Hi Miguel, I would agree with your interpretation of what the developer is saying. Indeed, in the SIESTA manual it says
"Note that Intel compilers default to high optimizations which tends to break SIESTA. We advice to use -fp-model source flag and not compile with higher optimizations than -O2." Following your pointer to the toolchain options I see that there is no option 'source', but 'precise' seems to do the same thing on 64-bit Linux. I'll give it a whirl. Cheers, Loris Miguel Costa <[email protected]> writes: > Hello Loris, > > if the issue is with -ftz, I suppose the simplest way to build without > it is to use set strict or precise to True in the easyconfig > toolchainopts... [1] > > Is the developer saying that Siesta relies on not flushing denormals > to zero? > > Miguel > > [1] > https://github.com/easybuilders/easybuild-framework/blob/ff78a54e4c292004fcd35d1d72b48a4cb2d1ead7/easybuild/toolchains/compiler/inteliccifort.py#L57 > > On Mon, Nov 4, 2019 at 4:20 PM Loris Bennett <[email protected]> > wrote: > > Loris Bennett <[email protected]> writes: > > > Hi, > > > > We have Siesta/4.1-b4-intel-2018b installed via EB. A user of mine > > informs me that the same version of Siesta with the same input, but I > > assume not installed via EB, runs fine on an HPC system at DTU, but not > > on our machine. He contacted a Siesta co-developer, also at DTU, and > > got the following response: > > > > "" > > NICK: Yes, this is bad. Since it errors out during the SCF cycle there is > > definitely something fishy going on. > > Could you ask them to recompile Siesta with these intel flags: > > -m64 -fPIC -O2 -xHost -prec-div -prec-sqrt -fp-model source -qopt-prefetch > > > > (they should try without -ftz) > > "" > > > > Does this seem plausible? If so, would a patch be the easiest way of > > tweaking the compile flags to build a test version? If so, which > > variables should I overwrite? > > Does anyone have any thoughts on the above, or can at least confirm that > the EB version works fine for them? > > Regards > > Loris > > -- > Dr. Loris Bennett (Mr.) > ZEDAT, Freie Universität Berlin Email [email protected] > -- Dr. Loris Bennett (Mr.) ZEDAT, Freie Universität Berlin Email [email protected]
