On Dienstag 09 Dezember 2008, Randy Heiland wrote:
> However, trying to run some of the pycuda examples seem problematic:
>
> heiland:~/Documents/workspace/pycuda/examples$ python demo.py
> [[ 1.46184683  1.11751497 -5.8999877  -7.09263182]
>   [-0.39740902  0.94425917 -0.61876231  0.38862616]
>   [-2.40514922  2.16720867 -0.15235817 -0.51888126]
>   [ 0.07973941  1.79443252  1.62692916  1.70564294]]
> [[ 0.73092341  0.55875748 -2.94999385 -3.54631591]
>   [-0.19870451  0.47212958 -0.30938116  0.19431308]
>   [-1.20257461  1.08360434 -0.07617909 -0.25944063]
>   [ 0.03986971  0.89721626  0.81346458  0.85282147]]
> [[  0.00000000e+00   0.00000000e+00   0.00000000e+00   3.00927702e-36]
>   [  1.59078493e-33   4.74793215e-38   1.88079096e-37   4.51463371e-36]
>   [  0.00000000e+00   1.88079096e-37   7.70560149e-34   0.00000000e+00]
>   [  0.00000000e+00   0.00000000e+00   1.88470114e-37   0.00000000e+00]]
> [[  0.00000000e+00   0.00000000e+00   0.00000000e+00   3.00927702e-36]
>   [  1.59078493e-33   4.74793215e-38   1.88079096e-37   4.51463371e-36]
>   [  0.00000000e+00   1.88079096e-37   7.70560149e-34   0.00000000e+00]
>   [  0.00000000e+00   0.00000000e+00   1.88470114e-37   0.00000000e+00]]
> terminate called after throwing an instance of 'cuda::error'
>    what():  cuMemFree failed: invalid value
> Abort trap

Weird. Go to src/cpp/cuda.hpp and uncomment CUDAPP_TRACE_CUDA, rm -Rf build, 
rebuild PyCuda, and post the output. Also try to reduce demo.py to a point 
where it doesn't crash, and post the smallest crashing/largest non-crashing 
versions.

I only see this kind of behavior if a CUDA context gets invalidated in the 
middle of program execution--this would be accompanied by "NVRM Xid" messages 
in the kernel log (use dmesg on Linux to see it). On Linux, this happens when 
there's a bug in a kernel and the GPU's memory protection trips.

Confused...

Andreas

Attachment: signature.asc
Description: This is a digitally signed message part.

_______________________________________________
PyCuda mailing list
[email protected]
http://tiker.net/mailman/listinfo/pycuda_tiker.net

Reply via email to