Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 12 Nov 2014 23:57:30 +0100
From: magnum <>
Subject: Re: Xeon Phi support?

On 2014-11-12 23:35, Solar Designer wrote:
> On Wed, Nov 12, 2014 at 09:39:07PM +0100, magnum wrote:
>> Solar/Openwall has a Xeon Phi 5110P in a dev box. The (admittedly old)
>> drivers currently installed on that box are *awful*, they fail compiling
>> many kernels and others segfault when running. This driver is worse than
>> Apple's OpenCL drivers for nvidia/AMD two years ago, which would
>> otherwise easily hold the world record in crappiness.
>> Right now I'm not interested in playing with the Phi for that reason alone.
> As you're aware, OpenCL is a wrong way to use Xeon Phi, at least
> currently.  Besides the poor state of OpenCL support for it, why buy the
> normally expensive Xeon Phi and then use it as if it were a cheap GPU?

Yes, I should have mentioned that. But I think there's potential for 
much better OpenCL performance too, with better drivers.

> Perhaps it'd help you appreciate the good things about Xeon Phi if you
> temporarily forget the existence of OpenCL support for it?
> We need to update JtR's CPU "formats" to use Xeon Phi intrinsics.  I did
> that for bitslice DES (unreleased; I should release now that there's
> some interest in Xeon Phi), and the rest of the team may easily do it
> for other SIMD'able hash types.  For a handful of them, such as SunMD5
> and scrypt, things might be slightly interesting performance-wise.

If you release that patch, it would probably help me getting started. I 
usually approach new stuff in a very disorganized manner but once I get 
the hang of it I could easily get the idea to spend several days with 
the Jumbo-specific intrinsics stuff.

> We need to support cross-compiling of jumbo for Xeon Phi, or/and to add
> support for OpenMP offload.  In my bitslice DES experiments, I was
> compiling 1.8.0 with my patches, not jumbo.  To compile jumbo, we first
> need to cross-compile some libraries.

We should make sure Jumbo's autoconf can produce a valid Makefile for 
cross compiling to Phi. Most, if not all, things should be in place 
already - as long as there is a toolchain and libs in place.


Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.