ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambcomp.c
Revision: 2.68
Committed: Thu Oct 23 18:19:14 2014 UTC (9 years, 6 months ago) by greg
Content type: text/plain
Branch: MAIN
CVS Tags: rad4R2P2
Changes since 2.67: +3 -3 lines
Log Message:
Fixed floating-point error under Windows caused by cos^2 + sin^2 > 1

File Contents

# User Rev Content
1 greg 1.1 #ifndef lint
2 greg 2.68 static const char RCSid[] = "$Id: ambcomp.c,v 2.67 2014/10/17 20:47:59 greg Exp $";
3 greg 1.1 #endif
4     /*
5     * Routines to compute "ambient" values using Monte Carlo
6 greg 2.9 *
7 greg 2.27 * 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 greg 2.46 * 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 greg 2.9 * Declarations of external symbols in ambient.h
16     */
17    
18 greg 2.10 #include "copyright.h"
19 greg 1.1
20     #include "ray.h"
21 greg 2.25 #include "ambient.h"
22     #include "random.h"
23 greg 1.1
24 greg 2.63 #ifndef OLDAMB
25 greg 1.1
26 greg 2.26 extern void SDsquare2disk(double ds[2], double seedx, double seedy);
27    
28     typedef struct {
29 greg 2.44 COLOR v; /* hemisphere sample value */
30 greg 2.47 float d; /* reciprocal distance (1/rt) */
31 greg 2.44 FVECT p; /* intersection point */
32     } AMBSAMP; /* sample value */
33    
34     typedef struct {
35 greg 2.26 RAY *rp; /* originating ray sample */
36     int ns; /* number of samples per axis */
37 greg 2.61 int sampOK; /* acquired full sample set? */
38 greg 2.26 COLOR acoef; /* division contribution coefficient */
39 greg 2.61 double acol[3]; /* accumulated color */
40     FVECT ux, uy; /* tangent axis unit vectors */
41 greg 2.44 AMBSAMP sa[1]; /* sample array (extends struct) */
42 greg 2.26 } AMBHEMI; /* ambient sample hemisphere */
43    
44 greg 2.56 #define AI(h,i,j) ((i)*(h)->ns + (j))
45     #define ambsam(h,i,j) (h)->sa[AI(h,i,j)]
46 greg 2.26
47 greg 2.27 typedef struct {
48 greg 2.35 FVECT r_i, r_i1, e_i, rcp, rI2_eJ2;
49     double I1, I2;
50 greg 2.27 } FFTRI; /* vectors and coefficients for Hessian calculation */
51    
52 greg 2.26
53 greg 2.61 static int
54     ambsample( /* initial ambient division sample */
55     AMBHEMI *hp,
56     int i,
57     int j,
58     int n
59 greg 2.26 )
60     {
61 greg 2.61 AMBSAMP *ap = &ambsam(hp,i,j);
62     RAY ar;
63 greg 2.41 int hlist[3], ii;
64     double spt[2], zd;
65 greg 2.61 /* generate hemispherical sample */
66 greg 2.26 /* ambient coefficient for weight */
67     if (ambacc > FTINY)
68 greg 2.61 setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
69 greg 2.26 else
70 greg 2.61 copycolor(ar.rcoef, hp->acoef);
71 greg 2.62 if (rayorigin(&ar, AMBIENT, hp->rp, ar.rcoef) < 0)
72 greg 2.41 return(0);
73 greg 2.26 if (ambacc > FTINY) {
74 greg 2.61 multcolor(ar.rcoef, hp->acoef);
75     scalecolor(ar.rcoef, 1./AVGREFL);
76 greg 2.41 }
77     hlist[0] = hp->rp->rno;
78 greg 2.46 hlist[1] = j;
79     hlist[2] = i;
80 greg 2.41 multisamp(spt, 2, urand(ilhash(hlist,3)+n));
81 greg 2.64 /* avoid coincident samples */
82     if (!n && (0 < i) & (i < hp->ns-1) &&
83     (0 < j) & (j < hp->ns-1)) {
84 greg 2.46 if ((spt[0] < 0.1) | (spt[0] >= 0.9))
85 greg 2.41 spt[0] = 0.1 + 0.8*frandom();
86 greg 2.46 if ((spt[1] < 0.1) | (spt[1] >= 0.9))
87 greg 2.41 spt[1] = 0.1 + 0.8*frandom();
88 greg 2.26 }
89 greg 2.46 SDsquare2disk(spt, (j+spt[1])/hp->ns, (i+spt[0])/hp->ns);
90 greg 2.26 zd = sqrt(1. - spt[0]*spt[0] - spt[1]*spt[1]);
91     for (ii = 3; ii--; )
92 greg 2.61 ar.rdir[ii] = spt[0]*hp->ux[ii] +
93 greg 2.26 spt[1]*hp->uy[ii] +
94     zd*hp->rp->ron[ii];
95 greg 2.61 checknorm(ar.rdir);
96 greg 2.56 dimlist[ndims++] = AI(hp,i,j) + 90171;
97 greg 2.61 rayvalue(&ar); /* evaluate ray */
98     ndims--;
99     if (ar.rt <= FTINY)
100     return(0); /* should never happen */
101     multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
102 greg 2.62 if (ar.rt*ap->d < 1.0) /* new/closer distance? */
103 greg 2.61 ap->d = 1.0/ar.rt;
104     if (!n) { /* record first vertex & value */
105     if (ar.rt > 10.0*thescene.cusize)
106     ar.rt = 10.0*thescene.cusize;
107     VSUM(ap->p, ar.rorg, ar.rdir, ar.rt);
108     copycolor(ap->v, ar.rcol);
109     } else { /* else update recorded value */
110     hp->acol[RED] -= colval(ap->v,RED);
111     hp->acol[GRN] -= colval(ap->v,GRN);
112     hp->acol[BLU] -= colval(ap->v,BLU);
113     zd = 1.0/(double)(n+1);
114     scalecolor(ar.rcol, zd);
115     zd *= (double)n;
116     scalecolor(ap->v, zd);
117     addcolor(ap->v, ar.rcol);
118     }
119     addcolor(hp->acol, ap->v); /* add to our sum */
120 greg 2.41 return(1);
121     }
122    
123    
124     /* Estimate errors based on ambient division differences */
125     static float *
126     getambdiffs(AMBHEMI *hp)
127     {
128 greg 2.55 float *earr = (float *)calloc(hp->ns*hp->ns, sizeof(float));
129 greg 2.41 float *ep;
130 greg 2.42 AMBSAMP *ap;
131 greg 2.41 double b, d2;
132     int i, j;
133    
134     if (earr == NULL) /* out of memory? */
