| 1 |
#ifndef lint |
| 2 |
static const char RCSid[] = "$Id: raypcalls.c,v 2.3 2003/07/03 15:00:19 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 processors, 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 undrstand 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. This permits all sorts of queries that |
| 27 |
* wouldn't be possible otherwise, without causing any real |
| 28 |
* memory overhead, since all the static data are shared |
| 29 |
* between processes. Rays are then traced using a simple |
| 30 |
* queuing mechanism, explained below. |
| 31 |
* |
| 32 |
* The ray queue holds as many rays as there are rendering |
| 33 |
* processes. Rays are queued and returned by a single |
| 34 |
* ray_pqueue() call. A ray_pqueue() return |
| 35 |
* value of 0 indicates that no rays are ready |
| 36 |
* and the queue is not yet full. A return value of 1 |
| 37 |
* indicates that a ray was returned, though it is probably |
| 38 |
* not the one you just requested. Rays may be identified by |
| 39 |
* the rno member of the RAY struct, which is incremented |
| 40 |
* by the rayorigin() call, or may be set explicitly by |
| 41 |
* the caller. Below is an example call sequence: |
| 42 |
* |
| 43 |
* myRay.rorg = ( ray origin point ) |
| 44 |
* myRay.rdir = ( normalized ray direction ) |
| 45 |
* myRay.rmax = ( maximum length, or zero for no limit ) |
| 46 |
* rayorigin(&myRay, NULL, PRIMARY, 1.0); |
| 47 |
* myRay.rno = ( my personal ray identifier ) |
| 48 |
* if (ray_pqueue(&myRay) == 1) |
| 49 |
* { do something with results } |
| 50 |
* |
| 51 |
* Note the differences between this and the simpler ray_trace() |
| 52 |
* call. In particular, the call may or may not return a value |
| 53 |
* in the passed ray structure. Also, you need to call rayorigin() |
| 54 |
* yourself, which is normally for you by ray_trace(). The |
| 55 |
* great thing is that ray_pqueue() will trace rays faster in |
| 56 |
* proportion to the number of CPUs you have available on your |
| 57 |
* system. If the ray queue is full before the call, ray_pqueue() |
| 58 |
* will block until a result is ready so it can queue this one. |
| 59 |
* The global int ray_pnidle indicates the number of currently idle |
| 60 |
* children. If you want to check for completed rays without blocking, |
| 61 |
* or get the results from rays that have been queued without |
| 62 |
* queuing any new ones, the ray_presult() call is for you: |
| 63 |
* |
| 64 |
* if (ray_presult(&myRay, 1) == 1) |
| 65 |
* { do something with results } |
| 66 |
* |
| 67 |
* If the second argument is 1, the call won't block when |
| 68 |
* results aren't ready, but will immediately return 0. |
| 69 |
* If the second argument is 0, the call will block |
| 70 |
* until a value is available, returning 0 only if the |
| 71 |
* queue is completely empty. A negative return value |
| 72 |
* indicates that a rendering process died. If this |
| 73 |
* happens, ray_close(0) is automatically called to close |
| 74 |
