nvidia gpu

[Giacomo Castiglioni]

Hi,

I am trying to use likwid 5.2.1 (commit: 233ab943543480cd46058b34616c174198ba0459) with a Tesla P100-PCIE-16GB without success.

I have a simple test program (attached) that runs fine without activating the markers, otherwise I get the following error.

likwid-perfctr -G 0 -W FLOPS_DP -m ./daxpy.x
--------------------------------------------------------------------------------
CPU name:   Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz
CPU type:   Intel Xeon Haswell EN/EP/EX processor
CPU clock:  2.60 GHz
--------------------------------------------------------------------------------
Error: function (*cudaFreePtr)(NULL) failed with error 46.
ERROR - [./src/nvmon.c:nvmon_addEventSet:468] Bad address.
Failed to add event set for GPU 0
Error setting up GPU Marker API.
Error starting up GPU Marker API.
terminate called after throwing an instance of 'std::runtime_error'
  what():  cudaAssert: all CUDA-capable devices are busy or unavailable: daxpy.cu:49

--------------------------------------------------------------------------------
GPU Marker API result file does not exist. This may happen if the application has not called LIKWID_GPUMARKER_CLOSE.


Any clue about what is going wrong?

Thank you in advance for your help, let me know if you need additional info.

Cheers,
Giacomo

[Thomas Gruber]

Hi,

sorry for the late answer.

I can run your code without any problems (with various GPUs). Which CUDA version are you using? Driver/Library mismatch?

According to https://developer.download.nvidia.cn/compute/DevZone/docs/html/C/doc/html/group__CUDART__MEMORY_gb17fef862d4d1fefb9dba35bd62a187e.html, there are only two possible errors for cudaFree:

- cudaErrorInvalidDevicePointer
This indicates that at least one device pointer passed to the API call is not a valid device pointer.

- cudaErrorInitializationError 
The API call failed because the CUDA driver and runtime could not be initialized

Please run your code with -V 3 and supply the output

Best,
Thomas