ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambcomp.c
Revision: 2.82
Committed: Thu Apr 12 20:07:09 2018 UTC (6 years, 1 month ago) by greg
Content type: text/plain
Branch: MAIN
CVS Tags: rad5R2
Changes since 2.81: +8 -8 lines
Log Message:
Further refinement on variance estimate

File Contents

# Content
1 #ifndef lint
2 static const char RCSid[] = "$Id: ambcomp.c,v 2.81 2018/04/12 18:02:45 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 + 1000.)
141 ar.rt = 10.0*thescene.cusize + 1000.;
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 variance based on ambient division differences */
160 static float *
161 getambdiffs(AMBHEMI *hp)
162 {
163 const double normf = 1./bright(hp->acoef);
164 float *earr = (float *)calloc(hp->ns*hp->ns, sizeof(float));
165 float *ep;
166 AMBSAMP *ap;
167 double b, b1, d2;
168 int i, j;
169
170 if (earr == NULL) /* out of memory? */
171 return(NULL);
172 /* sum squared neighbor diffs */
173 for (ap = hp->sa, ep = earr, i = 0; i < hp->ns; i++)
174 for (j = 0; j < hp->ns; j++, ap++, ep++) {
175 b = bright(ap[0].v);
176 if (i) { /* from above */
177 b1 = bright(ap[-hp->ns].v);
178 d2 = b - b1;
179 d2 *= d2*normf/(b + b1);
180 ep[0] += d2;
181 ep[-hp->ns] += d2;
182 }
183 if (!j) continue;
184 /* from behind */
185 b1 = bright(ap[-1].v);
186 d2 = b - b1;
187 d2 *= d2*normf/(b + b1);
188 ep[0] += d2;
189 ep[-1] += d2;
190 if (!i) continue;
191 /* diagonal */
192 b1 = bright(ap[-hp->ns-1].v);
193 d2 = b - b1;
194 d2 *= d2*normf/(b + b1);
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 float *ep;
222 int i, j, n, nss;
223
224 if (earr == NULL) /* just skip calc. if no memory */
225 return;
226 /* accumulate estimated variances */
227 for (ep = earr + hp->ns*hp->ns; ep > earr; )
228 e2rem += *--ep;
229 ep = earr; /* perform super-sampling */
230 for (i = 0; i < hp->ns; i++)
231 for (j = 0; j < hp->ns; j++) {
232 if (e2rem <= FTINY)
233 goto done; /* nothing left to do */
234 nss = *ep/e2rem*cnt + frandom();
235 for (n = 1; n <= nss && ambsample(hp,i,j,n); n++)
236 if (!--cnt) goto done;
237 e2rem -= *ep++; /* update remainder */
238 }
239 done:
240 free(earr);
241 }
242
243
244 static AMBHEMI *
245 samp_hemi( /* sample indirect hemisphere */
246 COLOR rcol,
247 RAY *r,
248 double wt
249 )
250 {
251 AMBHEMI *hp;
252 double d;
253 int n, i, j;
254 /* insignificance check */
255 if (bright(rcol) <= FTINY)
256 return(NULL);
257 /* set number of divisions */
258 if (ambacc <= FTINY &&
259 wt > (d = 0.8*intens(rcol)*r->rweight/(ambdiv*minweight)))
260 wt = d; /* avoid ray termination */
261 n = sqrt(ambdiv * wt) + 0.5;
262 i = 1 + 5*(ambacc > FTINY); /* minimum number of samples */
263 if (n < i)
264 n = i;
265 /* allocate sampling array */
266 hp = (AMBHEMI *)malloc(sizeof(AMBHEMI) + sizeof(AMBSAMP)*(n*n - 1));
267 if (hp == NULL)
268 error(SYSTEM, "out of memory in samp_hemi");
269 hp->rp = r;
270 hp->ns = n;
271 hp->acol[RED] = hp->acol[GRN] = hp->acol[BLU] = 0.0;
272 memset(hp->sa, 0, sizeof(AMBSAMP)*n*n);
273 hp->sampOK = 0;
274 /* assign coefficient */
275 copycolor(hp->acoef, rcol);
276 d = 1.0/(n*n);
277 scalecolor(hp->acoef, d);
278 /* make tangent plane axes */
279 if (!getperpendicular(hp->ux, r->ron, 1))
280 error(CONSISTENCY, "bad ray direction in samp_hemi");
281 VCROSS(hp->uy, r->ron, hp->ux);
282 /* sample divisions */
283 for (i = hp->ns; i--; )
284 for (j = hp->ns; j--; )
285 hp->sampOK += ambsample(hp, i, j, 0);
286 copycolor(rcol, hp->acol);
287 if (!hp->sampOK) { /* utter failure? */
288 free(hp);
289 return(NULL);
290 }
291 if (hp->sampOK < hp->ns*hp->ns) {
292 hp->sampOK *= -1; /* soft failure */
293 return(hp);
294 }
295 n = ambssamp*wt + 0.5;
296 if (n > 8) { /* perform super-sampling? */
297 ambsupersamp(hp, n);
298 copycolor(rcol, hp->acol);
299 }
300 return(hp); /* all is well */
301 }
302
303
304 /* Return brightness of farthest ambient sample */
305 static double
306 back_ambval(AMBHEMI *hp, const int n1, const int n2, const int n3)
307 {
308 if (hp->sa[n1].d <= hp->sa[n2].d) {
309 if (hp->sa[n1].d <= hp->sa[n3].d)
310 return(colval(hp->sa[n1].v,CIEY));
311 return(colval(hp->sa[n3].v,CIEY));
312 }
313 if (hp->sa[n2].d <= hp->sa[n3].d)
314 return(colval(hp->sa[n2].v,CIEY));
315 return(colval(hp->sa[n3].v,CIEY));
316 }
317
318
319 /* Compute vectors and coefficients for Hessian/gradient calcs */
320 static void
321 comp_fftri(FFTRI *ftp, AMBHEMI *hp, const int n0, const int n1)
322 {
323 double rdot_cp, dot_e, dot_er, rdot_r, rdot_r1, J2;
324 int ii;
325
326 VSUB(ftp->r_i, hp->sa[n0].p, hp->rp->rop);
327 VSUB(ftp->r_i1, hp->sa[n1].p, hp->rp->rop);
328 VSUB(ftp->e_i, hp->sa[n1].p, hp->sa[n0].p);
329 VCROSS(ftp->rcp, ftp->r_i, ftp->r_i1);
330 rdot_cp = 1.0/DOT(ftp->rcp,ftp->rcp);
331 dot_e = DOT(ftp->e_i,ftp->e_i);
332 dot_er = DOT(ftp->e_i, ftp->r_i);
333 rdot_r = 1.0/DOT(ftp->r_i,ftp->r_i);
334 rdot_r1 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
335 ftp->I1 = acos( DOT(ftp->r_i, ftp->r_i1) * sqrt(rdot_r*rdot_r1) ) *
336 sqrt( rdot_cp );
337 ftp->I2 = ( DOT(ftp->e_i, ftp->r_i1)*rdot_r1 - dot_er*rdot_r +
338 dot_e*ftp->I1 )*0.5*rdot_cp;
339 J2 = ( 0.5*(rdot_r - rdot_r1) - dot_er*ftp->I2 ) / dot_e;
340 for (ii = 3; ii--; )
341 ftp->rI2_eJ2[ii] = ftp->I2*ftp->r_i[ii] + J2*ftp->e_i[ii];
342 }
343
344
345 /* Compose 3x3 matrix from two vectors */
346 static void
347 compose_matrix(FVECT mat[3], FVECT va, FVECT vb)
348 {
349 mat[0][0] = 2.0*va[0]*vb[0];
350 mat[1][1] = 2.0*va[1]*vb[1];
351 mat[2][2] = 2.0*va[2]*vb[2];
352 mat[0][1] = mat[1][0] = va[0]*vb[1] + va[1]*vb[0];
353 mat[0][2] = mat[2][0] = va[0]*vb[2] + va[2]*vb[0];
354 mat[1][2] = mat[2][1] = va[1]*vb[2] + va[2]*vb[1];
355 }
356
357
358 /* Compute partial 3x3 Hessian matrix for edge */
359 static void
360 comp_hessian(FVECT hess[3], FFTRI *ftp, FVECT nrm)
361 {
362 FVECT ncp;
363 FVECT m1[3], m2[3], m3[3], m4[3];
364 double d1, d2, d3, d4;
365 double I3, J3, K3;
366 int i, j;
367 /* compute intermediate coefficients */
368 d1 = 1.0/DOT(ftp->r_i,ftp->r_i);
369 d2 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
370 d3 = 1.0/DOT(ftp->e_i,ftp->e_i);
371 d4 = DOT(ftp->e_i, ftp->r_i);
372 I3 = ( DOT(ftp->e_i, ftp->r_i1)*d2*d2 - d4*d1*d1 + 3.0/d3*ftp->I2 )
373 / ( 4.0*DOT(ftp->rcp,ftp->rcp) );
374 J3 = 0.25*d3*(d1*d1 - d2*d2) - d4*d3*I3;
375 K3 = d3*(ftp->I2 - I3/d1 - 2.0*d4*J3);
376 /* intermediate matrices */
377 VCROSS(ncp, nrm, ftp->e_i);
378 compose_matrix(m1, ncp, ftp->rI2_eJ2);
379 compose_matrix(m2, ftp->r_i, ftp->r_i);
380 compose_matrix(m3, ftp->e_i, ftp->e_i);
381 compose_matrix(m4, ftp->r_i, ftp->e_i);
382 d1 = DOT(nrm, ftp->rcp);
383 d2 = -d1*ftp->I2;
384 d1 *= 2.0;
385 for (i = 3; i--; ) /* final matrix sum */
386 for (j = 3; j--; ) {
387 hess[i][j] = m1[i][j] + d1*( I3*m2[i][j] + K3*m3[i][j] +
388 2.0*J3*m4[i][j] );
389 hess[i][j] += d2*(i==j);
390 hess[i][j] *= -1.0/PI;
391 }
392 }
393
394
395 /* Reverse hessian calculation result for edge in other direction */
396 static void
397 rev_hessian(FVECT hess[3])
398 {
399 int i;
400
401 for (i = 3; i--; ) {
402 hess[i][0] = -hess[i][0];
403 hess[i][1] = -hess[i][1];
