ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambcomp.c
Revision: 2.84
Committed: Tue Feb 26 00:37:54 2019 UTC (5 years, 2 months ago) by greg
Content type: text/plain
Branch: MAIN
Changes since 2.83: +3 -1 lines
Log Message:
Inserted safety valve to shut off super-sampling when ad drops below 64 samples

File Contents

# Content
1 #ifndef lint
2 static const char RCSid[] = "$Id: ambcomp.c,v 2.83 2018/11/13 19:58:33 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 if (hp->sampOK < 64)
297 return(hp); /* insufficient for super-sampling */
298 n = ambssamp*wt + 0.5;
299 if (n > 8) { /* perform super-sampling? */
300 ambsupersamp(hp, n);
301 copycolor(rcol, hp->acol);
302 }
303 return(hp); /* all is well */
304 }
305
306
307 /* Return brightness of farthest ambient sample */
308 static double
309 back_ambval(AMBHEMI *hp, const int n1, const int n2, const int n3)
310 {
311 if (hp->sa[n1].d <= hp->sa[n2].d) {
312 if (hp->sa[n1].d <= hp->sa[n3].d)
313 return(colval(hp->sa[n1].v,CIEY));
314 return(colval(hp->sa[n3].v,CIEY));
315 }
316 if (hp->sa[n2].d <= hp->sa[n3].d)
317 return(colval(hp->sa[n2].v,CIEY));
318 return(colval(hp->sa[n3].v,CIEY));
319 }
320
321
322 /* Compute vectors and coefficients for Hessian/gradient calcs */
323 static void
324 comp_fftri(FFTRI *ftp, AMBHEMI *hp, const int n0, const int n1)
325 {
326 double rdot_cp, dot_e, dot_er, rdot_r, rdot_r1, J2;
327 int ii;
328
329 VSUB(ftp->r_i, hp->sa[n0].p, hp->rp->rop);
330 VSUB(ftp->r_i1, hp->sa[n1].p, hp->rp->rop);
331 VSUB(ftp->e_i, hp->sa[n1].p, hp->sa[n0].p);
332 VCROSS(ftp->rcp, ftp->r_i, ftp->r_i1);
333 rdot_cp = 1.0/DOT(ftp->rcp,ftp->rcp);
334 dot_e = DOT(ftp->e_i,ftp->e_i);
335 dot_er = DOT(ftp->e_i, ftp->r_i);
336 rdot_r = 1.0/DOT(ftp->r_i,ftp->r_i);
337 rdot_r1 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
338 ftp->I1 = acos( DOT(ftp->r_i, ftp->r_i1) * sqrt(rdot_r*rdot_r1) ) *
339 sqrt( rdot_cp );
340 ftp->I2 = ( DOT(ftp->e_i, ftp->r_i1)*rdot_r1 - dot_er*rdot_r +
341 dot_e*ftp->I1 )*0.5*rdot_cp;
342 J2 = ( 0.5*(rdot_r - rdot_r1) - dot_er*ftp->I2 ) / dot_e;
343 for (ii = 3; ii--; )
344 ftp->rI2_eJ2[ii] = ftp->I2*ftp->r_i[ii] + J2*ftp->e_i[ii];
345 }
346
347
348 /* Compose 3x3 matrix from two vectors */
349 static void
350 compose_matrix(FVECT mat[3], FVECT va, FVECT vb)
351 {
352 mat[0][0] = 2.0*va[0]*vb[0];
353 mat[1][1] = 2.0*va[1]*vb[1];
354 mat[2][2] = 2.0*va[2]*vb[2];
355 mat[0][1] = mat[1][0] = va[0]*vb[1] + va[1]*vb[0];
356 mat[0][2] = mat[2][0] = va[0]*vb[2] + va[2]*vb[0];
357 mat[1][2] = mat[2][1] = va[1]*vb[2] + va[2]*vb[1];
358 }
359
360
361 /* Compute partial 3x3 Hessian matrix for edge */
362 static void
363 comp_hessian(FVECT hess[3], FFTRI *ftp, FVECT nrm)
364 {
365 FVECT ncp;
366 FVECT m1[3], m2[3], m3[3], m4[3];
367 double d1, d2, d3, d4;
368 double I3, J3, K3;
369 int i, j;
370 /* compute intermediate coefficients */
371 d1 = 1.0/DOT(ftp->r_i,ftp->r_i);
372 d2 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
373 d3 = 1.0/DOT(ftp->e_i,ftp->e_i);
374 d4 = DOT(ftp->e_i, ftp->r_i);
375 I3 = ( DOT(ftp->e_i, ftp->r_i1)*d2*d2 - d4*d1*d1 + 3.0/d3*ftp->I2 )
376 / ( 4.0*DOT(ftp->rcp,ftp->rcp) );
