[mvapich-discuss] Error trying to build mvapich2-1.0 with PGI 7.0-7

[David Gunter]

Thanks for pointing out the Python distutils connection to me.  The  
hint about the change from PGI 6 to PGI 7 was a clue as well.  I was  
able to add the flag "-noswitcherror" to my compile scripts which  
tells PGI 7 to only print warnings if unsupported flags are encountered.

So for now, I have been able to compile the 1.0 release with PGI.

Thanks for the help.

-david
--
David Gunter
HPC-4: HPC Environments: Parallel Tools Team
Los Alamos National Laboratory


On Sep 17, 2007, at 2:52 PM, Shaun Rowland wrote:

> David Gunter wrote:
>> I am trying to build the new mvapich2 with the PGI 7.0-7 compiler  
>> but run into a configuration and build error.  Specifically, the  
>> build scripts are trying to use gcc flags "-fno-strict-aliasing",  
>> "-pipe" and "-m64".  Where is it picking up those flags and how  
>> can I disable this behavior?
>
> This is something we will have to investigate some more I think. I can
> tell you where they are coming from however. If you enable TotalView
> support, the src/pm/mpd/Makefile will try to build the mtv.o file with
> the following:
>
>         @if [ "${DO_MTV}" = "yes" ] ; then \
>             currdir=`pwd` ; cd ${srcdir}  &&  ${PYTHON} ./ 
> mtv_setup.py build_ext --build-lib=$$currdir --build-temp=$$currdir ;\
>         fi
>
>
> This runs the mtv_setup.py program in the same directory. This program
> is what is adding the GCC compiler flags:
>
> [rowland at ro0-cr mpd]$ ./mtv_setup.py build_ext --build-lib=$PWD -- 
> build-temp=$PWD
> running build_ext
> building 'mtv' extension
> gcc -pthread -fno-strict-aliasing -DNDEBUG -O2 -g -pipe -m64 - 
> D_GNU_SOURCE -fPIC -fPIC -I/usr/include/python2.3 -c mtv.c -o /home/ 
> 7/rowland/work/mvapich2-trunk/src/pm/mpd/mtv.o
> gcc -pthread -shared /home/7/rowland/work/mvapich2-trunk/src/pm/mpd/ 
> mtv.o -o /home/7/rowland/work/mvapich2-trunk/src/pm/mpd/mtv.so
>
>
> For now you can try building without TotalView support. The problem  
> here
> (at least from what I've observed previously), is that PGI 6.x  
> would not
> die with an error if GCC flags were given, however PGI 7.x does. We
> don't currently have access to a PGI 7.x license (just 6.x), but I am
> sure that's the problem based on some past experiences.
>
> I'll see if there's anything we can do about this. It might be  
> possible
> for you to make a shell script named mtv_setup.py, and just execute
> similar commands to what I see above - but with the right compiler,  
> path
> ($PWD should work), and minus the GCC flags. That's very "hacky"  
> though.
> The reason for the Python script is so the right include path will be
> used I bet, based on the specific Python installation you have.
> -- 
> Shaun Rowland	rowland at cse.ohio-state.edu
> http://www.cse.ohio-state.edu/~rowland/

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20070917/90bed6fb/attachment-0001.html