Counting Flops for a code!

Question

This is really taking my time. I could not find a simple way to estimate FLOPS for a following code (the loop), How much FLOPS are for a single iteration of the loop:

float func(float * atominfo, float energygridItem, int xindex, int yindex)
{
   ...
   for (atomid=0; atomid<numatoms*4; atomid+=4) 
   {
       float dy = coory - atominfo[atomid+2];
       float dysqpdzsq = (dy * dy) + atominfo[atomid+3];
       float dx1 = coorx1 - atominfo[atomid+1];

       float s, y, t;
       s = atominfo[atomid] * (1.0f / sqrtf(dx1*dx1 + dysqpdzsq));
       y = s - energycomp1;
       t = energyvalx1 + y;
       energycomp1 = (t - energyvalx1)  - y;
       energyvalx1 = t;
    }
    ...
}

It looks simple but I got confused with some other numbers given earlier, so it would be great if someone can give an exact number.

Thanks.

Answer 1

I see (in order of increasing complexity):

• 8 additions (inc. subtractions)
• 3 multiplications
• 1 reciprocal-square-root

How these relate to each other depends strongly on the CPU family.

Answer 2

Try to either take intermediate assembly code or decompile exe.

Then count all floating point operations (in x86 assembly code they start with F prefix like FSIN).

Answer 3

I count 12 plus a sqrt (which is likely using Newton's method, which is a loop), but that depends on the data types of some variables that you did not specify, and the result of compilation (which may add more, or optimize out some operations).

I am counting each +, /, -, or * where the expression contains at least one floating point variable, so array indices and the loop invariant do not count, and those are integer operations.

Answer 4

Try using a performance measurement library like PAPI, they give abstractions to hardware counters which would be your best best to measure the FLOPS. PAPI_FLOPS.