* all child processes, and ray_pnprocs is set to zero. |
| 75 |
* |
| 76 |
* If you just want to fill the ray queue without checking for |
| 77 |
* results, check ray_pnidle and call ray_psend(): |
| 78 |
* |
| 79 |
* while (ray_pnidle) { |
| 80 |
* ( set up ray ) |
| 81 |
* ray_psend(&myRay); |
| 82 |
* } |
| 83 |
* |
| 84 |
* The ray_presult() and/or ray_pqueue() functions may then be |
| 85 |
* called to read back the results. |
| 86 |
* |
| 87 |
* When you are done, you may call ray_pdone(1) to close |
| 88 |
* all child processes and clean up memory used by Radiance. |
| 89 |
* Any queued ray calculations will be awaited and discarded. |
| 90 |
* As with ray_done(), ray_pdone(0) hangs onto data files |
| 91 |
* and fonts that are likely to be used in subsequent renderings. |
| 92 |
* Whether you want to bother cleaning up memory or not, you |
| 93 |
* should at least call ray_pclose(0) to clean the child processes. |
| 94 |
* |
| 95 |
* Warning: You cannot affect any of the rendering processes |
| 96 |
* by changing global parameter values onece ray_pinit() has |
| 97 |
* been called. Changing global parameters will have no effect |
| 98 |
* until the next call to ray_pinit(), which restarts everything. |
| 99 |
* If you just want to reap children so that you can alter the |
| 100 |
* rendering parameters without reloading the scene, use the |
| 101 |
* ray_pclose(0) and ray_popen(nproc) calls to close |
| 102 |
* then restart the child processes. |
| 103 |
* |
| 104 |
* Note: These routines are written to coordinate with the |
| 105 |
* definitions in raycalls.c, and in fact depend on them. |
| 106 |
* If you want to trace a ray and get a result synchronously, |
| 107 |
* use the ray_trace() call to compute it in the parent process. |
| 108 |
* |
| 109 |
* Note: One of the advantages of using separate processes |
| 110 |
* is that it gives the calling program some immunity from |
| 111 |
* fatal rendering errors. As discussed in raycalls.c, |
| 112 |
* Radiance tends to throw up its hands and exit at the |
| 113 |
* first sign of trouble, calling quit() to return control |
| 114 |
* to the system. Although you can avoid exit() with |
| 115 |
* your own longjmp() in quit(), the cleanup afterwards |
| 116 |
* is always suspect. Through the use of subprocesses, |
| 117 |
* we avoid this pitfall by closing the processes and |
| 118 |
* returning a negative value from ray_pqueue() or |
| 119 |
* ray_presult(). If you get a negative value from either |
| 120 |
* of these calls, you can assume that the processes have |
| 121 |
* been cleaned up with a call to ray_close(), though you |
| 122 |
* will have to call ray_pdone() yourself if you want to |
| 123 |
* free memory. Obviously, you cannot continue rendering, |
| 124 |
* but otherwise your process should not be compromised. |
| 125 |
*/ |
| 126 |
|
| 127 |
#include "ray.h" |
| 128 |
|
| 129 |
#include "selcall.h" |
| 130 |
|
| 131 |
#ifndef RAYQLEN |
| 132 |
#define RAYQLEN 16 /* # rays to send at once */ |
| 133 |
#endif |
| 134 |
|
| 135 |
#ifndef MAX_RPROCS |
| 136 |
#if (FD_SETSIZE/2-4 < 64) |
| 137 |
#define MAX_NPROCS (FD_SETSIZE/2-4) |
| 138 |
#else |
| 139 |
#define MAX_NPROCS 64 /* max. # rendering processes */ |
| 140 |
#endif |
| 141 |
#endif |
| 142 |
|
| 143 |
extern char *shm_boundary; /* boundary of shared memory */ |
| 144 |
|
| 145 |
int ray_pnprocs = 0; /* number of child processes */ |
| 146 |
int ray_pnidle = 0; /* number of idle children */ |
| 147 |
|
| 148 |
static struct child_proc { |
| 149 |
int pid; /* child process id */ |
| 150 |
int fd_send; /* write to child here */ |
| 151 |
int fd_recv; /* read from child here */ |
| 152 |
int npending; /* # rays in process */ |
| 153 |
unsigned long rno[RAYQLEN]; /* working on these rays */ |
| 154 |
} r_proc[MAX_NPROCS]; /* our child processes */ |
| 155 |
|
| 156 |
static RAY r_queue[2*RAYQLEN]; /* ray i/o buffer */ |
| 157 |
static int r_send_next; /* next send ray placement */ |
| 158 |
static int r_recv_first; /* position of first unreported ray */ |
| 159 |
static int r_recv_next; /* next receive ray placement */ |
| 160 |
|
| 161 |
#define sendq_full() (r_send_next >= RAYQLEN) |
| 162 |
|
| 163 |
|
| 164 |
void |
| 165 |
ray_pinit(otnm, nproc) /* initialize ray-tracing processes */ |
| 166 |
char *otnm; |
| 167 |
int nproc; |
| 168 |
{ |
| 169 |
if (nobjects > 0) /* close old calculation */ |
| 170 |
ray_pdone(0); |
| 171 |
|
| 172 |
ray_init(otnm); /* load the shared scene */ |
| 173 |
|
| 174 |
preload_objs(); /* preload auxiliary data */ |
| 175 |
|
| 176 |
/* set shared memory boundary */ |
| 177 |
shm_boundary = (char *)malloc(16); |
| 178 |
strcpy(shm_boundary, "SHM_BOUNDARY"); |
| 179 |
|
| 180 |
r_send_next = 0; /* set up queue */ |
| 181 |
r_recv_first = r_recv_next = RAYQLEN; |
| 182 |
|
| 183 |
ray_popen(nproc); /* fork children */ |
| 184 |
} |
| 185 |
|
| 186 |
|
| 187 |
static int |
| 188 |
ray_pflush() /* send queued rays to idle children */ |
| 189 |
{ |
| 190 |
int nc, n, nw, i, sfirst; |
| 191 |
|
| 192 |
if ((ray_pnidle <= 0 | r_send_next <= 0)) |
| 193 |
return(0); /* nothing we can send */ |
| 194 |
|
| 195 |
sfirst = 0; /* divvy up labor */ |
| 196 |
nc = ray_pnidle; |
| 197 |
for (i = ray_pnprocs; nc && i--; ) { |
| 198 |
if (r_proc[i].npending > 0) |
| 199 |
continue; /* child looks busy */ |
| 200 |
n = (r_send_next - sfirst)/nc--; |
| 201 |
if (!n) |
| 202 |
continue; |
| 203 |
/* smuggle set size in crtype */ |
| 204 |
r_queue[sfirst].crtype = n; |
| 205 |
nw = writebuf(r_proc[i].fd_send, (char *)&r_queue[sfirst], |
| 206 |
sizeof(RAY)*n); |
| 207 |
if (nw != sizeof(RAY)*n) |
| 208 |
return(-1); /* write error */ |
| 209 |
r_proc[i].npending = n; |
| 210 |
while (n--) /* record ray IDs */ |
| 211 |
r_proc[i].rno[n] = r_queue[sfirst+n].rno; |
| 212 |
sfirst += r_proc[i].npending; |
| 213 |
ray_pnidle--; /* now she's busy */ |
| 214 |
} |
| 215 |
if (sfirst != r_send_next) |
| 216 |
error(CONSISTENCY, "code screwup in ray_pflush"); |
| 217 |
r_send_next = 0; |
