I feel bad asking such a simple question.
How can we find the conversion factor, or the sampling interval, in order to convert the delays into real time ?
In the current and upcoming releases, this is not provided. However, in an upcoming version we plan to allow to set the sampling interval and also report the sampling rate. Note is that even with that data there will be no conversion from number of samples to actual time taken. A sampling profiler cannot capture this information - not on the GPU nor on a CPU. It’s always just a sample count. The conversion factor is only useful to double check the overall expected number of samples.
Does the latest ncu have the ability to show some estimate of time or cycles spent on each line of code? For example, if I collect a cycle metric (e.g. sm__cycles_active.avg) and set a fixed sample frequency, then I should be able to estimate time spent on each line.
Dan