CUDA: Support a "native" option for CUDA_ARCHITECTURES

At the moment, if you want to build a CUDA application target your own system's GPU, you need to either manually specify the architecture, e.g.

set_property(TARGET tgt PROPERTY CUDA_ARCHITECTURES 70)

or use the CUDA_SELECT_NVCC_ARCH_FLAGS mechanism (see also this issue).

That is too cumbersome. I don't see any reason why the following shouldn't work:

set_property(TARGET tgt PROPERTY CUDA_ARCHITECTURES auto)

which would just perform the equivalent of CUDA_SELECT_NVCC_ARCH_FLAGS behind the scenes (with caching of course).

Bikeshedding: Capitalization (auto? Auto? AUTO)? Use of system or system-devices or system-gpus or available-gpus?

Edited Mar 04, 2022 by Brad King

To upload designs, you'll need to enable LFS and have an admin enable hashed storage. More information