ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/ambient.c
Revision: 2.110
Committed: Fri Feb 19 22:05:46 2021 UTC (3 years, 3 months ago) by greg
Content type: text/plain
Branch: MAIN
Changes since 2.109: +5 -2 lines
Log Message:
perf: put maximum cap on gradient values to avoid over-extrapolation

File Contents

# Content
1 static const char RCSid[] = "$Id: ambient.c,v 2.109 2020/03/10 15:57:52 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 int rdepth = 0; /* ambient recursion */
276 COLOR acol, caustic;
277 int i, ok;
278 double d, l;
279
280 /* PMAP: Factor in ambient from photon map, if enabled and ray is
281 * ambient. Return as all ambient components accounted for, else
282 * continue. */
283 if (ambPmap(aval, r, rdepth))
284 return;
285
286 /* PMAP: Factor in specular-diffuse ambient (caustics) from photon
287 * map, if enabled and ray is primary, else caustic is zero. Continue
288 * with RADIANCE ambient calculation */
289 copycolor(caustic, aval);
290 ambPmapCaustic(caustic, r, rdepth);
291
292 if (ambdiv <= 0) /* no ambient calculation */
293 goto dumbamb;
294 /* check number of bounces */
295 if (rdepth >= ambounce)
296 goto dumbamb;
297 /* check ambient list */
298 if (ambincl != -1 && r->ro != NULL &&
299 ambincl != inset(ambset, r->ro->omod))
300 goto dumbamb;
301
302 if (ambacc <= FTINY) { /* no ambient storage */
303 FVECT uvd[2];
304 float dgrad[2], *dgp = NULL;
305
306 if (nrm != r->ron && DOT(nrm,r->ron) < 0.9999)
307 dgp = dgrad; /* compute rotational grad. */
308 copycolor(acol, aval);
309 rdepth++;
310 ok = doambient(acol, r, r->rweight,
311 uvd, NULL, NULL, dgp, NULL);
312 rdepth--;
313 if (!ok)
314 goto dumbamb;
315 if ((ok > 0) & (dgp != NULL)) { /* apply texture */
316 FVECT v1;
317 VCROSS(v1, r->ron, nrm);
318 d = 1.0;
319 for (i = 3; i--; )
320 d += v1[i] * (dgp[0]*uvd[0][i] + dgp[1]*uvd[1][i]);
321 if (d >= 0.05)
322 scalecolor(acol, d);
323 }
324 copycolor(aval, acol);
325
326 /* PMAP: add in caustic */
327 addcolor(aval, caustic);
328 return;
329 }
330
331 if (tracktime) /* sort to minimize thrashing */
332 sortambvals(0);
333 /* interpolate ambient value */
334 setcolor(acol, 0.0, 0.0, 0.0);
335 d = sumambient(acol, r, nrm, rdepth,
336 &atrunk, thescene.cuorg, thescene.cusize);
337
338 if (d > FTINY) {
339 d = 1.0/d;
340 scalecolor(acol, d);
341 multcolor(aval, acol);
342
343 /* PMAP: add in caustic */
344 addcolor(aval, caustic);
345 return;
346 }
347
348 rdepth++; /* need to cache new value */
349 ok = makeambient(acol, r, nrm, rdepth-1);
350 rdepth--;
351
352 if (ok) {
353 multcolor(aval, acol); /* computed new value */
354
355 /* PMAP: add in caustic */
356 addcolor(aval, caustic);
357 return;
358 }
359
360 dumbamb: /* return global value */
361 if ((ambvwt <= 0) | (navsum == 0)) {
362 multcolor(aval, ambval);
363
364 /* PMAP: add in caustic */
365 addcolor(aval, caustic);
366 return;
367 }
368
369 l = bright(ambval); /* average in computations */
370 if (l > FTINY) {
371 d = (log(l)*(double)ambvwt + avsum) /
372 (double)(ambvwt + navsum);
373 d = exp(d) / l;
374 scalecolor(aval, d);
375 multcolor(aval, ambval); /* apply color of ambval */
376 } else {
377 d = exp( avsum / (double)navsum );
378 scalecolor(aval, d); /* neutral color */
379 }
380 }
381
382
383 /* Plug a potential leak where ambient cache value is occluded */
384 static int
385 plugaleak(RAY *r, AMBVAL *ap, FVECT anorm, double ang)
386 {
387 const double cost70sq = 0.1169778; /* cos(70deg)^2 */
388 RAY rtst;
389 FVECT vdif;
390 double normdot, ndotd, nadotd;
391 double a, b, c, t[2];
392
393 ang += 2.*PI*(ang < 0); /* check direction flags */
394 if ( !(ap->corral>>(int)(ang*(16./PI)) & 1) )
395 return(0);
396 /*
397 * Generate test ray, targeting 20 degrees above sample point plane
398 * along surface normal from cache position. This should be high
399 * enough to miss local geometry we don't really care about.
400 */
401 VSUB(vdif, ap->pos, r->rop);
402 normdot = DOT(anorm, r->ron);
403 ndotd = DOT(vdif, r->ron);
404 nadotd = DOT(vdif, anorm);
405 a = normdot*normdot - cost70sq;
406 b = 2.0*(normdot*ndotd - nadotd*cost70sq);
407 c = ndotd*ndotd - DOT(vdif,vdif)*cost70sq;
408 if (quadratic(t, a, b, c) != 2)
409 return(1); /* should rarely happen */
410 if (t[1] <= FTINY)
411 return(0); /* should fail behind test */
412 rayorigin(&rtst, SHADOW, r, NULL);
413 VSUM(rtst.rdir, vdif, anorm, t[1]); /* further dist. > plane */
414 rtst.rmax = normalize(rtst.rdir); /* short ray test */
415 while (localhit(&rtst, &thescene)) { /* check for occluder */
416 OBJREC *m = findmaterial(rtst.ro);
417 if (m != NULL && !istransp(m->otype) && !isBSDFproxy(m) &&
418 (rtst.clipset == NULL ||
419 !inset(rtst.clipset, rtst.ro->omod)))
420 return(1); /* plug light leak */
421 VCOPY(rtst.rorg, rtst.rop); /* skip invisible surface */
422 rtst.rmax -= rtst.rot;
423 rayclear(&rtst);
424 }
425 return(0); /* seems we're OK */
426 }
427
428
429 static double
430 sumambient( /* get interpolated ambient value */
431 COLOR acol,
432 RAY *r,
433 FVECT rn,
434 int al,
435 AMBTREE *at,
436 FVECT c0,
437 double s
438 )
439 { /* initial limit is 10 degrees plus ambacc radians */
440 const double minangle = 10.0 * PI/180.;
441 double maxangle = minangle + ambacc;
442 double wsum = 0.0;
443 FVECT ck0;
444 int i, j;
445 AMBVAL *av;
446
447 if (at->kid != NULL) { /* sum children first */
448 s *= 0.5;
449 for (i = 0; i < 8; i++) {
450 for (j = 0; j < 3; j++) {
451 ck0[j] = c0[j];
452 if (1<<j & i)
453 ck0[j] += s;
454 if (r->rop[j] < ck0[j] - OCTSCALE*s)
455 break;
456 if (r->rop[j] > ck0[j] + (1.0+OCTSCALE)*s)
457 break;
458 }
459 if (j == 3)
460 wsum += sumambient(acol, r, rn, al,
461 at->kid+i, ck0, s);
462 }
463 /* good enough? */
464 if (wsum >= 0.05 && s > minarad*10.0)
465 return(wsum);
466 }
467 /* adjust maximum angle */
468 if (at->alist != NULL && (at->alist->lvl <= al) & (r->rweight < 0.6))
469 maxangle = (maxangle - PI/2.)*pow(r->rweight,0.13) + PI/2.;
470 /* sum this node */
471 for (av = at->alist; av != NULL; av = av->next) {
472 double u, v, d, delta_r2, delta_t2;
473 COLOR ct;
474 FVECT uvw[3];
475 /* record access */
476 if (tracktime)
477 av->latick = ambclock;
478 /*
479 * Ambient level test
480 */
481 if (av->lvl > al || /* list sorted, so this works */
482 (av->lvl == al) & (av->weight < 0.9*r->rweight))
483 break;
484 /*
485 * Direction test using unperturbed normal
486 */
487 decodedir(uvw[2], av->ndir);
488 d = DOT(uvw[2], r->ron);
489 if (d <= 0.0) /* >= 90 degrees */
490 continue;
491 delta_r2 = 2.0 - 2.0*d; /* approx. radians^2 */
492 if (delta_r2 >= maxangle*maxangle)
493 continue;
494 /*
495 * Modified ray behind test
496 */
497 VSUB(ck0, r->rop, av->pos);
498 d = DOT(ck0, uvw[2]);
