Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sun, 9 Dec 2012 05:49:22 +0400
From: Solar Designer <>
Subject: Re: bitslice DES on GPU

On Sun, Dec 09, 2012 at 05:34:03AM +0400, Solar Designer wrote:
> You'd need to show (post in here) how you encoded the 0xf0000000 and
> 0x0000000f constants in the kernels.  Being a fairly large immediate
> value, 0xf0000000 probably does not fit in one instruction's immediate
> value field (GCN appears to use a 16-bit field for those, and a 12-bit
> field for address offsets).

Oh, there's also field code 255, which means "Literal constant", "32-bit
constant from instruction stream".  Apparently, this is precisely the
case when a GCN instruction is 64-bit rather than 32-bit.  This might
apply to scalar rather than vector instructions only, though.  It's page
52 in my AMD_Southern_Islands_Instruction_Set_Architecture.pdf.

Anyhow, this:

> On the other hand, 0x0000000f is small
> enough to fit in an immediate offset field or even in a register number
> field (GCN allows for encoding of values in the -16 to +64 range in that
> way, via impossible register numbers).

is sufficient to explain why you found no common offsets for the two
different constants.


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.