Re: [Cfrg] How to (pre-)compute a ladder [revised version]

Hi Francisco,

I've updated my (java) implementations of X25519 and X448 in line with
your revised paper. My rough benchmark results for a complete ECDH are
-17.7% and -19.5% time required for X25519 and X448 respectively.

These are perhaps slightly more modest than the paper's estimates. Your
paper already anticipates that (uncounted) field add/sub are significant
in practice.

Part of the difference however is that my implementation of Algorithm 3
handles the final 'q' bits using simple doublings, and removes redundant
cswaps - just as Algorithm 5 does. Of course Alg. 3 is just as presented
in RFC 7748, but it might be a fairer comparison to first apply to Alg.
3 those of your optimizations that are applicable.

Also, my (32-bit) X448 implementation (Alg. 5) required a carry
propagation for 'A' (line 10), due to tight bounds on limbs when
squaring D, E. I did the same for (32-bit) X25519 although I haven't yet
proved it necessary.

An idea: for fixed-base X25519, why not precompute one doubling? i.e.
just calculate (k/2) * 2P, needing only 2 final doublings (you even
already have S of order 4). I am not sure whether this extends to
precomputing (q-1) doublings.

Another thought that occurred to me is that someone might get the idea
to use your fixed-base ladder with some other point. Couldn't this go
awry if that point were already "P + S" or similar? If so, a caution
somewhere in the paper on preconditions for the fixed point might be
advised.

Regards,
Pete Dettman

On 12/05/2017 7:14 AM, Francisco Rodriguez- Henriquez wrote:
> Dear CFRG community,
> 
> We would like to draw your attention to an improved version of our IACR
> pre-print 2017/264 now entitled:
> 
>     "How to (pre-)compute a ladder"
> 
> In this revised version, we present an improved differential addition
> formula that uses pre-computation to match the computational cost of the
> classical Montgomery differential addition.
> 
> Accordingly, our estimates suggest that a full implementation of our
> pre-computable ladder proposal should outperform state-of-the-art
> software implementations of the X25519 and X448 functions by a 40%
> speedup when working in the fixed-point scenario.
> 
> We would be delighted to receive feedback (including sightings of typos)
> from the CFRG community.
> 
> With best regards,
> 
> Thomaz Oliveira, Julio López, Hüseyin Hisil and Francisco
> Rodríguez-Henríquez
> 
> 
> _______________________________________________
> Cfrg mailing list
> Cfrg@irtf.org
> https://www.irtf.org/mailman/listinfo/cfrg
> 