[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20150524221734.GA4659@openwall.com>
Date: Mon, 25 May 2015 01:17:34 +0300
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: PHC: Parallel in OpenCL
On Sun, May 24, 2015 at 11:58:58PM +0200, Lukas Odzioba wrote:
> 2015-05-24 22:20 GMT+02:00 Solar Designer <solar@...nwall.com>:
> > In general, if USE_BITSELECT isn't helping on GCN, this suggests there's still room for
> > optimization... and then USE_BITSELECT will likely help.
>
> I agree with that, and this also means that current implementation is
> likely memory bound and that should be target for optimization on the
> first place.
This might or might not be the case. IIRC, we also had a case where
bitselect() was hurting performance because, as we found out later, we
had redundant calculations in the code, and the compiler was able to
"see through" simpler operations and optimize them out, but not through
bitselect(). I don't mean to imply this is what we have here; my point
is that reasons why bitselect() isn't helping yet may vary, but in
general, as we both have said above, this suggests there's something
else (possibly multiple things) to optimize first.
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.
