Some Radeon ROCm Packages Pending Review For Fedora

[phoronix]

Phoronix: Some Radeon ROCm Packages Pending Review For Fedora

Earlier this month was word that Fedora developers were looking at packaging Radeon Open Compute (ROCm) to make it easier for their distribution users to enjoy this open-source Radeon GPU computing software from OpenCL to a TensorFlow port. Some of the early packages of ROCm are now under review for Fedora...

http://www.phoronix.com/scan.php?page=news_item&px=ROCm-Some-Review-Fedora

[Azpegath]

Great work! Hopefully they can whip the rest into shape as well, within the near future.
How much would they need to upstream to be able to work on a upstream version?

[pal666]

How much would they need to upstream to be able to work on a upstream version?

upstream unhardcoded paths should be easy, upstream llvm is the task of upstream rocm devs and it is much harder

[brent]

I think they should keep their own variant of LLVM. LLVM upstream has a very slow release rate, and on top of that patch releases are rare. It doesn't look like this is going to change any time soon. Waiting half a year for a critical fix for e.g. a GPU hang simply isn't acceptable.

I'd actually prefer if Mesa used its own branch of LLVM as well. It's quite common that the latest LLVM release turns out to have some critical code generation bugs for AMDGPU and probably other targets as well. And they often aren't fixed until the next major release. Also, this way, Mesa wouldn't depend on the system-wide LLVM version being up to date. On some distributions, LLVM updates are slow.

Alternatively, Mesa and/or ROCm developers could put together a semi-official patch set that applies to the latest LLVM stable release(s) to fix critical issues in LLVM.

[darkbasic]

Forks are not a solution, they should simply make sure that LLVM bugfix releases get issued more often.

[cen1]

if Tom Stellard works on it there is nothing to worry about. He will get it done.