search

Gwinel / likwid

Automatically exported from code.google.com/p/likwid
GNU General Public License v3.0
0 stars 0 forks source link

Counting AVX #95

Closed GoogleCodeExporter closed 9 years ago

GoogleCodeExporter commented 9 years ago 
In the likwid homepage in Source tab(see below) in the FLOPS_AVX.txt file you 
mentioned: " Please note that the current flop measurements on SandyBridge are 
potentially wrong..."

Would you please let us know what the precise meaning of "wrong" is? Is that 
because they do not count Integer operations? and if a program does have only 
floating points vector operations, would the count be accurate?

Thank you for your time answering our question.

link for the FLOPS_AVX.txt : 
http://code.google.com/p/likwid/source/browse/groups/sandybridge/FLOPS_AVX.txt?r
=7bf5963556aa43865509044df985c23f1446f346

Original issue reported on code.google.com by L.aghaba...@gmail.com on 1 Mar 2013 at 9:24

GoogleCodeExporter commented 9 years ago 
Hi, the flop counts overcount, if the data is loaded from a cache level 
different from the L1 cache. As this is usually true the event is overcounted. 
See a blog entry on this here:

http://likwid-tools.blogspot.de/2012/02/intel-sandybridge-and-counting-flops.htm
l

This is a bug in the performance counter implementation on the chip.
This was not fixed in IvyBridge. In Haswell the flop events are still missing. 
Still they promised to fix this with Haswell.

Original comment by jan.trei...@gmail.com on 8 May 2013 at 8:17