ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambient.c
Revision: 2.111
Committed: Thu Nov 18 19:38:21 2021 UTC (2 years, 5 months ago) by greg
Content type: text/plain
Branch: MAIN
Changes since 2.110: +7 -3 lines
Log Message:
perf(rpict, rtrace, rvu): improved function of -aw rendering parameter

File Contents

# Content
1 static const char RCSid[] = "$Id: ambient.c,v 2.110 2021/02/19 22:05:46 greg Exp $";
2 /*
3 * ambient.c - routines dealing with ambient (inter-reflected) component.
4 *
5 * Declarations of external symbols in ambient.h
6 */
7
8 #include "copyright.h"
9
10 #include <string.h>
11
12 #include "platform.h"
13 #include "ray.h"
14 #include "otypes.h"
15 #include "otspecial.h"
16 #include "resolu.h"
17 #include "ambient.h"
18 #include "random.h"
19 #include "pmapamb.h"
20
21 #ifndef OCTSCALE
22 #define OCTSCALE 1.0 /* ceil((valid rad.)/(cube size)) */
23 #endif
24
25 extern char *shm_boundary; /* memory sharing boundary */
26
27 #ifndef MAXASET
28 #define MAXASET 4095 /* maximum number of elements in ambient set */
29 #endif
30 OBJECT ambset[MAXASET+1]={0}; /* ambient include/exclude set */
31
32 double maxarad; /* maximum ambient radius */
33 double minarad; /* minimum ambient radius */
34
35 static AMBTREE atrunk; /* our ambient trunk node */
36
37 static FILE *ambfp = NULL; /* ambient file pointer */
38 static int nunflshed = 0; /* number of unflushed ambient values */
39
40 #ifndef SORT_THRESH
41 #ifdef SMLMEM
42 #define SORT_THRESH ((16L<<20)/sizeof(AMBVAL))
43 #else
44 #define SORT_THRESH ((64L<<20)/sizeof(AMBVAL))
45 #endif
46 #endif
47 #ifndef SORT_INTVL
48 #define SORT_INTVL (SORT_THRESH<<1)
49 #endif
50 #ifndef MAX_SORT_INTVL
51 #define MAX_SORT_INTVL (SORT_INTVL<<6)
52 #endif
53
54
55 static double avsum = 0.; /* computed ambient value sum (log) */
56 static unsigned int navsum = 0; /* number of values in avsum */
57 static unsigned int nambvals = 0; /* total number of indirect values */
58 static unsigned int nambshare = 0; /* number of values from file */
59 static unsigned long ambclock = 0; /* ambient access clock */
60 static unsigned long lastsort = 0; /* time of last value sort */
61 static long sortintvl = SORT_INTVL; /* time until next sort */
62 static FILE *ambinp = NULL; /* auxiliary file for input */
63 static long lastpos = -1; /* last flush position */
64
65 #define MAXACLOCK (1L<<30) /* clock turnover value */
66 /*
67 * Track access times unless we are sharing ambient values
68 * through memory on a multiprocessor, when we want to avoid
69 * claiming our own memory (copy on write). Go ahead anyway
70 * if more than two thirds of our values are unshared.
71 * Compile with -Dtracktime=0 to turn this code off.
72 */
73 #ifndef tracktime
74 #define tracktime (shm_boundary == NULL || nambvals > 3*nambshare)
75 #endif
76
77 #define AMBFLUSH (BUFSIZ/AMBVALSIZ)
78
79 #define newambval() (AMBVAL *)malloc(sizeof(AMBVAL))
80
81 #define tfunc(lwr, x, upr) (((x)-(lwr))/((upr)-(lwr)))
82
83 static void initambfile(int creat);
84 static void avsave(AMBVAL *av);
85 static AMBVAL *avstore(AMBVAL *aval);
86 static AMBTREE *newambtree(void);
87 static void freeambtree(AMBTREE *atp);
88
89 typedef void unloadtf_t(AMBVAL *);
90 static unloadtf_t avinsert;
91 static unloadtf_t av2list;
92 static unloadtf_t avfree;
93 static void unloadatree(AMBTREE *at, unloadtf_t *f);
94
95 static int aposcmp(const void *avp1, const void *avp2);
96 static int avlmemi(AMBVAL *avaddr);
97 static void sortambvals(int always);
98
99 static int plugaleak(RAY *r, AMBVAL *ap, FVECT anorm, double ang);
100 static double sumambient(COLOR acol, RAY *r, FVECT rn, int al,
101 AMBTREE *at, FVECT c0, double s);
102 static int makeambient(COLOR acol, RAY *r, FVECT rn, int al);
103 static int extambient(COLOR cr, AMBVAL *ap, FVECT pv, FVECT nv,
