[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[ccp4bb]: Some naive questions about AltiVec
*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
Hi there,
Given the amount of enthusiasm out there for Mac OS X and the recent
discussion of benchmarks, I was wondering if anyone had attempted to
optimize CCP4 code for the G4 processor, in particular to exploit the
AltiVec/Velocity Engine. It seems to me that unless one attempts to
exploit these inbuilt features, the G4 is always going to be a bit of a
lame duck compared to AMD or Intel processors and one will only use Macs
because of the beauty of the operating system! From reading, e.g.
http://gravity.psu.edu/~khanna/autovp.html
I get the impression that the untapped potential for speed enhancements
is pretty large. Also, here is an excerpt from a NASA study on
computational fluid dynamics (accessible through the Web page above):
"Results indicate that the PowerMac G4 system has the capacity to be an
inexpensive high performance scientific computing platform. Much of that
potential is currently unrealized, however, due to the limited amount of
AltiVec support in FORTRAN. Without the parallel vector processing
capabilities of AltiVec, the G4 places near the end of the pack in
preformance tests using standard FORTRAN scientific codes. In limited
cases where AltiVec acceleration was available and tested under FORTRAN,
the G4 showed a clear advantage, with 4-7x greater performance and a
5-8x greater cost effectiveness than all other workstations evaluated.
Examples presented in this report show that only minor re-coding would
be necessary to implement AltiVec instructions if they were accessible
to standard FORTRAN programming."
The problem is that there only seems to be one automatic vectorization
program, i.e. VAST from Veridian. The combined FORTRAN/C optimizer costs
$1000! Absoft bundle VAST for FORTRAN with their ProFortran compiler for
$1300. This is obviously a bit much for an individual lab, so I wondered
if the CCP4 developers, or any other developers and implementers out
there, had thought about this issue. Obviously it means producing major
amounts of tailor-made code for a specific operating system, which is
maybe not true to the CCP4 philosophy.
Regards
Derek
--
Derek Logan tel: +46 46 222 1443
Associate professor fax: +46 46 222 4692
Molecular Biophysics
Lund University
Box 124, Lund, Sweden