Lev Givon <[email protected]> writes:
> I'm trying to access the memory associated with a GPUArray from within a
> compiled extension built using Cython's memoryview feature. According to the
> Cython documentation, it is possible to access C arrays using this feature;
> however, when I attempt to do so using a GPUArray's pointer with a GPU (and
> version of CUDA) that supports UVA, the extension segfaults. Does anyone have
> any thoughts as to why this might be occurring?

Three ideas:

- Check in /proc/self/maps whether the value of ptr makes sense/is
  backed by a map.

- I've never used UVA. From

  
http://docs.nvidia.com/cuda/cuda-c-programming-guide/#unified-virtual-address-space

  it seems that that device memory doesn't need to be allocated in any
  special way to be host-accessible?

- I dimly remember being terribly excited about the unified address
  space business, only to find out that it's not as exciting as I
  thought. IIRC, it just means that everybody is using the same
  pointers, but not that everyone can dereference every
  pointer. Specifically, I recall that the host wasn't allowed to
  dereference dev pointers. I might be wrong though, and would be happy
  to learn that I'm wrong.

Andreas

Attachment: pgpla6CT3a4_A.pgp
Description: PGP signature

_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda

Reply via email to