Openwall GNU/*/Linux - a small security-enhanced Linux distro for servers
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Thu, 27 Sep 2012 08:01:07 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: sha512crypt-opencl

On 27 Sep, 2012, at 0:06 , Claudio André <claudioandre.br@...il.com> wrote:

> Hi, thanks again.
> 
> 2012/9/26 Gifts <gifts.antichat@...il.com>
> Benchmark with 20 minutes JtR and 10 minutes cudaHashcat
> 
>  And it is interesting, that Intel OpenCl SDK can't vectorize this
> kernel. I hadn't look deeply into code yet, was this kernel optimized
> for GPU?
> 
> 
>  
> I suppose magnum have something interesting to say about Intel SDK. Let's hear from him.
> 
> I don't have any Intel hardware, so i can't say anything useful. Nevertheless, the kernel that runs on CPU is not optimized for GPUs. It gives acceptable results on AMD CPUs (i would say it was optimized for it).
> 
> Maybe, there are tricks one can use to help the Intel compiler "do the job". Anyone have any clues about it?
> 
> Claudio

I have seen Intel SDK auto-vectorize code that did not have any "hints" nor used vector types. That's impressing. I have also seen (judging from performance) Intel fail to vectorize code that is very simple and that do use vector types. That's a lot less impressing, even annoying. AMD APP seems to handle the latter case better. Note that this is regardless of hardware - you can use Intel SDK with AMD hardware or vice versa. I use to install both on any hardware that does not have a GPU.

Another caveat: I've seen indications that when Intel reports "kernel <xyz> was not vectorized" while building, it only means it was not auto-vectorized from scalar source. I mean, if the code does use vector data types it hopefully ended up as SIMD code despite that message. This is confusing, I'd rather have it silent.

magnum

Powered by blists - more mailing lists

Your e-mail address:

Powered by Openwall GNU/*/Linux - Powered by OpenVZ