[mario4tier]

to_numpy() already returns float64 for a float64 Series, and astype(float) then copied it again on every call, in the Function, abstract and stream APIs. With copy=False, astype converts only input that is not float64 yet, so results are unchanged.

Median time per call, before and after, for pandas float64 into ATR. Two source builds per host (dev and this branch, both with TA-Lib C 0.8.3), 3 processes each, Python 3.12, numpy 2.5.3, pandas 3.1.0, polars 1.44.2:

API	rows	i7-13700K	Ryzen 7 PRO 8840U
Function	10,000	21.9 → 18.0 µs, 18.0% faster	28.3 → 23.0 µs, 18.8% faster
Function	100,000	182 → 109 µs, 39.9% faster	202 → 127 µs, 37.2% faster
abstract	10,000	54.2 → 49.5 µs, 8.7% faster	64.1 → 55.7 µs, 13.0% faster
abstract	100,000	224 → 151 µs, 32.4% faster	259 → 161 µs, 37.7% faster
stream open	10,000	13.9 → 9.9 µs, 28.5% faster	16.3 → 12.1 µs, 25.4% faster
stream open	100,000	145 → 79 µs, 45.2% faster	154 → 94 µs, 38.8% faster

Opening a stream on a function that reads little history gains most: ADD at 100,000 rows is 47.6 → 1.5 µs, 96.8% faster. Polars gains more than pandas; nullable Float64, Arrow-backed and 2-D-built DataFrame columns gain too. Outputs were identical in every case, including leading NaN or null.

Two things change:

• An int64 Series still needs the conversion and now also pays for the keyword argument: up to 3.3% slower at 1,000 rows (stream open of ADD, 2.46 → 2.54 µs on the i7). NumPy input is unaffected.
• The kernels run without the GIL, so a thread writing to the same pandas Series during a call is now seen by the kernel, as it already is with NumPy input; the copy used to hide it. This matters most on free-threaded Python.

[mrjbq7]

Nice!

[mrjbq7]

Oh oops I thought this was to main branch!

[mrjbq7]

But I liked it, so I cherry-picked it to main.

[mario4tier]

No problem for the merge to dev. Glad that this is heading in the next release ... one way or another 😄