Detecting wrong version of nvprof when remote debugging with nvvp

Cuda version 10.2, running on Nvidia AGX Pegasus. DriveOS 5.2.3 for QNX.

When attempting to do a remote debug session nvvp detects v0.0 for nvprof, even though it has the correct version installed on the board:

image

I tried setting the environment variables for the nvvp session but doesn’t make a difference, is there any place where I can see a more detailed log?

Best,
Flavio.

Hi,

Can you please send below details to investigate the problem.

  1. Run below command on your target (QNX in your case) and send the output

/usr/local/cuda/bin/nvprof --query-cuda-info; /usr/local/cuda/bin/nvprof --version

  1. Run below command on your host (where you launched nvvp) and send the output

/usr/local/cuda/bin/nvprof --version

  1. Do you have multiple cuda toolkit version installed on target?
  2. If yes, make sure to selected correct toolkit location in “Create New Session” dialog, For example see below screenshot.

image

Hi, I have checked the versions and they are the same:

1 - Target (QNX):

# /usr/local/cuda/bin/nvprof --query-cuda-info 
0/(Xavier cc=7.2)
22740994/100663390,22740994/100663391,22740994/100663385,22740994/100663386,22740994/100663345,22740994/100663387,22740994/100663388,22740994/100663405,22740994/100663406,22740994/100663407,22740994/10066,
20971521,20971522,20971523,20971524,20971525,20971526,20971527,20971528,20971529,20971530,20971531,20971532,20971533,20971534,20971535,20971536,20971537,20971538,20971539,20971540,20971541,20971544,209715,

# /usr/local/cuda/bin/nvprof --version         
nvprof: NVIDIA (R) Cuda command line profiler
Copyright (c) 2012 - 2020 NVIDIA Corporation
Release version 10.2.244 (21)

2 - Host

$ /usr/local/cuda/bin/nvprof --version
nvprof: NVIDIA (R) Cuda command line profiler
Copyright (c) 2012 - 2020 NVIDIA Corporation
Release version 10.2.244 (21)

3 - Just one cuda toolkit and the path is correct in the session