From: "Deranged Mutant" <WlkngOwl@UNiX.asb.com>
Date: Sun, 21 Jan 1996 11:46:09 +0800
To: Wei Dai <weidai@eskimo.com>
Subject: Re: HAVAL (was Re: crypto benchmarks)
Message-ID: <199601202200.RAA09207@UNiX.asb.com>
MIME-Version: 1.0
Content-Type: text/plain


> Thanks.  It looks like F4 and F5 are improved.  Do you know how these
> optimizations can be done in general?  I tried playing with F2 
> as a multivariate polynomial with coefficients in GF(2) in Mathematica.  
> This seems to work and I found several equivalent expressions that take 
> 13 operations (the original also takes 13 operations).  Is there a tool 
> that can do this automaticly?

I did the optimizations by hand.  Simple rules of boolean arithmetic 
and logic (you know, things like Demorgan's Law applied to binary 
operations).  Other processor-related optimizations can be done by 
hand, such as add x,x instead of shl x,1.

I think I had the same proglems with F2 as well.  Couldn't find a way 
to optimize it reasonably.

 > The biggest problem I have with HAVAL now is that with 4 or 5 passes the
> transform functions are larger than 10k even with compiler optimzation for
> size.  Since the Pentium L1 instruction cache is only 8k, this makes HAVAL
> with 4 or 5 passes extremely slow.  Do you have ideas how I can fit the 
> transform functions into L1 cache?

You might do some creative optimization to use more registers than it 
does.  I haven't looked at it in a while.  The code was so huge and 
slow compared to optimized MD5 and SHS that I have up using it for an 
unfinished encrypted file system.

Rob.
 
--- "Mutant" Rob <wlkngowl@unix.asb.com>

Send a blank message with the subject "send pgp-key"
(not in quotes) for a copy of my PGP key.