[mvapich-discuss] GPU compute capability issue when compiling MVAPICH2 with CUDA

Jonathan Perkins perkinjo at cse.ohio-state.edu
Wed Mar 9 11:31:43 EST 2016


Hello Davide.  We've updated this in our newer releases of MVAPICH2.  If
you'd like to backport the change all you need to do is replace the use of
sm_13 with sm_20.  Please let us know if this helps or if you need any more
details.

On Wed, Mar 9, 2016 at 11:27 AM Davide Vanzo <vanzod at accre.vanderbilt.edu>
wrote:

> Hi all,
> I'm compiling MVAPICH2 2.1 with CUDA support. Here are the configure flags:
>
> ./configure --with-device=ch3:mrail \
>
>             --with-rdma=gen2 \
>
>             --with-ib-include=/usr/include/infiniband \
>
>             --with-ib-libpath=/usr/lib64 \
>
>             --enable-hwloc \
>
>             --enable-fortran=yes \
>
>             --enable-cxx \
>
>             --with-pm=slurm \
>
>             --with-slurm=/usr/scheduler/slurm \
>
>             --with-pmi=pmi2 \
>
>             --enable-cuda \
>
>             --with-cuda=/usr/local/cuda/x86_64/7.5 \
>
>             --prefix=${PREFIX}
>
>
> The problem is that during the configuration step the Makefile is
> generated with the following nvcc flags:
>
>
> NVCFLAGS = -cuda -arch sm_13 -maxrregcount 32
>
>
> despite having GPUs with compute capability 5.2. Since sm_13 is not
> supported by nvcc anymore, the build fails with the following error:
>
> nvcc -cuda -arch sm_13 -maxrregcount 32  -I/usr/local/cuda/x86_64/7.0/include -I/usr/include/infiniband   -I/home/vanzod/Building_zone/MVAPICH/build/src/mpl/include -I/home/vanzod/Building_zone/MVAPICH/source/src/mpl/include -I/home/vanzod/Building_zone/MVAPICH/source/src/openpa/src -I/home/vanzod/Building_zone/MVAPICH/build/src/openpa/src -D_REENTRANT -I/home/vanzod/Building_zone/MVAPICH/build/src/mpi/romio/include -I/usr/scheduler/slurm/include -I/usr/include/infiniband --output-file src/mpid/ch3/channels/mrail/src/cuda/pack_unpack.cpp ../source/src/mpid/ch3/channels/mrail/src/cuda/pack_unpack.cu
>
> nvcc fatal   : Value 'sm_13' is not defined for option 'gpu-architecture'
>
> make[2]: *** [src/mpid/ch3/channels/mrail/src/cuda/pack_unpack.cpp] Error 1
>
> make[2]: Leaving directory `/gpfs0/home/vanzod/Building_zone/MVAPICH/build'
>
> make[1]: *** [all-recursive] Error 1
>
> make[1]: Leaving directory `/gpfs0/home/vanzod/Building_zone/MVAPICH/build'
>
> make: *** [all] Error 2
>
>
> What am I missing?
>
> Thank you
>
> Davide
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20160309/01a2b518/attachment-0001.html>


More information about the mvapich-discuss mailing list