Zune-inflate on 0.2.0

[etemesi254]

@Shnatsel since you were interested, I almost finalized zune-inflate.

The speedups are nice over miniz-oxide but libdeflate absolutely floors us with its decode speeds.

May spend sometime tweaking the knobs to get mine faster but initial benchmarks can be found in the Readme.

[Shnatsel]

Nice! I'll take a closer look shortly.

A nit on the README: flate2 with miniz_oxide backend does use some unsafe - it has miniz_oxide write into an uninitialized buffer. If you need 100% safe decompression, you have to use miniz_oxide directly, but its streaming interface is more low-level than flate2.

[Shnatsel]

Since the performance is not as great as libdeflate yet, perhaps it makes sense to remove the remaining unsafe and position the library as somewhat slower than libdeflate but 100% safe Rust? It still beats miniz_oxide and makes using zune-inflate instead of miniz_ozide a no-brainer.

You can later re-add unsafe optimizations as an optional feature, once they have a bigger impact on performance. E.g. https://github.com/pseitz/lz4_flex does this.

I've naively replaced the implementations of the only two unsafe functions with the appropriate methods on slices, and hyperfine reports 4% performance impact because of that on the enwiki_part file.

[etemesi254]

345ca8a

[Shnatsel]

A benchmark against flate2 with zlib-ng backend would also be great. This is the option people usually turn to for more performance, since it's just a feature flag on the familiar flate2.

[etemesi254]

Done, we still beat it by a good margin, but libdeflate is really impressive.

[Shnatsel]

I believe zune-inflate is production-ready, considering all the different kinds of fuzzing we've done. There may still be room for API tweaks (#33) and performance optimizations, but it's already superior to miniz_oxide and should be ready for early adopters. I'd like to replace miniz_oxide with it in https://github.com/johannesvollmer/exrs

Is there anything else you'd like to do before publishing to crates.io? Publishing does not have to coincide with a public announcement.

[etemesi254]

Yes, there are some things I'd like before publishing.

1. A working CI/CD.
2. A benchmark library/program external to my computer, I was thinking something like using aws/google to run benchmarks and update a Github page with benchmarks, this is important because a selling point of the library is speed, one person's computer is not a measure of speed, and allows a good regression testing.
3. Licensing, MIT/Apache but should be explicit though.

[etemesi254]

Hi, all that is left is 2, I can't do that as I'm just a student with interesting inclinations, I was hoping for now you might :).

But ideally I've set up criterion in that calling cargo bench --workspace will produce a html report in target with all the criterion benchmarks in it, so one can peruse whatever benchmarks they want.

Currently limited to first class decoders, i.e jpeg and png and inflate, things I spent my time optimizing, not ppm, qoi for now.

[Shnatsel]

AWS has a free tier for compute: https://aws.amazon.com/premiumsupport/knowledge-center/free-tier-windows-instance/ (I'm sure it allows Linux as well, which is cheaper because you don't have to pay for a Windows license).

I'm not sure if the VMs created that way are noisy or not, however. The variance may come either from over-commit of host resources or from you getting different CPU platforms in different runs.

At least the CPU platform variance can be mitigated by running benchmarks on both master and the proposed PR in the same run, and using the relative values in place of the absolute values.

IIRC Github Actions isn't suitable for this specifically because the results are noisy, but that's just hearsay, I haven't verified it myself. Google Cloud also has an always-free VM instance tier but those don't even get an entire CPU core to themselves, that's useless for benchmarking. The same goes for Azure.

If AWS (or some other cloud) provides a suitable free tier, you can configure the VM to pull the repo and run the benchmarks on boot, and then shut down. On AWS you can trigger it e.g. through a call to AWS Lambda, which accepts an HTTP request and makes an API call to start up the VM. There may be better ways, I'm not very familiar with AWS.

I suggest investigating whether AWS allows non-shared-core VMs in the free tier, and if so, if the benchmarking runs are at all consistent.

If there's no free tier available, I suppose I could foot the bill for the infra for a while.

[etemesi254]

Ideally, I don't want per PR testing, just those that either promise a speed increase and probably a week or a month to just see how we are doing from the baseline.

I'll definitely check free tiers or see if there is a place I can get some free compute, but for now I think my machine may suffice.
But thanks for your help

[etemesi254]

With all the parameters met, and https://etemesi254.github.io/posts/Zune-Benchmarks/ which consequently leads to the benchmark page at https://etemesi254.github.io/assets/criterion/report/index.html

It follows that https://crates.io/crates/zune-inflate should exist.

PS for exr, I suggest only enabling zlib, saves some binary size by not including some huge crc table that is used for gzip decoding.

Happy programming

[Shnatsel]

The PR for exr is up: johannesvollmer/exrs#179