On Tue, 2005-04-05 at 18:17, Andy Polyakov wrote:
> > Current OpenSSL (0.9.8-dev) rc4speed throughput on a Nocona (Em64t, b4bit) 3.6GHz is 272Mb/s, while this version of RC4 code can archive 536Mb/s in RC4Speed.
> >
> >   Would you please review it?

>
> Cool conditional moves in unrolled loop. Have you considered/tried cmov
> instead of jump over move instuction? No, there is no conditional move
> with zero extention, but upper part is maintained zeroed, so that byte
> cmov shoud do... Well, I bet those jumps are seldom taken, so that
> branch prediction logic can make better job than cmov, but I have to ask:-)
>

Well, I tried use cmov here, it just slow down the throughput a lot...
> Or how about moving mozb (%rdi,%r10),%r8d upwards as movzb
> (%rdi,%r10),%r14b and make inter-register move between r8 and r14
> conditional?
>

I will try it.
> The reason I didn't attempt to unroll the RC4_CHAR loop was that I never
> had access to EM64T hardware and simply mechanically ported P4 loop from
> 32-bit implementation [where unrolling affected performance negatively]
> and tested it for correctness on Opteron.
>
> BTW, 272MBps at 3.6GHz? I get 262MBps out of [as just mentioned
> virtually identical] 32-bit code at 2.4GHz P4... A.
>


In fact, Your implement on EM64t isn't that slow if
we change the inc and dec to add and sub.

With that change the throughput boost from 272Mb/s to 396Mb/s.

I have not investigated the 32 bit P4 path yet,
But you should see performance gain on P4 with this change.

Zou Nan hai

__________________________________________________ ____________________
OpenSSL Project http://www.openssl.org
Development Mailing List openssl-dev@openssl.org
Automated List Manager majordomo@openssl.org