Can we use nsight compute / system (in command line) to get the sampled time or executed instructions (%) information inside a kernel?

As the figure shows, we use ncu to get the warp sampled percentage values in the source counter page.

Can we use command line to directly extract the values to a txt/csv file instead of only using GUI?

Is there a reasonable method to measure the device function time (performance) inside a global kernel?

The “Instructions Executed” column maps to the inst_executed metric. You can use CLI options like --page source to print Source page content on the command line. You can combine this with other options, like --metrics inst_executed or --csv . You can import an existing report with -i. You can familiarize yourself with all options here.

You can also use the Python Report Interfaceto parse the report from Python and access these metrics. Examples are on github.

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.