Speed up Image.blend()

[akx]

I was working on a project that uses Image.blend() and it showed up pretty red in a Viztracer timeline, so I decided to investigate a little.

Very small patch, but allows the C compiler to optimize the loops better:

• we know ysize and linesize can't change during iteration (but the compiler can't statically know that), so hoisting those to locals
• we know imOut is a fresh allocation and can't alias another pointer, so adding restrict to note that.
  • restrict should be available on MSVC in C99 mode without having to use the MS extension name __restrict; we'll see how it pans out in CI and adjust accordingly.

In addition, added a little fast-path if you pass in the same image object twice.

On my machine, this shows a 300% speed improvement (with blend_bench.py being a pytest-benchmark suite doing benchmark(Image.blend, im1, im2, 0.5) on two 2048x2048 RGB images).

(main) $ uv pip install -e . && uv run --no-sync pytest --benchmark-autosave blend_bench.py
[...]
Name (time in ms)        Min      Max    Mean  StdDev  Median     IQR  Outliers       OPS  Rounds  Iterations
-------------------------------------------------------------------------------------------------------------
test_blend_perf       9.1552  11.0795  9.7906  0.2351  9.7934  0.1972      17;5  102.1389      90           1

(faster-blend) $ uv pip install -e . && uv run --no-sync pytest --benchmark-autosave blend_bench.py
[...]
Name (time in ms)        Min     Max    Mean  StdDev  Median     IQR  Outliers       OPS  Rounds  Iterations
------------------------------------------------------------------------------------------------------------
test_blend_perf       2.1012  3.2935  2.2357  0.1044  2.2116  0.1040      43;7  447.2908     284           1
------------------------------------------------------------------------------------------------------------

(faster-blend) $ pytest-benchmark compare .benchmarks/Darwin-CPython-3.14-64bit/*.json --between=ops

----------------- benchmark: 1 tests, 2 sources ------------------
Name (time in ms)     0001_2d4bc04 OPS  0002_9da64fc OPS      ΔOPS
------------------------------------------------------------------
test_blend_perf               102.1389          447.2908   +337.9%
------------------------------------------------------------------

I also looked at using the integer math from AlphaComposite.c, but that only made things a little slower.

[hugovk]

Please can you share blend_bench.py or a similar benchmark?

[akx]

@hugovk Sure thing:

import pytest

from PIL import Image


@pytest.fixture(scope="module")
def images() -> tuple[Image.Image, Image.Image]:
    size = (2048, 2048)
    im1 = Image.new("RGB", size)
    im2 = Image.new("RGB", size)
    im1.frombytes(bytes(i % 256 for i in range(im1.width * im1.height * 3)))
    im2.frombytes(bytes((i * 7 + 31) % 256 for i in range(im2.width * im2.height * 3)))
    return im1, im2


def test_blend_perf(benchmark, images: tuple[Image.Image, Image.Image]) -> None:
    im1, im2 = images
    result = benchmark(Image.blend, im1, im2, 0.5)
    assert result.size == im1.size

[akx]

Btw, if this is merged, I can take a look at other similar loops and apply the same optimizations - looks like there could be lots of easy perf pickings 😊