Use LUT for linear to sRGB conversion

[thvdveld]

Using a LUT for the linear to sRGB conversion improves decoding performance for the decode_into benchmark:

decode_into LGF5]+Yk^6#M@-5c,1J5@[or[Q6./200
    time:   [766.58 µs 769.86 µs 776.05 µs]
    change: [-62.967% -62.784% -62.602%] (p = 0.00 < 0.05)
    Performance has improved.

However, the size of the LUT is 8.192 u8. The fast-linear-to-srgb feature flag was added, which is enabled by default. In cases where binary size is more important than performance, the feature can be disabled.

Another note in the implementation: using a lookup table reduces the accuracy of the conversion, compared to the original implementation. But all current tests pass. Higher accuracy can be achieved by increasing the LUT size.

[thvdveld]

Oh, I didn't see that there were more tests.

[thvdveld]

This test fails even on main on my machine:

 ---- tests::decode_blurhash_pixbuf stdout ----
thread 'tests::decode_blurhash_pixbuf' panicked at src/lib.rs:455:9:
assertion `left == right` failed
  left: [166, 198, 128, 222, 221]
 right: [77, 210, 4, 80, 15]

[rubdos]

This test fails even on main on my machine:

That might not be an issue; that test has failed randomly anyway. Can you maybe check the visuals for decoding? cargo test should yield you data/out.png, you can compare it to the version that's checked in in the repo. We need to clear that behaviour at some point... cargo test should overwrite files in the repo :'-)

[rubdos]

Cross-checked with #14, doesn't fail.

[rubdos]

Would this be a case for https://docs.rs/phf/latest/phf/ ? 🤔

[rubdos]

@thvdveld a rebase on main should make the tests pass with a higher probability ;-)

[thvdveld]

All checks have passed! 🎉

[rubdos]

On my armv7 phone:

decode_into LGF5]+Yk^6#M@-5c,1J5@[or[Q6./200
                        time:   [4.3007 ms 4.3054 ms 4.3104 ms]
                        change: [-70.320% -70.238% -70.170%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 1 outliers among 100 measurements (1.00%)
  1 (1.00%) high mild

That's 10ms faster :'-)