[vstinner]

Optimize json encoder for floats: up to 1.3x faster.

Avoid creating a temporary Unicode object when writing the JSON representation of a float: decode directly the ASCII-encoded string (PyOS_double_to_string()) into the writer.

• Issue: Speed up float encoding in json #159702

[vstinner]

Benchmark:

import pyperf
runner = pyperf.Runner()
for list_len in (1, 10, 100, 1000):
    runner.timeit(f'encode list of {list_len} floats',
        setup=f'import json; L=[1.0] * {list_len}; dumps=json.dumps',
        stmt='dumps(L)')

Results:

Benchmark	main	writer
encode list of 1 floats	1.34 us	1.28 us: 1.05x faster
encode list of 10 floats	2.65 us	2.34 us: 1.13x faster
encode list of 100 floats	11.6 us	9.20 us: 1.26x faster
encode list of 1000 floats	94.5 us	71.3 us: 1.33x faster
Geometric mean	(ref)	1.19x faster

[nineteendo]

I compared the performance on jsonyx (because I didn't want to re-build CPython):

case	before	after	difference
0.0	122.32 ns	67.52 ns	1.81x faster
1e-05	327.62 ns	275.35 ns	1.19x faster
5e-324	578.92 ns	528.56 ns	1.10x faster
1.7976931348623157e+308	2378.23 ns	2326.70 ns	1.02x faster

[vstinner]

cc @serhiy-storchaka