377 J3 = 0.25*d3*(d1*d1 - d2*d2) - d4*d3*I3;
378 K3 = d3*(ftp->I2 - I3/d1 - 2.0*d4*J3);
379 /* intermediate matrices */
380 VCROSS(ncp, nrm, ftp->e_i);
381 compose_matrix(m1, ncp, ftp->rI2_eJ2);
382 compose_matrix(m2, ftp->r_i, ftp->r_i);
383 compose_matrix(m3, ftp->e_i, ftp->e_i);
384 compose_matrix(m4, ftp->r_i, ftp->e_i);
385 d1 = DOT(nrm, ftp->rcp);
386 d2 = -d1*ftp->I2;
387 d1 *= 2.0;
388 for (i = 3; i--; ) /* final matrix sum */
389 for (j = 3; j--; ) {
390 hess[i][j] = m1[i][j] + d1*( I3*m2[i][j] + K3*m3[i][j] +
391 2.0*J3*m4[i][j] );
392 hess[i][j] += d2*(i==j);
393 hess[i][j] *= -1.0/PI;
394 }
395 }
396
397
398 /* Reverse hessian calculation result for edge in other direction */
399 static void
400 rev_hessian(FVECT hess[3])
401 {
402 int i;
403
404 for (i = 3; i--; ) {
405 hess[i][0] = -hess[i][0];
406 hess[i][1] = -hess[i][1];
407 hess[i][2] = -hess[i][2];
408 }
409 }
410
411
412 /* Add to radiometric Hessian from the given triangle */
413 static void
414 add2hessian(FVECT hess[3], FVECT ehess1[3],
415 FVECT ehess2[3], FVECT ehess3[3], double v)
416 {
417 int i, j;
418
419 for (i = 3; i--; )
420 for (j = 3; j--; )
421 hess[i][j] += v*( ehess1[i][j] + ehess2[i][j] + ehess3[i][j] );
422 }
423
424
425 /* Compute partial displacement form factor gradient for edge */
426 static void
427 comp_gradient(FVECT grad, FFTRI *ftp, FVECT nrm)
428 {
429 FVECT ncp;
430 double f1;
431 int i;
432
433 f1 = 2.0*DOT(nrm, ftp->rcp);
434 VCROSS(ncp, nrm, ftp->e_i);
435 for (i = 3; i--; )
436 grad[i] = (0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
437 }
438
439
440 /* Reverse gradient calculation result for edge in other direction */
441 static void
442 rev_gradient(FVECT grad)
443 {
444 grad[0] = -grad[0];
445 grad[1] = -grad[1];
446 grad[2] = -grad[2];
447 }
448
449
450 /* Add to displacement gradient from the given triangle */
451 static void
452 add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, double v)
453 {
454 int i;
455
456 for (i = 3; i--; )
457 grad[i] += v*( egrad1[i] + egrad2[i] + egrad3[i] );
458 }
459
460
461 /* Compute anisotropic radii and eigenvector directions */
462 static void
463 eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3])
464 {
465 double hess2[2][2];
466 FVECT a, b;
467 double evalue[2], slope1, xmag1;
468 int i;
469 /* project Hessian to sample plane */
470 for (i = 3; i--; ) {
471 a[i] = DOT(hessian[i], uv[0]);
472 b[i] = DOT(hessian[i], uv[1]);
473 }
474 hess2[0][0] = DOT(uv[0], a);
475 hess2[0][1] = DOT(uv[0], b);
476 hess2[1][0] = DOT(uv[1], a);
477 hess2[1][1] = DOT(uv[1], b);
478 /* compute eigenvalue(s) */
479 i = quadratic(evalue, 1.0, -hess2[0][0]-hess2[1][1],
480 hess2[0][0]*hess2[1][1]-hess2[0][1]*hess2[1][0]);
481 if (i == 1) /* double-root (circle) */
482 evalue[1] = evalue[0];
483 if (!i || ((evalue[0] = fabs(evalue[0])) <= FTINY*FTINY) |
484 ((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) ) {
485 ra[0] = ra[1] = maxarad;
486 return;
487 }
488 if (evalue[0] > evalue[1]) {
489 ra[0] = sqrt(sqrt(4.0/evalue[0]));
490 ra[1] = sqrt(sqrt(4.0/evalue[1]));
491 slope1 = evalue[1];
492 } else {
493 ra[0] = sqrt(sqrt(4.0/evalue[1]));
