Expose TORCH_XPU_API to Pytorch

[etaf]

Expose TORCH_XPU_API to Pytorch

[etaf]

This PR is relied upon stock Pytorch PR: pytorch/pytorch#126678

[EikanWang]

@etaf , is it critical to catch up PT 2.4?

[etaf]

@etaf , is it critical to catch up PT 2.4?

No, it's a performance optimization.

[etaf]

Since this Pytorch 2.4 release branch has been cut, I think we can start land this PR and the pytorch/pytorch#126678 that depend on it.

[EikanWang]

LGTM.

[EikanWang]

Since we have included the xpu/EmptyTensor, is ATen/EmptyTensor still required?

[etaf]

ATen/xpu/EmptyTensor.h is the exported interface of EmptyTensor.cpp
and ATen/EmptyTensor.h is for at::detail::empty_generic which is used in this EmptyTensor.cpp file.

[EikanWang]

Then my question will be why ATen/xpu/EmptyTensor.h includes ATen/EmptyTensor.h?

[etaf]

Sorry, but ATen/xpu/EmptyTensor.h does not include ATen/EmptyTensor.h. ATen/xpu/EmptyTensor.h is a copy of torch-xpu-ops/src/aten/EmptyTensor.h.

[etaf]

@fengyuan14 can you help review this PR?

[fengyuan14]

You can expose the var to PyTorch cmake,

torch-xpu-ops/src/CMakeLists.txt

Line 7 in a7c3bd2

set(ATen_XPU_INCLUDE_DIRS ${TORCH_XPU_OPS_ROOT}/src)

And append it into cmake var of Caffe2 or ATen include dirs.

[etaf]

You can expose the var to PyTorch cmake,

torch-xpu-ops/src/CMakeLists.txt

Line 7 in a7c3bd2

set(ATen_XPU_INCLUDE_DIRS ${TORCH_XPU_OPS_ROOT}/src)

And append it into cmake var of Caffe2 or ATen include dirs.

Wouldn't that make the two PRs dependent on each other?

[fengyuan14]

You can expose the var to PyTorch cmake,

torch-xpu-ops/src/CMakeLists.txt

Line 7 in a7c3bd2

set(ATen_XPU_INCLUDE_DIRS ${TORCH_XPU_OPS_ROOT}/src)

And append it into cmake var of Caffe2 or ATen include dirs.

Wouldn't that make the two PRs dependent on each other?

ATen_XPU_INCLUDE_DIR is the committed formal and stable API of torch-xpu-ops CMAKE. It is not implementation dependent.

[fengyuan14]

Can we do like this ? What do you think of it?

[fengyuan14]

Looks I have exposed it by CMAKE target torch-xpu-ops, like this,

You can retrieve them in caffe2/CMakeList.txt by get_target_property(ATen_XPU_INCLUDE_DIR torch-xpu-ops INCLUDE_DIRECTORIES)

[etaf]

The original design is to align the include path with CPU/CUDA, eg. ATen/xpu/EmptyTensor.h, can I still keep this path with your design ?