gh-133968: Add PyUnicodeWriter_WriteASCII() function

[vstinner]

Replace most PyUnicodeWriter_WriteUTF8() calls with PyUnicodeWriter_WriteASCII().

• Issue: Using the public PyUnicodeWriter C API made the json module slower #133968

---

📚 Documentation preview 📚: https://cpython-previews--133973.org.readthedocs.build/

[vstinner]

JSON benchmark: #133832 (comment)

Benchmark	ref	change
encode 100 booleans	7.15 us	6.54 us: 1.09x faster
encode 100 integers	11.6 us	11.7 us: 1.01x slower
encode 100 "ascii" strings	13.4 us	13.2 us: 1.02x faster
encode escaped string len=128	1.11 us	1.10 us: 1.01x faster
encode 1000 booleans	39.3 us	32.9 us: 1.19x faster
encode Unicode string len=1000	4.93 us	4.94 us: 1.00x slower
encode 10000 booleans	343 us	286 us: 1.20x faster
encode ascii string len=10000	28.5 us	28.8 us: 1.01x slower
encode escaped string len=9984	38.7 us	38.9 us: 1.00x slower
encode Unicode string len=10000	42.6 us	42.4 us: 1.00x faster
Geometric mean	(ref)	1.02x faster

Benchmark hidden because not significant (11): encode 100 floats, encode ascii string len=100, encode Unicode string len=100, encode 1000 integers, encode 1000 floats, encode 1000 "ascii" strings, encode ascii string len=1000, encode escaped string len=896, encode 10000 integers, encode 10000 floats, encode 10000 "ascii" strings

Up to 1.20x faster to encode booleans is interesting knowing that these strings are very short: "true" (4 characters) and "false" (5 characters).

[vstinner]

The PyUnicodeWriter_WriteASCII() function is faster than PyUnicodeWriter_WriteUTF8(), but has an undefined behavior if the input string contains non-ASCII characters.

@serhiy-storchaka: What do you think of this function?

[vstinner]

cc @ZeroIntensity

[ZeroIntensity]

Some nits

[serhiy-storchaka]

Well, we had _PyUnicodeWriter_WriteASCIIString for reasons.

But unicode_decode_utf8_writer is already optimized for ASCII. Can it be optimized even more? In theory, it can be made almost as fast as _PyUnicodeWriter_WriteASCIIString.

We can add private _PyUnicodeWriter_WriteASCII for now, to avoid regression in JSON encode, and then try to squeeze nanoseconds from PyUnicodeWriter_WriteUTF8. If we fail, we can add public PyUnicodeWriter_WriteASCII.

[vstinner]

But unicode_decode_utf8_writer is already optimized for ASCII. Can it be optimized even more?

I don't think that it can become as fast or faster than a function which takes ASCII string as argument. If we know that the input string is ASCII, there is no need to scan the string for non-ASCII characters, and we can take the fast path.

You're right that the UTF-8 decoder is already highly optimized.

[vstinner]

In short:

• PyUnicodeWriter_WriteUTF8() calls ascii_decode() which is an efficient ASCII decoder.
• PyUnicodeWriter_WriteASCII() calls memcpy().

It's hard to beat memcpy() performance!

[serhiy-storchaka]

Yes, although it was close, at least for moderately large strings. Could it be optimized even more? I don't know.

But decision about PyUnicodeWriter_WriteASCII should be made by the C API Workgroup. I'm not sure of my opinion yet. This API is unsafe.

[vstinner]

I created capi-workgroup/decisions#65 issue.

[vstinner]

Benchmark:

write_utf8 size=10: Mean +- std dev: 153 ns +- 1 ns
write_utf8 size=100: Mean +- std dev: 174 ns +- 1 ns
write_utf8 size=1,000: Mean +- std dev: 279 ns +- 0 ns
write_utf8 size=10,000: Mean +- std dev: 1.36 us +- 0.00 us

write_ascii size=10: Mean +- std dev: 141 ns +- 0 ns
write_ascii size=100: Mean +- std dev: 149 ns +- 0 ns
write_ascii size=1,000: Mean +- std dev: 176 ns +- 3 ns
write_ascii size=10,000: Mean +- std dev: 690 ns +- 8 ns

On long strings (10,000 bytes), PyUnicodeWriter_WriteASCII() is up to 2x faster (1.36 us => 690 ns) than PyUnicodeWriter_WriteUTF8().

Details

from _testcapi import PyUnicodeWriter
import pyperf

range_100 = range(100)

def bench_write_utf8(text, size):
    writer = PyUnicodeWriter(0)
    for _ in range_100:
        writer.write_utf8(text, size)
        writer.write_utf8(text, size)
        writer.write_utf8(text, size)
        writer.write_utf8(text, size)
        writer.write_utf8(text, size)
        writer.write_utf8(text, size)
        writer.write_utf8(text, size)
        writer.write_utf8(text, size)
        writer.write_utf8(text, size)
        writer.write_utf8(text, size)

def bench_write_ascii(text, size):
    writer = PyUnicodeWriter(0)
    for _ in range_100:
        writer.write_ascii(text, size)
        writer.write_ascii(text, size)
        writer.write_ascii(text, size)
        writer.write_ascii(text, size)
        writer.write_ascii(text, size)
        writer.write_ascii(text, size)
        writer.write_ascii(text, size)
        writer.write_ascii(text, size)
        writer.write_ascii(text, size)
        writer.write_ascii(text, size)

runner = pyperf.Runner()
sizes = (10, 100, 1_000, 10_000)

for size in sizes:
    text = b'x' * size
    runner.bench_func(f'write_utf8 size={size:,}', bench_write_utf8, text, size,
                      inner_loops=1_000)

for size in sizes:
    text = b'x' * size
    runner.bench_func(f'write_ascii size={size:,}', bench_write_ascii, text, size,
                      inner_loops=1_000)

[encukou]

Do we know where the bottleneck is for long strings?
Would it make sense have a version of find_first_nonascii that checks and copies in the same loop?

[vstinner]

Do we know where the bottleneck is for long strings?

WriteUTF8() has to check for non-ASCII characters: this check has a cost. That's the bottleneck.

Would it make sense have a version of find_first_nonascii that checks and copies in the same loop?

Maybe, I don't know if it would be faster.

[vstinner]

Would it make sense have a version of find_first_nonascii that checks and copies in the same loop?

I tried but failed to modify the code to copy while reading (checking if the string is encoded to ASCII). The code is quite complicated.

[picnixz]

I'm happy to have this function public. I always preferred using the faster versions of the writer API when I hardcoded strings, but they were private.

[ZeroIntensity]

Sorry for the late review, LGTM as well.

[vstinner]

I created capi-workgroup/decisions#65 issue.

The C API Working Group voted in favor of adding the function.

[bedevere-app]

GH-134974 is a backport of this pull request to the 3.14 branch.