« Return to Thread: 3.9.6

Re: 3.9.6

by Clint Whaley :: Rate this Message:

Reply to Author | View in Thread

Michael,

> 1) I have to build ATLAS 3.8.2 on an Core i7 and as mentioned above
>you have your own tuning setting in 3.9.6, so how much of a bad idea
>would it be to use the Core2 tuning info for 3.8.2 for the i7 core?
>Right now we get an unknown arch and have to do a full tune which is
>kind of annoying if you build ATLAS more than once a week on that CPU.

My *guess* is that your setting arch to Core2 would get you a fine library
using 3.9.6.  Unfortunately, 3.8.x *sucks* for the Core2, because 2-D register
blocking is critical on that machine.  I should have backported the 2-D kernel
to the 3.8 series when I wrote it (using a template provided by Yevgen
Voronenko), but at the time I thought I would be releasing 3.10 in only a few
months.  Unfortunately (in some respects), our research in threading was wildly
successful, which caused me to rewrite the threading subsystem, and then we
were getting good results by tuning lapack, etc., and I got assigned to a bunch
of committees, so that the 3.10 series is *still* not out.

I really should backport the 2-D kernels to 3.8 and issue a bug fix 3.8.3,
but I presently am so swamped that I have had almost no time for ATLAS work,
and what time I had I had to work on new stuff so that I don't block my
student's research . . .

> 2) A while back I asked about how to treat the Atom (at least the 64
>bit models with SSE3 and all that good stuff) and I did not recall
>that anyone answered. Somebody else did ask me off list a couple days
>about that, so it would be nice to get an answer. My instinct is to
>make everything being treated as a Core2 for now.

I would also guess that.  However, for all the questions, I recommend taking
the old pepsi challenge: install ATLAS one time w/o arch defs, and do
"make time".  Now install it where you force the arch equal to Core2
(the -A flag to configure; see xprint_enums for values) and use the defaults
and do "make time".  You can have xatl_bench directly compare the two
timings, as discussed in the install guide.

Cheers,
Clint

**************************************************************************
** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
**************************************************************************

------------------------------------------------------------------------------
Create and Deploy Rich Internet Apps outside the browser with Adobe(R)AIR(TM)
software. With Adobe AIR, Ajax developers can use existing skills and code to
build responsive, highly engaging applications that combine the power of local
resources and data with the reach of the web. Download the Adobe AIR SDK and
Ajax docs to start building applications today-http://p.sf.net/sfu/adobe-com
_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel

 « Return to Thread: 3.9.6