Ben Greenman

Proof of concept https://gist.github.com/LeifAndersen/0c5eff4a84cb56a2e78b2ca64cdf75e5

Bonus: automatically switch to the generic version if the normal version fails somehow (missing import, ...?)

Let's start with the examples we already have. Make a script that runs them all once & records the time taken. EDIT: FPBench should be good for this. We can...

Nope, no progress. We're looking for someone to take the lead on this project.

Ingredients for a new type: - [ ] write cast function to/from TrackedX - [ ] identify the error values for that type - [ ] collect list of operations...

@ashton314 try a quick version later this month --- for the JuliaCon talk. Partition in the main loop of shallow waters, or something similar. Don't make a public interface for...

Concretely: 1. `track(x)` to traverse the array `x` and replace all floats with TFs. Use `TrackedFloat64` as the default but allow changes. 2. `@track foo(...)` to replace all Float-like types...

Yes, that'd be great! Just like a Python decorator.

NB: setting `maxLogs` (#28 ) is a quick way to improve

Example log from Finch advection. Ran 10 times and got basically the same injection for each! `finch_interface.jl 1188` ``` Recording to rand-adv2d-recording_202305081839563... INJECT min (1.1111111111111064e-5, 1.1111111111111066e-5)... run_or_inject at /home/ben/.julia/dev/FloatTracker/src/TrackedFloat.jl:39 min...