SpM-lab
Long precompilation times
SparseIR is now super-fast to load. However, it does take quite long to precompile.
I don't understand why this should be, since we mostly do bog-standard linear algebra ... or is the quad-precision linear algebra just much slower?
Quite a small issue, but it is a little annoying :)
On my Macbook Pro M1
julia> @time Base.compilecache(Base.identify_package("SparseIR"))
[ Info: Precompiling SparseIR [4fe2279e-80f0-4adb-8463-ee114ff56b7d]
20.228946 seconds (3.60 k allocations: 326.414 KiB)
Yes, it is a bit long, although I was not aware of this issue.
Disabling quad-precision (i.e. replacing Float64x2 by Float64) reduces precompilation from ~45 to ~30 seconds on my laptop. I have tried in the past to speed this up, but don't really have a clue how to diagnose where the time is spent.
As of Julia 1.9, Julia introduced Package images.
https://docs.julialang.org/en/v1/devdocs/pkgimg/#pkgimages.
Package images are shared libraries that contain both code and data. Like .ji cache files, they are generated per package.
This feature lets us quickly load SparseIR.jl (i.e., using SparseIR). However, it increases the time for precompilation.
But, in the end, we will switch the backend of SparseIR.jl to C++/C. The precompilation time will be reduced then.
In 1.12, there are tools to profile precompilation https://docs.julialang.org/en/v1.12-dev/manual/performance-tips/#Reducing-precompilation-time. Haven't yet tried it though.