| Author |
Message |
Grigori Fursin
Guest
|
 Posted:
Thu Jul 21, 2005 2:18 pm Post subject:
question about FP_INS counter event on Athlon XP |
|
|
Hello all,
Just a quick question for those who are familiar with hardware counters
for Athlon XP. I am using the latest PAPI library for Linux and trying to
move
my software from P4 to Athlon XP and have a problem with the
event PAPI_FP_INS (retired x87 uops). Does anyone know if this event
actually exist on Athlon XP or it's my software problem?
Also, when I print all possible hardware events seen by PAPI,
I get more than 100 events on P4 while only 34 on Athlon.
Again, is it reasonable or it's a problem with my kernel configuration?
Thank you in advance,
Grigori
=============================================
Grigori Fursin, PhD, Research Associate
INRIA Futurs (France) and
ICSA, University of Edinburgh (United Kingdom)
http://homepages.inf.ed.ac.uk/gfursin
============================================= |
|
| Back to top |
|
 |
Mikael Pettersson
Guest
|
| Posted:
Thu Jul 21, 2005 2:54 pm Post subject:
Re: question about FP_INS counter event on Athlon XP |
|
|
In article <5MJDe.15621$hV3.5859@nntpserver.swip.net>,
Grigori Fursin <gfursin@SPAMMMMTRAPyahoo.co.uk> wrote:
| Quote: | Hello all,
Just a quick question for those who are familiar with hardware counters
for Athlon XP. I am using the latest PAPI library for Linux and trying to
move
my software from P4 to Athlon XP and have a problem with the
event PAPI_FP_INS (retired x87 uops). Does anyone know if this event
actually exist on Athlon XP or it's my software problem?
Also, when I print all possible hardware events seen by PAPI,
I get more than 100 events on P4 while only 34 on Athlon.
Again, is it reasonable or it's a problem with my kernel configuration?
|
The K7 cannot count FP ops, that ability came with the K8.
--
Mikael Pettersson (mikpe@csd.uu.se)
Computing Science Department, Uppsala University |
|
| Back to top |
|
|
Grumble
Guest
|
| Posted:
Thu Jul 21, 2005 3:15 pm Post subject:
Re: question about FP_INS counter event on Athlon XP |
|
|
Grigori Fursin wrote:
| Quote: | Just a quick question for those who are familiar with hardware counters
for Athlon XP. I am using the latest PAPI library for Linux and trying to
move my software from P4 to Athlon XP and have a problem with the
event PAPI_FP_INS (retired x87 uops). Does anyone know if this event
actually exist on Athlon XP or it's my software problem?
|
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22007.pdf
Appendix D
Retired instructions (includes exceptions, interrupts, resyncs)
Retired Ops
Retired branches (conditional, unconditional, exceptions, interrupts)
Retired branches mispredicted
Retired taken branches
Retired taken branches mispredicted
Retired far control transfers
Retired resync branches (only non-control transfer branches counted)
No "retired x87 uops" counter in K7.
| Quote: | Also, when I print all possible hardware events seen by PAPI,
I get more than 100 events on P4 while only 34 on Athlon.
Again, is it reasonable or it's a problem with my kernel configuration?
|
K7 only supports a small number of events.
<quote>
The Intel Pentium 4 processor performance metrics are a set of
quantities that are useful for tuning software performance when running
applications on the Pentium 4 and Intel Xeon processors. The metrics
are derived from the Pentium 4 and Intel Xeon processor performance
monitoring events, which are described in Chapter 15 and Appendix A
of the IA-32 Intel Architecture Software Developer's Manual, Volume 3.
</quote>
ftp://download.intel.com/design/Pentium4/manuals/25366816.pdf
More than 100 events sounds about right for NetBurst.
It seems to me oprofile only supports a small subset of all P4 events?
--
Regards, Grumble |
|
| Back to top |
|
|
Grigori Fursin
Guest
|
| Posted:
Thu Jul 21, 2005 4:03 pm Post subject:
Re: question about FP_INS counter event on Athlon XP |
|
|
Thanks everyone for quick response.
I somehow thought that Athlon XP had K8 architecture,
so now I see why my software doesn't work...
It's a shame that Athlon doesn't have FP counts -
it was a useful event for tuning numerical codes on P4 ...
Cheers,
Grigori |
|
| Back to top |
|
|
Anton Ertl
Guest
|
| Posted:
Fri Jul 22, 2005 12:15 am Post subject:
Re: question about FP_INS counter event on Athlon XP |
|
|
Grumble <devnull@kma.eu.org> writes:
| Quote: | Grigori Fursin wrote:
Also, when I print all possible hardware events seen by PAPI,
I get more than 100 events on P4 while only 34 on Athlon.
Again, is it reasonable or it's a problem with my kernel configuration?
K7 only supports a small number of events.
|
Well, if you look at all the sensible combinations of the various
fields in the event descriptor, you probably get several hundred
different events on the K7, but PAPI does not support all of them.
Anyway, the Pentium 4 supports more basic event types and additional
features (like tagging), so it's not surprising that you see more
events in PAPI.
If you want to use the full power of the performance counters, you
probably have to use the perfctr stuff (perfex or the library)
directly, but this is much harder to use than the PAPI stuff,
especially on the Pentium 4.
- anton
--
M. Anton Ertl Some things have to be seen to be believed
anton@mips.complang.tuwien.ac.at Most things have to be believed to be seen
http://www.complang.tuwien.ac.at/anton/home.html |
|
| Back to top |
|
|
Grigori Fursin
Guest
|
| Posted:
Fri Jul 22, 2005 4:15 pm Post subject:
Re: question about FP_INS counter event on Athlon XP |
|
|
I am potentially thinking of using perfctr in the future
but so far found PAPI very useful and easy to use
for quick experiments.
Thank you for your answer,
Grigori
"Anton Ertl" <anton@mips.complang.tuwien.ac.at> wrote in message
news:2005Jul21.234204@mips.complang.tuwien.ac.at...
| Quote: | Grumble <devnull@kma.eu.org> writes:
Grigori Fursin wrote:
Also, when I print all possible hardware events seen by PAPI,
I get more than 100 events on P4 while only 34 on Athlon.
Again, is it reasonable or it's a problem with my kernel configuration?
K7 only supports a small number of events.
Well, if you look at all the sensible combinations of the various
fields in the event descriptor, you probably get several hundred
different events on the K7, but PAPI does not support all of them.
Anyway, the Pentium 4 supports more basic event types and additional
features (like tagging), so it's not surprising that you see more
events in PAPI.
If you want to use the full power of the performance counters, you
probably have to use the perfctr stuff (perfex or the library)
directly, but this is much harder to use than the PAPI stuff,
especially on the Pentium 4.
- anton
--
M. Anton Ertl Some things have to be seen to be believed
anton@mips.complang.tuwien.ac.at Most things have to be believed to be seen
http://www.complang.tuwien.ac.at/anton/home.html |
|
|
| Back to top |
|
|
|
|
|
|
FAQ
Memberlist
Register
Profile
Log in to check your private messages
Log in