499 if (d < -minarad*ambacc-.001)
500 continue;
501 d /= av->rad[0];
502 delta_t2 = d*d;
503 if (delta_t2 >= ambacc*ambacc)
504 continue;
505 /*
506 * Elliptical radii test based on Hessian
507 */
508 decodedir(uvw[0], av->udir);
509 VCROSS(uvw[1], uvw[2], uvw[0]);
510 d = (u = DOT(ck0, uvw[0])) / av->rad[0];
511 delta_t2 += d*d;
512 d = (v = DOT(ck0, uvw[1])) / av->rad[1];
513 delta_t2 += d*d;
514 if (delta_t2 >= ambacc*ambacc)
515 continue;
516 /*
517 * Test for potential light leak
518 */
519 if (av->corral && plugaleak(r, av, uvw[2], atan2a(v,u)))
520 continue;
521 /*
522 * Extrapolate value and compute final weight (hat function)
523 */
524 if (!extambient(ct, av, r->rop, rn, uvw))
525 continue;
526 d = tfunc(maxangle, sqrt(delta_r2), 0.0) *
527 tfunc(ambacc, sqrt(delta_t2), 0.0);
528 scalecolor(ct, d);
529 addcolor(acol, ct);
530 wsum += d;
531 }
532 return(wsum);
533 }
534
535
536 static int
537 makeambient( /* make a new ambient value for storage */
538 COLOR acol,
539 RAY *r,
540 FVECT rn,
541 int al
542 )
543 {
544 AMBVAL amb;
545 FVECT uvw[3];
546 int i;
547
548 amb.weight = 1.0; /* compute weight */
549 for (i = al; i-- > 0; )
550 amb.weight *= AVGREFL;
551 if (r->rweight < 0.1*amb.weight) /* heuristic override */
552 amb.weight = 1.25*r->rweight;
553 setcolor(acol, AVGREFL, AVGREFL, AVGREFL);
554 /* compute ambient */
555 i = doambient(acol, r, amb.weight,
556 uvw, amb.rad, amb.gpos, amb.gdir, &amb.corral);
557 scalecolor(acol, 1./AVGREFL); /* undo assumed reflectance */
558 if (i <= 0 || amb.rad[0] <= FTINY) /* no Hessian or zero radius */
559 return(i);
560 /* store value */
561 VCOPY(amb.pos, r->rop);
562 amb.ndir = encodedir(r->ron);
563 amb.udir = encodedir(uvw[0]);
564 amb.lvl = al;
565 copycolor(amb.val, acol);
566 /* insert into tree */
567 avsave(&amb); /* and save to file */
568 if (rn != r->ron) { /* texture */
569 VCOPY(uvw[2], r->ron);
570 extambient(acol, &amb, r->rop, rn, uvw);
571 }
572 return(1);
573 }
574
575
576 static int
577 extambient( /* extrapolate value at pv, nv */
578 COLOR cr,
579 AMBVAL *ap,
580 FVECT pv,
581 FVECT nv,
582 FVECT uvw[3]
583 )
584 {
585 const double min_d = 0.05;
586 const double max_d = 20.;
587 static FVECT my_uvw[3];
588 FVECT v1;
589 int i;
590 double d = 1.0; /* zeroeth order */
591
592 if (uvw == NULL) { /* need local coordinates? */
593 decodedir(my_uvw[2], ap->ndir);
594 decodedir(my_uvw[0], ap->udir);
595 VCROSS(my_uvw[1], my_uvw[2], my_uvw[0]);
596 uvw = my_uvw;
597 }
598 for (i = 3; i--; ) /* gradient due to translation */
599 d += (pv[i] - ap->pos[i]) *
600 (ap->gpos[0]*uvw[0][i] + ap->gpos[1]*uvw[1][i]);
601
602 VCROSS(v1, uvw[2], nv); /* gradient due to rotation */
603 for (i = 3; i--; )
604 d += v1[i] * (ap->gdir[0]*uvw[0][i] + ap->gdir[1]*uvw[1][i]);
605
