[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <fca4249a5b7af62b4fc3ba82d2742cc9@smtp.hushmail.com>
Date: Tue, 19 Feb 2013 20:06:04 +0100
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: Bleeding device selection
On 19 Feb, 2013, at 19:39 , Sayantan Datta <std2048@...il.com> wrote:
> On Tue, Feb 19, 2013 at 11:40 PM, magnum <john.magnum@...hmail.com> wrote:
> Not yet. I fixed the worst problem in unstable (DES now asks device for max LWS). Should be good enough for Jumbo-8 now. Claudio seems to have fixed the worst issues in bleeding.
>
> Thanks for the patch. However we should not be using LWS greater than 64 even if the device supports higher LWS value because it will cause more wastage of local memory.
We don't! The value may only be decreased, and only if required.
> However if the device doesn't support LWS value 64 or the device doesn't have sufficient local memory for 64 LWS size,then in those cases we should reduce LWS value.
We do now.
> Also the minimum local memory required for this format to run is 2KB per compute unit. Was there any problem with LWS size fixed to 64 , because most of the device around have at least 16KB local memory which is sufficient for LWS size of 64.
"Most" is a key word - with this super trivial patch you don't have to rely on such assumptions. Also, my poor OpenCL/CPU driver only accepts LWS=1 for that format. It's not about local memory - I have no idea why. Probably just a poor implementation from Apple.
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.
