[gmx-developers] Coulomb LR energy totals on AMD???

[Erik Lindahl]

On Wednesday, September 10, 2003, at 08:46 AM, Anton Feenstra wrote:

> David van der Spoel wrote:
>> On Tue, 2003-09-09 at 16:06, Anton Feenstra wrote:
>> If it is amd only, then I presume that it has to be in one of either
>> PME, fftw or the ewald corrections. If you say it is reproducible do 
>> you
>> mean that you get the same weird result with the same run repeatedly?
>
> No, similar results in many different runs, i.e. not simulation system
> dependent (but all protein + heme + ligand + water, as far as I've
> checked). But, this is without PME or ewald...

Hm.. I have a vague memory that David or I fixed something related to 
that, but it was minor and definitely didn't affect forces. I could 
send you a copy of newer AMD loops if you want, but the calling 
interface has changed in two ways:

1. All inner loops have exactly the same interface.
2. You have to provide a mutex and a counter.

It's no big deal, but I don't have the time to backport or support it 
myself right now.

>
> It's 3Dnow only, because on our newer AMD, where SSE is used, the error
> disappears.
>
>> The line of testing should be first to turn off assembly in compiling
>> (this will also turn off some other assembly bits in pme.c), then, if
>> the error persists use CFLAGS=-g, a further option could be to upgrade
>> your fftw (2.1.5) or to compile fftw without the i386 hacks..
>
> Isn't there an envvar to disable assembly loops runtime?
>
>

NOASSEMBLYLOOPS=1.

Cheers,

Erik