ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambcomp.c
Revision: 2.83
Committed: Tue Nov 13 19:58:33 2018 UTC (5 years, 6 months ago) by greg
Content type: text/plain
Branch: MAIN
Changes since 2.82: +12 -11 lines
Log Message:
Added -orRxX options to rtrace for VR rendering

File Contents

# Content
1 #ifndef lint
2 static const char RCSid[] = "$Id: ambcomp.c,v 2.82 2018/04/12 20:07:09 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 */
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 zd = raydistance(&ar);
135 if (zd <= FTINY)
136 return(0); /* should never happen */
137 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
138 if (zd*ap->d < 1.0) /* new/closer distance? */
139 ap->d = 1.0/zd;
140 if (!n) { /* record first vertex & value */
141 if (zd > 10.0*thescene.cusize + 1000.)
142 zd = 10.0*thescene.cusize + 1000.;
143 VSUM(ap->p, ar.rorg, ar.rdir, zd);
144 copycolor(ap->v, ar.rcol);
145 } else { /* else update recorded value */
146 hp->acol[RED] -= colval(ap->v,RED);
147 hp->acol[GRN] -= colval(ap->v,GRN);
148 hp->acol[BLU] -= colval(ap->v,BLU);
149 zd = 1.0/(double)(n+1);
150 scalecolor(ar.rcol, zd);
151 zd *= (double)n;
152 scalecolor(ap->v, zd);
153 addcolor(ap->v, ar.rcol);
154 }
155 addcolor(hp->acol, ap->v); /* add to our sum */
156 return(1);
157 }
158
159
160 /* Estimate variance based on ambient division differences */
161 static float *
162 getambdiffs(AMBHEMI *hp)
163 {
164 const double normf = 1./bright(hp->acoef);
165 float *earr = (float *)calloc(hp->ns*hp->ns, sizeof(float));
166 float *ep;
167 AMBSAMP *ap;
168 double b, b1, d2;
169 int i, j;
170
171 if (earr == NULL) /* out of memory? */
172 return(NULL);
173 /* sum squared neighbor diffs */
174 for (ap = hp->sa, ep = earr, i = 0; i < hp->ns; i++)
175 for (j = 0; j < hp->ns; j++, ap++, ep++) {
176 b = bright(ap[0].v);
177 if (i) { /* from above */
178 b1 = bright(ap[-hp->ns].v);
179 d2 = b - b1;
180 d2 *= d2*normf/(b + b1);
181 ep[0] += d2;
182 ep[-hp->ns] += d2;
183 }
184 if (!j) continue;
185 /* from behind */
186 b1 = bright(ap[-1].v);
187 d2 = b - b1;
188 d2 *= d2*normf/(b + b1);
189 ep[0] += d2;
190 ep[-1] += d2;
191 if (!i) continue;
192 /* diagonal */
193 b1 = bright(ap[-hp->ns-1].v);
194 d2 = b - b1;
195 d2 *= d2*normf/(b + b1);
196 ep[0] += d2;
197 ep[-hp->ns-1] += d2;
198 }
199 /* correct for number of neighbors */
200 earr[0] *= 8./3.;
201 earr[hp->ns-1] *= 8./3.;
202 earr[(hp->ns-1)*hp->ns] *= 8./3.;
203 earr[(hp->ns-1)*hp->ns + hp->ns-1] *= 8./3.;
204 for (i = 1; i < hp->ns-1; i++) {
205 earr[i*hp->ns] *= 8./5.;
206 earr[i*hp->ns + hp->ns-1] *= 8./5.;
207 }
208 for (j = 1; j < hp->ns-1; j++) {
209 earr[j] *= 8./5.;
210 earr[(hp->ns-1)*hp->ns + j] *= 8./5.;
211 }
212 return(earr);
213 }
214
215
216 /* Perform super-sampling on hemisphere (introduces bias) */
217 static void
218 ambsupersamp(AMBHEMI *hp, int cnt)
219 {
220 float *earr = getambdiffs(hp);
221 double e2rem = 0;
222 float *ep;
223 int i, j, n, nss;
224
225 if (earr == NULL) /* just skip calc. if no memory */
226 return;
227 /* accumulate estimated variances */
228 for (ep = earr + hp->ns*hp->ns; ep > earr; )
229 e2rem += *--ep;
230 ep = earr; /* perform super-sampling */
231 for (i = 0; i < hp->ns; i++)
232 for (j = 0; j < hp->ns; j++) {
233 if (e2rem <= FTINY)
234 goto done; /* nothing left to do */
235 nss = *ep/e2rem*cnt + frandom();
236 for (n = 1; n <= nss && ambsample(hp,i,j,n); n++)
237 if (!--cnt) goto done;
238 e2rem -= *ep++; /* update remainder */
239 }
240 done:
241 free(earr);
242 }
243
244
245 static AMBHEMI *
246 samp_hemi( /* sample indirect hemisphere */
247 COLOR rcol,
248 RAY *r,
249 double wt
250 )
251 {
252 AMBHEMI *hp;
253 double d;
254 int n, i, j;
255 /* insignificance check */
256 if (bright(rcol) <= FTINY)
257 return(NULL);
258 /* set number of divisions */
259 if (ambacc <= FTINY &&
260 wt > (d = 0.8*intens(rcol)*r->rweight/(ambdiv*minweight)))
261 wt = d; /* avoid ray termination */
262 n = sqrt(ambdiv * wt) + 0.5;
263 i = 1 + 5*(ambacc > FTINY); /* minimum number of samples */
264 if (n < i)
265 n = i;
266 /* allocate sampling array */
267 hp = (AMBHEMI *)malloc(sizeof(AMBHEMI) + sizeof(AMBSAMP)*(n*n - 1));
268 if (hp == NULL)
269 error(SYSTEM, "out of memory in samp_hemi");
270 hp->rp = r;
271 hp->ns = n;
272 hp->acol[RED] = hp->acol[GRN] = hp->acol[BLU] = 0.0;
273 memset(hp->sa, 0, sizeof(AMBSAMP)*n*n);
274 hp->sampOK = 0;
275 /* assign coefficient */
276 copycolor(hp->acoef, rcol);
277 d = 1.0/(n*n);
278 scalecolor(hp->acoef, d);
279 /* make tangent plane axes */
280 if (!getperpendicular(hp->ux, r->ron, 1))
281 error(CONSISTENCY, "bad ray direction in samp_hemi");
282 VCROSS(hp->uy, r->ron, hp->ux);
283 /* sample divisions */
284 for (i = hp->ns; i--; )
285 for (j = hp->ns; j--; )
286 hp->sampOK += ambsample(hp, i, j, 0);
287 copycolor(rcol, hp->acol);
288 if (!hp->sampOK) { /* utter failure? */
289 free(hp);
290 return(NULL);
291 }
292 if (hp->sampOK < hp->ns*hp->ns) {
293 hp->sampOK *= -1; /* soft failure */
294 return(hp);
295 }
296 n = ambssamp*wt + 0.5;
297 if (n > 8) { /* perform super-sampling? */
298 ambsupersamp(hp, n);
299 copycolor(rcol, hp->acol);
300 }
301 return(hp); /* all is well */
302 }
303
304
305 /* Return brightness of farthest ambient sample */
306 static double
307 back_ambval(AMBHEMI *hp, const int n1, const int n2, const int n3)
308 {
309 if (hp->sa[n1].d <= hp->sa[n2].d) {
310 if (hp->sa[n1].d <= hp->sa[n3].d)
311 return(colval(hp->sa[n1].v,CIEY));
312 return(colval(hp->sa[n3].v,CIEY));
313 }
314 if (hp->sa[n2].d <= hp->sa[n3].d)
315 return(colval(hp->sa[n2].v,CIEY));
316 return(colval(hp->sa[n3].v,CIEY));
317 }
318
319
320 /* Compute vectors and coefficients for Hessian/gradient calcs */
321 static void
322 comp_fftri(FFTRI *ftp, AMBHEMI *hp, const int n0, const int n1)
323 {
324 double rdot_cp, dot_e, dot_er, rdot_r, rdot_r1, J2;
325 int ii;
326
327 VSUB(ftp->r_i, hp->sa[n0].p, hp->rp->rop);
328 VSUB(ftp->r_i1, hp->sa[n1].p, hp->rp->rop);
329 VSUB(ftp->e_i, hp->sa[n1].p, hp->sa[n0].p);
330 VCROSS(ftp->rcp, ftp->r_i, ftp->r_i1);
331 rdot_cp = 1.0/DOT(ftp->rcp,ftp->rcp);
332 dot_e = DOT(ftp->e_i,ftp->e_i);
333 dot_er = DOT(ftp->e_i, ftp->r_i);
334 rdot_r = 1.0/DOT(ftp->r_i,ftp->r_i);
335 rdot_r1 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
336 ftp->I1 = acos( DOT(ftp->r_i, ftp->r_i1) * sqrt(rdot_r*rdot_r1) ) *
337 sqrt( rdot_cp );
338 ftp->I2 = ( DOT(ftp->e_i, ftp->r_i1)*rdot_r1 - dot_er*rdot_r +
339 dot_e*ftp->I1 )*0.5*rdot_cp;
340 J2 = ( 0.5*(rdot_r - rdot_r1) - dot_er*ftp->I2 ) / dot_e;
341 for (ii = 3; ii--; )
342 ftp->rI2_eJ2[ii] = ftp->I2*ftp->r_i[ii] + J2*ftp->e_i[ii];
343 }
344
345
346 /* Compose 3x3 matrix from two vectors */
347 static void
348 compose_matrix(FVECT mat[3], FVECT va, FVECT vb)
349 {
350 mat[0][0] = 2.0*va[0]*vb[0];
351 mat[1][1] = 2.0*va[1]*vb[1];
352 mat[2][2] = 2.0*va[2]*vb[2];
353 mat[0][1] = mat[1][0] = va[0]*vb[1] + va[1]*vb[0];
354 mat[0][2] = mat[2][0] = va[0]*vb[2] + va[2]*vb[0];
355 mat[1][2] = mat[2][1] = va[1]*vb[2] + va[2]*vb[1];
356 }
357
358
359 /* Compute partial 3x3 Hessian matrix for edge */
360 static void
361 comp_hessian(FVECT hess[3], FFTRI *ftp, FVECT nrm)
362 {
363 FVECT ncp;
364 FVECT m1[3], m2[3], m3[3], m4[3];
365 double d1, d2, d3, d4;
366 double I3, J3, K3;
367 int i, j;
368 /* compute intermediate coefficients */
369 d1 = 1.0/DOT(ftp->r_i,ftp->r_i);
370 d2 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
