Mail Archives: djgpp/1997/09/28/21:12:12
On Sun, 28 Sep 1997, Vic wrote:
> hello. Do you remember Thomas Djafari's test? It was a DCT compressor to
> see the performance of different compilers. anyways, today I recompiled
> the thingie and it was 2 times slower! After playing around with the
> flags for a while, I realised it must be something else. The only thing
> I changed were the binutils. So I do a test:
> with -O6 -ffast-math -mpentium -fomit-frame-pointer -funroll-loops
> I get:
> With BNU 2.7:
> approx. 2939 million clocks compressing
> approx. 2561 million clocks decompressing
> With BNU 2.8.1 I get
> approx. 3326 million clocks compressing
> approx. 2965 million clocks decompressing
> Why is this slower when using BNU2.8.1?? what are the new things
> installed in it?
That's strange. I just ran a more-or-less "standard" benchmark, using
gcc-2.7.2.1 and PGCC/egcs-970910, both using BNU-2.7 (I did not upgrade to
2.8.1 as the PGCC page suggests). I ran two sets of tests: the v2.0 DP
FLOPS benchmark from ftp.nosc.mil, and LINPACK-C. Here's what I got
(results on the left are for egcs, on the right for gcc):
--
Pentium-100 MHz
gcc -O6 -mpentium -ffast-math -fomit-frame-pointer -fno-strength-reduce
i386-linux-gcc -O2 -m486 -ffast-math -fomit-frame-pointer -fno-strength-reduce
FLOPS C Program (Double Precision), V2.0 18 Dec 1992
(1) 15.1022/11.9219 MFLOPS
(2) 14.4005/12.0392 MFLOPS
(3) 22.8970/17.5147 MFLOPS
(4) 33.5495/23.6551 MFLOPS
LINPACK-C
Double-precision, rolled 12.60/10.01 MFLOPS
Double-precision, unrolled 13.04/12.14 MFLOPS
Single-precision, rolled 14.59/11.33 MFLOPS
Single-precision, unrolled 15.31/13.68 MFLOPS
-------------------------------------------------------------------
Orlando Alcantara Andico
WWW: http://www2.mozcom.com/~orly/ Email: orly AT mozcom DOT com
ICBM: 14 30 00 N 120 59 00 E POTS: (+632) 932-2385
- Raw text -