ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambcomp.c
Revision: 2.78
Committed: Tue Jan 9 00:51:51 2018 UTC (6 years, 4 months ago) by greg
Content type: text/plain
Branch: MAIN
Changes since 2.77: +8 -2 lines
Log Message:
Attempt to fix issue with low-angle ambient through trans and the like

File Contents

# Content
1 #ifndef lint
2 static const char RCSid[] = "$Id: ambcomp.c,v 2.77 2017/04/21 16:07:29 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 * Added book-keeping optimization to avoid calculations that would
12 * cancel due to traversal both directions on edges that are adjacent
13 * to same-valued triangles. This cuts about half of Hessian math.
14 *
15 * Declarations of external symbols in ambient.h
16 */
17
18 #include "copyright.h"
19
20 #include "ray.h"
21 #include "ambient.h"
22 #include "random.h"
23 #include "source.h"
24 #include "otypes.h"
25 #include "otspecial.h"
26
27 #ifndef OLDAMB
28
29 extern void SDsquare2disk(double ds[2], double seedx, double seedy);
30
31 typedef struct {
32 COLOR v; /* hemisphere sample value */
33 float d; /* reciprocal distance (1/rt) */
34 FVECT p; /* intersection point */
35 } AMBSAMP; /* sample value */
36
37 typedef struct {
38 RAY *rp; /* originating ray sample */
39 int ns; /* number of samples per axis */
40 int sampOK; /* acquired full sample set? */
41 COLOR acoef; /* division contribution coefficient */
42 double acol[3]; /* accumulated color */
43 FVECT ux, uy; /* tangent axis unit vectors */
44 AMBSAMP sa[1]; /* sample array (extends struct) */
45 } AMBHEMI; /* ambient sample hemisphere */
46
47 #define AI(h,i,j) ((i)*(h)->ns + (j))
48 #define ambsam(h,i,j) (h)->sa[AI(h,i,j)]
49
50 typedef struct {
51 FVECT r_i, r_i1, e_i, rcp, rI2_eJ2;
52 double I1, I2;
53 } FFTRI; /* vectors and coefficients for Hessian calculation */
54
55
56 static int
57 ambcollision( /* proposed direciton collides? */
58 AMBHEMI *hp,
59 int i,
60 int j,
61 FVECT dv
62 )
63 {
64 double cos_thresh;
65 int ii, jj;
66 /* min. spacing = 1/4th division */
67 cos_thresh = (PI/4.)/(double)hp->ns;
68 cos_thresh = 1. - .5*cos_thresh*cos_thresh;
69 /* check existing neighbors */
70 for (ii = i-1; ii <= i+1; ii++) {
71 if (ii < 0) continue;
72 if (ii >= hp->ns) break;
73 for (jj = j-1; jj <= j+1; jj++) {
74 AMBSAMP *ap;
75 FVECT avec;
76 double dprod;
77 if (jj < 0) continue;
78 if (jj >= hp->ns) break;
79 if ((ii==i) & (jj==j)) continue;
80 ap = &ambsam(hp,ii,jj);
81 if (ap->d <= .5/FHUGE)
82 continue; /* no one home */
83 VSUB(avec, ap->p, hp->rp->rop);
84 dprod = DOT(avec, dv);
85 if (dprod >= cos_thresh*VLEN(avec))
86 return(1); /* collision */
87 }
88 }
89 return(0); /* nothing to worry about */
90 }
91
92
93 static int
94 ambsample( /* initial ambient division sample */
95 AMBHEMI *hp,
96 int i,
97 int j,
98 int n
99 )
100 {
101 AMBSAMP *ap = &ambsam(hp,i,j);
102 RAY ar;
103 int hlist[3], ii;
104 double spt[2], zd;
105 /* generate hemispherical sample */
106 /* ambient coefficient for weight */
107 if (ambacc > FTINY)
108 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
109 else
110 copycolor(ar.rcoef, hp->acoef);
111 if (rayorigin(&ar, AMBIENT, hp->rp, ar.rcoef) < 0)
112 return(0);
113 if (ambacc > FTINY) {
114 multcolor(ar.rcoef, hp->acoef);
115 scalecolor(ar.rcoef, 1./AVGREFL);
116 }
117 hlist[0] = hp->rp->rno;
118 hlist[1] = j;
119 hlist[2] = i;
120 multisamp(spt, 2, urand(ilhash(hlist,3)+n));
121 resample:
122 SDsquare2disk(spt, (j+spt[1])/hp->ns, (i+spt[0])/hp->ns);
123 zd = sqrt(1. - spt[0]*spt[0] - spt[1]*spt[1]);
124 for (ii = 3; ii--; )
125 ar.rdir[ii] = spt[0]*hp->ux[ii] +
126 spt[1]*hp->uy[ii] +
127 zd*hp->rp->ron[ii];
128 checknorm(ar.rdir);
129 /* avoid coincident samples */
130 if (!n && ambcollision(hp, i, j, ar.rdir)) {
131 spt[0] = frandom(); spt[1] = frandom();
132 goto resample; /* reject this sample */
133 }
134 dimlist[ndims++] = AI(hp,i,j) + 90171;
135 rayvalue(&ar); /* evaluate ray */
136 ndims--;
137 if (ar.rt <= FTINY)
138 return(0); /* should never happen */
139 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
140 if (ar.rt*ap->d < 1.0) /* new/closer distance? */
141 ap->d = 1.0/ar.rt;
142 if (!n) { /* record first vertex & value */
143 if (ar.rt > 10.0*thescene.cusize + 1000.)
