Here is some information on intel performance counters from
Sameer Shende (Tau person):
> Yes, here's the link:
>http://www.cs.utk.edu/~ghenry/distrib/archive.htm#monitor
> Also, there is another package called PCL that I am
>trying to integrate with TAU. It gives a common API for access to
>counters on multiple platforms. Check out :
>http://www.fz-juelich.de/zam/PT/ReDec/SoftTools/PCL/PCL.html
Also, here is a suggestion for a simple templates-optimization
performance test: using some e.t. implementation (blitz/pooma/whatever),
do A+B+C+D, with the length of these arrays all zero (but set the
length to zero where the optimizer won't see it). Compare the
timing of this to the equivalent C code on arrays of zero length
(also would have to hide the zero length from the optimizer).
This would give you a handle on how well the compiler is optimizing
away expression templates overhead.
Another good test might be something like A+B+C+D on TinyVectors
or their equivalent, and compare to hand-unrolled C code.
Best of luck,
Todd
>
> Hello,
>
> As most of you probably realized, my post from yesterday was intended to be
> a private communication with Todd. My apologies to all the list recipients.
> This was much more embarrassing than realizing how much more glaring my
> errors had become in the version of my double precision code sent back to me
> from the Blitz subscription list than in the copy I first submitted.
>
> I am pressing ahead with my proposed extensions to the Stepanov abstraction
> suite. I would like this suite to serve as a broadly useful tool for
> evaluating the relative strengths of different compilers to handle
> expression template libraries.
>
> I will be keeping notes about my progress on my web site www.stokes.ca
> Suggestions are welcome.
>
> I have also received a degree of interest from several groups about working
> together on a common performance-assessment framework. Steve Karmesin from
> the Pooma group offered these comments:
>
> <<<
> I understand the desire to have such a system. POOMA (and other efforts in
> the Advanced Computing Lab here at LANL) is working with the TAU effort
> based at the University of Oregon in Eugene. Like everyone else, we want to
> be able to profile and trace our stuff, and the compiler supplied tools are
> inadequate at best.
> >>>
>
> He then goes on to say that even TAU currently has problems coping with the
> sophistication of expression template code.
>
> The first step is to open the discussion to see how a performance-assessment
> library operating at the C++ level could best complement tool-based
> approaches such as TAU.
>
> I will be surveying a variety of expression template projects to see what
> most needs to be done. Once again, comments and requests are encouraged.
>
> Allan
>
> --------------------- blitz-dev list --------------------------------
> * To subscribe/unsubscribe: mail to majordomo@oonumerics.org, with
> "subscribe blitz-dev" or "unsubscribe blitz-dev" in the body of the message
> * Blitz++ web page: http://oonumerics.org/blitz/
>
>
--------------------- blitz-dev list --------------------------------
* To subscribe/unsubscribe: mail to majordomo@oonumerics.org, with
"subscribe blitz-dev" or "unsubscribe blitz-dev" in the body of the message
* Blitz++ web page: http://oonumerics.org/blitz/
This archive was generated by hypermail 2b29 : Wed Feb 20 2002 - 04:30:08 EST