104 FVECT uvw[3]);
105
106 #ifdef F_SETLKW
107 static void aflock(int typ);
108 #endif
109
110
111 void
112 setambres( /* set ambient resolution */
113 int ar
114 )
115 {
116 ambres = ar < 0 ? 0 : ar; /* may be done already */
117 /* set min & max radii */
118 if (ar <= 0) {
119 minarad = 0;
120 maxarad = thescene.cusize*0.2;
121 } else {
122 minarad = thescene.cusize / ar;
123 maxarad = 64.0 * minarad; /* heuristic */
124 if (maxarad > thescene.cusize*0.2)
125 maxarad = thescene.cusize*0.2;
126 }
127 if (minarad <= FTINY)
128 minarad = 10.0*FTINY;
129 if (maxarad <= minarad)
130 maxarad = 64.0 * minarad;
131 }
132
133
134 void
135 setambacc( /* set ambient accuracy */
136 double newa
137 )
138 {
139 static double olda; /* remember previous setting here */
140
141 newa *= (newa > 0);
142 if (fabs(newa - olda) >= .05*(newa + olda)) {
143 ambacc = newa;
144 if (ambacc > FTINY && nambvals > 0)
145 sortambvals(1); /* rebuild tree */
146 }
147 }
148
149
150 void
151 setambient(void) /* initialize calculation */
152 {
153 int readonly = 0;
154 long flen;
155 AMBVAL amb;
156 /* make sure we're fresh */
157 ambdone();
158 /* init ambient limits */
159 setambres(ambres);
160 setambacc(ambacc);
161 if (ambfile == NULL || !ambfile[0])
162 return;
163 if (ambacc <= FTINY) {
164 sprintf(errmsg, "zero ambient accuracy so \"%s\" not opened",
165 ambfile);
166 error(WARNING, errmsg);
167 return;
168 }
169 /* open ambient file */
170 if ((ambfp = fopen(ambfile, "r+")) == NULL)
171 readonly = (ambfp = fopen(ambfile, "r")) != NULL;
172 if (ambfp != NULL) {
173 initambfile(0); /* file exists */
174 lastpos = ftell(ambfp);
175 while (readambval(&amb, ambfp))
176 avstore(&amb);
177 nambshare = nambvals; /* share loaded values */
178 if (readonly) {
179 sprintf(errmsg,
180 "loaded %u values from read-only ambient file",
181 nambvals);
182 error(WARNING, errmsg);
183 fclose(ambfp); /* close file so no writes */
184 ambfp = NULL;
185 return; /* avoid ambsync() */
186 }
187 /* align file pointer */
188 lastpos += (long)nambvals*AMBVALSIZ;
189 flen = lseek(fileno(ambfp), (off_t)0, SEEK_END);
190 if (flen != lastpos) {
191 sprintf(errmsg,
192 "ignoring last %ld values in ambient file (corrupted)",
193 (flen - lastpos)/AMBVALSIZ);
194 error(WARNING, errmsg);
195 fseek(ambfp, lastpos, SEEK_SET);
196 ftruncate(fileno(ambfp), (off_t)lastpos);
197 }
198 } else if ((ambfp = fopen(ambfile, "w+")) != NULL) {
199 initambfile(1); /* else create new file */
200 fflush(ambfp);
201 lastpos = ftell(ambfp);
202 } else {
203 sprintf(errmsg, "cannot open ambient file \"%s\"", ambfile);
204 error(SYSTEM, errmsg);
205 }
206 #ifdef F_SETLKW
207 aflock(F_UNLCK); /* release file */
208 #endif
209 }
210
211
212 void
213 ambdone(void) /* close ambient file and free memory */
214 {
215 if (ambfp != NULL) { /* close ambient file */
216 ambsync();
217 fclose(ambfp);
218 ambfp = NULL;
219 if (ambinp != NULL) {
220 fclose(ambinp);
221 ambinp = NULL;
222 }
223 lastpos = -1;
224 }
225 /* free ambient tree */
226 unloadatree(&atrunk, avfree);
227 /* reset state variables */
228 avsum = 0.;
229 navsum = 0;
230 nambvals = 0;
231 nambshare = 0;
232 ambclock = 0;
233 lastsort = 0;
234 sortintvl = SORT_INTVL;
235 }
236
237
238 void
239 ambnotify( /* record new modifier */
240 OBJECT obj
241 )
242 {
243 static int hitlimit = 0;
244 OBJREC *o;
245 char **amblp;
246
247 if (obj == OVOID) { /* starting over */
248 ambset[0] = 0;
249 hitlimit = 0;
250 return;
251 }
252 o = objptr(obj);
253 if (hitlimit || !ismodifier(o->otype))
254 return;
255 for (amblp = amblist; *amblp != NULL; amblp++)
256 if (!strcmp(o->oname, *amblp)) {
257 if (ambset[0] >= MAXASET) {
258 error(WARNING, "too many modifiers in ambient list");
259 hitlimit++;
260 return; /* should this be fatal? */
261 }
262 insertelem(ambset, obj);
