Inflate: not faster than miniz_oxide on exr benchmark

[Shnatsel]

git clone https://github.com/Shnatsel/exrs
cd exrs
git checkout zune-inflate
cargo bench read_single_image_non_parallel_zips_rgba

The Cargo.toml file is locked in to 0.2.1. Edit it to point to 0.2.2 and you'll see a performance regression, to the point that miniz_oxide used on the master branch becomes slightly faster than zune-inflate.

This is not reflected in the zune-inflate benchmark, so apparently it represents significantly different data patterns.

Edit: actually 0.2.1 doesn't pass tests, you have to upgrade to 0.2.2 for correctness.

[Shnatsel]

Improved by #41

[Shnatsel]

Half the time here is spent in various parts of RLE decoding, so closing in favor of #41

[etemesi254]

Looking into it

[Shnatsel]

If you profile the benchmark with Samply (either on Linux or on Mac) you can double-click the items in the flame graph to view the source code, and where the time is spent within the function line by line. I've found it helpful to understand where time is spent.

This only works while Samply is running, this won't work in profiles shared on the web. That's how I got this image.

[etemesi254]

I didn't know about samply.

Thanks for that,

was wondering how you actually got those values

[Shnatsel]

I've updated the article about bounds checks with instructions for using Samply on Linux.