Xeon phi, offload mode, cache miss analysis

[Bobo_S_]

I am using Xeon phi offload. I want to analyze the cache miss on MIC. How can I use amplex-cl to do that? 

I am using VTune 2016.

I tried 

amplxe-cl -target-system=mic-host-launch -collect advanced-hotspots

amplxe-cl -report hw-events 

But didn't get what I want. I am a little confused. 

Thanks.

[Peter_W_Intel · Accepted solution]

Analysis type "advanced-hotspots" only provides CPU time (clockticks) and Instruction Retires, you may use analysis type "general-exploration" which includes performance data on front-end and back-end (cache miss involved). For example:

>amplxe-cl -target-system=mic-host-launch -collect general-exploration -search-dir=path-of-bin-and-src -- app

[Bobo_S_]

Hi Peter Wang,

Thank you.

[Dmitry_P_Intel1]

Hello,

It is worth to look at the article to understand the metrics in general exploration on memory usage efficiency. Since L2 misses on KNC includes not only demand reads but also different types of prefetching we calculate and show L1 Hit rate and Estimated latency impact of L1 misses that being higher a particular threshold can indicate that L1 miss was not served with L2.

Thanks & Regards, Dmitry