request: libflame
BLIS does not export a full LAPACK (see https://github.com/JuliaLinearAlgebra/BLIS.jl/issues/3), but is supposed to be a part of flame https://github.com/flame/libflame. In LinearSolve.jl, I don't need (or want) libblastrampoline wrappers on here anyways, so if the jll exists it would be immediately useful to a large community and would finish (https://github.com/SciML/LinearSolve.jl/pull/431).
libflame on its own is probably doable, but I am not sure it can fully be integrated into an LBT-based switching system on its own. According to the manual, the lapack2flame compatibility layer only exports the following
--enable-lapack2flame provides a LAPACK 3.11 compatibility layer.
Note also this recent commit that updates on this work: https://github.com/flame/libflame/commit/74b56a226bf976be9a89e810b2c7bd0789c6d2f9
Does FLAME provide a complete compatiblity with LAPACK? It would be nice to include it in Ygg.
Does FLAME provide a complete compatiblity with LAPACK? It would be nice to include it in Ygg.
I actually just saw some discussion on this that happened in the BLIS discord server recently, and apparently, libflame is ABI compatible with LAPACK 3.11 when it is built using --enable-lapack, but is not ABI compatible using their older f2c LAPACK wrapper layer.
So I think we can get a LAPACK version of libflame into Ygg.
And might also then transparently work with LBT. What fun!
Any movement on this?
The initial script in https://github.com/JuliaPackaging/Yggdrasil/pull/8671 was working on all the platforms except for macOS - on there is was somehow creating a dylib that had absolutely no functions at all. No one could figure out exactly what was happening, and so it just kind of stalled.
That's fine since you almost certainly want to use Apple accelerate on that platform anyways, so if it's a no op there then it's fine
The libflame build script is merged now.