371 d3 = 1.0/DOT(ftp->e_i,ftp->e_i);
372 d4 = DOT(ftp->e_i, ftp->r_i);
373 I3 = ( DOT(ftp->e_i, ftp->r_i1)*d2*d2 - d4*d1*d1 + 3.0/d3*ftp->I2 )
374 / ( 4.0*DOT(ftp->rcp,ftp->rcp) );
375 J3 = 0.25*d3*(d1*d1 - d2*d2) - d4*d3*I3;
376 K3 = d3*(ftp->I2 - I3/d1 - 2.0*d4*J3);
377 /* intermediate matrices */
378 VCROSS(ncp, nrm, ftp->e_i);
379 compose_matrix(m1, ncp, ftp->rI2_eJ2);
380 compose_matrix(m2, ftp->r_i, ftp->r_i);
381 compose_matrix(m3, ftp->e_i, ftp->e_i);
382 compose_matrix(m4, ftp->r_i, ftp->e_i);
383 d1 = DOT(nrm, ftp->rcp);
384 d2 = -d1*ftp->I2;
385 d1 *= 2.0;
386 for (i = 3; i--; ) /* final matrix sum */
387 for (j = 3; j--; ) {
388 hess[i][j] = m1[i][j] + d1*( I3*m2[i][j] + K3*m3[i][j] +
389 2.0*J3*m4[i][j] );
390 hess[i][j] += d2*(i==j);
391 hess[i][j] *= -1.0/PI;
392 }
393 }
394
395
396 /* Reverse hessian calculation result for edge in other direction */
397 static void
398 rev_hessian(FVECT hess[3])
399 {
400 int i;
401
402 for (i = 3; i--; ) {
403 hess[i][0] = -hess[i][0];
404 hess[i][1] = -hess[i][1];
405 hess[i][2] = -hess[i][2];
406 }
407 }
408
409
410 /* Add to radiometric Hessian from the given triangle */
411 static void
412 add2hessian(FVECT hess[3], FVECT ehess1[3],
413 FVECT ehess2[3], FVECT ehess3[3], double v)
414 {
415 int i, j;
416
417 for (i = 3; i--; )
418 for (j = 3; j--; )
419 hess[i][j] += v*( ehess1[i][j] + ehess2[i][j] + ehess3[i][j] );
420 }
421
422
423 /* Compute partial displacement form factor gradient for edge */
424 static void
425 comp_gradient(FVECT grad, FFTRI *ftp, FVECT nrm)
426 {
427 FVECT ncp;
428 double f1;
429 int i;
430
431 f1 = 2.0*DOT(nrm, ftp->rcp);
432 VCROSS(ncp, nrm, ftp->e_i);
433 for (i = 3; i--; )
434 grad[i] = (0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
435 }
436
437
438 /* Reverse gradient calculation result for edge in other direction */
439 static void
440 rev_gradient(FVECT grad)
441 {
442 grad[0] = -grad[0];
443 grad[1] = -grad[1];
444 grad[2] = -grad[2];
445 }
446
447
448 /* Add to displacement gradient from the given triangle */
449 static void
450 add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, double v)
451 {
452 int i;
453
454 for (i = 3; i--; )
455 grad[i] += v*( egrad1[i] + egrad2[i] + egrad3[i] );
456 }
457
458
459 /* Compute anisotropic radii and eigenvector directions */
460 static void
461 eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3])
462 {
463 double hess2[2][2];
464 FVECT a, b;
465 double evalue[2], slope1, xmag1;
466 int i;
467 /* project Hessian to sample plane */
468 for (i = 3; i--; ) {
469 a[i] = DOT(hessian[i], uv[0]);
470 b[i] = DOT(hessian[i], uv[1]);
471 }
472 hess2[0][0] = DOT(uv[0], a);
473 hess2[0][1] = DOT(uv[0], b);
474 hess2[1][0] = DOT(uv[1], a);
475 hess2[1][1] = DOT(uv[1], b);
476 /* compute eigenvalue(s) */
477 i = quadratic(evalue, 1.0, -hess2[0][0]-hess2[1][1],
478 hess2[0][0]*hess2[1][1]-hess2[0][1]*hess2[1][0]);
479 if (i == 1) /* double-root (circle) */
480 evalue[1] = evalue[0];
481 if (!i || ((evalue[0] = fabs(evalue[0])) <= FTINY*FTINY) |
482 ((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) ) {
483 ra[0] = ra[1] = maxarad;
484 return;
485 }
486 if (evalue[0] > evalue[1]) {
487 ra[0] = sqrt(sqrt(4.0/evalue[0]));
488 ra[1] = sqrt(sqrt(4.0/evalue[1]));
489 slope1 = evalue[1];
490 } else {
491 ra[0] = sqrt(sqrt(4.0/evalue[1]));
492 ra[1] = sqrt(sqrt(4.0/evalue[0]));
493 slope1 = evalue[0];
494 }
495 /* compute unit eigenvectors */
496 if (fabs(hess2[0][1]) <= FTINY)
497 return; /* uv OK as is */
498 slope1 = (slope1 - hess2[0][0]) / hess2[0][1];
499 xmag1 = sqrt(1.0/(1.0 + slope1*slope1));
500 for (i = 3; i--; ) {
501 b[i] = xmag1*uv[0][i] + slope1*xmag1*uv[1][i];
502 a[i] = slope1*xmag1*uv[0][i] - xmag1*uv[1][i];
