I was wondering if it was yet possible to do OpenACC/CUDA linking? I’m writing up some demo codes/slides to show the various methods of allocating/using memory in OpenACC and an example of getting a deviceptr which is then passed on to CUDA would be handy.
CUDA libraries and OpenACC both operate on device arrays. OpenACC provides some mechanisms for interop with library calls: host_data construct gets device address for pointers inside acc data regions; deviceptr data clause passes pre-allocated device data to OpenACC regions and loops.
The same techniques can be used to share device data between OpenACC loops and custom CUDA C/C++/Fortran/etc. device code or any CUDA Library that uses CUDA device pointers.