Add Support for Single Precision

[rayegun]

Incomplete, posting now for codecov purposes. There are some outstanding issues on how to do promotion and conversion, since the CHOLMOD built-in capabilities are limited.

[ViralBShah]

Wow that is a huge PR!

[codecov]

Codecov Report

Attention: 35 lines in your changes are missing coverage. Please review.

Comparison is base (75081bc) 85.17% compared to head (9731b71) 85.00%.

Files	Patch %	Lines
src/solvers/cholmod.jl	80.87%	35 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #487      +/-   ##
==========================================
- Coverage   85.17%   85.00%   -0.17%     
==========================================
  Files          12       12              
  Lines        8866     8959      +93     
==========================================
+ Hits         7552     7616      +64     
- Misses       1314     1343      +29     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[rayegun]

The vast majority of it is trying to handle promotion at least somewhat sanely. Since mixed precision is not supported (and probably never will be).

[rayegun]

Curses, I have no idea what's going on with Float32 on 32-bit plats. Are there arcane architectural differences afoot here? Or just a bad wrap somewhere :/

[rayegun]

This seems to be an artifact of using sprand here perhaps? Does that make sense to you @ViralBShah? I suppose it's possible that the matrix generated here is close to non-posdef, and that minor architectural changes are causing the issue?

If I run

julia> Ftmp = SparseMatrixCSC{Float32, Int32}(sprandn(Tv, 100, 100, 0.1)); Ftmp = Ftmp'Ftmp + I;

julia> cholesky(Ftmp)

There's a 1 in 4 chance it fails on 32-bit Linux. It never fails using Float64. It also never fails on 64 bit platforms, no matter which datatypes or index types are used.

[ViralBShah]

Just added a heavy diagonal:

SparseMatrixCSC{Float32, Int32}(sprandn(Tv, 100, 100, 0.1)); Ftmp = Ftmp'Ftmp + 100I

Float32 matrices will be more likely to be ill-conditioned. Let's see if my commit helps.

[ViralBShah]

There's a 20min timeout. Should we reduce the set of tests run?

[ViralBShah]

Does the failing test on ComplexF32 suggest that there is a bug somewhere in the cholesky for single precision just added in the complex case?

[rayegun]

I need to trim the tests. I think I just located a bug in the wrapper, I'm running the tests on Ubuntu-32 right now. There was an assumption that only held by accident in the old code. Passing scale factors in sdmult and factorize was not really adhering to the documented API

[ViralBShah]

The tests are timing out on Windows after 6 hours. I have triggered the x64 test again. I believe it was going through successfully earlier in this PR, but perhaps the hanging is a sign of memory corruption.

[dkarrasch]

I think Windows tests have been hanging for a while, also in other PRs, on main there have been issues since f890a1e.

[rayegun]

The tests are currently massive for cholesky. I will work on reducing, we don't really need a complete product of Float32, Float64 x Real, Complex x Int32, Int64 for every test.

I just did that to naively check everything while making the changes

[ViralBShah]

The hangs since SuiteSparse 7.4 are a bit worrisome, but shouldn't hold this PR.

[rayegun]

Are those hangs really from SuiteSparse? The times for the SuiteSparse tests don't look that terrible. They're faster than my MacBook. Which definitely doesn't take 5 hours.

[rayegun]

Oh looks like indeed they happen after SuiteSparse 7.4 update. I wonder why 🤔 because the actual times don't look terrible.

I'll test on my Windows machine.

[ViralBShah]

With #490 clearly identifying the issue on Windows, I think it is safe to merge this.