PyCUDA Default CUDA Context overrides other CUDA contexts

[sjperkins]

The following mechanism

import pycuda.autoinit

creates a default CUDA context and pushes it on the CUDA context stack. Unfortunately, it does not pop the context until the program exits and this does not play nicely with other libraries using CUDA since their contexts, if initialised before PyMORESANE, will be replaced by PyMORESANE's, for all time (or at least the length of the program).

It would be nice if PyMORESANE handled it's context manually by pushing before executing CUDA code, and popping once its done:

import pycuda.tools

class ContextWrapper(object):
    def __init__(self, context):
        self._ctx = context

    def __enter__(self):
        self._ctx.push()
       return self

    def __exit__(self, evalue, etype, etraceback):
        self._ctx.pop()

__internal_context = None

def context_wrapper():
    if __internal_context is None:
        context =  pycuda.tools.make_default_context()
        context.pop() # Contexts seem to push on creation
       __internal_context = ContextWrapper(context)

   return __internal_context

def moresane_gpu_func(...):
   with context_wrapper():
       run_gpu_code()