Hi Bogdan, all, Bogdan Opanchuk <manti...@gmail.com> writes: > Both in PyCUDA and PyOpenCL constructors of GPU arrays have > ``strides`` keyword parameter, and you can create a non-contiguous > array, e.g.: > >>>> import pyopencl as cl >>>> from pyopencl.array import Array >>>> import numpy >>>> ctx = cl.create_some_context() >>>> a = Array(ctx, (100,100), numpy.float32, strides=(512,4)) >>>> a.strides > (512, 4) > > But you cannot get it back to CPU: > >>>> a.get() > ... > AssertionError: Array in get() must be contiguous > > And the size of the underlying buffer does not seem to correspond to > the strides: > >>>> a.base_data.size > 40000 > > So, my question is: is it just that the specific feature of > non-contiguous array support is missing at the moment, but in > principle can be implemented, or there was an actual design decision > of not supporting it at all? Or is it because the ``reorg`` fork of > compyte seems to support it, and we are waiting for it to be merged?
I thought it was important that the data structure is able to represent non-contiguous arrays, but they're not really supported at all by any of the built-in operations. As you note, the plan is that in the medium term bringing in the compyte array should fix this for many operations. Andreas
pgpJRv4Az7MFE.pgp
Description: PGP signature
_______________________________________________ PyCUDA mailing list PyCUDA@tiker.net http://lists.tiker.net/listinfo/pycuda
