Date: Wed, 26 Aug 1998 03:06:44 -0400 From: Harlan Stenn <Harlan.Stenn@pfcs.com> To: joelh@gnu.org Cc: garbanzo@hooked.net, mike@smith.net.au, entropy@compufit.at, wwoods@cybcon.com, freebsd-current@FreeBSD.ORG Subject: Re: gcc 2.8 Message-ID: <25448.904115204@brown.pfcs.com> In-Reply-To: Joel Ray Holveck's (joelh@gnu.org) message dated Tue, 25 Aug 1998 17:54:26. <199808252254.RAA01679@detlev.UUCP>
next in thread | previous in thread | raw e-mail | index | archive | help
> > I tested two packages.
> > One compares a number (6-8) byte move subroutines (memcpy, bcopy, a variety
> > of Duff's devices (using char, short, and int), and some other "fast" byte
> > copies I've snarfed over the years).
>
> Just to be sure: these are using their own recompiled bcopy etc, and
> not running the ones out of libc?
This benchmark uses whatever libc byte movers are detected (bcopy or
memcpy), and 5 or 6 different "source" versions.
> > I run a reasonable quantity of different size/alignments against each of
> > these, and report the CPU time of each one.
>
> Could you please post this data?
It's kinda boring, but if everybody is interested I'll post it. If only a
couple of folks are interested, I'll send it directly.
Basically, I added the following code to the byte-mover subroutine:
#ifdef BMOVESTATS
#define BMS_MAXSIZE 2049
int bms[BMS_MAXSIZE + 1][4][4];
#endif /* BMOVESTATS */
(I picked 2049 because this application hardly ever moves bigger chunks of
memory.)
... and inside the byte mover subroutine:
#ifdef BMOVESTATS
if((n = len) > BMS_MAXSIZE) {
n = BMS_MAXSIZE;}
++bms[n][((int)src) & 3][((int)dst) & 3];
#endif /* BMOVESTATS */
... and I call this subroutine before the executable exits:
void bmovestats(void)
{
#ifdef BMOVESTATS
int i;
int d;
int s;
FILE *fp;
#define BMOVE_STATFILE "bmovestats"
if((fp = fopen(BMOVE_STATFILE, "w")) == NULL) {
fp = stderr;}
for(i = 0; i <= BMS_MAXSIZE; ++i)
for (s = 0; s < 4; s++)
for (d = 0; d < 4; d++)
if(bms[i][s][d])
fprintf(fp, "%d\t%d\t%d\t%d\n", i, s, d, bms[i][s][d]);
if(fp != stderr) {
fclose(fp);}
#endif /* BMOVESTATS */
return;}
Then I ran several applications under various production conditions,
collected various "bmovestats" files, and then ran a little program that,
for each byte mover, basically ran the following loop (in this case, for
bcopy):
#ifdef HAVE_BCOPY
rewind(ifp);
(void)times(&t_b);
t_u = t_b.tms_utime;
while((rc = fscanf(ifp, "%d%d%d%d", &len, &sa, &da, &qty)) != EOF) {
sp = src + sa;
dp = dst + da;
for(i = 0; i < qty; i++) {
bcopy(sp, dp, len);}}
(void)times(&t_b);
printf("%8.0d", t_b.tms_utime - t_u - ovhd);
fflush(stdout);
#endif /* HAVE_BCOPY */
the "ovhd" value is the time to run the loop with *no* byte copy being
performed; I suspect a Really Good Compiler would pretty much invalidate
this benchmark. "ifp" points to the current "bmovestats" file.
H
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?25448.904115204>