PG OpenACC on hUMA Kaveri APU zero-copy?

Just wondering if on AMD unified memory architectures like the Kaveri APU, do the PG compilers support the zero-copy option between host <-> device?

Note that this option can be used within OpenCL 1.0 via CL_MEM_ALLOC_HOST_PTR and doesn’t quite require OpenCL 2.0’s SVM. See for example:

http://stackoverflow.com/questions/23378707/data-sharing-between-cpu-and-gpu-on-modern-x86-hardware-with-opencl-or-other-gpg

The beta version of the PG unified memory module currently available, only appears to support Nvidia/CUDA.

Hi gkhanna,

I believe it’s something we will add at some point, but I don’t have a firm time line. I’ve let our management know that you’re looking for this feature and added an RFE (TPR#21422).

  • Mat