135     return(NULL);
136     /* compute squared neighbor diffs */
137 greg 2.42 for (ap = hp->sa, ep = earr, i = 0; i < hp->ns; i++)
138     for (j = 0; j < hp->ns; j++, ap++, ep++) {
139     b = bright(ap[0].v);
140 greg 2.41 if (i) { /* from above */
141 greg 2.42 d2 = b - bright(ap[-hp->ns].v);
142 greg 2.41 d2 *= d2;
143     ep[0] += d2;
144     ep[-hp->ns] += d2;
145     }
146 greg 2.55 if (!j) continue;
147     /* from behind */
148     d2 = b - bright(ap[-1].v);
149     d2 *= d2;
150     ep[0] += d2;
151     ep[-1] += d2;
152     if (!i) continue;
153     /* diagonal */
154     d2 = b - bright(ap[-hp->ns-1].v);
155     d2 *= d2;
156     ep[0] += d2;
157     ep[-hp->ns-1] += d2;
158 greg 2.41 }
159     /* correct for number of neighbors */
160 greg 2.55 earr[0] *= 8./3.;
161     earr[hp->ns-1] *= 8./3.;
162     earr[(hp->ns-1)*hp->ns] *= 8./3.;
163     earr[(hp->ns-1)*hp->ns + hp->ns-1] *= 8./3.;
164 greg 2.41 for (i = 1; i < hp->ns-1; i++) {
165 greg 2.55 earr[i*hp->ns] *= 8./5.;
166     earr[i*hp->ns + hp->ns-1] *= 8./5.;
167 greg 2.41 }
168     for (j = 1; j < hp->ns-1; j++) {
169 greg 2.55 earr[j] *= 8./5.;
170     earr[(hp->ns-1)*hp->ns + j] *= 8./5.;
171 greg 2.41 }
172     return(earr);
173     }
174    
175    
176 greg 2.43 /* Perform super-sampling on hemisphere (introduces bias) */
177 greg 2.41 static void
178 greg 2.61 ambsupersamp(AMBHEMI *hp, int cnt)
179 greg 2.41 {
180     float *earr = getambdiffs(hp);
181 greg 2.54 double e2rem = 0;
182 greg 2.41 AMBSAMP *ap;
183     float *ep;
184 greg 2.55 int i, j, n, nss;
185 greg 2.41
186     if (earr == NULL) /* just skip calc. if no memory */
187     return;
188 greg 2.54 /* accumulate estimated variances */
189 greg 2.55 for (ep = earr + hp->ns*hp->ns; ep > earr; )
190     e2rem += *--ep;
191 greg 2.41 ep = earr; /* perform super-sampling */
192     for (ap = hp->sa, i = 0; i < hp->ns; i++)
193     for (j = 0; j < hp->ns; j++, ap++) {
194 greg 2.55 if (e2rem <= FTINY)
195     goto done; /* nothing left to do */
196     nss = *ep/e2rem*cnt + frandom();
197 greg 2.62 for (n = 1; n <= nss && ambsample(hp,i,j,n); n++)
198     --cnt;
199 greg 2.61 e2rem -= *ep++; /* update remainder */
200 greg 2.41 }
201 greg 2.55 done:
202 greg 2.41 free(earr);
203     }
204    
205    
206 greg 2.61 static AMBHEMI *
207     samp_hemi( /* sample indirect hemisphere */
208     COLOR rcol,
209     RAY *r,
210     double wt
211     )
212     {
213     AMBHEMI *hp;
214     double d;
215     int n, i, j;
216     /* set number of divisions */
217     if (ambacc <= FTINY &&
218     wt > (d = 0.8*intens(rcol)*r->rweight/(ambdiv*minweight)))
219     wt = d; /* avoid ray termination */
220     n = sqrt(ambdiv * wt) + 0.5;
221 greg 2.68 i = 1 + 5*(ambacc > FTINY); /* minimum number of samples */
222 greg 2.61 if (n < i)
223     n = i;
224     /* allocate sampling array */
225     hp = (AMBHEMI *)malloc(sizeof(AMBHEMI) + sizeof(AMBSAMP)*(n*n - 1));
226     if (hp == NULL)
227     error(SYSTEM, "out of memory in samp_hemi");
228     hp->rp = r;
229     hp->ns = n;
230     hp->acol[RED] = hp->acol[GRN] = hp->acol[BLU] = 0.0;
231 greg 2.62 memset(hp->sa, 0, sizeof(AMBSAMP)*n*n);
232 greg 2.61 hp->sampOK = 0;
233     /* assign coefficient */
234     copycolor(hp->acoef, rcol);
235     d = 1.0/(n*n);
236     scalecolor(hp->acoef, d);
237     /* make tangent plane axes */
238     hp->uy[0] = 0.5 - frandom();
239     hp->uy[1] = 0.5 - frandom();
240     hp->uy[2] = 0.5 - frandom();
241     for (i = 3; i--; )
242     if ((-0.6 < r->ron[i]) & (r->ron[i] < 0.6))
243     break;
244     if (i < 0)
245     error(CONSISTENCY, "bad ray direction in samp_hemi");
246     hp->uy[i] = 1.0;
247     VCROSS(hp->ux, hp->uy, r->ron);
248     normalize(hp->ux);
249     VCROSS(hp->uy, r->ron, hp->ux);
250     /* sample divisions */
251     for (i = hp->ns; i--; )
252     for (j = hp->ns; j--; )
253     hp->sampOK += ambsample(hp, i, j, 0);
254     copycolor(rcol, hp->acol);
255     if (!hp->sampOK) { /* utter failure? */
256     free(hp);
257     return(NULL);
258     }
259     if (hp->sampOK < hp->ns*hp->ns) {
260     hp->sampOK *= -1; /* soft failure */
261     return(hp);
262     }
263     n = ambssamp*wt + 0.5;
264     if (n > 8) { /* perform super-sampling? */
265     ambsupersamp(hp, n);
266     copycolor(rcol, hp->acol);
267     }
268     return(hp); /* all is well */
269     }
270    
271    
272 greg 2.46 /* Return brightness of farthest ambient sample */
273     static double
274 greg 2.56 back_ambval(AMBHEMI *hp, const int n1, const int n2, const int n3)
275 greg 2.46 {
276 greg 2.56 if (hp->sa[n1].d <= hp->sa[n2].d) {
277     if (hp->sa[n1].d <= hp->sa[n3].d)
278     return(colval(hp->sa[n1].v,CIEY));
279     return(colval(hp->sa[n3].v,CIEY));
280     }
281     if (hp->sa[n2].d <= hp->sa[n3].d)
282     return(colval(hp->sa[n2].v,CIEY));
283     return(colval(hp->sa[n3].v,CIEY));
284 greg 2.46 }
285    
286    
287 greg 2.27 /* Compute vectors and coefficients for Hessian/gradient calcs */
288     static void
289 greg 2.56 comp_fftri(FFTRI *ftp, AMBHEMI *hp, const int n0, const int n1)
290 greg 2.27 {
291 greg 2.56 double rdot_cp, dot_e, dot_er, rdot_r, rdot_r1, J2;
292     int ii;
293    
294     VSUB(ftp->r_i, hp->sa[n0].p, hp->rp->rop);
295     VSUB(ftp->r_i1, hp->sa[n1].p, hp->rp->rop);
296     VSUB(ftp->e_i, hp->sa[n1].p, hp->sa[n0].p);
297 greg 2.35 VCROSS(ftp->rcp, ftp->r_i, ftp->r_i1);
298     rdot_cp = 1.0/DOT(ftp->rcp,ftp->rcp);
299 greg 2.27 dot_e = DOT(ftp->e_i,ftp->e_i);
300     dot_er = DOT(ftp->e_i, ftp->r_i);
301 greg 2.32 rdot_r = 1.0/DOT(ftp->r_i,ftp->r_i);
302     rdot_r1 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
303     ftp->I1 = acos( DOT(ftp->r_i, ftp->r_i1) * sqrt(rdot_r*rdot_r1) ) *
304 greg 2.35 sqrt( rdot_cp );
305 greg 2.32 ftp->I2 = ( DOT(ftp->e_i, ftp->r_i1)*rdot_r1 - dot_er*rdot_r +
306 greg 2.35 dot_e*ftp->I1 )*0.5*rdot_cp;
307 greg 2.32 J2 = ( 0.5*(rdot_r - rdot_r1) - dot_er*ftp->I2 ) / dot_e;
308 greg 2.46 for (ii = 3; ii--; )
309     ftp->rI2_eJ2[ii] = ftp->I2*ftp->r_i[ii] + J2*ftp->e_i[ii];
310 greg 2.27 }
311    
312    
313 greg 2.28 /* Compose 3x3 matrix from two vectors */
314 greg 2.27 static void
315     compose_matrix(FVECT mat[3], FVECT va, FVECT vb)
316     {
317     mat[0][0] = 2.0*va[0]*vb[0];
318     mat[1][1] = 2.0*va[1]*vb[1];
319     mat[2][2] = 2.0*va[2]*vb[2];
320     mat[0][1] = mat[1][0] = va[0]*vb[1] + va[1]*vb[0];
321     mat[0][2] = mat[2][0] = va[0]*vb[2] + va[2]*vb[0];
322     mat[1][2] = mat[2][1] = va[1]*vb[2] + va[2]*vb[1];
323     }
324    
325    
326     /* Compute partial 3x3 Hessian matrix for edge */
327     static void
328     comp_hessian(FVECT hess[3], FFTRI *ftp, FVECT nrm)
329     {
330 greg 2.35 FVECT ncp;
331 greg 2.27 FVECT m1[3], m2[3], m3[3], m4[3];
332     double d1, d2, d3, d4;
333     double I3, J3, K3;
334     int i, j;
335     /* compute intermediate coefficients */
336     d1 = 1.0/DOT(ftp->r_i,ftp->r_i);
337     d2 = 1.0/DOT(ftp->r_i1,ftp->r_i1);
338     d3 = 1.0/DOT(ftp->e_i,ftp->e_i);
339     d4 = DOT(ftp->e_i, ftp->r_i);
340 greg 2.35 I3 = ( DOT(ftp->e_i, ftp->r_i1)*d2*d2 - d4*d1*d1 + 3.0/d3*ftp->I2 )
341     / ( 4.0*DOT(ftp->rcp,ftp->rcp) );
342 greg 2.27 J3 = 0.25*d3*(d1*d1 - d2*d2) - d4*d3*I3;
343     K3 = d3*(ftp->I2 - I3/d1 - 2.0*d4*J3);
344     /* intermediate matrices */
345 greg 2.35 VCROSS(ncp, nrm, ftp->e_i);
346     compose_matrix(m1, ncp, ftp->rI2_eJ2);
347 greg 2.27 compose_matrix(m2, ftp->r_i, ftp->r_i);
348     compose_matrix(m3, ftp->e_i, ftp->e_i);
349     compose_matrix(m4, ftp->r_i, ftp->e_i);
350 greg 2.35 d1 = DOT(nrm, ftp->rcp);
351 greg 2.27 d2 = -d1*ftp->I2;
352     d1 *= 2.0;
353     for (i = 3; i--; ) /* final matrix sum */
354     for (j = 3; j--; ) {
355     hess[i][j] = m1[i][j] + d1*( I3*m2[i][j] + K3*m3[i][j] +
356     2.0*J3*m4[i][j] );
357     hess[i][j] += d2*(i==j);
358 greg 2.46 hess[i][j] *= -1.0/PI;
359 greg 2.27 }
360     }
361    
362    
363     /* Reverse hessian calculation result for edge in other direction */
364     static void
365     rev_hessian(FVECT hess[3])
366     {
367     int i;
368    
369     for (i = 3; i--; ) {
370     hess[i][0] = -hess[i][0];
371     hess[i][1] = -hess[i][1];
372     hess[i][2] = -hess[i][2];
373     }
374     }
375    
376    
377     /* Add to radiometric Hessian from the given triangle */
378     static void
379     add2hessian(FVECT hess[3], FVECT ehess1[3],
380 greg 2.46 FVECT ehess2[3], FVECT ehess3[3], double v)
381 greg 2.27 {
382     int i, j;
383    
384     for (i = 3; i--; )
385     for (j = 3; j--; )
386     hess[i][j] += v*( ehess1[i][j] + ehess2[i][j] + ehess3[i][j] );
387     }
388    
389    
390     /* Compute partial displacement form factor gradient for edge */
391     static void
392     comp_gradient(FVECT grad, FFTRI *ftp, FVECT nrm)
393     {
394 greg 2.35 FVECT ncp;
395 greg 2.27 double f1;
396     int i;
397    
398 greg 2.35 f1 = 2.0*DOT(nrm, ftp->rcp);
399     VCROSS(ncp, nrm, ftp->e_i);
400 greg 2.27 for (i = 3; i--; )
401 greg 2.46 grad[i] = (0.5/PI)*( ftp->I1*ncp[i] + f1*ftp->rI2_eJ2[i] );
402 greg 2.27 }
403    
404    
405     /* Reverse gradient calculation result for edge in other direction */
406     static void
407     rev_gradient(FVECT grad)
408     {
409     grad[0] = -grad[0];
410     grad[1] = -grad[1];
411     grad[2] = -grad[2];
412     }
413    
414    
415     /* Add to displacement gradient from the given triangle */
416     static void
417 greg 2.46 add2gradient(FVECT grad, FVECT egrad1, FVECT egrad2, FVECT egrad3, double v)
418 greg 2.27 {
419     int i;
420    
421     for (i = 3; i--; )
422     grad[i] += v*( egrad1[i] + egrad2[i] + egrad3[i] );
423     }
424    
425    
426     /* Compute anisotropic radii and eigenvector directions */
427 greg 2.53 static void
428 greg 2.27 eigenvectors(FVECT uv[2], float ra[2], FVECT hessian[3])
429     {
430     double hess2[2][2];
431     FVECT a, b;
432     double evalue[2], slope1, xmag1;
433     int i;
434     /* project Hessian to sample plane */
435     for (i = 3; i--; ) {
436     a[i] = DOT(hessian[i], uv[0]);
437     b[i] = DOT(hessian[i], uv[1]);
438     }
439     hess2[0][0] = DOT(uv[0], a);
440     hess2[0][1] = DOT(uv[0], b);
441     hess2[1][0] = DOT(uv[1], a);
442     hess2[1][1] = DOT(uv[1], b);
443 greg 2.38 /* compute eigenvalue(s) */
444     i = quadratic(evalue, 1.0, -hess2[0][0]-hess2[1][1],
445     hess2[0][0]*hess2[1][1]-hess2[0][1]*hess2[1][0]);
446     if (i == 1) /* double-root (circle) */
447     evalue[1] = evalue[0];
448     if (!i || ((evalue[0] = fabs(evalue[0])) <= FTINY*FTINY) |
449 greg 2.53 ((evalue[1] = fabs(evalue[1])) <= FTINY*FTINY) ) {
450     ra[0] = ra[1] = maxarad;
451     return;
452     }
453 greg 2.27 if (evalue[0] > evalue[1]) {
454 greg 2.29 ra[0] = sqrt(sqrt(4.0/evalue[0]));
455     ra[1] = sqrt(sqrt(4.0/evalue[1]));
456 greg 2.27 slope1 = evalue[1];
457     } else {
458 greg 2.29 ra[0] = sqrt(sqrt(4.0/evalue[1]));
459     ra[1] = sqrt(sqrt(4.0/evalue[0]));
460 greg 2.27 slope1 = evalue[0];
461     }
462     /* compute unit eigenvectors */
463     if (fabs(hess2[0][1]) <= FTINY)
464     return; /* uv OK as is */
465     slope1 = (slope1 - hess2[0][0]) / hess2[0][1];
466     xmag1 = sqrt(1.0/(1.0 + slope1*slope1));
467     for (i = 3; i--; ) {
468     b[i] = xmag1*uv[0][i] + slope1*xmag1*uv[1][i];
469     a[i] = slope1*xmag1*uv[0][i] - xmag1*uv[1][i];
470     }
471     VCOPY(uv[0], a);
472     VCOPY(uv[1], b);
473     }
474    
475    
476 greg 2.26 static void
477     ambHessian( /* anisotropic radii & pos. gradient */
478     AMBHEMI *hp,
479     FVECT uv[2], /* returned */
480 greg 2.28 float ra[2], /* returned (optional) */
481     float pg[2] /* returned (optional) */
482 greg 2.26 )
483     {
484 greg 2.27 static char memerrmsg[] = "out of memory in ambHessian()";
485     FVECT (*hessrow)[3] = NULL;
486     FVECT *gradrow = NULL;
487     FVECT hessian[3];
488     FVECT gradient;
489     FFTRI fftr;
490     int i, j;
491     /* be sure to assign unit vectors */
492     VCOPY(uv[0], hp->ux);
493     VCOPY(uv[1], hp->uy);
494     /* clock-wise vertex traversal from sample POV */
495     if (ra != NULL) { /* initialize Hessian row buffer */
496 greg 2.28 hessrow = (FVECT (*)[3])malloc(sizeof(FVECT)*3*(hp->ns-1));
497 greg 2.27 if (hessrow == NULL)
498     error(SYSTEM, memerrmsg);
499     memset(hessian, 0, sizeof(hessian));
500     } else if (pg == NULL) /* bogus call? */
501     return;
502     if (pg != NULL) { /* initialize form factor row buffer */
503 greg 2.28 gradrow = (FVECT *)malloc(sizeof(FVECT)*(hp->ns-1));
504 greg 2.27 if (gradrow == NULL)
505     error(SYSTEM, memerrmsg);
506     memset(gradient, 0, sizeof(gradient));
507     }
508     /* compute first row of edges */
509     for (j = 0; j < hp->ns-1; j++) {
510 greg 2.56 comp_fftri(&fftr, hp, AI(hp,0,j), AI(hp,0,j+1));
511 greg 2.27 if (hessrow != NULL)
512     comp_hessian(hessrow[j], &fftr, hp->rp->ron);
513     if (gradrow != NULL)
514     comp_gradient(gradrow[j], &fftr, hp->rp->ron);
515     }
516     /* sum each row of triangles */
517     for (i = 0; i < hp->ns-1; i++) {
518     FVECT hesscol[3]; /* compute first vertical edge */
519     FVECT gradcol;
520 greg 2.56 comp_fftri(&fftr, hp, AI(hp,i,0), AI(hp,i+1,0));
521 greg 2.27 if (hessrow != NULL)
522     comp_hessian(hesscol, &fftr, hp->rp->ron);
523     if (gradrow != NULL)
524     comp_gradient(gradcol, &fftr, hp->rp->ron);
525     for (j = 0; j < hp->ns-1; j++) {
526     FVECT hessdia[3]; /* compute triangle contributions */
527     FVECT graddia;
528 greg 2.46 double backg;
529 greg 2.56 backg = back_ambval(hp, AI(hp,i,j),
530     AI(hp,i,j+1), AI(hp,i+1,j));
531 greg 2.27 /* diagonal (inner) edge */
532 greg 2.56 comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j));
533 greg 2.27 if (hessrow != NULL) {
534     comp_hessian(hessdia, &fftr, hp->rp->ron);
535     rev_hessian(hesscol);
536     add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
537     }
538 greg 2.39 if (gradrow != NULL) {
539 greg 2.27 comp_gradient(graddia, &fftr, hp->rp->ron);
540     rev_gradient(gradcol);
541     add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
542     }
543     /* initialize edge in next row */
544 greg 2.56 comp_fftri(&fftr, hp, AI(hp,i+1,j+1), AI(hp,i+1,j));
545 greg 2.27 if (hessrow != NULL)
546     comp_hessian(hessrow[j], &fftr, hp->rp->ron);
547     if (gradrow != NULL)
548     comp_gradient(gradrow[j], &fftr, hp->rp->ron);
549     /* new column edge & paired triangle */
550 greg 2.56 backg = back_ambval(hp, AI(hp,i+1,j+1),
551     AI(hp,i+1,j), AI(hp,i,j+1));
552     comp_fftri(&fftr, hp, AI(hp,i,j+1), AI(hp,i+1,j+1));
553 greg 2.27 if (hessrow != NULL) {
554     comp_hessian(hesscol, &fftr, hp->rp->ron);
555     rev_hessian(hessdia);
556     add2hessian(hessian, hessrow[j], hessdia, hesscol, backg);
557     if (i < hp->ns-2)
558     rev_hessian(hessrow[j]);
559     }
560     if (gradrow != NULL) {
561     comp_gradient(gradcol, &fftr, hp->rp->ron);
562     rev_gradient(graddia);
563     add2gradient(gradient, gradrow[j], graddia, gradcol, backg);
564     if (i < hp->ns-2)
565     rev_gradient(gradrow[j]);
566     }
567     }
568     }
569     /* release row buffers */
570     if (hessrow != NULL) free(hessrow);
571     if (gradrow != NULL) free(gradrow);
572    
573     if (ra != NULL) /* extract eigenvectors & radii */
574     eigenvectors(uv, ra, hessian);
575 greg 2.32 if (pg != NULL) { /* tangential position gradient */
576     pg[0] = DOT(gradient, uv[0]);
577     pg[1] = DOT(gradient, uv[1]);
578 greg 2.27 }
579     }
580    
581    
582     /* Compute direction gradient from a hemispherical sampling */
583     static void
584     ambdirgrad(AMBHEMI *hp, FVECT uv[2], float dg[2])
585     {
586 greg 2.41 AMBSAMP *ap;
587     double dgsum[2];
588     int n;
589     FVECT vd;
590     double gfact;
591 greg 2.27
592 greg 2.29 dgsum[0] = dgsum[1] = 0.0; /* sum values times -tan(theta) */
593 greg 2.27 for (ap = hp->sa, n = hp->ns*hp->ns; n--; ap++) {
594     /* use vector for azimuth + 90deg */
595     VSUB(vd, ap->p, hp->rp->rop);
596 greg 2.29 /* brightness over cosine factor */
597     gfact = colval(ap->v,CIEY) / DOT(hp->rp->ron, vd);
598 greg 2.40 /* sine = proj_radius/vd_length */
599     dgsum[0] -= DOT(uv[1], vd) * gfact;
600     dgsum[1] += DOT(uv[0], vd) * gfact;
601 greg 2.26 }
602 greg 2.29 dg[0] = dgsum[0] / (hp->ns*hp->ns);
603     dg[1] = dgsum[1] / (hp->ns*hp->ns);
604 greg 2.26 }
605    
606 greg 2.27
607 greg 2.49 /* Compute potential light leak direction flags for cache value */
608     static uint32
609     ambcorral(AMBHEMI *hp, FVECT uv[2], const double r0, const double r1)
610 greg 2.47 {
611 greg 2.50 const double max_d = 1.0/(minarad*ambacc + 0.001);
612 greg 2.66 const double ang_res = 0.5*PI/hp->ns;
613     const double ang_step = ang_res/((int)(16/PI*ang_res) + 1.01);
614 greg 2.51 double avg_d = 0;
615 greg 2.50 uint32 flgs = 0;
616 greg 2.58 FVECT vec;
617 greg 2.62 double u, v;
618 greg 2.58 double ang, a1;
619 greg 2.50 int i, j;
620 greg 2.52 /* don't bother for a few samples */
621     if (hp->ns < 12)
622     return(0);
623     /* check distances overhead */
624 greg 2.51 for (i = hp->ns*3/4; i-- > hp->ns>>2; )
625     for (j = hp->ns*3/4; j-- > hp->ns>>2; )
626     avg_d += ambsam(hp,i,j).d;
627     avg_d *= 4.0/(hp->ns*hp->ns);
628 greg 2.52 if (avg_d*r0 >= 1.0) /* ceiling too low for corral? */
629     return(0);
630     if (avg_d >= max_d) /* insurance */
631 greg 2.51 return(0);
632     /* else circle around perimeter */
633 greg 2.47 for (i = 0; i < hp->ns; i++)
634     for (j = 0; j < hp->ns; j += !i|(i==hp->ns-1) ? 1 : hp->ns-1) {
635     AMBSAMP *ap = &ambsam(hp,i,j);
636 greg 2.50 if ((ap->d <= FTINY) | (ap->d >= max_d))
637     continue; /* too far or too near */
638 greg 2.47 VSUB(vec, ap->p, hp->rp->rop);
639 greg 2.62 u = DOT(vec, uv[0]);
640     v = DOT(vec, uv[1]);
641     if ((r0*r0*u*u + r1*r1*v*v) * ap->d*ap->d <= u*u + v*v)
642 greg 2.49 continue; /* occluder outside ellipse */
643     ang = atan2a(v, u); /* else set direction flags */
644 greg 2.66 for (a1 = ang-ang_res; a1 <= ang+ang_res; a1 += ang_step)
645 greg 2.50 flgs |= 1L<<(int)(16/PI*(a1 + 2.*PI*(a1 < 0)));
646 greg 2.47 }
647 greg 2.58 /* add low-angle incident (< 20deg) */
648     if (fabs(hp->rp->rod) <= 0.342) {
649     u = -DOT(hp->rp->rdir, uv[0]);
650     v = -DOT(hp->rp->rdir, uv[1]);
651     if ((r0*r0*u*u + r1*r1*v*v) > hp->rp->rot*hp->rp->rot) {
652     ang = atan2a(v, u);
653     ang += 2.*PI*(ang < 0);
654     ang *= 16/PI;
655     if ((ang < .5) | (ang >= 31.5))
656     flgs |= 0x80000001;
657     else
658     flgs |= 3L<<(int)(ang-.5);
659     }
660     }
661 greg 2.49 return(flgs);
662 greg 2.47 }
663    
664    
665 greg 2.26 int
666     doambient( /* compute ambient component */
667     COLOR rcol, /* input/output color */
668     RAY *r,
669     double wt,
670 greg 2.27 FVECT uv[2], /* returned (optional) */
671     float ra[2], /* returned (optional) */
672     float pg[2], /* returned (optional) */
673 greg 2.49 float dg[2], /* returned (optional) */
674     uint32 *crlp /* returned (optional) */
675 greg 2.26 )
676     {
677 greg 2.61 AMBHEMI *hp = samp_hemi(rcol, r, wt);
678 greg 2.41 FVECT my_uv[2];
679 greg 2.61 double d, K;
680 greg 2.41 AMBSAMP *ap;
681 greg 2.61 int i;
682     /* clear return values */
683 greg 2.26 if (uv != NULL)
684     memset(uv, 0, sizeof(FVECT)*2);
685     if (ra != NULL)
686     ra[0] = ra[1] = 0.0;
687     if (pg != NULL)
688     pg[0] = pg[1] = 0.0;
689     if (dg != NULL)
690     dg[0] = dg[1] = 0.0;
691 greg 2.49 if (crlp != NULL)
692     *crlp = 0;
693 greg 2.61 if (hp == NULL) /* sampling falure? */
694     return(0);
695    
696     if ((ra == NULL) & (pg == NULL) & (dg == NULL) ||
697 greg 2.68 (hp->sampOK < 0) | (hp->ns < 6)) {
698 greg 2.61 free(hp); /* Hessian not requested/possible */
699     return(-1); /* value-only return value */
700 greg 2.26 }
701 greg 2.61 if ((d = bright(rcol)) > FTINY) { /* normalize Y values */
702 greg 2.45 d = 0.99*(hp->ns*hp->ns)/d;
703 greg 2.38 K = 0.01;
704 greg 2.45 } else { /* or fall back on geometric Hessian */
705 greg 2.38 K = 1.0;
706     pg = NULL;
707     dg = NULL;
708 greg 2.53 crlp = NULL;
709 greg 2.38 }
710 greg 2.29 ap = hp->sa; /* relative Y channel from here on... */
711 greg 2.26 for (i = hp->ns*hp->ns; i--; ap++)
712 greg 2.38 colval(ap->v,CIEY) = bright(ap->v)*d + K;
713 greg 2.26
714     if (uv == NULL) /* make sure we have axis pointers */
715     uv = my_uv;
716     /* compute radii & pos. gradient */
717     ambHessian(hp, uv, ra, pg);
718 greg 2.29
719 greg 2.26 if (dg != NULL) /* compute direction gradient */
720     ambdirgrad(hp, uv, dg);
721 greg 2.29
722 greg 2.28 if (ra != NULL) { /* scale/clamp radii */
723 greg 2.35 if (pg != NULL) {
724     if (ra[0]*(d = fabs(pg[0])) > 1.0)
725     ra[0] = 1.0/d;
726     if (ra[1]*(d = fabs(pg[1])) > 1.0)
727     ra[1] = 1.0/d;
728 greg 2.48 if (ra[0] > ra[1])
729     ra[0] = ra[1];
730 greg 2.35 }
731 greg 2.29 if (ra[0] < minarad) {
732     ra[0] = minarad;
733     if (ra[1] < minarad)
734     ra[1] = minarad;
735     }
736 greg 2.60 ra[0] *= d = 1.0/sqrt(wt);
737 greg 2.26 if ((ra[1] *= d) > 2.0*ra[0])
738     ra[1] = 2.0*ra[0];
739 greg 2.28 if (ra[1] > maxarad) {
740     ra[1] = maxarad;
741     if (ra[0] > maxarad)
742     ra[0] = maxarad;
743     }
744 greg 2.53 /* flag encroached directions */
745 greg 2.57 if ((wt >= 0.89*AVGREFL) & (crlp != NULL))
746 greg 2.49 *crlp = ambcorral(hp, uv, ra[0]*ambacc, ra[1]*ambacc);
747 greg 2.35 if (pg != NULL) { /* cap gradient if necessary */
748     d = pg[0]*pg[0]*ra[0]*ra[0] + pg[1]*pg[1]*ra[1]*ra[1];
749     if (d > 1.0) {
750     d = 1.0/sqrt(d);
751     pg[0] *= d;
752     pg[1] *= d;
753     }
754     }
755 greg 2.26 }
756     free(hp); /* clean up and return */
757     return(1);
758     }
759    
760    
761 greg 2.25 #else /* ! NEWAMB */
762 greg 1.1
763    
764 greg 2.15 void
765 greg 2.14 inithemi( /* initialize sampling hemisphere */
766 greg 2.23 AMBHEMI *hp,
767 greg 2.16 COLOR ac,
768 greg 2.14 RAY *r,
769     double wt
770     )
771 greg 1.1 {
772 greg 2.16 double d;
773 greg 2.23 int i;
774 greg 2.14 /* set number of divisions */
775 greg 2.16 if (ambacc <= FTINY &&
776 greg 2.20 wt > (d = 0.8*intens(ac)*r->rweight/(ambdiv*minweight)))
777 greg 2.16 wt = d; /* avoid ray termination */
778     hp->nt = sqrt(ambdiv * wt / PI) + 0.5;
779 greg 2.14 i = ambacc > FTINY ? 3 : 1; /* minimum number of samples */
780     if (hp->nt < i)
781     hp->nt = i;
782     hp->np = PI * hp->nt + 0.5;
783     /* set number of super-samples */
784 greg 2.15 hp->ns = ambssamp * wt + 0.5;
785 greg 2.16 /* assign coefficient */
786 greg 2.14 copycolor(hp->acoef, ac);
787 greg 2.16 d = 1.0/(hp->nt*hp->np);
788     scalecolor(hp->acoef, d);
789 greg 2.14 /* make axes */
790     VCOPY(hp->uz, r->ron);
791     hp->uy[0] = hp->uy[1] = hp->uy[2] = 0.0;
792     for (i = 0; i < 3; i++)
793     if (hp->uz[i] < 0.6 && hp->uz[i] > -0.6)
794     break;
795     if (i >= 3)
796     error(CONSISTENCY, "bad ray direction in inithemi");
797     hp->uy[i] = 1.0;
798     fcross(hp->ux, hp->uy, hp->uz);
799     normalize(hp->ux);
800     fcross(hp->uy, hp->uz, hp->ux);
801 greg 1.1 }
802    
803    
804 greg 2.9 int
805 greg 2.14 divsample( /* sample a division */
806 greg 2.23 AMBSAMP *dp,
807 greg 2.14 AMBHEMI *h,
808     RAY *r
809     )
810 greg 1.1 {
811     RAY ar;
812 greg 1.11 int hlist[3];
813     double spt[2];
814 greg 1.1 double xd, yd, zd;
815     double b2;
816     double phi;
817 greg 2.23 int i;
818 greg 2.15 /* ambient coefficient for weight */
819 greg 2.16 if (ambacc > FTINY)
820     setcolor(ar.rcoef, AVGREFL, AVGREFL, AVGREFL);
821     else
822     copycolor(ar.rcoef, h->acoef);
823 greg 2.14 if (rayorigin(&ar, AMBIENT, r, ar.rcoef) < 0)
824 greg 1.4 return(-1);
825 greg 2.17 if (ambacc > FTINY) {
826     multcolor(ar.rcoef, h->acoef);
827     scalecolor(ar.rcoef, 1./AVGREFL);
828     }
829 greg 1.1 hlist[0] = r->rno;
830     hlist[1] = dp->t;
831     hlist[2] = dp->p;
832 greg 1.13 multisamp(spt, 2, urand(ilhash(hlist,3)+dp->n));
833 greg 1.11 zd = sqrt((dp->t + spt[0])/h->nt);
834     phi = 2.0*PI * (dp->p + spt[1])/h->np;
835 gwlarson 2.8 xd = tcos(phi) * zd;
836     yd = tsin(phi) * zd;
837 greg 1.1 zd = sqrt(1.0 - zd*zd);
838 greg 1.2 for (i = 0; i < 3; i++)
839     ar.rdir[i] = xd*h->ux[i] +
840     yd*h->uy[i] +
841     zd*h->uz[i];
842 greg 2.22 checknorm(ar.rdir);
843 greg 1.2 dimlist[ndims++] = dp->t*h->np + dp->p + 90171;
844 greg 1.1 rayvalue(&ar);
845     ndims--;
846 greg 2.16 multcolor(ar.rcol, ar.rcoef); /* apply coefficient */
847 greg 1.1 addcolor(dp->v, ar.rcol);
848 greg 2.9 /* use rt to improve gradient calc */
849     if (ar.rt > FTINY && ar.rt < FHUGE)
850     dp->r += 1.0/ar.rt;
851 greg 1.1 /* (re)initialize error */
