gh-155300: speed up float repr() / str() with new Żmij algorithm

[chirizxc]

🐉 Żmij

• 🐉 Żmij Algorithm

• 🐉 Żmij benchmark

• 🐉 Żmij licence (MIT)

• dtoa-benchmark where this algorithm is benched

My machine

OS: Windows 10 Pro

CPU: Intel Core i5-11300H @ 3.11GHz

Summary

When I search for ussie on GitHub using keyword "zmij", I didn't find a single issue 🧐

Diff is large, but that's mainly because we're vendor algorithm itself into CPython.

On my machine, this results in a 4.09× speedup for repr(x) and a 4.33× speedup for str(x)

Benchmark

I used this short benchmark:

import random
import struct
import sys
import time

random.seed(1)
xs = []
while len(xs) < 10_000:  # noqa: PLR2004
    x = struct.unpack("<d", struct.pack("<Q", random.getrandbits(64)))[0]
    if x == x:  # skip NaN  # noqa: PLR0124
        xs.append(x)

ROUNDS = 50


def bench_repr():  # noqa: ANN201
    best = float("inf")
    for _ in range(ROUNDS):
        t0 = time.perf_counter()
        for x in xs:
            repr(x)
        best = min(best, time.perf_counter() - t0)
    return best


def bench_str():  # noqa: ANN201
    best = float("inf")
    for _ in range(ROUNDS):
        t0 = time.perf_counter()
        for x in xs:
            str(x)
        best = min(best, time.perf_counter() - t0)
    return best


n = len(xs)
print(f"{sys.version.split()[0]}")
print(f"  repr(x):        {bench_repr() / n * 1e9:8.1f} ns/op")
print(f"  str(x):         {bench_str() / n * 1e9:8.1f} ns/op")

• Issue: Use Żmij algorithm for speed up float repr() / str() #155300

[picnixz]

Closing because it's not discussed.