When I am profiling DNN workloads the total time mentioned along gpu__time_duration.sum column name, and the sum of all the elements inside this column don’t match in scale at all. Which time to consider as the actual runtime?
If you are having the issue that Metric Unit is switching between nanoseconds, microseconds, milliseconds, etc. then the command line option --print-unit=base will force gpu__time_duration.sum to use nanoseconds as the Metric Unit.
--print-units arg (=auto) Set scaling of metric units. Allowed values:
auto (default)
(Scale metrics to fitting magnitude)
base
(Show metrics with their base unit)