[molpro-user] cross-compiling 2012.1 using PGI compiler for Sandy Bridge architecture

[Kirk Peterson]

Andy,

probably this is just closely related to my problems with the ifort compiler as well.  I'm running Rocks 6.0 on my cluster, so the OS and libraries are identical on both the head node and the compute nodes.  ldd certainly doesn't disagree with this.  You're correct about the licensing issue, the PGI compiler can only be run on a single, registered machine (the license contains the host id #), which is different than ifort. As I mentioned in a previous email, I'm going to upgrade my gcc compiler (from 4.4.6 to 4.7.2) to see if that helps with any of this.

best wishes,

Kirk

On Dec 5, 2012, at 6:01 AM, Andy May <MayAJ1 at Cardiff.ac.uk> wrote:

> Kirk,
> 
> I think -tp=sandybridge is sufficient since according to the man page:
> 
> "-tp=target by itself will compile for a 32-bit or 64-bit target depending on whether the 32-bit or 64-bit compiler is invoked from your command line path."
> 
> and I assume you're using the 64-bit compiler executables.
> 
> The problem is not easy to track down because the hardware, and possibly OS, shared libraries etc. most likely differ between the compile node and run node. It would be nice to know if the same problem is seen when compiling and running all on the same node, but I'm guessing there is an issue of where the compiler license is located otherwise you wouldn't be cross compiling.
> 
> It might be worth running 'ldd bin/molpro.exe' and checking the versions of the outputted shared libraries between the compile and run nodes to see if there is anything that stands out.
> 
> Best wishes,
> 
> Andy
> 
> 
> On 07/11/12 16:10, Kirk Peterson wrote:
>> Dear Jacek,
>> 
>> sorry, I should have been a bit more specific - I was actually trying -tp=sandybridge-64
>> 
>> -Kirk
>> 
>> On Nov 7, 2012, at 2:55 AM, Jacek Antoni Klos <jklos at umd.edu> wrote:
>> 
>>> Dear Kirk,
>>> 
>>> I don't have a PGI compiler, but have you tried -tp=sandybridge-64 to specify 64bit architecture? Or -tp=sandybridge is by default 64bit?
>>> 
>>> Best
>>> Jacek
>>> 
>>> On Nov 6, 2012, at 6:01 PM, Kirk Peterson wrote:
>>> 
>>>> Dear fellow molpro users,
>>>> 
>>>> while I'm waiting for an upgrade to my Intel fortran compiler, I thought I would compile Molpro 2012.1 with my PGI (v12.10) compiler specifying -tp=sandybridge so that I could get the most speed out of some new multicore Intel nodes I just installed.  The (mpp, ga-mpich2) build seemed to go just fine but all the test jobs fail with something like this in the output:
>>>> 
>>>> Library entry O      S cc-pVDZ              selected for orbital group  1
>>>> Library entry O      P cc-pVDZ              selected for orbital group  1
>>>> Library entry O      D cc-pVDZ              selected for orbital group  1
>>>> ? Error
>>>> ? Unexpected non-adapted orbital
>>>> ? The problem occurs in binput:msomat
>>>> 
>>>> 
>>>> Has anyone experience this problem, which is certainly some sort of optimization bug?  Compiling binput.F w/o optimization doesn't help.  The program compiles and runs just fine without specifying a target architecture.
>>>> 
>>>> best wishes,
>>>> 
>>>> -Kirk
>>>> _______________________________________________
>>>> Molpro-user mailing list
>>>> Molpro-user at molpro.net
>>>> http://www.molpro.net/mailman/listinfo/molpro-user
>>> 
>>> Jacek Klos, Research Assistant Professor
>>> Department of Chemistry
>>> University of Maryland
>>> College Park, MD 20742
>>> tel:301-405-1384
>>> fax:301-314-9121
>>> jklos at umd.edu
>>> 
>>> 
>>> 
>> 
>> _______________________________________________
>> Molpro-user mailing list
>> Molpro-user at molpro.net
>> http://www.molpro.net/mailman/listinfo/molpro-user
>>