ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/radiance/ray/src/rt/raypcalls.c
Revision: 2.33
Committed: Mon Jun 15 15:44:03 2020 UTC (3 years, 11 months ago) by greg
Content type: text/plain
Branch: MAIN
Changes since 2.32: +7 -2 lines
Log Message:
Optimize rtrace for throughput, especially with ambient cache

File Contents

# Content
1 #ifndef lint
2 static const char RCSid[] = "$Id: raypcalls.c,v 2.32 2020/04/30 19:30:48 greg Exp $";
3 #endif
4 /*
5 * raypcalls.c - interface for parallel rendering using Radiance
6 *
7 * External symbols declared in ray.h
8 */
9
10 #include "copyright.h"
11
12 /*
13 * These calls are designed similarly to the ones in raycalls.c,
14 * but allow for multiple rendering processes on the same host
15 * machine. There is no sense in specifying more child processes
16 * than you have processor cores, but one child may help by allowing
17 * asynchronous ray computation in an interactive program, and
18 * will protect the caller from fatal rendering errors.
19 *
20 * You should first read and understand the header in raycalls.c,
21 * as some things are explained there that are not repated here.
22 *
23 * The first step is opening one or more rendering processes
24 * with a call to ray_pinit(oct, nproc). Before calling fork(),
25 * ray_pinit() loads the octree and data structures into the
26 * caller's memory, and ray_popen() synchronizes the ambient
27 * file, if any. Shared memory permits all sorts of queries
28 * that wouldn't be possible otherwise without causing any real
29 * memory overhead, since all the static data are shared
30 * between processes. Rays are traced using a simple
31 * queuing mechanism, explained below.
32 *
33 * The ray queue buffers RAYQLEN rays before sending to
34 * children, each of which may internally buffer RAYQLEN rays
35 * during evaluation. Rays are not returned in the order
36 * they are sent when multiple processes are open.
37 *
38 * Rays are queued and returned by a single
39 * ray_pqueue() call. A ray_pqueue() return
40 * value of 0 indicates that no rays are ready
41 * and the queue is not yet full. A return value of 1
42 * indicates that a ray was returned, though it is probably
43 * not the one you just requested. Rays may be identified by
44 * the rno member of the RAY struct, which is incremented
45 * by the rayorigin() call, or may be set explicitly by
46 * the caller. Below is an example call sequence:
47 *
48 * myRay.rorg = ( ray origin point )
49 * myRay.rdir = ( normalized ray direction )
50 * myRay.rmax = ( maximum length, or zero for no limit )
51 * rayorigin(&myRay, PRIMARY, NULL, NULL);
52 * myRay.rno = ( my personal ray identifier )
53 * if (ray_pqueue(&myRay) == 1)
54 * { do something with results }
55 *
56 * Note the differences between this and the simpler ray_trace()
57 * call. In particular, the call may or may not return a value
58 * in the passed ray structure. Also, you need to call rayorigin()
59 * yourself, which is normally called for you by ray_trace(). The
60 * benefit is that ray_pqueue() will trace rays faster in
61 * proportion to the number of CPUs you have available on your
62 * system. If the ray queue is full before the call, ray_pqueue()
63 * will block until a result is ready so it can queue this one.
64 * The global int ray_pnidle indicates the number of currently idle
65 * children. If you want to check for completed rays without blocking,
66 * or get the results from rays that have been queued without
67 * queuing any new ones, the ray_presult() call is for you:
68 *
69 * if (ray_presult(&myRay, 1) == 1)
70 * { do something with results }
71 *
72 * If the second argument is 1, the call won't block when
73 * results aren't ready, but will immediately return 0.
74 * If the second argument is 0, the call will block
75 * until a value is available, returning 0 only if the
76 * queue is completely empty. Setting the second argument
77 * to -1 returns 0 unless a ray is ready in the queue and
78 * no system calls are needed. A negative return value
79 * indicates that a rendering process died. If this
80 * happens, ray_pclose(0) is automatically called to close
81 * all child processes, and ray_pnprocs is set to zero.
82 *
83 * If you just want to fill the ray queue without checking for
84 * results, check ray_pnidle and call ray_psend():
85 *
86 * while (ray_pnidle) {
87 * ( set up ray )
88 * ray_psend(&myRay);
89 * }
90 *
91 * Note that it is a mistake to call ra_psend() when
92 * ray_pnidle is zero, and nothing will be sent in
93 * this case. Otherwise, the ray_presult() and/or ray_pqueue()
94 * functions may be called subsequently to read back the results
95 * of rays queued by ray_psend().
96 *
97 * When you are done, you may call ray_pdone(1) to close
98 * all child processes and clean up memory used by Radiance.
99 * Any queued ray calculations will be awaited and discarded.
100 * As with ray_done(), ray_pdone(0) hangs onto data files
101 * and fonts that are likely to be used in subsequent renderings.
102 * Whether you need to clean up memory or not, you should
103 * at least call ray_pclose(0) to await the child processes.
104 * The caller should define a quit() function that calls
105 * ray_pclose(0) if ray_pnprocs > 0.
106 *
107 * Warning: You cannot affect any of the rendering processes
108 * by changing global parameter values onece ray_pinit() has
109 * been called. Changing global parameters will have no effect
110 * until the next call to ray_pinit(), which restarts everything.
111 * If you just want to reap children so that you can alter the
112 * rendering parameters without reloading the scene, use the
113 * ray_pclose(0) and ray_popen(nproc) calls to close
114 * then restart the child processes after the changes are made.
115 *
116 * Note: These routines are written to coordinate with the
117 * definitions in raycalls.c, and in fact depend on them.
118 * If you want to trace a ray and get a result synchronously,
119 * use the ray_trace() call to compute it in the parent process.
120 * This will not interfere with any subprocess calculations,
121 * but beware that a fatal error may end with a call to quit().
122 *
123 * Note: One of the advantages of using separate processes
124 * is that it gives the calling program some immunity from
125 * fatal rendering errors. As discussed in raycalls.c,
126 * Radiance tends to throw up its hands and exit at the
127 * first sign of trouble, calling quit() to return control
128 * to the top level. Although you can avoid exit() with
129 * your own longjmp() in quit(), the cleanup afterwards
130 * is always suspect. Through the use of subprocesses,
131 * we avoid this pitfall by closing the processes and
132 * returning a negative value from ray_pqueue() or
133 * ray_presult(). If you get a negative value from either
134 * of these calls, you can assume that the processes have
135 * been cleaned up with a call to ray_pclose(), though you
136 * will have to call ray_pdone() yourself if you want to
137 * free memory. Obviously, you cannot continue rendering
138 * without risking further errors, but otherwise your
139 * process should not be compromised.
140 */
141
142 #include "rtprocess.h"
143 #include "ray.h"
144 #include "ambient.h"
145 #include <sys/types.h>
146 #include <sys/wait.h>
147 #include "selcall.h"
148
149 #ifndef RAYQLEN
150 #define RAYQLEN 24 /* # rays to send at once */
151 #endif
152
153 #ifndef MAX_RPROCS
154 #if (FD_SETSIZE/2-4 < 64)
155 #define MAX_NPROCS (FD_SETSIZE/2-4)
156 #else
157 #define MAX_NPROCS 64 /* max. # rendering processes */
158 #endif
159 #endif
160
161 extern char *shm_boundary; /* boundary of shared memory */
162
163 int ray_pnprocs = 0; /* number of child processes */
164 int ray_pnidle = 0; /* number of idle children */
