Plures/uarray

[Unknown, Quansight] [Quansight-Labs/uarray] Pull request opened by peterbell10
<https://github.com/Quansight-Labs/uarray/pull/255|\#255 Accelerate multimethod-calls with Vectorcall protocol>
Vectorcall is a new argument passing convention in CPython that was formalized in PEP-590, and made fully public in Python 3.9. It has a number of speed advantages over tp_call which make it desirable to use in uarray. It's also existed in some form since long before 3.9 as CPython's internal "FastCall" convention.

This PR provides shims to make Python 3.9-style Vectorcall available in CPython >= 3.6 and in PyPy as a wrapper around PyObject_Call. It also uses Vectorcall in places where previously I had to explicitly create a new PyTuple object to make the call.

In a simple benchmark, this gives a noticeable speedup on all CPython versions >= 3.7 and is neutral for Python 3.6.

Benchmark code

from timeit import Timer
import uarray as ua

class Backend:
    __ua_domain__ = 'bench'
    def __ua_function__(self, func, args, kwargs):
        return None

mm = ua.generate_multimethod(lambda a: (a,), lambda a, kw, d: (a, kw), 'bench')
ua.register_backend(Backend())

t = Timer('mm(None)', globals=dict(mm=mm))

results = []
n, _ = t.autorange()
n //= 10
for _ in range(10):
    t.timeit(n//10)  # Warm-up run
    results.extend([t.timeit(n) for _ in range(10)])

time = min(results) / n

print(f'{time*1e9:.2f} ns uarray overhead')