606 if (d < min_d) /* clamp min/max scaling */
607 d = min_d;
608 else if (d > max_d)
609 d = max_d;
610 copycolor(cr, ap->val);
611 scalecolor(cr, d);
612 return(d > min_d);
613 }
614
615
616 static void
617 avinsert( /* insert ambient value in our tree */
618 AMBVAL *av
619 )
620 {
621 AMBTREE *at;
622 AMBVAL *ap;
623 AMBVAL avh;
624 FVECT ck0;
625 double s;
626 int branch;
627 int i;
628
629 if (av->rad[0] <= FTINY)
630 error(CONSISTENCY, "zero ambient radius in avinsert");
631 at = &atrunk;
632 VCOPY(ck0, thescene.cuorg);
633 s = thescene.cusize;
634 while (s*(OCTSCALE/2) > av->rad[1]*ambacc) {
635 if (at->kid == NULL)
636 if ((at->kid = newambtree()) == NULL)
637 error(SYSTEM, "out of memory in avinsert");
638 s *= 0.5;
639 branch = 0;
640 for (i = 0; i < 3; i++)
641 if (av->pos[i] > ck0[i] + s) {
642 ck0[i] += s;
643 branch |= 1 << i;
644 }
645 at = at->kid + branch;
646 }
647 avh.next = at->alist; /* order by increasing level */
648 for (ap = &avh; ap->next != NULL; ap = ap->next)
649 if ( ap->next->lvl > av->lvl ||
650 (ap->next->lvl == av->lvl) &
651 (ap->next->weight <= av->weight) )
652 break;
653 av->next = ap->next;
654 ap->next = (AMBVAL*)av;
655 at->alist = avh.next;
656 }
657
658
659 static void
660 initambfile( /* initialize ambient file */
661 int cre8
662 )
663 {
664 extern char *progname, *octname;
665 static char *mybuf = NULL;
666
667 #ifdef F_SETLKW
668 aflock(cre8 ? F_WRLCK : F_RDLCK);
669 #endif
670 SET_FILE_BINARY(ambfp);
671 if (mybuf == NULL)
672 mybuf = (char *)bmalloc(BUFSIZ+8);
673 setbuf(ambfp, mybuf);
674 if (cre8) { /* new file */
675 newheader("RADIANCE", ambfp);
676 fprintf(ambfp, "%s -av %g %g %g -aw %d -ab %d -aa %g ",
677 progname, colval(ambval,RED),
678 colval(ambval,GRN), colval(ambval,BLU),
679 ambvwt, ambounce, ambacc);
680 fprintf(ambfp, "-ad %d -as %d -ar %d ",
681 ambdiv, ambssamp, ambres);
682 if (octname != NULL)
683 fputs(octname, ambfp);
684 fputc('\n', ambfp);
685 fprintf(ambfp, "SOFTWARE= %s\n", VersionID);
686 fputnow(ambfp);
687 fputformat(AMBFMT, ambfp);
688 fputc('\n', ambfp);
689 putambmagic(ambfp);
690 } else if (checkheader(ambfp, AMBFMT, NULL) < 0 || !hasambmagic(ambfp))
691 error(USER, "bad ambient file");
692 }
693
694
695 static void
696 avsave( /* insert and save an ambient value */
697 AMBVAL *av
698 )
699 {
700 avstore(av);
701 if (ambfp == NULL)
702 return;
703 if (writambval(av, ambfp) < 0)
704 goto writerr;
705 if (++nunflshed >= AMBFLUSH)
706 if (ambsync() == EOF)
707 goto writerr;
708 return;
709 writerr:
710 error(SYSTEM, "error writing to ambient file");
711 }
712
713
714 static AMBVAL *
715 avstore( /* allocate memory and save aval */
716 AMBVAL *aval
717 )
718 {
719 AMBVAL *av;
720 double d;
721
722 if ((av = newambval()) == NULL)
723 error(SYSTEM, "out of memory in avstore");
724 *av = *aval;
