Average value of metric in multiple kernel invocations

I would like to know if there any option or capability in Nsight Compute to calculate the average value of a metric over multiple kernel invocations? For example, Assume that a kernel regex name is profiled and in there are 100 invocations and the metric is IPC. So, I see multiple IPC values such as 1.3, 1.23, 1.45, … and I want to get an average value. I know that nvprof shows such value.

In Nsight Compute, I import the file and then choose the RAW view and then export as CSV file. I then have to manually open the CSV file and do some Excel works. For large number of files and metrics, this task becomes cumbersome. Any suggestion for that?

You can use the Nsight Compute –print-summary per-kernel command line option to get the summary (minimum, maximum and average metric values) for each kernel type across all the kernel invocations.

1 Like