[matmen]

Resizing was the one benchmark row where ImageScript was slower than Jimp. On GitHub's runner the numbers were 30.6 vs. 87.1 ops/s for halving a 1920×1080 photo bilinearly.

Jimp's bilinear reads 2×2 pixels per output pixel, without antialiasing. Ours is a proper filter like Pillow's and reads 4×4. That makes it slower by design, but the bigger cost was the implementation: float arithmetic in a loop over the filter taps, with the horizontal pass first.

Changes

• Integer arithmetic: weights have 14 fractional bits and are rounded so they still sum to one. Every output index gets the same number of taps, so the passes unroll by that count.
• Vertical pass first: its weights are constant for a whole row, which makes it the cheaper pass to run on the larger image.
• Opaque images skip premultiplying. With non-negative filters (bilinear, area) and up to 4 taps per output row, red and blue share one 32-bit multiplication (SWAR with 8-bit weights and 16-bit lanes). If the pass finds a pixel that isn't opaque, the image is resampled premultiplied instead.
• clone() copies in the array's constructor instead of into a zeroed array, which makes it about 50% faster.
• Fix: when enlarging with RESIZE_AREA, the box filter's weights could all be 0, so whole rows and columns came out transparent. Those pixels now take the pixel in range.

Speed (this LXC, CPU time)

1920×1080 photo	before	after
bilinear to half	49 ms	15 ms
area to half	33 ms	9 ms
bilinear 3.7× smaller	41 ms	24 ms
bilinear 2× larger	392 ms	89 ms
bicubic to half	105 ms	78 ms
bilinear to half, image with alpha	—	20–40% faster

The benchmark run locally on the same machine:

	resize	clone
ImageScript (this PR)	51.2 ops/s	711 ops/s
ImageScript 3.2.0	14.9 ops/s	596 ops/s
Jimp	39.7 ops/s	474 ops/s

Results

The output is no longer bit-identical: it differs from 3.2.0 by at most 1 per channel. This was checked on:

• 1313 corpus images, each at 27 combinations of mode and size (shrinking, enlarging, and changing only one axis)
• noise, the worst case for rounded weights

Larger differences show up in only two places:

• the color of pixels with alpha 0 or 1, where it can't be seen
• the transparent lines from the area bug, which are now filled

Every other operation is unchanged (hashes compared with master over the corpus). tests/targets/{resize,effects,readme}.png are regenerated: resize differs by ≤1, and effects like sharpen and color-dodge amplify that to a few levels in some pixels.

The changelog entry is under 3.2.1.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YJgRK5qpwJEbHPuFv6kgAQ