Understanding the non-cudf kernels in Nsight System Profling results

[AsaMyth]

Hi,
I used Nsight sys to profile the second query in https://github.com/NVIDIA/spark-rapids-examples/blob/main/examples/SQL%2BDF-Examples/micro-benchmarks/notebooks/micro-benchmarks-gpu.ipynb

In the result I'm seeing some non-cuda kernels, especially the DeviceMergeSort and static_kernel, their function name doesn't has the cuda perfix, so I have no idea how to find the source of them. If I zommed into the desription, I can see some cudf name though the whole description doesn't quite make sense as shown in the second screenshot.
what are they used for? and where can I find the source of them? Thanks.

[jlowe]

DeviceMergeSort and static_kernel typically groups kernels associated with CUB, whereas _kernel_agent typically groups kernels from Thrust. CUB and Thrust are tools leveraged by libcudf and spark-rapids-jni. See https://nvidia.github.io/cccl/cub/index.html and https://nvidia.github.io/cccl/thrust/, respectively. CUB and Thrust are part of CUDA Core Compute Libraries (CCCL), see https://github.com/NVIDIA/cccl.

[AsaMyth]

Thanks for the clear and prompt anwser!