CUDA/Fortran and Kepler read-only data cache


Is is possible to enforce CUDA/Fortran compiler to use the read-only data cache (Kepler architecture) following the simple way it is done in CUDA/C by declaring an input data pointer with const and restrict ? Is there an attribute to a device array dummy argument that could do that ?


Hi Pierre,

While we don’t have LDG support into CUDA Fortran yet, we have been looking at using the “INTENT(IN)” attribute as a way to express C’s “const restrict” keywords. Expect this support to be available in the near future.

  • Mat

Hi All,

I can’t seem to find it documented anywhere, but “intent(in)” can now be used to get the same behaviour as “const restrict”. I’ve only tested it in 15.4 so I’m not sure when it became available. Just letting everyone know because it gave me a 10-15% performance improvement for my kernels.