3.;
200 }
201 return(earr);
202 }
203
204
205 /* Perform super-sampling on hemisphere */
206 static void
207 ambsupersamp(double acol[3], AMBHEMI *hp, int cnt)
208 {
209 float *earr = getambdiffs(hp);
210 double e2sum = 0;
211 AMBSAMP *ap;
212 RAY ar;
213 COLOR asum;
214 float *ep;
215 int i, j, n;
216
217 if (earr == NULL) /* just skip calc. if no memory */
218 return;
219 /* add up estimated variances */
220 for (ep = earr + hp->ns*hp->ns; ep-- > earr; )
221 e2sum += *ep;
222 ep = earr; /* perform super-sampling */
223 for (ap = hp->sa, i = 0; i < hp->ns; i++)
224 for (j = 0; j < hp->ns; j++, ap++) {
225 int nss = *ep/e2sum*cnt + frandom();
226 setcolor(asum, 0., 0., 0.);
227 for (n = 1; n <= nss; n++) {
228 if (!prepambsamp(&ar, hp, i, j, n)) {
229 nss = n-1;
230 break;
231 }
232 dimlist[ndims++] = i*hp->ns + j + 90171;
233 rayvalue(&ar); /* evaluate super-sample */
234 ndims--;
235 multcolor(ar.rcol, ar.rcoef);
236 addcolor(asum, ar.rcol);
237 }
238 if (nss) { /* update returned ambient value */
239 const double ssf = 1./(nss + 1);
240 for (n = 3; n--; )
241 acol[n] += ssf*colval(asum,n) +
242 (ssf - 1.)*colval(ap->v,n);
243 }
244 e2sum -= *ep++; /* update remainders */
245 cnt -= nss;
246 }
247 free(earr);
248 }
249
250
251 /* Compute vectors and coefficients for Hessian/gradient calcs */
252 static void
253 comp_fftri(FFTRI *ftp, FVECT ap0, FVECT ap1, FVECT rop)
254 {
255 double rdot_cp, dot_e, dot_er, rdot_r, rdot_r1, J2;
256 int i;
257
258 VSUB(ftp->r_i, ap0, rop);
259 VSUB(ftp->r_i1, ap1, rop);
260 VSUB(ftp->e_i, ap1, ap0);
261 VCROSS(ftp->rcp, ftp->r_i, ftp->r_i1);
262 rdot_cp = 1.0/DOT(ftp->rcp,ftp->rcp);
263 dot_e = DOT(ftp->e_i,ftp->e_i);
264 dot_er = DOT(ftp->e_i, ftp->r_i);
265 rdot_r = 1.0/DOT(ftp->r_i,ftp->r_i);
266 rdot_r1 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
267 ftp->I1 = acos( DOT(ftp->r_i, ftp->r_i1) * sqrt(rdot_r*rdot_r1) ) *
268 sqrt( rdot_cp );
269 ftp->I2 = ( DOT(ftp->e_i, ftp->r_i1)*rdot_r1 - dot_er*rdot_r +
270 dot_e*ftp->I1 )*0.5*rdot_cp;
271 J2 = ( 0.5*(rdot_r - rdot_r1) - dot_er*ftp->I2 ) / dot_e;
272 for (i = 3; i--; )
273 ftp->rI2_eJ2[i] = ftp->I2*ftp->r_i[i] + J2*ftp->e_i[i];
274 }
275
276
277 /* Compose 3x3 matrix from two vectors */
278 static void
279 compose_matrix(FVECT mat[3], FVECT va, FVECT vb)
280 {
281 mat[0][0] = 2.0*va[0]*vb[0];
282 mat[1][1] = 2.0*va[1]*vb[1];
283 mat[2][2] = 2.0*va[2]*vb[2];
284 mat[0][1] = mat[1][0] = va[0]*vb[1] + va[1]*vb[0];
285 mat[0][2] = mat[2][0] = va[0]*vb[2] + va[2]*vb[0];
286 mat[1][2] = mat[2][1] = va[1]*vb[2] + va[2]*vb[1];
287 }
288
289
290 /* Compute partial 3x3 Hessian matrix for edge */
291 static void
292 comp_hessian(FVECT hess[3], FFTRI *ftp, FVECT nrm)
293 {
294 FVECT ncp;
295 FVECT m1[3], m2[3], m3[3], m4[3];
296 double d1, d2, d3, d4;
297 double I3, J3, K3;
298 int i, j;
299 /* compute intermediate coefficients */
300 d1 = 1.0/DOT(ftp->r_i,ftp->r_i);
301 d2 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
302 d3 = 1.0/DOT(ftp->e_i,ftp->e_i);
303 d4 = DOT(ftp->e_i, ftp->r_i);
304 I3 = ( DOT(ftp->e_i, ftp->r_i1)*d2*d2 - d4*d1*d1 + 3.0/d3*ftp->I2 )
305 / ( 4.0*DOT(ftp->rcp,ftp->rcp) );
306 J3 = 0.25*d3*(d1*d1 - d2*d2) - d4*d3*I3;
307 K3 = d3*(ftp->I2 - I3/d1 - 2.0*d4*J3);
308 /* intermediate matrices */
309 VCROSS(ncp, nrm, ftp->e_i);
310 compose_matrix(m1, ncp, ftp->rI2_eJ2);
311 compose_matrix(m2, ftp->r_i, ftp->r_i);
312 compose_matrix(m3, ftp->e_i, ftp->e_i);
313 compose_matrix(m4, ftp->r_i, ftp->e_i);
314 d1 = DOT(nrm, ftp->rcp);
315 d2 = -d1*ftp->I2;
316 d1 *= 2.0;
317 for (i = 3; i--; ) /* final matrix sum */
318 for (j = 3; j--; ) {
319 hess[i][j] = m1[i][j] + d1*( I3*m2[i][j] + K3*m3[i][j] +
320 2.0*J3*m4[i][j] );
321 hess[i][j] += d2*(i==j);
322 hess[i][j] *= 1.0/PI;
323 }
324 }
325
326
327 /* Reverse hessian calculation result for edge in other direction */
328 static void
329 rev_hessian(FVECT hess[3])
330 {
331 int i;
332
333 for (i = 3; i--; ) {
334 hess[i][0] = -hess[i][0];
335 hess[i][1] = -hess[i][1];
336 hess[i][2] = -hess[i][2];
337 }
338 }
339
340
341 /* Add to radiometric Hessian from the given triangle */
342 static void
343 add2hessian(FVECT hess[3], FVECT ehess1[3],
344 FVECT ehess2[3], FVECT ehess3[3], COLORV v)
345 {
346 int i, j;
347
348 for (i = 3; i--; )
349 for (j = 3; j--; )
350 hess[i][j] += v*( ehess1[i][j] + ehess2[i][j] + ehess3[i][j] );
351 }
352
353
354 /* Compute partial displacement form factor gradient for edge */
355 static void
356 comp_gradient(FVECT grad, FFTRI *ftp, FVECT nrm)
357 {
358 FVECT ncp;
359 double f1;
360 int i;
361
362 f1 = 2.0*DOT(nrm, ftp->rcp);
363 VCROSS(ncp, nrm, ftp->e_i);
364 for (i = 3; i--; )
365 grad[i] = (-0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
366 }
367
368
369 /* Reverse gradient calculation result for edge in other direction */
370 static void
371 rev_gradient(FVECT grad)
372 {
373 grad[0] = -grad[0];
374 grad[1] = -grad[1];
375 grad[2] = -grad[2];
376 }
377
378
379 /* Add to displacement gradient from the given triangle */
380 static void
381 add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, COLORV v)
382 {
383 int i;
384
385 for (i = 3; i--; )
386 grad[i] += v*( egrad1[i] + egrad2[i] + egrad3[i] );
387 }
388
389
390 /* Return brightness of furthest ambient sample */
391 static COLORV
392 back_ambval(AMBSAMP *ap1, AMBSAMP *ap2, AMBSAMP *ap3, FVECT orig)
393 {
394 COLORV vback;
395 FVECT vec;
396 double d2, d2best;
397
398 VSUB(vec, ap1->p, orig);
399 d2best = DOT(vec,vec);
400 vback = colval(ap1->v,CIEY);
401 VSUB(vec, ap2->p, orig);
402 d2 = DOT(vec,vec);
403 if (d2 > d2best) {
404 d2best = d2;
405 vback = colval(ap2->v,CIEY);
406 }
407 VSUB(vec, ap3->p, orig);
408 d2 = DOT(vec,vec);
409 if (d2 > d2best)
410 return(colval(ap3->v,CIEY));
411 return(vback);
412 }
413
414
415 /* Compute anisotropic radii and eigenvector directions */
416 static int
417 eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3])
418 {
419 double hess2[2][2];
420 FVECT a, b;
421 double evalue[2], slope1, xmag1;
422 int i;
423 /* project Hessian to sample plane */
424 for (i = 3; i--; ) {
425 a[i] = DOT(hessian[i], uv[0]);
426 b[i] = DOT(hessian[i], uv[1]);
427 }
428 hess2[0][0] = DOT(uv[0], a);
429 hess2[0][1] = DOT(uv[0], b);
430 hess2[1][0] = DOT(uv[1], a);
431 hess2[1][1] = DOT(uv[1], b);
432 /* compute eigenvalue(s) */
433 i = quadratic(evalue, 1.0, -hess2[0][0]-hess2[1][1],
434 hess2[0][0]*hess2[1][1]-hess2[0][1]*hess2[1][0]);
435 if (i == 1) /* double-root (circle) */
436 evalue[1] = evalue[0];
437 if (!i || ((evalue[0] = fabs(evalue[0])) <= FTINY*FTINY) |
438 ((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) )
439 error(INTERNAL, "bad eigenvalue calculation");
440
441 if (evalue[0] > evalue[1]) {
442 ra[0] = sqrt(sqrt(4.0/evalue[0]));
443 ra[1] = sqrt(sqrt(4.0/evalue[1]));
444 slope1 = evalue[1];
445 } else {
446 ra[0] = sqrt(sqrt(4.0/evalue[1]));
447 ra[1] = sqrt(sqrt(4.0/evalue[0]));
448 slope1 = evalue[0];
449 }
450 /* compute unit eigenvectors */
451 if (fabs(hess2[0][1]) <= FTINY)
452 return; /* uv OK as is */
453 slope1 = (slope1 - hess2[0][0]) / hess2[0][1];
454 xmag1 = sqrt(1.0/(1.0 + slope1*slope1));
455 for (i = 3; i--; ) {
456 b[i] = xmag1*uv[0][i] + slope1*xmag1*uv[1][i];
457 a[i] = slope1*xmag1*uv[0][i] - xmag1*uv[1][i];
458 }
459 VCOPY(uv[0], a);
460 VCOPY(uv[1], b);
461 }
462
463
464 static void
465 ambHessian( /* anisotropic radii & pos. gradient */
466 AMBHEMI *hp,
467 FVECT uv[2], /* returned */
468 float ra[2], /* returned (optional) */
469 float pg[2] /* returned (optional) */
470 )
471 {
472 static char memerrmsg[] = "out of memory in ambHessian()";
473 FVECT (*hessrow)[3] = NULL;
474 FVECT *gradrow = NULL;
475 FVECT hessian[3];
476 FVECT gradient;
477 FFTRI fftr;
478 int i, j;
479 /* be sure to assign unit vectors */
480 VCOPY(uv[0], hp->ux);
481 VCOPY(uv[1], hp->uy);
482 /* clock-wise vertex traversal from sample POV */
483 if (ra != NULL) { /* initialize Hessian row buffer */
484 hessrow = (FVECT (*)[3])malloc(sizeof(FVECT)*3*(hp->ns-1));
485 if (hessrow == NULL)
486 error(SYSTEM, memerrmsg);
487 memset(hessian, 0, sizeof(hessian));
488 } else if (pg == NULL) /* bogus call? */
489 return;
490 if (pg != NULL) { /* initialize form factor row buffer */
491 gradrow = (FVECT *)malloc(sizeof(FVECT)*(hp->ns-1));
492 if (gradrow == NULL)
493 error(SYSTEM, memerrmsg);
494 memset(gradient, 0, sizeof(gradient));
495 }
496 /* compute first row of edges */
497 for (j = 0; j < hp->ns-1; j++) {
498 comp_fftri(&fftr, ambsamp(hp,0,j).p,
499 ambsamp(hp,0,j+1).p, hp->rp->rop);
500 if (hessrow != NULL)
501 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
502 if (gradrow != NULL)
503 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
504 }
505 /* sum each row of triangles */
506 for (i = 0; i < hp->ns-1; i++) {
507 FVECT hesscol[3]; /* compute first vertical edge */
508 FVECT gradcol;
509 comp_fftri(&fftr, ambsamp(hp,i,0).p,
510 ambsamp(hp,i+1,0).p, hp->rp->rop);
511 if (hessrow != NULL)
512 comp_hessian(hesscol, &fftr, hp->rp->ron);
513 if (gradrow != NULL)
514 comp_gradient(gradcol, &fftr, hp->rp->ron);
515 for (j = 0; j < hp->ns-1; j++) {
516 FVECT hessdia[3]; /* compute triangle contributions */
517 FVECT graddia;
518 COLORV backg;
519 backg = back_ambval(&ambsamp(hp,i,j), &ambsamp(hp,i,j+1),
520 &ambsamp(hp,i+1,j), hp->rp->rop);
521 /* diagonal (inner) edge */
522 comp_fftri(&fftr, ambsamp(hp,i,j+1).p,
523 ambsamp(hp,i+1,j).p, hp->rp->rop);
524 if (hessrow != NULL) {
525 comp_hessian(hessdia, &fftr, hp->rp->ron);
526 rev_hessian(hesscol);
527 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
528 }
529 if (gradrow != NULL) {
530 comp_gradient(graddia, &fftr, hp->rp->ron);
531 rev_gradient(gradcol);
532 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
533 }
534 /* initialize edge in next row */
535 comp_fftri(&fftr, ambsamp(hp,i+1,j+1).p,
536 ambsamp(hp,i+1,j).p, hp->rp->rop);
537 if (hessrow != NULL)
538 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
539 if (gradrow != NULL)
540 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
541 /* new column edge & paired triangle */
542 backg = back_ambval(&ambsamp(hp,i,j+1), &ambsamp(hp,i+1,j+1),
543 &ambsamp(hp,i+1,j), hp->rp->rop);
544 comp_fftri(&fftr, ambsamp(hp,i,j+1).p, ambsamp(hp,i+1,j+1).p,
545 hp->rp->rop);
546 if (hessrow != NULL) {
547 comp_hessian(hesscol, &fftr, hp->rp->ron);
548 rev_hessian(hessdia);
549 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
550 if (i < hp->ns-2)
551 rev_hessian(hessrow[j]);
552 }
553 if (gradrow != NULL) {
554 comp_gradient(gradcol, &fftr, hp->rp->ron);
555 rev_gradient(graddia);
556 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
557 if (i < hp->ns-2)
558 rev_gradient(gradrow[j]);
559 }
560 }
561 }
562 /* release row buffers */
563 if (hessrow != NULL) free(hessrow);
564 if (gradrow != NULL) free(gradrow);
565
566 if (ra != NULL) /* extract eigenvectors & radii */
567 eigenvectors(uv, ra, hessian);
568 if (pg != NULL) { /* tangential position gradient */
569 pg[0] = DOT(gradient, uv[0]);
570 pg[1] = DOT(gradient, uv[1]);
571 }
572 }
573
574
575 /* Compute direction gradient from a hemispherical sampling */
576 static void
577 ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
578 {
579 AMBSAMP *ap;
580 double dgsum[2];
581 int n;
582 FVECT vd;
583 double gfact;
584
585 dgsum[0] = dgsum[1] = 0.0; /* sum values times -tan(theta) */
586 for (ap = hp->sa, n = hp->ns*hp->ns; n--; ap++) {
587 /* use vector for azimuth + 90deg */
588 VSUB(vd, ap->p, hp->rp->rop);
589 /* brightness over cosine factor */
590 gfact = colval(ap->v,CIEY) / DOT(hp->rp->ron, vd);
591 /* sine = proj_radius/vd_length */
592 dgsum[0] -= DOT(uv[1], vd) * gfact;
593 dgsum[1] += DOT(uv[0], vd) * gfact;
594 }
595 dg[0] = dgsum[0] / (hp->ns*hp->ns);
596 dg[1] = dgsum[1] / (hp->ns*hp->ns);
597 }
598
599
600 int
601 doambient( /* compute ambient component */
602 COLOR rcol, /* input/output color */
603 RAY *r,
604 double wt,
605 FVECT uv[2], /* returned (optional) */
606 float ra[2], /* returned (optional) */
607 float pg[2], /* returned (optional) */
608 float dg[2] /* returned (optional) */
609 )
610 {
611 AMBHEMI *hp = inithemi(rcol, r, wt);
612 int cnt = 0;
613 FVECT my_uv[2];
614 double d, K, acol[3];
615 AMBSAMP *ap;
616 int i, j;
617 /* check/initialize */
618 if (hp == NULL)
619 return(0);
620 if (uv != NULL)
621 memset(uv, 0, sizeof(FVECT)*2);
622 if (ra != NULL)
623 ra[0] = ra[1] = 0.0;
624 if (pg != NULL)
625 pg[0] = pg[1] = 0.0;
626 if (dg != NULL)
627 dg[0] = dg[1] = 0.0;
628 /* sample the hemisphere */
629 acol[0] = acol[1] = acol[2] = 0.0;
630 for (i = hp->ns; i--; )
631 for (j = hp->ns; j--; )
632 if ((ap = ambsample(hp, i, j)) != NULL) {
633 addcolor(acol, ap->v);
634 ++cnt;
635 }
636 if (!cnt) {
637 setcolor(rcol, 0.0, 0.0, 0.0);
638 free(hp);
639 return(0); /* no valid samples */
640 }
641 if (cnt < hp->ns*hp->ns) { /* incomplete sampling? */
642 copycolor(rcol, acol);
643 free(hp);
644 return(-1); /* return value w/o Hessian */
645 }
646 cnt = ambssamp*wt + 0.5; /* perform super-sampling? */
647 if (cnt > 0)
648 ambsupersamp(acol, hp, cnt);
649 copycolor(rcol, acol); /* final indirect irradiance/PI */
650 if ((ra == NULL) & (pg == NULL) & (dg == NULL)) {
651 free(hp);
652 return(-1); /* no radius or gradient calc. */
653 }
654 if (bright(acol) > FTINY) { /* normalize Y values */
655 d = 0.99*cnt/bright(acol);
656 K = 0.01;
657 } else { /* geometric Hessian fall-back */
658 d = 0.0;
659 K = 1.0;
660 pg = NULL;
661 dg = NULL;
662 }
663 ap = hp->sa; /* relative Y channel from here on... */
664 for (i = hp->ns*hp->ns; i--; ap++)
665 colval(ap->v,CIEY) = bright(ap->v)*d + K;
666
667 if (uv == NULL) /* make sure we have axis pointers */
668 uv = my_uv;
669 /* compute radii & pos. gradient */
670 ambHessian(hp, uv, ra, pg);
671
672 if (dg != NULL) /* compute direction gradient */
673 ambdirgrad(hp, uv, dg);
674
675 if (ra != NULL) { /* scale/clamp radii */
676 if (pg != NULL) {
677 if (ra[0]*(d = fabs(pg[0])) > 1.0)
678 ra[0] = 1.0/d;
679 if (ra[1]*(d = fabs(pg[1])) > 1.0)
680 ra[1] = 1.0/d;
681 if (ra[0] > ra[1])
682 ra[0] = ra[1];
683 }
684 if (ra[0] < minarad) {
685 ra[0] = minarad;
686 if (ra[1] < minarad)
687 ra[1] = minarad;
688 }
689 ra[0] *= d = 1.0/sqrt(sqrt(wt));
690 if ((ra[1] *= d) > 2.0*ra[0])
691 ra[1] = 2.0*ra[0];
692 if (ra[1] > maxarad) {
693 ra[1] = maxarad;
694 if (ra[0] > maxarad)
695 ra[0] = maxarad;
696 }
697 if (pg != NULL) { /* cap gradient if necessary */
698 d = pg[0]*pg[0]*ra[0]*ra[0] + pg[1]*pg[1]*ra[1]*ra[1];
699 if (d > 1.0) {
700 d = 1.0/sqrt(d);
701 pg[0] *= d;
702 pg[1] *= d;
703 }
704 }
705 }
706 free(hp); /* clean up and return */
707 return(1);
708 }
709
710
711 #else /* ! NEWAMB */
712
713
714 void
715 inithemi( /* initialize sampling hemisphere */
716 AMBHEMI *hp,
717 COLOR ac,
718 RAY *r,
719 double wt
720 )
721 {
722 double d;
723 int i;
724 /* set number of divisions */
725 if (ambacc <= FTINY &&
726 wt > (d = 0.8*intens(ac)*r->rweight/(ambdiv*minweight)))
727 wt = d; /* avoid ray termination */
728 hp->nt = sqrt(ambdiv * wt / PI) + 0.5;
729 i = ambacc > FTINY ? 3 : 1; /* minimum number of samples */
730 if (hp->nt < i)
731 hp->nt = i;
732 hp->np = PI * hp->nt + 0.5;
733 /* set number of super-samples */
734 hp->ns = ambssamp * wt + 0.5;
735 /* assign coefficient */
736 copycolor(hp->acoef, ac);
737 d = 1.0/(hp->nt*hp->np);
738 scalecolor(hp->acoef, d);
739 /* make axes */
740 VCOPY(hp->uz, r->ron);
741 hp->uy[0] = hp->uy[1] = hp->uy[2] = 0.0;
742 for (i = 0; i < 3; i++)
743 if (hp->uz[i] < 0.6 && hp->uz[i] > -0.6)
744 break;
745 if (i >= 3)
746 error(CONSISTENCY, "bad ray direction in inithemi");
747 hp->uy[i] = 1.0;
748 fcross(hp->ux, hp->uy, hp->uz);
749 normalize(hp->ux);
750 fcross(hp->uy, hp->uz, hp->ux);
751 }
752
753
754 int
755 divsample( /* sample a division */
756 AMBSAMP *dp,
757 AMBHEMI *h,
758 RAY *r
759 )
760 {
761 RAY ar;
762 int hlist[3];
763 double spt[2];
764 double xd, yd, zd;
765 double b2;
766 double phi;
767 int i;
768 /* ambient coefficient for weight */
769 if (ambacc > FTINY)
770 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
771 else
772 copycolor(ar.rcoef, h->acoef);
773 if (rayorigin(&ar, AMBIENT, r, ar.rcoef) < 0)
774 return(-1);
775 if (ambacc > FTINY) {
776 multcolor(ar.rcoef, h->acoef);
777 scalecolor(ar.rcoef, 1./AVGREFL);
778 }
779 hlist[0] = r->rno;
780 hlist[1] = dp->t;
781 hlist[2] = dp->p;
782 multisamp(spt, 2, urand(ilhash(hlist,3)+dp->n));
783 zd = sqrt((dp->t + spt[0])/h->nt);
784 phi = 2.0*PI * (dp->p + spt[1])/h->np;
785 xd = tcos(phi) * zd;
786 yd = tsin(phi) * zd;
787 zd = sqrt(1.0 - zd*zd);
788 for (i = 0; i < 3; i++)
789 ar.rdir[i] = xd*h->ux[i] +
790 yd*h->uy[i] +
791 zd*h->uz[i];
792 checknorm(ar.rdir);
793 dimlist[ndims++] = dp->t*h->np + dp->p + 90171;
794 rayvalue(&ar);
795 ndims--;
796 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
797 addcolor(dp->v, ar.rcol);
798 /* use rt to improve gradient calc */
799 if (ar.rt > FTINY && ar.rt < FHUGE)
800 dp->r += 1.0/ar.rt;
801 /* (re)initialize error */
802 if (dp->n++) {
803 b2 = bright(dp->v)/dp->n - bright(ar.rcol);
804 b2 = b2*b2 + dp->k*((dp->n-1)*(dp->n-1));
805 dp->k = b2/(dp->n*dp->n);
806 } else
807 dp->k = 0.0;
808 return(0);
809 }
810
811
812 static int
813 ambcmp( /* decreasing order */
814 const void *p1,
815 const void *p2
816 )
817 {
818 const AMBSAMP *d1 = (const AMBSAMP *)p1;
819 const AMBSAMP *d2 = (const AMBSAMP *)p2;
820
821 if (d1->k < d2->k)
822 return(1);
823 if (d1->k > d2->k)
824 return(-1);
825 return(0);
826 }
827
828
829 static int
830 ambnorm( /* standard order */
831 const void *p1,
832 const void *p2
833 )
834 {
835 const AMBSAMP *d1 = (const AMBSAMP *)p1;
836 const AMBSAMP *d2 = (const AMBSAMP *)p2;
837 int c;
838
839 if ( (c = d1->t - d2->t) )
840 return(c);
841 return(d1->p - d2->p);
842 }
843
844
845 double
846 doambient( /* compute ambient component */
847 COLOR rcol,
848 RAY *r,
849 double wt,
850 FVECT pg,
851 FVECT dg
852 )
853 {
854 double b, d=0;
855 AMBHEMI hemi;
856 AMBSAMP *div;
857 AMBSAMP dnew;
858 double acol[3];
859 AMBSAMP *dp;
860 double arad;
861 int divcnt;
862 int i, j;
863 /* initialize hemisphere */
864 inithemi(&hemi, rcol, r, wt);
865 divcnt = hemi.nt * hemi.np;
866 /* initialize */
867 if (pg != NULL)
868 pg[0] = pg[1] = pg[2] = 0.0;
869 if (dg != NULL)
870 dg[0] = dg[1] = dg[2] = 0.0;
871 setcolor(rcol, 0.0, 0.0, 0.0);
872 if (divcnt == 0)
873 return(0.0);
874 /* allocate super-samples */
875 if (hemi.ns > 0 || pg != NULL || dg != NULL) {
876 div = (AMBSAMP *)malloc(divcnt*sizeof(AMBSAMP));
877 if (div == NULL)
878 error(SYSTEM, "out of memory in doambient");
879 } else
880 div = NULL;
881 /* sample the divisions */
882 arad = 0.0;
883 acol[0] = acol[1] = acol[2] = 0.0;
884 if ((dp = div) == NULL)
885 dp = &dnew;
886 divcnt = 0;
887 for (i = 0; i < hemi.nt; i++)
888 for (j = 0; j < hemi.np; j++) {
889 dp->t = i; dp->p = j;
890 setcolor(dp->v, 0.0, 0.0, 0.0);
891 dp->r = 0.0;
892 dp->n = 0;
893 if (divsample(dp, &hemi, r) < 0) {
894 if (div != NULL)
895 dp++;
896 continue;
897 }
898 arad += dp->r;
899 divcnt++;
900 if (div != NULL)
901 dp++;
902 else
903 addcolor(acol, dp->v);
904 }
905 if (!divcnt) {
906 if (div != NULL)
907 free((void *)div);
908 return(0.0); /* no samples taken */
909 }
910 if (divcnt < hemi.nt*hemi.np) {
911 pg = dg = NULL; /* incomplete sampling */
912 hemi.ns = 0;
913 } else if (arad > FTINY && divcnt/arad < minarad) {
914 hemi.ns = 0; /* close enough */
915 } else if (hemi.ns > 0) { /* else perform super-sampling? */
916 comperrs(div, &hemi); /* compute errors */
917 qsort(div, divcnt, sizeof(AMBSAMP), ambcmp); /* sort divs */
918 /* super-sample */
919 for (i = hemi.ns; i > 0; i--) {
920 dnew = *div;
921 if (divsample(&dnew, &hemi, r) < 0) {
922 dp++;
923 continue;
924 }
925 dp = div; /* reinsert */
926 j = divcnt < i ? divcnt : i;
927 while (--j > 0 && dnew.k < dp[1].k) {
928 *dp = *(dp+1);
929 dp++;
930 }
931 *dp = dnew;
932 }
933 if (pg != NULL || dg != NULL) /* restore order */
934 qsort(div, divcnt, sizeof(AMBSAMP), ambnorm);
935 }
936 /* compute returned values */
937 if (div != NULL) {
938 arad = 0.0; /* note: divcnt may be < nt*np */
939 for (i = hemi.nt*hemi.np, dp = div; i-- > 0; dp++) {
940 arad += dp->r;
941 if (dp->n > 1) {
942 b = 1.0/dp->n;
943 scalecolor(dp->v, b);
944 dp->r *= b;
945 dp->n = 1;
946 }
947 addcolor(acol, dp->v);
948 }
949 b = bright(acol);
950 if (b > FTINY) {
951 b = 1.0/b; /* compute & normalize gradient(s) */
952 if (pg != NULL) {
953 posgradient(pg, div, &hemi);
954 for (i = 0; i < 3; i++)
955 pg[i] *= b;
956 }
957 if (dg != NULL) {
958 dirgradient(dg, div, &hemi);
959 for (i = 0; i < 3; i++)
960 dg[i] *= b;
961 }
962 }
963 free((void *)div);
964 }
965 copycolor(rcol, acol);
966 if (arad <= FTINY)
967 arad = maxarad;
968 else
969 arad = (divcnt+hemi.ns)/arad;
970 if (pg != NULL) { /* reduce radius if gradient large */
971 d = DOT(pg,pg);
972 if (d*arad*arad > 1.0)
973 arad = 1.0/sqrt(d);
974 }
975 if (arad < minarad) {
976 arad = minarad;
977 if (pg != NULL && d*arad*arad > 1.0) { /* cap gradient */
978 d = 1.0/arad/sqrt(d);
979 for (i = 0; i < 3; i++)
980 pg[i] *= d;
981 }
982 }
983 if ((arad /= sqrt(wt)) > maxarad)
984 arad = maxarad;
985 return(arad);
986 }
987
988
989 void
990 comperrs( /* compute initial error estimates */
991 AMBSAMP *da, /* assumes standard ordering */
992 AMBHEMI *hp
993 )
994 {
995 double b, b2;
996 int i, j;
997 AMBSAMP *dp;
998 /* sum differences from neighbors */
999 dp = da;
1000 for (i = 0; i < hp->nt; i++)
1001 for (j = 0; j < hp->np; j++) {
1002 #ifdef DEBUG
1003 if (dp->t != i || dp->p != j)
1004 error(CONSISTENCY,
1005 "division order in comperrs");
1006 #endif
1007 b = bright(dp[0].v);
1008 if (i > 0) { /* from above */
1009 b2 = bright(dp[-hp->np].v) - b;
1010 b2 *= b2 * 0.25;
1011 dp[0].k += b2;
1012 dp[-hp->np].k += b2;
1013 }
1014 if (j > 0) { /* from behind */
1015 b2 = bright(dp[-1].v) - b;
1016 b2 *= b2 * 0.25;
1017 dp[0].k += b2;
1018 dp[-1].k += b2;
1019 } else { /* around */
1020 b2 = bright(dp[hp->np-1].v) - b;
1021 b2 *= b2 * 0.25;
1022 dp[0].k += b2;
1023 dp[hp->np-1].k += b2;
1024 }
1025 dp++;
1026 }
1027 /* divide by number of neighbors */
1028 dp = da;
1029 for (j = 0; j < hp->np; j++) /* top row */
1030 (dp++)->k *= 1.0/3.0;
1031 if (hp->nt < 2)
1032 return;
1033 for (i = 1; i < hp->nt-1; i++) /* central region */
1034 for (j = 0; j < hp->np; j++)
1035 (dp++)->k *= 0.25;
1036 for (j = 0; j < hp->np; j++) /* bottom row */
1037 (dp++)->k *= 1.0/3.0;
1038 }
1039
1040
1041 void
1042 posgradient( /* compute position gradient */
1043 FVECT gv,
1044 AMBSAMP *da, /* assumes standard ordering */
1045 AMBHEMI *hp
1046 )
1047 {
1048 int i, j;
1049 double nextsine, lastsine, b, d;
1050 double mag0, mag1;
1051 double phi, cosp, sinp, xd, yd;
1052 AMBSAMP *dp;
1053
1054 xd = yd = 0.0;
1055 for (j = 0; j < hp->np; j++) {
1056 dp = da + j;
1057 mag0 = mag1 = 0.0;
1058 lastsine = 0.0;
1059 for (i = 0; i < hp->nt; i++) {
1060 #ifdef DEBUG
1061 if (dp->t != i || dp->p != j)
1062 error(CONSISTENCY,
1063 "division order in posgradient");
1064 #endif
1065 b = bright(dp->v);
1066 if (i > 0) {
1067 d = dp[-hp->np].r;
1068 if (dp[0].r > d) d = dp[0].r;
1069 /* sin(t)*cos(t)^2 */
1070 d *= lastsine * (1.0 - (double)i/hp->nt);
1071 mag0 += d*(b - bright(dp[-hp->np].v));
1072 }
1073 nextsine = sqrt((double)(i+1)/hp->nt);
1074 if (j > 0) {
1075 d = dp[-1].r;
1076 if (dp[0].r > d) d = dp[0].r;
1077 mag1 += d * (nextsine - lastsine) *
1078 (b - bright(dp[-1].v));
1079 } else {
1080 d = dp[hp->np-1].r;
1081 if (dp[0].r > d) d = dp[0].r;
1082 mag1 += d * (nextsine - lastsine) *
1083 (b - bright(dp[hp->np-1].v));
1084 }
1085 dp += hp->np;
1086 lastsine = nextsine;
1087 }
1088 mag0 *= 2.0*PI / hp->np;
1089 phi = 2.0*PI * (double)j/hp->np;
1090 cosp = tcos(phi); sinp = tsin(phi);
1091 xd += mag0*cosp - mag1*sinp;
1092 yd += mag0*sinp + mag1*cosp;
1093 }
1094 for (i = 0; i < 3; i++)
1095 gv[i] = (xd*hp->ux[i] + yd*hp->uy[i])*(hp->nt*hp->np)/PI;
1096 }
1097
1098
1099 void
1100 dirgradient( /* compute direction gradient */
1101 FVECT gv,
1102 AMBSAMP *da, /* assumes standard ordering */
1103 AMBHEMI *hp
1104 )
1105 {
1106 int i, j;
1107 double mag;
1108 double phi, xd, yd;
1109 AMBSAMP *dp;
1110
1111 xd = yd = 0.0;
1112 for (j = 0; j < hp->np; j++) {
1113 dp = da + j;
1114 mag = 0.0;
1115 for (i = 0; i < hp->nt; i++) {
1116 #ifdef DEBUG
1117 if (dp->t != i || dp->p != j)
1118 error(CONSISTENCY,
1119 "division order in dirgradient");
1120 #endif
1121 /* tan(t) */
1122 mag += bright(dp->v)/sqrt(hp->nt/(i+.5) - 1.0);
1123 dp += hp->np;
1124 }
1125 phi = 2.0*PI * (j+.5)/hp->np + PI/2.0;
1126 xd += mag * tcos(phi);
1127 yd += mag * tsin(phi);
1128 }
1129 for (i = 0; i < 3; i++)
1130 gv[i] = xd*hp->ux[i] + yd*hp->uy[i];
1131 }
1132
1133 #endif /* ! NEWAMB */