494 ra[1] = sqrt(sqrt(4.0/evalue[0]));
495 slope1 = evalue[0];
496 }
497 /* compute unit eigenvectors */
498 if (fabs(hess2[0][1]) <= FTINY)
499 return; /* uv OK as is */
500 slope1 = (slope1 - hess2[0][0]) / hess2[0][1];
501 xmag1 = sqrt(1.0/(1.0 + slope1*slope1));
502 for (i = 3; i--; ) {
503 b[i] = xmag1*uv[0][i] + slope1*xmag1*uv[1][i];
504 a[i] = slope1*xmag1*uv[0][i] - xmag1*uv[1][i];
505 }
506 VCOPY(uv[0], a);
507 VCOPY(uv[1], b);
508 }
509
510
511 static void
512 ambHessian( /* anisotropic radii & pos. gradient */
513 AMBHEMI *hp,
514 FVECT uv[2], /* returned */
515 float ra[2], /* returned (optional) */
516 float pg[2] /* returned (optional) */
517 )
518 {
519 static char memerrmsg[] = "out of memory in ambHessian()";
520 FVECT (*hessrow)[3] = NULL;
521 FVECT *gradrow = NULL;
522 FVECT hessian[3];
523 FVECT gradient;
524 FFTRI fftr;
525 int i, j;
526 /* be sure to assign unit vectors */
527 VCOPY(uv[0], hp->ux);
528 VCOPY(uv[1], hp->uy);
529 /* clock-wise vertex traversal from sample POV */
530 if (ra != NULL) { /* initialize Hessian row buffer */
531 hessrow = (FVECT (*)[3])malloc(sizeof(FVECT)*3*(hp->ns-1));
532 if (hessrow == NULL)
533 error(SYSTEM, memerrmsg);
534 memset(hessian, 0, sizeof(hessian));
535 } else if (pg == NULL) /* bogus call? */
536 return;
537 if (pg != NULL) { /* initialize form factor row buffer */
538 gradrow = (FVECT *)malloc(sizeof(FVECT)*(hp->ns-1));
539 if (gradrow == NULL)
540 error(SYSTEM, memerrmsg);
541 memset(gradient, 0, sizeof(gradient));
542 }
543 /* compute first row of edges */
544 for (j = 0; j < hp->ns-1; j++) {
545 comp_fftri(&fftr, hp, AI(hp,0,j), AI(hp,0,j+1));
546 if (hessrow != NULL)
547 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
548 if (gradrow != NULL)
549 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
550 }
551 /* sum each row of triangles */
552 for (i = 0; i < hp->ns-1; i++) {
553 FVECT hesscol[3]; /* compute first vertical edge */
554 FVECT gradcol;
555 comp_fftri(&fftr, hp, AI(hp,i,0), AI(hp,i+1,0));
556 if (hessrow != NULL)
557 comp_hessian(hesscol, &fftr, hp->rp->ron);
558 if (gradrow != NULL)
559 comp_gradient(gradcol, &fftr, hp->rp->ron);
560 for (j = 0; j < hp->ns-1; j++) {
561 FVECT hessdia[3]; /* compute triangle contributions */
562 FVECT graddia;
563 double backg;
564 backg = back_ambval(hp, AI(hp,i,j),
565 AI(hp,i,j+1), AI(hp,i+1,j));
566 /* diagonal (inner) edge */
567 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j));
568 if (hessrow != NULL) {
569 comp_hessian(hessdia, &fftr, hp->rp->ron);
570 rev_hessian(hesscol);
571 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
572 }
573 if (gradrow != NULL) {
574 comp_gradient(graddia, &fftr, hp->rp->ron);
575 rev_gradient(gradcol);
576 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
577 }
578 /* initialize edge in next row */
579 comp_fftri(&fftr, hp, AI(hp,i+1,j+1), AI(hp,i+1,j));
580 if (hessrow != NULL)
581 comp_hessian(hessrow[j], &fftr, hp->rp->ron);
582 if (gradrow != NULL)
583 comp_gradient(gradrow[j], &fftr, hp->rp->ron);
584 /* new column edge & paired triangle */
585 backg = back_ambval(hp, AI(hp,i+1,j+1),
586 AI(hp,i+1,j), AI(hp,i,j+1));
587 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j+1));
588 if (hessrow != NULL) {
589 comp_hessian(hesscol, &fftr, hp->rp->ron);
590 rev_hessian(hessdia);
591 add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
592 if (i < hp->ns-2)
593 rev_hessian(hessrow[j]);
594 }
595 if (gradrow != NULL) {
596 comp_gradient(gradcol, &fftr, hp->rp->ron);
597 rev_gradient(graddia);
598 add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
599 if (i < hp->ns-2)
600 rev_gradient(gradrow[j]);
601 }
602 }
603 }
604 /* release row buffers */
605 if (hessrow != NULL) free(hessrow);
606 if (gradrow != NULL) free(gradrow);
607
608 if (ra != NULL) /* extract eigenvectors & radii */
609 eigenvectors(uv, ra, hessian);
610 if (pg != NULL) { /* tangential position gradient */
611 pg[0] = DOT(gradient, uv[0]);
612 pg[1] = DOT(gradient, uv[1]);
613 }
614 }
615
616
617 /* Compute direction gradient from a hemispherical sampling */
618 static void
619 ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
620 {
621 AMBSAMP *ap;
622 double dgsum[2];
623 int n;
624 FVECT vd;
625 double gfact;
626
627 dgsum[0] = dgsum[1] = 0.0; /* sum values times -tan(theta) */
628 for (ap = hp->sa, n = hp->ns*hp->ns; n--; ap++) {
629 /* use vector for azimuth + 90deg */
630 VSUB(vd, ap->p, hp->rp->rop);
631 /* brightness over cosine factor */
632 gfact = colval(ap->v,CIEY) / DOT(hp->rp->ron, vd);
633 /* sine = proj_radius/vd_length */
634 dgsum[0] -= DOT(uv[1], vd) * gfact;
635 dgsum[1] += DOT(uv[0], vd) * gfact;
636 }
637 dg[0] = dgsum[0] / (hp->ns*hp->ns);
638 dg[1] = dgsum[1] / (hp->ns*hp->ns);
639 }
640
641
642 /* Compute potential light leak direction flags for cache value */
643 static uint32
644 ambcorral(AMBHEMI *hp, FVECT uv[2], const double r0, const double r1)
645 {
646 const double max_d = 1.0/(minarad*ambacc + 0.001);
647 const double ang_res = 0.5*PI/hp->ns;
648 const double ang_step = ang_res/((int)(16/PI*ang_res) + 1.01);
649 double avg_d = 0;
650 uint32 flgs = 0;
651 FVECT vec;
652 double u, v;
653 double ang, a1;
654 int i, j;
655 /* don't bother for a few samples */
656 if (hp->ns < 8)
657 return(0);
658 /* check distances overhead */
659 for (i = hp->ns*3/4; i-- > hp->ns>>2; )
660 for (j = hp->ns*3/4; j-- > hp->ns>>2; )
661 avg_d += ambsam(hp,i,j).d;
662 avg_d *= 4.0/(hp->ns*hp->ns);
663 if (avg_d*r0 >= 1.0) /* ceiling too low for corral? */
664 return(0);
665 if (avg_d >= max_d) /* insurance */
666 return(0);
667 /* else circle around perimeter */
668 for (i = 0; i < hp->ns; i++)
669 for (j = 0; j < hp->ns; j += !i|(i==hp->ns-1) ? 1 : hp->ns-1) {
670 AMBSAMP *ap = &ambsam(hp,i,j);
671 if ((ap->d <= FTINY) | (ap->d >= max_d))
672 continue; /* too far or too near */
673 VSUB(vec, ap->p, hp->rp->rop);
674 u = DOT(vec, uv[0]);
675 v = DOT(vec, uv[1]);
676 if ((r0*r0*u*u + r1*r1*v*v) * ap->d*ap->d <= u*u + v*v)
677 continue; /* occluder outside ellipse */
678 ang = atan2a(v, u); /* else set direction flags */
679 for (a1 = ang-ang_res; a1 <= ang+ang_res; a1 += ang_step)
680 flgs |= 1L<<(int)(16/PI*(a1 + 2.*PI*(a1 < 0)));
681 }
682 return(flgs);
683 }
684
685
686 int
687 doambient( /* compute ambient component */
688 COLOR rcol, /* input/output color */