144 ar.rt = 10.0*thescene.cusize + 1000.;
145 VSUM(ap->p, ar.rorg, ar.rdir, ar.rt);
146 copycolor(ap->v, ar.rcol);
147 } else { /* else update recorded value */
148 hp->acol[RED] -= colval(ap->v,RED);
149 hp->acol[GRN] -= colval(ap->v,GRN);
150 hp->acol[BLU] -= colval(ap->v,BLU);
151 zd = 1.0/(double)(n+1);
152 scalecolor(ar.rcol, zd);
153 zd *= (double)n;
154 scalecolor(ap->v, zd);
155 addcolor(ap->v, ar.rcol);
156 }
157 addcolor(hp->acol, ap->v); /* add to our sum */
158 return(1);
159 }
160
161
162 /* Estimate errors based on ambient division differences */
163 static float *
164 getambdiffs(AMBHEMI *hp)
165 {
166 const double normf = 1./bright(hp->acoef);
167 float *earr = (float *)calloc(hp->ns*hp->ns, sizeof(float));
168 float *ep;
169 AMBSAMP *ap;
170 double b, d2;
171 int i, j;
172
173 if (earr == NULL) /* out of memory? */
174 return(NULL);
175 /* compute squared neighbor diffs */
176 for (ap = hp->sa, ep = earr, i = 0; i < hp->ns; i++)
177 for (j = 0; j < hp->ns; j++, ap++, ep++) {
178 b = bright(ap[0].v);
179 if (i) { /* from above */
180 d2 = normf*(b - bright(ap[-hp->ns].v));
181 d2 *= d2;
182 ep[0] += d2;
183 ep[-hp->ns] += d2;
184 }
185 if (!j) continue;
186 /* from behind */
187 d2 = normf*(b - bright(ap[-1].v));
188 d2 *= d2;
189 ep[0] += d2;
190 ep[-1] += d2;
191 if (!i) continue;
192 /* diagonal */
193 d2 = normf*(b - bright(ap[-hp->ns-1].v));
194 d2 *= d2;
195 ep[0] += d2;
196 ep[-hp->ns-1] += d2;
197 }
198 /* correct for number of neighbors */
199 earr[0] *= 8./3.;
200 earr[hp->ns-1] *= 8./3.;
201 earr[(hp->ns-1)*hp->ns] *= 8./3.;
202 earr[(hp->ns-1)*hp->ns + hp->ns-1] *= 8./3.;
203 for (i = 1; i < hp->ns-1; i++) {
204 earr[i*hp->ns] *= 8./5.;
205 earr[i*hp->ns + hp->ns-1] *= 8./5.;
206 }
207 for (j = 1; j < hp->ns-1; j++) {
208 earr[j] *= 8./5.;
209 earr[(hp->ns-1)*hp->ns + j] *= 8./5.;
210 }
211 return(earr);
212 }
213
214
215 /* Perform super-sampling on hemisphere (introduces bias) */
216 static void
217 ambsupersamp(AMBHEMI *hp, int cnt)
218 {
219 float *earr = getambdiffs(hp);
220 double e2rem = 0;
221 AMBSAMP *ap;
222 float *ep;
223 int i, j, n, nss;
224
225 if (earr == NULL) /* just skip calc. if no memory */
226 return;
227 /* accumulate estimated variances */
228 for (ep = earr + hp->ns*hp->ns; ep > earr; )
229 e2rem += *--ep;
230 ep = earr; /* perform super-sampling */
231 for (ap = hp->sa, i = 0; i < hp->ns; i++)
232 for (j = 0; j < hp->ns; j++, ap++) {
233 if (e2rem <= FTINY)
234 goto done; /* nothing left to do */
235 nss = *ep/e2rem*cnt + frandom();
236 for (n = 1; n <= nss && ambsample(hp,i,j,n); n++)
237 if (!--cnt) goto done;
238 e2rem -= *ep++; /* update remainder */
239 }
240 done:
241 free(earr);
242 }
243
244
245 static AMBHEMI *
246 samp_hemi( /* sample indirect hemisphere */
247 COLOR rcol,
248 RAY *r,
249 double wt
250 )
251 {
252 AMBHEMI *hp;
253 double d;
254 int n, i, j;
255 /* insignificance check */
256 if (bright(rcol) <= FTINY)
257 return(NULL);
258 /* set number of divisions */
259 if (ambacc <= FTINY &&
260 wt > (d = 0.8*intens(rcol)*r->rweight/(ambdiv*minweight)))
261 wt = d; /* avoid ray termination */
262 n = sqrt(ambdiv * wt) + 0.5;
263 i = 1 + 5*(ambacc > FTINY); /* minimum number of samples */
264 if (n < i)
265 n = i;
266 /* allocate sampling array */
267 hp = (AMBHEMI *)malloc(sizeof(AMBHEMI) + sizeof(AMBSAMP)*(n*n - 1));
268 if (hp == NULL)
269 error(SYSTEM, "out of memory in samp_hemi");
270 hp->rp = r;
271 hp->ns = n;
272 hp->acol[RED] = hp->acol[GRN] = hp->acol[BLU] = 0.0;
273 memset(hp->sa, 0, sizeof(AMBSAMP)*n*n);
274 hp->sampOK = 0;
275 /* assign coefficient */
276 copycolor(hp->acoef, rcol);
277 d = 1.0/(n*n);
278 scalecolor(hp->acoef, d);
279 /* make tangent plane axes */
280 if (!getperpendicular(hp->ux, r->ron, 1))
281 error(CONSISTENCY, "bad ray direction in samp_hemi");
282 VCROSS(hp->uy, r->ron, hp->ux);
283 /* sample divisions */
284 for (i = hp->ns; i--; )
285 for (j = hp->ns; j--; )
286 hp->sampOK += ambsample(hp, i, j, 0);
287 copycolor(rcol, hp->acol);
288 if (!hp->sampOK) { /* utter failure? */
289 free(hp);
290 return(NULL);
291 }
292 if (hp->sampOK < hp->ns*hp->ns) {
293 hp->sampOK *= -1; /* soft failure */
294 return(hp);
295 }
