1 |
/* Copyright (c) 1998 Silicon Graphics, Inc. */ |
2 |
|
3 |
#ifndef lint |
4 |
static char SCCSid[] = "$SunId$ SGI"; |
5 |
#endif |
6 |
|
7 |
/* |
8 |
* Routines for tracking beam compuatations |
9 |
*/ |
10 |
|
11 |
#include "rholo.h" |
12 |
#include <sys/types.h> |
13 |
|
14 |
#ifndef NFRAG2CHUNK |
15 |
#define NFRAG2CHUNK 4096 /* number of fragments to start chunking */ |
16 |
#endif |
17 |
|
18 |
#ifndef abs |
19 |
#define abs(x) ((x) > 0 ? (x) : -(x)) |
20 |
#endif |
21 |
#ifndef sgn |
22 |
#define sgn(x) ((x) > 0 ? 1 : (x) < 0 ? -1 : 0) |
23 |
#endif |
24 |
|
25 |
#define rchunk(n) (((n)+(RPACKSIZ/2))/RPACKSIZ) |
26 |
|
27 |
extern time_t time(); |
28 |
|
29 |
static PACKHEAD *complist=NULL; /* list of beams to compute */ |
30 |
static int complen=0; /* length of complist */ |
31 |
static int listpos=0; /* current list position for next_packet */ |
32 |
static int lastin= -1; /* last ordered position in list */ |
33 |
static int chunky=0; /* clump beams together on disk */ |
34 |
|
35 |
|
36 |
int |
37 |
beamcmp(b0, b1) /* comparison for compute order */ |
38 |
register PACKHEAD *b0, *b1; |
39 |
{ |
40 |
BEAMI *bip0, *bip1; |
41 |
register long c; |
42 |
/* first check desired quantities */ |
43 |
if (chunky) |
44 |
c = rchunk(b1->nr)*(rchunk(b0->nc)+1L) - |
45 |
rchunk(b0->nr)*(rchunk(b1->nc)+1L); |
46 |
else |
47 |
c = b1->nr*(b0->nc+1L) - b0->nr*(b1->nc+1L); |
48 |
if (c > 0) return(1); |
49 |
if (c < 0) return(-1); |
50 |
/* only one file, so skip the following: */ |
51 |
#if 0 |
52 |
/* next, check file descriptors */ |
53 |
c = hdlist[b0->hd]->fd - hdlist[b1->hd]->fd; |
54 |
if (c) return(c); |
55 |
#endif |
56 |
/* finally, check file positions */ |
57 |
bip0 = &hdlist[b0->hd]->bi[b0->bi]; |
58 |
bip1 = &hdlist[b1->hd]->bi[b1->bi]; |
59 |
/* put diskless beams last */ |
60 |
if (!bip0->nrd) |
61 |
return(bip1->nrd > 0); |
62 |
if (!bip1->nrd) |
63 |
return(-1); |
64 |
c = bip0->fo - bip1->fo; |
65 |
return(c < 0 ? -1 : c > 0); |
66 |
} |
67 |
|
68 |
|
69 |
int |
70 |
beamidcmp(b0, b1) /* comparison for beam searching */ |
71 |
register PACKHEAD *b0, *b1; |
72 |
{ |
73 |
register int c = b0->hd - b1->hd; |
74 |
|
75 |
if (c) return(c); |
76 |
return(b0->bi - b1->bi); |
77 |
} |
78 |
|
79 |
|
80 |
int |
81 |
dispbeam(b, hb) /* display a holodeck beam */ |
82 |
register BEAM *b; |
83 |
register HDBEAMI *hb; |
84 |
{ |
85 |
static int n = 0; |
86 |
static PACKHEAD *p = NULL; |
87 |
|
88 |
if (b == NULL) |
89 |
return; |
90 |
if (b->nrm > n) { /* (re)allocate packet holder */ |
91 |
n = b->nrm; |
92 |
if (p == NULL) p = (PACKHEAD *)malloc(packsiz(n)); |
93 |
else p = (PACKHEAD *)realloc((char *)p, packsiz(n)); |
94 |
if (p == NULL) |
95 |
error(SYSTEM, "out of memory in dispbeam"); |
96 |
} |
97 |
/* assign packet fields */ |
98 |
bcopy((char *)hdbray(b), (char *)packra(p), b->nrm*sizeof(RAYVAL)); |
99 |
p->nr = p->nc = b->nrm; |
100 |
for (p->hd = 0; hdlist[p->hd] != hb->h; p->hd++) |
101 |
if (hdlist[p->hd] == NULL) |
102 |
error(CONSISTENCY, "unregistered holodeck in dispbeam"); |
103 |
p->bi = hb->b; |
104 |
disp_packet(p); /* display it */ |
105 |
} |
106 |
|
107 |
|
108 |
bundle_set(op, clist, nents) /* bundle set operation */ |
109 |
int op; |
110 |
PACKHEAD *clist; |
111 |
int nents; |
112 |
{ |
113 |
int oldnr, n; |
114 |
HDBEAMI *hbarr; |
115 |
register PACKHEAD *csm; |
116 |