689 RAY *r,
690 double wt,
691 FVECT uv[2], /* returned (optional) */
692 float ra[2], /* returned (optional) */
693 float pg[2], /* returned (optional) */
694 float dg[2], /* returned (optional) */
695 uint32 *crlp /* returned (optional) */
696 )
697 {
698 AMBHEMI *hp = samp_hemi(rcol, r, wt);
699 FVECT my_uv[2];
700 double d, K;
701 AMBSAMP *ap;
702 int i;
703 /* clear return values */
704 if (uv != NULL)
705 memset(uv, 0, sizeof(FVECT)*2);
706 if (ra != NULL)
707 ra[0] = ra[1] = 0.0;
708 if (pg != NULL)
709 pg[0] = pg[1] = 0.0;
710 if (dg != NULL)
711 dg[0] = dg[1] = 0.0;
712 if (crlp != NULL)
713 *crlp = 0;
714 if (hp == NULL) /* sampling falure? */
715 return(0);
716
717 if ((ra == NULL) & (pg == NULL) & (dg == NULL) ||
718 (hp->sampOK < 0) | (hp->ns < 6)) {
719 free(hp); /* Hessian not requested/possible */
720 return(-1); /* value-only return value */
721 }
722 if ((d = bright(rcol)) > FTINY) { /* normalize Y values */
723 d = 0.99*(hp->ns*hp->ns)/d;
724 K = 0.01;
725 } else { /* or fall back on geometric Hessian */
726 K = 1.0;
727 pg = NULL;
728 dg = NULL;
729 crlp = NULL;
730 }
731 ap = hp->sa; /* relative Y channel from here on... */
732 for (i = hp->ns*hp->ns; i--; ap++)
733 colval(ap->v,CIEY) = bright(ap->v)*d + K;
734
735 if (uv == NULL) /* make sure we have axis pointers */
736 uv = my_uv;
737 /* compute radii & pos. gradient */
738 ambHessian(hp, uv, ra, pg);
739
740 if (dg != NULL) /* compute direction gradient */
741 ambdirgrad(hp, uv, dg);
742
743 if (ra != NULL) { /* scale/clamp radii */
744 if (pg != NULL) {
745 if (ra[0]*(d = fabs(pg[0])) > 1.0)
746 ra[0] = 1.0/d;
747 if (ra[1]*(d = fabs(pg[1])) > 1.0)
748 ra[1] = 1.0/d;
749 if (ra[0] > ra[1])
750 ra[0] = ra[1];
751 }
752 if (ra[0] < minarad) {
753 ra[0] = minarad;
754 if (ra[1] < minarad)
755 ra[1] = minarad;
756 }
757 ra[0] *= d = 1.0/sqrt(wt);
758 if ((ra[1] *= d) > 2.0*ra[0])
759 ra[1] = 2.0*ra[0];
760 if (ra[1] > maxarad) {
761 ra[1] = maxarad;
762 if (ra[0] > maxarad)
763 ra[0] = maxarad;
764 }
765 /* flag encroached directions */
766 if (crlp != NULL)
767 *crlp = ambcorral(hp, uv, ra[0]*ambacc, ra[1]*ambacc);
768 if (pg != NULL) { /* cap gradient if necessary */
769 d = pg[0]*pg[0]*ra[0]*ra[0] + pg[1]*pg[1]*ra[1]*ra[1];
770 if (d > 1.0) {
771 d = 1.0/sqrt(d);
772 pg[0] *= d;
773 pg[1] *= d;
774 }
775 }
776 }
777 free(hp); /* clean up and return */
778 return(1);
779 }
780
781
782 #else /* ! NEWAMB */
783
784
785 void
786 inithemi( /* initialize sampling hemisphere */
787 AMBHEMI *hp,
788 COLOR ac,
789 RAY *r,
790 double wt
791 )
792 {
793 double d;
794 int i;
795 /* set number of divisions */
796 if (ambacc <= FTINY &&
797 wt > (d = 0.8*intens(ac)*r->rweight/(ambdiv*minweight)))
798 wt = d; /* avoid ray termination */
799 hp->nt = sqrt(ambdiv * wt / PI) + 0.5;
800 i = ambacc > FTINY ? 3 : 1; /* minimum number of samples */
801 if (hp->nt < i)
802 hp->nt = i;
803 hp->np = PI * hp->nt + 0.5;
804 /* set number of super-samples */
805 hp->ns = ambssamp * wt + 0.5;
806 /* assign coefficient */
807 copycolor(hp->acoef, ac);
808 d = 1.0/(hp->nt*hp->np);
809 scalecolor(hp->acoef, d);
810 /* make axes */
