How does Halide deal with the fact that mobile processors are using ARM architectures, while desktops are more likely to use x86? Does it have some vast knowledge bank where it can intelligently optimise for the different hardware architectures/core counts/execution contexts, etc?

Also - does it turn out to be faster to use something like Halide to do this instead of using GPU processing? If so, is that because it takes too much time to shift data around inside the device?

@harrymellsop compiler has different stages and code generation is one of it. Code for ARM and x86 can be individually optimized.

@harrymellsop I think Halide can be used to generate code that works well on GPU.