725 av->latick = ambclock;
726 av->next = NULL;
727 nambvals++;
728 d = bright(av->val);
729 if (d > FTINY) { /* add to log sum for averaging */
730 avsum += log(d);
731 navsum++;
732 }
733 avinsert(av); /* insert in our cache tree */
734 return(av);
735 }
736
737
738 #define ATALLOCSZ 512 /* #/8 trees to allocate at once */
739
740 static AMBTREE *atfreelist = NULL; /* free ambient tree structures */
741
742
743 static AMBTREE *
744 newambtree(void) /* allocate 8 ambient tree structs */
745 {
746 AMBTREE *atp, *upperlim;
747
748 if (atfreelist == NULL) { /* get more nodes */
749 atfreelist = (AMBTREE *)malloc(ATALLOCSZ*8*sizeof(AMBTREE));
750 if (atfreelist == NULL)
751 return(NULL);
752 /* link new free list */
753 upperlim = atfreelist + 8*(ATALLOCSZ-1);
754 for (atp = atfreelist; atp < upperlim; atp += 8)
755 atp->kid = atp + 8;
756 atp->kid = NULL;
757 }
758 atp = atfreelist;
759 atfreelist = atp->kid;
760 memset(atp, 0, 8*sizeof(AMBTREE));
761 return(atp);
762 }
763
764
765 static void
766 freeambtree( /* free 8 ambient tree structs */
767 AMBTREE *atp
768 )
769 {
770 atp->kid = atfreelist;
771 atfreelist = atp;
772 }
773
774
775 static void
776 unloadatree( /* unload an ambient value tree */
777 AMBTREE *at,
778 unloadtf_t *f
779 )
780 {
781 AMBVAL *av;
782 int i;
783 /* transfer values at this node */
784 for (av = at->alist; av != NULL; av = at->alist) {
785 at->alist = av->next;
786 av->next = NULL;
787 (*f)(av);
788 }
789 if (at->kid == NULL)
790 return;
791 for (i = 0; i < 8; i++) /* transfer and free children */
792 unloadatree(at->kid+i, f);
793 freeambtree(at->kid);
794 at->kid = NULL;
795 }
796
797
798 static struct avl {
799 AMBVAL *p;
800 unsigned long t;
801 } *avlist1; /* ambient value list with ticks */
802 static AMBVAL **avlist2; /* memory positions for sorting */
803 static int i_avlist; /* index for lists */
804
805 static int alatcmp(const void *av1, const void *av2);
806
807 static void
808 avfree(AMBVAL *av)
809 {
810 free(av);
811 }
812
813 static void
814 av2list(
815 AMBVAL *av
816 )
817 {
818 #ifdef DEBUG
819 if (i_avlist >= nambvals)
820 error(CONSISTENCY, "too many ambient values in av2list1");
821 #endif
822 avlist1[i_avlist].p = avlist2[i_avlist] = (AMBVAL*)av;
823 avlist1[i_avlist++].t = av->latick;
824 }
825
826
827 static int
828 alatcmp( /* compare ambient values for MRA */
829 const void *av1,
830 const void *av2
831 )
832 {
833 long lc = ((struct avl *)av2)->t - ((struct avl *)av1)->t;
834 return(lc<0 ? -1 : lc>0 ? 1 : 0);
835 }
836
837
838 /* GW NOTE 2002/10/3:
839 * I used to compare AMBVAL pointers, but found that this was the
840 * cause of a serious consistency error with gcc, since the optimizer
841 * uses some dangerous trick in pointer subtraction that
842 * assumes pointers differ by exact struct size increments.
