[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: compaq fortran compiler



I do not have any explicit test suites for fortran, but one of the
primary pieces of software I use is an 88,000 line linear analysis
program.  It primarily spends its time doing floating point stuff,
such as inverting matrices.  I normally run the program under Tru64,
or the Tru64 binary on Alpha Linux.

After grabbing the Debianized fortran from Compaq, I decided to
recompile the program under Alpha Linux.  I am pleased to report that
the program compiled without a single problem, though it would not
link statically.

Under both Linux and Tru64 I compiled the program with the following
flags: -fast -tune host -arch host -assume accuracy_sensitive -O5

Running the program on a 533mhz 164sx I get the following timings for
a particularly hairy model:

Compaq Fortran for Linux:	2:16.55 minutes
Tru64 binary under Linux:	2:14.80
Tru64 binary under Tru64:	2:20.67 (a different machine, but
					identical hardware)

To me, these are identical times, as I was not controlling for machine 
load, though for all tests the computers were not doing anything else
terribly taxing.  I am pleased to see that Compaq Fortran for Linux
appears to perform just as well as Fortran for Tru64.  I would be even 
happier if I did not have access to Tru64.

All three instances get the same answer to the number of reported
decimal places.

Just for comparison I built the same program with f2c and egcs-1.1.2,
and the results were a bit more disturbing.  g77 does not work well
compiling this program.

f2c (egcs) Alpha Linux:		6:29.01 minutes
f2c (egcs) Intel Linux:		4:18.17

The Intel box is a dual processor 488mhz dual Celeron, though the
program is only single threaded.  To me, these numbers should
demonstrate the wisdom of Compaq's decision to release their fortran
for Linux.

-- 
Jeff Lessem.


Reply to: