I want to getting CUPTI parameters from CUDA kernel for comparing Nsys output, because I am setting various parameters to same CUDA kernel.
Nsys CUPTI parameters outputs are CorrelationID, Graph Node ID, Stream etc.
Is there any method to getting those same parameters from CUDA kernel source code?
Thank you for your comments.
I want to get these information in both sides (Nsys and CUDA kernel code) for mapping cuda kernel behaviour.
Is it available?
One option would be to use CUpti_ActivityKernel9::gridId and access the gridId from the kernel using the ptx special register %gridid. I do not know if there is a CUDA intrinsic for gridid. You may have to use inline PTX. I also do not know gridid’s behavior for CUDA graph nodes. GridId has historically been used by tools for CDP trace and by the debugger akin to correlationId. gridid is owned by the driver. correlationId is owned by CUPTI.