503 }
504 VCOPY(uv[0], a);
505 VCOPY(uv[1], b);
506 }
507
508
509 static void
510 ambHessian( /* anisotropic radii & pos. gradient */
511 AMBHEMI *hp,
512 FVECT uv[2], /* returned */
513 float ra[2], /* returned (optional) */
514 float pg[2] /* returned (optional) */
515 )
516 {
517 static char memerrmsg[] = "out of memory in ambHessian()";
518 FVECT (*hessrow)[3] = NULL;
519 FVECT *gradrow = NULL;
520 FVECT hessian[3];
521 FVECT gradient;
522 FFTRI fftr;
523 int i, j;
524 /* be sure to assign unit vectors */
525 VCOPY(uv[0], hp->ux);
526 VCOPY(uv[1], hp->uy);
527 /* clock-wise vertex traversal from sample POV */
528 if (ra != NULL) { /* initialize Hessian row buffer */
529 hessrow = (FVECT (*)[3])malloc(sizeof(FVECT)*3*(hp->ns-1));
530 if (hessrow == NULL)
531 error(SYSTEM, memerrmsg);
532 memset(hessian, 0, sizeof(hessian));
533 } else if (pg == NULL) /* bogus call? */
534 return;
535 if (pg != NULL) { /* initialize form factor row buffer */
536 gradrow = (FVECT *)malloc(sizeof(FVECT)*(hp->ns-1));
537 if (gradrow == NULL)
538 error(SYSTEM, memerrmsg);
539 memset(gradient, 0, sizeof(gradient));
540 }
541 /* compute first row of edges */
542 for (j = 0; j < hp->ns-1; j++) {
543 comp_fftri(&fftr, hp, AI(hp,0,j), AI(hp,0,j+1));
544 if (hessrow != NULL)
545 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
546 if (gradrow != NULL)
547 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
548 }
549 /* sum each row of triangles */
550 for (i = 0; i < hp->ns-1; i++) {
551 FVECT hesscol[3]; /* compute first vertical edge */
552 FVECT gradcol;
553 comp_fftri(&fftr, hp, AI(hp,i,0), AI(hp,i+1,0));
554 if (hessrow != NULL)
555 comp_hessian(hesscol, &fftr, hp->rp->ron);
556 if (gradrow != NULL)
557 comp_gradient(gradcol, &fftr, hp->rp->ron);
558 for (j = 0; j < hp->ns-1; j++) {
559 FVECT hessdia[3]; /* compute triangle contributions */
560 FVECT graddia;
561 double backg;
562 backg = back_ambval(hp, AI(hp,i,j),
563 AI(hp,i,j+1), AI(hp,i+1,j));
564 /* diagonal (inner) edge */
565 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j));
566 if (hessrow != NULL) {
567 comp_hessian(hessdia, &fftr, hp->rp->ron);
568 rev_hessian(hesscol);
569 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
570 }
571 if (gradrow != NULL) {
572 comp_gradient(graddia, &fftr, hp->rp->ron);
573 rev_gradient(gradcol);
574 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
575 }
576 /* initialize edge in next row */
577 comp_fftri(&fftr, hp, AI(hp,i+1,j+1), AI(hp,i+1,j));
578 if (hessrow != NULL)
579 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
580 if (gradrow != NULL)
581 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
582 /* new column edge & paired triangle */
583 backg = back_ambval(hp, AI(hp,i+1,j+1),
584 AI(hp,i+1,j), AI(hp,i,j+1));
585 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j+1));
586 if (hessrow != NULL) {
587 comp_hessian(hesscol, &fftr, hp->rp->ron);
588 rev_hessian(hessdia);
589 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
590 if (i < hp->ns-2)
591 rev_hessian(hessrow[j]);
592 }
593 if (gradrow != NULL) {
594 comp_gradient(gradcol, &fftr, hp->rp->ron);
595 rev_gradient(graddia);
596 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
597 if (i < hp->ns-2)
598 rev_gradient(gradrow[j]);
599 }
600 }
601 }
602 /* release row buffers */
603 if (hessrow != NULL) free(hessrow);
604 if (gradrow != NULL) free(gradrow);
605
606 if (ra != NULL) /* extract eigenvectors & radii */
607 eigenvectors(uv, ra, hessian);
608 if (pg != NULL) { /* tangential position gradient */
609 pg[0] = DOT(gradient, uv[0]);
610 pg[1] = DOT(gradient, uv[1]);
611 }
612 }
613
614
615 /* Compute direction gradient from a hemispherical sampling */
