AMBER Archive (2009)

Subject: Re: [AMBER] Compiling AmberTools 1.2 with intel compilers for Mac Osx in 64 bits

From: David Watson (dewatson_at_olemiss.edu)
Date: Thu Jul 30 2009 - 14:10:39 CDT

On Jul 30, 2009, at 8:39 AM, Alan wrote:
>
> Starting always with fresh uncompressed ambertools folder.
>

I think the main thing is to perform a 'make -f Makefile_at clean' and
then rerun the ./configure_at line with the chosen options.
>
> I didn't try with gcc and MKL so far. That said, both 32 and 64 bits
> compilation with "configure_at icc" worked with MKL, however doing
> "make -f
> Makefile_at test" I had some surprises:
>
> 1) With 32 bits and MKL, almost everything worked fine, even the
> FAILURES
> (when I compared with my ambertools gcc without MKL and 32 bits)
> excepted
> for '1rrb_vac_mwcovarmat_evecs.dat.dif' in 'ptraj_matrix'. I got the
> absolute values identical to 1rrb_vac_distcovarmat_evecs.dat.save
> except for
> the sign where with MKL they were inverted (like a matrix multiplied
> by -1),
> e.g.:
>
> 537c537
> < 0.00237 -0.00485 0.00133 0.00259 -0.00563 0.00178
> 0.00236 (save)
>> -0.00237 0.00485 -0.00133 -0.00259 0.00563 -0.00178
> -0.00236 (mine)
>

This is my experience exactly.
As far as the other failures, they are within the absolute/relative
margins listed in the comments.

Another oddity is that use of MKL results in this:
=====================================================
Running test to do simple lmod optimization

1c1
< Glob. min. E = -122.793 kcal/mol 

---
 > Glob. min. E         =     -127.803 kcal/mol
    FAILED (probably OK if energy is -115 to -125)

=====================================================

> Is it OK?
>

I have no idea.

> The compilation with 32 bits with gcc and without MKL has not such a  
> diff.
>
> If compiled with icc and NO mkl I got no such a diff either.
>

I get tons of pragma warnings with icc.

>
> CONCLUSION 2: Intel MKL libs 64 bits doesn't work.
>

Exactly my experience.

>
>>
>>> NOTE2:
>>> I didn't find more documentation about './configure_at -h' for  
>>> options
>>> '-mpi' or '-openmp'. It would enable some programmes in ambertools  
>>> to use
>>> multi-cores? Which programmes could take advantage of these options?
>>>
>>
> Any comments about NOTE2 above?

You could dig around in the source to find out which files use OPENMP  
defines.

You may compile with gcc-4.2/g++-4.2 if you use the './configure_at - 
openmp' step followed by a few edits to the resulting config.h:

FLIBS= $(LIBDIR)/libsym.a $(LIBDIR)/arpack.a $(LIBDIR)/f2c.a -L$ 
(MKL_HOME)/lib/32 $(MKL_HOME)/lib/32/libmkl_solver.a -lm
kl_intel -lmkl_intel_thread -lmkl_core -fopenmp -lpthread
CC=gcc-4.2
CXX=g++-4.2

Now the interesting thing is that the lmod test gets much worse:
=====================================================
Running test to do simple lmod optimization

1c1
< Glob. min. E         =     -122.793 kcal/mol
---
 > Glob. min. E         =     -135.362 kcal/mol
    FAILED (probably OK if energy is -115 to -125)

=====================================================

Otherwise, the same test failures appear as without openmp.

_______________________________________________
AMBER mailing list
AMBER_at_ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber