ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambcomp.c
Revision: 2.77
Committed: Fri Apr 21 16:07:29 2017 UTC (7 years ago) by greg
Content type: text/plain
Branch: MAIN
CVS Tags: rad5R1
Changes since 2.76: +9 -5 lines
Log Message:
Fixed issue where ambient super-samples were being left off deep ray trees

File Contents

# Content
1 #ifndef lint
2 static const char RCSid[] = "$Id: ambcomp.c,v 2.76 2017/01/26 16:46:58 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 errors 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, d2;
168 int i, j;
169
170 if (earr == NULL) /* out of memory? */
171 return(NULL);
172 /* compute 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 d2 = normf*(b - bright(ap[-hp->ns].v));
178 d2 *= d2;
179 ep[0] += d2;
180 ep[-hp->ns] += d2;
181 }
182 if (!j) continue;
183 /* from behind */
184 d2 = normf*(b - bright(ap[-1].v));
185 d2 *= d2;
186 ep[0] += d2;
187 ep[-1] += d2;
188 if (!i) continue;
189 /* diagonal */
190 d2 = normf*(b - bright(ap[-hp->ns-1].v));
191 d2 *= d2;
192 ep[0] += d2;
193 ep[-hp->ns-1] += d2;
194 }
195 /* correct for number of neighbors */
196 earr[0] *= 8./3.;
197 earr[hp->ns-1] *= 8./3.;
198 earr[(hp->ns-1)*hp->ns] *= 8./3.;
199 earr[(hp->ns-1)*hp->ns + hp->ns-1] *= 8./3.;
200 for (i = 1; i < hp->ns-1; i++) {
201 earr[i*hp->ns] *= 8./5.;
202 earr[i*hp->ns + hp->ns-1] *= 8./5.;
203 }
204 for (j = 1; j < hp->ns-1; j++) {
205 earr[j] *= 8./5.;
206 earr[(hp->ns-1)*hp->ns + j] *= 8./5.;
207 }
208 return(earr);
209 }
210
211
212 /* Perform super-sampling on hemisphere (introduces bias) */
213 static void
214 ambsupersamp(AMBHEMI *hp, int cnt)
215 {
216 float *earr = getambdiffs(hp);
217 double e2rem = 0;
218 AMBSAMP *ap;
219 float *ep;
220 int i, j, n, nss;
221
222 if (earr == NULL) /* just skip calc. if no memory */
223 return;
224 /* accumulate estimated variances */
225 for (ep = earr + hp->ns*hp->ns; ep > earr; )
226 e2rem += *--ep;
227 ep = earr; /* perform super-sampling */
228 for (ap = hp->sa, i = 0; i < hp->ns; i++)
229 for (j = 0; j < hp->ns; j++, ap++) {
230 if (e2rem <= FTINY)
231 goto done; /* nothing left to do */
232 nss = *ep/e2rem*cnt + frandom();
233 for (n = 1; n <= nss && ambsample(hp,i,j,n); n++)
234 if (!--cnt) goto done;
235 e2rem -= *ep++; /* update remainder */
236 }
237 done:
238 free(earr);
239 }
240
241
242 static AMBHEMI *
243 samp_hemi( /* sample indirect hemisphere */
244 COLOR rcol,
245 RAY *r,
246 double wt
247 )
248 {
249 AMBHEMI *hp;
250 double d;
251 int n, i, j;
252 /* insignificance check */
253 if (bright(rcol) <= FTINY)
254 return(NULL);
255 /* set number of divisions */
256 if (ambacc <= FTINY &&
257 wt > (d = 0.8*intens(rcol)*r->rweight/(ambdiv*minweight)))
258 wt = d; /* avoid ray termination */
259 n = sqrt(ambdiv * wt) + 0.5;
260 i = 1 + 5*(ambacc > FTINY); /* minimum number of samples */
261 if (n < i)
262 n = i;
263 /* allocate sampling array */
264 hp = (AMBHEMI *)malloc(sizeof(AMBHEMI) + sizeof(AMBSAMP)*(n*n - 1));
265 if (hp == NULL)
266 error(SYSTEM, "out of memory in samp_hemi");
267 hp->rp = r;
268 hp->ns = n;
269 hp->acol[RED] = hp->acol[GRN] = hp->acol[BLU] = 0.0;
270 memset(hp->sa, 0, sizeof(AMBSAMP)*n*n);
271 hp->sampOK = 0;
272 /* assign coefficient */
273 copycolor(hp->acoef, rcol);
274 d = 1.0/(n*n);
275 scalecolor(hp->acoef, d);
276 /* make tangent plane axes */
277 if (!getperpendicular(hp->ux, r->ron, 1))
278 error(CONSISTENCY, "bad ray direction in samp_hemi");
279 VCROSS(hp->uy, r->ron, hp->ux);
280 /* sample divisions */
281 for (i = hp->ns; i--; )
282 for (j = hp->ns; j--; )
283 hp->sampOK += ambsample(hp, i, j, 0);