811 VCOPY(hp->uz, r->ron);
812 hp->uy[0] = hp->uy[1] = hp->uy[2] = 0.0;
813 for (i = 0; i < 3; i++)
814 if (hp->uz[i] < 0.6 && hp->uz[i] > -0.6)
815 break;
816 if (i >= 3)
817 error(CONSISTENCY, "bad ray direction in inithemi");
818 hp->uy[i] = 1.0;
819 fcross(hp->ux, hp->uy, hp->uz);
820 normalize(hp->ux);
821 fcross(hp->uy, hp->uz, hp->ux);
822 }
823
824
825 int
826 divsample( /* sample a division */
827 AMBSAMP *dp,
828 AMBHEMI *h,
829 RAY *r
830 )
831 {
832 RAY ar;
833 int hlist[3];
834 double spt[2];
835 double xd, yd, zd;
836 double b2;
837 double phi;
838 int i;
839 /* ambient coefficient for weight */
840 if (ambacc > FTINY)
841 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
842 else
843 copycolor(ar.rcoef, h->acoef);
844 if (rayorigin(&ar, AMBIENT, r, ar.rcoef) < 0)
845 return(-1);
846 if (ambacc > FTINY) {
847 multcolor(ar.rcoef, h->acoef);
848 scalecolor(ar.rcoef, 1./AVGREFL);
849 }
850 hlist[0] = r->rno;
851 hlist[1] = dp->t;
852 hlist[2] = dp->p;
853 multisamp(spt, 2, urand(ilhash(hlist,3)+dp->n));
854 zd = sqrt((dp->t + spt[0])/h->nt);
855 phi = 2.0*PI * (dp->p + spt[1])/h->np;
856 xd = tcos(phi) * zd;
857 yd = tsin(phi) * zd;
858 zd = sqrt(1.0 - zd*zd);
859 for (i = 0; i < 3; i++)
860 ar.rdir[i] = xd*h->ux[i] +
861 yd*h->uy[i] +
862 zd*h->uz[i];
863 checknorm(ar.rdir);
864 dimlist[ndims++] = dp->t*h->np + dp->p + 90171;
865 rayvalue(&ar);
866 ndims--;
867 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
868 addcolor(dp->v, ar.rcol);
869 /* use rxt to improve gradient calc */
870 if (ar.rxt > FTINY && ar.rxt < FHUGE)
871 dp->r += 1.0/ar.rxt;
872 /* (re)initialize error */
873 if (dp->n++) {
874 b2 = bright(dp->v)/dp->n - bright(ar.rcol);
875 b2 = b2*b2 + dp->k*((dp->n-1)*(dp->n-1));
876 dp->k = b2/(dp->n*dp->n);
877 } else
878 dp->k = 0.0;
879 return(0);
880 }
881
882
883 static int
884 ambcmp( /* decreasing order */
885 const void *p1,
886 const void *p2
887 )
888 {
889 const AMBSAMP *d1 = (const AMBSAMP *)p1;
890 const AMBSAMP *d2 = (const AMBSAMP *)p2;
891
892 if (d1->k < d2->k)
893 return(1);
894 if (d1->k > d2->k)
895 return(-1);
896 return(0);
897 }
898
899
900 static int
901 ambnorm( /* standard order */
902 const void *p1,
903 const void *p2
904 )
905 {
906 const AMBSAMP *d1 = (const AMBSAMP *)p1;
907 const AMBSAMP *d2 = (const AMBSAMP *)p2;
908 int c;
909
910 if ( (c = d1->t - d2->t) )
911 return(c);
912 return(d1->p - d2->p);
913 }
914
915
916 double
917 doambient( /* compute ambient component */
918 COLOR rcol,
919 RAY *r,
920 double wt,
921 FVECT pg,
922 FVECT dg
923 )
924 {
925 double b, d=0;
926 AMBHEMI hemi;
927 AMBSAMP *div;
928 AMBSAMP dnew;
929 double acol[3];
930 AMBSAMP *dp;
931 double arad;
932 int divcnt;
933 int i, j;
934 /* initialize hemisphere */
935 inithemi(&hemi, rcol, r, wt);
936 divcnt = hemi.nt * hemi.np;
937 /* initialize */
938 if (pg != NULL)
939 pg[0] = pg[1] = pg[2] = 0.0;
940 if (dg != NULL)
941 dg[0] = dg[1] = dg[2] = 0.0;
942 setcolor(rcol, 0.0, 0.0, 0.0);
943 if (divcnt == 0)
944 return(0.0);
945 /* allocate super-samples */
946 if (hemi.ns > 0 || pg != NULL || dg != NULL) {
947 div = (AMBSAMP *)malloc(divcnt*sizeof(AMBSAMP));
948 if (div == NULL)
