* Re: Strange results on code speed
@ 1997-12-10 5:31 Laurent Bonnaud
0 siblings, 0 replies; 3+ messages in thread
From: Laurent Bonnaud @ 1997-12-10 5:31 UTC (permalink / raw)
To: egcs; +Cc: bonnaud
> You might want to retry your tests with -O4 or so. At -O2 function
>inlining is not done, I believe, and that might well make a big
>difference.
According to the doc, -O3 is the highest optimization level available
and -O3 does inlining indeed. In fact i used the -O2 option on
purpose, as the code is written with inline directives where they are
needed. Nevertheless, i tried your suggestion, but the speed
difference is not significant.
--
Laurent.
^ permalink raw reply [flat|nested] 3+ messages in thread
* Strange results on code speed
1997-12-09 16:25 Laurent Bonnaud
@ 1997-12-10 0:16 ` Mark Mitchell
0 siblings, 0 replies; 3+ messages in thread
From: Mark Mitchell @ 1997-12-10 0:16 UTC (permalink / raw)
To: egcs
Lauren --
You might want to retry your tests with -O4 or so. At -O2 function
inlining is not done, I believe, and that might well make a big
difference.
--
Mark Mitchell mmitchell@usa.net
Stanford University http://www.stanford.edu
^ permalink raw reply [flat|nested] 3+ messages in thread
* Strange results on code speed
@ 1997-12-09 16:25 Laurent Bonnaud
1997-12-10 0:16 ` Mark Mitchell
0 siblings, 1 reply; 3+ messages in thread
From: Laurent Bonnaud @ 1997-12-09 16:25 UTC (permalink / raw)
To: egcs
Hi,
since i read about a new optimization in the egcs-971207 snapshot, i
tried two C++ benchmarks :
ftp://ftp.kai.com/pub/benchmarks/stepanov_v1p2.C
and
ftp://ftp.kai.com/pub/benchmarks/oopack_v1p7.C
which measure the relative speed of the same computations coded in C
and in various C++ styles (ideally it should be 1.0).
I got strange results. To sum up, the new snapshot performs incredibly
better than egcs-1.0 on the Stepanov benchmark, but only on
sparc-solaris (and not on i586-linux which is even worse than g++ 2.7).
On the other benchmark, results vary wildly with the architecture, but
the overall result is a bit disappointing and is better for the i86.
The good news is that this new snapshot fixes all bugs i reported.
Thanks everybody for your work !
* stepanov_v1p2.C :
- on an Ultrasparc 170MHz
egcs-971207 -O2 (the very good result)
test absolute additions ratio with
number time per second test0
0 0.92sec 54.35M 1.00
1 0.92sec 54.35M 1.00
2 0.92sec 54.35M 1.00
3 1.24sec 40.32M 1.35
4 1.22sec 40.98M 1.33
5 1.23sec 40.65M 1.34
6 1.23sec 40.65M 1.34
7 1.23sec 40.65M 1.34
8 1.23sec 40.65M 1.34
9 1.23sec 40.65M 1.34
10 1.23sec 40.65M 1.34
11 1.23sec 40.65M 1.34
12 1.22sec 40.98M 1.33
mean: 1.15sec 43.50M 1.25
egcs-1.0 -O2 (still better than g++-2.7.2.1 except for test1)
0 0.92sec 54.35M 1.00
1 3.65sec 13.70M 3.97
2 4.26sec 11.74M 4.63
3 3.96sec 12.63M 4.30
4 5.48sec 9.12M 5.96
5 3.96sec 12.63M 4.30
6 5.49sec 9.11M 5.97
7 5.18sec 9.65M 5.63
8 5.48sec 9.12M 5.96
9 5.17sec 9.67M 5.62
10 6.06sec 8.25M 6.59
11 3.76sec 13.30M 4.09
12 6.40sec 7.81M 6.96
mean: 4.24sec 11.78M 4.61
- on a 586 166MHz
egcs-971207 -O2 (not as good as on sparc)
0 3.63sec 13.77M 1.00
1 6.67sec 7.50M 1.84
2 5.72sec 8.74M 1.58
3 8.01sec 6.24M 2.21
4 6.09sec 8.21M 1.68
5 8.20sec 6.10M 2.26
6 6.10sec 8.20M 1.68
7 9.76sec 5.12M 2.69
8 9.52sec 5.25M 2.62
9 9.72sec 5.14M 2.68
10 9.51sec 5.26M 2.62
11 11.61sec 4.31M 3.20
12 9.72sec 5.14M 2.68
mean: 7.69sec 6.50M 2.12
egcs-1.0 -O2 (slightly worse except for tests 8, 10 and 12)
0 3.63sec 13.77M 1.00
1 6.67sec 7.50M 1.84
2 7.43sec 6.73M 2.05
3 9.85sec 5.08M 2.71
4 8.38sec 5.97M 2.31
5 9.71sec 5.15M 2.67
6 8.38sec 5.97M 2.31
7 10.08sec 4.96M 2.78
8 8.83sec 5.66M 2.43
9 9.83sec 5.09M 2.71
10 8.76sec 5.71M 2.41
11 12.84sec 3.89M 3.54
12 9.14sec 5.47M 2.52
mean: 8.43sec 5.93M 2.32
g++-2.7.2.3 -O2 (the C speed is much better, but the C++ optimization makes up for it !)
0 2.29sec 21.83M 1.00
1 3.81sec 13.12M 1.66
2 8.76sec 5.71M 3.83
3 10.08sec 4.96M 4.40
4 14.27sec 3.50M 6.23
5 14.11sec 3.54M 6.16
6 18.27sec 2.74M 7.98
7 14.83sec 3.37M 6.48
8 19.78sec 2.53M 8.64
9 14.84sec 3.37M 6.48
10 19.79sec 2.53M 8.64
11 14.83sec 3.37M 6.48
12 18.88sec 2.65M 8.24
mean: 11.59sec 4.31M 5.06
* oopack_v1p7.C
- on an Ultrasparc 170MHz
egcs-971207 -O2
Seconds Mflops
Test Iterations C OOP C OOP Ratio
---- ---------- ----------- ----------- -----
Max 5000 0.3 0.4 16.7 12.2 1.4
Matrix 50 0.5 2.5 27.2 5.1 5.3
Complex 2000 0.3 1.5 45.7 10.7 4.3
Iterator 5000 0.2 0.6 47.6 17.2 2.8
egcs-971207 -O2 -mcpu=v8
Max 5000 0.2 0.4 20.8 12.5 1.7
Matrix 50 0.5 0.5 26.6 26.0 1.0
Complex 2000 0.4 1.1 42.1 14.4 2.9
Iterator 5000 0.2 0.7 41.7 15.2 2.8
The -mcpu=v8 option improves dramatically the Matrix test, but only for
the OOP style (same thing for g++ 2.7) ! For the Complex test, the
switch has opposite effects for both styles !
- on a 586 166MHz the result is better :
egcs-971207 -O2 -mcpu=pentium
Max 5000 0.6 0.6 9.1 8.3 1.1
Matrix 50 0.7 0.9 17.4 13.6 1.3
Complex 2000 0.8 1.4 20.5 11.1 1.8
Iterator 5000 0.5 0.6 18.5 16.4 1.1
So what do you think ? If egcs could always perform as well on i86 as
on the sparc and vice-versa, depending on the code, it would really .
--
Laurent.
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~1997-12-10 5:31 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
1997-12-10 5:31 Strange results on code speed Laurent Bonnaud
-- strict thread matches above, loose matches on Subject: below --
1997-12-09 16:25 Laurent Bonnaud
1997-12-10 0:16 ` Mark Mitchell
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).