Hi,
TensorRT engine generate NVTX events by default.
I know because I profiled the entire TensorRT application with Nsight CLI and I get all the NVTX events in the timeline and in the sqlite file ect…
Now I would like nsight to start profiling when NVTX Domain (named TensorRT) is invoked.
After reading the documentation I found the --capture-range option.
nsys profile --stats=true --trace=cuda,cudnn,nvtx,cublas,opengl,osrt --export=sqlite --capture-range=nvtx --stop-on-range-end=true --cudabacktrace=true -x true --output=my_report my_engine
But when I add this option I get the following error:
The application terminated before the collection started. No report was generated.
Like nsight couldn’t catch the NVTX events…
TensorRT Version: 7.2
CUDA Version: 11.1
Nsight Version: 11.1