Avoid nested par_iter()s in ed25519_verify_cpu

[mvines]

Empirically removing the nested par_iter()s in this file avoids a
deadlock within rayon when another par_iter() is running in parallel.

Fixes #764

[mvines]

This may have a slight impact on CPU signature verification performance but verify_packet() is still run in parallel. Any benchmarks you want to see run @garious before landing this?

[aeyakovenko]

That’s lame. I think rayon has an option for a local thread pool

[garious]

How about chaining all the par iterators and calling collect() once?

[garious]

This is all we have for benchmarking that module: https://github.com/solana-labs/solana/blob/master/benches/sigverify.rs

[mvines]

Nope nope nope. This is wrong. rayon is working fine, the problem is that ed25519_verify_cpu() is grabbing all the rayon threads and nothing else runs 😊

[rob-solana]

is there an issue covering "ed25519_verify_cpu() is grabbing all the rayon threads"?

[mvines]

Yep, that's just #764.