Tips on translating to PyCUDA

[ceich]

I have some image reconstruction code that uses PyOpenCL but I need to use PyCUDA (targeting Jetson AGX Xavier which doesn't support PyOpenCL). I have found a somewhat helpful page, but could use a few more tips. Thanks!

[isuruf]

Slightly related. POCL with CUDA support works just fine on the Xavier and therefore you can keep using PyOpenCL.

[ceich]

Sorry, I'm a POCL n00b; how do I add CUDA support?

[ceich]

I began with this (3-year-old) answer saying that POCL is not supported on Xavier.

I (just this week) asked what an alternative was and was told PyCUDA.

[isuruf]

There's an issue with using both CUDA device and CPU device as mentioned in pocl/pocl#853,
but if you disable the CPU device, CUDA works. I don't have access to a Xavier, so I can't debug that.

[isuruf]

If you are using conda to install pyopencl (see https://documen.tician.de/pyopencl/misc.html#installing-from-conda-forge)
then you can use conda install pocl-cuda to get a pre-built pocl binary with CUDA support.
To avoid the CPU device, you can do export POCL_DEVICES=cuda

[inducer]

I agree with @isuruf that pocl should help offer OpenCL even where Nvidia does not. And with OpenCL being a much less stateful (i.e. much nicer!) API than CUDA, I think this is generally the technically preferable solution.

That said, if you must do a port, it's not too hard to come up with an abstraction layer, e.g.

• https://github.com/inducer/pyopencl/blob/main/pyopencl/_cluda.py
• https://reikna.publicfields.net/en/latest/introduction.html#cluda