BUG: fix overflow in np.round of float16 - #32916

"""Time np.round for float16/float32/float64.

usage: python bench_round_pr.py out.json
Run it once per NumPy build (e.g. with PYTHONPATH pointing at each build,
pinned to one core: taskset -c 6 python bench_round_pr.py out.json).
"""
import json
import sys
import time

import numpy as np


def seconds_per_call(f):
    f()
    k = 1
    while True:  # repeat until one timing takes >= 20 ms
        t0 = time.perf_counter()
        for _ in range(k):
            f()
        if time.perf_counter() - t0 >= 0.02 or k >= 100_000:
            break
        k *= 2

    def one():
        t0 = time.perf_counter()
        for _ in range(k):
            f()
        return (time.perf_counter() - t0) / k

    return min(one() for _ in range(9))


res = {}
for dtype in (np.float16, np.float32, np.float64):
    for decimals in (-2, 2, 6):
        for n in (1000, 100_000, 1_000_000):
            rng = np.random.default_rng(12345)
            a = (rng.random(n) * 200 - 100).astype(dtype)
            out = np.empty_like(a)
            cases = {
                "round": lambda: np.round(a, decimals),
                "round_out": lambda: np.round(a, decimals, out=out),
                "round_strided": lambda: np.round(a[::2], decimals),
            }
            for name, f in cases.items():
                key = f"{dtype.__name__}|decimals={decimals}|n={n}|{name}"
                res[key] = seconds_per_call(f)

with open(sys.argv[1], "w") as fh:
    json.dump(res, fh)

"""usage: python compare_round_pr.py base1.json [base2.json ...] -- new1.json [new2.json ...]
Takes the min over the runs of each build and prints the speedup (base / new)."""
import collections
import json
import statistics
import sys

i = sys.argv.index("--")
base = [json.load(open(f)) for f in sys.argv[1:i]]
new = [json.load(open(f)) for f in sys.argv[i + 1:]]
groups = collections.defaultdict(list)
for key in base[0]:
    speedup = min(r[key] for r in base) / min(r[key] for r in new)
    dtype, _, _, name = key.split("|")
    groups[(dtype, name)].append(speedup)
for (dtype, name), v in sorted(groups.items()):
    print(f"{dtype:8s} {name:14s} median {statistics.median(v):6.2f}x  "
          f"min {min(v):6.2f}x  max {max(v):6.2f}x")