I am trying to profile vectorized_elementwise_kernel in GNMT application from the source code below using Nsight compute.
I got the following command after i profile the nsight sys.
ncu --kernel-name vectorized_elementwise_kernel --launch-skip 127032 --launch-count 1 python3 -u train.py --local_rank=0 --seed 2 --train-global-batch-size 1024
However, when i run ncu, it gives me this error, unrecognised option ‘–kernel-name’
Do you know to to successfully profile the kernel from pytorch code using nsight-compute?
Thanks!
The --kernel-name option was added in Nsight Compute version 2021.1. Looks like you are using an older version of Nsight Compute. You can use the --kernel-regex option.
Please refer the “Command Line Options” section of the Nsight Compute CLI document in your local Nsight Compute installation.
Thanks for your response! I am currently using Nsight Compute version of 2022.1.0. I switched to --kernel-regex option and it is working. :-) I will run the entire application through to the completion to check the output.
The --kernel-name option is supported in Nsight Compute version of 2022.1.0.
I switched to --kernel-regex option and it is working.
Did you switch when using the older version of Nsight Compute?
Sorry for the confusion. Right, i am using the older version of Nsight Compute. Version 2020.3.0.0. Now i understand why --kernel-name is not supported. Thanks!