Measuring FLOPs of an application with the linux perf tool

Question

I want to measure the ammount of floating point and arithmetic operations executed by some application with 'perf', the new command line interface command to the linux performance counter subsystem. (For testing purposes I use a simple dummy app which I created, see below).

Because I could not find any 'perf' events defined for measuring FP and integer operations, I started digging in the raw hardware event codes (to be used with -rNNN, where NNN is hexadecimal value of the event code). So my real problem is that, the codes I found for retired instructions (INST_RETIRED) do not make the distinction between FP and other instructions (X87 and MMX/SSE). When I tried to use the appropriate umasks to the particular code I found out that somehow 'perf' does not understand or support the umask inclusion. i tried with:

% perf stat -e rC0 ./a.out

which gives me the instructions retired, but

% perf stat -e rC002 ./a.out 

which should give me the X87 instructions executed says I supplied wrong parameters. Maybe so, but what is the correct way to use umasks of raw hardware events with 'perf'? in general what is the way to get the exact number of floating point and integer operations a program executed using the perf tool?

Many thanks, Konstantin Boyanov

---

Here is my test app:

int main(void){
  float  numbers[1000];
  float res1;
  double doubles[1000];
  double res2;

  int i,j=3,k=42;

  for(i=0;i<1000;i++){
    numbers[i] = (i+k)*j;
    doubles[i] = (i+j)*k;
    res1 = numbers[i]/(float)k;
    res2 = doubles[i]/(float)j;
  }
}

Answer 1

The event to use depends on the processor. You can use libpfm4 (http://perfmon2.git.sourceforge.net/git/gitweb-index.cgi) to determine which are the available events (using the showevinfo program) and then check_events from the same distribution to figure out the raw codes for the event. My Sandy Bridge CPU supports the FP_COMP_OPS_EXE event which I have empirically found corresponds closely to the FLOP count.

Answer 2

I'm not sure about perf, but oprofile has floating point events for many processors. There may be some overlap, as INST_RETIRED is a valid oprofile event too.