Denis, 2017-09-16 18:30 GMT-04:00 Denis Davydov <davyd...@gmail.com>:
> @Bruno I don't any experience with cuda, but do we need anything special > but > $spack install cuda > to use it? > > I wouldn't use spack to install cuda because there is nothing to compile. You need binaries and you need drivers that depends on your kernel. If you try to do something a little bit too exotic you get into trouble when you upgrade your kernel or your drivers (I have been there...). > I guess one way is to use directly cuda compiler wrappers. It also looks > like openmpi can be compiled with cuda support (then maybe MPI wrappers can > be used directly?). > cuda support in openmpi allows you to by-pass the cpu when you want to send a MPI message from one GPU to another GPU. I don't think it changes the wrapper. > If you can roughly tell me how this shall work, I can try blindly extend > deal.II package in Spack for CUDA and wait for you to test. > I'll try to do it myself. I'll you know if I hit a road block. > > p.s. i guess whole lot of dependencies like Trilinos, PETSc, SLEPc would > need to be turned off though... > I don't think so. Basically, nvcc is a pre-processor that sends all the code to gcc. If there is no cuda code, you basically have a regular gcc compiler. 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.