852     if (dp->n++) {
853     b2 = bright(dp->v)/dp->n - bright(ar.rcol);
854     b2 = b2*b2 + dp->k*((dp->n-1)*(dp->n-1));
855     dp->k = b2/(dp->n*dp->n);
856     } else
857     dp->k = 0.0;
858 greg 1.4 return(0);
859 greg 1.1 }
860    
861    
862 greg 2.14 static int
863     ambcmp( /* decreasing order */
864     const void *p1,
865     const void *p2
866     )
867     {
868     const AMBSAMP *d1 = (const AMBSAMP *)p1;
869     const AMBSAMP *d2 = (const AMBSAMP *)p2;
870    
871     if (d1->k < d2->k)
872     return(1);
873     if (d1->k > d2->k)
874     return(-1);
875     return(0);
876     }
877    
878    
879     static int
880     ambnorm( /* standard order */
881     const void *p1,
882     const void *p2
883     )
884     {
885     const AMBSAMP *d1 = (const AMBSAMP *)p1;
886     const AMBSAMP *d2 = (const AMBSAMP *)p2;
887 greg 2.23 int c;
888 greg 2.14
889     if ( (c = d1->t - d2->t) )
890     return(c);
891     return(d1->p - d2->p);
892     }
893    
894    
895 greg 1.1 double
896 greg 2.14 doambient( /* compute ambient component */
897 greg 2.23 COLOR rcol,
898 greg 2.14 RAY *r,
899     double wt,
900     FVECT pg,
901     FVECT dg
902     )
903 greg 1.1 {
904 greg 2.24 double b, d=0;
905 greg 1.1 AMBHEMI hemi;
906     AMBSAMP *div;
907     AMBSAMP dnew;
908 greg 2.23 double acol[3];
909     AMBSAMP *dp;
910 greg 1.1 double arad;
911 greg 2.19 int divcnt;
912 greg 2.23 int i, j;
913 greg 1.1 /* initialize hemisphere */
914 greg 2.23 inithemi(&hemi, rcol, r, wt);
915 greg 2.19 divcnt = hemi.nt * hemi.np;
916 greg 2.17 /* initialize */
917     if (pg != NULL)
918     pg[0] = pg[1] = pg[2] = 0.0;
919     if (dg != NULL)
920     dg[0] = dg[1] = dg[2] = 0.0;
921 greg 2.23 setcolor(rcol, 0.0, 0.0, 0.0);
922 greg 2.19 if (divcnt == 0)
923 greg 1.1 return(0.0);
924 greg 2.14 /* allocate super-samples */
925 greg 2.15 if (hemi.ns > 0 || pg != NULL || dg != NULL) {
926 greg 2.19 div = (AMBSAMP *)malloc(divcnt*sizeof(AMBSAMP));
927 greg 1.1 if (div == NULL)
928     error(SYSTEM, "out of memory in doambient");
929     } else
930     div = NULL;
931     /* sample the divisions */
932     arad = 0.0;
933 greg 2.23 acol[0] = acol[1] = acol[2] = 0.0;
934 greg 1.1 if ((dp = div) == NULL)
935     dp = &dnew;
936 greg 2.19 divcnt = 0;
937 greg 1.1 for (i = 0; i < hemi.nt; i++)
938     for (j = 0; j < hemi.np; j++) {
939     dp->t = i; dp->p = j;
940     setcolor(dp->v, 0.0, 0.0, 0.0);
941 greg 1.2 dp->r = 0.0;
942 greg 1.1 dp->n = 0;
943 greg 2.16 if (divsample(dp, &hemi, r) < 0) {
944 greg 2.19 if (div != NULL)
945     dp++;
946 greg 2.16 continue;
947     }
948 greg 2.6 arad += dp->r;
949 greg 2.19 divcnt++;
950 greg 1.1 if (div != NULL)
951     dp++;
952 greg 2.6 else
953 greg 1.1 addcolor(acol, dp->v);
954     }
955 greg 2.21 if (!divcnt) {
956     if (div != NULL)
957     free((void *)div);
958 greg 2.19 return(0.0); /* no samples taken */
959 greg 2.21 }
960 greg 2.19 if (divcnt < hemi.nt*hemi.np) {
961     pg = dg = NULL; /* incomplete sampling */
962     hemi.ns = 0;
963     } else if (arad > FTINY && divcnt/arad < minarad) {
964 greg 2.15 hemi.ns = 0; /* close enough */
965 greg 2.19 } else if (hemi.ns > 0) { /* else perform super-sampling? */
966 greg 1.4 comperrs(div, &hemi); /* compute errors */
967 greg 2.19 qsort(div, divcnt, sizeof(AMBSAMP), ambcmp); /* sort divs */
968 greg 1.1 /* super-sample */
969 greg 2.15 for (i = hemi.ns; i > 0; i--) {
970 schorsch 2.11 dnew = *div;
971 greg 2.16 if (divsample(&dnew, &hemi, r) < 0) {
972     dp++;
973     continue;
974     }
975     dp = div; /* reinsert */
976 greg 2.19 j = divcnt < i ? divcnt : i;
977 greg 1.1 while (--j > 0 && dnew.k < dp[1].k) {
978 schorsch 2.11 *dp = *(dp+1);
979 greg 1.1 dp++;
980     }
981 schorsch 2.11 *dp = dnew;
982 greg 1.1 }
983 greg 1.2 if (pg != NULL || dg != NULL) /* restore order */
984 greg 2.19 qsort(div, divcnt, sizeof(AMBSAMP), ambnorm);
985 greg 1.1 }
986     /* compute returned values */
987 greg 1.3 if (div != NULL) {
988 greg 2.19 arad = 0.0; /* note: divcnt may be < nt*np */
989     for (i = hemi.nt*hemi.np, dp = div; i-- > 0; dp++) {
990 greg 1.3 arad += dp->r;
991     if (dp->n > 1) {
992     b = 1.0/dp->n;
993     scalecolor(dp->v, b);
994     dp->r *= b;
995     dp->n = 1;
996     }
997     addcolor(acol, dp->v);
998     }
999 greg 1.5 b = bright(acol);
1000 greg 1.6 if (b > FTINY) {
1001 greg 2.17 b = 1.0/b; /* compute & normalize gradient(s) */
1002 greg 1.6 if (pg != NULL) {
1003     posgradient(pg, div, &hemi);
1004     for (i = 0; i < 3; i++)
1005     pg[i] *= b;
1006     }
1007     if (dg != NULL) {
1008     dirgradient(dg, div, &hemi);
1009     for (i = 0; i < 3; i++)
1010     dg[i] *= b;
1011     }
1012 greg 1.5 }
1013 greg 2.9 free((void *)div);
1014 greg 1.3 }
1015 greg 2.23 copycolor(rcol, acol);
1016 greg 1.1 if (arad <= FTINY)
1017 greg 1.16 arad = maxarad;
1018 greg 2.3 else
1019 greg 2.19 arad = (divcnt+hemi.ns)/arad;
