I am profiling a certain kernel using NSight Compute (OpenCL version). The profiling is taking quite some time and at some point I get the message:
==WARNING== Launching the workload is taking more time than expected
and a suggestion to run the profiler with range replay rather than single kernel replay. Now, my question is not about how to avoid/overcome this per se, but rather what are the causes for this occurring. That is:
What exactly does NSight Compute “expect”?
How does NSight Compute set the “expectation”? What does it base it on?
Why would a plain-vanilla kernel - with nothing “extraordinary” in terms of instructions used etc. - go outside the range of these “expectations”?
Can this have something to do with the kernel being JIT-compiled rather having been compiled apriori?
Notes:
I get this even when using an ‘application’ which is merely a test harness and just builds and runs that kernel on inputs loaded from a binary file - without any other activity (gpu-kernel-runner).
If you only have information about the CUDA version of NSight Compute, I believe that’s a good enough answer, since NSight Compute OpenCL and CUDA seem almost identical.
This is unlikely to be due to the kernel execution itself taking an excessively long time. i.e. I’ve noticed this for kernels taking under 0.1 sec for example.