296 n = ambssamp*wt + 0.5;
297 if (n > 8) { /* perform super-sampling? */
298 ambsupersamp(hp, n);
299 copycolor(rcol, hp->acol);
300 }
301 return(hp); /* all is well */
302 }
303
304
305 /* Return brightness of farthest ambient sample */
306 static double
307 back_ambval(AMBHEMI *hp, const int n1, const int n2, const int n3)
308 {
309 if (hp->sa[n1].d <= hp->sa[n2].d) {
310 if (hp->sa[n1].d <= hp->sa[n3].d)
311 return(colval(hp->sa[n1].v,CIEY));
312 return(colval(hp->sa[n3].v,CIEY));
313 }
314 if (hp->sa[n2].d <= hp->sa[n3].d)
315 return(colval(hp->sa[n2].v,CIEY));
316 return(colval(hp->sa[n3].v,CIEY));
317 }
318
319
320 /* Compute vectors and coefficients for Hessian/gradient calcs */
321 static void
322 comp_fftri(FFTRI *ftp, AMBHEMI *hp, const int n0, const int n1)
323 {
324 double rdot_cp, dot_e, dot_er, rdot_r, rdot_r1, J2;
325 int ii;
326
327 VSUB(ftp->r_i, hp->sa[n0].p, hp->rp->rop);
328 VSUB(ftp->r_i1, hp->sa[n1].p, hp->rp->rop);
329 VSUB(ftp->e_i, hp->sa[n1].p, hp->sa[n0].p);
330 VCROSS(ftp->rcp, ftp->r_i, ftp->r_i1);
331 rdot_cp = 1.0/DOT(ftp->rcp,ftp->rcp);
332 dot_e = DOT(ftp->e_i,ftp->e_i);
333 dot_er = DOT(ftp->e_i, ftp->r_i);
334 rdot_r = 1.0/DOT(ftp->r_i,ftp->r_i);
335 rdot_r1 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
336 ftp->I1 = acos( DOT(ftp->r_i, ftp->r_i1) * sqrt(rdot_r*rdot_r1) ) *
337 sqrt( rdot_cp );
338 ftp->I2 = ( DOT(ftp->e_i, ftp->r_i1)*rdot_r1 - dot_er*rdot_r +
339 dot_e*ftp->I1 )*0.5*rdot_cp;
340 J2 = ( 0.5*(rdot_r - rdot_r1) - dot_er*ftp->I2 ) / dot_e;
341 for (ii = 3; ii--; )
342 ftp->rI2_eJ2[ii] = ftp->I2*ftp->r_i[ii] + J2*ftp->e_i[ii];
343 }
344
345
346 /* Compose 3x3 matrix from two vectors */
347 static void
348 compose_matrix(FVECT mat[3], FVECT va, FVECT vb)
349 {
350 mat[0][0] = 2.0*va[0]*vb[0];
351 mat[1][1] = 2.0*va[1]*vb[1];
352 mat[2][2] = 2.0*va[2]*vb[2];
353 mat[0][1] = mat[1][0] = va[0]*vb[1] + va[1]*vb[0];
354 mat[0][2] = mat[2][0] = va[0]*vb[2] + va[2]*vb[0];
355 mat[1][2] = mat[2][1] = va[1]*vb[2] + va[2]*vb[1];
356 }
357
358
359 /* Compute partial 3x3 Hessian matrix for edge */
360 static void
361 comp_hessian(FVECT hess[3], FFTRI *ftp, FVECT nrm)
362 {
363 FVECT ncp;
364 FVECT m1[3], m2[3], m3[3], m4[3];
365 double d1, d2, d3, d4;
366 double I3, J3, K3;
367 int i, j;
368 /* compute intermediate coefficients */
369 d1 = 1.0/DOT(ftp->r_i,ftp->r_i);
370 d2 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
371 d3 = 1.0/DOT(ftp->e_i,ftp->e_i);
372 d4 = DOT(ftp->e_i, ftp->r_i);
373 I3 = ( DOT(ftp->e_i, ftp->r_i1)*d2*d2 - d4*d1*d1 + 3.0/d3*ftp->I2 )
374 / ( 4.0*DOT(ftp->rcp,ftp->rcp) );
375 J3 = 0.25*d3*(d1*d1 - d2*d2) - d4*d3*I3;
376 K3 = d3*(ftp->I2 - I3/d1 - 2.0*d4*J3);
377 /* intermediate matrices */
378 VCROSS(ncp, nrm, ftp->e_i);
379 compose_matrix(m1, ncp, ftp->rI2_eJ2);
380 compose_matrix(m2, ftp->r_i, ftp->r_i);
381 compose_matrix(m3, ftp->e_i, ftp->e_i);
382 compose_matrix(m4, ftp->r_i, ftp->e_i);
383 d1 = DOT(nrm, ftp->rcp);
384 d2 = -d1*ftp->I2;
385 d1 *= 2.0;
386 for (i = 3; i--; ) /* final matrix sum */
387 for (j = 3; j--; ) {
388 hess[i][j] = m1[i][j] + d1*( I3*m2[i][j] + K3*m3[i][j] +
389 2.0*J3*m4[i][j] );
390 hess[i][j] += d2*(i==j);
391 hess[i][j] *= -1.0/PI;
392 }
393 }
394
395
396 /* Reverse hessian calculation result for edge in other direction */
397 static void
398 rev_hessian(FVECT hess[3])
399 {
400 int i;
401
402 for (i = 3; i--; ) {
403 hess[i][0] = -hess[i][0];
404 hess[i][1] = -hess[i][1];
405 hess[i][2] = -hess[i][2];
406 }
407 }
408
409
410 /* Add to radiometric Hessian from the given triangle */
411 static void
412 add2hessian(FVECT hess[3], FVECT ehess1[3],
413 FVECT ehess2[3], FVECT ehess3[3], double v)
414 {
415 int i, j;
416
417 for (i = 3; i--; )
418 for (j = 3; j--; )
419 hess[i][j] += v*( ehess1[i][j] + ehess2[i][j] + ehess3[i][j] );
420 }
421
422
423 /* Compute partial displacement form factor gradient for edge */
424 static void
425 comp_gradient(FVECT grad, FFTRI *ftp, FVECT nrm)
426 {
427 FVECT ncp;
428 double f1;
429 int i;
430
431 f1 = 2.0*DOT(nrm, ftp->rcp);
432 VCROSS(ncp, nrm, ftp->e_i);
433 for (i = 3; i--; )
434 grad[i] = (0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
435 }
436
437
