On Thu, 15 Sep 2011 12:52:30 -0400, "Andrew J. Hesford"
Dear Dr. Kloeckner,
I have discovered a bug in your PyOpenCL implementation of 32-bit vector integers. In
pyopencl/array.py, on line 62, you list the tuple ('int', np.uint32) instead of
the correct ('int', np.int32). As a consequence, attempts to make an int3 vector
type actually make a uint3 vector type. Replacing the incorrect tuple with the correct one
appears to fix the problem.
Fixed in PyCUDA and PyOpenCL git.