165 int ray_pnbatch = 0; /* throughput over responsiveness? */
166
167 static struct child_proc {
168 RT_PID pid; /* child process id */
169 int fd_send; /* write to child here */
170 int fd_recv; /* read from child here */
171 int npending; /* # rays in process */
172 RNUMBER rno[RAYQLEN]; /* working on these rays */
173 } r_proc[MAX_NPROCS]; /* our child processes */
174
175 static RAY r_queue[2*RAYQLEN]; /* ray i/o buffer */
176 static int r_send_next = 0; /* next send ray placement */
177 static int r_recv_first = RAYQLEN; /* position of first unreported ray */
178 static int r_recv_next = RAYQLEN; /* next received ray placement */
179
180 static int samplestep = 1; /* sample step size */
181
182 #define sendq_full() (r_send_next >= RAYQLEN)
183
184 static int ray_pflush(void);
185 static void ray_pchild(int fd_in, int fd_out);
186
187
188 void
189 ray_pinit( /* initialize ray-tracing processes */
190 char *otnm,
191 int nproc
192 )
193 {
194 if (nobjects > 0) /* close old calculation */
195 ray_pdone(0);
196
197 ray_init(otnm); /* load the shared scene */
198
199 ray_popen(nproc); /* fork children */
200 }
201
202
203 static int
204 ray_pflush(void) /* send queued rays to idle children */
205 {
206 int nc, n, nw, i, sfirst;
207
208 if ((ray_pnidle <= 0) | (r_send_next <= 0))
209 return(0); /* nothing we can send */
210
211 sfirst = 0; /* divvy up labor */
212 nc = ray_pnidle;
213 for (i = ray_pnprocs; nc && i--; ) {
214 if (r_proc[i].npending > 0)
215 continue; /* child looks busy */
216 n = r_send_next - sfirst;
217 if (ray_pnbatch)
218 nc--; /* maximize bundling for batch calc */
219 else
220 n /= nc--; /* distribute work for interactivity */
221 if (!n)
222 continue;
223 /* smuggle set size in crtype */
224 r_queue[sfirst].crtype = n;
225 nw = writebuf(r_proc[i].fd_send, (char *)&r_queue[sfirst],
226 sizeof(RAY)*n);
227 if (nw != sizeof(RAY)*n)
228 return(-1); /* write error */
229 r_proc[i].npending = n;
230 while (n--) /* record ray IDs */
231 r_proc[i].rno[n] = r_queue[sfirst+n].rno;
232 sfirst += r_proc[i].npending;
233 ray_pnidle--; /* now she's busy */
234 }
235 if (sfirst != r_send_next)
236 error(CONSISTENCY, "code screwup in ray_pflush()");
237 r_send_next = 0;
238 return(sfirst); /* return total # sent */
239 }
240
241
242 int
243 ray_psend( /* add a ray to our send queue */
244 RAY *r
245 )
246 {
247 int rv;
248
249 if ((r == NULL) | (ray_pnidle <= 0))
250 return(0);
251 /* flush output if necessary */
252 if (sendq_full() && (rv = ray_pflush()) <= 0)
253 return(rv);
254
255 r_queue[r_send_next++] = *r;
256 return(1);
257 }
258
259
260 int
261 ray_pqueue( /* queue a ray for computation */
262 RAY *r
263 )
264 {
265 if (r == NULL)
266 return(0);
267 /* check for full send queue */
268 if (sendq_full()) {
269 RAY mySend = *r;
270 /* wait for a result */
271 if (ray_presult(r, 0) <= 0)
272 return(-1);
273 /* put new ray in queue */
274 r_queue[r_send_next++] = mySend;
275
276 return(1);
277 }
278 /* else add ray to send queue */
279 r_queue[r_send_next++] = *r;
280 /* check for returned ray... */
281 if (r_recv_first >= r_recv_next)
282 return(0);
283 /* ...one is sitting in queue */
284 *r = r_queue[r_recv_first++];
285 return(1);
286 }
287
288
289 int
290 ray_presult( /* check for a completed ray */
291 RAY *r,
292 int poll
293 )
294 {
295 static struct timeval tpoll; /* zero timeval struct */
296 static fd_set readset, errset;
297 int n, ok;
298 int pn;
299
300 if (r == NULL)
301 return(0);
302 /* check queued results first */
303 if (r_recv_first < r_recv_next) {
304 *r = r_queue[r_recv_first++];
305 return(1);
306 }
307 if (poll < 0) /* immediate polling mode? */
308 return(0);
309
310 n = ray_pnprocs - ray_pnidle; /* pending before flush? */
311
312 if (ray_pflush() < 0) /* send new rays to process */
313 return(-1);
314 /* reset receive queue */
315 r_recv_first = r_recv_next = RAYQLEN;
316
317 if (!poll) /* count newly sent unless polling */
318 n = ray_pnprocs - ray_pnidle;
319 if (n <= 0) /* return if nothing to await */
320 return(0);
321 if (!poll && ray_pnprocs == 1) /* one process -> skip select() */
322 FD_SET(r_proc[0].fd_recv, &readset);
323
324 getready: /* any children waiting for us? */
325 for (pn = ray_pnprocs; pn--; )
326 if (FD_ISSET(r_proc[pn].fd_recv, &readset) ||
327 FD_ISSET(r_proc[pn].fd_recv, &errset))
328 break;
329 /* call select() if we must */
330 if (pn < 0) {
331 FD_ZERO(&readset); FD_ZERO(&errset); n = 0;
332 for (pn = ray_pnprocs; pn--; ) {
333 if (r_proc[pn].npending > 0)
334 FD_SET(r_proc[pn].fd_recv, &readset);
335 FD_SET(r_proc[pn].fd_recv, &errset);
336 if (r_proc[pn].fd_recv >= n)
337 n = r_proc[pn].fd_recv + 1;
338 }
339 /* find out who is ready */
340 while ((n = select(n, &readset, (fd_set *)NULL, &errset,
341 poll ? &tpoll : (struct timeval *)NULL)) < 0)
342 if (errno != EINTR) {
343 error(WARNING,
344 "select call failed in ray_presult()");
345 ray_pclose(0);
346 return(-1);
347 }
348 if (n > 0) /* go back and get it */
349 goto getready;
350 return(0); /* else poll came up empty */
351 }
352 if (r_recv_next + r_proc[pn].npending > sizeof(r_queue)/sizeof(RAY))
353 error(CONSISTENCY, "buffer shortage in ray_presult()");
354
355 /* read rendered ray data */
356 n = readbuf(r_proc[pn].fd_recv, (char *)&r_queue[r_recv_next],
357 sizeof(RAY)*r_proc[pn].npending);
358 if (n > 0) {
359 r_recv_next += n/sizeof(RAY);
360 ok = (n == sizeof(RAY)*r_proc[pn].npending);
361 } else
362 ok = 0;
363 /* reset child's status */
364 FD_CLR(r_proc[pn].fd_recv, &readset);
365 if (n <= 0)
366 FD_CLR(r_proc[pn].fd_recv, &errset);
367 r_proc[pn].npending = 0;
368 ray_pnidle++;
369 /* check for rendering errors */
370 if (!ok) {
371 ray_pclose(0); /* process died -- clean up */
372 return(-1);
373 }
374 /* preen returned rays */
375 for (n = r_recv_next - r_recv_first; n--; ) {
376 RAY *rp = &r_queue[r_recv_first + n];
377 rp->rno = r_proc[pn].rno[n];
378 rp->parent = NULL;
379 rp->newcset = rp->clipset = NULL;
380 rp->rox = NULL;
381 rp->slights = NULL;
382 }
383 /* return first ray received */
384 *r = r_queue[r_recv_first++];
385 return(1);
386 }
387
388
389 void
390 ray_pdone( /* reap children and free data */
391 int freall
392 )
393 {
394 ray_pclose(0); /* close child processes */
395
396 if (shm_boundary != NULL) { /* clear shared memory boundary */
397 free((void *)shm_boundary);
398 shm_boundary = NULL;
399 }
400
401 ray_done(freall); /* free rendering data */
402 }
403
404
405 static void
406 ray_pchild( /* process rays (never returns) */
407 int fd_in,
408 int fd_out
409 )
410 {
411 int n;
412 int i;
413 /* flag child process for quit() */
414 ray_pnprocs = -1;
415 /* read each ray request set */
416 while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) {
417 int n2;
418 if (n < sizeof(RAY))
419 break;
420 /* get smuggled set length */
421 n2 = sizeof(RAY)*r_queue[0].crtype - n;
422 if (n2 < 0)
423 error(INTERNAL, "buffer over-read in ray_pchild()");
424 if (n2 > 0) { /* read the rest of the set */
425 i = readbuf(fd_in, (char *)r_queue + n, n2);
426 if (i != n2)
427 break;
428 n += n2;
429 }
430 n /= sizeof(RAY);
431 /* evaluate rays */
432 for (i = 0; i < n; i++) {
433 r_queue[i].crtype = r_queue[i].rtype;
434 r_queue[i].parent = NULL;
435 r_queue[i].clipset = NULL;
436 r_queue[i].slights = NULL;
437 r_queue[i].rlvl = 0;
438 samplendx += samplestep;
439 rayclear(&r_queue[i]);
440 rayvalue(&r_queue[i]);
441 }
442 /* write back our results */
443 i = writebuf(fd_out, (char *)r_queue, sizeof(RAY)*n);
444 if (i != sizeof(RAY)*n)
445 error(SYSTEM, "write error in ray_pchild()");
446 }
447 if (n)
448 error(SYSTEM, "read error in ray_pchild()");
449 ambsync();
450 quit(0); /* normal exit */
451 }
452
453
454 void
455 ray_popen( /* open the specified # processes */
456 int nadd
457 )
458 {
459 /* check if our table has room */
460 if (ray_pnprocs + nadd > MAX_NPROCS)
461 nadd = MAX_NPROCS - ray_pnprocs;
462 if (nadd <= 0)
463 return;
464 ambsync(); /* load any new ambient values */
465 if (shm_boundary == NULL) { /* first child process? */
466 preload_objs(); /* preload auxiliary data */
467 /* set shared memory boundary */
468 shm_boundary = (char *)malloc(16);
469 strcpy(shm_boundary, "SHM_BOUNDARY");
470 }
471 fflush(NULL); /* clear pending output */
472 samplestep = ray_pnprocs + nadd;
473 while (nadd--) { /* fork each new process */
474 int p0[2], p1[2];
475 if (pipe(p0) < 0 || pipe(p1) < 0)
476 error(SYSTEM, "cannot create pipe");
477 if ((r_proc[ray_pnprocs].pid = fork()) == 0) {
478 int pn; /* close others' descriptors */
479 for (pn = ray_pnprocs; pn--; ) {
480 close(r_proc[pn].fd_send);
481 close(r_proc[pn].fd_recv);
482 }
483 close(p0[0]); close(p1[1]);
484 close(0); /* don't share stdin */
485 /* following call never returns */
486 ray_pchild(p1[0], p0[1]);
487 }
488 if (r_proc[ray_pnprocs].pid < 0)
489 error(SYSTEM, "cannot fork child process");
490 close(p1[0]); close(p0[1]);
491 if (rand_samp) /* decorrelate random sequence */
492 srandom(random());
493 else
494 samplendx++;
495 /*
496 * Close write stream on exec to avoid multiprocessing deadlock.
497 * No use in read stream without it, so set flag there as well.
498 */
499 fcntl(p1[1], F_SETFD, FD_CLOEXEC);
500 fcntl(p0[0], F_SETFD, FD_CLOEXEC);
501 r_proc[ray_pnprocs].fd_send = p1[1];
502 r_proc[ray_pnprocs].fd_recv = p0[0];
503 r_proc[ray_pnprocs].npending = 0;
504 ray_pnprocs++;
505 ray_pnidle++;
506 }
507 }
508
509
510 void
511 ray_pclose( /* close one or more child processes */
512 int nsub
513 )
514 {
515 static int inclose = 0;
516 RAY res;
517 int i, status = 0;
518 /* check recursion */
519 if (inclose)
520 return;
521 inclose++;
522 /* check no child / in child */
523 if (ray_pnprocs <= 0)
524 return;
525 /* check argument */
526 if ((nsub <= 0) | (nsub > ray_pnprocs))
527 nsub = ray_pnprocs;
528 /* clear our ray queue */
529 while (ray_presult(&res,0) > 0)
530 ;
531 r_send_next = 0; /* hard reset in case of error */
532 r_recv_first = r_recv_next = RAYQLEN;
533 /* close send pipes */
534 for (i = ray_pnprocs-nsub; i < ray_pnprocs; i++)
535 close(r_proc[i].fd_send);
536
537 if (nsub == 1) { /* awaiting single process? */
538 if (waitpid(r_proc[ray_pnprocs-1].pid, &status, 0) < 0)
539 status = 127<<8;
540 close(r_proc[ray_pnprocs-1].fd_recv);
541 } else /* else unordered wait */
542 for (i = 0; i < nsub; ) {
543 int j, mystatus;
544 RT_PID pid = wait(&mystatus);
545 for (j = ray_pnprocs-nsub; j < ray_pnprocs; j++)
546 if (r_proc[j].pid == pid) {
547 if (mystatus)
548 status = mystatus;
549 close(r_proc[j].fd_recv);
550 ++i;
551 }
552 }
553 ray_pnprocs -= nsub;
554 ray_pnidle -= nsub;
555 if (status) {
556 sprintf(errmsg, "rendering process exited with code %d", status>>8);
557 error(WARNING, errmsg);
558 }
559 inclose--;
560 }