1020 greg 1.15 if (pg != NULL) { /* reduce radius if gradient large */
1021     d = DOT(pg,pg);
1022     if (d*arad*arad > 1.0)
1023     arad = 1.0/sqrt(d);
1024     }
1025 greg 1.16 if (arad < minarad) {
1026 greg 1.1 arad = minarad;
1027 greg 1.16 if (pg != NULL && d*arad*arad > 1.0) { /* cap gradient */
1028     d = 1.0/arad/sqrt(d);
1029     for (i = 0; i < 3; i++)
1030     pg[i] *= d;
1031     }
1032     }
1033 greg 2.3 if ((arad /= sqrt(wt)) > maxarad)
1034     arad = maxarad;
1035     return(arad);
1036 greg 1.1 }
1037    
1038    
1039 greg 2.9 void
1040 greg 2.14 comperrs( /* compute initial error estimates */
1041     AMBSAMP *da, /* assumes standard ordering */
1042 greg 2.23 AMBHEMI *hp
1043 greg 2.14 )
1044 greg 1.1 {
1045     double b, b2;
1046     int i, j;
1047 greg 2.23 AMBSAMP *dp;
1048 greg 1.1 /* sum differences from neighbors */
1049     dp = da;
1050     for (i = 0; i < hp->nt; i++)
1051     for (j = 0; j < hp->np; j++) {
1052 greg 1.6 #ifdef DEBUG
1053     if (dp->t != i || dp->p != j)
1054     error(CONSISTENCY,
1055     "division order in comperrs");
1056     #endif
1057 greg 1.1 b = bright(dp[0].v);
1058     if (i > 0) { /* from above */
1059     b2 = bright(dp[-hp->np].v) - b;
1060     b2 *= b2 * 0.25;
1061     dp[0].k += b2;
1062     dp[-hp->np].k += b2;
1063     }
1064     if (j > 0) { /* from behind */
1065     b2 = bright(dp[-1].v) - b;
1066     b2 *= b2 * 0.25;
1067     dp[0].k += b2;
1068     dp[-1].k += b2;
1069 greg 1.4 } else { /* around */
1070     b2 = bright(dp[hp->np-1].v) - b;
1071 greg 1.1 b2 *= b2 * 0.25;
1072     dp[0].k += b2;
1073 greg 1.4 dp[hp->np-1].k += b2;
1074 greg 1.1 }
1075     dp++;
1076     }
1077     /* divide by number of neighbors */
1078     dp = da;
1079     for (j = 0; j < hp->np; j++) /* top row */
1080     (dp++)->k *= 1.0/3.0;
1081     if (hp->nt < 2)
1082     return;
1083     for (i = 1; i < hp->nt-1; i++) /* central region */
1084     for (j = 0; j < hp->np; j++)
1085     (dp++)->k *= 0.25;
1086     for (j = 0; j < hp->np; j++) /* bottom row */
1087     (dp++)->k *= 1.0/3.0;
1088     }
1089    
1090    
1091 greg 2.9 void
1092 greg 2.14 posgradient( /* compute position gradient */
1093     FVECT gv,
1094     AMBSAMP *da, /* assumes standard ordering */
1095 greg 2.23 AMBHEMI *hp
1096 greg 2.14 )
1097 greg 1.1 {
1098 greg 2.23 int i, j;
1099 greg 2.2 double nextsine, lastsine, b, d;
1100 greg 1.2 double mag0, mag1;
1101     double phi, cosp, sinp, xd, yd;
1102 greg 2.23 AMBSAMP *dp;
1103 greg 1.2
1104     xd = yd = 0.0;
1105     for (j = 0; j < hp->np; j++) {
1106     dp = da + j;
1107     mag0 = mag1 = 0.0;
1108 greg 2.2 lastsine = 0.0;
1109 greg 1.2 for (i = 0; i < hp->nt; i++) {
1110     #ifdef DEBUG
1111     if (dp->t != i || dp->p != j)
1112     error(CONSISTENCY,
1113     "division order in posgradient");
1114     #endif
1115     b = bright(dp->v);
1116     if (i > 0) {
1117     d = dp[-hp->np].r;
1118     if (dp[0].r > d) d = dp[0].r;
1119 greg 2.2 /* sin(t)*cos(t)^2 */
1120     d *= lastsine * (1.0 - (double)i/hp->nt);
1121 greg 1.2 mag0 += d*(b - bright(dp[-hp->np].v));
1122     }
1123 greg 2.2 nextsine = sqrt((double)(i+1)/hp->nt);
1124 greg 1.2 if (j > 0) {
1125     d = dp[-1].r;
1126     if (dp[0].r > d) d = dp[0].r;
1127 greg 2.2 mag1 += d * (nextsine - lastsine) *
1128     (b - bright(dp[-1].v));
1129 greg 1.2 } else {
1130     d = dp[hp->np-1].r;
1131     if (dp[0].r > d) d = dp[0].r;
1132 greg 2.2 mag1 += d * (nextsine - lastsine) *
1133     (b - bright(dp[hp->np-1].v));
1134 greg 1.2 }
1135     dp += hp->np;
1136 greg 2.2 lastsine = nextsine;
1137 greg 1.2 }
1138 greg 2.2 mag0 *= 2.0*PI / hp->np;
1139 greg 1.2 phi = 2.0*PI * (double)j/hp->np;
1140 gwlarson 2.8 cosp = tcos(phi); sinp = tsin(phi);
1141 greg 1.2 xd += mag0*cosp - mag1*sinp;
1142     yd += mag0*sinp + mag1*cosp;
1143     }
1144     for (i = 0; i < 3; i++)
1145 greg 2.16 gv[i] = (xd*hp->ux[i] + yd*hp->uy[i])*(hp->nt*hp->np)/PI;
1146 greg 1.1 }
1147    
1148    
1149 greg 2.9 void
1150 greg 2.14 dirgradient( /* compute direction gradient */
1151     FVECT gv,
1152     AMBSAMP *da, /* assumes standard ordering */
1153 greg 2.23 AMBHEMI *hp
1154 greg 2.14 )
1155 greg 1.1 {
1156 greg 2.23 int i, j;
1157 greg 1.2 double mag;
1158     double phi, xd, yd;
1159 greg 2.23 AMBSAMP *dp;
1160 greg 1.2
1161     xd = yd = 0.0;
1162     for (j = 0; j < hp->np; j++) {
1163     dp = da + j;
1164     mag = 0.0;
1165     for (i = 0; i < hp->nt; i++) {
1166     #ifdef DEBUG
1167     if (dp->t != i || dp->p != j)
1168     error(CONSISTENCY,
1169     "division order in dirgradient");
1170     #endif
1171 greg 2.2 /* tan(t) */
1172     mag += bright(dp->v)/sqrt(hp->nt/(i+.5) - 1.0);
1173 greg 1.2 dp += hp->np;
1174     }
1175     phi = 2.0*PI * (j+.5)/hp->np + PI/2.0;
1176 gwlarson 2.8 xd += mag * tcos(phi);
1177     yd += mag * tsin(phi);
1178 greg 1.2 }
1179     for (i = 0; i < 3; i++)
1180 greg 2.16 gv[i] = xd*hp->ux[i] + yd*hp->uy[i];
1181 greg 1.1 }
1182 greg 2.25
1183     #endif /* ! NEWAMB */