616 static void
617 ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
618 {
619 AMBSAMP *ap;
620 double dgsum[2];
621 int n;
622 FVECT vd;
623 double gfact;
624
625 dgsum[0] = dgsum[1] = 0.0; /* sum values times -tan(theta) */
626 for (ap = hp->sa, n = hp->ns*hp->ns; n--; ap++) {
627 /* use vector for azimuth + 90deg */
628 VSUB(vd, ap->p, hp->rp->rop);
629 /* brightness over cosine factor */
630 gfact = colval(ap->v,CIEY) / DOT(hp->rp->ron, vd);
631 /* sine = proj_radius/vd_length */
632 dgsum[0] -= DOT(uv[1], vd) * gfact;
633 dgsum[1] += DOT(uv[0], vd) * gfact;
634 }
635 dg[0] = dgsum[0] / (hp->ns*hp->ns);
636 dg[1] = dgsum[1] / (hp->ns*hp->ns);
637 }
638
639
640 /* Compute potential light leak direction flags for cache value */
641 static uint32
642 ambcorral(AMBHEMI *hp, FVECT uv[2], const double r0, const double r1)
643 {
644 const double max_d = 1.0/(minarad*ambacc + 0.001);
645 const double ang_res = 0.5*PI/hp->ns;
646 const double ang_step = ang_res/((int)(16/PI*ang_res) + 1.01);
647 double avg_d = 0;
648 uint32 flgs = 0;
649 FVECT vec;
650 double u, v;
651 double ang, a1;
652 int i, j;
653 /* don't bother for a few samples */
654 if (hp->ns < 8)
655 return(0);
656 /* check distances overhead */
657 for (i = hp->ns*3/4; i-- > hp->ns>>2; )
658 for (j = hp->ns*3/4; j-- > hp->ns>>2; )
659 avg_d += ambsam(hp,i,j).d;
660 avg_d *= 4.0/(hp->ns*hp->ns);
661 if (avg_d*r0 >= 1.0) /* ceiling too low for corral? */
662 return(0);
663 if (avg_d >= max_d) /* insurance */
664 return(0);
665 /* else circle around perimeter */
666 for (i = 0; i < hp->ns; i++)
667 for (j = 0; j < hp->ns; j += !i|(i==hp->ns-1) ? 1 : hp->ns-1) {
668 AMBSAMP *ap = &ambsam(hp,i,j);
669 if ((ap->d <= FTINY) | (ap->d >= max_d))
670 continue; /* too far or too near */
671 VSUB(vec, ap->p, hp->rp->rop);
672 u = DOT(vec, uv[0]);
673 v = DOT(vec, uv[1]);
674 if ((r0*r0*u*u + r1*r1*v*v) * ap->d*ap->d <= u*u + v*v)
675 continue; /* occluder outside ellipse */
676 ang = atan2a(v, u); /* else set direction flags */
677 for (a1 = ang-ang_res; a1 <= ang+ang_res; a1 += ang_step)
678 flgs |= 1L<<(int)(16/PI*(a1 + 2.*PI*(a1 < 0)));
679 }
680 return(flgs);
681 }
682
683
684 int
685 doambient( /* compute ambient component */
686 COLOR rcol, /* input/output color */
687 RAY *r,
688 double wt,
689 FVECT uv[2], /* returned (optional) */
690 float ra[2], /* returned (optional) */
691 float pg[2], /* returned (optional) */
692 float dg[2], /* returned (optional) */
693 uint32 *crlp /* returned (optional) */
694 )
695 {
696 AMBHEMI *hp = samp_hemi(rcol, r, wt);
697 FVECT my_uv[2];
698 double d, K;
699 AMBSAMP *ap;
700 int i;
701 /* clear return values */
702 if (uv != NULL)
703 memset(uv, 0, sizeof(FVECT)*2);
704 if (ra != NULL)
705 ra[0] = ra[1] = 0.0;
706 if (pg != NULL)
707 pg[0] = pg[1] = 0.0;
708 if (dg != NULL)
709 dg[0] = dg[1] = 0.0;
710 if (crlp != NULL)
711 *crlp = 0;
712 if (hp == NULL) /* sampling falure? */
713 return(0);
714
715 if ((ra == NULL) & (pg == NULL) & (dg == NULL) ||
716 (hp->sampOK < 0) | (hp->ns < 6)) {
717 free(hp); /* Hessian not requested/possible */
718 return(-1); /* value-only return value */
719 }
720 if ((d = bright(rcol)) > FTINY) { /* normalize Y values */
721 d = 0.99*(hp->ns*hp->ns)/d;
722 K = 0.01;
723 } else { /* or fall back on geometric Hessian */
724 K = 1.0;
725 pg = NULL;
726 dg = NULL;
727 crlp = NULL;
728 }
729 ap = hp->sa; /* relative Y channel from here on... */
730 for (i = hp->ns*hp->ns; i--; ap++)
731 colval(ap->v,CIEY) = bright(ap->v)*d + K;
732
733 if (uv == NULL) /* make sure we have axis pointers */
734 uv = my_uv;
735 /* compute radii & pos. gradient */
736 ambHessian(hp, uv, ra, pg);
737
738 if (dg != NULL) /* compute direction gradient */
