Re: (re)compiling debian packages with ccc
Bas Zoetekouw <bas@debian.org> writes:
> I'm not sure if the gap has become smaller, but ccc still generated code
> that is a lot faster than gcc-3.2.
>
> For example, take a look at these results from SCIbench
> (http://math.nist.gov/scimark2), generated on an quadruple-proc EV67
> machine (running Tru64 Unix btw, not Linux):
>
> Compaq C compiler, V6.4-014
> CFLAGS = -arch ev67 -fast -O4
> | Composite Score: 195.47
> GNU C compiler, V3.2.1
> CFLAGS = -O3 -fomit-frame-pointer -ffast-math -mcpu=ev67
> | Composite Score: 137.18
Well, you can beef up gcc's score a bit with -funroll-loops and
-fprefetch-array-loops (ccc does these optimizations without needing
to be kicked). On an 800MHz ev68 (with some background load) I get:
gcc version 3.2.2 20030131 (Debian prerelease)
CFLAGS = -O3 -fomit-frame-pointer -ffast-math -mcpu=ev67 -funroll-loops -fprefetch-loop-arrays
| Composite Score: 149.17
| FFT Mflops: 226.05 (N=1024)
| SOR Mflops: 186.61 (100 x 100)
| MonteCarlo: Mflops: 57.64
| Sparse matmult Mflops: 106.21 (N=1000, nz=5000)
| LU Mflops: 169.36 (M=100, N=100)
Compaq C V6.5-001
CFLAGS = -arch ev67 -fast -O4
| Composite Score: 170.43
| FFT Mflops: 247.80 (N=1024)
| SOR Mflops: 226.62 (100 x 100)
| MonteCarlo: Mflops: 53.78
| Sparse matmult Mflops: 92.64 (N=1000, nz=5000)
| LU Mflops: 231.30 (M=100, N=100)
Well, still 14% better, but this is with a benchmark from an area
where ccc really excels. However, this benchmark is way from
representive for most stuff in Debian; it uses lots of floating point,
a tiny code working set, large data working sets, and simple data
access patterns. (BTW, I guess it could be sped up a lot with the
superpage kernel patch
(http://shimizu-lab.dt.u-tokai.ac.jp/lsp.html).)
It would really be nice if there was a useful free benchmark, like
specint, with real world code from diverse areas... does any body know
one?
--
Falk
Reply to: