Fedora Developers Look At Packaging Up The Radeon Open Compute Stack (ROCm)

[bridgman]

Yeah, I hate the idea of having binaries in different places depending on who builds them. If everyone thinks that putting our AMD-built binaries in /usr/local would be OK then I would prefer to always use that location instead.

[schwarzman]

Yeah, I hate the idea of having binaries in different places depending on who builds them. If everyone thinks that putting our AMD-built binaries in /usr/local would be OK then I would prefer to always use that location instead.

Personally I think /opt is better for 3rd party packages. However I don't care much because I don't use 3rd party packages from vendors often - I really prefer distro-supplied packages. These are usually easier to work with even if I have to rebuild them to apply a few extra patches.

Tom Stellard now works for Red Hat and did some initial packaging of ROCm for Fedora. How about working with Tom+Fedora to see what issues distros are having with the ROCm build process? For example Tom mentioned that "hcc" is hard to package and he is waiting for HIP to drop its hcc dependency (if I understood him correctly). Bonus points if you could set up some wiki page (somewhere in the ROCm github repo pages?) where distro issues/missing mainline patches are listed so other distros get a bit of help as well.

Other than Fedora maybe you should also approach Ubuntu/Debian. When Fedora+Debian are able to package ROCm this should mean all major issues have been resolved and other distros should be able to use their scripts as a blue print.

I know that reaching out to big communities is not easy but I think there is a lot of interest on using Tensorflow with distro-supplied packages. IMHO having distro-supplied packages would be huge deal and worth a LOT of marketing dollars. Hopefully you can make a point internally to get some AMD funding. This can really be a chance to counter the argument "everyone is using CUDA -> need to buy Nvidia".

[nanonyme]

Yeah, I hate the idea of having binaries in different places depending on who builds them. If everyone thinks that putting our AMD-built binaries in /usr/local would be OK then I would prefer to always use that location instead.

That sounds fine to me if you're not using the native packaging system of that distro. If you are, then totally directly under /usr.