739 ambdirgrad(hp, uv, dg);
740
741 if (ra != NULL) { /* scale/clamp radii */
742 if (pg != NULL) {
743 if (ra[0]*(d = fabs(pg[0])) > 1.0)
744 ra[0] = 1.0/d;
745 if (ra[1]*(d = fabs(pg[1])) > 1.0)
746 ra[1] = 1.0/d;
747 if (ra[0] > ra[1])
748 ra[0] = ra[1];
749 }
750 if (ra[0] < minarad) {
751 ra[0] = minarad;
752 if (ra[1] < minarad)
753 ra[1] = minarad;
754 }
755 ra[0] *= d = 1.0/sqrt(wt);
756 if ((ra[1] *= d) > 2.0*ra[0])
757 ra[1] = 2.0*ra[0];
758 if (ra[1] > maxarad) {
759 ra[1] = maxarad;
760 if (ra[0] > maxarad)
761 ra[0] = maxarad;
762 }
763 /* flag encroached directions */
764 if (crlp != NULL)
765 *crlp = ambcorral(hp, uv, ra[0]*ambacc, ra[1]*ambacc);
766 if (pg != NULL) { /* cap gradient if necessary */
767 d = pg[0]*pg[0]*ra[0]*ra[0] + pg[1]*pg[1]*ra[1]*ra[1];
768 if (d > 1.0) {
769 d = 1.0/sqrt(d);
770 pg[0] *= d;
771 pg[1] *= d;
772 }
773 }
774 }
775 free(hp); /* clean up and return */
776 return(1);
777 }
778
779
780 #else /* ! NEWAMB */
781
782
783 void
784 inithemi( /* initialize sampling hemisphere */
785 AMBHEMI *hp,
786 COLOR ac,
787 RAY *r,
788 double wt
789 )
790 {
791 double d;
792 int i;
793 /* set number of divisions */
794 if (ambacc <= FTINY &&
795 wt > (d = 0.8*intens(ac)*r->rweight/(ambdiv*minweight)))
796 wt = d; /* avoid ray termination */
797 hp->nt = sqrt(ambdiv * wt / PI) + 0.5;
798 i = ambacc > FTINY ? 3 : 1; /* minimum number of samples */
799 if (hp->nt < i)
800 hp->nt = i;
801 hp->np = PI * hp->nt + 0.5;
802 /* set number of super-samples */
803 hp->ns = ambssamp * wt + 0.5;
804 /* assign coefficient */
805 copycolor(hp->acoef, ac);
806 d = 1.0/(hp->nt*hp->np);
807 scalecolor(hp->acoef, d);
808 /* make axes */
809 VCOPY(hp->uz, r->ron);
810 hp->uy[0] = hp->uy[1] = hp->uy[2] = 0.0;
811 for (i = 0; i < 3; i++)
812 if (hp->uz[i] < 0.6 && hp->uz[i] > -0.6)
813 break;
814 if (i >= 3)
815 error(CONSISTENCY, "bad ray direction in inithemi");
816 hp->uy[i] = 1.0;
817 fcross(hp->ux, hp->uy, hp->uz);
818 normalize(hp->ux);
819 fcross(hp->uy, hp->uz, hp->ux);
820 }
821
822
823 int
824 divsample( /* sample a division */
825 AMBSAMP *dp,
826 AMBHEMI *h,
827 RAY *r
828 )
829 {
830 RAY ar;
831 int hlist[3];
832 double spt[2];
833 double xd, yd, zd;
834 double b2;
835 double phi;
836 int i;
837 /* ambient coefficient for weight */
838 if (ambacc > FTINY)
839 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
840 else
841 copycolor(ar.rcoef, h->acoef);
842 if (rayorigin(&ar, AMBIENT, r, ar.rcoef) < 0)
843 return(-1);
844 if (ambacc > FTINY) {
845 multcolor(ar.rcoef, h->acoef);
846 scalecolor(ar.rcoef, 1./AVGREFL);
847 }
848 hlist[0] = r->rno;
849 hlist[1] = dp->t;
850 hlist[2] = dp->p;
851 multisamp(spt, 2, urand(ilhash(hlist,3)+dp->n));
852 zd = sqrt((dp->t + spt[0])/h->nt);
853 phi = 2.0*PI * (dp->p + spt[1])/h->np;
854 xd = tcos(phi) * zd;
855 yd = tsin(phi) * zd;
856 zd = sqrt(1.0 - zd*zd);
857 for (i = 0; i < 3; i++)
858 ar.rdir[i] = xd*h->ux[i] +
859 yd*h->uy[i] +
860 zd*h->uz[i];
861 checknorm(ar.rdir);
862 dimlist[ndims++] = dp->t*h->np + dp->p + 90171;
863 rayvalue(&ar);
864 ndims--;
865 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
866 addcolor(dp->v, ar.rcol);
867 /* use rxt to improve gradient calc */
868 if (ar.rxt > FTINY && ar.rxt < FHUGE)
869 dp->r += 1.0/ar.rxt;
870 /* (re)initialize error */
871 if (dp->n++) {
872 b2 = bright(dp->v)/dp->n - bright(ar.rcol);
873 b2 = b2*b2 + dp->k*((dp->n-1)*(dp->n-1));
874 dp->k = b2/(dp->n*dp->n);
875 } else
876 dp->k = 0.0;
877 return(0);
878 }
879
880
881 static int
882 ambcmp( /* decreasing order */
883 const void *p1,
884 const void *p2
885 )
886 {
887 const AMBSAMP *d1 = (const AMBSAMP *)p1;
888 const AMBSAMP *d2 = (const AMBSAMP *)p2;
889
890 if (d1->k < d2->k)
