GPU cuda autocompilation in IMP
I added an apparently operational cuda auto-build (any .cu file placed in src or src/internal should work). It seems to work fine but let me know if it causes havoc.
Important note - the cuda build system hates the IMP namespace, specifically. Perhaps its some internal cuda macro. It means we cannot include <modulename>_config.h in any cuda files, for now, and need to add a bypass if we need to access an IMP method or object from cuda code.
Barak
p.s. forgot to mention that for building a certain model with cuda, add *-DIMP_CUDA=<module_name1:module_name2:etc>* to cmake
On Thu, Apr 30, 2015 at 12:53 AM, Barak Raveh barak.raveh@gmail.com wrote:
> I added an apparently operational cuda auto-build (any .cu file placed in > src or src/internal should work). It seems to work fine but let me know if > it causes havoc. > > Important note - the cuda build system hates the IMP namespace, > specifically. Perhaps its some internal cuda macro. It means we cannot > include <modulename>_config.h in any cuda files, for now, and need to add a > bypass if we need to access an IMP method or object from cuda code. > > Barak > >
participants (1)
-
Barak Raveh