I'm trying to code a kernel profiling on my source code, and it needs to profile multiple metrics on CUDA kernels. I've tried PAPI with the CUDA component, but it's doesn't worked out. I also noticed the NVidia CUPTI API, but the sample only profiles one metric per kernel and I'm a bit confused on how to change the sample source code to perform under multiple events.
PS: the nvprof tool doesn't help, because the profiling must happen on my source code.
I've faced this same problem - I suspect you are working with CUDA on Linux? I know the tools on Windows are better supported than Linux, which is disappointing.
But, here's a work-around that we found to be quite useful in addition to the tools that NVIDIA does support on Linux (the one we had hope for was NV Perfkit, but it never worked correctly for us). So, instead we used Systemtap (https://sourceware.org/systemtap/ ), which is a general tool for profiling the Linux kernel, drivers (even third party BLOBs), and your kernel modules. It does also support user space profiling, but we did all our work with kernel modules.
It would be better if NVIDIA would just support Linux the same as Windows, but in the mean time, Linux has plenty of tools like systemtap that can help. Hope that this tip helps you.
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts