PIL._binary: Use int.to_bytes() for faster conversion

[akx]

int.to_bytes() has been a thing since Python 3.2, and it's faster than struct.pack having to parse that spec and all that on every invocation. This micro-benchmarks to be 1.2x to 1.5x faster on my machine for essentially free.

Fun fact

There's an upcoming CPython 3.16 optimization that will further help o8() memory-wise (since the int.to_bytes(..., 1) call will end up just returning one of the singleton byteses).

Fun related fact

We could borrow that idea, namely something like

 from struct import unpack_from
 
+_o8_bytes = tuple(bytes(range(256))[n : n + 1] for n in range(256))
+
 
 def i8(c: bytes) -> int:
     return c[0]
 
 
 def o8(i: int) -> bytes:
-    return (i & 255).to_bytes()
+    return _o8_bytes[i & 255]

but it felt a little too much. It would save about 42 bytes of allocations for every o8() call before 3.16 lands, as well as some time (microbenchmarked to be 2.6x faster), but it's a bit weird - I suspect better gains would be had by looking at the call sites for o8() to see if they really need to pack a single byte at a time. (The weird construction for the tuple ensures we do actually get references to the immortal singleton byteses, not new allocations.)

[akx]

I suspect our benchmark suite isn't really testing the write side of plugins that use o8() (BMP, DDS, FLI, GIF, GIMP Gradients, ICO, JPEG, Palm, PCX, PNG, PPM, QOI, SGI, TGA, TIFF, XPM, XVThumb), so CodSpeed doesn't say much about this one.

[hugovk]

Please can you share your microbenchmarks?

[akx]

@hugovk

Just a simple pytest-benchmark benchmark:

from PIL._binary import *


def test_o8(benchmark):
    benchmark(o8, 123)


def test_o16le(benchmark):
    benchmark(o16le, 12345)


def test_o16be(benchmark):
    benchmark(o16be, 12345)

def test_o32le(benchmark):
    benchmark(o32le, 12345678)

def test_o32be(benchmark):
    benchmark(o32be, 12345678)

[hugovk]

Helper	struct.pack	int.to_bytes	Speedup
o8	68.3 ns	37.1 ns	1.84×
o16le	57.3 ns	49.0 ns	1.17×
o16be	57.0 ns	50.5 ns	1.13×
o32le	59.1 ns	50.3 ns	1.17×
o32be	62.4 ns	52.8 ns	1.18×