ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambcomp.c
Revision: 2.75
Committed: Sat Oct 15 14:54:39 2016 UTC (7 years, 7 months ago) by greg
Content type: text/plain
Branch: MAIN
Changes since 2.74: +4 -4 lines
Log Message:
Increased minimum sampling spacing slightly -- rejection still less than 1%

File Contents

# Content
1 #ifndef lint
2 static const char RCSid[] = "$Id: ambcomp.c,v 2.74 2016/10/14 19:15:34 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
24 #ifndef OLDAMB
25
26 extern void SDsquare2disk(double ds[2], double seedx, double seedy);
27
28 typedef struct {
29 COLOR v; /* hemisphere sample value */
30 float d; /* reciprocal distance (1/rt) */
31 FVECT p; /* intersection point */
32 } AMBSAMP; /* sample value */
33
34 typedef struct {
35 RAY *rp; /* originating ray sample */
36 int ns; /* number of samples per axis */
37 int sampOK; /* acquired full sample set? */
38 COLOR acoef; /* division contribution coefficient */
39 double acol[3]; /* accumulated color */
40 FVECT ux, uy; /* tangent axis unit vectors */
41 AMBSAMP sa[1]; /* sample array (extends struct) */
42 } AMBHEMI; /* ambient sample hemisphere */
43
44 #define AI(h,i,j) ((i)*(h)->ns + (j))
45 #define ambsam(h,i,j) (h)->sa[AI(h,i,j)]
46
47 typedef struct {
48 FVECT r_i, r_i1, e_i, rcp, rI2_eJ2;
49 double I1, I2;
50 } FFTRI; /* vectors and coefficients for Hessian calculation */
51
52
53 static int
54 ambcollision( /* proposed direciton collides? */
55 AMBHEMI *hp,
56 int i,
57 int j,
58 FVECT dv
59 )
60 {
61 double cos_thresh;
62 int ii, jj;
63 /* min. spacing = 1/4th division */
64 cos_thresh = (PI/4.)/(double)hp->ns;
65 cos_thresh = 1. - .5*cos_thresh*cos_thresh;
66 /* check existing neighbors */
67 for (ii = i-1; ii <= i+1; ii++) {
68 if (ii < 0) continue;
69 if (ii >= hp->ns) break;
70 for (jj = j-1; jj <= j+1; jj++) {
71 AMBSAMP *ap;
72 FVECT avec;
73 double dprod;
74 if (jj < 0) continue;
75 if (jj >= hp->ns) break;
76 if ((ii==i) & (jj==j)) continue;
77 ap = &ambsam(hp,ii,jj);
78 if (ap->d <= .5/FHUGE)
79 continue; /* no one home */
80 VSUB(avec, ap->p, hp->rp->rop);
81 dprod = DOT(avec, dv);
82 if (dprod >= cos_thresh*VLEN(avec))
83 return(1); /* collision */
84 }
85 }
86 return(0); /* nothing to worry about */
87 }
88
89
90 static int
91 ambsample( /* initial ambient division sample */
92 AMBHEMI *hp,
93 int i,
94 int j,
95 int n
96 )
97 {
98 AMBSAMP *ap = &ambsam(hp,i,j);
99 RAY ar;
100 int hlist[3], ii;
101 double spt[2], zd;
102 /* generate hemispherical sample */
103 /* ambient coefficient for weight */
104 if (ambacc > FTINY)
105 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
106 else
107 copycolor(ar.rcoef, hp->acoef);
108 if (rayorigin(&ar, AMBIENT, hp->rp, ar.rcoef) < 0)
109 return(0);
110 if (ambacc > FTINY) {
111 multcolor(ar.rcoef, hp->acoef);
112 scalecolor(ar.rcoef, 1./AVGREFL);
113 }
114 hlist[0] = hp->rp->rno;
115 hlist[1] = j;
116 hlist[2] = i;
117 multisamp(spt, 2, urand(ilhash(hlist,3)+n));
118 resample:
119 SDsquare2disk(spt, (j+spt[1])/hp->ns, (i+spt[0])/hp->ns);
120 zd = sqrt(1. - spt[0]*spt[0] - spt[1]*spt[1]);
121 for (ii = 3; ii--; )
122 ar.rdir[ii] = spt[0]*hp->ux[ii] +
123 spt[1]*hp->uy[ii] +
124 zd*hp->rp->ron[ii];
125 checknorm(ar.rdir);
126 /* avoid coincident samples */
127 if (!n && ambcollision(hp, i, j, ar.rdir)) {
128 spt[0] = frandom(); spt[1] = frandom();
129 goto resample; /* reject this sample */
130 }
131 dimlist[ndims++] = AI(hp,i,j) + 90171;
132 rayvalue(&ar); /* evaluate ray */
133 ndims--;
134 if (ar.rt <= FTINY)
135 return(0); /* should never happen */
136 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
137 if (ar.rt*ap->d < 1.0) /* new/closer distance? */
138 ap->d = 1.0/ar.rt;
139 if (!n) { /* record first vertex & value */
140 if (ar.rt > 10.0*thescene.cusize)
141 ar.rt = 10.0*thescene.cusize;
142 VSUM(ap->p, ar.rorg, ar.rdir, ar.rt);
143 copycolor(ap->v, ar.rcol);
144 } else { /* else update recorded value */
145 hp->acol[RED] -= colval(ap->v,RED);
146 hp->acol[GRN] -= colval(ap->v,GRN);
147 hp->acol[BLU] -= colval(ap->v,BLU);
148 zd = 1.0/(double)(n+1);
