Compiling NWChem 6.3 on Cray XC30 (ARCHER) with Intel

Clicked A Few Times

6:33:03 AM PDT - Wed, May 21st 2014

Hi,

Just a note about compiling NWChem on the UK national supercomputing service ARCHER using Intel compilers in case anyone else tries to do the same. As a first attempt I compiled a basic MPI-TS build with the following settings:

export NWCHEM_TOP=...

export NWCHEM_TARGET=LINUX64

export USE_MPI=y

export USE_MPIF=y

export USE_MPIF4=y

export LIBMPI=" "

export ARMCI_NETWORK=MPI-TS

export USE_64TO32=y 

make FC=ftn nwchem_config NWCHEM_MODULES=all &> make_config.log 

make 64_to_32 &> make_64to32.log 

make FC=ftn &> make.log

This seemed to compile OK but in certain circumstances it gives incorrect results, in particular for the force (as printed in force.dat) on the point charge of the following input:

title "nwchem" 

memory 100 mb

print medium

start nwchem

geometry print units bohr nocenter noautoz

symmetry c1 tol 0.0

zn 0.00000000000000 0.00000000000000 0.00000000000000

end

bq units au

force force.dat

 0.00000000000000 3.77945332702064 0.00000000000000 2.0000000000

end

basis print

zn library 3-21g

end

charge 0

set scf:converged false

task scf gradient

I traced the error back to the file src/property/hnd_elfcon.F where the standard Intel compiler optimisations cause the loop in the subroutine multi_reduce to give an incorrect result. I tried switching the i and j loops but the errors persisted (albeit with different numbers). However the errors can be avoided by moving hnd_elfcon.o from OBJ_OPTIMIZE to OBJ in src/property/GNUmakefile, as a result of the lower optimisation level used.

I am not sure whether this is an XC30-specific issue or simply a compiler problem. My Intel build on my local machine is not affected, but I am using a different compiler version (13.0.1 vs 14.0.1 on Archer).

NB: the standard NWChem module on ARCHER is compiled with GNU compilers and so is not affected by this problem.

Tom

Forum Vet

9:40:50 AM PDT - Wed, May 21st 2014
Tom Thanks for the detailed bug report Edo

Forum Vet

5:27:40 PM PDT - Wed, May 21st 2014

Tom
I was able to reproduce your failure with ifort 14.0.1.
I have tried ifort 14.0.3, too; ifort 14.0.3 generates correct code, instead.

You can avoid the Intel 14.0.1 failure by adding the following line just before the call to mat_reduce


!DEC$ NOINLINE

Cheers, Edo

PS. Here is a more complete diff ouput


$ svn diff hnd_elfcon.F
Index: hnd_elfcon.F
===================================================================
--- hnd_elfcon.F	(revision 25615)
+++ hnd_elfcon.F	(working copy)
@@ -90,7 +90,7 @@
                call int_1eelec(basis,ishell,basis,jshell,maxscr,
      &                         dbl_mb(k_scr),nint,dbl_mb(k_buf),
      &                         nder,points,npt)
-c
+!DEC$ NOINLINE
                call multi_reduce(dbl_mb(k_buf),dbl_mb(k_dens),
      &                           elfval,idim,jdim,npt*(max(nder*3,1)))
             end if  ! mod parallel loop

Clicked A Few Times

1:43:24 AM PDT - Thu, May 22nd 2014
Hi Edo, OK, thanks for clearing that up. I'll let the Archer people know they should upgrade! Cheers Tom

Forum >> NWChem's corner >> Compiling NWChem