Counting FLOPs on Intel Sandy Bridge

[wzhang80]

Hi,

I am a new user of PAPI. I am now trying to use PAPI to measure the FLOPs of my code. Unfortunately,
all of my available resources are based on Intel Sandy Bridge.

What I learn from here (http://icl.cs.utk.edu/projects/papi/wik ... SandyFlops) is that the
FLOPs counts of PAPI on Intel Sandy Bridge have serious overcounting problem and thus I cannot use the
FLOPs counts and rates of PAPI as an absolute measure of the amount of work done.

My question is will this FLOPs overcounting problem on Intel Sandy Bridge be fixed in future versions of PAPI?
Is there an approximate date? I think this problem is still not solved in PAPI 5.3.

Or is there any other way to accurately measure the FLOPs on Intel Sandy Bridge? Now I can only rely on manually
calculated results:-(

Many thanks! I truly appreciate your reply.

Best regards,
Wentao

[xphi512]

I think the only way to measure FLOPs/cycle on a CPU is to calculate the FLOPs needed for your algorithm by hand - that shouldn't be a problem at all. Then you just have to count the CPU Cycles with PAPI_TOT_CYC needed for your algorithm.

Thereby you don't get problems with counting FLOPs which are computed by the CPU, but may not be needed for your algorithm, e.g. using AVX register with padding.