263 return;
264 }
265 }
266
267
268 void
269 multambient( /* compute ambient component & multiply by coef. */
270 COLOR aval,
271 RAY *r,
272 FVECT nrm
273 )
274 {
275 static double logAvgAbsorp = 1;
276 static int rdepth = 0; /* ambient recursion */
277 COLOR acol, caustic;
278 int i, ok;
279 double d, l;
280
281 /* PMAP: Factor in ambient from photon map, if enabled and ray is
282 * ambient. Return as all ambient components accounted for, else
283 * continue. */
284 if (ambPmap(aval, r, rdepth))
285 return;
286
287 if (logAvgAbsorp > 0) /* exclude in -aw to avoid growth */
288 logAvgAbsorp = log(1.-AVGREFL);
289
290 /* PMAP: Factor in specular-diffuse ambient (caustics) from photon
291 * map, if enabled and ray is primary, else caustic is zero. Continue
292 * with RADIANCE ambient calculation */
293 copycolor(caustic, aval);
294 ambPmapCaustic(caustic, r, rdepth);
295
296 if (ambdiv <= 0) /* no ambient calculation */
297 goto dumbamb;
298 /* check number of bounces */
299 if (rdepth >= ambounce)
300 goto dumbamb;
301 /* check ambient list */
302 if (ambincl != -1 && r->ro != NULL &&
303 ambincl != inset(ambset, r->ro->omod))
304 goto dumbamb;
305
306 if (ambacc <= FTINY) { /* no ambient storage */
307 FVECT uvd[2];
308 float dgrad[2], *dgp = NULL;
309
310 if (nrm != r->ron && DOT(nrm,r->ron) < 0.9999)
311 dgp = dgrad; /* compute rotational grad. */
312 copycolor(acol, aval);
313 rdepth++;
314 ok = doambient(acol, r, r->rweight,
315 uvd, NULL, NULL, dgp, NULL);
316 rdepth--;
317 if (!ok)
318 goto dumbamb;
319 if ((ok > 0) & (dgp != NULL)) { /* apply texture */
320 FVECT v1;
321 VCROSS(v1, r->ron, nrm);
322 d = 1.0;
323 for (i = 3; i--; )
324 d += v1[i] * (dgp[0]*uvd[0][i] + dgp[1]*uvd[1][i]);
325 if (d >= 0.05)
326 scalecolor(acol, d);
327 }
328 copycolor(aval, acol);
329
330 /* PMAP: add in caustic */
331 addcolor(aval, caustic);
332 return;
333 }
334
335 if (tracktime) /* sort to minimize thrashing */
336 sortambvals(0);
337 /* interpolate ambient value */
338 setcolor(acol, 0.0, 0.0, 0.0);
339 d = sumambient(acol, r, nrm, rdepth,
340 &atrunk, thescene.cuorg, thescene.cusize);
341
342 if (d > FTINY) {
343 d = 1.0/d;
344 scalecolor(acol, d);
345 multcolor(aval, acol);
346
347 /* PMAP: add in caustic */
348 addcolor(aval, caustic);
349 return;
350 }
351
352 rdepth++; /* need to cache new value */
353 ok = makeambient(acol, r, nrm, rdepth-1);
354 rdepth--;
355
356 if (ok) {
357 multcolor(aval, acol); /* computed new value */
358
359 /* PMAP: add in caustic */
360 addcolor(aval, caustic);
361 return;
362 }
363
364 dumbamb: /* return global value */
365 if ((ambvwt <= 0) | (navsum == 0)) {
366 multcolor(aval, ambval);
367
368 /* PMAP: add in caustic */
369 addcolor(aval, caustic);
370 return;
371 }
372
373 l = bright(ambval); /* average in computations */
374 if (l > FTINY) {
375 d = (log(l)*(double)ambvwt + avsum + logAvgAbsorp*navsum) /
376 (double)(ambvwt + navsum);
377 d = exp(d) / l;
378 scalecolor(aval, d);
379 multcolor(aval, ambval); /* apply color of ambval */
380 } else {
381 d = exp( avsum/(double)navsum + logAvgAbsorp );
382 scalecolor(aval, d); /* neutral color */
383 }
384 }
385
386
387 /* Plug a potential leak where ambient cache value is occluded */
388 static int
389 plugaleak(RAY *r, AMBVAL *ap, FVECT anorm, double ang)
390 {
391 const double cost70sq = 0.1169778; /* cos(70deg)^2 */
392 RAY rtst;
393 FVECT vdif;
394 double normdot, ndotd, nadotd;
395 double a, b, c, t[2];
396
397 ang += 2.*PI*(ang < 0); /* check direction flags */
398 if ( !(ap->corral>>(int)(ang*(16./PI)) & 1) )
399 return(0);
400 /*
401 * Generate test ray, targeting 20 degrees above sample point plane
402 * along surface normal from cache position. This should be high
403 * enough to miss local geometry we don't really care about.
