How to use a device pointer in PyCUDA kernel call?

[villekf]

This is a similar question as outlined here: #347 (comment)

I have an Arrayfire array that allows to output the device pointer as int. Is it possible to somehow use this device pointer in a kernel call? Or make a GPUArray from it? In PyOpenCL this can be achieved with cl.MemoryObject.from_int_ptr, but is there an equivalent function in PyCUDA?

[pierrepaleo]

You can create a GPUArray from a device pointer, which can be latter passed to kernels.

import numpy as np
import arrayfire as af
if af.get_active_backend() != "cuda":
    af.set_backend("cuda")
import pycuda.gpuarray as garray

af_arr = af.to_array(np.ones(50, dtype="f"))
pycuda_arr = garray.empty(af_arr.shape, np.float32, gpudata=af_arr.device_ptr())
assert pycuda_arr.ptr == af_arr.device_ptr()
pycuda_arr += 1
print(af_arr.to_ndarray())

Beware, from my (admittedly very limited) experience with arrayfire, it seems that arrayfire almost-always re-allocates memory (even for in-place operations like af_arr += np.float32(1)) ; so the arrayfire pointer is likely to change.