Hang on hang on let me write a CUDA kernel for this. This is going to be really huge.