404 */
405 VSUB(vdif, ap->pos, r->rop);
406 normdot = DOT(anorm, r->ron);
407 ndotd = DOT(vdif, r->ron);
408 nadotd = DOT(vdif, anorm);
409 a = normdot*normdot - cost70sq;
410 b = 2.0*(normdot*ndotd - nadotd*cost70sq);
411 c = ndotd*ndotd - DOT(vdif,vdif)*cost70sq;
412 if (quadratic(t, a, b, c) != 2)
413 return(1); /* should rarely happen */
414 if (t[1] <= FTINY)
415 return(0); /* should fail behind test */
416 rayorigin(&rtst, SHADOW, r, NULL);
417 VSUM(rtst.rdir, vdif, anorm, t[1]); /* further dist. > plane */
418 rtst.rmax = normalize(rtst.rdir); /* short ray test */
419 while (localhit(&rtst, &thescene)) { /* check for occluder */
420 OBJREC *m = findmaterial(rtst.ro);
421 if (m != NULL && !istransp(m->otype) && !isBSDFproxy(m) &&
422 (rtst.clipset == NULL ||
423 !inset(rtst.clipset, rtst.ro->omod)))
424 return(1); /* plug light leak */
425 VCOPY(rtst.rorg, rtst.rop); /* skip invisible surface */
426 rtst.rmax -= rtst.rot;
427 rayclear(&rtst);
428 }
429 return(0); /* seems we're OK */
430 }
431
432
433 static double
434 sumambient( /* get interpolated ambient value */
435 COLOR acol,
436 RAY *r,
437 FVECT rn,
438 int al,
439 AMBTREE *at,
440 FVECT c0,
441 double s
442 )
443 { /* initial limit is 10 degrees plus ambacc radians */
444 const double minangle = 10.0 * PI/180.;
445 double maxangle = minangle + ambacc;
446 double wsum = 0.0;
447 FVECT ck0;
448 int i, j;
449 AMBVAL *av;
450
451 if (at->kid != NULL) { /* sum children first */
452 s *= 0.5;
453 for (i = 0; i < 8; i++) {
454 for (j = 0; j < 3; j++) {
455 ck0[j] = c0[j];
456 if (1<<j & i)
457 ck0[j] += s;
458 if (r->rop[j] < ck0[j] - OCTSCALE*s)
459 break;
460 if (r->rop[j] > ck0[j] + (1.0+OCTSCALE)*s)
461 break;
462 }
463 if (j == 3)
464 wsum += sumambient(acol, r, rn, al,
465 at->kid+i, ck0, s);
466 }
467 /* good enough? */
468 if (wsum >= 0.05 && s > minarad*10.0)
469 return(wsum);
470 }
471 /* adjust maximum angle */
472 if (at->alist != NULL && (at->alist->lvl <= al) & (r->rweight < 0.6))
473 maxangle = (maxangle - PI/2.)*pow(r->rweight,0.13) + PI/2.;
474 /* sum this node */
475 for (av = at->alist; av != NULL; av = av->next) {
476 double u, v, d, delta_r2, delta_t2;
477 COLOR ct;
478 FVECT uvw[3];
479 /* record access */
480 if (tracktime)
481 av->latick = ambclock;
482 /*
483 * Ambient level test
484 */
485 if (av->lvl > al || /* list sorted, so this works */
486 (av->lvl == al) & (av->weight < 0.9*r->rweight))
487 break;
488 /*
489 * Direction test using unperturbed normal
490 */
491 decodedir(uvw[2], av->ndir);
492 d = DOT(uvw[2], r->ron);
493 if (d <= 0.0) /* >= 90 degrees */
494 continue;
495 delta_r2 = 2.0 - 2.0*d; /* approx. radians^2 */
496 if (delta_r2 >= maxangle*maxangle)
497 continue;
498 /*
499 * Modified ray behind test
500 */
501 VSUB(ck0, r->rop, av->pos);
502 d = DOT(ck0, uvw[2]);
503 if (d < -minarad*ambacc-.001)
504 continue;
505 d /= av->rad[0];
506 delta_t2 = d*d;
507 if (delta_t2 >= ambacc*ambacc)
508 continue;
509 /*
510 * Elliptical radii test based on Hessian
511 */
512 decodedir(uvw[0], av->udir);
513 VCROSS(uvw[1], uvw[2], uvw[0]);
514 d = (u = DOT(ck0, uvw[0])) / av->rad[0];
515 delta_t2 += d*d;
516 d = (v = DOT(ck0, uvw[1])) / av->rad[1];
517 delta_t2 += d*d;
518 if (delta_t2 >= ambacc*ambacc)
519 continue;
520 /*
521 * Test for potential light leak
522 */
523 if (av->corral && plugaleak(r, av, uvw[2], atan2a(v,u)))
