3 |
|
#endif |
4 |
|
/* |
5 |
|
* General component matrix combiner, operating on a row at a time. |
6 |
+ |
* |
7 |
+ |
* Multi-processing mode under Unix creates children that each work |
8 |
+ |
* on one input row at a time, fed by the single parent. Output to |
9 |
+ |
* a shared stdout is assured because each child waits for parent |
10 |
+ |
* to be ready with next input row before it outputs the previous result. |
11 |
+ |
* The assumption (checked in spawn_children()) is that the parent |
12 |
+ |
* will remain blocked until the child it is feeding has finished |
13 |
+ |
* flushing the previous row to stdout. The final output row is |
14 |
+ |
* less guaranteed to be in order, so the parent sleeps for a few seconds |
15 |
+ |
* between each child pipe closure, in hopes that this allows enough |
16 |
+ |
* time for row output to finish in each process. |
17 |
|
*/ |
18 |
|
|
19 |
|
#include <math.h> |
554 |
|
static int |
555 |
|
spawned_children(int np) |
556 |
|
{ |
557 |
< |
long inpwidth = 0; |
557 |
> |
size_t recsize = 0; |
558 |
|
int i, rv; |
559 |
|
|
560 |
|
#if defined(_WIN32) || defined(_WIN64) |
563 |
|
np = 1; |
564 |
|
} else |
565 |
|
#endif |
566 |
< |
if ((in_nrows > 0) & (np > in_nrows)) |
567 |
< |
np = in_nrows; |
566 |
> |
if ((in_nrows > 0) & (np*4 > in_nrows)) |
567 |
> |
np = in_nrows/4; |
568 |
|
/* we'll be doing a row at a time */ |
569 |
|
for (i = 0; i < nmats; i++) { |
570 |
|
mop[i].imx.nrows = 1; |
571 |
|
if (!rmx_prepare(&mop[i].imx)) |
572 |
|
goto memerror; |
573 |
< |
inpwidth += rmx_array_size(&mop[i].imx); |
573 |
> |
recsize += rmx_array_size(&mop[i].imx); |
574 |
|
if (mop[i].rmp != &mop[i].imx) { |
575 |
|
mop[i].rmp->nrows = 1; |
576 |
|
if (!rmx_prepare(mop[i].rmp)) |
577 |
|
goto memerror; |
567 |
– |
inpwidth += rmx_array_size(mop[i].rmp); |
578 |
|
} |
579 |
|
} |
580 |
|
/* prep output row buffer */ |
591 |
|
mop[nmats].imx.nrows = 1; |
592 |
|
if (!rmx_prepare(&mop[nmats].imx)) |
593 |
|
goto memerror; |
594 |
< |
if (np <= 1) { /* single process return point */ |
594 |
> |
if (np <= 1) { /* single process return */ |
595 |
|
#ifdef getc_unlocked |
596 |
|
for (i = 0; i < nmats; i++) |
597 |
|
flockfile(mop[i].infp); |
611 |
|
cproc[i].pid = -1; |
612 |
|
rv = open_process(&cproc[i], NULL); |
613 |
|
if (rv <= 0) break; |
614 |
< |
if (!i && 2*rv >= inpwidth) { |
614 |
> |
if (!i && 2*rv >= recsize) { |
615 |
|
fputs("Problem too small for multi-processing\n", |
616 |
|
stderr); |
617 |
+ |
close_processes(cproc, 1); |
618 |
|
exit(1); |
619 |
|
} |
620 |
|
} |
610 |
– |
if (rv > 0) |
611 |
– |
return(1); /* parent return value */ |
621 |
|
if (rv < 0) { |
622 |
|
perror("fork"); |
623 |
+ |
close_processes(cproc, i); |
624 |
|
exit(1); |
625 |
|
} |
626 |
+ |
if (rv > 0) /* parent return? */ |
627 |
+ |
return(1); |
628 |
|
inchild = i; /* our child index */ |
629 |
|
while (i-- > 0) /* don't share siblings' pipes */ |
630 |
|
close(cproc[i].w); |
637 |
|
fclose(mop[i].infp); /* ! pclose() */ |
638 |
|
mop[i].infp = stdin; |
639 |
|
mop[i].imx.dtype = DTdouble; |
640 |
+ |
mop[i].imx.pflags &= ~RMF_SWAPIN; |
641 |
|
} |
642 |
|
return(0); /* child return */ |
643 |
|
memerror: |
682 |
|
if (fflush(ofp) == EOF) |
683 |
|
return(0); |
684 |
|
} |
685 |
< |
for (i = 0; i < nchildren; i++) { |
686 |
< |
sleep(2); /* try to maintain order */ |
685 |
> |
for (i = 0; i < nchildren; i++) { /* maintain output order */ |
686 |
> |
sleep(1+(mop[nmats].rmp->ncols*mop[nmats].rmp->ncomp >> 15)); |
687 |
|
fclose(outfp[i]); |
688 |
|
} |
689 |
|
free(outfp); |