CUFFT Unable to Profile in Nsight Systems?

Developer Tools Nsight Systems Profiling Linux Targets
1

I see that there are options for the trace switch in nsys CLI for cudnn and cublas, but there is none for cufft. Is there a reason why this would be the case?

I have a call that looks like this

cufftExecD2Z(fftPlanHandle, (cufftDoubleReal *)&(*gpu_x)[0],
(cufftDoubleComplex *)&(*gpu_y)[0]);

When I do a nsys stats report on the gputrace, I see that this call seems to have been replaced by “regular_fft_r2c”

Is there a reason why the profiler is using the name of what I presume to be an internal to cufftExecD2Z function call as opposed to cufftExecD2Z? Could it also be related to why there is no cufft switch option for the --trace flag in “nsys profile”

2

Nsys does not yet formally trace cuFFT. We use interception to mark specific APIs that we want to mark on the timeline for cuDNN and cuBLAS, but not for the other cu* libraries.

It is on our roadmap.

What you are seeing in the stats output is information from the sample stack trace gathered by nsys using the linux perf functionality.