949 error(SYSTEM, "out of memory in doambient");
950 } else
951 div = NULL;
952 /* sample the divisions */
953 arad = 0.0;
954 acol[0] = acol[1] = acol[2] = 0.0;
955 if ((dp = div) == NULL)
956 dp = &dnew;
957 divcnt = 0;
958 for (i = 0; i < hemi.nt; i++)
959 for (j = 0; j < hemi.np; j++) {
960 dp->t = i; dp->p = j;
961 setcolor(dp->v, 0.0, 0.0, 0.0);
962 dp->r = 0.0;
963 dp->n = 0;
964 if (divsample(dp, &hemi, r) < 0) {
965 if (div != NULL)
966 dp++;
967 continue;
968 }
969 arad += dp->r;
970 divcnt++;
971 if (div != NULL)
972 dp++;
973 else
974 addcolor(acol, dp->v);
975 }
976 if (!divcnt) {
977 if (div != NULL)
978 free((void *)div);
979 return(0.0); /* no samples taken */
980 }
981 if (divcnt < hemi.nt*hemi.np) {
982 pg = dg = NULL; /* incomplete sampling */
983 hemi.ns = 0;
984 } else if (arad > FTINY && divcnt/arad < minarad) {
985 hemi.ns = 0; /* close enough */
986 } else if (hemi.ns > 0) { /* else perform super-sampling? */
987 comperrs(div, &hemi); /* compute errors */
988 qsort(div, divcnt, sizeof(AMBSAMP), ambcmp); /* sort divs */
989 /* super-sample */
990 for (i = hemi.ns; i > 0; i--) {
991 dnew = *div;
992 if (divsample(&dnew, &hemi, r) < 0) {
993 dp++;
994 continue;
995 }
996 dp = div; /* reinsert */
997 j = divcnt < i ? divcnt : i;
998 while (--j > 0 && dnew.k < dp[1].k) {
999 *dp = *(dp+1);
1000 dp++;
1001 }
1002 *dp = dnew;
1003 }
1004 if (pg != NULL || dg != NULL) /* restore order */
1005 qsort(div, divcnt, sizeof(AMBSAMP), ambnorm);
1006 }
1007 /* compute returned values */
1008 if (div != NULL) {
1009 arad = 0.0; /* note: divcnt may be < nt*np */
1010 for (i = hemi.nt*hemi.np, dp = div; i-- > 0; dp++) {
1011 arad += dp->r;
1012 if (dp->n > 1) {
1013 b = 1.0/dp->n;
1014 scalecolor(dp->v, b);
1015 dp->r *= b;
1016 dp->n = 1;
1017 }
1018 addcolor(acol, dp->v);
1019 }
1020 b = bright(acol);
1021 if (b > FTINY) {
1022 b = 1.0/b; /* compute & normalize gradient(s) */
1023 if (pg != NULL) {
1024 posgradient(pg, div, &hemi);
1025 for (i = 0; i < 3; i++)
1026 pg[i] *= b;
1027 }
1028 if (dg != NULL) {
1029 dirgradient(dg, div, &hemi);
1030 for (i = 0; i < 3; i++)
1031 dg[i] *= b;
1032 }
1033 }
1034 free((void *)div);
1035 }
1036 copycolor(rcol, acol);
1037 if (arad <= FTINY)
1038 arad = maxarad;
1039 else
1040 arad = (divcnt+hemi.ns)/arad;
1041 if (pg != NULL) { /* reduce radius if gradient large */
1042 d = DOT(pg,pg);
1043 if (d*arad*arad > 1.0)
1044 arad = 1.0/sqrt(d);
1045 }
1046 if (arad < minarad) {
1047 arad = minarad;
1048 if (pg != NULL && d*arad*arad > 1.0) { /* cap gradient */
1049 d = 1.0/arad/sqrt(d);
1050 for (i = 0; i < 3; i++)
1051 pg[i] *= d;
1052 }
1053 }
1054 if ((arad /= sqrt(wt)) > maxarad)
1055 arad = maxarad;
1056 return(arad);
1057 }
1058
1059
1060 void
1061 comperrs( /* compute initial error estimates */
1062 AMBSAMP *da, /* assumes standard ordering */
1063 AMBHEMI *hp
1064 )
1065 {
1066 double b, b2;
1067 int i, j;
1068 AMBSAMP *dp;
1069 /* sum differences from neighbors */
1070 dp = da;
1071 for (i = 0; i < hp->nt; i++)
1072 for (j = 0; j < hp->np; j++) {
1073 #ifdef DEBUG
1074 if (dp->t != i || dp->p != j)
1075 error(CONSISTENCY,
1076 "division order in comperrs");
