[<prev] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20240919152817.GA7108@openwall.com>
Date: Thu, 19 Sep 2024 17:28:17 +0200
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: Better-performing ???25519??? elliptic-curve cryptography
Hi Frank,
On Sat, Sep 14, 2024 at 12:55:05PM +0200, Frank Dittrich wrote:
> I stumbled across this yesterday, it might be interesting for john
> development.
>
> https://blog.fefe.de/?ts=981af8c1
>
> Automatic translation into English should do a decent job, at least
> Google Chrome did (not perfect, but good enouh).
>
> The links refer to documents written in English anyway:
>
>
> https://www.amazon.science/blog/better-performing-25519-elliptic-curve-cryptography
>
> https://github.com/aws/aws-lc
>
> https://eprint.iacr.org/2022/1303.pdf
>
> https://iacr.org/submit/files/slides/2024/tches/tches2024/1_11/slides.pdf
Thank you! I only skimmed, but I think this refers to optimizations
within the typical single input/output APIs. SIMD is used to some
extent anyway, but with multiple inputs/outputs much better throughput
should be achievable by that means instead. Some of the same
optimizations (more algorithm-level than code-level) could be reusable
within that model as well, but this would need to be re-evaluated.
We do not currently have a multi-input/output host code implementation
of anything ECC in our tree. I was considering getting one in while
optimizing tezos-opencl almost 3 years ago, when it still performed ECC
on host, but I ended up moving ECC to OpenCL instead, which thus
achieved the multi-input/output property without explicit SIMD in the
source code. We could revisit the explicit SIMD for non-OpenCL formats,
but like I say this is mostly orthogonal and far more promising than the
micro-optimizations you refer to, yet is unneeded when we use OpenCL.
Alexander
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
