[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20121209015309.GA1358@openwall.com>
Date: Sun, 9 Dec 2012 05:53:09 +0400
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: GCN: indexed access to VGPRs
On Sat, Dec 08, 2012 at 08:48:19PM -0500, Alain Espinosa wrote:
> On 12/8/12, Solar Designer <solar@...nwall.com> wrote:
> > On page 53 in AMD_Southern_Islands_Instruction_Set_Architecture.pdf
> > (SHA-1: c64c24077eab93b070d291da3624a8d988d70d5d), there's a mention
> > that VGPRs may be indexed - something I had thought was not possible.
>
> I think Nvidia cards supports this too, but using indexed registers is
> slower than using "static" registers. In some fast kernels i get a
> ~50% speedup changing to normal registers.
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? -
> In OpenCL when you use a private array the compiler may use indexed registers.
Cool.
As to indexed access being slower, this is no surprise. For bcrypt, we
don't have the option to access the registers directly, because we need
them for the S-boxes.
Alexander
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.
