Could you please explain what does this error means:
========= Program hit error 1 on CUDA API call to cuPointerGetAttribute
========= Saved host backtrace up to driver entry point at error
========= Host Frame:/usr/lib64/libcuda.so.1 (cuPointerGetAttribute + 0x1b5) [0x13a9e5]
========= Host Frame:../../bin/wrf.PGI.MVAPICH2.Alexey+Carl.single.exe (__pgi_uacc_cuda_shared + 0xa6) [0x221bece]
I launched my program under cuda-memcheck and got a huge bundle of such errors. The same errors appear under cuda-gdb
It’s a benign error having to how we’re getting device pointer information. I noticed this myself a few weeks ago. Michael is aware of it and said he’d clean it up in a future release. I just added TPR#20633 to track the issue so we remember to update this post once it’s fixed.
The OpenACC runtime calls cuPointerGetAttribute to test whether the pointer is part of a cuda ‘unified memory’ or cuda managed memory object. If it is, then the openacc runtime will let the cuda memory manager handle it implicitly. If it is not, then cuPointerGetAttribute returns an error code, rather than simply returning that the pointer is not managed. The 14.7 release will include a way to remove this error message, but it will also remove the ability to use cuda unified memory or managed memory with OpenACC. Setting the environment variable PGI_ACC_NOSHARED to a nonzero value will remove the calls to this routine.