When they said procedural textures I was thinking something else, but this does look interesting none-the-less.

What I want to see... I'll call 'sampler shaders'. Instead of using a normal sampler, you could create and use a 'Sampler Shader'. When the GPU gets to the point in a shader where it needs to sample a texture, it would stop (conceptually only) and execute the samper shader. The sampler shader would be executed over a range (say a 16x16 or 32x32 block, whatever the driver/hardware thinks is necessary) to produce unfiltered texels which would be stored in a cache (type, size, policies, etc... would be driver handled). The shader performing the sampling would then be able to read from the cache, perform whatever filtering is necessary, and continue on. The sampler shader would have a minimum blocking size (much like a compute shader) and access to on-chip shared memory. This way complex procedural texture data could be created and used in real-time, dynamically adjusting its resolution/quality based on its usage in the scene. Not only procedural texturing but higher quality texture compression (wavelet, VQ, fractal, whatever), would be rather trivial...

Alas this seems to be something else... /sigh

Hope this means direct compute will get feature parity with opencl/cuda.