| 218 |
return(sfirst); /* return total # sent */ |
| 219 |
} |
| 220 |
|
| 221 |
|
| 222 |
void |
| 223 |
ray_psend(r) /* add a ray to our send queue */ |
| 224 |
RAY *r; |
| 225 |
{ |
| 226 |
if (r == NULL) |
| 227 |
return; |
| 228 |
/* flush output if necessary */ |
| 229 |
if (sendq_full() && ray_pflush() <= 0) |
| 230 |
error(INTERNAL, "ray_pflush failed in ray_psend"); |
| 231 |
|
| 232 |
r_queue[r_send_next] = *r; |
| 233 |
r_send_next++; |
| 234 |
} |
| 235 |
|
| 236 |
|
| 237 |
int |
| 238 |
ray_pqueue(r) /* queue a ray for computation */ |
| 239 |
RAY *r; |
| 240 |
{ |
| 241 |
if (r == NULL) |
| 242 |
return(0); |
| 243 |
/* check for full send queue */ |
| 244 |
if (sendq_full()) { |
| 245 |
RAY mySend; |
| 246 |
int rval; |
| 247 |
mySend = *r; |
| 248 |
/* wait for a result */ |
| 249 |
rval = ray_presult(r, 0); |
| 250 |
/* put new ray in queue */ |
| 251 |
r_queue[r_send_next] = mySend; |
| 252 |
r_send_next++; |
| 253 |
return(rval); /* done */ |
| 254 |
} |
| 255 |
/* add ray to send queue */ |
| 256 |
r_queue[r_send_next] = *r; |
| 257 |
r_send_next++; |
| 258 |
/* check for returned ray... */ |
| 259 |
if (r_recv_first >= r_recv_next) |
| 260 |
return(0); |
| 261 |
/* ...one is sitting in queue */ |
| 262 |
*r = r_queue[r_recv_first]; |
| 263 |
r_recv_first++; |
| 264 |
return(1); |
| 265 |
} |
| 266 |
|
| 267 |
|
| 268 |
int |
| 269 |
ray_presult(r, poll) /* check for a completed ray */ |
| 270 |
RAY *r; |
| 271 |
int poll; |
| 272 |
{ |
| 273 |
static struct timeval tpoll; /* zero timeval struct */ |
| 274 |
static fd_set readset, errset; |
| 275 |
int n, ok; |
| 276 |
register int pn; |
| 277 |
|
| 278 |
if (r == NULL) |
| 279 |
return(0); |
| 280 |
/* check queued results first */ |
| 281 |
if (r_recv_first < r_recv_next) { |
| 282 |
*r = r_queue[r_recv_first]; |
| 283 |
r_recv_first++; |
| 284 |
return(1); |
| 285 |
} |
| 286 |
n = ray_pnprocs - ray_pnidle; /* pending before flush? */ |
| 287 |
|
| 288 |
if (ray_pflush() < 0) /* send new rays to process */ |
| 289 |
return(-1); |
| 290 |
/* reset receive queue */ |
| 291 |
r_recv_first = r_recv_next = RAYQLEN; |
| 292 |
|
| 293 |
if (!poll) /* count newly sent unless polling */ |
| 294 |
n = ray_pnprocs - ray_pnidle; |
| 295 |
if (n <= 0) /* return if nothing to await */ |
| 296 |
return(0); |
| 297 |
getready: /* any children waiting for us? */ |
| 298 |
for (pn = ray_pnprocs; pn--; ) |
| 299 |
if (FD_ISSET(r_proc[pn].fd_recv, &readset) || |
| 300 |
FD_ISSET(r_proc[pn].fd_recv, &errset)) |
| 301 |
break; |
| 302 |
/* call select if we must */ |
| 303 |
if (pn < 0) { |
| 304 |
FD_ZERO(&readset); FD_ZERO(&errset); n = 0; |
| 305 |
for (pn = ray_pnprocs; pn--; ) { |
| 306 |
if (r_proc[pn].npending > 0) |
| 307 |
FD_SET(r_proc[pn].fd_recv, &readset); |
| 308 |
FD_SET(r_proc[pn].fd_recv, &errset); |
| 309 |
if (r_proc[pn].fd_recv >= n) |
| 310 |
n = r_proc[pn].fd_recv + 1; |
| 311 |
} |
| 312 |
/* find out who is ready */ |
| 313 |
while ((n = select(n, &readset, (fd_set *)NULL, &errset, |