284 copycolor(rcol, hp->acol);
285 if (!hp->sampOK) { /* utter failure? */
286 free(hp);
287 return(NULL);
288 }
289 if (hp->sampOK < hp->ns*hp->ns) {
290 hp->sampOK *= -1; /* soft failure */
291 return(hp);
292 }
293 n = ambssamp*wt + 0.5;
294 if (n > 8) { /* perform super-sampling? */
295 ambsupersamp(hp, n);
296 copycolor(rcol, hp->acol);
297 }
298 return(hp); /* all is well */
299 }
300
301
302 /* Return brightness of farthest ambient sample */
303 static double
304 back_ambval(AMBHEMI *hp, const int n1, const int n2, const int n3)
305 {
306 if (hp->sa[n1].d <= hp->sa[n2].d) {
307 if (hp->sa[n1].d <= hp->sa[n3].d)
308 return(colval(hp->sa[n1].v,CIEY));
309 return(colval(hp->sa[n3].v,CIEY));
310 }
311 if (hp->sa[n2].d <= hp->sa[n3].d)
312 return(colval(hp->sa[n2].v,CIEY));
313 return(colval(hp->sa[n3].v,CIEY));
314 }
315
316
317 /* Compute vectors and coefficients for Hessian/gradient calcs */
318 static void
319 comp_fftri(FFTRI *ftp, AMBHEMI *hp, const int n0, const int n1)
320 {
321 double rdot_cp, dot_e, dot_er, rdot_r, rdot_r1, J2;
322 int ii;
323
324 VSUB(ftp->r_i, hp->sa[n0].p, hp->rp->rop);
325 VSUB(ftp->r_i1, hp->sa[n1].p, hp->rp->rop);
326 VSUB(ftp->e_i, hp->sa[n1].p, hp->sa[n0].p);
327 VCROSS(ftp->rcp, ftp->r_i, ftp->r_i1);
328 rdot_cp = 1.0/DOT(ftp->rcp,ftp->rcp);
329 dot_e = DOT(ftp->e_i,ftp->e_i);
330 dot_er = DOT(ftp->e_i, ftp->r_i);
331 rdot_r = 1.0/DOT(ftp->r_i,ftp->r_i);
332 rdot_r1 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
333 ftp->I1 = acos( DOT(ftp->r_i, ftp->r_i1) * sqrt(rdot_r*rdot_r1) ) *
334 sqrt( rdot_cp );
335 ftp->I2 = ( DOT(ftp->e_i, ftp->r_i1)*rdot_r1 - dot_er*rdot_r +
336 dot_e*ftp->I1 )*0.5*rdot_cp;
337 J2 = ( 0.5*(rdot_r - rdot_r1) - dot_er*ftp->I2 ) / dot_e;
338 for (ii = 3; ii--; )
339 ftp->rI2_eJ2[ii] = ftp->I2*ftp->r_i[ii] + J2*ftp->e_i[ii];
340 }
341
342
343 /* Compose 3x3 matrix from two vectors */
344 static void
345 compose_matrix(FVECT mat[3], FVECT va, FVECT vb)
346 {
347 mat[0][0] = 2.0*va[0]*vb[0];
348 mat[1][1] = 2.0*va[1]*vb[1];
349 mat[2][2] = 2.0*va[2]*vb[2];
350 mat[0][1] = mat[1][0] = va[0]*vb[1] + va[1]*vb[0];
351 mat[0][2] = mat[2][0] = va[0]*vb[2] + va[2]*vb[0];
352 mat[1][2] = mat[2][1] = va[1]*vb[2] + va[2]*vb[1];
353 }
354
355
356 /* Compute partial 3x3 Hessian matrix for edge */
357 static void
358 comp_hessian(FVECT hess[3], FFTRI *ftp, FVECT nrm)
359 {
360 FVECT ncp;
361 FVECT m1[3], m2[3], m3[3], m4[3];
362 double d1, d2, d3, d4;
363 double I3, J3, K3;
364 int i, j;
365 /* compute intermediate coefficients */
366 d1 = 1.0/DOT(ftp->r_i,ftp->r_i);
367 d2 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
368 d3 = 1.0/DOT(ftp->e_i,ftp->e_i);
369 d4 = DOT(ftp->e_i, ftp->r_i);
370 I3 = ( DOT(ftp->e_i, ftp->r_i1)*d2*d2 - d4*d1*d1 + 3.0/d3*ftp->I2 )
371 / ( 4.0*DOT(ftp->rcp,ftp->rcp) );
372 J3 = 0.25*d3*(d1*d1 - d2*d2) - d4*d3*I3;
373 K3 = d3*(ftp->I2 - I3/d1 - 2.0*d4*J3);
374 /* intermediate matrices */
375 VCROSS(ncp, nrm, ftp->e_i);
376 compose_matrix(m1, ncp, ftp->rI2_eJ2);
377 compose_matrix(m2, ftp->r_i, ftp->r_i);
378 compose_matrix(m3, ftp->e_i, ftp->e_i);
379 compose_matrix(m4, ftp->r_i, ftp->e_i);
380 d1 = DOT(nrm, ftp->rcp);
381 d2 = -d1*ftp->I2;
382 d1 *= 2.0;
383 for (i = 3; i--; ) /* final matrix sum */
384 for (j = 3; j--; ) {
385 hess[i][j] = m1[i][j] + d1*( I3*m2[i][j] + K3*m3[i][j] +
386 2.0*J3*m4[i][j] );
387 hess[i][j] += d2*(i==j);
388 hess[i][j] *= -1.0/PI;
389 }
390 }
391
392
393 /* Reverse hessian calculation result for edge in other direction */
394 static void
395 rev_hessian(FVECT hess[3])
396 {
397 int i;
398
399 for (i = 3; i--; ) {
400 hess[i][0] = -hess[i][0];
