[gmx-developers] random123 library portability issues

[Szilárd Páll]

PS: Forgot to mention that perhaps as far as I can tell the only
architecture of immediate relevancy (if we don't want to break support
on reasonably new architectures wrt 4.6) are Sparc64s, in particular K
computer and FX10 machines.
--
Szilárd


On Fri, May 30, 2014 at 12:43 PM, Szilárd Páll <pall.szilard at gmail.com> wrote:
> Hi,
>
> A few days ago I did some sanity checks by compiling and
> regressiontesting the 5.0 branch on ARM. This exercise made me realize
> that the recently included Random123 library is now the single most
> limiting factor when it comes to portability. In fact, as it is not
> optional, v5.0 is ATM practically limited to x86* and Powerpc64 (BGQ)
> architectures -- based on the docs (see http://goo.gl/YdKofG)
> Random123 has not been tested (let alone optimized) on anything else.
> Additionally, note that the code won't even compile on non-supported
> architectures, so in its current state GROMACS 5.0 won't compile on
> anything else but x86* and PPC64, e.g. with gcc compilation will be
> aborted by the "#error" in features/gccfeatures.h:38. Removing these
> sanity checks does let the code compile on ARM and regressiontests do
> pass, though.
>
> I've exchanged a couple of emails with the developers and they suggest
> that in practice Threefry should work on ARM (and perhaps other
> architectures too), but testing/validation is definitely necessary.
> Performance is not optimized on unsupported architectures, I assume,
> so some optimization may not hurt too - especially if there are cases
> where we generate many PRNGs per step (BTW has the potential overhead
> been assessed?).
>
> Cheers,
> --
> Szilárd