VTune : any way to only profile specific application functions/code sections using HPC Perf. Charac?

[mikeitexpert]

In my app, I have an initialization function running at beginning of the execution which I don't need to be profiled using HPC Performance Characterization analysis (in VTune). Also there are some other functions in the main body of the application which I want to exclude from data sampling and execution. 

Please let me know how I can possibly mark them up to be excluded from data sampling and analysis.

Many thanks.

[Bernard · Accepted solution]

Please explore so called "ITT API".

You can insert markers into your code and narrow down the sampling to those specific regions (i.e function scope, base-block scope, loop scope). Be aware, that at the micro-architectural level does not exist binding between specfic block of machine code to the PMCs (performance counters).

Such a binding exists only for thread 0 and thread 1 (SMT) and for kernel/user mode sampling. In my opinion where the narrowing scope markers are used the VTune would simply filter out the samples belonging to other code blocks.

https://software.intel.com/content/www/us/en/develop/documentation/vtune-help/top/api-support/instrumentation-and-tracing-technology-apis.html

[Dmitry_P_Intel1]

Hello,

If you have an MPI application you can also use standard MPI_Pcontrol function to pause/resume collection specifying 0 or 1 as the function parameter value.

Thanks & Regards, Dmitry

[JananiC_Intel]

Hi,

Thanks for accepting the solution. If you need any additional information, please submit a new question as this thread will no longer be monitored.

Regards,

Janani Chandran

[Bernard]

>>>Also there are some other functions in the main body of the application which I want to exclude from data sampling and execution. >>>

If you would like to achieve a "total" exclusion of some basic-blocks from profiling, then you may switch to counting mode and use ITT API to mark the regions some specific regions (i.e. basic blocks).

Probably there will be a some overhead of the call to the ITT API markers so the profiled region (by the counting mode) shall have a longer cycle count than  ITT API markers.