Clarification about benchmarks

[consensys]

Clarification about benchmarks

[yelhousni]

Hi @jon-chuang, I indeed got confused with the naming convention. BenchmarkG1AffineDouble is indeed in Jacobian coordinates and the reported timings for gurvy double are in extended Jacobian. I will update the corresponding data.

To be fair to all libs, I am updating the post as follows:

• correct arkworks-rs timings following a assembly bug (arkworks-rs/algebra#192)
• add relic assembly preset (x64-pbc-bls12-381)
• update kilic/bls12-381 after recent imporovements
• update constantine after recent imporovements
• add iden3/ffiasm to the benchmark

Thank you for opening the issue! Let me know if you notice anything else.

[jon-chuang]

Hi @yelhousni thanks!

I'm not too familiar with the extended jacobian coordinates, do you know how they are used in gurvy specifically? Meaning, are they used in scalar mul or msm for instance? Just curious as they seem to provide some nice speed results, I was wondering if it's been tested to work well as part of bigger routines.

Currently we are making some simple changes to improve performance across the board arkworks-rs/algebra#205 prompted by these benchmarks, it would be nice to include these improvements while measuring the asm for arkworks.🙏

Another issue is one should use the feature "n_fold" for more accurate benchmarking... The current benchmark framework is not very good for ns-level measurement and hopefully we will be changing it soon.

[yelhousni]

Extended Jacobian coordinates are indeed used in gurvy for MSM.
Let me know when you finish improving arkworks-rs and I will update the benchmark.
I am using n_fold feature for cargo bench.