optimization: support for 2-D/3-D arrays with strides >= for better memory bandwidth utilization about pyopencl

[inducer]

PyOpenCL's arrays are, for now, restricted to being contiguous for most operations.

To my mind, the correct way to implement operations on multi-dimensional arrays would use loopy.

[nevion]

loopy is a cool thing to have learned about - I see libraries like that pop-up from academic labs in a few different contexts.

But what we're talking about on the kernel side is probably no code changes to anything that takes in pitches/strides in bytes - ie it assumes unpacked layout. If the stride is a bad value for certain GPUs than one would want to set it >= the packed value and this can greatly change memory bandwidth utilization / bank conflicts.

We just need a small bit more support to make allocating those arrays not be too painful - of course it's best if it works on everything that is currently supported. The most important things are
-get matrix values
-set from or to values from a packed numpy array or another clarray
-set to constant value