Performance analysis of Polyhedron/gas_dyn

[Tobias Burnus]

Hi Janne,

Janne Blomqvist wrote:
> I spent some time with oprofile
Thanks
> in subroutine EOS the line
>     CS(:NODES) = SQRT(CGAMMA*PRES(:NODES)/DENS(:NODES))
This already became faster by Richard's patch in December 2006, cf. PR
21466.


Regarding other compilers, I have the following timings for gas_dyn (AMD64):

gfortran 11.62s [100%], ifort 9.1 8.58s [73%], ifort 10beta 7.23s [62%],
sunf95 12.06s [103%]

And for fatigue:

gfortran 12.87 [100], ifort9 12.01 [93], sunf95 8.34[64], pathscale 6.12
[47]

Maybe someone good at profiling and assembler could have a look at the
file produced by ifort or sunf95 to see what these compilers are doing
better.

> http://www.polyhedron.com/
I think since those days, gfortran improve a bit, cf.
http://www.suse.de/~gcctest/c++bench/polyhedron/polyhedron-summary.txt-2-0.html

(I wonder whether yesterdays's upwards surge for gas_dyn is correct, if
yes, one should try to find the cupid. For test_fpu we also have a
regression since January.)

> http://physik.fu-berlin.de/~tburnus/gcc-trunk/benchmark/
Since the patch in PR 31699 fixes the crashes in two test cases, today's
timings for gfortran should be ok.

Tobias