Data types

A recording moves in the type it was stored in. The conversion to float32 happens once, on the device.

int16 on disk  →  int16 in RAM  →  int16 uploaded  →  float32 on the device  →  float32 out

Why it matters: an int16 recording moves half the bytes a float32 copy would, at every step where bytes are the bottleneck.

What you get back

Outputs are float32 for every input, so every recording processed comes out in the same form. It matches how scipy.signal returns floats for integer input (scipy’s are float64).

Exceptions

Input	What happens
float64	Converted to float32 on the host: it halves the upload, and wgpu has no f64
Python lists and other NumPy dtypes	Become float32 on the host first
Fits that read host data (Whitening.fit_*, PCA, PPCA)	float32 on the host first
Synapse’s in-memory detectors	float32 on the host first

The Python reference says so on each function where it applies.

Writing integers

dspk.io.write_raw writes an integer array as the stored integers (value = stored * gain), and its format defaults to the array’s dtype, so nothing is converted.