Single precision (float32, f32) support is missing

[phansel]

Is your feature request related to a problem? Please describe.
optimization-engine appears to only support doubles (float64) as the native floating-point format for problem specification and code generation.

Passing any f32 arguments to OpEn results in various E0631, e.g.

...   = note: expected closure signature `for<'a, 'b> fn(&'a [f64], &'b mut f64) -> _`
              found closure signature `fn(&[f32], &mut f32) -> _`
note: required by a bound in `optimization_engine::Problem`...

casadi/casadi#4001

A comment on issue #4001 indicates it's possible to request that casadi use floats during code generation using the option "casadi_real".

Describe the solution you'd like
Add some float32 support

Describe alternatives you've considered
Converting all float32 problems to integer problems.

[wolfd]

I had the same issue. I'm using this crate in a limited way (just using the PANOC solver), so I did a hasty (and AI-assisted) attempt to make everything generic across f32/f64. Things seem to kinda work, but with the loss in precision there's a lot of questions of what the new tolerances and constants should be.

https://github.com/AutoPallet/optimization-engine
https://github.com/AutoPallet/lbfgs-rs

[alphaville]

@phansel @wolfd I'll see if I can implement this. Indeed, there can be implications on convergence. @wolfd, could you submit a couple of pull requests - one here and one on lbfgs-rs?

[wolfd]

I will, I admit it makes the code a lot uglier 😅

#373
korken89/lbfgs-rs#9

[alphaville]

I'm currently working on this issue in PR #399. We have already updated lbfgs-rs to support generic float types.