arrayfire-python icon indicating copy to clipboard operation
arrayfire-python copied to clipboard

OpenCL interoperability does not work on 64-bit OS

Open villekf opened this issue 2 months ago • 2 comments

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 uses c_int_t, as well as a few other functions in the same file. This can be fixed by replacing the c_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, replace safe_call(backend.get().afcl_add_device_context(dev, ctx, que)) with safe_call(backend.get().afcl_add_device_context(ct.c_int64(dev), ct.c_int64(ctx), ct.c_int64(que))) (using c_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.

villekf avatar Apr 09 '24 14:04 villekf