Chih-Che, Did you get everything to work or you gave up? You are the first non-developer trying to install deal.II with CUDA, so I am really interested to know which problems you are encountering.
Best, Bruno 2017-08-23 20:49 GMT-04:00 Bruno Turcksin <bruno.turck...@gmail.com>: > Chih-Che, > > 2017-08-23 20:09 GMT-04:00 Chih-Che Chueh <chue...@gmail.com>: > > That's still not working with getting the same error. Is this just > because > > it can't find the BLAS and LAPACK libraies? > Did you reconfigure Trilinos in an empty directory? If not do it, you > never know what has been cached by CMake... > > The problem with MKL is that instead of having one or two libraries, > they have split BLAS and LAPACK into a bunch of small libraries and > you need to find which libraries you need to include. You can use this > website to help you: > https://software.intel.com/en-us/articles/intel-mkl-link-line-advisor > Look at the libraries that are needed at linking time and add them to > BLAS_LIBRARY_NAMES or LAPACK_LIBRARY_NAMES. It shouldn't matter which > one. > > > By the way, before I installed the trilinos, I module load many things: > > > > [chueh@icy]$ module load intel/comp-15.0.0 > > [chueh@icy]$ module load intel/mkl-11.1.2 > > [chueh@icy]$ module load openmpi-1.6.5/intel-15.0 > Maybe pgi was a better idea :-D There are two problems with intel 15: > 1) we don't test deal.II with intel15 so I am not sure if it will > works but I guess it should work > 2) the big problem is how are you going to use CUDA? I didn't know > that nvcc supported icc but apparently it kinda does but it looks > pretty bad https://devtalk.nvidia.com/default/topic/947888/intel- > icc-compiler-and-c-11/?offset=4 > For your information, deal now requires C++11 so it won't work. Don't > you have gcc 5 on your machine? > > > [chueh@icy ~]$ module av > > > > ------------------------------------------------------------ > ----------------------------------------------------------- > > /work2/opt/Modules/modulefiles > > ------------------------------------------------------------ > ----------------------------------------------------------- > > cmake/3.9.1 hdf5/gcc-4.4.6_1.8.14 > > intel/comp-15.0 module-git > > ncl_ncarg/6.3.0 netcdf/hdf5_pgi-15.1_4.3.3.1 > > pgi/15.10/x86_64 > > dot hdf5/intel-15.0_1.10.0 > > intel/comp-15.0.0 module-info > > netcdf/hdf5_gnu_4.3.3.1 null > pgi_mpi/15.10 > > grads/2.0.2 hdf5/intel-15.0_1.8.14 > intel/mkl-11.1 > > modules netcdf/hdf5_intel-15.0_4.3.3.1 > > opengrads/2.0.2.oga.2 python/3.5.3 > > grads/2.1.a3 hdf5/pgi-15.1_1.8.14 > > intel/mkl-11.1.2 ncl_ncarg/6.1.2 > > netcdf/hdf5_intel-15.0_4.4.1.1 openmpi-1.6.5/intel-15.0 use.own > Are these all the modules that are available to you :-( Ideally, you > have something like openmpi/gcc5 (gcc 4.8 or 4.9 would also work not 6 > or later because they are not compatible with CUDA8). Now your problem > is that pgi-15 is probably too old to be useful for C++ code and > intel-15 does not work well with cuda. Personally, I would compile my > own gcc 5.4. It's actually pretty easy because you only need C++ and > C. You should _not_ compile support for the other languages. Then, you > load openmpi-1.6.5/intel-15.0 and use this > https://www.open-mpi.org/faq/?category=mpi-apps#override- > wrappers-after-v1.0 > to change the underlying compiler from intel to gcc 5. What I am > wondering is where is your CUDA module? How do you call nvcc? If you > want to use gcc 5 as the underlying compiler, you will need to do this > https://stackoverflow.com/questions/8317510/default- > host-compiler-used-by-nvcc-for-linux. > > Best, > > Bruno > -- The deal.II project is located at http://www.dealii.org/ For mailing list/forum options, see https://groups.google.com/d/forum/dealii?hl=en --- You received this message because you are subscribed to the Google Groups "deal.II User Group" group. To unsubscribe from this group and stop receiving emails from it, send an email to dealii+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