401 hess[i][1] = -hess[i][1];
402 hess[i][2] = -hess[i][2];
403 }
404 }
405
406
407 /* Add to radiometric Hessian from the given triangle */
408 static void
409 add2hessian(FVECT hess[3], FVECT ehess1[3],
410 FVECT ehess2[3], FVECT ehess3[3], double v)
411 {
412 int i, j;
413
414 for (i = 3; i--; )
415 for (j = 3; j--; )
416 hess[i][j] += v*( ehess1[i][j] + ehess2[i][j] + ehess3[i][j] );
417 }
418
419
420 /* Compute partial displacement form factor gradient for edge */
421 static void
422 comp_gradient(FVECT grad, FFTRI *ftp, FVECT nrm)
423 {
424 FVECT ncp;
425 double f1;
426 int i;
427
428 f1 = 2.0*DOT(nrm, ftp->rcp);
429 VCROSS(ncp, nrm, ftp->e_i);
430 for (i = 3; i--; )
431 grad[i] = (0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
432 }
433
434
435 /* Reverse gradient calculation result for edge in other direction */
436 static void
437 rev_gradient(FVECT grad)
438 {
439 grad[0] = -grad[0];
440 grad[1] = -grad[1];
441 grad[2] = -grad[2];
442 }
443
444
445 /* Add to displacement gradient from the given triangle */
446 static void
447 add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, double v)
448 {
449 int i;
450
451 for (i = 3; i--; )
452 grad[i] += v*( egrad1[i] + egrad2[i] + egrad3[i] );
453 }
454
455
456 /* Compute anisotropic radii and eigenvector directions */
457 static void
458 eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3])
459 {
460 double hess2[2][2];
461 FVECT a, b;
462 double evalue[2], slope1, xmag1;
463 int i;
464 /* project Hessian to sample plane */
465 for (i = 3; i--; ) {
466 a[i] = DOT(hessian[i], uv[0]);
467 b[i] = DOT(hessian[i], uv[1]);
468 }
469 hess2[0][0] = DOT(uv[0], a);
470 hess2[0][1] = DOT(uv[0], b);
471 hess2[1][0] = DOT(uv[1], a);
472 hess2[1][1] = DOT(uv[1], b);
473 /* compute eigenvalue(s) */
474 i = quadratic(evalue, 1.0, -hess2[0][0]-hess2[1][1],
475 hess2[0][0]*hess2[1][1]-hess2[0][1]*hess2[1][0]);
476 if (i == 1) /* double-root (circle) */
477 evalue[1] = evalue[0];
478 if (!i || ((evalue[0] = fabs(evalue[0])) <= FTINY*FTINY) |
479 ((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) ) {
480 ra[0] = ra[1] = maxarad;
481 return;
482 }
483 if (evalue[0] > evalue[1]) {
484 ra[0] = sqrt(sqrt(4.0/evalue[0]));
485 ra[1] = sqrt(sqrt(4.0/evalue[1]));
486 slope1 = evalue[1];
487 } else {
488 ra[0] = sqrt(sqrt(4.0/evalue[1]));
489 ra[1] = sqrt(sqrt(4.0/evalue[0]));
490 slope1 = evalue[0];
491 }
492 /* compute unit eigenvectors */
493 if (fabs(hess2[0][1]) <= FTINY)
494 return; /* uv OK as is */
495 slope1 = (slope1 - hess2[0][0]) / hess2[0][1];
496 xmag1 = sqrt(1.0/(1.0 + slope1*slope1));
497 for (i = 3; i--; ) {
498 b[i] = xmag1*uv[0][i] + slope1*xmag1*uv[1][i];
499 a[i] = slope1*xmag1*uv[0][i] - xmag1*uv[1][i];
500 }
501 VCOPY(uv[0], a);
502 VCOPY(uv[1], b);
503 }
504
505
506 static void
507 ambHessian( /* anisotropic radii & pos. gradient */
508 AMBHEMI *hp,
509 FVECT uv[2], /* returned */
510 float ra[2], /* returned (optional) */
511 float pg[2] /* returned (optional) */
512 )
513 {
514 static char memerrmsg[] = "out of memory in ambHessian()";
515 FVECT (*hessrow)[3] = NULL;
516 FVECT *gradrow = NULL;
517 FVECT hessian[3];
518 FVECT gradient;
519 FFTRI fftr;
520 int i, j;
521 /* be sure to assign unit vectors */
522 VCOPY(uv[0], hp->ux);
523 VCOPY(uv[1], hp->uy);
524 /* clock-wise vertex traversal from sample POV */
525 if (ra != NULL) { /* initialize Hessian row buffer */
526 hessrow = (FVECT (*)[3])malloc(sizeof(FVECT)*3*(hp->ns-1));
527 if (hessrow == NULL)
528 error(SYSTEM, memerrmsg);
529 memset(hessian, 0, sizeof(hessian));
530 } else if (pg == NULL) /* bogus call? */
531 return;
