Just to muddy the waters, I'm exploring generalizing Phil's PAPI interface
to perfmon2 (found in the PAPI perfmon.c file) to use Perfmon event tables
wherever possible (including PAPI/perfctr) substrates. This will hopefully
normalize native event naming and allow all (or at least several) of us to
work on maintaining a common set of event descriptions.
I've got prototype code working on Opteron. More to come...
- dan

> -----Original Message-----
> From: [EMAIL PROTECTED] [mailto:perfmon-
> [EMAIL PROTECTED] On Behalf Of Philip J. Mucci
> Sent: Wednesday, January 10, 2007 7:12 AM
> To: [EMAIL PROTECTED]
> Cc: [EMAIL PROTECTED]; [EMAIL PROTECTED]
> Subject: Re: [perfmon] perfmon and OProfile
> 
> HI Stefane,
> 
> Couldn't all this be solved by having opcontrol/ophelp use the pfmlib
> event tables directory and ignoring everything in the events directory
> of oprofile?
> 
> Phil
> 
> On Wed, 2007-01-10 at 02:42 -0800, Stephane Eranian wrote:
> > Hello,
> >
> > On Mon, Jan 08, 2007 at 01:17:40PM -0800, Stephane Eranian wrote:
> > > This little exercise raised a few questions:
> > >
> > >  - the code that uses the v2.3 interface (for all architectures) now
> relies
> > >    on libpfm. You pass to pfm_dispatch_events() OProfile event names.
> So
> > >    that assumes that OProfile and libpfm events match. I am not sure
> this
> > >    is the case for all events on all x86 processors. We need to solve
> this.
> > >
> >
> > After taking a closer look, I came to the conclusion that we cannot use
> > libpfm to do the event assignment. There are several complictaed issues
> > related to:
> >     1/ Event names
> >
> >        OProfile and libpfm may not use the same exact name for the same
> event.
> >        Finding an event in libpfm is based upon string matching.
> >
> >     2/ Handling of unit masks
> >
> >        In OProfile unit masks are provide using an hexadecimal bitmask.
> >        In libpfm, they are provided using a string.
> >
> > Reconciling both of those would require a lot of work.
> >
> > An alternative is to let OProfile manage names and unit masks and have
> > the daemon simply manage the mapping of OProfile counter indexes to
> > perfmon counter indexes. Such mapping is trivial for many processors.
> > It is more difficult for the P4. This approach is fairly similar to what
> > we have today for OProfile on IA-64.
> >
> > --
> > -Stephane
> > _______________________________________________
> > perfmon mailing list
> > [email protected]
> > http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/
> 
> _______________________________________________
> perfmon mailing list
> [email protected]
> http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/

_______________________________________________
perfmon mailing list
[email protected]
http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/

Reply via email to