| 314 |
poll ? &tpoll : (struct timeval *)NULL)) < 0) |
| 315 |
if (errno != EINTR) { |
| 316 |
error(WARNING, |
| 317 |
"select call failed in ray_presult"); |
| 318 |
ray_pclose(0); |
| 319 |
return(-1); |
| 320 |
} |
| 321 |
if (n > 0) /* go back and get it */ |
| 322 |
goto getready; |
| 323 |
return(0); /* else poll came up empty */ |
| 324 |
} |
| 325 |
if (r_recv_next + r_proc[pn].npending > sizeof(r_queue)/sizeof(RAY)) |
| 326 |
error(CONSISTENCY, "buffer shortage in ray_presult()"); |
| 327 |
|
| 328 |
/* read rendered ray data */ |
| 329 |
n = readbuf(r_proc[pn].fd_recv, (char *)&r_queue[r_recv_next], |
| 330 |
sizeof(RAY)*r_proc[pn].npending); |
| 331 |
if (n > 0) { |
| 332 |
r_recv_next += n/sizeof(RAY); |
| 333 |
ok = (n == sizeof(RAY)*r_proc[pn].npending); |
| 334 |
} else |
| 335 |
ok = 0; |
| 336 |
/* reset child's status */ |
| 337 |
FD_CLR(r_proc[pn].fd_recv, &readset); |
| 338 |
if (n <= 0) |
| 339 |
FD_CLR(r_proc[pn].fd_recv, &errset); |
| 340 |
r_proc[pn].npending = 0; |
| 341 |
ray_pnidle++; |
| 342 |
/* check for rendering errors */ |
| 343 |
if (!ok) { |
| 344 |
ray_pclose(0); /* process died -- clean up */ |
| 345 |
return(-1); |
| 346 |
} |
| 347 |
/* preen returned rays */ |
| 348 |
for (n = r_recv_next - r_recv_first; n--; ) { |
| 349 |
register RAY *rp = &r_queue[r_recv_first + n]; |
| 350 |
rp->rno = r_proc[pn].rno[n]; |
| 351 |
rp->parent = NULL; |
| 352 |
rp->newcset = rp->clipset = NULL; |
| 353 |
rp->rox = NULL; |
| 354 |
rp->slights = NULL; |
| 355 |
} |
| 356 |
/* return first ray received */ |
| 357 |
*r = r_queue[r_recv_first]; |
| 358 |
r_recv_first++; |
| 359 |
return(1); |
| 360 |
} |
| 361 |
|
| 362 |
|
| 363 |
void |
| 364 |
ray_pdone(freall) /* reap children and free data */ |
| 365 |
int freall; |
| 366 |
{ |
| 367 |
ray_pclose(0); /* close child processes */ |
| 368 |
|
| 369 |
if (shm_boundary != NULL) { /* clear shared memory boundary */ |
| 370 |
free((void *)shm_boundary); |
| 371 |
shm_boundary = NULL; |
| 372 |
} |
| 373 |
ray_done(freall); /* free rendering data */ |
| 374 |
} |
| 375 |
|
| 376 |
|
| 377 |
static void |
| 378 |
ray_pchild(fd_in, fd_out) /* process rays (never returns) */ |
| 379 |
int fd_in; |
| 380 |
int fd_out; |
| 381 |
{ |
| 382 |
int n; |
| 383 |
register int i; |
| 384 |
/* read each ray request set */ |
| 385 |
while ((n = read(fd_in, (char *)r_queue, sizeof(r_queue))) > 0) { |
| 386 |
int n2; |
| 387 |
if (n % sizeof(RAY)) |
| 388 |
break; |
| 389 |
n /= sizeof(RAY); |
| 390 |
/* get smuggled set length */ |
| 391 |
n2 = r_queue[0].crtype - n; |
| 392 |
if (n2 < 0) |
| 393 |
error(INTERNAL, "buffer over-read in ray_pchild"); |
| 394 |
if (n2 > 0) { /* read the rest of the set */ |
| 395 |
i = readbuf(fd_in, (char *)(r_queue+n), |
| 396 |
sizeof(RAY)*n2); |
| 397 |
if (i != sizeof(RAY)*n2) |
| 398 |
break; |
| 399 |
n += n2; |
| 400 |
} |
| 401 |
/* evaluate rays */ |
| 402 |
for (i = 0; i < n; i++) { |
| 403 |
r_queue[i].crtype = r_queue[i].rtype; |
| 404 |
