Decompose dot_xpu_mkl into mul and sum in non oneMKL path

[Silv3S]

If oneMKL is not available, replace CPU fallback with dot to mul+sum decomposition to avoid unnecessary data copies between devices.

[Copilot]

Pull request overview

This PR addresses XPU torch.dot failing for torch.int64 by decomposing the operation into elementwise mul followed by sum, and it removes the CPU fallback when oneMKL is unavailable to avoid device↔host copies.

Changes:

• Add an explicit Long guard in the oneMKL path to run mul + sum instead of dot_xpu_mkl.
• Replace the non-oneMKL CPU fallback with mul + sum directly on XPU.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

dot_xpu_mkl only dispatches floating/complex (see AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES... in src/ATen/native/xpu/mkl/BlasImpl.cpp), so in USE_ONEMKL_XPU builds this function will still throw for other integer dtypes (e.g., Int/Short/Byte) even after the Long guard. If torch.dot is expected to work for all integral types, consider expanding this guard to cover all integral (non-bool) dtypes (or adding an explicit error for the unsupported ones).

Suggested change

	if (self.scalar_type() == at::ScalarType::Long) {
	if (c10::isIntegralType(self.scalar_type(), /*includeBool=*/false)) {

[CuiYifeng]

@Silv3S Since oneMKL dot doesn't support all integrate types, similar issues can also occur when the input tensor is another integer. Suggestions for two purposes:

• To fix UT only => skipping failed cases just like CUDA did
• To support comprehensive functionality => add support not only for long

[Silv3S]

Thanks for review. I think first suggestion is better than extending functionality for ints. User can invoke mul and sum that supports int inputs if needed

[guangyey]

I notice that CUDA's impl dot_cuda doesn't support long data type yet. Why do we need this.

[Silv3S]

Thanks for review @guangyey. You're right - we don't need it. I assumed that it should be implemented based on failing ut from open issue. Then I checked that dot for int64 is also not covered by CUDA, as they just skip this ut.

But as general improvement I'd consider replacing existing CPU fallback with mul+sum for non-oneMKL path to keep calculations on device

#if defined(USE_ONEMKL_XPU)
  return at::native::xpu::dot_xpu_mkl(self, other);
#else
  // return at::native::dot(self.cpu(), other.cpu()).to(self.device());
  return at::mul(self, other).sum();
#endif

[CuiYifeng]

@Silv3S Since oneMKL dot doesn't support all integrate types, similar issues can also occur when the input tensor is another integer. Suggestions for two purposes:

• To fix UT only => skipping failed cases just like CUDA did
• To support comprehensive functionality => add support not only for long

[CuiYifeng]

Just a reminder: not sure whether this mathematically equivalent approach will overflow for large inputs.
For functionality first, this approach is acceptable.

[Silv3S]

Agree. They are numerically close, but it's not exact match. I checked multiple cases and they pass for 1e-5 tol (in fp32), but to not introduce any instabilities maybe it's better to leave the CPU fallback as is.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.