[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <4FE78363.6040607@gmail.com>
Date: Sun, 24 Jun 2012 18:15:15 -0300
From: Claudio André <claudioandre.br@...il.com>
To: john-dev@...ts.openwall.com
Subject: Re: Off-topic (was: bf-opencl )
Hi, my two (incomplete) cents.
My code has 7k IL, 6k ISA code lines (i mean i looked closer on ISA, on
IL i checked line count on gedit). But I can get 20k ISA lines using
imaginative strategies.
If you haven't tried, i think AMD Kernel Analyzer is the best shot. And
[1] is very good news if you want to dig into it.
JtR target is the best performance, so, it means experiment with IL. It
is not my king of pleasure programming to low level, hope it is yours.
Of course, i agree that you can get the best possible results if you
change/develop your IL, ISA or PTX code. But it is something i 'm not
interested because:
1. lock in.
2. it is much, much, i really mean much, harder.
3. AMD software is weak [2][3] (but i suppose it is not going to stay
this way for ever).
4. NVIDIA software is not perfect [4] (but it is improving).
I trust understanding IL is going to be a lonely trip, but using the
Kernel Analyzer, you might discover a lot of new things. Even, how to
patch your kernel and get better results.
[1] http://www.digipedia.pl/usenet/thread/11427/9774/
[2] A terrible sin: http://devgurus.amd.com/thread/158824. See [
*MicahVillmow <http://devgurus.amd.com/people/MicahVillmow>*15/03/2012
09:36]
[3] http://www.golubev.com/blog/?p=223
[4] http://homepages.dcc.ufmg.br/~fernando/
<http://homepages.dcc.ufmg.br/%7Efernando/> (He researches on nvcc
improvement, if you want to get a "feeling" from a specialist on
compilers area, he is nice guy to ask for some advice).
Claudio
Em 24-06-2012 02:45, SAYANTAN DATTA escreveu:
> Hi Alexander,
>
> I'm trying to get my grasp on AMD IL but I am unable to understand a
> few concepts. Can you explain what is the purpose of using
> UAVs(Unordered Access View) ?What are scatter /gather loads ? Also how
> am I supposed to optimize the codes using IL? I mean the generated IL
> has almost 10k lines of instructions in the most time consuming loop.
> So do I have to read each instruction to find out the bottleneck? Any
> suggestions is greately appreciated? I think at the moment you are
> busy with the release of jumbo-6. So there is no hurry,reply when you
> have the time.
>
> Regards,
> Sayantan
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.
