Tests for matmul for complex datatypes.

[PawelSwider2000]

Unskip test that were passing after implementing matmuls for complex datatypes.

Some test still fails but with different errors, unrelated to matmuls like: # NotImplementedError: The operator 'aten::cholesky_inverse.out' is not currently implemented for the XPU device.

Created as separate PR (implementation: #1992) as this will require additional PR in upstream torch

[Copilot]

Pull Request Overview

This PR enables complex datatype support for matrix multiplication operations on XPU devices by implementing oneMKL-based BLAS operations and unskipping relevant tests that were previously failing due to lack of complex matmul support.

• Implements complex datatype matrix multiplication operations using oneMKL BLAS library
• Unskips hundreds of tests that were disabled due to missing complex matmul support
• Updates error message comments to reflect that only double precision (not complex) matmul is unsupported in oneDNN

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
test/xpu/skip_list_common.py	Removes complex64/complex128 test skips and updates comments to reflect current oneDNN limitations
src/ATen/native/xpu/Blas.cpp	Implements oneMKL-based complex matrix multiplication operations (mm, bmm, addmm, baddbmm)

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

[nitpick] The comment should be more descriptive about what operations are affected. Consider: "RuntimeError: Double datatype matmul is not supported in oneDNN for certain operations"

Suggested change

	# RuntimeError: Double datatype matmul is not supported in oneDNN
	# RuntimeError: Double datatype matmul is not supported in oneDNN for certain operations (e.g., matmul, convolution)

[Copilot]

The clone().detach() operation creates an unnecessary copy of the tensor data. Since the tensor is already made contiguous and conjugate-resolved, consider using the tensor directly or only cloning when modification is actually needed to avoid the performance overhead.

Suggested change

	at::Tensor self_cont = self.contiguous().resolve_conj().clone().detach();
	at::Tensor self_cont = self.contiguous().resolve_conj();

[Copilot]

Similar to the addmm implementation, this clone().detach() operation creates an unnecessary copy. Consider avoiding the clone unless the tensor needs to be modified in-place to improve performance.

Suggested change

	at::Tensor self_cont = self.contiguous().resolve_conj().clone().detach();
	at::Tensor self_cont = self.contiguous().resolve_conj();

[CuiYifeng]

Although there are many test cases require complex MatMul for assertion, we don't need to enable all these cases at once. Activating the test case for directly testing Complex MatMul is sufficient.

[PawelSwider2000]

Although there are many test cases require complex MatMul for assertion, we don't need to enable all these cases at once. Activating the test case for directly testing Complex MatMul is sufficient.

Sure, I think we should run this test on CI and check them, if they are passing, I do not see reason to now unskip them. However if they are failing and are not directly testing MatMul they could stay skipped for now.