Float32 input and output¶
Float32 is supported end to end where the kernels allow it, with no silent conversions anywhere.
What works¶
| Call | f32 input | Notes |
|---|---|---|
extract_features 2D / list |
yes | Native narrow read, float64 accumulation; matches float64 on the same values within rtol 1e-4 (measured max 8.8e-6) |
extract_features_ragged CSR |
yes | Same agreement |
extract_features + minimal / core33 subsets |
yes | Gathered from a full core33 row, still zero-copy |
extract_features + extended / full / non-raw views |
no | Raises ValueError (f32 kernels implement core33 only); pass float64. This path previously panicked across the FFI boundary. |
sliding_features, extract_features_mc, streaming |
no | ValueError; pass float64 |
Precision rules¶
- Accumulation is always float64. Float32 samples widen exactly on read, so the only differences vs the float64 pipeline are f32-arithmetic rounding inside a few kernels (worst measured: second differences).
- The float64 path is unchanged and bit-identical release to release.
out_dtype="float32"casts on write (default"float64").precisionaccepts"float64"/"float32"(anything else raisesValueError) but is currently informational: the input dtype governs the read path either way. It never triggers a silent downcast.