843 */
844 static int
845 aposcmp( /* compare ambient value positions */
846 const void *avp1,
847 const void *avp2
848 )
849 {
850 long diff = *(char * const *)avp1 - *(char * const *)avp2;
851 if (diff < 0)
852 return(-1);
853 return(diff > 0);
854 }
855
856
857 static int
858 avlmemi( /* find list position from address */
859 AMBVAL *avaddr
860 )
861 {
862 AMBVAL **avlpp;
863
864 avlpp = (AMBVAL **)bsearch(&avaddr, avlist2,
865 nambvals, sizeof(AMBVAL *), aposcmp);
866 if (avlpp == NULL)
867 error(CONSISTENCY, "address not found in avlmemi");
868 return(avlpp - avlist2);
869 }
870
871
872 static void
873 sortambvals( /* resort ambient values */
874 int always
875 )
876 {
877 AMBTREE oldatrunk;
878 AMBVAL tav, *tap, *pnext;
879 int i, j;
880 /* see if it's time yet */
881 if (!always && (ambclock++ < lastsort+sortintvl ||
882 nambvals < SORT_THRESH))
883 return;
884 /*
885 * The idea here is to minimize memory thrashing
886 * in VM systems by improving reference locality.
887 * We do this by periodically sorting our stored ambient
888 * values in memory in order of most recently to least
889 * recently accessed. This ordering was chosen so that new
890 * ambient values (which tend to be less important) go into
891 * higher memory with the infrequently accessed values.
892 * Since we expect our values to need sorting less
893 * frequently as the process continues, we double our
894 * waiting interval after each call.
895 * This routine is also called by setambacc() with
896 * the "always" parameter set to 1 so that the ambient
897 * tree will be rebuilt with the new accuracy parameter.
898 */
899 if (tracktime) { /* allocate pointer arrays to sort */
900 avlist2 = (AMBVAL **)malloc(nambvals*sizeof(AMBVAL *));
901 avlist1 = (struct avl *)malloc(nambvals*sizeof(struct avl));
902 } else {
903 avlist2 = NULL;
904 avlist1 = NULL;
905 }
906 if (avlist1 == NULL) { /* no time tracking -- rebuild tree? */
907 if (avlist2 != NULL)
908 free(avlist2);
909 if (always) { /* rebuild without sorting */
910 oldatrunk = atrunk;
911 atrunk.alist = NULL;
912 atrunk.kid = NULL;
913 unloadatree(&oldatrunk, avinsert);
914 }
915 } else { /* sort memory by last access time */
916 /*
917 * Sorting memory is tricky because it isn't contiguous.
918 * We have to sort an array of pointers by MRA and also
919 * by memory position. We then copy values in "loops"
920 * to minimize memory hits. Nevertheless, we will visit
921 * everyone at least twice, and this is an expensive process
922 * when we're thrashing, which is when we need to do it.
