[verekia]

BC5 error is 3.8–60% lower (MSE) on every real texture, and GPU time is unchanged. All three backends are updated: f16 WGSL, f32 WGSL and WebGL GLSL.

Changes

• Inset pass-1 levels: pass 1 assigns pixels to levels against the min/max range narrowed by ~5.5/256 of the span on each side (scale ×7.3125/7, offset 11/32). The top and bottom levels now gather every pixel near the extremes, not just the single extreme pixel, so the least-squares refit lands much closer to the best endpoints. An exhaustive endpoint search showed the plain min/max seed leaving 0.7–6 dB on the table.
• Lossless flat blocks: blocks spanning 7 or fewer values get a 7-wide endpoint window, so every value lands exactly on a palette level.
• Cheaper refit: the least-squares solve is now a plain line fit from the pass-1 moments (~12 ops per channel). It is accepted whenever it stays in 6-interp mode. The old accept-if-better pricing changed nothing on the inset partition, and dropping it pays for the offset round.
• Offset round: both endpoints shift by the rounded mean leftover error of the shipped levels. Pass 2, now a loop over quads, only has to add ΣL. A whole-level shift moves every palette entry equally, so a block can never get worse, whichever decoder reads it.

Results

Measured on /eval (M3, hardware-decoded quality, interleaved timing):

Content	MSE vs main
Normal maps (rock/wood, 1K–4K)	−3.8 … −6.3%
Colour / roughness / AO / packed	−3.8 … −9%
Displacement (1K–4K)	−9 … −60%
Synthetic normal test card	−2.9% (−4.8% CPU-decoded)

• GPU time: f16 +0.6% geomean, which is within timer quantization; f32 −1.1%.
• WebGL: output is byte-identical to the f32 WGSL on 13 images, including odd sizes and flipY. The constants are exact dyadic values so both compilers fold them identically.
• /test: 0 failures. BC5 PSNR floors are raised to the new levels, and the GPU now beats the CPU reference by 0.18–0.30 dB on every row checked against it.
• bun test: 225 pass.

Notes

• Apple GPUs decode BC4/BC5 with BC7-style weights (0, 9, 18, 27, 37, 46, 55, 64)/64 rather than exact sevenths, so hardware-decoded PSNR on M3 sits about 0.07 dB below CPU-decoded. The encoder still targets the spec palette.
• A full second refit round would add about 0.1 dB more but costs 18% GPU time at 1K, so it was rejected. The offset round gets half that gain for almost nothing.

🤖 Generated with Claude Code