register int i; |
117 |
/* search for common members */ |
118 |
for (csm = clist+nents; csm-- > clist; ) |
119 |
csm->nc = -1; |
120 |
qsort((char *)clist, nents, sizeof(PACKHEAD), beamidcmp); |
121 |
for (i = 0; i < complen; i++) { |
122 |
csm = (PACKHEAD *)bsearch((char *)(complist+i), (char *)clist, |
123 |
nents, sizeof(PACKHEAD), beamidcmp); |
124 |
if (csm == NULL) |
125 |
continue; |
126 |
oldnr = complist[i].nr; |
127 |
csm->nc = complist[i].nc; |
128 |
switch (op) { |
129 |
case BS_ADD: /* add to count */ |
130 |
complist[i].nr += csm->nr; |
131 |
csm->nr = 0; |
132 |
break; |
133 |
case BS_ADJ: /* reset count */ |
134 |
complist[i].nr = csm->nr; |
135 |
csm->nr = 0; |
136 |
break; |
137 |
case BS_DEL: /* delete count */ |
138 |
if (csm->nr == 0 || csm->nr >= complist[i].nr) |
139 |
complist[i].nr = 0; |
140 |
else |
141 |
complist[i].nr -= csm->nr; |
142 |
break; |
143 |
} |
144 |
if (complist[i].nr != oldnr) |
145 |
lastin = -1; /* flag sort */ |
146 |
} |
147 |
/* record computed rays for uncommon beams */ |
148 |
for (csm = clist+nents; csm-- > clist; ) |
149 |
if (csm->nc < 0) |
150 |
csm->nc = bnrays(hdlist[csm->hd], csm->bi); |
151 |
/* complete list operations */ |
152 |
switch (op) { |
153 |
case BS_NEW: /* new computation set */ |
154 |
listpos = 0; lastin = -1; |
155 |
if (complen) /* free old list */ |
156 |
free((char *)complist); |
157 |
complist = NULL; |
158 |
if (!(complen = nents)) |
159 |
return; |
160 |
complist = (PACKHEAD *)malloc(nents*sizeof(PACKHEAD)); |
161 |
if (complist == NULL) |
162 |
goto memerr; |
163 |
bcopy((char *)clist, (char *)complist, nents*sizeof(PACKHEAD)); |
164 |
break; |
165 |
case BS_ADD: /* add to computation set */ |
166 |
case BS_ADJ: /* adjust set quantities */ |
167 |
if (nents <= 0) |
168 |
return; |
169 |
sortcomplist(); /* sort updated list & new entries */ |
170 |
qsort((char *)clist, nents, sizeof(PACKHEAD), beamcmp); |
171 |
/* what can't we satisfy? */ |
172 |
for (i = nents, csm = clist; i-- && csm->nr > csm->nc; csm++) |
173 |
; |
174 |
n = csm - clist; |
175 |
if (op == BS_ADJ) { /* don't regenerate adjusted beams */ |
176 |
for (++i; i-- && csm->nr > 0; csm++) |
177 |
; |
178 |
nents = csm - clist; |
179 |
} |
180 |
if (n) { /* allocate space for merged list */ |
181 |
PACKHEAD *newlist; |
182 |
newlist = (PACKHEAD *)malloc( |
183 |
(complen+n)*sizeof(PACKHEAD) ); |
184 |
if (newlist == NULL) |
185 |
goto memerr; |
186 |
/* merge lists */ |
187 |
mergeclists(newlist, clist, n, complist, complen); |
188 |
if (complen) |
189 |
free((char *)complist); |
190 |
complist = newlist; |
191 |
complen += n; |
192 |
} |
193 |
listpos = 0; |
194 |
lastin = complen-1; /* list is now sorted */ |
195 |
break; |
196 |
case BS_DEL: /* delete from computation set */ |
197 |
return; /* already done */ |
198 |
default: |
199 |
error(CONSISTENCY, "bundle_set called with unknown operation"); |
200 |
} |
201 |
if (outdev == NULL || !nents) /* nothing to display? */ |
202 |
return; |
203 |
/* load and display beams we have */ |
204 |
hbarr = (HDBEAMI *)malloc(nents*sizeof(HDBEAMI)); |
205 |
for (i = nents; i--; ) { |
206 |
hbarr[i].h = hdlist[clist[i].hd]; |
207 |
hbarr[i].b = clist[i].bi; |
208 |
} |
209 |
hdloadbeams(hbarr, nents, dispbeam); |
210 |
free((char *)hbarr); |
211 |
if (hdfragflags&FF_READ) { |
212 |
listpos = 0; |
213 |
lastin = -1; /* need to re-sort list */ |
