Re: AES in LibTomCrypt

From: \(lurking\) (bubba_at_aol.com)
Date: 06/14/03 

Date: 13 Jun 2003 21:37:53 -0500

"Tom St Denis" <tomstdenis@iahu.ca> wrote in message news:vYuGa.83171$j9%.53257@news04.bloor.is.net.cable.rogers.com...
> (lurking) wrote:
> > "Paul Rubin" <http://phr.cx@NOSPAM.invalid> wrote in message news:7x8ys692cc.fsf@ruckus.brouhaha.com...
> >
> >>"Brian Gladman" <fake@nowhere.org> writes:
> >>
> >>>A fast implementation of AES can achieve 22 cycles per byte in C.
> >>
> >>Is that possible without contorting the C implementation or the
> >>compiler flags for the sake of one particular cpu?
> >
> >
> > Buiding Tom's LTC version 0.85 with the Intel compiler gave 26
> > cycles per byte for both encrypt and decrypt.

Oops, forgot to name the system. That was on a P4 running Windows.
So there is no surprise there, the Intel compiler and gcc are close.

> On what? GCCv3.2 on a P4 gives 25 cycles per byte [iirc].
>
> > Tom's C code takes 21 cycles per byte to encrypt and 23 cycles
> > per byte to decrypt on AMD64.
>
> Keen...I assume the routines passed the test vectors on the AMD64 box
> right? Always keen to hear which new machines it works on.

Certainly everything passes on AMD64. To confirm that the sanity
checks are really in place, properly compiled in, and working, I
just commented out a portion of the AES code and confirmed that it failed.

> BTW [cheap plug] v0.86 of the lib [due Monday the 16th] will sport
> better MSVC support and a couple of non-ISO C things [like "#line 0
> "bn_xxx.c"] fixed up.
>
> Tom
>