[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <89b47336e8f12fd7b85e5531b9b77c34@smtp.hushmail.com>
Date: Tue, 25 Aug 2015 16:08:09 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: 7z-opencl crashes with auto-tuning
On 2015-08-25 14:39, Solar Designer wrote:
> On Tue, Aug 25, 2015 at 03:27:47PM +0300, Solar Designer wrote:
>> The 7z-opencl format passes test when run with explicit GWS and LWS, but
>> crashes during/after my patched auto-tuning:
>
> When targeting -dev=1, rar-opencl similarly crashes with the auto-tuning
> patch, at the same point (when starting to tune GWS for the second
> time). However, surprisingly it doesn't crash on -dev=5 (and actually
> gets tuned better than before). 7z-opencl crashes on both.
>
> There are probably changes needed to these two formats, or to how the
> auto-tuning is revised - or both.
>
> Maybe more formats are broken with that patch. I only tested a few.
>
> I hope someone will take it from here.
I will. I'll probably just commit your patches and then continue working
with it. My initial tests with your patches looked very good: It doesn't
seem to slow things down much but end speed is sometimes a good deal better.
Long ago, we did:
1. tune LWS using some arbitrary GWS
2. tune GWS using LWS from 1
This was never very good so we changed it a year ago or more to:
1. tune GWS using LWS=NULL
2. tune LWS using GWS from 1
And now you added a third step, which obviously can make quite some
difference for some formats:
3. re-tune GWS using LWS from 2
I thought about this at times but had the idea it would be too slow for
too little gain but you seem to have proved me wrong. There's also a
lot of cleaning-up to do (for both old and new code). And we should make
it even simpler to use if possible.
Thanks!
magnum
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.
