Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 20 Apr 2012 22:59:54 +0300
From: Milen Rangelov <>
Subject: Re: cl_khr_byte_addressable_store

Well especially for RAR on AMD, I had several attempts around that idea and
they ended much slower than the vectorized, bitwise magic version. But you
should leave it just because 4xxx is not supported. I know sometimes it's
hard and it could get VERY UGLY (my rar kernel is frightening). Nvidia may
have no problems with it, but AMD is not the case..

On Fri, Apr 20, 2012 at 10:50 PM, magnum <> wrote:

> On 04/20/2012 09:32 PM, Milen Rangelov wrote:
> > On Fri, Apr 20, 2012 at 7:17 PM, magnum <>
> wrote:
> >> On 04/20/2012 05:12 PM, Lukas Odzioba wrote:
> >>> We will have to add:
> >>> #pragma OPENCL EXTENSION cl_khr_byte_addressable_store: enable
> >>> to kernel for NVIDIA devices (tested on gtx460).
> >>
> >> By the way, this was just now added to RAR too. If I'm really lucky this
> >> fixes it for 7970.
> >>
> >> I had no idea this was required until a couple of days ago and my cards
> >> do fine without it (thus violating the spec. as I read it).
> >
> > Don't use that extension, at least not on AMD. You would be surprised
> with
> > the result. It's just bad. I'd rather do some bitwise macros, they work
> > almost equally well on Nvidia and AMD. Byte addressable stores are a big
> > disaster especially for vectorized code.
> That is confusing and even worse I trust you so this sounds pretty bad.
> Are you effectively saying I really should not write to char pointers
> inside any kernel? I have played with the thought of getting rid of that
> but it would get pretty silly for the inner loop of RAR - or use a
> sh!tload of registers.
> magnum

Content of type "text/html" skipped

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.