524 continue;
525 /*
526 * Extrapolate value and compute final weight (hat function)
527 */
528 if (!extambient(ct, av, r->rop, rn, uvw))
529 continue;
530 d = tfunc(maxangle, sqrt(delta_r2), 0.0) *
531 tfunc(ambacc, sqrt(delta_t2), 0.0);
532 scalecolor(ct, d);
533 addcolor(acol, ct);
534 wsum += d;
535 }
536 return(wsum);
537 }
538
539
540 static int
541 makeambient( /* make a new ambient value for storage */
542 COLOR acol,
543 RAY *r,
544 FVECT rn,
545 int al
546 )
547 {
548 AMBVAL amb;
549 FVECT uvw[3];
550 int i;
551
552 amb.weight = 1.0; /* compute weight */
553 for (i = al; i-- > 0; )
554 amb.weight *= AVGREFL;
555 if (r->rweight < 0.1*amb.weight) /* heuristic override */
556 amb.weight = 1.25*r->rweight;
557 setcolor(acol, AVGREFL, AVGREFL, AVGREFL);
558 /* compute ambient */
559 i = doambient(acol, r, amb.weight,
560 uvw, amb.rad, amb.gpos, amb.gdir, &amb.corral);
561 scalecolor(acol, 1./AVGREFL); /* undo assumed reflectance */
562 if (i <= 0 || amb.rad[0] <= FTINY) /* no Hessian or zero radius */
563 return(i);
564 /* store value */
565 VCOPY(amb.pos, r->rop);
566 amb.ndir = encodedir(r->ron);
567 amb.udir = encodedir(uvw[0]);
568 amb.lvl = al;
569 copycolor(amb.val, acol);
570 /* insert into tree */
571 avsave(&amb); /* and save to file */
572 if (rn != r->ron) { /* texture */
573 VCOPY(uvw[2], r->ron);
574 extambient(acol, &amb, r->rop, rn, uvw);
575 }
576 return(1);
577 }
578
579
580 static int
581 extambient( /* extrapolate value at pv, nv */
582 COLOR cr,
583 AMBVAL *ap,
584 FVECT pv,
585 FVECT nv,
586 FVECT uvw[3]
587 )
588 {
589 const double min_d = 0.05;
590 const double max_d = 20.;
591 static FVECT my_uvw[3];
592 FVECT v1;
593 int i;
594 double d = 1.0; /* zeroeth order */
595
596 if (uvw == NULL) { /* need local coordinates? */
597 decodedir(my_uvw[2], ap->ndir);
598 decodedir(my_uvw[0], ap->udir);
599 VCROSS(my_uvw[1], my_uvw[2], my_uvw[0]);
600 uvw = my_uvw;
601 }
602 for (i = 3; i--; ) /* gradient due to translation */
603 d += (pv[i] - ap->pos[i]) *
604 (ap->gpos[0]*uvw[0][i] + ap->gpos[1]*uvw[1][i]);
605
606 VCROSS(v1, uvw[2], nv); /* gradient due to rotation */
607 for (i = 3; i--; )
608 d += v1[i] * (ap->gdir[0]*uvw[0][i] + ap->gdir[1]*uvw[1][i]);
609
610 if (d < min_d) /* clamp min/max scaling */
611 d = min_d;
612 else if (d > max_d)
613 d = max_d;
614 copycolor(cr, ap->val);
615 scalecolor(cr, d);
616 return(d > min_d);
617 }
618
619
620 static void
621 avinsert( /* insert ambient value in our tree */
622 AMBVAL *av
623 )
624 {
625 AMBTREE *at;
626 AMBVAL *ap;
627 AMBVAL avh;
628 FVECT ck0;
629 double s;
630 int branch;
631 int i;
632
633 if (av->rad[0] <= FTINY)
634 error(CONSISTENCY, "zero ambient radius in avinsert");
635 at = &atrunk;
636 VCOPY(ck0, thescene.cuorg);
637 s = thescene.cusize;
638 while (s*(OCTSCALE/2) > av->rad[1]*ambacc) {
639 if (at->kid == NULL)
640 if ((at->kid = newambtree()) == NULL)
641 error(SYSTEM, "out of memory in avinsert");
642 s *= 0.5;
643 branch = 0;
644 for (i = 0; i < 3; i++)
645 if (av->pos[i] > ck0[i] + s) {
646 ck0[i] += s;
647 branch |= 1 << i;
648 }
649 at = at->kid + branch;
650 }
651 avh.next = at->alist; /* order by increasing level */
652 for (ap = &avh; ap->next != NULL; ap = ap->next)
653 if ( ap->next->lvl > av->lvl ||
654 (ap->next->lvl == av->lvl) &
655 (ap->next->weight <= av->weight) )
656 break;
657 av->next = ap->next;
