On 3/23/21 9:04 AM, Thomas Schwinge wrote: > What is your opinion about backporting that (plus Tobias' documentation > update, plus corresponding web 'changes.html' updates?) to release > branches, so that nvptx offloading users may use GCC 10/9/8 with CUDA > 11.0+?
Hmm, so you're talking about making -misa=sm_35 the default for gcc 10/9/8? Let's see, users of the latest nvidia consumer microarchitecture: ampere are forced to use a cuda11.0+ toolchain, which means that the sm_30 default is wrong for them. Cuda 5.x supports both sm_30 and sm_35, so users should be fine. Cuda 4.x supports neither sm_30 nor sm_35, so users also shouldn't observe a difference. There might be users of cuda 4.x or older and a specific newer driver that supports sm_30 but not sm_35, and those will be the only ones affected. I think we can explicitly unsupport that scenario. So, in conclusion, LGTM. Thanks, - Tom