1077 #endif
1078 b = bright(dp[0].v);
1079 if (i > 0) { /* from above */
1080 b2 = bright(dp[-hp->np].v) - b;
1081 b2 *= b2 * 0.25;
1082 dp[0].k += b2;
1083 dp[-hp->np].k += b2;
1084 }
1085 if (j > 0) { /* from behind */
1086 b2 = bright(dp[-1].v) - b;
1087 b2 *= b2 * 0.25;
1088 dp[0].k += b2;
1089 dp[-1].k += b2;
1090 } else { /* around */
1091 b2 = bright(dp[hp->np-1].v) - b;
1092 b2 *= b2 * 0.25;
1093 dp[0].k += b2;
1094 dp[hp->np-1].k += b2;
1095 }
1096 dp++;
1097 }
1098 /* divide by number of neighbors */
1099 dp = da;
1100 for (j = 0; j < hp->np; j++) /* top row */
1101 (dp++)->k *= 1.0/3.0;
1102 if (hp->nt < 2)
1103 return;
1104 for (i = 1; i < hp->nt-1; i++) /* central region */
1105 for (j = 0; j < hp->np; j++)
1106 (dp++)->k *= 0.25;
1107 for (j = 0; j < hp->np; j++) /* bottom row */
1108 (dp++)->k *= 1.0/3.0;
1109 }
1110
1111
1112 void
1113 posgradient( /* compute position gradient */
1114 FVECT gv,
1115 AMBSAMP *da, /* assumes standard ordering */
1116 AMBHEMI *hp
1117 )
1118 {
1119 int i, j;
1120 double nextsine, lastsine, b, d;
1121 double mag0, mag1;
1122 double phi, cosp, sinp, xd, yd;
1123 AMBSAMP *dp;
1124
1125 xd = yd = 0.0;
1126 for (j = 0; j < hp->np; j++) {
1127 dp = da + j;
1128 mag0 = mag1 = 0.0;
1129 lastsine = 0.0;
1130 for (i = 0; i < hp->nt; i++) {
1131 #ifdef DEBUG
1132 if (dp->t != i || dp->p != j)
1133 error(CONSISTENCY,
1134 "division order in posgradient");
1135 #endif
1136 b = bright(dp->v);
1137 if (i > 0) {
1138 d = dp[-hp->np].r;
1139 if (dp[0].r > d) d = dp[0].r;
1140 /* sin(t)*cos(t)^2 */
1141 d *= lastsine * (1.0 - (double)i/hp->nt);
1142 mag0 += d*(b - bright(dp[-hp->np].v));
1143 }
1144 nextsine = sqrt((double)(i+1)/hp->nt);
1145 if (j > 0) {
1146 d = dp[-1].r;
1147 if (dp[0].r > d) d = dp[0].r;
1148 mag1 += d * (nextsine - lastsine) *
1149 (b - bright(dp[-1].v));
1150 } else {
1151 d = dp[hp->np-1].r;
1152 if (dp[0].r > d) d = dp[0].r;
1153 mag1 += d * (nextsine - lastsine) *
1154 (b - bright(dp[hp->np-1].v));
1155 }
1156 dp += hp->np;
1157 lastsine = nextsine;
1158 }
1159 mag0 *= 2.0*PI / hp->np;
1160 phi = 2.0*PI * (double)j/hp->np;
1161 cosp = tcos(phi); sinp = tsin(phi);
1162 xd += mag0*cosp - mag1*sinp;
1163 yd += mag0*sinp + mag1*cosp;
1164 }
1165 for (i = 0; i < 3; i++)
1166 gv[i] = (xd*hp->ux[i] + yd*hp->uy[i])*(hp->nt*hp->np)/PI;
1167 }
1168
1169
1170 void
1171 dirgradient( /* compute direction gradient */
1172 FVECT gv,
1173 AMBSAMP *da, /* assumes standard ordering */
1174 AMBHEMI *hp
1175 )
1176 {
1177 int i, j;
1178 double mag;
1179 double phi, xd, yd;
1180 AMBSAMP *dp;
1181
1182 xd = yd = 0.0;
1183 for (j = 0; j < hp->np; j++) {
1184 dp = da + j;
1185 mag = 0.0;
1186 for (i = 0; i < hp->nt; i++) {
1187 #ifdef DEBUG
1188 if (dp->t != i || dp->p != j)
1189 error(CONSISTENCY,
1190 "division order in dirgradient");
1191 #endif
1192 /* tan(t) */
1193 mag += bright(dp->v)/sqrt(hp->nt/(i+.5) - 1.0);
1194 dp += hp->np;
1195 }
1196 phi = 2.0*PI * (j+.5)/hp->np + PI/2.0;
1197 xd += mag * tcos(phi);
1198 yd += mag * tsin(phi);
1199 }
1200 for (i = 0; i < 3; i++)
1201 gv[i] = xd*hp->ux[i] + yd*hp->uy[i];
1202 }
1203
1204 #endif /* ! NEWAMB */