[gmx-users] bug in 3.1.5_beta Altivec loops?

[Erik Lindahl]

Choon Peng wrote:

>Dear all,
>
>  Anyone came across funny numerical results obtained from version
>3.1.5_beta on G4 using Altivec?
>For a NPT simulation (NVT is OK), I actually got a value of200 bar for
>pressure while the same beta installed on a PIII clustergave the more
>acceptable result of 3.8 (correct result is 2.3)!Also, quantities like
>total energy, temperature, etc are also quiteway off.
>
>Initially I suspected there's general bugs in the beta release but
>after confirming with a PIII (using SSE) installation, the prime
>suspect is now the Altivec calculations. I know they're faster than in
>3.1.4 due to more tuning and will inaccuracies creep in as a result of
>too aggressive tuning for performance?
>
>Thanks.
>  
>
Hi Coon-Peng,

Pressure is *extremely* sensitive to minor numerical differences, but 
send me
an example if the error in energies and a small input file to reproduce it.

Cheers,

Erik