Re: gfortran

[lsorense@csclub.uwaterloo.ca (Lennart Sorensen)]

On Mon, Oct 24, 2005 at 10:32:56PM +0200, Dominique Rousset wrote:
> I've been a silent subscriber to this list for the past 3+ months, and now, I 
> am coming with my question...
> 
> I am benchmarking a fortran code of my own (simulation of seismic wave 
> propagation)
> whose output is a direct access binary file.
> I just modified it a little from Intel fortran90 to gfortran.
> I succesfully ran it at work on a PowerPC linux, an ultrasparc linux and a 
> Pentium 4 (old, fully 32 bits) linux.  The results are consistent and 
> consitent (not really equal, of course) to those obtained on Intel proc using 
> Intel compiler.
> 
> I tested it at home on my Athlon64.  Dual boot one in 32 bits and the other 
> one in 64 bits, both debian etch.
> 
> I am surprised that the output fiile in 64 bits mode is larger of 808 bytes.
> It's just one excess byte per record. 

Were you expecting an extra 4 bytes per record?

> Any clue about that ?
> Do you thing I must issue a bug report to the maintainer, to the upstream 
> developer team ?
> 
> Thanks
> D.
> 
> PS : computation time is 15% longer in 32 bits mode, using the same compiler 
> options.

Do the results seem to match as well numerically?  I imagine you are
probably doing floating point, which is certainly expected to be faster
in 64bit mode since it uses sse rather than x87 instructions.

Could you post your code?  Might be easier for someone to help if they
knew what the code was.

Len Sorensen