[SCS] link against libblastrampoline #9608
Conversation
kalmarek
changed the title
|
This needs a new version number (and then some changes to SCS.jl), so let's just wait for the next upstream release. |
|
there's also a build failure due to libblastrampoline not found on aarch64-unknown-freebsd, is that expected @giordano ? |
|
@ararslan what version of libblastrampoline supports aarch64-freebsd? Is it easier to just skip that platform? |
|
Should we move freebsd-aarch64 to experimental until Julia 1.12 is out? |
The experimental option didn't work very well because it's a moving target, I prefer being explicit. My general approach with building a package for such a niche platform like aarch64-freebsd is "if it works easily, awesome, if it takes too much effort to make it work then just exclude it and move on". |
I believe it's 5.11.0. I can do a rebuild of 5.4.0 if needed. |
should we filter platforms and create two separate deps with different compat for blastrampoline? What is the compat we should use when targeting julia-1.10 and higher? |
|
I believe the only thing to do for now is to filter out freebsd aarch64. |
|
IIUC, libblastrampoline is the rare example where bumping the compat is not a problem. I went through a similar situation recently and that was my takeaway. Perhaps @giordano can confirm? |
|
stdlibs are always bound to a specific version of Julia. I think you did recently on suitesparse, which is a stdlib itself, so the problem doesn't really matter there. |
|
Ah right, okay. Looking at the builders that depend on libblastrampoline, the most common compat setting is 5.4, so perhaps it could be worthwhile to rebuild that anyway. |
|
@ararslan let me know when rebuilt is ready! |
|
Both 1.10.5 and 1.11.0 have libblastrampoline 5.11.0. However 1.10.0 has LBT 5.4. |
|
Trying LBT update in in #9641 |
|
Thanks Viral! Sorry for not following up on this sooner, I've had basically no time for anything lately. |
| @@ -52,6 +57,6 @@ for platform in platforms | |||
| cuda_deps = CUDA.required_dependencies(platform) | |||
|
|
|||
| build_tarballs(ARGS, name, version, sources, script, [platform], products, | |||
| [dependencies; cuda_deps], julia_compat="1.6", | |||
| [dependencies; cuda_deps], julia_compat="1.10", | |||
There was a problem hiding this comment.
BTW, what's the reason to bump up the Julia compat to 1.10? Is it for LBT? If so, we can probably do 1.8, just to keep it on the conservative side.
My understanding is that you want these things to be maximally backwards compatible, even if only for installation purposes (and the upstream packages can have tighter Julia version bounds).
There was a problem hiding this comment.
thanks for the comment, I'll change this to 1.8 then!
|
@ViralBShah No, I think we'll wait for a new upstream version |
whenever it makes sense;
No changes to scsmkl, since it is already linked against mkl directly.
CC: @odow