404 hess[i][2] = -hess[i][2];
405 }
406 }
407
408
409 /* Add to radiometric Hessian from the given triangle */
410 static void
411 add2hessian(FVECT hess[3], FVECT ehess1[3],
412 FVECT ehess2[3], FVECT ehess3[3], double v)
413 {
414 int i, j;
415
416 for (i = 3; i--; )
417 for (j = 3; j--; )
418 hess[i][j] += v*( ehess1[i][j] + ehess2[i][j] + ehess3[i][j] );
419 }
420
421
422 /* Compute partial displacement form factor gradient for edge */
423 static void
424 comp_gradient(FVECT grad, FFTRI *ftp, FVECT nrm)
425 {
426 FVECT ncp;
427 double f1;
428 int i;
429
430 f1 = 2.0*DOT(nrm, ftp->rcp);
431 VCROSS(ncp, nrm, ftp->e_i);
432 for (i = 3; i--; )
433 grad[i] = (0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
434 }
435
436
437 /* Reverse gradient calculation result for edge in other direction */
438 static void
439 rev_gradient(FVECT grad)
440 {
441 grad[0] = -grad[0];
442 grad[1] = -grad[1];
443 grad[2] = -grad[2];
444 }
445
446
447 /* Add to displacement gradient from the given triangle */
448 static void
449 add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, double v)
450 {
451 int i;
452
453 for (i = 3; i--; )
454 grad[i] += v*( egrad1[i] + egrad2[i] + egrad3[i] );
455 }
456
457
458 /* Compute anisotropic radii and eigenvector directions */
459 static void
460 eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3])
461 {
462 double hess2[2][2];
463 FVECT a, b;
464 double evalue[2], slope1, xmag1;
465 int i;
466 /* project Hessian to sample plane */
467 for (i = 3; i--; ) {
468 a[i] = DOT(hessian[i], uv[0]);
469 b[i] = DOT(hessian[i], uv[1]);
470 }
471 hess2[0][0] = DOT(uv[0], a);
472 hess2[0][1] = DOT(uv[0], b);
473 hess2[1][0] = DOT(uv[1], a);
474 hess2[1][1] = DOT(uv[1], b);
475 /* compute eigenvalue(s) */
476 i = quadratic(evalue, 1.0, -hess2[0][0]-hess2[1][1],
477 hess2[0][0]*hess2[1][1]-hess2[0][1]*hess2[1][0]);
478 if (i == 1) /* double-root (circle) */
479 evalue[1] = evalue[0];
480 if (!i || ((evalue[0] = fabs(evalue[0])) <= FTINY*FTINY) |
481 ((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) ) {
482 ra[0] = ra[1] = maxarad;
483 return;
484 }
485 if (evalue[0] > evalue[1]) {
486 ra[0] = sqrt(sqrt(4.0/evalue[0]));
487 ra[1] = sqrt(sqrt(4.0/evalue[1]));
488 slope1 = evalue[1];
489 } else {
490 ra[0] = sqrt(sqrt(4.0/evalue[1]));
491 ra[1] = sqrt(sqrt(4.0/evalue[0]));
492 slope1 = evalue[0];
493 }
494 /* compute unit eigenvectors */
495 if (fabs(hess2[0][1]) <= FTINY)
496 return; /* uv OK as is */
497 slope1 = (slope1 - hess2[0][0]) / hess2[0][1];
498 xmag1 = sqrt(1.0/(1.0 + slope1*slope1));
499 for (i = 3; i--; ) {
500 b[i] = xmag1*uv[0][i] + slope1*xmag1*uv[1][i];
501 a[i] = slope1*xmag1*uv[0][i] - xmag1*uv[1][i];
502 }
503 VCOPY(uv[0], a);
504 VCOPY(uv[1], b);
505 }
506
507
508 static void
509 ambHessian( /* anisotropic radii & pos. gradient */
510 AMBHEMI *hp,
511 FVECT uv[2], /* returned */
512 float ra[2], /* returned (optional) */
513 float pg[2] /* returned (optional) */
514 )
515 {
516 static char memerrmsg[] = "out of memory in ambHessian()";
517 FVECT (*hessrow)[3] = NULL;
518 FVECT *gradrow = NULL;
519 FVECT hessian[3];
520 FVECT gradient;
521 FFTRI fftr;
522 int i, j;
523 /* be sure to assign unit vectors */
524 VCOPY(uv[0], hp->ux);
525 VCOPY(uv[1], hp->uy);
526 /* clock-wise vertex traversal from sample POV */
527 if (ra != NULL) { /* initialize Hessian row buffer */
528 hessrow = (FVECT (*)[3])malloc(sizeof(FVECT)*3*(hp->ns-1));
529 if (hessrow == NULL)
530 error(SYSTEM, memerrmsg);
531 memset(hessian, 0, sizeof(hessian));