532 if (pg != NULL) { /* initialize form factor row buffer */
533 gradrow = (FVECT *)malloc(sizeof(FVECT)*(hp->ns-1));
534 if (gradrow == NULL)
535 error(SYSTEM, memerrmsg);
536 memset(gradient, 0, sizeof(gradient));
537 }
538 /* compute first row of edges */
539 for (j = 0; j < hp->ns-1; j++) {
540 comp_fftri(&fftr, hp, AI(hp,0,j), AI(hp,0,j+1));
541 if (hessrow != NULL)
542 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
543 if (gradrow != NULL)
544 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
545 }
546 /* sum each row of triangles */
547 for (i = 0; i < hp->ns-1; i++) {
548 FVECT hesscol[3]; /* compute first vertical edge */
549 FVECT gradcol;
550 comp_fftri(&fftr, hp, AI(hp,i,0), AI(hp,i+1,0));
551 if (hessrow != NULL)
552 comp_hessian(hesscol, &fftr, hp->rp->ron);
553 if (gradrow != NULL)
554 comp_gradient(gradcol, &fftr, hp->rp->ron);
555 for (j = 0; j < hp->ns-1; j++) {
556 FVECT hessdia[3]; /* compute triangle contributions */
557 FVECT graddia;
558 double backg;
559 backg = back_ambval(hp, AI(hp,i,j),
560 AI(hp,i,j+1), AI(hp,i+1,j));
561 /* diagonal (inner) edge */
562 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j));
563 if (hessrow != NULL) {
564 comp_hessian(hessdia, &fftr, hp->rp->ron);
565 rev_hessian(hesscol);
566 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
567 }
568 if (gradrow != NULL) {
569 comp_gradient(graddia, &fftr, hp->rp->ron);
570 rev_gradient(gradcol);
571 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
572 }
573 /* initialize edge in next row */
574 comp_fftri(&fftr, hp, AI(hp,i+1,j+1), AI(hp,i+1,j));
575 if (hessrow != NULL)
576 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
577 if (gradrow != NULL)
578 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
579 /* new column edge & paired triangle */
580 backg = back_ambval(hp, AI(hp,i+1,j+1),
581 AI(hp,i+1,j), AI(hp,i,j+1));
582 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j+1));
583 if (hessrow != NULL) {
584 comp_hessian(hesscol, &fftr, hp->rp->ron);
585 rev_hessian(hessdia);
586 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
587 if (i < hp->ns-2)
588 rev_hessian(hessrow[j]);
589 }
590 if (gradrow != NULL) {
591 comp_gradient(gradcol, &fftr, hp->rp->ron);
592 rev_gradient(graddia);
593 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
594 if (i < hp->ns-2)
595 rev_gradient(gradrow[j]);
596 }
597 }
598 }
599 /* release row buffers */
600 if (hessrow != NULL) free(hessrow);
601 if (gradrow != NULL) free(gradrow);
602
603 if (ra != NULL) /* extract eigenvectors & radii */
604 eigenvectors(uv, ra, hessian);
605 if (pg != NULL) { /* tangential position gradient */
606 pg[0] = DOT(gradient, uv[0]);
607 pg[1] = DOT(gradient, uv[1]);
608 }
609 }
610
611
612 /* Compute direction gradient from a hemispherical sampling */
613 static void
614 ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
615 {
616 AMBSAMP *ap;
617 double dgsum[2];
618 int n;
619 FVECT vd;
620 double gfact;
621
622 dgsum[0] = dgsum[1] = 0.0; /* sum values times -tan(theta) */
623 for (ap = hp->sa, n = hp->ns*hp->ns; n--; ap++) {
624 /* use vector for azimuth + 90deg */
625 VSUB(vd, ap->p, hp->rp->rop);
626 /* brightness over cosine factor */
627 gfact = colval(ap->v,CIEY) / DOT(hp->rp->ron, vd);
628 /* sine = proj_radius/vd_length */
629 dgsum[0] -= DOT(uv[1], vd) * gfact;
630 dgsum[1] += DOT(uv[0], vd) * gfact;
631 }
632 dg[0] = dgsum[0] / (hp->ns*hp->ns);
633 dg[1] = dgsum[1] / (hp->ns*hp->ns);
634 }
635
636
637 /* Compute potential light leak direction flags for cache value */
638 static uint32
639 ambcorral(AMBHEMI *hp, FVECT uv[2], const double r0, const double r1)
640 {
641 const double max_d = 1.0/(minarad*ambacc + 0.001);
642 const double ang_res = 0.5*PI/hp->ns;
643 const double ang_step = ang_res/((int)(16/PI*ang_res) + 1.01);
644 double avg_d = 0;
645 uint32 flgs = 0;
646 FVECT vec;
647 double u, v;
648 double ang, a1;
649 int i, j;
650 /* don't bother for a few samples */
651 if (hp->ns < 8)
652 return(0);
653 /* check distances overhead */
654 for (i = hp->ns*3/4; i-- > hp->ns>>2; )
655 for (j = hp->ns*3/4; j-- > hp->ns>>2; )
656 avg_d += ambsam(hp,i,j).d;
657 avg_d *= 4.0/(hp->ns*hp->ns);
658 if (avg_d*r0 >= 1.0) /* ceiling too low for corral? */
659 return(0);
660 if (avg_d >= max_d) /* insurance */
661 return(0);
662 /* else circle around perimeter */
663 for (i = 0; i < hp->ns; i++)
664 for (j = 0; j < hp->ns; j += !i|(i==hp->ns-1) ? 1 : hp->ns-1) {
665 AMBSAMP *ap = &ambsam(hp,i,j);
666 if ((ap->d <= FTINY) | (ap->d >= max_d))
667 continue; /* too far or too near */
668 VSUB(vec, ap->p, hp->rp->rop);
669 u = DOT(vec, uv[0]);
670 v = DOT(vec, uv[1]);
671 if ((r0*r0*u*u + r1*r1*v*v) * ap->d*ap->d <= u*u + v*v)
672 continue; /* occluder outside ellipse */
673 ang = atan2a(v, u); /* else set direction flags */
674 for (a1 = ang-ang_res; a1 <= ang+ang_res; a1 += ang_step)
675 flgs |= 1L<<(int)(16/PI*(a1 + 2.*PI*(a1 < 0)));
676 }
677 /* add low-angle incident (< 20deg) */
678 if (fabs(hp->rp->rod) <= 0.342) {
679 u = -DOT(hp->rp->rdir, uv[0]);
680 v = -DOT(hp->rp->rdir, uv[1]);
681 if ((r0*r0*u*u + r1*r1*v*v) > hp->rp->rot*hp->rp->rot) {
682 ang = atan2a(v, u);
683 ang += 2.*PI*(ang < 0);
684 ang *= 16/PI;
685 if ((ang < .5) | (ang >= 31.5))
686 flgs |= 0x80000001;
687 else
688 flgs |= 3L<<(int)(ang-.5);
689 }
690 }
691 return(flgs);
692 }
693
694
695 int
696 doambient( /* compute ambient component */
697 COLOR rcol, /* input/output color */
698 RAY *r,
699 double wt,
700 FVECT uv[2], /* returned (optional) */
701 float ra[2], /* returned (optional) */
702 float pg[2], /* returned (optional) */
703 float dg[2], /* returned (optional) */
704 uint32 *crlp /* returned (optional) */
705 )
706 {
707 AMBHEMI *hp = samp_hemi(rcol, r, wt);
708 FVECT my_uv[2];
709 double d, K;
710 AMBSAMP *ap;
711 int i;
712 /* clear return values */
713 if (uv != NULL)
714 memset(uv, 0, sizeof(FVECT)*2);
715 if (ra != NULL)
716 ra[0] = ra[1] = 0.0;
717 if (pg != NULL)
718 pg[0] = pg[1] = 0.0;
719 if (dg != NULL)
720 dg[0] = dg[1] = 0.0;
721 if (crlp != NULL)
722 *crlp = 0;
723 if (hp == NULL) /* sampling falure? */
724 return(0);
725
726 if ((ra == NULL) & (pg == NULL) & (dg == NULL) ||
727 (hp->sampOK < 0) | (hp->ns < 6)) {
728 free(hp); /* Hessian not requested/possible */
729 return(-1); /* value-only return value */
730 }
731 if ((d = bright(rcol)) > FTINY) { /* normalize Y values */
732 d = 0.99*(hp->ns*hp->ns)/d;
733 K = 0.01;
734 } else { /* or fall back on geometric Hessian */
735 K = 1.0;
736 pg = NULL;
737 dg = NULL;
738 crlp = NULL;
739 }
740 ap = hp->sa; /* relative Y channel from here on... */
741 for (i = hp->ns*hp->ns; i--; ap++)
742 colval(ap->v,CIEY) = bright(ap->v)*d + K;
743
744 if (uv == NULL) /* make sure we have axis pointers */
745 uv = my_uv;
746 /* compute radii & pos. gradient */
747 ambHessian(hp, uv, ra, pg);
748
749 if (dg != NULL) /* compute direction gradient */
750 ambdirgrad(hp, uv, dg);
751
752 if (ra != NULL) { /* scale/clamp radii */
753 if (pg != NULL) {
754 if (ra[0]*(d = fabs(pg[0])) > 1.0)
755 ra[0] = 1.0/d;
756 if (ra[1]*(d = fabs(pg[1])) > 1.0)
757 ra[1] = 1.0/d;
758 if (ra[0] > ra[1])
759 ra[0] = ra[1];
760 }
761 if (ra[0] < minarad) {
762 ra[0] = minarad;
