Date: Sun, 22 Jun 2014 16:35:37 -0400
From: Bill Cox <waywardgeek@...il.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] TwoCats multiplication chain

On Sat, Jun 21, 2014 at 5:06 PM, Solar Designer <solar@...nwall.com> wrote:

> Bill, all -
>
> Here's one more idea I'd like to have recorded in here, and it's about
> multiplication latency hardening in general, not TwoCats specifically.
>
> An attack which I think we haven't mentioned yet is speculative
> computation of multiple possibilities of the next MUL before a previous
> MUL completes.
>
> For a simplified example, suppose there's exactly one MUL output bit
> that limits the overall speed (it arrives late and/or is needed early,
> overall more so than the rest of bits do/are).  We can trivially
> eliminate this bottleneck, exposing whatever next bottleneck there may
> be, by starting to compute the next MUL for both possible values of this
> one bit before its actual value is known.  Ditto for the next bottleneck
> bit, and so on.  To do it for e.g. 10 bits at once, we need 1024
> multipliers - this is likely still affordable in terms of die area (it
> may be less than the die area we're forcing the attacker to spend on
> memory), but it might or might not be affordable in terms of routing and
> extra MUXes (which introduce delays, too).
>

This is the basic idea behind a carry select adder:

http://en.wikipedia.org/wiki/Carry-select_adder

More generally, the speedup technique you are describing is application of
what I used to call "Shannon Decomposition", though Wikipedia seems to
prefer Bool's expansion:

http://en.wikipedia.org/wiki/Boole's_expansion_theorem

This can be a very effective technique for speeding up logic, especially
when there is a clearly identifiable critical path, such as a carry chain.
 I've written code to automate this, and it can speed up carry-chain adders
a ton.

I think I could accelerate multiplication chains built using the last
multiplier I designed by about 2X.  The LSB of the upper output word had
about 1/2 the latency of the MSB.    It is possible that Intel and AMD used
something as lame as my Booth encoded multiplier.  I hate to assume they
decided to design something faster, even if they are so speed competitive
with each other.

However, I think it is far more likely they did something awesome like a
Booth encoded Wallace Tree multiplier.  Wallace Tree multipliers generate
outputs at nearly the same time on all of the upper 32 bits, so there would
be little chance to speed up those designs from just a logic redesign.  The
last stage is typically a fast carry-select adder, which is why the output
bits are computed so close in time to each other.  Here's a paper on
multiplier verification that leaks a hint that Intel is comfortable using
Wallace Tree multipliers:

http://www.cs.utexas.edu/users/hunt/FMCAD/FMCAD13/papers/80-Relational-STE-and-Theorem-Proving.pdf

If you can come up with a way to significantly reduce the propagation delay
through multiplication chains compared to chains of Booth Encoded Wallace
Tree multipliers with registers between them, I will be very impressed!
 I've scanned some recent papers that save an adder or two, but there
aren't any practical implementations that demonstrate a significant speedup
over the venerable Wallace Trees.

Bill

Content of type "text/html" skipped

Powered by blists - more mailing lists