149 scalecolor(ar.rcol, zd);
150 zd *= (double)n;
151 scalecolor(ap->v, zd);
152 addcolor(ap->v, ar.rcol);
153 }
154 addcolor(hp->acol, ap->v); /* add to our sum */
155 return(1);
156 }
157
158
159 /* Estimate errors based on ambient division differences */
160 static float *
161 getambdiffs(AMBHEMI *hp)
162 {
163 float *earr = (float *)calloc(hp->ns*hp->ns, sizeof(float));
164 float *ep;
165 AMBSAMP *ap;
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 (ap = hp->sa, ep = earr, i = 0; i < hp->ns; i++)
173 for (j = 0; j < hp->ns; j++, ap++, ep++) {
174 b = bright(ap[0].v);
175 if (i) { /* from above */
176 d2 = b - bright(ap[-hp->ns].v);
177 d2 *= d2;
178 ep[0] += d2;
179 ep[-hp->ns] += d2;
180 }
181 if (!j) continue;
182 /* from behind */
183 d2 = b - bright(ap[-1].v);
184 d2 *= d2;
185 ep[0] += d2;
186 ep[-1] += d2;
187 if (!i) continue;
188 /* diagonal */
189 d2 = b - bright(ap[-hp->ns-1].v);
190 d2 *= d2;
191 ep[0] += d2;
192 ep[-hp->ns-1] += d2;
193 }
194 /* correct for number of neighbors */
195 earr[0] *= 8./3.;
196 earr[hp->ns-1] *= 8./3.;
197 earr[(hp->ns-1)*hp->ns] *= 8./3.;
198 earr[(hp->ns-1)*hp->ns + hp->ns-1] *= 8./3.;
199 for (i = 1; i < hp->ns-1; i++) {
200 earr[i*hp->ns] *= 8./5.;
201 earr[i*hp->ns + hp->ns-1] *= 8./5.;
202 }
203 for (j = 1; j < hp->ns-1; j++) {
204 earr[j] *= 8./5.;
205 earr[(hp->ns-1)*hp->ns + j] *= 8./5.;
206 }
207 return(earr);
208 }
209
210
211 /* Perform super-sampling on hemisphere (introduces bias) */
212 static void
213 ambsupersamp(AMBHEMI *hp, int cnt)
214 {
215 float *earr = getambdiffs(hp);
216 double e2rem = 0;
217 AMBSAMP *ap;
218 float *ep;
219 int i, j, n, nss;
220
221 if (earr == NULL) /* just skip calc. if no memory */
222 return;
223 /* accumulate estimated variances */
224 for (ep = earr + hp->ns*hp->ns; ep > earr; )
225 e2rem += *--ep;
226 ep = earr; /* perform super-sampling */
227 for (ap = hp->sa, i = 0; i < hp->ns; i++)
228 for (j = 0; j < hp->ns; j++, ap++) {
229 if (e2rem <= FTINY)
230 goto done; /* nothing left to do */
231 nss = *ep/e2rem*cnt + frandom();
232 for (n = 1; n <= nss && ambsample(hp,i,j,n); n++)
233 --cnt;
234 e2rem -= *ep++; /* update remainder */
235 }
236 done:
237 free(earr);
238 }
239
240
241 static AMBHEMI *
242 samp_hemi( /* sample indirect hemisphere */
243 COLOR rcol,
244 RAY *r,
245 double wt
246 )
247 {
248 AMBHEMI *hp;
249 double d;
250 int n, i, j;
251 /* set number of divisions */
252 if (ambacc <= FTINY &&
253 wt > (d = 0.8*intens(rcol)*r->rweight/(ambdiv*minweight)))
254 wt = d; /* avoid ray termination */
255 n = sqrt(ambdiv * wt) + 0.5;
256 i = 1 + 5*(ambacc > FTINY); /* minimum number of samples */
257 if (n < i)
258 n = i;
259 /* allocate sampling array */
260 hp = (AMBHEMI *)malloc(sizeof(AMBHEMI) + sizeof(AMBSAMP)*(n*n - 1));
261 if (hp == NULL)
262 error(SYSTEM, "out of memory in samp_hemi");
263 hp->rp = r;
264 hp->ns = n;
265 hp->acol[RED] = hp->acol[GRN] = hp->acol[BLU] = 0.0;
266 memset(hp->sa, 0, sizeof(AMBSAMP)*n*n);
267 hp->sampOK = 0;
268 /* assign coefficient */
269 copycolor(hp->acoef, rcol);
270 d = 1.0/(n*n);
271 scalecolor(hp->acoef, d);
272 /* make tangent plane axes */
273 if (!getperpendicular(hp->ux, r->ron, 1))
274 error(CONSISTENCY, "bad ray direction in samp_hemi");
275 VCROSS(hp->uy, r->ron, hp->ux);
276 /* sample divisions */
277 for (i = hp->ns; i--; )
278 for (j = hp->ns; j--; )
279 hp->sampOK += ambsample(hp, i, j, 0);
280 copycolor(rcol, hp->acol);
281 if (!hp->sampOK) { /* utter failure? */
282 free(hp);
283 return(NULL);
284 }
285 if (hp->sampOK < hp->ns*hp->ns) {
286 hp->sampOK *= -1; /* soft failure */
287 return(hp);
288 }
289 n = ambssamp*wt + 0.5;
290 if (n > 8) { /* perform super-sampling? */
291 ambsupersamp(hp, n);
292 copycolor(rcol, hp->acol);
293 }
294 return(hp); /* all is well */
295 }
296
297
298 /* Return brightness of farthest ambient sample */
299 static double
300 back_ambval(AMBHEMI *hp, const int n1, const int n2, const int n3)
301 {
302 if (hp->sa[n1].d <= hp->sa[n2].d) {
303 if (hp->sa[n1].d <= hp->sa[n3].d)
304 return(colval(hp->sa[n1].v,CIEY));
305 return(colval(hp->sa[n3].v,CIEY));
306 }
307 if (hp->sa[n2].d <= hp->sa[n3].d)
308 return(colval(hp->sa[n2].v,CIEY));
309 return(colval(hp->sa[n3].v,CIEY));
310 }
311
312
313 /* Compute vectors and coefficients for Hessian/gradient calcs */
314 static void
315 comp_fftri(FFTRI *ftp, AMBHEMI *hp, const int n0, const int n1)
316 {
317 double rdot_cp, dot_e, dot_er, rdot_r, rdot_r1, J2;
318 int ii;
319
320 VSUB(ftp->r_i, hp->sa[n0].p, hp->rp->rop);
321 VSUB(ftp->r_i1, hp->sa[n1].p, hp->rp->rop);
322 VSUB(ftp->e_i, hp->sa[n1].p, hp->sa[n0].p);
323 VCROSS(ftp->rcp, ftp->r_i, ftp->r_i1);
324 rdot_cp = 1.0/DOT(ftp->rcp,ftp->rcp);
325 dot_e = DOT(ftp->e_i,ftp->e_i);
326 dot_er = DOT(ftp->e_i, ftp->r_i);
327 rdot_r = 1.0/DOT(ftp->r_i,ftp->r_i);
328 rdot_r1 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
329 ftp->I1 = acos( DOT(ftp->r_i, ftp->r_i1) * sqrt(rdot_r*rdot_r1) ) *
330 sqrt( rdot_cp );
331 ftp->I2 = ( DOT(ftp->e_i, ftp->r_i1)*rdot_r1 - dot_er*rdot_r +
332 dot_e*ftp->I1 )*0.5*rdot_cp;
333 J2 = ( 0.5*(rdot_r - rdot_r1) - dot_er*ftp->I2 ) / dot_e;
334 for (ii = 3; ii--; )
335 ftp->rI2_eJ2[ii] = ftp->I2*ftp->r_i[ii] + J2*ftp->e_i[ii];
336 }
337
338
339 /* Compose 3x3 matrix from two vectors */
340 static void
341 compose_matrix(FVECT mat[3], FVECT va, FVECT vb)
342 {
343 mat[0][0] = 2.0*va[0]*vb[0];
344 mat[1][1] = 2.0*va[1]*vb[1];
345 mat[2][2] = 2.0*va[2]*vb[2];
346 mat[0][1] = mat[1][0] = va[0]*vb[1] + va[1]*vb[0];
347 mat[0][2] = mat[2][0] = va[0]*vb[2] + va[2]*vb[0];
348 mat[1][2] = mat[2][1] = va[1]*vb[2] + va[2]*vb[1];
349 }
350
351
352 /* Compute partial 3x3 Hessian matrix for edge */
353 static void
354 comp_hessian(FVECT hess[3], FFTRI *ftp, FVECT nrm)
355 {
356 FVECT ncp;
357 FVECT m1[3], m2[3], m3[3], m4[3];
358 double d1, d2, d3, d4;
359 double I3, J3, K3;
360 int i, j;
361 /* compute intermediate coefficients */
362 d1 = 1.0/DOT(ftp->r_i,ftp->r_i);
363 d2 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
364 d3 = 1.0/DOT(ftp->e_i,ftp->e_i);
365 d4 = DOT(ftp->e_i, ftp->r_i);
366 I3 = ( DOT(ftp->e_i, ftp->r_i1)*d2*d2 - d4*d1*d1 + 3.0/d3*ftp->I2 )
367 / ( 4.0*DOT(ftp->rcp,ftp->rcp) );
368 J3 = 0.25*d3*(d1*d1 - d2*d2) - d4*d3*I3;
369 K3 = d3*(ftp->I2 - I3/d1 - 2.0*d4*J3);
370 /* intermediate matrices */
371 VCROSS(ncp, nrm, ftp->e_i);
372 compose_matrix(m1, ncp, ftp->rI2_eJ2);
373 compose_matrix(m2, ftp->r_i, ftp->r_i);
374 compose_matrix(m3, ftp->e_i, ftp->e_i);
375 compose_matrix(m4, ftp->r_i, ftp->e_i);
376 d1 = DOT(nrm, ftp->rcp);
377 d2 = -d1*ftp->I2;
378 d1 *= 2.0;
379 for (i = 3; i--; ) /* final matrix sum */
380 for (j = 3; j--; ) {
381 hess[i][j] = m1[i][j] + d1*( I3*m2[i][j] + K3*m3[i][j] +
382 2.0*J3*m4[i][j] );
383 hess[i][j] += d2*(i==j);
384 hess[i][j] *= -1.0/PI;
385 }
386 }
387
388
389 /* Reverse hessian calculation result for edge in other direction */
390 static void
391 rev_hessian(FVECT hess[3])
392 {
393 int i;
394
395 for (i = 3; i--; ) {
396 hess[i][0] = -hess[i][0];
397 hess[i][1] = -hess[i][1];
398 hess[i][2] = -hess[i][2];
399 }
400 }
401
402
403 /* Add to radiometric Hessian from the given triangle */
404 static void
405 add2hessian(FVECT hess[3], FVECT ehess1[3],
406 FVECT ehess2[3], FVECT ehess3[3], double v)
407 {
408 int i, j;
409
410 for (i = 3; i--; )
411 for (j = 3; j--; )
412 hess[i][j] += v*( ehess1[i][j] + ehess2[i][j] + ehess3[i][j] );
413 }
414
415
416 /* Compute partial displacement form factor gradient for edge */
417 static void
418 comp_gradient(FVECT grad, FFTRI *ftp, FVECT nrm)
419 {
420 FVECT ncp;
421 double f1;
422 int i;
423
424 f1 = 2.0*DOT(nrm, ftp->rcp);
425 VCROSS(ncp, nrm, ftp->e_i);
426 for (i = 3; i--; )
427 grad[i] = (0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
428 }
429
430
431 /* Reverse gradient calculation result for edge in other direction */
432 static void
433 rev_gradient(FVECT grad)
434 {
435 grad[0] = -grad[0];
436 grad[1] = -grad[1];
437 grad[2] = -grad[2];
438 }
439
440
441 /* Add to displacement gradient from the given triangle */
442 static void
443 add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, double v)
444 {
445 int i;
446
447 for (i = 3; i--; )
448 grad[i] += v*( egrad1[i] + egrad2[i] + egrad3[i] );
449 }
450
451
452 /* Compute anisotropic radii and eigenvector directions */
453 static void
454 eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3])
455 {
456 double hess2[2][2];
457 FVECT a, b;
458 double evalue[2], slope1, xmag1;
459 int i;
460 /* project Hessian to sample plane */
461 for (i = 3; i--; ) {
462 a[i] = DOT(hessian[i], uv[0]);
463 b[i] = DOT(hessian[i], uv[1]);
464 }
465 hess2[0][0] = DOT(uv[0], a);
466 hess2[0][1] = DOT(uv[0], b);
467 hess2[1][0] = DOT(uv[1], a);
468 hess2[1][1] = DOT(uv[1], b);
469 /* compute eigenvalue(s) */
470 i = quadratic(evalue, 1.0, -hess2[0][0]-hess2[1][1],
471 hess2[0][0]*hess2[1][1]-hess2[0][1]*hess2[1][0]);
472 if (i == 1) /* double-root (circle) */
473 evalue[1] = evalue[0];
474 if (!i || ((evalue[0] = fabs(evalue[0])) <= FTINY*FTINY) |
475 ((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) ) {
476 ra[0] = ra[1] = maxarad;
477 return;
478 }
479 if (evalue[0] > evalue[1]) {
480 ra[0] = sqrt(sqrt(4.0/evalue[0]));
481 ra[1] = sqrt(sqrt(4.0/evalue[1]));
482 slope1 = evalue[1];
483 } else {
484 ra[0] = sqrt(sqrt(4.0/evalue[1]));
485 ra[1] = sqrt(sqrt(4.0/evalue[0]));
486 slope1 = evalue[0];
487 }
488 /* compute unit eigenvectors */
489 if (fabs(hess2[0][1]) <= FTINY)
490 return; /* uv OK as is */
491 slope1 = (slope1 - hess2[0][0]) / hess2[0][1];
492 xmag1 = sqrt(1.0/(1.0 + slope1*slope1));
493 for (i = 3; i--; ) {
494 b[i] = xmag1*uv[0][i] + slope1*xmag1*uv[1][i];
495 a[i] = slope1*xmag1*uv[0][i] - xmag1*uv[1][i];
496 }
497 VCOPY(uv[0], a);
498 VCOPY(uv[1], b);
499 }
500
501
502 static void
503 ambHessian( /* anisotropic radii & pos. gradient */
504 AMBHEMI *hp,
505 FVECT uv[2], /* returned */
506 float ra[2], /* returned (optional) */
507 float pg[2] /* returned (optional) */
508 )
509 {
510 static char memerrmsg[] = "out of memory in ambHessian()";
511 FVECT (*hessrow)[3] = NULL;
512 FVECT *gradrow = NULL;
513 FVECT hessian[3];
514 FVECT gradient;
515 FFTRI fftr;
516 int i, j;
517 /* be sure to assign unit vectors */
518 VCOPY(uv[0], hp->ux);
519 VCOPY(uv[1], hp->uy);
520 /* clock-wise vertex traversal from sample POV */
521 if (ra != NULL) { /* initialize Hessian row buffer */
522 hessrow = (FVECT (*)[3])malloc(sizeof(FVECT)*3*(hp->ns-1));
523 if (hessrow == NULL)
524 error(SYSTEM, memerrmsg);
525 memset(hessian, 0, sizeof(hessian));
526 } else if (pg == NULL) /* bogus call? */
527 return;
528 if (pg != NULL) { /* initialize form factor row buffer */
529 gradrow = (FVECT *)malloc(sizeof(FVECT)*(hp->ns-1));
530 if (gradrow == NULL)
531 error(SYSTEM, memerrmsg);
532 memset(gradient, 0, sizeof(gradient));
533 }
534 /* compute first row of edges */
535 for (j = 0; j < hp->ns-1; j++) {
536 comp_fftri(&fftr, hp, AI(hp,0,j), AI(hp,0,j+1));
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 }
542 /* sum each row of triangles */
543 for (i = 0; i < hp->ns-1; i++) {
544 FVECT hesscol[3]; /* compute first vertical edge */
545 FVECT gradcol;
546 comp_fftri(&fftr, hp, AI(hp,i,0), AI(hp,i+1,0));
547 if (hessrow != NULL)
548 comp_hessian(hesscol, &fftr, hp->rp->ron);
549 if (gradrow != NULL)
550 comp_gradient(gradcol, &fftr, hp->rp->ron);
551 for (j = 0; j < hp->ns-1; j++) {
552 FVECT hessdia[3]; /* compute triangle contributions */
553 FVECT graddia;
554 double backg;
555 backg = back_ambval(hp, AI(hp,i,j),
556 AI(hp,i,j+1), AI(hp,i+1,j));
557 /* diagonal (inner) edge */
558 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j));
559 if (hessrow != NULL) {
560 comp_hessian(hessdia, &fftr, hp->rp->ron);
561 rev_hessian(hesscol);
562 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
563 }
564 if (gradrow != NULL) {
565 comp_gradient(graddia, &fftr, hp->rp->ron);
566 rev_gradient(gradcol);
567 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
568 }
569 /* initialize edge in next row */
570 comp_fftri(&fftr, hp, AI(hp,i+1,j+1), AI(hp,i+1,j));
571 if (hessrow != NULL)
572 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
573 if (gradrow != NULL)
