[cfe-dev] Code share CUDA + libomptarget

[C Bergström via cfe-dev]

CUDA, OMP4 and anyone else "offloading" will have similar goals at
some point/layer.. (This is obvious and accurate statement, right?)

To reiterate some previous comments -
A "good" (my opinion) offloading library should probably have 3 layers
of internal API separation

Top - programming model specific (CUDA, OMP. etc)
Middle - Some common denominator which many things end up equating to
Bottom - Target specific
----------
I don't know what mechanism the CUDA support in clang/llvm is using to
offload kernels to the device, but that functionality is probably(???)
substantially similar to what's being proposed for OMP right now
http://reviews.llvm.org/D14031#8e6fbb36

I'd really love to see a new project (liboffload) be created which
aims to address the goal of offloading when and where it's needed. I
don't want to be like other people who gripe about API design and
haven't submitted a patch to fix this, but...

/* I'm just hoping to avoid more of a train wreak */

Thoughts? Suggestions? Anyone else *really* care?