[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <7c21b8ffe256f9233853c2aff9572aeb@smtp.hushmail.com>
Date: Sun, 3 Feb 2013 12:40:07 +0100
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: Avoiding AMD ASIC hangs
On 3 Feb, 2013, at 7:57 , Solar Designer <solar@...nwall.com> wrote:
> On Sat, Feb 02, 2013 at 10:49:38PM -0800, Brian Wallace wrote:
>> Solar requested that I implement a split kernel to pwsafe-opencl to avoid
>> AMD ASIC hangs. Specifically, it was requested I ask magnum about the
>> right way to go about this. I won't be able to implement this right away,
>> but I should be able to in the coming week. How should I go about this?
>
> As I wrote in the original thread on pwsafe, before I saw your posting above:
>
> Brian - basically, individual kernel invocations should be taking no
> more than 200ms, preferably much less. This means that with a large
> GWS, you need to be computing only a fraction of the 2048 iterations per
> kernel invocation. Please store intermediate results in global memory.
>
> Some examples:
>
> fgrep HASH_LOOPS opencl_*.c opencl/*
The key to success is to not transfer anything to/from kernel between loop invocations but store the data needed in global memory. Then pick a HASH_LOOPS that takes about 10-20 ms per call for the loop kernel.
http://www.openwall.com/lists/john-dev/2012/09/26/5
For processing events within the loop, see
http://www.openwall.com/lists/john-dev/2012/12/20/1
magnum
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.