214 |
} |
215 |
return; |
216 |
memerr: |
217 |
error(SYSTEM, "out of memory in bundle_set"); |
218 |
} |
219 |
|
220 |
|
221 |
double |
222 |
beamvolume(hp, bi) /* compute approximate volume of a beam */ |
223 |
HOLO *hp; |
224 |
int bi; |
225 |
{ |
226 |
GCOORD gc[2]; |
227 |
FVECT cp[4], edgeA, edgeB, cent[2]; |
228 |
FVECT v, crossp[2], diffv; |
229 |
double vol[2]; |
230 |
register int i; |
231 |
/* get grid coordinates */ |
232 |
if (!hdbcoord(gc, hp, bi)) |
233 |
error(CONSISTENCY, "bad beam index in beamvolume"); |
234 |
for (i = 0; i < 2; i++) { /* compute cell area vectors */ |
235 |
hdcell(cp, hp, gc+i); |
236 |
VSUM(edgeA, cp[1], cp[0], -1.0); |
237 |
VSUM(edgeB, cp[3], cp[1], -1.0); |
238 |
fcross(crossp[i], edgeA, edgeB); |
239 |
/* compute center */ |
240 |
cent[i][0] = 0.5*(cp[0][0] + cp[2][0]); |
241 |
cent[i][1] = 0.5*(cp[0][1] + cp[2][1]); |
242 |
cent[i][2] = 0.5*(cp[0][2] + cp[2][2]); |
243 |
} |
244 |
/* compute difference vector */ |
245 |
VSUM(diffv, cent[1], cent[0], -1.0); |
246 |
for (i = 0; i < 2; i++) { /* compute volume contributions */ |
247 |
vol[i] = 0.5*DOT(crossp[i], diffv); |
248 |
if (vol[i] < 0.) vol[i] = -vol[i]; |
249 |
} |
250 |
return(vol[0] + vol[1]); /* return total volume */ |
251 |
} |
252 |
|
253 |
|
254 |
init_global() /* initialize global ray computation */ |
255 |
{ |
256 |
long wtotal = 0; |
257 |
double frac; |
258 |
int i; |
259 |
register int j, k; |
260 |
/* free old list and empty queue */ |
261 |
if (complen > 0) { |
262 |
free((char *)complist); |
263 |
done_packets(flush_queue()); |
264 |
} |
265 |
/* reseed random number generator */ |
266 |
srandom(time(NULL)); |
267 |
/* allocate beam list */ |
268 |
complen = 0; |
269 |
for (j = 0; hdlist[j] != NULL; j++) |
270 |
complen += nbeams(hdlist[j]); |
271 |
complist = (PACKHEAD *)malloc(complen*sizeof(PACKHEAD)); |
272 |
if (complist == NULL) |
273 |
error(SYSTEM, "out of memory in init_global"); |
274 |
/* compute beam weights */ |
275 |
k = 0; |
276 |
for (j = 0; hdlist[j] != NULL; j++) { |
277 |
frac = 512. * VLEN(hdlist[j]->wg[0]) * |
278 |
VLEN(hdlist[j]->wg[1]) * |
279 |
VLEN(hdlist[j]->wg[2]); |
280 |
for (i = nbeams(hdlist[j]); i > 0; i--) { |
281 |
complist[k].hd = j; |
282 |
complist[k].bi = i; |
283 |
complist[k].nr = frac*beamvolume(hdlist[j], i) + 0.5; |
284 |
complist[k].nc = bnrays(hdlist[j], i); |
285 |
wtotal += complist[k++].nr; |
286 |
} |
287 |
} |
288 |
/* adjust weights */ |
289 |
if (vdef(DISKSPACE)) |
290 |
frac = 1024.*1024.*vflt(DISKSPACE) / (wtotal*sizeof(RAYVAL)); |
291 |
else |
292 |
frac = 1024.*1024.*16384. / (wtotal*sizeof(RAYVAL)); |
293 |
while (k--) |
294 |
complist[k].nr = frac*complist[k].nr + 0.5; |
295 |
listpos = 0; lastin = -1; /* perform initial sort */ |
296 |
sortcomplist(); |
297 |
/* no view vicinity */ |
298 |
myeye.rng = 0; |
299 |
} |
300 |
|
301 |
|
302 |
mergeclists(cdest, cl1, n1, cl2, n2) /* merge two sorted lists */ |
303 |
register PACKHEAD *cdest; |
304 |
register PACKHEAD *cl1, *cl2; |
305 |
int n1, n2; |
306 |
{ |
307 |
register int cmp; |
308 |
|
309 |
while (n1 | n2) { |
310 |
if (!n1) cmp = 1; |
311 |
else if (!n2) cmp = -1; |
312 |
else cmp = beamcmp(cl1, cl2); |
313 |
if (cmp > 0) { |
314 |
copystruct(cdest, cl2); |
315 |
cl2++; n2--; |
316 |
} else { |
317 |
copystruct(cdest, cl1); |
318 |
cl1++; n1--; |
319 |
} |
320 |
cdest++; |
321 |
} |
322 |
} |
323 |
|
324 |
|
325 |
sortcomplist() /* fix our list order */ |
326 |
{ |
327 |
PACKHEAD *list2; |
328 |
int listlen; |
329 |
register int i; |
330 |
|
331 |
if (complen <= 0) /* check to see if there is even a list */ |
332 |
return; |
333 |
if (!chunky) /* check to see if fragment list is full */ |
334 |
if (!hdfragOK(hdlist[0]->fd, &listlen, NULL) |
335 |
#if NFRAG2CHUNK |
336 |
|| listlen >= NFRAG2CHUNK |
337 |
#endif |
338 |
) { |
339 |
#ifdef DEBUG |
340 |
error(WARNING, "using chunky comparison mode"); |
341 |
#endif |
342 |
chunky++; /* use "chunky" comparison */ |
343 |
lastin = -1; /* need to re-sort list */ |
344 |
} |
345 |
#ifdef DEBUG |
346 |
else |
347 |
fprintf(stderr, "sortcomplist: %d fragments\n", |
348 |
listlen); |
349 |
#endif |
350 |
if (lastin < 0 || listpos*4 >= complen*3) |
351 |
qsort((char *)complist, complen, sizeof(PACKHEAD), beamcmp); |
352 |
else if (listpos) { /* else sort and merge sublist */ |
353 |
list2 = (PACKHEAD *)malloc(listpos*sizeof(PACKHEAD)); |
354 |
if (list2 == NULL) |
355 |
error(SYSTEM, "out of memory in sortcomplist"); |
356 |
bcopy((char *)complist,(char *)list2,listpos*sizeof(PACKHEAD)); |
357 |
qsort((char *)list2, listpos, sizeof(PACKHEAD), beamcmp); |
358 |
mergeclists(complist, list2, listpos, |
359 |
complist+listpos, complen-listpos); |
360 |
free((char *)list2); |
361 |
} |
362 |
/* drop satisfied requests */ |
363 |
for (i = complen; i-- && complist[i].nr <= complist[i].nc; ) |
364 |
; |
365 |
if (i < 0) { |
366 |
free((char *)complist); |
367 |
complist = NULL; |
368 |
complen = 0; |
369 |
} else if (i < complen-1) { |
370 |
list2 = (PACKHEAD *)realloc((char *)complist, |
371 |
(i+1)*sizeof(PACKHEAD)); |
372 |
if (list2 != NULL) |
373 |
complist = list2; |
374 |
complen = i+1; |
375 |
} |
376 |
listpos = 0; lastin = i; |
377 |
} |
378 |
|
379 |
|
380 |
/* |
381 |
* The following routine works on the assumption that the bundle weights are |
382 |
* more or less evenly distributed, such that computing a packet causes |
383 |
* a given bundle to move way down in the computation order. We keep |
384 |
* track of where the computed bundle with the highest priority would end |
385 |
* up, and if we get further in our compute list than this, we re-sort the |
386 |
* list and start again from the beginning. Since |
387 |
* a merge sort is used, the sorting costs are minimal. |
388 |
*/ |
389 |
next_packet(p, n) /* prepare packet for computation */ |
390 |
register PACKET *p; |
391 |
int n; |
392 |
{ |
393 |
register int i; |
394 |
|
395 |
if (listpos > lastin) /* time to sort the list */ |
396 |
sortcomplist(); |
397 |
if (complen <= 0) |
398 |
return(0); |
399 |
p->hd = complist[listpos].hd; |
400 |
p->bi = complist[listpos].bi; |
401 |
p->nc = complist[listpos].nc; |
402 |
p->nr = complist[listpos].nr - p->nc; |
403 |
if (p->nr <= 0) |
404 |
return(0); |
405 |
DCHECK(n < 1 | n > RPACKSIZ, |
406 |
CONSISTENCY, "next_packet called with bad n value"); |
407 |
if (p->nr > n) |
408 |
p->nr = n; |
409 |
complist[listpos].nc += p->nr; /* find where this one would go */ |
410 |
if (hdgetbeam(hdlist[p->hd], p->bi) != NULL) |
411 |
hdfreefrag(hdlist[p->hd], p->bi); |
412 |
while (lastin > listpos && |
413 |
beamcmp(complist+lastin, complist+listpos) > 0) |
414 |
lastin--; |
415 |
listpos++; |
416 |
return(1); |
417 |
} |