GPGPU Programming in .Net - Microsoft Accelerator VS OpenCL.net [VS Others]

Question

I am developing a program (in .Net) that has a lot of image processing (like 2D Convolutions, ...) and I want to do them with GPGPU (parallel processing) So, I have some choices: Microsoft Accelerator, OpenCL.Net, ... Which one is better for me?

Some things important for me: Speed, Compatibility (e.g. if GPU does not support that, what happens?), Easy to learn, Work with both Nvidia and ATI , ...

And I heard that I can create my own function in OpenCL and run it for many array cells, but I don't know if I can do that in Microsoft Accelerator or I have to work with its functions...