Fix --enable_arm_neon_nchwc being a silent no-op on Windows ARM64

[Copilot]

Description

setup_arm_neon_nchwc() exported MLAS_USE_ARM_NEON_NCHWC with a single set(... PARENT_SCOPE). On Windows ARM64 it is called from inside setup_mlas_source_for_windows(), so the value landed in that function's local scope and was discarded on return — it never reached the directory-scope target_compile_definitions(). Linux/macOS aarch64 call it at directory scope, where one hop is enough.

Changes in cmake/onnxruntime_mlas.cmake:

• New helper mlas_add_private_compile_definitions() — accumulates definitions in a directory property. Directory properties are unaffected by CMake function scopes, so call depth stops mattering.
• Merge step before consumption — the property is folded back into mlas_private_compile_definitions just before the foreach(mlas_target ...) loop, and ahead of onnxruntime_unittests.cmake being included. This is why the fix is not simply target_compile_definitions(onnxruntime_mlas ...) inside the function: the mlas test and benchmark targets also consume this list.
• MLAS_USE_SVE in the Windows branch switched to the same helper. It happens to work today (its PARENT_SCOPE is one hop from directory scope) but carries the same latent trap.

Directory-scope list(APPEND ...) sites (Linux/macOS aarch64, RVV) are untouched.

function(mlas_add_private_compile_definitions)
  set_property(DIRECTORY APPEND PROPERTY mlas_private_compile_definitions_from_functions ${ARGN})
endfunction()

Motivation and Context

Every #if defined(MLAS_USE_ARM_NEON_NCHWC) guard evaluated false on Windows ARM64: MlasNchwcGetBlockSize() returned 1, NchwcTransformer never ran, and sconv_nchwc_kernel_neon.cpp / spool_nchwc_kernel_neon.cpp compiled into empty translation units. The flag was accepted and silently dropped, producing a binary identical to one built without it.

Because the failure mode is a missing define rather than a build error, win-qnn-arm64-ci-pipeline.yml — which passes --enable_arm_neon_nchwc explicitly for NCHWc coverage — has been testing the non-NCHWc path. That pipeline now exercises the code paths it was added to cover.

• Fixes [Build] --enable_arm_neon_nchwc is silently a no-op on Windows ARM64 (MLAS_USE_ARM_NEON_NCHWC dropped by CMake function scope) #32060

[azure-pipelines]

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

[Copilot]

Pull request overview

Fixes Windows ARM64 NCHWc builds silently dropping required MLAS compile definitions across nested CMake function scopes.

Changes:

• Adds a directory-property helper for scope-independent compile definitions.
• Migrates ARM NCHWc and Windows SVE definitions to the helper.
• Merges accumulated definitions before applying them to MLAS targets and tests.

---

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

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.