532 } else if (pg == NULL) /* bogus call? */
533 return;
534 if (pg != NULL) { /* initialize form factor row buffer */
535 gradrow = (FVECT *)malloc(sizeof(FVECT)*(hp->ns-1));
536 if (gradrow == NULL)
537 error(SYSTEM, memerrmsg);
538 memset(gradient, 0, sizeof(gradient));
539 }
540 /* compute first row of edges */
541 for (j = 0; j < hp->ns-1; j++) {
542 comp_fftri(&fftr, hp, AI(hp,0,j), AI(hp,0,j+1));
543 if (hessrow != NULL)
544 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
545 if (gradrow != NULL)
546 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
547 }
548 /* sum each row of triangles */
549 for (i = 0; i < hp->ns-1; i++) {
550 FVECT hesscol[3]; /* compute first vertical edge */
551 FVECT gradcol;
552 comp_fftri(&fftr, hp, AI(hp,i,0), AI(hp,i+1,0));
553 if (hessrow != NULL)
554 comp_hessian(hesscol, &fftr, hp->rp->ron);
555 if (gradrow != NULL)
556 comp_gradient(gradcol, &fftr, hp->rp->ron);
557 for (j = 0; j < hp->ns-1; j++) {
558 FVECT hessdia[3]; /* compute triangle contributions */
559 FVECT graddia;
560 double backg;
561 backg = back_ambval(hp, AI(hp,i,j),
562 AI(hp,i,j+1), AI(hp,i+1,j));
563 /* diagonal (inner) edge */
564 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j));
565 if (hessrow != NULL) {
566 comp_hessian(hessdia, &fftr, hp->rp->ron);
567 rev_hessian(hesscol);
568 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
569 }
570 if (gradrow != NULL) {
571 comp_gradient(graddia, &fftr, hp->rp->ron);
572 rev_gradient(gradcol);
573 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
574 }
575 /* initialize edge in next row */
576 comp_fftri(&fftr, hp, AI(hp,i+1,j+1), AI(hp,i+1,j));
577 if (hessrow != NULL)
578 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
579 if (gradrow != NULL)
580 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
581 /* new column edge & paired triangle */
582 backg = back_ambval(hp, AI(hp,i+1,j+1),
583 AI(hp,i+1,j), AI(hp,i,j+1));
584 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j+1));
585 if (hessrow != NULL) {
586 comp_hessian(hesscol, &fftr, hp->rp->ron);
587 rev_hessian(hessdia);
588 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
589 if (i < hp->ns-2)
590 rev_hessian(hessrow[j]);
591 }
592 if (gradrow != NULL) {
593 comp_gradient(gradcol, &fftr, hp->rp->ron);
594 rev_gradient(graddia);
595 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
596 if (i < hp->ns-2)
597 rev_gradient(gradrow[j]);
598 }
599 }
600 }
601 /* release row buffers */
602 if (hessrow != NULL) free(hessrow);
603 if (gradrow != NULL) free(gradrow);
604
605 if (ra != NULL) /* extract eigenvectors & radii */
606 eigenvectors(uv, ra, hessian);
607 if (pg != NULL) { /* tangential position gradient */
608 pg[0] = DOT(gradient, uv[0]);
609 pg[1] = DOT(gradient, uv[1]);
610 }
611 }
612
613
614 /* Compute direction gradient from a hemispherical sampling */
615 static void
616 ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
617 {
618 AMBSAMP *ap;
619 double dgsum[2];
620 int n;
621 FVECT vd;
622 double gfact;
623
624 dgsum[0] = dgsum[1] = 0.0; /* sum values times -tan(theta) */
625 for (ap = hp->sa, n = hp->ns*hp->ns; n--; ap++) {
626 /* use vector for azimuth + 90deg */
627 VSUB(vd, ap->p, hp->rp->rop);
628 /* brightness over cosine factor */
629 gfact = colval(ap->v,CIEY) / DOT(hp->rp->ron, vd);
630 /* sine = proj_radius/vd_length */
631 dgsum[0] -= DOT(uv[1], vd) * gfact;
632 dgsum[1] += DOT(uv[0], vd) * gfact;
633 }
634 dg[0] = dgsum[0] / (hp->ns*hp->ns);
635 dg[1] = dgsum[1] / (hp->ns*hp->ns);
636 }
637
638
