how to get report about load/store and flop operations of source code

[SB17]

Hello

There is a need to understand how many load / store operations and how many flop contains the source code. Is it possible to get such statistics by the Intel compiler ?

[TimP · Accepted solution]

What's relatively easy to get is opt-report4 (for current releases; closest thing previously was opt-report6).  This will give you some of that for vectorized loops.

A compiler has no chance of getting you total counts which depend on adding up loop counts and intervening code.

As experts have said again recently, if you want that, your own assessment is probably best.  

There are some counters accessible by VTune et al. for certain CPU models, but that may not be what you're looking for.

Questions include do you count AVX2 fma as 1 or 2 flops, how do you count parallel and scalar simd instructions, .... for which you won't find much unanimity.

[SB17]

You answer create many right question :)

Turns out that static analysis based on compilator is too imprecise

Thanks for your reply