658 ap->next = (AMBVAL*)av;
659 at->alist = avh.next;
660 }
661
662
663 static void
664 initambfile( /* initialize ambient file */
665 int cre8
666 )
667 {
668 extern char *progname, *octname;
669 static char *mybuf = NULL;
670
671 #ifdef F_SETLKW
672 aflock(cre8 ? F_WRLCK : F_RDLCK);
673 #endif
674 SET_FILE_BINARY(ambfp);
675 if (mybuf == NULL)
676 mybuf = (char *)bmalloc(BUFSIZ+8);
677 setbuf(ambfp, mybuf);
678 if (cre8) { /* new file */
679 newheader("RADIANCE", ambfp);
680 fprintf(ambfp, "%s -av %g %g %g -aw %d -ab %d -aa %g ",
681 progname, colval(ambval,RED),
682 colval(ambval,GRN), colval(ambval,BLU),
683 ambvwt, ambounce, ambacc);
684 fprintf(ambfp, "-ad %d -as %d -ar %d ",
685 ambdiv, ambssamp, ambres);
686 if (octname != NULL)
687 fputs(octname, ambfp);
688 fputc('\n', ambfp);
689 fprintf(ambfp, "SOFTWARE= %s\n", VersionID);
690 fputnow(ambfp);
691 fputformat(AMBFMT, ambfp);
692 fputc('\n', ambfp);
693 putambmagic(ambfp);
694 } else if (checkheader(ambfp, AMBFMT, NULL) < 0 || !hasambmagic(ambfp))
695 error(USER, "bad ambient file");
696 }
697
698
699 static void
700 avsave( /* insert and save an ambient value */
701 AMBVAL *av
702 )
703 {
704 avstore(av);
705 if (ambfp == NULL)
706 return;
707 if (writambval(av, ambfp) < 0)
708 goto writerr;
709 if (++nunflshed >= AMBFLUSH)
710 if (ambsync() == EOF)
711 goto writerr;
712 return;
713 writerr:
714 error(SYSTEM, "error writing to ambient file");
715 }
716
717
718 static AMBVAL *
719 avstore( /* allocate memory and save aval */
720 AMBVAL *aval
721 )
722 {
723 AMBVAL *av;
724 double d;
725
726 if ((av = newambval()) == NULL)
727 error(SYSTEM, "out of memory in avstore");
728 *av = *aval;
729 av->latick = ambclock;
730 av->next = NULL;
731 nambvals++;
732 d = bright(av->val);
733 if (d > FTINY) { /* add to log sum for averaging */
734 avsum += log(d);
735 navsum++;
736 }
737 avinsert(av); /* insert in our cache tree */
738 return(av);
739 }
740
741
742 #define ATALLOCSZ 512 /* #/8 trees to allocate at once */
743
744 static AMBTREE *atfreelist = NULL; /* free ambient tree structures */
745
746
747 static AMBTREE *
748 newambtree(void) /* allocate 8 ambient tree structs */
749 {
750 AMBTREE *atp, *upperlim;
751
752 if (atfreelist == NULL) { /* get more nodes */
753 atfreelist = (AMBTREE *)malloc(ATALLOCSZ*8*sizeof(AMBTREE));
754 if (atfreelist == NULL)
755 return(NULL);
756 /* link new free list */
757 upperlim = atfreelist + 8*(ATALLOCSZ-1);
758 for (atp = atfreelist; atp < upperlim; atp += 8)
759 atp->kid = atp + 8;
760 atp->kid = NULL;
761 }
762 atp = atfreelist;
763 atfreelist = atp->kid;
764 memset(atp, 0, 8*sizeof(AMBTREE));
765 return(atp);
766 }
767
768
769 static void
770 freeambtree( /* free 8 ambient tree structs */
771 AMBTREE *atp
772 )
773 {
774 atp->kid = atfreelist;
775 atfreelist = atp;
776 }
777
778
779 static void
780 unloadatree( /* unload an ambient value tree */
781 AMBTREE *at,
782 unloadtf_t *f
783 )
784 {
785 AMBVAL *av;
786 int i;
787 /* transfer values at this node */
788 for (av = at->alist; av != NULL; av = at->alist) {
789 at->alist = av->next;
790 av->next = NULL;
791 (*f)(av);
792 }
793 if (at->kid == NULL)
794 return;
795 for (i = 0; i < 8; i++) /* transfer and free children */
796 unloadatree(at->kid+i, f);
797 freeambtree(at->kid);
798 at->kid = NULL;
799 }
800
801
802 static struct avl {
803 AMBVAL *p;
804 unsigned long t;