574 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
575 /* new column edge & paired triangle */
576 backg = back_ambval(hp, AI(hp,i+1,j+1),
577 AI(hp,i+1,j), AI(hp,i,j+1));
578 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j+1));
579 if (hessrow != NULL) {
580 comp_hessian(hesscol, &fftr, hp->rp->ron);
581 rev_hessian(hessdia);
582 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
583 if (i < hp->ns-2)
584 rev_hessian(hessrow[j]);
585 }
586 if (gradrow != NULL) {
587 comp_gradient(gradcol, &fftr, hp->rp->ron);
588 rev_gradient(graddia);
589 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
590 if (i < hp->ns-2)
591 rev_gradient(gradrow[j]);
592 }
593 }
594 }
595 /* release row buffers */
596 if (hessrow != NULL) free(hessrow);
597 if (gradrow != NULL) free(gradrow);
598
599 if (ra != NULL) /* extract eigenvectors & radii */
600 eigenvectors(uv, ra, hessian);
601 if (pg != NULL) { /* tangential position gradient */
602 pg[0] = DOT(gradient, uv[0]);
603 pg[1] = DOT(gradient, uv[1]);
604 }
605 }
606
607
608 /* Compute direction gradient from a hemispherical sampling */
609 static void
610 ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
611 {
612 AMBSAMP *ap;
613 double dgsum[2];
614 int n;
615 FVECT vd;
616 double gfact;
617
618 dgsum[0] = dgsum[1] = 0.0; /* sum values times -tan(theta) */
619 for (ap = hp->sa, n = hp->ns*hp->ns; n--; ap++) {
620 /* use vector for azimuth + 90deg */
621 VSUB(vd, ap->p, hp->rp->rop);
622 /* brightness over cosine factor */
623 gfact = colval(ap->v,CIEY) / DOT(hp->rp->ron, vd);
624 /* sine = proj_radius/vd_length */
625 dgsum[0] -= DOT(uv[1], vd) * gfact;
626 dgsum[1] += DOT(uv[0], vd) * gfact;
627 }
628 dg[0] = dgsum[0] / (hp->ns*hp->ns);
629 dg[1] = dgsum[1] / (hp->ns*hp->ns);
630 }
631
632
633 /* Compute potential light leak direction flags for cache value */
634 static uint32
635 ambcorral(AMBHEMI *hp, FVECT uv[2], const double r0, const double r1)
636 {
637 const double max_d = 1.0/(minarad*ambacc + 0.001);
638 const double ang_res = 0.5*PI/hp->ns;
639 const double ang_step = ang_res/((int)(16/PI*ang_res) + 1.01);
640 double avg_d = 0;
641 uint32 flgs = 0;
642 FVECT vec;
643 double u, v;
644 double ang, a1;
645 int i, j;
646 /* don't bother for a few samples */
647 if (hp->ns < 8)
648 return(0);
649 /* check distances overhead */
650 for (i = hp->ns*3/4; i-- > hp->ns>>2; )
651 for (j = hp->ns*3/4; j-- > hp->ns>>2; )
652 avg_d += ambsam(hp,i,j).d;
653 avg_d *= 4.0/(hp->ns*hp->ns);
654 if (avg_d*r0 >= 1.0) /* ceiling too low for corral? */
655 return(0);
656 if (avg_d >= max_d) /* insurance */
657 return(0);
658 /* else circle around perimeter */
659 for (i = 0; i < hp->ns; i++)
660 for (j = 0; j < hp->ns; j += !i|(i==hp->ns-1) ? 1 : hp->ns-1) {
661 AMBSAMP *ap = &ambsam(hp,i,j);
662 if ((ap->d <= FTINY) | (ap->d >= max_d))
663 continue; /* too far or too near */
664 VSUB(vec, ap->p, hp->rp->rop);
665 u = DOT(vec, uv[0]);
666 v = DOT(vec, uv[1]);
667 if ((r0*r0*u*u + r1*r1*v*v) * ap->d*ap->d <= u*u + v*v)
668 continue; /* occluder outside ellipse */
669 ang = atan2a(v, u); /* else set direction flags */
670 for (a1 = ang-ang_res; a1 <= ang+ang_res; a1 += ang_step)
671 flgs |= 1L<<(int)(16/PI*(a1 + 2.*PI*(a1 < 0)));
672 }
673 /* add low-angle incident (< 20deg) */
674 if (fabs(hp->rp->rod) <= 0.342) {
675 u = -DOT(hp->rp->rdir, uv[0]);
676 v = -DOT(hp->rp->rdir, uv[1]);
677 if ((r0*r0*u*u + r1*r1*v*v) > hp->rp->rot*hp->rp->rot) {
678 ang = atan2a(v, u);
679 ang += 2.*PI*(ang < 0);
680 ang *= 16/PI;
681 if ((ang < .5) | (ang >= 31.5))
682 flgs |= 0x80000001;
683 else
684 flgs |= 3L<<(int)(ang-.5);
685 }
686 }
687 return(flgs);
688 }
689
690
691 int
692 doambient( /* compute ambient component */
693 COLOR rcol, /* input/output color */
694 RAY *r,
695 double wt,
696 FVECT uv[2], /* returned (optional) */
697 float ra[2], /* returned (optional) */
698 float pg[2], /* returned (optional) */
699 float dg[2], /* returned (optional) */
700 uint32 *crlp /* returned (optional) */
701 )
702 {
703 AMBHEMI *hp = samp_hemi(rcol, r, wt);
704 FVECT my_uv[2];
705 double d, K;
706 AMBSAMP *ap;
707 int i;
708 /* clear return values */
709 if (uv != NULL)
710 memset(uv, 0, sizeof(FVECT)*2);
711 if (ra != NULL)
712 ra[0] = ra[1] = 0.0;
713 if (pg != NULL)
714 pg[0] = pg[1] = 0.0;
715 if (dg != NULL)
716 dg[0] = dg[1] = 0.0;
717 if (crlp != NULL)
718 *crlp = 0;
719 if (hp == NULL) /* sampling falure? */
720 return(0);
721
722 if ((ra == NULL) & (pg == NULL) & (dg == NULL) ||
723 (hp->sampOK < 0) | (hp->ns < 6)) {
724 free(hp); /* Hessian not requested/possible */
725 return(-1); /* value-only return value */
726 }
727 if ((d = bright(rcol)) > FTINY) { /* normalize Y values */
728 d = 0.99*(hp->ns*hp->ns)/d;
729 K = 0.01;
730 } else { /* or fall back on geometric Hessian */
731 K = 1.0;
732 pg = NULL;
733 dg = NULL;
734 crlp = NULL;
735 }
736 ap = hp->sa; /* relative Y channel from here on... */
737 for (i = hp->ns*hp->ns; i--; ap++)
738 colval(ap->v,CIEY) = bright(ap->v)*d + K;
739
740 if (uv == NULL) /* make sure we have axis pointers */
741 uv = my_uv;
742 /* compute radii & pos. gradient */
743 ambHessian(hp, uv, ra, pg);
744
745 if (dg != NULL) /* compute direction gradient */
746 ambdirgrad(hp, uv, dg);
747
748 if (ra != NULL) { /* scale/clamp radii */
749 if (pg != NULL) {
750 if (ra[0]*(d = fabs(pg[0])) > 1.0)
751 ra[0] = 1.0/d;
752 if (ra[1]*(d = fabs(pg[1])) > 1.0)
753 ra[1] = 1.0/d;
754 if (ra[0] > ra[1])
755 ra[0] = ra[1];
756 }
757 if (ra[0] < minarad) {
758 ra[0] = minarad;
759 if (ra[1] < minarad)
760 ra[1] = minarad;
761 }
762 ra[0] *= d = 1.0/sqrt(wt);
763 if ((ra[1] *= d) > 2.0*ra[0])
764 ra[1] = 2.0*ra[0];
765 if (ra[1] > maxarad) {
766 ra[1] = maxarad;
767 if (ra[0] > maxarad)
768 ra[0] = maxarad;
769 }
770 /* flag encroached directions */
771 if (crlp != NULL)
772 *crlp = ambcorral(hp, uv, ra[0]*ambacc, ra[1]*ambacc);
773 if (pg != NULL) { /* cap gradient if necessary */
774 d = pg[0]*pg[0]*ra[0]*ra[0] + pg[1]*pg[1]*ra[1]*ra[1];
775 if (d > 1.0) {
776 d = 1.0/sqrt(d);
777 pg[0] *= d;
778 pg[1] *= d;
779 }
780 }
781 }
782 free(hp); /* clean up and return */
783 return(1);
784 }
785
786
787 #else /* ! NEWAMB */
788
789
790 void
791 inithemi( /* initialize sampling hemisphere */
792 AMBHEMI *hp,
793 COLOR ac,
794 RAY *r,
795 double wt
796 )
797 {
798 double d;
799 int i;
800 /* set number of divisions */
801 if (ambacc <= FTINY &&
802 wt > (d = 0.8*intens(ac)*r->rweight/(ambdiv*minweight)))
803 wt = d; /* avoid ray termination */
804 hp->nt = sqrt(ambdiv * wt / PI) + 0.5;
805 i = ambacc > FTINY ? 3 : 1; /* minimum number of samples */
806 if (hp->nt < i)
807 hp->nt = i;
808 hp->np = PI * hp->nt + 0.5;
809 /* set number of super-samples */
810 hp->ns = ambssamp * wt + 0.5;
811 /* assign coefficient */
812 copycolor(hp->acoef, ac);
813 d = 1.0/(hp->nt*hp->np);
814 scalecolor(hp->acoef, d);
815 /* make axes */
816 VCOPY(hp->uz, r->ron);
817 hp->uy[0] = hp->uy[1] = hp->uy[2] = 0.0;
818 for (i = 0; i < 3; i++)
819 if (hp->uz[i] < 0.6 && hp->uz[i] > -0.6)
820 break;
821 if (i >= 3)
822 error(CONSISTENCY, "bad ray direction in inithemi");
823 hp->uy[i] = 1.0;
824 fcross(hp->ux, hp->uy, hp->uz);
825 normalize(hp->ux);
826 fcross(hp->uy, hp->uz, hp->ux);
827 }
828
829
830 int
831 divsample( /* sample a division */
832 AMBSAMP *dp,
833 AMBHEMI *h,
834 RAY *r
835 )
836 {
837 RAY ar;
838 int hlist[3];
839 double spt[2];
840 double xd, yd, zd;
841 double b2;
842 double phi;
843 int i;
844 /* ambient coefficient for weight */
845 if (ambacc > FTINY)
846 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
847 else
848 copycolor(ar.rcoef, h->acoef);
849 if (rayorigin(&ar, AMBIENT, r, ar.rcoef) < 0)
850 return(-1);
851 if (ambacc > FTINY) {
852 multcolor(ar.rcoef, h->acoef);
853 scalecolor(ar.rcoef, 1./AVGREFL);
854 }
855 hlist[0] = r->rno;
856 hlist[1] = dp->t;
857 hlist[2] = dp->p;
858 multisamp(spt, 2, urand(ilhash(hlist,3)+dp->n));
859 zd = sqrt((dp->t + spt[0])/h->nt);
860 phi = 2.0*PI * (dp->p + spt[1])/h->np;
861 xd = tcos(phi) * zd;
862 yd = tsin(phi) * zd;
863 zd = sqrt(1.0 - zd*zd);
864 for (i = 0; i < 3; i++)
865 ar.rdir[i] = xd*h->ux[i] +
866 yd*h->uy[i] +
867 zd*h->uz[i];
868 checknorm(ar.rdir);
