[Feature]: controlling the profiling scope

[maartenarnst]

Suggestion Description

The documentation for rocprofiler-v1 indicates that the profiling scope can be controlled by "dispatch range" and "kernel name". Here, the "dispatch range" is an index or a range of indices of dispatch ranges. However, when using roctx markup, the ranges are labelled with strings by their names, and not with numbers by indices. Indeed, roctxRangeStart takes the name as input, and it returns the index, but this index is typically an internal number in the code, not accessible to the user.

Thus, a useful feature may be to provide the ability to control the profiling scope by passing the name of the roctx range.

The feature would be similar to --nvtx-include and --nvtx-exclude that can be passed to ncu on Cuda.

A related question is how the profiling scope should/can be controlled in rocprofiler-v2.

Operating System

No response

GPU

No response

ROCm Component

No response