438 /* Reverse gradient calculation result for edge in other direction */
439 static void
440 rev_gradient(FVECT grad)
441 {
442 grad[0] = -grad[0];
443 grad[1] = -grad[1];
444 grad[2] = -grad[2];
445 }
446
447
448 /* Add to displacement gradient from the given triangle */
449 static void
450 add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, double v)
451 {
452 int i;
453
454 for (i = 3; i--; )
455 grad[i] += v*( egrad1[i] + egrad2[i] + egrad3[i] );
456 }
457
458
459 /* Compute anisotropic radii and eigenvector directions */
460 static void
461 eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3])
462 {
463 double hess2[2][2];
464 FVECT a, b;
465 double evalue[2], slope1, xmag1;
466 int i;
467 /* project Hessian to sample plane */
468 for (i = 3; i--; ) {
469 a[i] = DOT(hessian[i], uv[0]);
470 b[i] = DOT(hessian[i], uv[1]);
471 }
472 hess2[0][0] = DOT(uv[0], a);
473 hess2[0][1] = DOT(uv[0], b);
474 hess2[1][0] = DOT(uv[1], a);
475 hess2[1][1] = DOT(uv[1], b);
476 /* compute eigenvalue(s) */
477 i = quadratic(evalue, 1.0, -hess2[0][0]-hess2[1][1],
478 hess2[0][0]*hess2[1][1]-hess2[0][1]*hess2[1][0]);
479 if (i == 1) /* double-root (circle) */
480 evalue[1] = evalue[0];
481 if (!i || ((evalue[0] = fabs(evalue[0])) <= FTINY*FTINY) |
482 ((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) ) {
483 ra[0] = ra[1] = maxarad;
484 return;
485 }
486 if (evalue[0] > evalue[1]) {
487 ra[0] = sqrt(sqrt(4.0/evalue[0]));
488 ra[1] = sqrt(sqrt(4.0/evalue[1]));
489 slope1 = evalue[1];
490 } else {
491 ra[0] = sqrt(sqrt(4.0/evalue[1]));
492 ra[1] = sqrt(sqrt(4.0/evalue[0]));
493 slope1 = evalue[0];
494 }
495 /* compute unit eigenvectors */
496 if (fabs(hess2[0][1]) <= FTINY)
497 return; /* uv OK as is */
498 slope1 = (slope1 - hess2[0][0]) / hess2[0][1];
499 xmag1 = sqrt(1.0/(1.0 + slope1*slope1));
500 for (i = 3; i--; ) {
501 b[i] = xmag1*uv[0][i] + slope1*xmag1*uv[1][i];
502 a[i] = slope1*xmag1*uv[0][i] - xmag1*uv[1][i];
503 }
504 VCOPY(uv[0], a);
505 VCOPY(uv[1], b);
506 }
507
508
509 static void
510 ambHessian( /* anisotropic radii & pos. gradient */
511 AMBHEMI *hp,
512 FVECT uv[2], /* returned */
513 float ra[2], /* returned (optional) */
514 float pg[2] /* returned (optional) */
515 )
516 {
517 static char memerrmsg[] = "out of memory in ambHessian()";
518 FVECT (*hessrow)[3] = NULL;
519 FVECT *gradrow = NULL;
520 FVECT hessian[3];
521 FVECT gradient;
522 FFTRI fftr;
523 int i, j;
524 /* be sure to assign unit vectors */
525 VCOPY(uv[0], hp->ux);
526 VCOPY(uv[1], hp->uy);
527 /* clock-wise vertex traversal from sample POV */
528 if (ra != NULL) { /* initialize Hessian row buffer */
529 hessrow = (FVECT (*)[3])malloc(sizeof(FVECT)*3*(hp->ns-1));
530 if (hessrow == NULL)
531 error(SYSTEM, memerrmsg);
532 memset(hessian, 0, sizeof(hessian));
533 } else if (pg == NULL) /* bogus call? */
534 return;
535 if (pg != NULL) { /* initialize form factor row buffer */
536 gradrow = (FVECT *)malloc(sizeof(FVECT)*(hp->ns-1));
537 if (gradrow == NULL)
538 error(SYSTEM, memerrmsg);
539 memset(gradient, 0, sizeof(gradient));
540 }
541 /* compute first row of edges */
542 for (j = 0; j < hp->ns-1; j++) {
543 comp_fftri(&fftr, hp, AI(hp,0,j), AI(hp,0,j+1));
544 if (hessrow != NULL)
545 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
546 if (gradrow != NULL)
547 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
548 }
549 /* sum each row of triangles */
550 for (i = 0; i < hp->ns-1; i++) {
551 FVECT hesscol[3]; /* compute first vertical edge */
552 FVECT gradcol;
553 comp_fftri(&fftr, hp, AI(hp,i,0), AI(hp,i+1,0));
554 if (hessrow != NULL)
555 comp_hessian(hesscol, &fftr, hp->rp->ron);
556 if (gradrow != NULL)
557 comp_gradient(gradcol, &fftr, hp->rp->ron);
558 for (j = 0; j < hp->ns-1; j++) {
559 FVECT hessdia[3]; /* compute triangle contributions */
560 FVECT graddia;
561 double backg;
562 backg = back_ambval(hp, AI(hp,i,j),
563 AI(hp,i,j+1), AI(hp,i+1,j));
564 /* diagonal (inner) edge */
565 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j));
566 if (hessrow != NULL) {
567 comp_hessian(hessdia, &fftr, hp->rp->ron);
568 rev_hessian(hesscol);
569 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
570 }
571 if (gradrow != NULL) {
572 comp_gradient(graddia, &fftr, hp->rp->ron);
