Intel PCM - measure time using Performance Counters

[pranith]

Hi,

I am trying to time my piece of code using performance counters.

I see that there we cannot use the cycle count from the cores due to turbo boost and various other frequency variations in the CPU.

For this I think I should use the Invariant TSC value provided by "getInvariantTSC()". But, there is no corresponding Invariant TSC frequency value which I need to calculate the time.

How do I get this frequency? Or should I use QueryPerformanceCounter/Frequency() on Windows?

Thanks,

Pranith

[Patrick_F_Intel1]

Hello Pranith,
The answerdepends on what accuracy you need and what you mean by 'time my piece of code'.
In my experience, the TSC can provide nanosecond resolution with little overhead.
QueryPerformanceCounter can provide about microsecond resolution with more overhead.
You can look http://stackoverflow.com/questions/1739259/how-to-use-queryperformancecounter for an intro to QueryPerformanceCounter().
I can't find a reference to getInvariantTSC().
If you have an invariant TSC (see http://software.intel.com/en-us/forums/showthread.php?t=101951) then the TSC runs at the same rate regardless of Cstate, Pstate, turbo mode, etc.
Pat

[pranith]

By 'time my piece of code', I meant:

main()

{

warmup();

startCounters(start);

stopCounters(end);

timeToExecute = f(start, end);

}

I was referring to cpucounters.h in Intel PCM for getInvariantTSC(). That gives us the invariant clock ticks. But to get time from those ticks, I need the frequency at which this clock ticks.

I wanted to use performance counters to do this, but if I cannot get the invariant clock frequency, then I would use QueryPerformanceCounter/Frequency to time my code.

[Patrick_F_Intel1 · Accepted solution]

You can get the TSC frequency from:
uint64 PCM::getNominalFrequency()

You can get the TSC with the Windows*intrinsic:
#include
unsigned __int64 __rdtsc(void);

You can check if the TSC is invariant with the above link (see http://software.intel.com/en-us/forums/showthread.php?t=101951)

Pat

[alexbhatti]

Hi Pranith,

It's true in that the concept of core cycles aren't able to be used in a reliable manner anymore, since the Turbo Boost or power states alter the frequency. The TSC Invariant was developed to address this issue through ticking at a set rate, regardless of CPU frequency However, the frequency of ticking isn't available through an API.

The two most commonly used methods to deal with this are:

CPUID 0x15/0x16 For newer Intel processors, they will give users the crystal clock as well as the multiplier, allowing you to determine TSC frequency. TSC frequency.

Calibration with QueryPerformanceCounter - On Windows, you can record both __rdtsc() and QueryPerformanceCounter values over a known time interval (for example 100 ms) and derive the TSC frequency once at startup.

If you're just looking for wall clock timing, using QPC is the most straightforward option since in modern Windows it's already supported by the standard TSC.

If you're switching between minutes, seconds and hours when doing benchmarking, it's a good idea to have it's a good idea to have a quick reference, such as an hour calculator

It can help you check your calculations when you convert TSC ticks into human-readable time.

This method provides stable and repeatable measurements across both power and core states.
Best Regards