Open villekf opened 5 months ago
Note that the PyOpenCL parts have been commented in the tests: https://github.com/arrayfire/arrayfire-python/blob/master/tests/simple/interop.py#L78, although the issue that is mentioned there has been closed over six years ago.
Contrary to my earlier mention, this also affects Linux systems.
Currently the OpenCL interoperability does not work on 64-bit Windows machine. The reason for this is the use of
c_int_t
which defaults to long on Windows (which on the other hand is 32-bit). This can be problematic with things like device IDs as they can easily go over 2^32/2-1. For example https://github.com/arrayfire/arrayfire-python/blob/master/arrayfire/opencl.py#L88 usesc_int_t
, as well as a few other functions in the same file. This can be fixed by replacing thec_int_t
with, for example,c_void_ptr_t
as it is used with the context. Is there by the way some specific reason why context is treated differently?There is also another issue. When the context is created by PyOpenCL, the input device, context and queue values are input as python ints in order to use the same context and queue, but on Windows you get
<class 'OverflowError'>: int too long to convert
errors then. These can be avoided by casting them to, for example,c_int64
-types. For example in https://github.com/arrayfire/arrayfire-python/blob/master/arrayfire/opencl.py#L156, replacesafe_call(backend.get().afcl_add_device_context(dev, ctx, que))
withsafe_call(backend.get().afcl_add_device_context(ct.c_int64(dev), ct.c_int64(ctx), ct.c_int64(que)))
(usingc_void_p
seems to work too).I haven't tested these changes extensively, but they do seem to get things working on 64-bit Windows. I can provide a PR if needed.