Make codegen backend agnostic minus fallbacks

[bdhirsh]

This PR removes all xla-specific logic from the pytorch codegen, except for the CPU fallback logic (working on in a later PR here. The corresponding pytorch-side changes are pytorch/pytorch#58064 and pytorch/pytorch#58568.

Main changes:

• Rename aten_xla_type.h -> XLANativeFunctions.h, which is a little more in-line with how we name other in-tree generated files. Also updated the corresponding documentation.
• I deleted InitializeAtenBindings() mostly because it looks like legacy (it's a no-op), and it would be an extra hoop to force every external backend to implement the same function
• Removed the AtenXlaType class. XLA lowering are now just defined under torch_xla::{op}.
• I also started passing a few xla macros and headers directly through the yaml, instead of hardcoding them in the codegen. I'm actually hoping to remove them though, since the logging information seems redundant: they're used in the generate out/inplace wrappers, but those wrapper just call into the handwritten xla functional lowering, which also has its own logging information.

[bdhirsh]

Huh, not actually sure why these are showing up in the diff, since they're already there in master.

[bdhirsh]

fixed

[bdhirsh]

@ailzhang @JackCaoG I'm actually hoping that I can just completely remove this, but I wanted to confirm with you. Right now, the auto-generated add_out() kernel for XLA has the same logging behavior as the original XLA codegen, and looks like this:

at::Tensor wrapper_Tensor_add(const at::Tensor &self, const at::Tensor &other,
                              const at::Scalar &alpha) {
  // This is the actual XLA add kernel
  return torch_xla::add(self, other, alpha);
}

}  // anonymous namespace
at::Tensor &wrapper_out_add_out(const at::Tensor &self, const at::Tensor &other,
                                const at::Scalar &alpha, at::Tensor &out) {
  XLA_FN_TRACK(3);
  TF_VLOG(3) << "XLA wrapper_out_add_out :"
             << " self=" << self.toString() << " other=" << other.toString()
             << " out=" << out.toString();

  auto wrapper_out_add_out_tmp = wrapper_Tensor_add(self, other, alpha);
  at::_copy_from_and_resize(wrapper_out_add_out_tmp, out);
  return out;
}

That logging information in the generated add_out kernel seems unnecessary though, since it ends up calling directly into torch_xla::add, which has its own logging too..

Passing C++ macros directly into yaml so the codegen can plop them into a C++ function is also pretty fragile; what if another backend wants some other more custom logging.

So I think it would be cleaner to kill that code-generated logging and just expect all logging to be done in the functional kernel lowerings, but I wanted to see what your thoughts are!

[JackCaoG]

I agree that auto-generated add_out() should not have logging and counter if the actual kernel has the logging and counter already.

[JackCaoG]

Mostly LGTM, some minor nits.

[JackCaoG]

Could you change 2. to 3. and update the order in below.

[JackCaoG]

Not directly related to your change but could you update "These declarations are one to one mapping" to something similar to "These declarations are usually one to one mapping". We sometimes reuse XLATensor for multiple atenXlaType functions.

[bdhirsh]

Fixed the docs. I also removed the logging from the generated inplace/out wrappers, which simplifies some of the codegen and makes the yaml more robust.

[JackCaoG]

Thanks @bdhirsh !