Open teojgo opened 8 months ago
Thanks for the output. It seems to me that despite calling cudaDeviceSynchronize()
before LIKWID_NVMARKER_STOP
, the counts are not finished yet. In LIKWID_NVMARKER_CLOSE
, the library finally stops the counting (last lines of debugging output after nvmon_markerStopRegion
) and reads still 166 fadd instructions although the GPU should be inactive after the execution of LIKWID_NVMARKER_STOP
. I have to investigate what's going on there. Thanks for the test code.
I played around a little but no solution yet.
I played around a little but no solution yet.
Have you managed to reproduce the problem?
Yes but I couldn't find any problematic parts until now.
My holidays start tomorrow, so there will be no activity in the next month. I just wanted to let you know that it is not forgotten.
Describe the bug When running a simple CUDA vector addition kernel, the expected number of floating point additions is equal to the number of elements in the vector. Nvidia Nsight Compute gives the correct result for the same metric.
To Reproduce The following is the
vectorAdd.cu
source code (without error checking to simplify it):The above can be easily compiled using
nvcc
as follows:And then executed with:
likwid-perfctr -W FLOPS_SP -m ./vectorAdd
The output is giving:
Running using Nvidia Nsight Compute:
To Reproduce with a LIKWID command