BLS12-381 signature verification

[aumetra]

Closes #2094

This PR adds common BLS12-381 operations as VM intrinsics.

[aumetra]

Oh yeah, MSRV for tests.. I forgot.

[webmaster128]

Great work. Some first thoughts, not a full review

[aumetra]

Well, we still need a good metric as how to judge point aggregations.
I'm not quite sure about a good benchmark here to find out how much gas should be consumed.

@webmaster128 any ideas about this?

[webmaster128]

The point aggregations are a list of sums. Aggregating n points means m = n-1 sums. In Ethereum this is just a constant 500/800 gas for BLS12_G1ADD/BLS12_G2ADD. EDIT: well, this is EIP-2537 – not implemented

As we execute them in parallel, we get machine-specific performance. It would be good to measure using only a small number of cores (let's say 4) and assume production systems usually perform better.

To benchmark this we can calculate m sums with m = 1, 2, 4, 8, 16, 32, 64, 128 with arbitrary points. Drand generates a point in G1 (the signature) for every round, see https://github.com/noislabs/nois-contracts/blob/main/tests/src/bot.ts#L20-L21. Doing the same for drand network ID 8990e7a9aaed2ffed73dbd7092123d6f289930540d7651336225dc172e51b2ce (classic mainnet) gives you points in G2.

[aumetra]

So about the mismatch of the benchmarks in CI and the values in the environment: The benches for that were run locally by me.

But it was run as follows:

• CPU: AMD Ryzen 7 7700X 8-Core Processor at stock clock rates
• Rayon limited to 4 threads via RAYON_NUM_THREADS

So it wasn't run with the full horse-power of the CPU parallelism, but with the full clock speed. The differences are probably clock speed and parallelism (since IIRC the CircleCI medium class only has two cores instead of 4).

But we can pretty much assume that production environment processors will be at least as fast as the Ryzen with 4 cores utilized IMO.

[webmaster128]

Super nice. I did not read it all yet, but some comments for today

[webmaster128]

🐎🐎🐎🐎🐎🐎🐎