763 if (ra[1] < minarad)
764 ra[1] = minarad;
765 }
766 ra[0] *= d = 1.0/sqrt(wt);
767 if ((ra[1] *= d) > 2.0*ra[0])
768 ra[1] = 2.0*ra[0];
769 if (ra[1] > maxarad) {
770 ra[1] = maxarad;
771 if (ra[0] > maxarad)
772 ra[0] = maxarad;
773 }
774 /* flag encroached directions */
775 if (crlp != NULL)
776 *crlp = ambcorral(hp, uv, ra[0]*ambacc, ra[1]*ambacc);
777 if (pg != NULL) { /* cap gradient if necessary */
778 d = pg[0]*pg[0]*ra[0]*ra[0] + pg[1]*pg[1]*ra[1]*ra[1];
779 if (d > 1.0) {
780 d = 1.0/sqrt(d);
781 pg[0] *= d;
782 pg[1] *= d;
783 }
784 }
785 }
786 free(hp); /* clean up and return */
787 return(1);
788 }
789
790
791 #else /* ! NEWAMB */
792
793
794 void
795 inithemi( /* initialize sampling hemisphere */
796 AMBHEMI *hp,
797 COLOR ac,
798 RAY *r,
799 double wt
800 )
801 {
802 double d;
803 int i;
804 /* set number of divisions */
805 if (ambacc <= FTINY &&
806 wt > (d = 0.8*intens(ac)*r->rweight/(ambdiv*minweight)))
807 wt = d; /* avoid ray termination */
808 hp->nt = sqrt(ambdiv * wt / PI) + 0.5;
809 i = ambacc > FTINY ? 3 : 1; /* minimum number of samples */
810 if (hp->nt < i)
811 hp->nt = i;
812 hp->np = PI * hp->nt + 0.5;
813 /* set number of super-samples */
814 hp->ns = ambssamp * wt + 0.5;
815 /* assign coefficient */
816 copycolor(hp->acoef, ac);
817 d = 1.0/(hp->nt*hp->np);
818 scalecolor(hp->acoef, d);
819 /* make axes */
820 VCOPY(hp->uz, r->ron);
821 hp->uy[0] = hp->uy[1] = hp->uy[2] = 0.0;
822 for (i = 0; i < 3; i++)
823 if (hp->uz[i] < 0.6 && hp->uz[i] > -0.6)
824 break;
825 if (i >= 3)
826 error(CONSISTENCY, "bad ray direction in inithemi");
827 hp->uy[i] = 1.0;
828 fcross(hp->ux, hp->uy, hp->uz);
829 normalize(hp->ux);
830 fcross(hp->uy, hp->uz, hp->ux);
831 }
832
833
834 int
835 divsample( /* sample a division */
836 AMBSAMP *dp,
837 AMBHEMI *h,
838 RAY *r
839 )
840 {
841 RAY ar;
842 int hlist[3];
843 double spt[2];
844 double xd, yd, zd;
845 double b2;
846 double phi;
847 int i;
848 /* ambient coefficient for weight */
849 if (ambacc > FTINY)
850 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
851 else
852 copycolor(ar.rcoef, h->acoef);
853 if (rayorigin(&ar, AMBIENT, r, ar.rcoef) < 0)
854 return(-1);
855 if (ambacc > FTINY) {
856 multcolor(ar.rcoef, h->acoef);
857 scalecolor(ar.rcoef, 1./AVGREFL);
858 }
859 hlist[0] = r->rno;
860 hlist[1] = dp->t;
861 hlist[2] = dp->p;
862 multisamp(spt, 2, urand(ilhash(hlist,3)+dp->n));
863 zd = sqrt((dp->t + spt[0])/h->nt);
864 phi = 2.0*PI * (dp->p + spt[1])/h->np;
865 xd = tcos(phi) * zd;
866 yd = tsin(phi) * zd;
867 zd = sqrt(1.0 - zd*zd);
868 for (i = 0; i < 3; i++)
869 ar.rdir[i] = xd*h->ux[i] +
870 yd*h->uy[i] +
871 zd*h->uz[i];
872 checknorm(ar.rdir);
873 dimlist[ndims++] = dp->t*h->np + dp->p + 90171;
874 rayvalue(&ar);
875 ndims--;
876 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
877 addcolor(dp->v, ar.rcol);
878 /* use rt to improve gradient calc */
879 if (ar.rt > FTINY && ar.rt < FHUGE)
880 dp->r += 1.0/ar.rt;
881 /* (re)initialize error */
882 if (dp->n++) {
883 b2 = bright(dp->v)/dp->n - bright(ar.rcol);
884 b2 = b2*b2 + dp->k*((dp->n-1)*(dp->n-1));
885 dp->k = b2/(dp->n*dp->n);
886 } else
887 dp->k = 0.0;
888 return(0);
889 }
890
891
892 static int
893 ambcmp( /* decreasing order */
894 const void *p1,
895 const void *p2
896 )
897 {
898 const AMBSAMP *d1 = (const AMBSAMP *)p1;
899 const AMBSAMP *d2 = (const AMBSAMP *)p2;
900
901 if (d1->k < d2->k)
902 return(1);
903 if (d1->k > d2->k)
904 return(-1);
905 return(0);
906 }
907
908
909 static int
910 ambnorm( /* standard order */
911 const void *p1,
912 const void *p2
913 )
914 {
915 const AMBSAMP *d1 = (const AMBSAMP *)p1;
