[tor-bugs] #16467 [Tor]: Faster Ed25519 implementation.

[Tor Bug Tracker & Wiki]

#16467: Faster Ed25519 implementation.
-----------------------------+------------------------------------
     Reporter:  yawning      |      Owner:
         Type:  enhancement  |     Status:  new
     Priority:  normal       |  Milestone:  Tor: 0.2.7.x-final
    Component:  Tor          |    Version:  Tor: 0.2.7
   Resolution:               |   Keywords:  performance, tor-relay
Actual Points:               |  Parent ID:  #9663
       Points:               |
-----------------------------+------------------------------------

Comment (by yawning):

 Making #9663 the parent for now, since there's a cross-dependency there,
 and I've been doing the integration as:
 https://github.com/Yawning/tor/compare/feature9663

 For those that are too lazy to look at #9663 to pull out things relevant
 here:

 The "on my desktop with TurboBoost etc" x86_64 performance gains look
 like:
  * Generate public key: 60.96 usec -> 22.22 usec
  * Sign a short message: 62.16 usec -> 23.84 usec
  * Verify signature: 161.53 usec -> 77.68 usec
  * Blind a public key: 139.10 usec -> 70.78 usec
  * Convert public point from curve25519: 9.81 usec -> 5.30 usec

 TODO/Open questions:
  * Write the self-test/fallback code.
  * Actually use the batch verification code.
  * Is it worth potential pain to enable SSE2 on ia32 targets? (SSE2
 performance appears to be worse with x86_64).  Benchmarks needed, it is
 probably an ok assumption that the "portable" 32 bit code outperforms
 `ref10` though.
  * Someone who cares maybe should unbreak the Visual Studio build
 (Apparently the VS version of -donna doesn't have SSE2 assembly, but it
 should build).

--
Ticket URL: <https://trac.torproject.org/projects/tor/ticket/16467#comment:1>
Tor Bug Tracker & Wiki <https://trac.torproject.org/>
The Tor Project: anonymity online