923 */
924 #ifdef DEBUG
925 sprintf(errmsg, "sorting %u ambient values at ambclock=%lu...",
926 nambvals, ambclock);
927 eputs(errmsg);
928 #endif
929 i_avlist = 0;
930 unloadatree(&atrunk, av2list); /* empty current tree */
931 #ifdef DEBUG
932 if (i_avlist < nambvals)
933 error(CONSISTENCY, "missing ambient values in sortambvals");
934 #endif
935 qsort(avlist1, nambvals, sizeof(struct avl), alatcmp);
936 qsort(avlist2, nambvals, sizeof(AMBVAL *), aposcmp);
937 for (i = 0; i < nambvals; i++) {
938 if (avlist1[i].p == NULL)
939 continue;
940 tap = avlist2[i];
941 tav = *tap;
942 for (j = i; (pnext = avlist1[j].p) != tap;
943 j = avlmemi(pnext)) {
944 *(avlist2[j]) = *pnext;
945 avinsert(avlist2[j]);
946 avlist1[j].p = NULL;
947 }
948 *(avlist2[j]) = tav;
949 avinsert(avlist2[j]);
950 avlist1[j].p = NULL;
951 }
952 free(avlist1);
953 free(avlist2);
954 /* compute new sort interval */
955 sortintvl = ambclock - lastsort;
956 if (sortintvl >= MAX_SORT_INTVL/2)
957 sortintvl = MAX_SORT_INTVL;
958 else
959 sortintvl <<= 1; /* wait twice as long next */
960 #ifdef DEBUG
961 eputs("done\n");
962 #endif
963 }
964 if (ambclock >= MAXACLOCK)
965 ambclock = MAXACLOCK/2;
966 lastsort = ambclock;
967 }
968
969
970 #ifdef F_SETLKW
971
972 static void
973 aflock( /* lock/unlock ambient file */
974 int typ
975 )
976 {
977 static struct flock fls; /* static so initialized to zeroes */
978
979 if (typ == fls.l_type) /* already called? */
980 return;
981
982 fls.l_type = typ;
983 do
984 if (fcntl(fileno(ambfp), F_SETLKW, &fls) != -1)
985 return;
986 while (errno == EINTR);
987
988 error(SYSTEM, "cannot (un)lock ambient file");
989 }
990
991
992 int
993 ambsync(void) /* synchronize ambient file */
994 {
995 long flen;
996 AMBVAL avs;
997 int n;
998
999 if (ambfp == NULL) /* no ambient file? */
1000 return(0);
1001 /* gain appropriate access */
1002 aflock(nunflshed ? F_WRLCK : F_RDLCK);
1003 /* see if file has grown */
1004 if ((flen = lseek(fileno(ambfp), (off_t)0, SEEK_END)) < 0)
1005 goto seekerr;
1006 if ((n = flen - lastpos) > 0) { /* file has grown */
1007 if (ambinp == NULL) { /* get new file pointer */
1008 ambinp = fopen(ambfile, "rb");
1009 if (ambinp == NULL)
1010 error(SYSTEM, "fopen failed in ambsync");
1011 }
1012 if (fseek(ambinp, lastpos, SEEK_SET) < 0)
1013 goto seekerr;
1014 while (n >= AMBVALSIZ) { /* load contributed values */
1015 if (!readambval(&avs, ambinp)) {
1016 sprintf(errmsg,
1017 "ambient file \"%s\" corrupted near character %ld",
1018 ambfile, flen - n);
1019 error(WARNING, errmsg);
1020 break;
1021 }
1022 avstore(&avs);
1023 n -= AMBVALSIZ;
1024 }
1025 lastpos = flen - n; /* check alignment */
1026 if (n && lseek(fileno(ambfp), (off_t)lastpos, SEEK_SET) < 0)
1027 goto seekerr;
1028 }
1029 n = fflush(ambfp); /* calls write() at last */
1030 lastpos += (long)nunflshed*AMBVALSIZ;
1031 aflock(F_UNLCK); /* release file */
1032 nunflshed = 0;
1033 return(n);
1034 seekerr:
1035 error(SYSTEM, "seek failed in ambsync");
1036 return(EOF); /* pro forma return */
1037 }
1038
1039 #else /* ! F_SETLKW */
1040
1041 int
1042 ambsync(void) /* flush ambient file */
1043 {
1044 if (ambfp == NULL)
1045 return(0);
1046 nunflshed = 0;
1047 return(fflush(ambfp));
1048 }
1049
1050 #endif /* ! F_SETLKW */