869 dimlist[ndims++] = dp->t*h->np + dp->p + 90171;
870 rayvalue(&ar);
871 ndims--;
872 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
873 addcolor(dp->v, ar.rcol);
874 /* use rt to improve gradient calc */
875 if (ar.rt > FTINY && ar.rt < FHUGE)
876 dp->r += 1.0/ar.rt;
877 /* (re)initialize error */
878 if (dp->n++) {
879 b2 = bright(dp->v)/dp->n - bright(ar.rcol);
880 b2 = b2*b2 + dp->k*((dp->n-1)*(dp->n-1));
881 dp->k = b2/(dp->n*dp->n);
882 } else
883 dp->k = 0.0;
884 return(0);
885 }
886
887
888 static int
889 ambcmp( /* decreasing order */
890 const void *p1,
891 const void *p2
892 )
893 {
894 const AMBSAMP *d1 = (const AMBSAMP *)p1;
895 const AMBSAMP *d2 = (const AMBSAMP *)p2;
896
897 if (d1->k < d2->k)
898 return(1);
899 if (d1->k > d2->k)
900 return(-1);
901 return(0);
902 }
903
904
905 static int
906 ambnorm( /* standard order */
907 const void *p1,
908 const void *p2
909 )
910 {
911 const AMBSAMP *d1 = (const AMBSAMP *)p1;
912 const AMBSAMP *d2 = (const AMBSAMP *)p2;
913 int c;
914
915 if ( (c = d1->t - d2->t) )
916 return(c);
917 return(d1->p - d2->p);
918 }
919
920
921 double
922 doambient( /* compute ambient component */
923 COLOR rcol,
924 RAY *r,
925 double wt,
926 FVECT pg,
927 FVECT dg
928 )
929 {
930 double b, d=0;
931 AMBHEMI hemi;
932 AMBSAMP *div;
933 AMBSAMP dnew;
934 double acol[3];
935 AMBSAMP *dp;
936 double arad;
937 int divcnt;
938 int i, j;
939 /* initialize hemisphere */
940 inithemi(&hemi, rcol, r, wt);
941 divcnt = hemi.nt * hemi.np;
942 /* initialize */
943 if (pg != NULL)
944 pg[0] = pg[1] = pg[2] = 0.0;
945 if (dg != NULL)
946 dg[0] = dg[1] = dg[2] = 0.0;
947 setcolor(rcol, 0.0, 0.0, 0.0);
948 if (divcnt == 0)
949 return(0.0);
950 /* allocate super-samples */
951 if (hemi.ns > 0 || pg != NULL || dg != NULL) {
952 div = (AMBSAMP *)malloc(divcnt*sizeof(AMBSAMP));
953 if (div == NULL)
954 error(SYSTEM, "out of memory in doambient");
955 } else
956 div = NULL;
957 /* sample the divisions */
958 arad = 0.0;
959 acol[0] = acol[1] = acol[2] = 0.0;
960 if ((dp = div) == NULL)
961 dp = &dnew;
962 divcnt = 0;
963 for (i = 0; i < hemi.nt; i++)
964 for (j = 0; j < hemi.np; j++) {
965 dp->t = i; dp->p = j;
966 setcolor(dp->v, 0.0, 0.0, 0.0);
967 dp->r = 0.0;
968 dp->n = 0;
969 if (divsample(dp, &hemi, r) < 0) {
970 if (div != NULL)
971 dp++;
972 continue;
973 }
974 arad += dp->r;
975 divcnt++;
976 if (div != NULL)
977 dp++;
978 else
979 addcolor(acol, dp->v);
980 }
981 if (!divcnt) {
982 if (div != NULL)
983 free((void *)div);
984 return(0.0); /* no samples taken */
985 }
986 if (divcnt < hemi.nt*hemi.np) {
987 pg = dg = NULL; /* incomplete sampling */
988 hemi.ns = 0;
989 } else if (arad > FTINY && divcnt/arad < minarad) {
990 hemi.ns = 0; /* close enough */
991 } else if (hemi.ns > 0) { /* else perform super-sampling? */
992 comperrs(div, &hemi); /* compute errors */
993 qsort(div, divcnt, sizeof(AMBSAMP), ambcmp); /* sort divs */
994 /* super-sample */
995 for (i = hemi.ns; i > 0; i--) {
996 dnew = *div;
997 if (divsample(&dnew, &hemi, r) < 0) {
998 dp++;
999 continue;
1000 }
1001 dp = div; /* reinsert */
1002 j = divcnt < i ? divcnt : i;
1003 while (--j > 0 && dnew.k < dp[1].k) {
1004 *dp = *(dp+1);
1005 dp++;
1006 }
1007 *dp = dnew;
1008 }
1009 if (pg != NULL || dg != NULL) /* restore order */
1010 qsort(div, divcnt, sizeof(AMBSAMP), ambnorm);
1011 }
1012 /* compute returned values */
1013 if (div != NULL) {
1014 arad = 0.0; /* note: divcnt may be < nt*np */
1015 for (i = hemi.nt*hemi.np, dp = div; i-- > 0; dp++) {
1016 arad += dp->r;
1017 if (dp->n > 1) {
1018 b = 1.0/dp->n;
1019 scalecolor(dp->v, b);
1020 dp->r *= b;
1021 dp->n = 1;
1022 }
1023 addcolor(acol, dp->v);
1024 }
1025 b = bright(acol);
1026 if (b > FTINY) {
1027 b = 1.0/b; /* compute & normalize gradient(s) */
1028 if (pg != NULL) {
1029 posgradient(pg, div, &hemi);
1030 for (i = 0; i < 3; i++)
1031 pg[i] *= b;
1032 }
1033 if (dg != NULL) {
1034 dirgradient(dg, div, &hemi);
1035 for (i = 0; i < 3; i++)
1036 dg[i] *= b;
1037 }
1038 }
1039 free((void *)div);
1040 }
1041 copycolor(rcol, acol);
1042 if (arad <= FTINY)
