Hi! As far as I can tell, we (GCC) don't currently state the minimum CUDA version supported: for nvptx target, and especially then for OpenACC/OpenMP nvptx offloading. <https://gcc.gnu.org/install/specific.html#nvptx-x-none>, <https://gcc.gnu.org/wiki/nvptx>, <https://gcc.gnu.org/wiki/Offloading> all are silent on this topic, for example.
Do distribution maintainers have an opinion on this maybe, what they need to support (for upcoming GCC 11, especially)? (Matthias, others?) "CUDA" here primarily means libcuda (CUDA Driver API and corresponding Nvidia Linux kernel driver). A related topic then is which Nvidia GPU devices are supported by GCC. Due to the generic nature of PTX code generation as well as pretty generic hardware access via the CUDA Driver API in libgomp, this is primarily dictated by whichever Nvidia GPU devices are supported by libcuda, which GCC doesn't influence. (A related topic also is PR96005 "Add possibility to use newer ptx isa", etc., but that's for a separate discussion thread.) I can confirm that as of present, things are still working fine with many-years-old CUDA 6.5/libcuda 346.46 (2014), with Nvidia Tesla K40c GPU, for example. Until recently (but not anymore), I've been running testing with a Dell Precision M4700 laptop's Nvidia Quadro K1000M GPU (GK107 core). If I remember correctly, that one supports Compute Capability 3.0, which Tom recently removed (default) support for via PR97348 "[nvptx] Make -misa=sm_35 the default". For avoidance of doubt: that's OK as far as I'm concerned; not using this laptop anymore for GCC testing, and, as has been noted, 3.5 now is the minimum version with CUDA 11 anyway. (CUDA 11 still supports now, but does deprecate compute capabilities 3.5, 3.7, 5.0, by the way.) As another data point, I noticed that Tom recently removed XFAILs from GCC testcases "for ptxas crash fixed in cuda 8.0". (CUDA 8.0 has been released in 2016.) It's of course nice if old software versions (2014 CUDA 6.5) and hardware devices (Compute Capability 3.5) continue to work with 2020 GCC, but what is our claim what actually is supported, and thus tested (by somebody)? Grüße Thomas ----------------- Mentor Graphics (Deutschland) GmbH, Arnulfstraße 201, 80634 München / Germany Registergericht München HRB 106955, Geschäftsführer: Thomas Heurung, Alexander Walter