573 rev_gradient(gradcol);
574 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
575 }
576 /* initialize edge in next row */
577 comp_fftri(&fftr, hp, AI(hp,i+1,j+1), AI(hp,i+1,j));
578 if (hessrow != NULL)
579 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
580 if (gradrow != NULL)
581 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
582 /* new column edge & paired triangle */
583 backg = back_ambval(hp, AI(hp,i+1,j+1),
584 AI(hp,i+1,j), AI(hp,i,j+1));
585 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j+1));
586 if (hessrow != NULL) {
587 comp_hessian(hesscol, &fftr, hp->rp->ron);
588 rev_hessian(hessdia);
589 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
590 if (i < hp->ns-2)
591 rev_hessian(hessrow[j]);
592 }
593 if (gradrow != NULL) {
594 comp_gradient(gradcol, &fftr, hp->rp->ron);
595 rev_gradient(graddia);
596 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
597 if (i < hp->ns-2)
598 rev_gradient(gradrow[j]);
599 }
600 }
601 }
602 /* release row buffers */
603 if (hessrow != NULL) free(hessrow);
604 if (gradrow != NULL) free(gradrow);
605
606 if (ra != NULL) /* extract eigenvectors & radii */
607 eigenvectors(uv, ra, hessian);
608 if (pg != NULL) { /* tangential position gradient */
609 pg[0] = DOT(gradient, uv[0]);
610 pg[1] = DOT(gradient, uv[1]);
611 }
612 }
613
614
615 /* Compute direction gradient from a hemispherical sampling */
616 static void
617 ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
618 {
619 AMBSAMP *ap;
620 double dgsum[2];
621 int n;
622 FVECT vd;
623 double gfact;
624
625 dgsum[0] = dgsum[1] = 0.0; /* sum values times -tan(theta) */
626 for (ap = hp->sa, n = hp->ns*hp->ns; n--; ap++) {
627 /* use vector for azimuth + 90deg */
628 VSUB(vd, ap->p, hp->rp->rop);
629 /* brightness over cosine factor */
630 gfact = colval(ap->v,CIEY) / DOT(hp->rp->ron, vd);
631 /* sine = proj_radius/vd_length */
632 dgsum[0] -= DOT(uv[1], vd) * gfact;
633 dgsum[1] += DOT(uv[0], vd) * gfact;
634 }
635 dg[0] = dgsum[0] / (hp->ns*hp->ns);
636 dg[1] = dgsum[1] / (hp->ns*hp->ns);
637 }
638
639
640 /* Compute potential light leak direction flags for cache value */
641 static uint32
642 ambcorral(AMBHEMI *hp, FVECT uv[2], const double r0, const double r1)
643 {
644 const double max_d = 1.0/(minarad*ambacc + 0.001);
645 const double ang_res = 0.5*PI/hp->ns;
646 const double ang_step = ang_res/((int)(16/PI*ang_res) + 1.01);
647 double avg_d = 0;
648 uint32 flgs = 0;
649 FVECT vec;
650 double u, v;
651 double ang, a1;
652 OBJREC *m;
653 int i, j;
654 /* don't bother for a few samples */
655 if (hp->ns < 8)
656 return(0);
657 /* check distances overhead */
658 for (i = hp->ns*3/4; i-- > hp->ns>>2; )
659 for (j = hp->ns*3/4; j-- > hp->ns>>2; )
660 avg_d += ambsam(hp,i,j).d;
661 avg_d *= 4.0/(hp->ns*hp->ns);
662 if (avg_d*r0 >= 1.0) /* ceiling too low for corral? */
663 return(0);
664 if (avg_d >= max_d) /* insurance */
665 return(0);
666 /* else circle around perimeter */
667 for (i = 0; i < hp->ns; i++)
668 for (j = 0; j < hp->ns; j += !i|(i==hp->ns-1) ? 1 : hp->ns-1) {
669 AMBSAMP *ap = &ambsam(hp,i,j);
670 if ((ap->d <= FTINY) | (ap->d >= max_d))
671 continue; /* too far or too near */
672 VSUB(vec, ap->p, hp->rp->rop);
673 u = DOT(vec, uv[0]);
674 v = DOT(vec, uv[1]);
675 if ((r0*r0*u*u + r1*r1*v*v) * ap->d*ap->d <= u*u + v*v)
676 continue; /* occluder outside ellipse */
677 ang = atan2a(v, u); /* else set direction flags */
678 for (a1 = ang-ang_res; a1 <= ang+ang_res; a1 += ang_step)
679 flgs |= 1L<<(int)(16/PI*(a1 + 2.*PI*(a1 < 0)));
680 }
681 /* add low-angle incident (< 20deg) */
682 if (fabs(hp->rp->rod) <= 0.342 && hp->rp->parent != NULL &&
683 (m = findmaterial(hp->rp->parent->ro)) != NULL &&
684 isopaque(m->otype)) {
685 u = -DOT(hp->rp->rdir, uv[0]);
686 v = -DOT(hp->rp->rdir, uv[1]);
687 if ((r0*r0*u*u + r1*r1*v*v) > hp->rp->rot*hp->rp->rot) {
688 ang = atan2a(v, u);
689 ang += 2.*PI*(ang < 0);
690 ang *= 16/PI;
691 if ((ang < .5) | (ang >= 31.5))
692 flgs |= 0x80000001;
693 else
694 flgs |= 3L<<(int)(ang-.5);
695 }
696 }
697 return(flgs);
698 }
699
700
701 int
702 doambient( /* compute ambient component */
703 COLOR rcol, /* input/output color */
704 RAY *r,
705 double wt,
706 FVECT uv[2], /* returned (optional) */
707 float ra[2], /* returned (optional) */
708 float pg[2], /* returned (optional) */
709 float dg[2], /* returned (optional) */
710 uint32 *crlp /* returned (optional) */
711 )
712 {
713 AMBHEMI *hp = samp_hemi(rcol, r, wt);
714 FVECT my_uv[2];
715 double d, K;
716 AMBSAMP *ap;
717 int i;
718 /* clear return values */
719 if (uv != NULL)
720 memset(uv, 0, sizeof(FVECT)*2);
721 if (ra != NULL)
722 ra[0] = ra[1] = 0.0;
723 if (pg != NULL)
724 pg[0] = pg[1] = 0.0;
725 if (dg != NULL)
726 dg[0] = dg[1] = 0.0;
727 if (crlp != NULL)
728 *crlp = 0;
729 if (hp == NULL) /* sampling falure? */
730 return(0);
731
732 if ((ra == NULL) & (pg == NULL) & (dg == NULL) ||
733 (hp->sampOK < 0) | (hp->ns < 6)) {
734 free(hp); /* Hessian not requested/possible */
735 return(-1); /* value-only return value */
736 }
737 if ((d = bright(rcol)) > FTINY) { /* normalize Y values */
738 d = 0.99*(hp->ns*hp->ns)/d;
739 K = 0.01;
740 } else { /* or fall back on geometric Hessian */
741 K = 1.0;
742 pg = NULL;
743 dg = NULL;
744 crlp = NULL;
745 }
746 ap = hp->sa; /* relative Y channel from here on... */
747 for (i = hp->ns*hp->ns; i--; ap++)
748 colval(ap->v,CIEY) = bright(ap->v)*d + K;
749
750 if (uv == NULL) /* make sure we have axis pointers */
751 uv = my_uv;
752 /* compute radii & pos. gradient */
753 ambHessian(hp, uv, ra, pg);
754
755 if (dg != NULL) /* compute direction gradient */
756 ambdirgrad(hp, uv, dg);
757
758 if (ra != NULL) { /* scale/clamp radii */
759 if (pg != NULL) {
760 if (ra[0]*(d = fabs(pg[0])) > 1.0)
761 ra[0] = 1.0/d;
762 if (ra[1]*(d = fabs(pg[1])) > 1.0)
763 ra[1] = 1.0/d;
764 if (ra[0] > ra[1])
765 ra[0] = ra[1];
766 }
767 if (ra[0] < minarad) {
768 ra[0] = minarad;
769 if (ra[1] < minarad)
770 ra[1] = minarad;
771 }
772 ra[0] *= d = 1.0/sqrt(wt);
773 if ((ra[1] *= d) > 2.0*ra[0])
774 ra[1] = 2.0*ra[0];
775 if (ra[1] > maxarad) {
776 ra[1] = maxarad;
777 if (ra[0] > maxarad)
778 ra[0] = maxarad;
779 }
780 /* flag encroached directions */
781 if (crlp != NULL)
782 *crlp = ambcorral(hp, uv, ra[0]*ambacc, ra[1]*ambacc);
783 if (pg != NULL) { /* cap gradient if necessary */
784 d = pg[0]*pg[0]*ra[0]*ra[0] + pg[1]*pg[1]*ra[1]*ra[1];
785 if (d > 1.0) {
786 d = 1.0/sqrt(d);
787 pg[0] *= d;
788 pg[1] *= d;
789 }
790 }
791 }
792 free(hp); /* clean up and return */
793 return(1);
794 }
795
796
797 #else /* ! NEWAMB */
798
799
800 void
801 inithemi( /* initialize sampling hemisphere */
802 AMBHEMI *hp,
803 COLOR ac,
804 RAY *r,
805 double wt
806 )
807 {
808 double d;
809 int i;
810 /* set number of divisions */
811 if (ambacc <= FTINY &&
812 wt > (d = 0.8*intens(ac)*r->rweight/(ambdiv*minweight)))
813 wt = d; /* avoid ray termination */
814 hp->nt = sqrt(ambdiv * wt / PI) + 0.5;
815 i = ambacc > FTINY ? 3 : 1; /* minimum number of samples */
816 if (hp->nt < i)
817 hp->nt = i;
818 hp->np = PI * hp->nt + 0.5;
819 /* set number of super-samples */
820 hp->ns = ambssamp * wt + 0.5;
821 /* assign coefficient */
822 copycolor(hp->acoef, ac);
823 d = 1.0/(hp->nt*hp->np);
824 scalecolor(hp->acoef, d);
825 /* make axes */
826 VCOPY(hp->uz, r->ron);
827 hp->uy[0] = hp->uy[1] = hp->uy[2] = 0.0;
828 for (i = 0; i < 3; i++)
829 if (hp->uz[i] < 0.6 && hp->uz[i] > -0.6)
830 break;
831 if (i >= 3)
832 error(CONSISTENCY, "bad ray direction in inithemi");
833 hp->uy[i] = 1.0;
834 fcross(hp->ux, hp->uy, hp->uz);
835 normalize(hp->ux);
836 fcross(hp->uy, hp->uz, hp->ux);
837 }
838
839
840 int
841 divsample( /* sample a division */
842 AMBSAMP *dp,
843 AMBHEMI *h,
844 RAY *r
845 )
846 {
847 RAY ar;
848 int hlist[3];
849 double spt[2];
850 double xd, yd, zd;
851 double b2;
852 double phi;
853 int i;
854 /* ambient coefficient for weight */
855 if (ambacc > FTINY)
856 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
857 else
858 copycolor(ar.rcoef, h->acoef);
859 if (rayorigin(&ar, AMBIENT, r, ar.rcoef) < 0)
860 return(-1);
861 if (ambacc > FTINY) {
862 multcolor(ar.rcoef, h->acoef);
863 scalecolor(ar.rcoef, 1./AVGREFL);
864 }
865 hlist[0] = r->rno;
866 hlist[1] = dp->t;
867 hlist[2] = dp->p;
868 multisamp(spt, 2, urand(ilhash(hlist,3)+dp->n));
869 zd = sqrt((dp->t + spt[0])/h->nt);
870 phi = 2.0*PI * (dp->p + spt[1])/h->np;
871 xd = tcos(phi) * zd;
872 yd = tsin(phi) * zd;
873 zd = sqrt(1.0 - zd*zd);
874 for (i = 0; i < 3; i++)
875 ar.rdir[i] = xd*h->ux[i] +
876 yd*h->uy[i] +
877 zd*h->uz[i];
878 checknorm(ar.rdir);
879 dimlist[ndims++] = dp->t*h->np + dp->p + 90171;
880 rayvalue(&ar);
881 ndims--;
882 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
883 addcolor(dp->v, ar.rcol);
884 /* use rt to improve gradient calc */
885 if (ar.rt > FTINY && ar.rt < FHUGE)
886 dp->r += 1.0/ar.rt;
887 /* (re)initialize error */
888 if (dp->n++) {
889 b2 = bright(dp->v)/dp->n - bright(ar.rcol);
890 b2 = b2*b2 + dp->k*((dp->n-1)*(dp->n-1));
891 dp->k = b2/(dp->n*dp->n);
892 } else
893 dp->k = 0.0;
894 return(0);
895 }
896
897
898 static int
899 ambcmp( /* decreasing order */
900 const void *p1,
901 const void *p2
902 )
903 {
904 const AMBSAMP *d1 = (const AMBSAMP *)p1;
905 const AMBSAMP *d2 = (const AMBSAMP *)p2;
906
907 if (d1->k < d2->k)
908 return(1);
909 if (d1->k > d2->k)
910 return(-1);
911 return(0);
912 }
913
914
915 static int
916 ambnorm( /* standard order */
917 const void *p1,
918 const void *p2
919 )
920 {
921 const AMBSAMP *d1 = (const AMBSAMP *)p1;
922 const AMBSAMP *d2 = (const AMBSAMP *)p2;
923 int c;
924
925 if ( (c = d1->t - d2->t) )
926 return(c);
927 return(d1->p - d2->p);
928 }
929
930
931 double
932 doambient( /* compute ambient component */
933 COLOR rcol,
934 RAY *r,
935 double wt,
936 FVECT pg,
937 FVECT dg
938 )
939 {
940 double b, d=0;
941 AMBHEMI hemi;
942 AMBSAMP *div;
943 AMBSAMP dnew;
944 double acol[3];
945 AMBSAMP *dp;
946 double arad;
947 int divcnt;
948 int i, j;
949 /* initialize hemisphere */
950 inithemi(&hemi, rcol, r, wt);
951 divcnt = hemi.nt * hemi.np;
952 /* initialize */
953 if (pg != NULL)
954 pg[0] = pg[1] = pg[2] = 0.0;
955 if (dg != NULL)
956 dg[0] = dg[1] = dg[2] = 0.0;
957 setcolor(rcol, 0.0, 0.0, 0.0);
958 if (divcnt == 0)
959 return(0.0);
960 /* allocate super-samples */
961 if (hemi.ns > 0 || pg != NULL || dg != NULL) {
962 div = (AMBSAMP *)malloc(divcnt*sizeof(AMBSAMP));
963 if (div == NULL)
964 error(SYSTEM, "out of memory in doambient");
965 } else
966 div = NULL;
967 /* sample the divisions */
968 arad = 0.0;
969 acol[0] = acol[1] = acol[2] = 0.0;
970 if ((dp = div) == NULL)
971 dp = &dnew;
972 divcnt = 0;
973 for (i = 0; i < hemi.nt; i++)
974 for (j = 0; j < hemi.np; j++) {
975 dp->t = i; dp->p = j;
976 setcolor(dp->v, 0.0, 0.0, 0.0);
977 dp->r = 0.0;
978 dp->n = 0;
979 if (divsample(dp, &hemi, r) < 0) {
980 if (div != NULL)
981 dp++;
982 continue;
983 }
984 arad += dp->r;
985 divcnt++;
986 if (div != NULL)
987 dp++;
988 else
989 addcolor(acol, dp->v);
990 }
991 if (!divcnt) {
992 if (div != NULL)
993 free((void *)div);
994 return(0.0); /* no samples taken */
995 }
996 if (divcnt < hemi.nt*hemi.np) {
997 pg = dg = NULL; /* incomplete sampling */
998 hemi.ns = 0;
999 } else if (arad > FTINY && divcnt/arad < minarad) {
1000 hemi.ns = 0; /* close enough */
1001 } else if (hemi.ns > 0) { /* else perform super-sampling? */
1002 comperrs(div, &hemi); /* compute errors */
1003 qsort(div, divcnt, sizeof(AMBSAMP), ambcmp); /* sort divs */
1004 /* super-sample */
1005 for (i = hemi.ns; i > 0; i--) {
1006 dnew = *div;
1007 if (divsample(&dnew, &hemi, r) < 0) {
1008 dp++;
1009 continue;
1010 }
1011 dp = div; /* reinsert */
1012 j = divcnt < i ? divcnt : i;
1013 while (--j > 0 && dnew.k < dp[1].k) {
1014 *dp = *(dp+1);
1015 dp++;
1016 }
1017 *dp = dnew;
1018 }
1019 if (pg != NULL || dg != NULL) /* restore order */
1020 qsort(div, divcnt, sizeof(AMBSAMP), ambnorm);
1021 }
1022 /* compute returned values */
1023 if (div != NULL) {
1024 arad = 0.0; /* note: divcnt may be < nt*np */
1025 for (i = hemi.nt*hemi.np, dp = div; i-- > 0; dp++) {
1026 arad += dp->r;
1027 if (dp->n > 1) {
1028 b = 1.0/dp->n;
1029 scalecolor(dp->v, b);
1030 dp->r *= b;
1031 dp->n = 1;
1032 }
1033 addcolor(acol, dp->v);
1034 }
1035 b = bright(acol);
1036 if (b > FTINY) {
1037 b = 1.0/b; /* compute & normalize gradient(s) */
1038 if (pg != NULL) {
1039 posgradient(pg, div, &hemi);
1040 for (i = 0; i < 3; i++)
1041 pg[i] *= b;
1042 }
1043 if (dg != NULL) {
1044 dirgradient(dg, div, &hemi);
1045 for (i = 0; i < 3; i++)