916 const AMBSAMP *d2 = (const AMBSAMP *)p2;
917 int c;
918
919 if ( (c = d1->t - d2->t) )
920 return(c);
921 return(d1->p - d2->p);
922 }
923
924
925 double
926 doambient( /* compute ambient component */
927 COLOR rcol,
928 RAY *r,
929 double wt,
930 FVECT pg,
931 FVECT dg
932 )
933 {
934 double b, d=0;
935 AMBHEMI hemi;
936 AMBSAMP *div;
937 AMBSAMP dnew;
938 double acol[3];
939 AMBSAMP *dp;
940 double arad;
941 int divcnt;
942 int i, j;
943 /* initialize hemisphere */
944 inithemi(&hemi, rcol, r, wt);
945 divcnt = hemi.nt * hemi.np;
946 /* initialize */
947 if (pg != NULL)
948 pg[0] = pg[1] = pg[2] = 0.0;
949 if (dg != NULL)
950 dg[0] = dg[1] = dg[2] = 0.0;
951 setcolor(rcol, 0.0, 0.0, 0.0);
952 if (divcnt == 0)
953 return(0.0);
954 /* allocate super-samples */
955 if (hemi.ns > 0 || pg != NULL || dg != NULL) {
956 div = (AMBSAMP *)malloc(divcnt*sizeof(AMBSAMP));
957 if (div == NULL)
958 error(SYSTEM, "out of memory in doambient");
959 } else
960 div = NULL;
961 /* sample the divisions */
962 arad = 0.0;
963 acol[0] = acol[1] = acol[2] = 0.0;
964 if ((dp = div) == NULL)
965 dp = &dnew;
966 divcnt = 0;
967 for (i = 0; i < hemi.nt; i++)
968 for (j = 0; j < hemi.np; j++) {
969 dp->t = i; dp->p = j;
970 setcolor(dp->v, 0.0, 0.0, 0.0);
971 dp->r = 0.0;
972 dp->n = 0;
973 if (divsample(dp, &hemi, r) < 0) {
974 if (div != NULL)
975 dp++;
976 continue;
977 }
978 arad += dp->r;
979 divcnt++;
980 if (div != NULL)
981 dp++;
982 else
983 addcolor(acol, dp->v);
984 }
985 if (!divcnt) {
986 if (div != NULL)
987 free((void *)div);
988 return(0.0); /* no samples taken */
989 }
990 if (divcnt < hemi.nt*hemi.np) {
991 pg = dg = NULL; /* incomplete sampling */
992 hemi.ns = 0;
993 } else if (arad > FTINY && divcnt/arad < minarad) {
994 hemi.ns = 0; /* close enough */
995 } else if (hemi.ns > 0) { /* else perform super-sampling? */
996 comperrs(div, &hemi); /* compute errors */
997 qsort(div, divcnt, sizeof(AMBSAMP), ambcmp); /* sort divs */
998 /* super-sample */
999 for (i = hemi.ns; i > 0; i--) {
1000 dnew = *div;
1001 if (divsample(&dnew, &hemi, r) < 0) {
1002 dp++;
1003 continue;
1004 }
1005 dp = div; /* reinsert */
1006 j = divcnt < i ? divcnt : i;
1007 while (--j > 0 && dnew.k < dp[1].k) {
1008 *dp = *(dp+1);
1009 dp++;
1010 }
1011 *dp = dnew;
1012 }
1013 if (pg != NULL || dg != NULL) /* restore order */
1014 qsort(div, divcnt, sizeof(AMBSAMP), ambnorm);
1015 }
1016 /* compute returned values */
1017 if (div != NULL) {
1018 arad = 0.0; /* note: divcnt may be < nt*np */
1019 for (i = hemi.nt*hemi.np, dp = div; i-- > 0; dp++) {
1020 arad += dp->r;
1021 if (dp->n > 1) {
1022 b = 1.0/dp->n;
1023 scalecolor(dp->v, b);
1024 dp->r *= b;
1025 dp->n = 1;
1026 }
1027 addcolor(acol, dp->v);
1028 }
1029 b = bright(acol);
1030 if (b > FTINY) {
1031 b = 1.0/b; /* compute & normalize gradient(s) */
1032 if (pg != NULL) {
1033 posgradient(pg, div, &hemi);
1034 for (i = 0; i < 3; i++)
1035 pg[i] *= b;
1036 }
1037 if (dg != NULL) {
1038 dirgradient(dg, div, &hemi);
1039 for (i = 0; i < 3; i++)
1040 dg[i] *= b;
1041 }
1042 }
1043 free((void *)div);
1044 }
1045 copycolor(rcol, acol);
1046 if (arad <= FTINY)
1047 arad = maxarad;
1048 else
1049 arad = (divcnt+hemi.ns)/arad;
1050 if (pg != NULL) { /* reduce radius if gradient large */
1051 d = DOT(pg,pg);
1052 if (d*arad*arad > 1.0)
1053 arad = 1.0/sqrt(d);
1054 }
1055 if (arad < minarad) {
1056 arad = minarad;
1057 if (pg != NULL && d*arad*arad > 1.0) { /* cap gradient */
1058 d = 1.0/arad/sqrt(d);
1059 for (i = 0; i < 3; i++)
1060 pg[i] *= d;
1061 }
1062 }
1063 if ((arad /= sqrt(wt)) > maxarad)
1064 arad = maxarad;