891 return(1);
892 if (d1->k > d2->k)
893 return(-1);
894 return(0);
895 }
896
897
898 static int
899 ambnorm( /* standard order */
900 const void *p1,
901 const void *p2
902 )
903 {
904 const AMBSAMP *d1 = (const AMBSAMP *)p1;
905 const AMBSAMP *d2 = (const AMBSAMP *)p2;
906 int c;
907
908 if ( (c = d1->t - d2->t) )
909 return(c);
910 return(d1->p - d2->p);
911 }
912
913
914 double
915 doambient( /* compute ambient component */
916 COLOR rcol,
917 RAY *r,
918 double wt,
919 FVECT pg,
920 FVECT dg
921 )
922 {
923 double b, d=0;
924 AMBHEMI hemi;
925 AMBSAMP *div;
926 AMBSAMP dnew;
927 double acol[3];
928 AMBSAMP *dp;
929 double arad;
930 int divcnt;
931 int i, j;
932 /* initialize hemisphere */
933 inithemi(&hemi, rcol, r, wt);
934 divcnt = hemi.nt * hemi.np;
935 /* initialize */
936 if (pg != NULL)
937 pg[0] = pg[1] = pg[2] = 0.0;
938 if (dg != NULL)
939 dg[0] = dg[1] = dg[2] = 0.0;
940 setcolor(rcol, 0.0, 0.0, 0.0);
941 if (divcnt == 0)
942 return(0.0);
943 /* allocate super-samples */
944 if (hemi.ns > 0 || pg != NULL || dg != NULL) {
945 div = (AMBSAMP *)malloc(divcnt*sizeof(AMBSAMP));
946 if (div == NULL)
947 error(SYSTEM, "out of memory in doambient");
948 } else
949 div = NULL;
950 /* sample the divisions */
951 arad = 0.0;
952 acol[0] = acol[1] = acol[2] = 0.0;
953 if ((dp = div) == NULL)
954 dp = &dnew;
955 divcnt = 0;
956 for (i = 0; i < hemi.nt; i++)
957 for (j = 0; j < hemi.np; j++) {
958 dp->t = i; dp->p = j;
959 setcolor(dp->v, 0.0, 0.0, 0.0);
960 dp->r = 0.0;
961 dp->n = 0;
962 if (divsample(dp, &hemi, r) < 0) {
963 if (div != NULL)
964 dp++;
965 continue;
966 }
967 arad += dp->r;
968 divcnt++;
969 if (div != NULL)
970 dp++;
971 else
972 addcolor(acol, dp->v);
973 }
974 if (!divcnt) {
975 if (div != NULL)
976 free((void *)div);
977 return(0.0); /* no samples taken */
978 }
979 if (divcnt < hemi.nt*hemi.np) {
980 pg = dg = NULL; /* incomplete sampling */
981 hemi.ns = 0;
982 } else if (arad > FTINY && divcnt/arad < minarad) {
983 hemi.ns = 0; /* close enough */
984 } else if (hemi.ns > 0) { /* else perform super-sampling? */
985 comperrs(div, &hemi); /* compute errors */
986 qsort(div, divcnt, sizeof(AMBSAMP), ambcmp); /* sort divs */
987 /* super-sample */
988 for (i = hemi.ns; i > 0; i--) {
989 dnew = *div;
990 if (divsample(&dnew, &hemi, r) < 0) {
991 dp++;
992 continue;
993 }
994 dp = div; /* reinsert */
995 j = divcnt < i ? divcnt : i;
996 while (--j > 0 && dnew.k < dp[1].k) {
997 *dp = *(dp+1);
998 dp++;
999 }
1000 *dp = dnew;
1001 }
1002 if (pg != NULL || dg != NULL) /* restore order */
1003 qsort(div, divcnt, sizeof(AMBSAMP), ambnorm);
1004 }
1005 /* compute returned values */
1006 if (div != NULL) {
1007 arad = 0.0; /* note: divcnt may be < nt*np */
1008 for (i = hemi.nt*hemi.np, dp = div; i-- > 0; dp++) {
1009 arad += dp->r;
1010 if (dp->n > 1) {
1011 b = 1.0/dp->n;
1012 scalecolor(dp->v, b);
1013 dp->r *= b;
1014 dp->n = 1;
1015 }
1016 addcolor(acol, dp->v);
1017 }
1018 b = bright(acol);
1019 if (b > FTINY) {
1020 b = 1.0/b; /* compute & normalize gradient(s) */
1021 if (pg != NULL) {
1022 posgradient(pg, div, &hemi);
1023 for (i = 0; i < 3; i++)
1024 pg[i] *= b;
1025 }
1026 if (dg != NULL) {
1027 dirgradient(dg, div, &hemi);
1028 for (i = 0; i < 3; i++)
1029 dg[i] *= b;
1030 }
1031 }
1032 free((void *)div);
1033 }
1034 copycolor(rcol, acol);
1035 if (arad <= FTINY)
1036 arad = maxarad;
1037 else
1038 arad = (divcnt+hemi.ns)/arad;
1039 if (pg != NULL) { /* reduce radius if gradient large */
1040 d = DOT(pg,pg);
1041 if (d*arad*arad > 1.0)
1042 arad = 1.0/sqrt(d);
1043 }
1044 if (arad < minarad) {
1045 arad = minarad;
1046 if (pg != NULL && d*arad*arad > 1.0) { /* cap gradient */
1047 d = 1.0/arad/sqrt(d);
1048 for (i = 0; i < 3; i++)