639 /* Compute potential light leak direction flags for cache value */
640 static uint32
641 ambcorral(AMBHEMI *hp, FVECT uv[2], const double r0, const double r1)
642 {
643 const double max_d = 1.0/(minarad*ambacc + 0.001);
644 const double ang_res = 0.5*PI/hp->ns;
645 const double ang_step = ang_res/((int)(16/PI*ang_res) + 1.01);
646 double avg_d = 0;
647 uint32 flgs = 0;
648 FVECT vec;
649 double u, v;
650 double ang, a1;
651 int i, j;
652 /* don't bother for a few samples */
653 if (hp->ns < 8)
654 return(0);
655 /* check distances overhead */
656 for (i = hp->ns*3/4; i-- > hp->ns>>2; )
657 for (j = hp->ns*3/4; j-- > hp->ns>>2; )
658 avg_d += ambsam(hp,i,j).d;
659 avg_d *= 4.0/(hp->ns*hp->ns);
660 if (avg_d*r0 >= 1.0) /* ceiling too low for corral? */
661 return(0);
662 if (avg_d >= max_d) /* insurance */
663 return(0);
664 /* else circle around perimeter */
665 for (i = 0; i < hp->ns; i++)
666 for (j = 0; j < hp->ns; j += !i|(i==hp->ns-1) ? 1 : hp->ns-1) {
667 AMBSAMP *ap = &ambsam(hp,i,j);
668 if ((ap->d <= FTINY) | (ap->d >= max_d))
669 continue; /* too far or too near */
670 VSUB(vec, ap->p, hp->rp->rop);
671 u = DOT(vec, uv[0]);
672 v = DOT(vec, uv[1]);
673 if ((r0*r0*u*u + r1*r1*v*v) * ap->d*ap->d <= u*u + v*v)
674 continue; /* occluder outside ellipse */
675 ang = atan2a(v, u); /* else set direction flags */
676 for (a1 = ang-ang_res; a1 <= ang+ang_res; a1 += ang_step)
677 flgs |= 1L<<(int)(16/PI*(a1 + 2.*PI*(a1 < 0)));
678 }
679 return(flgs);
680 }
681
682
683 int
684 doambient( /* compute ambient component */
685 COLOR rcol, /* input/output color */
686 RAY *r,
687 double wt,
688 FVECT uv[2], /* returned (optional) */
689 float ra[2], /* returned (optional) */
690 float pg[2], /* returned (optional) */
691 float dg[2], /* returned (optional) */
692 uint32 *crlp /* returned (optional) */
693 )
694 {
695 AMBHEMI *hp = samp_hemi(rcol, r, wt);
696 FVECT my_uv[2];
697 double d, K;
698 AMBSAMP *ap;
699 int i;
700 /* clear return values */
701 if (uv != NULL)
702 memset(uv, 0, sizeof(FVECT)*2);
703 if (ra != NULL)
704 ra[0] = ra[1] = 0.0;
705 if (pg != NULL)
706 pg[0] = pg[1] = 0.0;
707 if (dg != NULL)
708 dg[0] = dg[1] = 0.0;
709 if (crlp != NULL)
710 *crlp = 0;
711 if (hp == NULL) /* sampling falure? */
712 return(0);
713
714 if ((ra == NULL) & (pg == NULL) & (dg == NULL) ||
715 (hp->sampOK < 0) | (hp->ns < 6)) {
716 free(hp); /* Hessian not requested/possible */
717 return(-1); /* value-only return value */
718 }
719 if ((d = bright(rcol)) > FTINY) { /* normalize Y values */
720 d = 0.99*(hp->ns*hp->ns)/d;
721 K = 0.01;
722 } else { /* or fall back on geometric Hessian */
723 K = 1.0;
724 pg = NULL;
725 dg = NULL;
726 crlp = NULL;
727 }
728 ap = hp->sa; /* relative Y channel from here on... */
729 for (i = hp->ns*hp->ns; i--; ap++)
730 colval(ap->v,CIEY) = bright(ap->v)*d + K;
731
732 if (uv == NULL) /* make sure we have axis pointers */
733 uv = my_uv;
734 /* compute radii & pos. gradient */
735 ambHessian(hp, uv, ra, pg);
736
737 if (dg != NULL) /* compute direction gradient */
738 ambdirgrad(hp, uv, dg);
739
740 if (ra != NULL) { /* scale/clamp radii */
741 if (pg != NULL) {
742 if (ra[0]*(d = fabs(pg[0])) > 1.0)
743 ra[0] = 1.0/d;
744 if (ra[1]*(d = fabs(pg[1])) > 1.0)
745 ra[1] = 1.0/d;
746 if (ra[0] > ra[1])
747 ra[0] = ra[1];
748 }
749 if (ra[0] < minarad) {
750 ra[0] = minarad;
751 if (ra[1] < minarad)
752 ra[1] = minarad;
753 }
754 ra[0] *= d = 1.0/sqrt(wt);
755 if ((ra[1] *= d) > 2.0*ra[0])
756 ra[1] = 2.0*ra[0];
757 if (ra[1] > maxarad) {
758 ra[1] = maxarad;
759 if (ra[0] > maxarad)
