PyOpenCL executor: accept device scalars for ValueArgs

[kaushikcfd]

-Closes #452.

• Closes ArrayContext.np.sqrt fails for device scalars meshmode#237.

[inducer]

Thanks! This was surprisingly lightweight. 😄 Two things below.

[inducer]

Thanks! Will go in after #451.

[inducer]

Actually, scratch that. Goes in now. :) Thanks again.

[inducer]

I think I will revert this; it does not properly do what it claims to do, because it looks at arguments at a point when it shouldn't. Case in point, if we change the test to:

def test_pyopencl_execution_accepts_device_scalars(ctx_factory):
    import pyopencl.array as cla

    ctx = ctx_factory()
    cq = cl.CommandQueue(ctx)

    knl = lp.make_kernel("{:}",
                         """
                         y = 2*x
                         """)

    evt, (out,) = knl(cq, x=cla.to_device(cq, np.asarray(21)))
    np.testing.assert_allclose(out.get(), 42)

    # adding this bit:
    evt, (out,) = knl(cq, x=np.int32(21))
    np.testing.assert_allclose(out.get(), 42)

things break, because the kernel to be executed got specialized to device scalars when it shouldn't have. Fixing this is non-trivial without a major hit to invocation performance, because to do this properly, we would need to compare the types of all arguments on every invocation. We can currently avoid that for kernels that do not have runtime-typed variables, resulting in a major cost savings that I do not want to give up.

[inducer]

#797 contains the revert.