Quoting Alex S <iwt...@gmail.com> (from Sun, 22 Mar 2020 17:53:47 +0300):
Alexander Leidinger wrote:First step would be to get CUDA support in FreeBSD.Ahem, I have a little patch, consisting of several functions copy-pasted from the Linux driver, which is supposed to enable core CUDA functionality: https://github.com/shkhln/nvshim/issues/1#issuecomment-600358438. This won't work for applications depending on unified memory support, but otherwise should be good enough.
Could you plesae describe a little bit more detailed what your shim does? I have the impression it maps some linux glibc symbols which are used in a NVidia library for CUDA which only exists as a linux lib to FreeBSD symbols and when some FreeBSD binary then tries to load the NVidia CUDA lib it is then using the FreeBSD libc/m/pthread.
If this is correct, what would be in your opinion the correct way to get official CUDA support? Something like the following or is more needed?
- implement nvida-uvm - implement os_lock_user_pages- submit the above to NVidia and ask for the CUDA lob to be compiled for FreeBSD
Bye, Alexander. -- http://www.Leidinger.net alexan...@leidinger.net: PGP 0x8F31830F9F2772BF http://www.FreeBSD.org netch...@freebsd.org : PGP 0x8F31830F9F2772BF
pgpiPy9_HIm84.pgp
Description: Digitale PGP-Signatur