760 ra[0] = maxarad;
761 }
762 /* flag encroached directions */
763 if (crlp != NULL)
764 *crlp = ambcorral(hp, uv, ra[0]*ambacc, ra[1]*ambacc);
765 if (pg != NULL) { /* cap gradient if necessary */
766 d = pg[0]*pg[0]*ra[0]*ra[0] + pg[1]*pg[1]*ra[1]*ra[1];
767 if (d > 1.0) {
768 d = 1.0/sqrt(d);
769 pg[0] *= d;
770 pg[1] *= d;
771 }
772 }
773 }
774 free(hp); /* clean up and return */
775 return(1);
776 }
777
778
779 #else /* ! NEWAMB */
780
781
782 void
783 inithemi( /* initialize sampling hemisphere */
784 AMBHEMI *hp,
785 COLOR ac,
786 RAY *r,
787 double wt
788 )
789 {
790 double d;
791 int i;
792 /* set number of divisions */
793 if (ambacc <= FTINY &&
794 wt > (d = 0.8*intens(ac)*r->rweight/(ambdiv*minweight)))
795 wt = d; /* avoid ray termination */
796 hp->nt = sqrt(ambdiv * wt / PI) + 0.5;
797 i = ambacc > FTINY ? 3 : 1; /* minimum number of samples */
798 if (hp->nt < i)
799 hp->nt = i;
800 hp->np = PI * hp->nt + 0.5;
801 /* set number of super-samples */
802 hp->ns = ambssamp * wt + 0.5;
803 /* assign coefficient */
804 copycolor(hp->acoef, ac);
805 d = 1.0/(hp->nt*hp->np);
806 scalecolor(hp->acoef, d);
807 /* make axes */
808 VCOPY(hp->uz, r->ron);
809 hp->uy[0] = hp->uy[1] = hp->uy[2] = 0.0;
810 for (i = 0; i < 3; i++)
811 if (hp->uz[i] < 0.6 && hp->uz[i] > -0.6)
812 break;
813 if (i >= 3)
814 error(CONSISTENCY, "bad ray direction in inithemi");
815 hp->uy[i] = 1.0;
816 fcross(hp->ux, hp->uy, hp->uz);
817 normalize(hp->ux);
818 fcross(hp->uy, hp->uz, hp->ux);
819 }
820
821
822 int
823 divsample( /* sample a division */
824 AMBSAMP *dp,
825 AMBHEMI *h,
826 RAY *r
827 )
828 {
829 RAY ar;
830 int hlist[3];
831 double spt[2];
832 double xd, yd, zd;
833 double b2;
834 double phi;
835 int i;
836 /* ambient coefficient for weight */
837 if (ambacc > FTINY)
838 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
839 else
840 copycolor(ar.rcoef, h->acoef);
841 if (rayorigin(&ar, AMBIENT, r, ar.rcoef) < 0)
842 return(-1);
843 if (ambacc > FTINY) {
844 multcolor(ar.rcoef, h->acoef);
845 scalecolor(ar.rcoef, 1./AVGREFL);
846 }
847 hlist[0] = r->rno;
848 hlist[1] = dp->t;
849 hlist[2] = dp->p;
850 multisamp(spt, 2, urand(ilhash(hlist,3)+dp->n));
851 zd = sqrt((dp->t + spt[0])/h->nt);
852 phi = 2.0*PI * (dp->p + spt[1])/h->np;
853 xd = tcos(phi) * zd;
854 yd = tsin(phi) * zd;
855 zd = sqrt(1.0 - zd*zd);
856 for (i = 0; i < 3; i++)
857 ar.rdir[i] = xd*h->ux[i] +
858 yd*h->uy[i] +
859 zd*h->uz[i];
860 checknorm(ar.rdir);
861 dimlist[ndims++] = dp->t*h->np + dp->p + 90171;
862 rayvalue(&ar);
863 ndims--;
864 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
865 addcolor(dp->v, ar.rcol);
866 /* use rt to improve gradient calc */
867 if (ar.rt > FTINY && ar.rt < FHUGE)
868 dp->r += 1.0/ar.rt;
869 /* (re)initialize error */
870 if (dp->n++) {
871 b2 = bright(dp->v)/dp->n - bright(ar.rcol);
872 b2 = b2*b2 + dp->k*((dp->n-1)*(dp->n-1));
873 dp->k = b2/(dp->n*dp->n);
874 } else
875 dp->k = 0.0;
876 return(0);
877 }
878
879
880 static int
881 ambcmp( /* decreasing order */
882 const void *p1,
883 const void *p2
884 )
885 {
886 const AMBSAMP *d1 = (const AMBSAMP *)p1;
887 const AMBSAMP *d2 = (const AMBSAMP *)p2;
888
889 if (d1->k < d2->k)
890 return(1);
891 if (d1->k > d2->k)
892 return(-1);
893 return(0);
894 }
895
896
897 static int
898 ambnorm( /* standard order */
899 const void *p1,
900 const void *p2
901 )
902 {
903 const AMBSAMP *d1 = (const AMBSAMP *)p1;
904 const AMBSAMP *d2 = (const AMBSAMP *)p2;
905 int c;
906
907 if ( (c = d1->t - d2->t) )
908 return(c);
