MAGMA: update to v2.9.0 #11237
MAGMA: update to v2.9.0 #11237
Conversation
|
Small note on the merge conflict: since it's just whitespace, I don't want to trigger another 12h CI just for that, so I'll wait just for reviewers. |
|
Why is ccache not working? |
| # This flag reduces the size of the compiled binaries; if | ||
| # they become over 2GB (e.g. due to targeting too many | ||
| # compute_XX), linking fails. | ||
| # See: https://github.com/NixOS/nixpkgs/pull/220402 | ||
| export NVCC_PREPEND_FLAGS+=' -Xfatbin=-compress-all' |
There was a problem hiding this comment.
Oooh, that sounds very interesting!
There was a problem hiding this comment.
Yeah, the builds in 5614799 were failing after I expanded the gencode targets in 72a7f0c to better support the compute capabilities in several popular GPU cards (see e.g. here).
Then I found out that NixOS uses that flag for all CUDA-related builds, especially for MAGMA itself, and indeed it works pretty well 😃
Despite increasing the compiled targets, which would have increased the library size, the decompressed artifacts sizes were actually lower than before.
|
Looks good to me, and it's great to see more and more packages being able to do cross-compilation of CUDA code (even if it's not as straightforward as I wished)! |
You mean resolving the merge conflict in |
|
No, I was referring to the fact that the rebuilds take over one hour, despite the fact we use ccache to reduce the time for rebuilds. For reference, a rebuild of llvm with warm cache should take here ~5-10 minutes (the strictly build-only time, then the auditor takes a lot longer), and that's a pretty large project, I'm surprised something else takes over one hour to rebuild. |
Oh, I see; haven't got any clue as well. In any case, thanks for merging! 🙏 |
I've done a little restructuring in addition to the version update. The changes enable
aarch64builds and optimal compatibility across a wide range of datacenter and consumer GPUs.