1049 pg[i] *= d;
1050 }
1051 }
1052 if ((arad /= sqrt(wt)) > maxarad)
1053 arad = maxarad;
1054 return(arad);
1055 }
1056
1057
1058 void
1059 comperrs( /* compute initial error estimates */
1060 AMBSAMP *da, /* assumes standard ordering */
1061 AMBHEMI *hp
1062 )
1063 {
1064 double b, b2;
1065 int i, j;
1066 AMBSAMP *dp;
1067 /* sum differences from neighbors */
1068 dp = da;
1069 for (i = 0; i < hp->nt; i++)
1070 for (j = 0; j < hp->np; j++) {
1071 #ifdef DEBUG
1072 if (dp->t != i || dp->p != j)
1073 error(CONSISTENCY,
1074 "division order in comperrs");
1075 #endif
1076 b = bright(dp[0].v);
1077 if (i > 0) { /* from above */
1078 b2 = bright(dp[-hp->np].v) - b;
1079 b2 *= b2 * 0.25;
1080 dp[0].k += b2;
1081 dp[-hp->np].k += b2;
1082 }
1083 if (j > 0) { /* from behind */
1084 b2 = bright(dp[-1].v) - b;
1085 b2 *= b2 * 0.25;
1086 dp[0].k += b2;
1087 dp[-1].k += b2;
1088 } else { /* around */
1089 b2 = bright(dp[hp->np-1].v) - b;
1090 b2 *= b2 * 0.25;
1091 dp[0].k += b2;
1092 dp[hp->np-1].k += b2;
1093 }
1094 dp++;
1095 }
1096 /* divide by number of neighbors */
1097 dp = da;
1098 for (j = 0; j < hp->np; j++) /* top row */
1099 (dp++)->k *= 1.0/3.0;
1100 if (hp->nt < 2)
1101 return;
1102 for (i = 1; i < hp->nt-1; i++) /* central region */
1103 for (j = 0; j < hp->np; j++)
1104 (dp++)->k *= 0.25;
1105 for (j = 0; j < hp->np; j++) /* bottom row */
1106 (dp++)->k *= 1.0/3.0;
1107 }
1108
1109
1110 void
1111 posgradient( /* compute position gradient */
1112 FVECT gv,
1113 AMBSAMP *da, /* assumes standard ordering */
1114 AMBHEMI *hp
1115 )
1116 {
1117 int i, j;
1118 double nextsine, lastsine, b, d;
1119 double mag0, mag1;
1120 double phi, cosp, sinp, xd, yd;
1121 AMBSAMP *dp;
1122
1123 xd = yd = 0.0;
1124 for (j = 0; j < hp->np; j++) {
1125 dp = da + j;
1126 mag0 = mag1 = 0.0;
1127 lastsine = 0.0;
1128 for (i = 0; i < hp->nt; i++) {
1129 #ifdef DEBUG
1130 if (dp->t != i || dp->p != j)
1131 error(CONSISTENCY,
1132 "division order in posgradient");
1133 #endif
1134 b = bright(dp->v);
1135 if (i > 0) {
1136 d = dp[-hp->np].r;
1137 if (dp[0].r > d) d = dp[0].r;
1138 /* sin(t)*cos(t)^2 */
1139 d *= lastsine * (1.0 - (double)i/hp->nt);
1140 mag0 += d*(b - bright(dp[-hp->np].v));
1141 }
1142 nextsine = sqrt((double)(i+1)/hp->nt);
1143 if (j > 0) {
1144 d = dp[-1].r;
1145 if (dp[0].r > d) d = dp[0].r;
1146 mag1 += d * (nextsine - lastsine) *
1147 (b - bright(dp[-1].v));
1148 } else {
1149 d = dp[hp->np-1].r;
1150 if (dp[0].r > d) d = dp[0].r;
1151 mag1 += d * (nextsine - lastsine) *
1152 (b - bright(dp[hp->np-1].v));
1153 }
1154 dp += hp->np;
1155 lastsine = nextsine;
1156 }
1157 mag0 *= 2.0*PI / hp->np;
1158 phi = 2.0*PI * (double)j/hp->np;
1159 cosp = tcos(phi); sinp = tsin(phi);
1160 xd += mag0*cosp - mag1*sinp;
1161 yd += mag0*sinp + mag1*cosp;
1162 }
1163 for (i = 0; i < 3; i++)
1164 gv[i] = (xd*hp->ux[i] + yd*hp->uy[i])*(hp->nt*hp->np)/PI;
1165 }
1166
1167
1168 void
1169 dirgradient( /* compute direction gradient */
1170 FVECT gv,
1171 AMBSAMP *da, /* assumes standard ordering */
1172 AMBHEMI *hp
1173 )
1174 {
1175 int i, j;
1176 double mag;
1177 double phi, xd, yd;
1178 AMBSAMP *dp;
1179
1180 xd = yd = 0.0;
1181 for (j = 0; j < hp->np; j++) {
1182 dp = da + j;
1183 mag = 0.0;
1184 for (i = 0; i < hp->nt; i++) {
1185 #ifdef DEBUG
1186 if (dp->t != i || dp->p != j)
1187 error(CONSISTENCY,
1188 "division order in dirgradient");
1189 #endif
1190 /* tan(t) */
1191 mag += bright(dp->v)/sqrt(hp->nt/(i+.5) - 1.0);
1192 dp += hp->np;
1193 }
1194 phi = 2.0*PI * (j+.5)/hp->np + PI/2.0;
1195 xd += mag * tcos(phi);
1196 yd += mag * tsin(phi);
1197 }
1198 for (i = 0; i < 3; i++)
1199 gv[i] = xd*hp->ux[i] + yd*hp->uy[i];
1200 }
1201
1202 #endif /* ! NEWAMB */