1043 arad = maxarad;
1044 else
1045 arad = (divcnt+hemi.ns)/arad;
1046 if (pg != NULL) { /* reduce radius if gradient large */
1047 d = DOT(pg,pg);
1048 if (d*arad*arad > 1.0)
1049 arad = 1.0/sqrt(d);
1050 }
1051 if (arad < minarad) {
1052 arad = minarad;
1053 if (pg != NULL && d*arad*arad > 1.0) { /* cap gradient */
1054 d = 1.0/arad/sqrt(d);
1055 for (i = 0; i < 3; i++)
1056 pg[i] *= d;
1057 }
1058 }
1059 if ((arad /= sqrt(wt)) > maxarad)
1060 arad = maxarad;
1061 return(arad);
1062 }
1063
1064
1065 void
1066 comperrs( /* compute initial error estimates */
1067 AMBSAMP *da, /* assumes standard ordering */
1068 AMBHEMI *hp
1069 )
1070 {
1071 double b, b2;
1072 int i, j;
1073 AMBSAMP *dp;
1074 /* sum differences from neighbors */
1075 dp = da;
1076 for (i = 0; i < hp->nt; i++)
1077 for (j = 0; j < hp->np; j++) {
1078 #ifdef DEBUG
1079 if (dp->t != i || dp->p != j)
1080 error(CONSISTENCY,
1081 "division order in comperrs");
1082 #endif
1083 b = bright(dp[0].v);
1084 if (i > 0) { /* from above */
1085 b2 = bright(dp[-hp->np].v) - b;
1086 b2 *= b2 * 0.25;
1087 dp[0].k += b2;
1088 dp[-hp->np].k += b2;
1089 }
1090 if (j > 0) { /* from behind */
1091 b2 = bright(dp[-1].v) - b;
1092 b2 *= b2 * 0.25;
1093 dp[0].k += b2;
1094 dp[-1].k += b2;
1095 } else { /* around */
1096 b2 = bright(dp[hp->np-1].v) - b;
1097 b2 *= b2 * 0.25;
1098 dp[0].k += b2;
1099 dp[hp->np-1].k += b2;
1100 }
1101 dp++;
1102 }
1103 /* divide by number of neighbors */
1104 dp = da;
1105 for (j = 0; j < hp->np; j++) /* top row */
1106 (dp++)->k *= 1.0/3.0;
1107 if (hp->nt < 2)
1108 return;
1109 for (i = 1; i < hp->nt-1; i++) /* central region */
1110 for (j = 0; j < hp->np; j++)
1111 (dp++)->k *= 0.25;
1112 for (j = 0; j < hp->np; j++) /* bottom row */
1113 (dp++)->k *= 1.0/3.0;
1114 }
1115
1116
1117 void
1118 posgradient( /* compute position gradient */
1119 FVECT gv,
1120 AMBSAMP *da, /* assumes standard ordering */
1121 AMBHEMI *hp
1122 )
1123 {
1124 int i, j;
1125 double nextsine, lastsine, b, d;
1126 double mag0, mag1;
1127 double phi, cosp, sinp, xd, yd;
1128 AMBSAMP *dp;
1129
1130 xd = yd = 0.0;
1131 for (j = 0; j < hp->np; j++) {
1132 dp = da + j;
1133 mag0 = mag1 = 0.0;
1134 lastsine = 0.0;
1135 for (i = 0; i < hp->nt; i++) {
1136 #ifdef DEBUG
1137 if (dp->t != i || dp->p != j)
1138 error(CONSISTENCY,
1139 "division order in posgradient");
1140 #endif
1141 b = bright(dp->v);
1142 if (i > 0) {
1143 d = dp[-hp->np].r;
1144 if (dp[0].r > d) d = dp[0].r;
1145 /* sin(t)*cos(t)^2 */
1146 d *= lastsine * (1.0 - (double)i/hp->nt);
1147 mag0 += d*(b - bright(dp[-hp->np].v));
1148 }
1149 nextsine = sqrt((double)(i+1)/hp->nt);
1150 if (j > 0) {
1151 d = dp[-1].r;
1152 if (dp[0].r > d) d = dp[0].r;
1153 mag1 += d * (nextsine - lastsine) *
1154 (b - bright(dp[-1].v));
1155 } else {
1156 d = dp[hp->np-1].r;
1157 if (dp[0].r > d) d = dp[0].r;
1158 mag1 += d * (nextsine - lastsine) *
1159 (b - bright(dp[hp->np-1].v));
1160 }
1161 dp += hp->np;
1162 lastsine = nextsine;
1163 }
1164 mag0 *= 2.0*PI / hp->np;
1165 phi = 2.0*PI * (double)j/hp->np;
1166 cosp = tcos(phi); sinp = tsin(phi);
1167 xd += mag0*cosp - mag1*sinp;
1168 yd += mag0*sinp + mag1*cosp;
1169 }
1170 for (i = 0; i < 3; i++)
1171 gv[i] = (xd*hp->ux[i] + yd*hp->uy[i])*(hp->nt*hp->np)/PI;
1172 }
1173
1174
1175 void
1176 dirgradient( /* compute direction gradient */
1177 FVECT gv,
1178 AMBSAMP *da, /* assumes standard ordering */
1179 AMBHEMI *hp
1180 )
1181 {
1182 int i, j;
1183 double mag;
1184 double phi, xd, yd;
1185 AMBSAMP *dp;
1186
1187 xd = yd = 0.0;
1188 for (j = 0; j < hp->np; j++) {
1189 dp = da + j;
1190 mag = 0.0;
1191 for (i = 0; i < hp->nt; i++) {
1192 #ifdef DEBUG
1193 if (dp->t != i || dp->p != j)
1194 error(CONSISTENCY,
1195 "division order in dirgradient");
1196 #endif
1197 /* tan(t) */
1198 mag += bright(dp->v)/sqrt(hp->nt/(i+.5) - 1.0);
1199 dp += hp->np;
1200 }
1201 phi = 2.0*PI * (j+.5)/hp->np + PI/2.0;
1202 xd += mag * tcos(phi);
1203 yd += mag * tsin(phi);
1204 }
1205 for (i = 0; i < 3; i++)
1206 gv[i] = xd*hp->ux[i] + yd*hp->uy[i];
1207 }
1208
1209 #endif /* ! NEWAMB */