[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CA+TsHUCRMnK1eQ3uYNBRpzjt6BvSzfNpo1E07FEKPh_gbtTF0Q@mail.gmail.com>
Date: Sun, 9 Dec 2012 14:24:32 +0530
From: Sayantan Datta <std2048@...il.com>
To: john-dev@...ts.openwall.com
Subject: Re: GCN: indexed access to VGPRs
On Sun, Dec 9, 2012 at 7:43 AM, Solar Designer <solar@...nwall.com> wrote:
> Sayantan -
>
> On Sat, Dec 08, 2012 at 09:08:33PM -0500, Alain Espinosa wrote:
> > On 12/8/12, Solar Designer <solar@...nwall.com> wrote:
> > > How do you access the indexed registers? (And it's the same register
> > > file, just accessed differently - right?) Is it via a private array as
> > > you mention below? -
> >
> > Via the private array in OpenCL and then checking the code generated.
> > Is the same register file as you point out.
>
> Would you try this for bcrypt-opencl? e.g. try to implement two hashes
> per work-item, by having S-boxes of one of them in a local array (like
> it's done now) and the other in a private array. In fact, you can start
> by using a private array only (as an experiment).
>
> Alexander
>
Yes I would. But I think 1024 integers in private array would eventually
end up in global memory. I think the private arrays will/could stay in
VGPRs given that their size is small enough so that it doesn't span across
multiple VGPRs. Which means arrays less than or equal to 512/32=16
elements.
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.
