On Sun, 2010-02-07 at 23:31 -0800, Sylv1 wrote: > Hey, > > can anyone tell me more about the issues for the lookup stage with the > new design? > I would like to develop a lookup tool for the table on my local > computer and would like to no more about what i should take care > about. > > Thanks for the help, > Cheers > > Sylvain > > --- On Sat, 2/6/10, sascha <sascha.kriss...@web.de> wrote: > > From: sascha <sascha.kriss...@web.de> > Subject: Re: [A51] lookup tool for tables > To: a51@lists.reflextor.com > Date: Saturday, February 6, 2010, 11:18 AM > > > the lookup tool is being worked on. with the latest changes in > table layout > there have come up some extra issues. and you wouldnt want a > lookup tool > that crashes regularly. > > > _______________________________________________ > A51 mailing list > A51@lists.reflextor.com > http://lists.lists.reflextor.com/cgi-bin/mailman/listinfo/a51
Lookup tools have already been written. I have committed an early version to SVN : http://reflextor.com/trac/a51/browser/tinkering/A5Util - this code is only for 'educational' use. Lookup on the CPU is much to slow, so a GPU version will be needed. Please note that this tool uses a different compressed format for the tables, and may not be suitable for use with nvidia generated tables. (The unused bits in the starting points are at different locations) On the ATI front I have done some work with bypassing Brook+ runtime layer, and taken direct control of the buffers / stream management by using CAL. I was hoping to see some speed increase by avoiding some data copying, but that wasn't a real problem. Profiling shows that the ATI code leaves the GPU idle 3-7% of the time now. So there isn't much more to gain. What comes next is to write a better inter process communication system for table generation. The python wrapper seems to have problems with handling > 8000 messages a second, and processes just dies on me. After this i will have to write a better job management system, so the GPUs can handle smaller work loads. Currently the ATI code needs a full pipeline to operate. Ideally such a system would be able to shift jobs between GPUs so that when there is little work to do, it will coalesce these jobs to a single GPU. At this point I should be ready for writing GPU lookup code. Well this is my time line for GPU lookup. F _______________________________________________ A51 mailing list A51@lists.reflextor.com http://lists.lists.reflextor.com/cgi-bin/mailman/listinfo/a51
