Short examples

First, a kernel that implements a 3x3 median filter on a 2-d array:

and the interface specification that lets us call it:

interface median3x3 {
    kernel median3x3(in int32 width, in int32 rowwidth, in float32* a, out float32* ret );
    alias first as median3x3(in int32 width, in int32 rowwidth, in float32* a, in float32* ret );
    alias step as median3x3(in int32 width, in int32 rowwidth, resident float32* a, resident float32* ret );
    alias last as median3x3(in int32 width, in int32 rowwidth, resident float32* a, out float32* ret );
};

Note the first/step/last aliases to the kernel function which hint buffers to be downloaded (first), used locally (step), and then returned (last).

Now, the wrapper code for loading that in an even prettier way:

and a driver program that compares running the median filter 10 times on each of an increasing series of image sizes.

A lot of the code in this example is included to create the following picture, which shows on my machine that the OpenCL median 3x3 is about 8 times faster than Numpy over a fairly wide range of array sizes, and that the same kernel using the resident buffer hinting show above runs about 30 times faster than Numpy.