805 } *avlist1; /* ambient value list with ticks */
806 static AMBVAL **avlist2; /* memory positions for sorting */
807 static int i_avlist; /* index for lists */
808
809 static int alatcmp(const void *av1, const void *av2);
810
811 static void
812 avfree(AMBVAL *av)
813 {
814 free(av);
815 }
816
817 static void
818 av2list(
819 AMBVAL *av
820 )
821 {
822 #ifdef DEBUG
823 if (i_avlist >= nambvals)
824 error(CONSISTENCY, "too many ambient values in av2list1");
825 #endif
826 avlist1[i_avlist].p = avlist2[i_avlist] = (AMBVAL*)av;
827 avlist1[i_avlist++].t = av->latick;
828 }
829
830
831 static int
832 alatcmp( /* compare ambient values for MRA */
833 const void *av1,
834 const void *av2
835 )
836 {
837 long lc = ((struct avl *)av2)->t - ((struct avl *)av1)->t;
838 return(lc<0 ? -1 : lc>0 ? 1 : 0);
839 }
840
841
842 /* GW NOTE 2002/10/3:
843 * I used to compare AMBVAL pointers, but found that this was the
844 * cause of a serious consistency error with gcc, since the optimizer
845 * uses some dangerous trick in pointer subtraction that
846 * assumes pointers differ by exact struct size increments.
847 */
848 static int
849 aposcmp( /* compare ambient value positions */
850 const void *avp1,
851 const void *avp2
852 )
853 {
854 long diff = *(char * const *)avp1 - *(char * const *)avp2;
855 if (diff < 0)
856 return(-1);
857 return(diff > 0);
858 }
859
860
861 static int
862 avlmemi( /* find list position from address */
863 AMBVAL *avaddr
864 )
865 {
866 AMBVAL **avlpp;
867
868 avlpp = (AMBVAL **)bsearch(&avaddr, avlist2,
869 nambvals, sizeof(AMBVAL *), aposcmp);
870 if (avlpp == NULL)
871 error(CONSISTENCY, "address not found in avlmemi");
872 return(avlpp - avlist2);
873 }
874
875
876 static void
877 sortambvals( /* resort ambient values */
878 int always
879 )
880 {
881 AMBTREE oldatrunk;
882 AMBVAL tav, *tap, *pnext;
883 int i, j;
884 /* see if it's time yet */
885 if (!always && (ambclock++ < lastsort+sortintvl ||
886 nambvals < SORT_THRESH))
887 return;
888 /*
889 * The idea here is to minimize memory thrashing
890 * in VM systems by improving reference locality.
891 * We do this by periodically sorting our stored ambient
892 * values in memory in order of most recently to least
893 * recently accessed. This ordering was chosen so that new
894 * ambient values (which tend to be less important) go into
895 * higher memory with the infrequently accessed values.
896 * Since we expect our values to need sorting less
897 * frequently as the process continues, we double our
898 * waiting interval after each call.
899 * This routine is also called by setambacc() with
900 * the "always" parameter set to 1 so that the ambient
901 * tree will be rebuilt with the new accuracy parameter.
902 */
903 if (tracktime) { /* allocate pointer arrays to sort */
904 avlist2 = (AMBVAL **)malloc(nambvals*sizeof(AMBVAL *));
905 avlist1 = (struct avl *)malloc(nambvals*sizeof(struct avl));
906 } else {
907 avlist2 = NULL;
908 avlist1 = NULL;
909 }
910 if (avlist1 == NULL) { /* no time tracking -- rebuild tree? */
911 if (avlist2 != NULL)
912 free(avlist2);
913 if (always) { /* rebuild without sorting */
914 oldatrunk = atrunk;
915 atrunk.alist = NULL;
916 atrunk.kid = NULL;
917 unloadatree(&oldatrunk, avinsert);
918 }
919 } else { /* sort memory by last access time */
920 /*
921 * Sorting memory is tricky because it isn't contiguous.
