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?
Here is the Cython code for a sample extension that uses memoryviews:
cdef class Wrapper:
cdef long n
cdef void *ptr
cdef double[:] view
def __cinit__(self, unsigned long long addr, long n):
self.n = n
self.ptr = <void *>addr
self.view = <double[:n]>self.ptr
def __getitem__(self, int i):
return self.view[i]
def __setitem__(self, int i, double val):
self.view[i] = val
When the above is compiled into an extension named my_ext using Cython 0.19.1
and Python 2.7.4 on Linux and passed the pointer associated with a numpy array
of doubles
as follows, I can use it to access the array's contents:
import numpy as np
import my_ext
x = np.arange(10, dtype=np.double)
w = my_ext.Wrapper(x.ctypes.data, x.size)
print w[0] # prints 0.0
w[0] = 100.0
print w[0] # prints 100.0
Attempting to do so with a GPUArray results in a segfault:
import pycuda.autoinit
import pycuda.gpuarray as gpuarray
x_gpu = gpuarray.arange(10, dtype=np.double)
w = my_ext.Wrapper(x.ptr, x.size)
print w[0] # segfaults
I'm using pycuda 2013.1.1 built with CUDA 5.0.35.
--
Lev Givon
http://www.columbia.edu/~lev/
http://lebedov.github.com/
_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda
