[MLAS][KleidiAI] Apply fp16 split independent fixes

[Laan33]

Description

This PR splits out a small set of independent fixes and test-scope updates from the larger CPU fp16 enablement work.

Changes:

• Adds KleidiAI Conv2D regression coverage for LHS indirection cache behaviour when the per-thread padding buffer grows.
• Runs that regression in both short and long Conv2D test paths when KleidiAI convolution is available.
• Disables WebGPU for the fp16 Attention test case that uses mask_index, which WebGPU Attention does not support.
• Renames the dynamic QGEMM batch-count parameter from BatchSize to BatchN for consistency with MLAS batch naming.

Motivation and Context

These changes are independent from the wider fp16 MLAS/API and CPU EP enablement work in #28487.

The Conv2D regression exercises the small-CI -> large-CI -> small-CI sequence that can expose stale cached LHS pointer tables after the KleidiAI padding buffer grows. The Attention test exclusion keeps the test targeted to providers that support the tested mask_index input combination.

Note: This PR is the first offshoot from #28487.

[Copilot]

Pull request overview

Splits a small set of independent fixes from the larger Arm64 fp16 MLAS/KleidiAI enablement effort: a refactored KleidiAI Conv2D regression test, a WebGPU exclusion for an unsupported fp16 Attention case, and a parameter rename for naming consistency.

Changes:

• Extracted the KleidiAI LHS-pointer-cache-by-pad-growth regression into a named helper, gated on USE_KLEIDIAI, and invoked it from both ExecuteShort and the existing long-test location.
• Disabled WebGPU for AttentionBatch1_Float16 since WebGPU Attention does not support mask_index.
• Renamed BatchSize parameter to BatchN in ArmKleidiAI::MlasDynamicQGemmBatch to match the public MlasDynamicQGemmBatch declaration in mlas.h.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
onnxruntime/test/mlas/unittest/test_conv2d.h	Factor KleidiAI pad-growth regression into a helper and call it from short/long Conv2D tests; expand mlasi.h include guard to USE_KLEIDIAI.
onnxruntime/test/contrib_ops/attention_op_test.cc	Disable WebGPU EP for the fp16 mask_index Attention test.
onnxruntime/core/mlas/lib/kleidiai/qgemm_kleidiai.cpp	Rename BatchSize → BatchN to match declaration in mlas.h/mlasi_kleidiai.h.

---

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

[hariharans29]

Review — PR #28720 (MLAS/KleidiAI fp16-split independent fixes)

Verdict: approve. This is a clean split of three unrelated, low-risk changes from the larger fp16 enablement work in #28487. Each is small, locally justified, and easy to verify.

What's here

1. BatchSize → BatchN rename in ArmKleidiAI::MlasDynamicQGemmBatch to match the public MlasDynamicQGemmBatch declaration in mlas.h and mlasi_kleidiai.h. Pure cosmetic, zero behavioral change.
2. Conv2D KleidiAI pad-growth regression extracted from the long-test body into a named helper TestKleidiAiLhsPointerCacheByPadGrowth(), gated on USE_KLEIDIAI, and invoked from both ExecuteShort() and the existing long-test call site. The include-guard for mlasi.h was widened from MLAS_TARGET_AMD64 to MLAS_TARGET_AMD64 || USE_KLEIDIAI to make GetMlasPlatform().MlasConvPrepareOverride visible on aarch64.
3. WebGPU disabled for AttentionBatch1_Float16 because WebGPU Attention does not support the mask_index input. Plumbed via the existing disable_webgpu parameter on RunAttentionTest.

Things that are right

• The Conv2D regression is now in ExecuteShort too, so default onnxruntime_mlas_test runs catch a future regression — previously it only ran in the long path. That's a real coverage win for issue [cpu] Loading certain models leads to global error state on M4 Max #26669.
• The USE_KLEIDIAI gate on the helper body and the explicit MlasConvPrepareOverride == nullptr runtime check together make the new test a no-op on builds where KleidiAI is compiled in but not active. Correct layering.
• The rename keeps the parameter type and ordering identical, so callers don't have to change.
• The disable_webgpu plumbing reuses the existing 4-flag tail of RunAttentionTest rather than inventing new infrastructure.

Minor comments

1. #else return; #endif is dead. The function returns at the end either way:

void TestKleidiAiLhsPointerCacheByPadGrowth() {
#if defined(USE_KLEIDIAI)
    ...
#else
    return;   // ← unreachable; function ends immediately below
#endif
}

Drop the else branch entirely — leave the #if defined(USE_KLEIDIAI) block and let the function be empty otherwise. Stylistic only.

2. The regression-test comment migrated from the long-test body to the helper definition, but the new helper body has a much shorter version. The original comment explained the smaller-CI → larger-CI → smaller-CI sequence, the pad-buffer growth, the cache-invalidation hazard, and the issue link. The new helper comment is two sentences and drops the issue link. Please preserve the link to issue #26669 in the helper's comment — it's exactly the kind of context that gets lost otherwise.

3. for (int i = 0; i < 4; ++i) is unexplained at the helper site. The original code had a comment "Repeat a few times to increase the likelihood of triggering a reallocation". Keep that comment on the loop in the helper too.

4. The disable_webgpu flag is set via a positional argument at position 14 of RunAttentionTest. That call site is now:

RunAttentionTest(input_data, weight_data, bias_data, mask_index_data, output_data,
                 batch_size, sequence_length, hidden_size, number_of_heads, true /*use_float16*/,
                 false, false, 0, nullptr, nullptr,
                 AttentionMaskType::MASK_1D_KEY_SEQ_LEN, 0, 0,
                 false /*disable_cpu*/, false /*disable_cuda*/, false /*disable_dml*/, disable_webgpu);

Three positional falses before the one you actually care about. The /*disable_cpu*/ etc. inline comments are doing real work here — please don't lose them in a future cleanup. If RunAttentionTest ever grows another optional, consider a follow-up to convert these tail flags to a small struct AttentionTestDisableFlags { bool cpu, cuda, dml, webgpu; };.

5. The rename touches 7 lines in qgemm_kleidiai.cpp. Confirm there are no other callers or comments in the KleidiAI subtree that still say BatchSize for this function — grep_search for BatchSize in mlas/lib/kleidiai/ if you haven't already, to make sure header/impl consistency is complete.

Bottom line

Merge after the trivial dead-return cleanup and (ideally) restoring the issue-link comment on the Conv2D helper. Neither is a blocker. 84/85 CI checks passing.

[Laan33]

There are no other uses for BatchSize in the KleidiAI subtree for the MlasDynamicQGemmBatch function.

[hariharans29]

Re-review — PR #28720 (commit fbf60fc)

Verdict: approve once CI is green.

The new commit addresses all three review comments from the prior round:

• Issue #26669 link restored in the helper's comment.
• Loop-iteration rationale restored ("Repeat a few times to increase the likelihood of triggering a reallocation").
• Dead #else return; removed — the helper is now #if defined(USE_KLEIDIAI) ... #endif with an empty fallback, which is the right shape.

No new code introduced, no new concerns. CI was at 2/67 at fetch time (early in the run); wait for it to complete, then merge.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.