922 * We have to sort an array of pointers by MRA and also
923 * by memory position. We then copy values in "loops"
924 * to minimize memory hits. Nevertheless, we will visit
925 * everyone at least twice, and this is an expensive process
926 * when we're thrashing, which is when we need to do it.
927 */
928 #ifdef DEBUG
929 sprintf(errmsg, "sorting %u ambient values at ambclock=%lu...",
930 nambvals, ambclock);
931 eputs(errmsg);
932 #endif
933 i_avlist = 0;
934 unloadatree(&atrunk, av2list); /* empty current tree */
935 #ifdef DEBUG
936 if (i_avlist < nambvals)
937 error(CONSISTENCY, "missing ambient values in sortambvals");
938 #endif
939 qsort(avlist1, nambvals, sizeof(struct avl), alatcmp);
940 qsort(avlist2, nambvals, sizeof(AMBVAL *), aposcmp);
941 for (i = 0; i < nambvals; i++) {
942 if (avlist1[i].p == NULL)
943 continue;
944 tap = avlist2[i];
945 tav = *tap;
946 for (j = i; (pnext = avlist1[j].p) != tap;
947 j = avlmemi(pnext)) {
948 *(avlist2[j]) = *pnext;
949 avinsert(avlist2[j]);
950 avlist1[j].p = NULL;
951 }
952 *(avlist2[j]) = tav;
953 avinsert(avlist2[j]);
954 avlist1[j].p = NULL;
955 }
956 free(avlist1);
957 free(avlist2);
958 /* compute new sort interval */
959 sortintvl = ambclock - lastsort;
960 if (sortintvl >= MAX_SORT_INTVL/2)
961 sortintvl = MAX_SORT_INTVL;
962 else
963 sortintvl <<= 1; /* wait twice as long next */
964 #ifdef DEBUG
965 eputs("done\n");
966 #endif
967 }
968 if (ambclock >= MAXACLOCK)
969 ambclock = MAXACLOCK/2;
970 lastsort = ambclock;
971 }
972
973
974 #ifdef F_SETLKW
975
976 static void
977 aflock( /* lock/unlock ambient file */
978 int typ
979 )
980 {
981 static struct flock fls; /* static so initialized to zeroes */
982
983 if (typ == fls.l_type) /* already called? */
984 return;
985
986 fls.l_type = typ;
987 do
988 if (fcntl(fileno(ambfp), F_SETLKW, &fls) != -1)
989 return;
990 while (errno == EINTR);
991
992 error(SYSTEM, "cannot (un)lock ambient file");
993 }
994
995
996 int
997 ambsync(void) /* synchronize ambient file */
998 {
999 long flen;
1000 AMBVAL avs;
1001 int n;
1002
1003 if (ambfp == NULL) /* no ambient file? */
1004 return(0);
1005 /* gain appropriate access */
1006 aflock(nunflshed ? F_WRLCK : F_RDLCK);
1007 /* see if file has grown */
1008 if ((flen = lseek(fileno(ambfp), (off_t)0, SEEK_END)) < 0)
1009 goto seekerr;
1010 if ((n = flen - lastpos) > 0) { /* file has grown */
1011 if (ambinp == NULL) { /* get new file pointer */
1012 ambinp = fopen(ambfile, "rb");
1013 if (ambinp == NULL)
1014 error(SYSTEM, "fopen failed in ambsync");
1015 }
1016 if (fseek(ambinp, lastpos, SEEK_SET) < 0)
1017 goto seekerr;
1018 while (n >= AMBVALSIZ) { /* load contributed values */
1019 if (!readambval(&avs, ambinp)) {
1020 sprintf(errmsg,
1021 "ambient file \"%s\" corrupted near character %ld",
1022 ambfile, flen - n);
1023 error(WARNING, errmsg);
1024 break;
1025 }
1026 avstore(&avs);
1027 n -= AMBVALSIZ;
1028 }
1029 lastpos = flen - n; /* check alignment */
1030 if (n && lseek(fileno(ambfp), (off_t)lastpos, SEEK_SET) < 0)
1031 goto seekerr;
1032 }
1033 n = fflush(ambfp); /* calls write() at last */
1034 lastpos += (long)nunflshed*AMBVALSIZ;
1035 aflock(F_UNLCK); /* release file */
1036 nunflshed = 0;
1037 return(n);
1038 seekerr:
1039 error(SYSTEM, "seek failed in ambsync");
1040 return(EOF); /* pro forma return */
1041 }
1042
1043 #else /* ! F_SETLKW */
1044
1045 int
1046 ambsync(void) /* flush ambient file */
1047 {
1048 if (ambfp == NULL)
1049 return(0);
1050 nunflshed = 0;
1051 return(fflush(ambfp));
1052 }
1053
1054 #endif /* ! F_SETLKW */