909 return(d1->p - d2->p);
910 }
911
912
913 double
914 doambient( /* compute ambient component */
915 COLOR rcol,
916 RAY *r,
917 double wt,
918 FVECT pg,
919 FVECT dg
920 )
921 {
922 double b, d=0;
923 AMBHEMI hemi;
924 AMBSAMP *div;
925 AMBSAMP dnew;
926 double acol[3];
927 AMBSAMP *dp;
928 double arad;
929 int divcnt;
930 int i, j;
931 /* initialize hemisphere */
932 inithemi(&hemi, rcol, r, wt);
933 divcnt = hemi.nt * hemi.np;
934 /* initialize */
935 if (pg != NULL)
936 pg[0] = pg[1] = pg[2] = 0.0;
937 if (dg != NULL)
938 dg[0] = dg[1] = dg[2] = 0.0;
939 setcolor(rcol, 0.0, 0.0, 0.0);
940 if (divcnt == 0)
941 return(0.0);
942 /* allocate super-samples */
943 if (hemi.ns > 0 || pg != NULL || dg != NULL) {
944 div = (AMBSAMP *)malloc(divcnt*sizeof(AMBSAMP));
945 if (div == NULL)
946 error(SYSTEM, "out of memory in doambient");
947 } else
948 div = NULL;
949 /* sample the divisions */
950 arad = 0.0;
951 acol[0] = acol[1] = acol[2] = 0.0;
952 if ((dp = div) == NULL)
953 dp = &dnew;
954 divcnt = 0;
955 for (i = 0; i < hemi.nt; i++)
956 for (j = 0; j < hemi.np; j++) {
957 dp->t = i; dp->p = j;
958 setcolor(dp->v, 0.0, 0.0, 0.0);
959 dp->r = 0.0;
960 dp->n = 0;
961 if (divsample(dp, &hemi, r) < 0) {
962 if (div != NULL)
963 dp++;
964 continue;
965 }
966 arad += dp->r;
967 divcnt++;
968 if (div != NULL)
969 dp++;
970 else
971 addcolor(acol, dp->v);
972 }
973 if (!divcnt) {
974 if (div != NULL)
975 free((void *)div);
976 return(0.0); /* no samples taken */
977 }
978 if (divcnt < hemi.nt*hemi.np) {
979 pg = dg = NULL; /* incomplete sampling */
980 hemi.ns = 0;
981 } else if (arad > FTINY && divcnt/arad < minarad) {
982 hemi.ns = 0; /* close enough */
983 } else if (hemi.ns > 0) { /* else perform super-sampling? */
984 comperrs(div, &hemi); /* compute errors */
985 qsort(div, divcnt, sizeof(AMBSAMP), ambcmp); /* sort divs */
986 /* super-sample */
987 for (i = hemi.ns; i > 0; i--) {
988 dnew = *div;
989 if (divsample(&dnew, &hemi, r) < 0) {
990 dp++;
991 continue;
992 }
993 dp = div; /* reinsert */
994 j = divcnt < i ? divcnt : i;
995 while (--j > 0 && dnew.k < dp[1].k) {
996 *dp = *(dp+1);
997 dp++;
998 }
999 *dp = dnew;
1000 }
1001 if (pg != NULL || dg != NULL) /* restore order */
1002 qsort(div, divcnt, sizeof(AMBSAMP), ambnorm);
1003 }
1004 /* compute returned values */
1005 if (div != NULL) {
1006 arad = 0.0; /* note: divcnt may be < nt*np */
1007 for (i = hemi.nt*hemi.np, dp = div; i-- > 0; dp++) {
1008 arad += dp->r;
1009 if (dp->n > 1) {
1010 b = 1.0/dp->n;
1011 scalecolor(dp->v, b);
1012 dp->r *= b;
1013 dp->n = 1;
1014 }
1015 addcolor(acol, dp->v);
1016 }
1017 b = bright(acol);
1018 if (b > FTINY) {
1019 b = 1.0/b; /* compute & normalize gradient(s) */
1020 if (pg != NULL) {
1021 posgradient(pg, div, &hemi);
1022 for (i = 0; i < 3; i++)
1023 pg[i] *= b;
1024 }
1025 if (dg != NULL) {
1026 dirgradient(dg, div, &hemi);
1027 for (i = 0; i < 3; i++)
1028 dg[i] *= b;
1029 }
1030 }
1031 free((void *)div);
1032 }
1033 copycolor(rcol, acol);
1034 if (arad <= FTINY)
1035 arad = maxarad;
1036 else
1037 arad = (divcnt+hemi.ns)/arad;
1038 if (pg != NULL) { /* reduce radius if gradient large */
1039 d = DOT(pg,pg);
1040 if (d*arad*arad > 1.0)
1041 arad = 1.0/sqrt(d);
1042 }
1043 if (arad < minarad) {
1044 arad = minarad;
1045 if (pg != NULL && d*arad*arad > 1.0) { /* cap gradient */
1046 d = 1.0/arad/sqrt(d);
1047 for (i = 0; i < 3; i++)
1048 pg[i] *= d;
1049 }
1050 }
1051 if ((arad /= sqrt(wt)) > maxarad)
1052 arad = maxarad;
1053 return(arad);
1054 }
1055
1056
1057 void