1065 return(arad);
1066 }
1067
1068
1069 void
1070 comperrs( /* compute initial error estimates */
1071 AMBSAMP *da, /* assumes standard ordering */
1072 AMBHEMI *hp
1073 )
1074 {
1075 double b, b2;
1076 int i, j;
1077 AMBSAMP *dp;
1078 /* sum differences from neighbors */
1079 dp = da;
1080 for (i = 0; i < hp->nt; i++)
1081 for (j = 0; j < hp->np; j++) {
1082 #ifdef DEBUG
1083 if (dp->t != i || dp->p != j)
1084 error(CONSISTENCY,
1085 "division order in comperrs");
1086 #endif
1087 b = bright(dp[0].v);
1088 if (i > 0) { /* from above */
1089 b2 = bright(dp[-hp->np].v) - b;
1090 b2 *= b2 * 0.25;
1091 dp[0].k += b2;
1092 dp[-hp->np].k += b2;
1093 }
1094 if (j > 0) { /* from behind */
1095 b2 = bright(dp[-1].v) - b;
1096 b2 *= b2 * 0.25;
1097 dp[0].k += b2;
1098 dp[-1].k += b2;
1099 } else { /* around */
1100 b2 = bright(dp[hp->np-1].v) - b;
1101 b2 *= b2 * 0.25;
1102 dp[0].k += b2;
1103 dp[hp->np-1].k += b2;
1104 }
1105 dp++;
1106 }
1107 /* divide by number of neighbors */
1108 dp = da;
1109 for (j = 0; j < hp->np; j++) /* top row */
1110 (dp++)->k *= 1.0/3.0;
1111 if (hp->nt < 2)
1112 return;
1113 for (i = 1; i < hp->nt-1; i++) /* central region */
1114 for (j = 0; j < hp->np; j++)
1115 (dp++)->k *= 0.25;
1116 for (j = 0; j < hp->np; j++) /* bottom row */
1117 (dp++)->k *= 1.0/3.0;
1118 }
1119
1120
1121 void
1122 posgradient( /* compute position gradient */
1123 FVECT gv,
1124 AMBSAMP *da, /* assumes standard ordering */
1125 AMBHEMI *hp
1126 )
1127 {
1128 int i, j;
1129 double nextsine, lastsine, b, d;
1130 double mag0, mag1;
1131 double phi, cosp, sinp, xd, yd;
1132 AMBSAMP *dp;
1133
1134 xd = yd = 0.0;
1135 for (j = 0; j < hp->np; j++) {
1136 dp = da + j;
1137 mag0 = mag1 = 0.0;
1138 lastsine = 0.0;
1139 for (i = 0; i < hp->nt; i++) {
1140 #ifdef DEBUG
1141 if (dp->t != i || dp->p != j)
1142 error(CONSISTENCY,
1143 "division order in posgradient");
1144 #endif
1145 b = bright(dp->v);
1146 if (i > 0) {
1147 d = dp[-hp->np].r;
1148 if (dp[0].r > d) d = dp[0].r;
1149 /* sin(t)*cos(t)^2 */
1150 d *= lastsine * (1.0 - (double)i/hp->nt);
1151 mag0 += d*(b - bright(dp[-hp->np].v));
1152 }
1153 nextsine = sqrt((double)(i+1)/hp->nt);
1154 if (j > 0) {
1155 d = dp[-1].r;
1156 if (dp[0].r > d) d = dp[0].r;
1157 mag1 += d * (nextsine - lastsine) *
1158 (b - bright(dp[-1].v));
1159 } else {
1160 d = dp[hp->np-1].r;
1161 if (dp[0].r > d) d = dp[0].r;
1162 mag1 += d * (nextsine - lastsine) *
1163 (b - bright(dp[hp->np-1].v));
1164 }
1165 dp += hp->np;
1166 lastsine = nextsine;
1167 }
1168 mag0 *= 2.0*PI / hp->np;
1169 phi = 2.0*PI * (double)j/hp->np;
1170 cosp = tcos(phi); sinp = tsin(phi);
1171 xd += mag0*cosp - mag1*sinp;
1172 yd += mag0*sinp + mag1*cosp;
1173 }
1174 for (i = 0; i < 3; i++)
1175 gv[i] = (xd*hp->ux[i] + yd*hp->uy[i])*(hp->nt*hp->np)/PI;
1176 }
1177
1178
1179 void
1180 dirgradient( /* compute direction gradient */
1181 FVECT gv,
1182 AMBSAMP *da, /* assumes standard ordering */
1183 AMBHEMI *hp
1184 )
1185 {
1186 int i, j;
1187 double mag;
1188 double phi, xd, yd;
1189 AMBSAMP *dp;
1190
1191 xd = yd = 0.0;
1192 for (j = 0; j < hp->np; j++) {
1193 dp = da + j;
1194 mag = 0.0;
1195 for (i = 0; i < hp->nt; i++) {
1196 #ifdef DEBUG
1197 if (dp->t != i || dp->p != j)
1198 error(CONSISTENCY,
1199 "division order in dirgradient");
1200 #endif
1201 /* tan(t) */
1202 mag += bright(dp->v)/sqrt(hp->nt/(i+.5) - 1.0);
1203 dp += hp->np;
1204 }
1205 phi = 2.0*PI * (j+.5)/hp->np + PI/2.0;
1206 xd += mag * tcos(phi);
1207 yd += mag * tsin(phi);
1208 }
1209 for (i = 0; i < 3; i++)
1210 gv[i] = xd*hp->ux[i] + yd*hp->uy[i];
1211 }
1212
1213 #endif /* ! NEWAMB */