1046 dg[i] *= b;
1047 }
1048 }
1049 free((void *)div);
1050 }
1051 copycolor(rcol, acol);
1052 if (arad <= FTINY)
1053 arad = maxarad;
1054 else
1055 arad = (divcnt+hemi.ns)/arad;
1056 if (pg != NULL) { /* reduce radius if gradient large */
1057 d = DOT(pg,pg);
1058 if (d*arad*arad > 1.0)
1059 arad = 1.0/sqrt(d);
1060 }
1061 if (arad < minarad) {
1062 arad = minarad;
1063 if (pg != NULL && d*arad*arad > 1.0) { /* cap gradient */
1064 d = 1.0/arad/sqrt(d);
1065 for (i = 0; i < 3; i++)
1066 pg[i] *= d;
1067 }
1068 }
1069 if ((arad /= sqrt(wt)) > maxarad)
1070 arad = maxarad;
1071 return(arad);
1072 }
1073
1074
1075 void
1076 comperrs( /* compute initial error estimates */
1077 AMBSAMP *da, /* assumes standard ordering */
1078 AMBHEMI *hp
1079 )
1080 {
1081 double b, b2;
1082 int i, j;
1083 AMBSAMP *dp;
1084 /* sum differences from neighbors */
1085 dp = da;
1086 for (i = 0; i < hp->nt; i++)
1087 for (j = 0; j < hp->np; j++) {
1088 #ifdef DEBUG
1089 if (dp->t != i || dp->p != j)
1090 error(CONSISTENCY,
1091 "division order in comperrs");
1092 #endif
1093 b = bright(dp[0].v);
1094 if (i > 0) { /* from above */
1095 b2 = bright(dp[-hp->np].v) - b;
1096 b2 *= b2 * 0.25;
1097 dp[0].k += b2;
1098 dp[-hp->np].k += b2;
1099 }
1100 if (j > 0) { /* from behind */
1101 b2 = bright(dp[-1].v) - b;
1102 b2 *= b2 * 0.25;
1103 dp[0].k += b2;
1104 dp[-1].k += b2;
1105 } else { /* around */
1106 b2 = bright(dp[hp->np-1].v) - b;
1107 b2 *= b2 * 0.25;
1108 dp[0].k += b2;
1109 dp[hp->np-1].k += b2;
1110 }
1111 dp++;
1112 }
1113 /* divide by number of neighbors */
1114 dp = da;
1115 for (j = 0; j < hp->np; j++) /* top row */
1116 (dp++)->k *= 1.0/3.0;
1117 if (hp->nt < 2)
1118 return;
1119 for (i = 1; i < hp->nt-1; i++) /* central region */
1120 for (j = 0; j < hp->np; j++)
1121 (dp++)->k *= 0.25;
1122 for (j = 0; j < hp->np; j++) /* bottom row */
1123 (dp++)->k *= 1.0/3.0;
1124 }
1125
1126
1127 void
1128 posgradient( /* compute position gradient */
1129 FVECT gv,
1130 AMBSAMP *da, /* assumes standard ordering */
1131 AMBHEMI *hp
1132 )
1133 {
1134 int i, j;
1135 double nextsine, lastsine, b, d;
1136 double mag0, mag1;
1137 double phi, cosp, sinp, xd, yd;
1138 AMBSAMP *dp;
1139
1140 xd = yd = 0.0;
1141 for (j = 0; j < hp->np; j++) {
1142 dp = da + j;
1143 mag0 = mag1 = 0.0;
1144 lastsine = 0.0;
1145 for (i = 0; i < hp->nt; i++) {
1146 #ifdef DEBUG
1147 if (dp->t != i || dp->p != j)
1148 error(CONSISTENCY,
1149 "division order in posgradient");
1150 #endif
1151 b = bright(dp->v);
1152 if (i > 0) {
1153 d = dp[-hp->np].r;
1154 if (dp[0].r > d) d = dp[0].r;
1155 /* sin(t)*cos(t)^2 */
1156 d *= lastsine * (1.0 - (double)i/hp->nt);
1157 mag0 += d*(b - bright(dp[-hp->np].v));
1158 }
1159 nextsine = sqrt((double)(i+1)/hp->nt);
1160 if (j > 0) {
1161 d = dp[-1].r;
1162 if (dp[0].r > d) d = dp[0].r;
1163 mag1 += d * (nextsine - lastsine) *
1164 (b - bright(dp[-1].v));
1165 } else {
1166 d = dp[hp->np-1].r;
1167 if (dp[0].r > d) d = dp[0].r;
1168 mag1 += d * (nextsine - lastsine) *
1169 (b - bright(dp[hp->np-1].v));
1170 }
1171 dp += hp->np;
1172 lastsine = nextsine;
1173 }
1174 mag0 *= 2.0*PI / hp->np;
1175 phi = 2.0*PI * (double)j/hp->np;
1176 cosp = tcos(phi); sinp = tsin(phi);
1177 xd += mag0*cosp - mag1*sinp;
1178 yd += mag0*sinp + mag1*cosp;
1179 }
1180 for (i = 0; i < 3; i++)
1181 gv[i] = (xd*hp->ux[i] + yd*hp->uy[i])*(hp->nt*hp->np)/PI;
1182 }
1183
1184
1185 void
1186 dirgradient( /* compute direction gradient */
1187 FVECT gv,
1188 AMBSAMP *da, /* assumes standard ordering */
1189 AMBHEMI *hp
1190 )
1191 {
1192 int i, j;
1193 double mag;
1194 double phi, xd, yd;
1195 AMBSAMP *dp;
1196
1197 xd = yd = 0.0;
1198 for (j = 0; j < hp->np; j++) {
1199 dp = da + j;
1200 mag = 0.0;
1201 for (i = 0; i < hp->nt; i++) {
1202 #ifdef DEBUG
1203 if (dp->t != i || dp->p != j)
1204 error(CONSISTENCY,
1205 "division order in dirgradient");
1206 #endif
1207 /* tan(t) */
1208 mag += bright(dp->v)/sqrt(hp->nt/(i+.5) - 1.0);
1209 dp += hp->np;
1210 }
1211 phi = 2.0*PI * (j+.5)/hp->np + PI/2.0;
1212 xd += mag * tcos(phi);
1213 yd += mag * tsin(phi);
1214 }
1215 for (i = 0; i < 3; i++)
1216 gv[i] = xd*hp->ux[i] + yd*hp->uy[i];
1217 }
1218
1219 #endif /* ! NEWAMB */