1058 comperrs( /* compute initial error estimates */
1059 AMBSAMP *da, /* assumes standard ordering */
1060 AMBHEMI *hp
1061 )
1062 {
1063 double b, b2;
1064 int i, j;
1065 AMBSAMP *dp;
1066 /* sum differences from neighbors */
1067 dp = da;
1068 for (i = 0; i < hp->nt; i++)
1069 for (j = 0; j < hp->np; j++) {
1070 #ifdef DEBUG
1071 if (dp->t != i || dp->p != j)
1072 error(CONSISTENCY,
1073 "division order in comperrs");
1074 #endif
1075 b = bright(dp[0].v);
1076 if (i > 0) { /* from above */
1077 b2 = bright(dp[-hp->np].v) - b;
1078 b2 *= b2 * 0.25;
1079 dp[0].k += b2;
1080 dp[-hp->np].k += b2;
1081 }
1082 if (j > 0) { /* from behind */
1083 b2 = bright(dp[-1].v) - b;
1084 b2 *= b2 * 0.25;
1085 dp[0].k += b2;
1086 dp[-1].k += b2;
1087 } else { /* around */
1088 b2 = bright(dp[hp->np-1].v) - b;
1089 b2 *= b2 * 0.25;
1090 dp[0].k += b2;
1091 dp[hp->np-1].k += b2;
1092 }
1093 dp++;
1094 }
1095 /* divide by number of neighbors */
1096 dp = da;
1097 for (j = 0; j < hp->np; j++) /* top row */
1098 (dp++)->k *= 1.0/3.0;
1099 if (hp->nt < 2)
1100 return;
1101 for (i = 1; i < hp->nt-1; i++) /* central region */
1102 for (j = 0; j < hp->np; j++)
1103 (dp++)->k *= 0.25;
1104 for (j = 0; j < hp->np; j++) /* bottom row */
1105 (dp++)->k *= 1.0/3.0;
1106 }
1107
1108
1109 void
1110 posgradient( /* compute position gradient */
1111 FVECT gv,
1112 AMBSAMP *da, /* assumes standard ordering */
1113 AMBHEMI *hp
1114 )
1115 {
1116 int i, j;
1117 double nextsine, lastsine, b, d;
1118 double mag0, mag1;
1119 double phi, cosp, sinp, xd, yd;
1120 AMBSAMP *dp;
1121
1122 xd = yd = 0.0;
1123 for (j = 0; j < hp->np; j++) {
1124 dp = da + j;
1125 mag0 = mag1 = 0.0;
1126 lastsine = 0.0;
1127 for (i = 0; i < hp->nt; i++) {
1128 #ifdef DEBUG
1129 if (dp->t != i || dp->p != j)
1130 error(CONSISTENCY,
1131 "division order in posgradient");
1132 #endif
1133 b = bright(dp->v);
1134 if (i > 0) {
1135 d = dp[-hp->np].r;
1136 if (dp[0].r > d) d = dp[0].r;
1137 /* sin(t)*cos(t)^2 */
1138 d *= lastsine * (1.0 - (double)i/hp->nt);
1139 mag0 += d*(b - bright(dp[-hp->np].v));
1140 }
1141 nextsine = sqrt((double)(i+1)/hp->nt);
1142 if (j > 0) {
1143 d = dp[-1].r;
1144 if (dp[0].r > d) d = dp[0].r;
1145 mag1 += d * (nextsine - lastsine) *
1146 (b - bright(dp[-1].v));
1147 } else {
1148 d = dp[hp->np-1].r;
1149 if (dp[0].r > d) d = dp[0].r;
1150 mag1 += d * (nextsine - lastsine) *
1151 (b - bright(dp[hp->np-1].v));
1152 }
1153 dp += hp->np;
1154 lastsine = nextsine;
1155 }
1156 mag0 *= 2.0*PI / hp->np;
1157 phi = 2.0*PI * (double)j/hp->np;
1158 cosp = tcos(phi); sinp = tsin(phi);
1159 xd += mag0*cosp - mag1*sinp;
1160 yd += mag0*sinp + mag1*cosp;
1161 }
1162 for (i = 0; i < 3; i++)
1163 gv[i] = (xd*hp->ux[i] + yd*hp->uy[i])*(hp->nt*hp->np)/PI;
1164 }
1165
1166
1167 void
1168 dirgradient( /* compute direction gradient */
1169 FVECT gv,
1170 AMBSAMP *da, /* assumes standard ordering */
1171 AMBHEMI *hp
1172 )
1173 {
1174 int i, j;
1175 double mag;
1176 double phi, xd, yd;
1177 AMBSAMP *dp;
1178
1179 xd = yd = 0.0;
1180 for (j = 0; j < hp->np; j++) {
1181 dp = da + j;
1182 mag = 0.0;
1183 for (i = 0; i < hp->nt; i++) {
1184 #ifdef DEBUG
1185 if (dp->t != i || dp->p != j)
1186 error(CONSISTENCY,
1187 "division order in dirgradient");
1188 #endif
1189 /* tan(t) */
1190 mag += bright(dp->v)/sqrt(hp->nt/(i+.5) - 1.0);
1191 dp += hp->np;
1192 }
1193 phi = 2.0*PI * (j+.5)/hp->np + PI/2.0;
1194 xd += mag * tcos(phi);
1195 yd += mag * tsin(phi);
1196 }
1197 for (i = 0; i < 3; i++)
1198 gv[i] = xd*hp->ux[i] + yd*hp->uy[i];
1199 }
1200
1201 #endif /* ! NEWAMB */