OpenCL tests fail gracefully if the OpenCL library can't be loaded

[michel-steuwer]

This PR makes all tests ineracting with the OpenCL executor fail gracefully if the executor library can not be loaded from disk (e.g. when it is not available on a specific development platform).

This PR guards against all code that failes do to a UnsatisfiedLinkError that is called from within a test function.

[Bastacyclop]

@michel-steuwer That looks fine for TestsWithExecutor, but we also use util.withExecutor sometimes (e.g. for shine.DPIA.Primitives.Scatter), as well as util.ExecuteOpenCL when using host code generation.

[michel-steuwer]

@michel-steuwer That looks fine for TestsWithExecutor, but we also use util.withExecutor sometimes (e.g. for shine.DPIA.Primitives.Scatter), as well as util.ExecuteOpenCL when using host code generation.

I wasn't aware of this, but I have added support for this now.

[Bastacyclop]

For util.ExecuteOpenCL maybe we can try running something like clang -lOpenCL trivial.c. If it fails, OpenCL is not installed? Maybe we actually then put the isOpenCLAvailable in a global variable and use that for all three cases? (TestsWithExecutor, util.withExecutor, and util.ExecuteOpenCL)

[umazalakain]

What about exposing this opt out switch to the user? That way we keep things explicit and not engage in guess work.

[michel-steuwer]

What about exposing this opt out switch to the user? That way we keep things explicit and not engage in guess work.

We could have an environment variable to opt-in or opt-out. That would also be good going forward with CUDA-specific tests coming.

@umazalakain, @Bastacyclop what should the defaults be? I.e. do we opt-in into OpenCL and CUDA tests, or opt-out?

[Bastacyclop]

@umazalakain @michel-steuwer I like automatic support detection, maybe we can detect support by default while exposing override switches to bypass detection?

[Bastacyclop]

So if clang -lOpenCL test.c works on my machine, OpenCL tests will be enabled by default. But I can still set RISE_NO_OPENCL=1 or whatever variable to disable them.

[umazalakain]

Sorry, somehow I missed these notifications. Automatic support detection + override switch sounds great!

[umazalakain]

Maybe the only risk of having it be automatic is that the user might not be aware of whether they are running or not, mostly because currently the tests output a lot of information (I've been thinking of getting rid of all those print statements left in the tests, they make it more difficult to scroll through the output)

[michel-steuwer]

We should aim for having a configuration with three states:
RISE_OPENCL=ENABLE, RISE_OPENCL=TRY, RISE_OPENCL=DISABLE

We should also make this work for CUDA.