[MLAS] Add int8 extreme-value coverage for ARM64 SymmQgemm

[kjg0724 (kjg0724)]

Summary

The ARM64 symmetric quantized GEMM tests (MlasSymmQgemmTest) only ever feed A and B with MatrixGuardBuffer::GetBuffer()'s default fill, which is confined to [21, 64) — even ExecuteLong's large M/N/K sweep never touches an int8 extreme. This adds SymmQgemmS8SignedInputTest, mirroring the QgemmS8U8SignedInputTest pattern from #29787, with A and B both explicitly filled with int8 extremes (-128, -1, 0, 1, 127, ...) across a K/M/N/offa grid sized around the kernel's actual block structure (PackedK=16, StrideM=4, N aligned to 16).

This is a test-only change; nothing under onnxruntime/core/mlas/lib/ is touched.

While adding these tests, I found a correctness bug in the plain-NEON (non-dotprod) MlasSymQgemmS8KernelNeon kernel and filed it separately as #31573. The new tests here pass because they exercise the SDOT dispatch (default on both machines I verified this on); the NEON dispatch only gets exercised on ARM64 cores without dot-product support.

Testing

onnxruntime_mlas_test, full suite, no regressions:

• Apple M1 (macOS): 29227/29227 passed
• Neoverse-N1 (Oracle Cloud A1, Ubuntu): 36020/36020 passed

*SymmQGemmS8_Int32_SignedInput*: 1248/1248 passed on both (624 cases x SingleThread/Threaded).

[azure-pipelines]

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

[Copilot]

🟡 Changes recommended

The new extreme-value tests should be skipped or conditionally registered on ARM64 hosts without NEON DotProd support to avoid known failures from the non-dotprod SymmQgemm NEON kernel (#31573).

Once you've addressed the issues Copilot identified, you can request another Copilot review.

This review doesn't count toward merge requirements. Sign up for the private preview to control whether Copilot approvals count.

Pull request overview

Adds additional MLAS unit-test coverage for ARM64 symmetric quantized GEMM (SymmQgemm) by introducing a new short-execute test suite that explicitly feeds extreme int8 values into both A and B, ensuring corner cases like INT8_MIN/INT8_MAX are exercised (instead of relying on the default [21,63] fill pattern).

Changes:

• Register a new SymmQgemmS8SignedInputTest suite for both single-threaded and threadpool-threaded short-execute runs.
• Implement SymmQgemmS8SignedInputTest fixture that fills A/B with explicit extreme int8 patterns across a small M/N/K/offa grid.

File summaries

File	Description
onnxruntime/test/mlas/unittest/test_symm_qgemm.cpp	Registers the new signed-input short-execute test suite (single-thread + threaded).
onnxruntime/test/mlas/unittest/test_symm_qgemm_fixture.h	Adds the new SymmQgemmS8SignedInputTest fixture that fills inputs with extreme int8 values.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 1
• Review effort level: Lite

We're testing this review assessment. Please use 👍 or 👎 to tell us if it's correct.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.

Suppressed comments (2)

onnxruntime/test/mlas/unittest/test_symm_qgemm_fixture.h:136

• The new signed-input tests register under the suite name "SymmQGemm...", but the existing SymmQgemm suites (e.g., MlasSymmQgemmTest::GetTestSuiteName) use "SymmQgemm...". Using a different capitalization makes filtering/triage inconsistent across SymmQgemm tests.

    testing::RegisterTest(
        Threaded ? "SymmQGemmS8_Int32_SignedInput_Threaded" : "SymmQGemmS8_Int32_SignedInput_SingleThread",
        test_name.c_str(),

onnxruntime/test/mlas/unittest/test_symm_qgemm_fixture.h:149

• These tests are expected to skip entirely on ARM64 hosts without FEAT_DotProd; as written, they still register 624 (per mode) individual tests and each one hits GTEST_SKIP. Consider short-circuiting registration when dot-product isn’t available to avoid noisy output and unnecessary per-test overhead on such hosts.

  static size_t RegisterShortExecuteTests() {
    size_t test_registered = 0;