为什么perf显示的浮点事件少于预期的？

Question

I'm writing a C program that does a 50x50 matrix multiplication.

I compiled it to assembly and I see it has at least 2 floating point instructions (mulsd and addsd) in the most inner loop. The loop is executed 125000 times, so I was expecting perf stat to give me at least 250000 floating point operations counted.

However when I run the following command,

perf stat -e r530110 -e r531010 -e r532010 -e r534010 -e r538010 ./matmul

I only get the following counts

           448 r530110                                                     
             0 r531010                                                     
             0 r532010                                                     
             0 r534010                                                     
 <not counted> r538010                 

   0.001082287 seconds time elapsed

My compiler does not produce SSE instructions so the other zero counts are expected. However, I only got 448 floating point operations.

What is happening?

Answer 1

I've figured it out. I misunderstood: mulsd and addsd are actually SSE instructions.

The stat r530110 (FP_COMPS_OPS_EXE:X87) does not count mulsd or addsd .

I needed to look at r538010 (FP_COMPS_OPS:SSE_SCALAR_DOUBLE).

My guess for the reason those were <not counted> in my program was because the program finished too quickly. When I increased the size of the matrix (so the program runs longer), the floating point event count came out correct.