r_queue[i].parent = NULL; |
| 405 |
r_queue[i].clipset = NULL; |
| 406 |
r_queue[i].slights = NULL; |
| 407 |
r_queue[i].revf = raytrace; |
| 408 |
samplendx++; |
| 409 |
rayclear(&r_queue[i]); |
| 410 |
rayvalue(&r_queue[i]); |
| 411 |
} |
| 412 |
/* write back our results */ |
| 413 |
i = writebuf(fd_out, (char *)r_queue, sizeof(RAY)*n); |
| 414 |
if (i != sizeof(RAY)*n) |
| 415 |
error(SYSTEM, "write error in ray_pchild"); |
| 416 |
} |
| 417 |
if (n) |
| 418 |
error(SYSTEM, "read error in ray_pchild"); |
| 419 |
ambsync(); |
| 420 |
quit(0); /* normal exit */ |
| 421 |
} |
| 422 |
|
| 423 |
|
| 424 |
void |
| 425 |
ray_popen(nadd) /* open the specified # processes */ |
| 426 |
int nadd; |
| 427 |
{ |
| 428 |
/* check if our table has room */ |
| 429 |
if (ray_pnprocs + nadd > MAX_NPROCS) |
| 430 |
nadd = MAX_NPROCS - ray_pnprocs; |
| 431 |
if (nadd <= 0) |
| 432 |
return; |
| 433 |
fflush(stderr); /* clear pending output */ |
| 434 |
fflush(stdout); |
| 435 |
while (nadd--) { /* fork each new process */ |
| 436 |
int p0[2], p1[2]; |
| 437 |
if (pipe(p0) < 0 || pipe(p1) < 0) |
| 438 |
error(SYSTEM, "cannot create pipe"); |
| 439 |
if ((r_proc[ray_pnprocs].pid = fork()) == 0) { |
| 440 |
int pn; /* close others' descriptors */ |
| 441 |
for (pn = ray_pnprocs; pn--; ) { |
| 442 |
close(r_proc[pn].fd_send); |
| 443 |
close(r_proc[pn].fd_recv); |
| 444 |
} |
| 445 |
close(p0[0]); close(p1[1]); |
| 446 |
/* following call never returns */ |
| 447 |
ray_pchild(p1[0], p0[1]); |
| 448 |
} |
| 449 |
if (r_proc[ray_pnprocs].pid < 0) |
| 450 |
error(SYSTEM, "cannot fork child process"); |
| 451 |
close(p1[0]); close(p0[1]); |
| 452 |
r_proc[ray_pnprocs].fd_send = p1[1]; |
| 453 |
r_proc[ray_pnprocs].fd_recv = p0[0]; |
| 454 |
r_proc[ray_pnprocs].npending = 0; |
| 455 |
ray_pnprocs++; |
| 456 |
ray_pnidle++; |
| 457 |
} |
| 458 |
} |
| 459 |
|
| 460 |
|
| 461 |
void |
| 462 |
ray_pclose(nsub) /* close one or more child processes */ |
| 463 |
int nsub; |
| 464 |
{ |
| 465 |
static int inclose = 0; |
| 466 |
RAY res; |
| 467 |
/* check recursion */ |
| 468 |
if (inclose) |
| 469 |
return; |
| 470 |
inclose++; |
| 471 |
/* check argument */ |
| 472 |
if ((nsub <= 0 | nsub > ray_pnprocs)) |
| 473 |
nsub = ray_pnprocs; |
| 474 |
/* clear our ray queue */ |
| 475 |
while (ray_presult(&res,0) > 0) |
| 476 |
; |
| 477 |
/* clean up children */ |
| 478 |
while (nsub--) { |
| 479 |
int status; |
| 480 |
ray_pnprocs--; |
| 481 |
close(r_proc[ray_pnprocs].fd_recv); |
| 482 |
close(r_proc[ray_pnprocs].fd_send); |
| 483 |
while (wait(&status) != r_proc[ray_pnprocs].pid) |
| 484 |
; |
| 485 |
if (status) { |
| 486 |
sprintf(errmsg, |
| 487 |
"rendering process %d exited with code %d", |
| 488 |
r_proc[ray_pnprocs].pid, status>>8); |
| 489 |
error(WARNING, errmsg); |
| 490 |
} |
| 491 |
ray_pnidle--; |
| 492 |
} |
| 493 |
inclose--; |
| 494 |
} |
| 495 |
|
| 496 |
|
| 497 |
void |
| 498 |
quit(ec) /* make sure exit is called */ |
| 499 |
int ec; |
| 500 |
{ |
| 501 |
exit(ec); |
| 502 |
} |
