Fix Where discarding the sign of a selected -0.0

[Kevin Xu (keevin-xu)]

Description

Replaces the two != T{} comparisons in Where's merge step with a WasSelected helper that also
admits negative zero, and adds four regression tests to WhereOpTest.

Where<T>::Compute builds an X_selection and a Y_selection, each holding the default T{}
where the element was not selected, then merges them with
(X_selection != default value) ? X_selection : Y_selection. For floats the default is +0.0, and
-0.0 == 0.0. So a -0.0 selected from X compares equal to the default, the merge treats it as
unselected, and takes Y_selection, which holds +0.0 at that position. At equal shapes
Y_selection is the untested fall-through of that merge, which is why a -0.0 from Y survives
there but not when Y is the broadcast scalar operand of MergeScalarAndVector.

A negative zero can only be present in a selection tensor if it was selected, since the default has
its sign bit clear, so admitting it is sufficient. Selecting +0.0 from X still falls through to
Y_selection, which holds +0.0 at that position, so nothing else changes.

For the CPU kernel the affected registered floating-point types are float and double.
MLFloat16 and BFloat16 are present in the registration block but commented out, and the
std::string specialization uses a separate overload where the empty string is the default and is
indistinguishable between the two selection arms, so it is unaffected.

OpTester compares floating-point outputs numerically and -0.0f == 0.0f, so a lost sign is
invisible to it; the new tests check the sign bit through SetCustomOutputVerifier, as in #31477.
Three of the four fail before this change and all pass after, with onnxruntime_provider_test green
on macOS arm64 (5553 passed, 0 failed).

Motivation and Context

Where is elementwise selection rather than arithmetic, so the selected element should be
preserved, and +0.0 and -0.0 are distinct IEEE-754 values with distinct bit patterns.
onnx.reference preserves the negative zero in both cases; ONNX Runtime does not.

Fixes #32191.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Kevin Xu (keevin-xu)]

@microsoft-github-policy-service agree

[Copilot]

Pull request overview

Fixes CPU Where so selected negative-zero values retain their sign.

Changes:

• Adds signed-zero-aware selection detection.
• Adds regression tests for equal-shape and broadcast paths.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
onnxruntime/core/providers/cpu/tensor/where_op.cc	Preserves selected negative zero during merging.
onnxruntime/test/providers/cpu/tensor/where_op_test.cc	Adds float signed-zero regression coverage.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.