cache miss and floating point operation

[Fabrizio_A_]

Salve a tutti,

sono interessato a valutare i cache miss e il numero di operazioni floating point con riferimento ad un programma MPI.

Da riga di comando ho provato:

mpirun -n 4 -l amplxe-cl -result-dir XXX -collect hotspots eseguibile

Il tipo di analisi "HOTSPOTS" è sufficiente per ottenere quanto desiderato?

Se sì. una volta aperti i risultati usando l'interfaccia grafica (amplxe-gui), dove sono reperibili le informazioni di cui sopra?

Se no, che comando è necessario utilizzare?

Grazie a tutti in anticipo!

[Bernard]

Can you use Google translate?

[David_A_Intel1]

Hi Fabrizio!

Unfortunately, it is not that easy.  Modern processors have many ways of performing "float-point operations."  Therefore, I refer you to this article. 

[Bernard]

>>>FP_COMP_OPS_EXE.X87>>>

This option refers  only to x87 floating-point code. Modern compilers usually emit SSE/AVX code for target architecture which is newer that Pentium 3.

[Bernard]

What CPU do you have?

[Bernard]

Then you should look for SSE/AVX generated code unless you have some old library compiled into x87 code.

[Peter_W_Intel]

Command line is OK. I don't know what processor you used. You may use "amplxe-runss - event-list | grep FP_" to find them, which are supported by your processor.

About using events to measure cache miss, please read right tuning guides from this.

[Bernard]

>>>So, you need to know what processor you work,>>>

OP is using  Intel Core i7 3610QM.

[Peter_W_Intel]

@iliyapolak

It's hard to say that PMU event name is strictly same as doc described, that was why I provided the way to list supported events...you can find similar one (you want) from the list...for example, use key words "x87", "SSE_DOUBLE",etc to find

[Bernard]

Ok I understand it.