Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

Subject: Re: [OMPI users] Install BLACS and ScaLAPACK on Leopard
From: Linwei Wang (lxw4393_at_[hidden])
Date: 2008-05-07 17:25:44


Dear sir,

   Thanks very much for your detailed guideline~
   I'm now trying to follow it out~
   I've installed gcc 4.3 & openmpi~
   When compiling CLAPACK, I'm trying to use the optimized BLAS
library by ATLAS, so I set the BLASLIB in the make.inc as:
   BLASLIB = ../../libcblaswr.a -lcblas -latlas
   then build the libraries (also before that, I built the f2clib
following the guideline in netlib
   It went well, but when I tried to built the blas testing code, it
generates errors for "undefined symbols"
   looks like those should be in the f2clib, but I already built it....
  "gcc sblat2.o \
          ../../F2CLIBS/libf2c.a -lm -o ../xblat2s
Undefined symbols:
   "_f2c_ssbmv", referenced from:
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schk2_ in sblat2.o
   "_f2c_sgbmv", referenced from:
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schke_ in sblat2.o
       _schk1_ in sblat2.o
......."

On the other side, when compiling ATLAS, I did the configure as you
said and "make build" went well.
But when I tried "make check" for testing, it again give errors for
"undefined symbols"...

"d: warning in /Users/maomaowlw/ATLAS/build/lib/liblapack.a, file is
not of required architecture
Undefined symbols:
   "_ATL_slauum", referenced from:
       _test_inv in sinvtst.o
   "_ATL_strtri", referenced from:
       _test_inv in sinvtst.o
   "_ATL_spotrf", referenced from:
       _test_inv in sinvtst.o
   "_ATL_sgetrf", referenced from:
       _test_inv in sinvtst.o
   "_ATL_sgetri", referenced from:
       _test_inv in sinvtst.o
"

I'm not sure where is the problem? Can you provide any help?

Thanks again!

Linwei

On May 6, 2008, at 11:11 AM, Gregory John Orris wrote:

> Points to clarify if I may, having gone through this relatively
> recently:
> g77 and gfortran are NOT one and the same.
> gfortran from sourceforge works well, but it is based on gnu gcc 4.3
> and not on the gnu gcc 4.0.1 that comes with Leopard.
> Your best bet is to download the ENTIRE gcc package from sourceforge
> and install it into /usr/local. This includes gcc, g++, and gfortran.
>
> Then you will need to do a number of things to actually get a reliable
> set of packages all compiled from the same version of gcc 4.3.
> Why? Because 4.3 seems to be notoriously faster. AND, I had a lot of
> problems integrating the 4.0.1 libs with the 4.3 libs without errors
> 1. download CLAPACK-3.1.1 from netlib And compile
> 2. Download ATLAS-1.8 from dourceforge (netlib is a little behind
> here) and configure it with the --with-netlib-lapack=your just
> compiled lapack from CLAPACK
> 3. Download OpenMPI 1.2.6 and install it also so that openMPI will
> have the fortran not installed with Leopard.
> 4. NOW you can compile BLACS and ScaLAPACK
>
> In all of this you will need to do a couple of additional things like
> set the env's
> setenv LDFLAGS "-L/usr/local/lib/x86_64"
> setenv DYLD_LIBRARY_PATH "your openmpi path"
> setenv LD_LIBRARY_PATH "your openmpi path"
>
> Do all this right and make sure you compile with the -m64 -mtune=core2
> flags and you will be golden.
>
> So what will you have---
> A new cblas, atlas, lapack, openmpi, fortran, c, c++, blacs, and
> scalapack.
> All on the same version of gnu c.
>
> Alternatively you can buy and use the intel compiler. It is
> significantly faster than gfortran, but it has a host of other
> problems associated with it.
> But if you follow the outline above, you will be left with the best
> that's available. I have lots more info on this, but time is short.
>
> FINALLY, and this is important, DO NOT FORGET ABOUT THE small STACK
> size on Mac's when using gfortran. It's so small that it's useless for
> large parallel jobs.
>
>
> On May 6, 2008, at 10:09 AM, Jeff Squyres wrote:
>
>> FWIW, I'm not a fortran expert, but if you built your Fortran
>> libraries with g77 and then tried to link against them with gfortran,
>> you might run into problems.
>>
>> My advice would be to use a single fortran compiler for building
>> everything: Open MPI, your libraries, your apps. I prefer gfortran
>> because it's more modern, but I have not done any performance
>> evaluations of gfortran vs. g77 -- I have heard [unverified]
>> anecdotes
>> that gfortran is "slower" than g77 -- google around and see what the
>> recent buzz is.
>>
>> FWIW: I tend to use the gnu suite from http://hpc.sourceforge.net/ --
>> it contains pre-built gcc/g++/gfortran binaries and libraries for
>> Leopard.
>>
>>
>> On May 5, 2008, at 2:59 PM, Linwei Wang wrote:
>>
>>> Dear Reeder,
>>>
>>> It does not work. I do think they are from the fortran programs
>>> I'm using (they are files included from the BLACS installation
>>> package, not written by my own.
>>>
>>> The thing is last time when I was using g77, it caused no
>>> problem...
>>>
>>> thanks for your help.
>>>
>>> Linwei.
>>>
>>> On May 5, 2008, at 2:33 PM, Doug Reeder wrote:
>>>
>>>> _s_wsle, _e_wsle
>>>
>>> _______________________________________________
>>> users mailing list
>>> users_at_[hidden]
>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>
>>
>> --
>> Jeff Squyres
>> Cisco Systems
>>
>> _______________________________________________
>> users mailing list
>> users_at_[hidden]
>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>
>
> _______________________________________________
> users mailing list
> users_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/users