Add missing MatMulInteger int8 x uint8 support

[rivkastroh]

Description

This PR adds support for the MatMulInteger operator when
input A is int8 and input B is uint8, and adds unit tests
to cover this type combination.

According to the ONNX specification for MatMulInteger, the type
constraints are:

• T1 ∈ {int8, uint8}
• T2 ∈ {int8, uint8}
• T3 = int32

This means all four combinations (T1, T2) = (int8,int8), (int8,uint8), (uint8,int8), (uint8,uint8) are valid. However, the implementation
was missing the (int8, uint8) registration, which caused a
NOT_IMPLEMENTED error at runtime for such models.

This PR aligns the kernel registration and tests with the ONNX spec.

Motivation and Context

Fixes #26743

Testing

• Added unit tests for the A=int8, B=uint8 combination:
  • MatmulIntegerOpTest.MatMulInteger_int8_uint8_2D
  • MatmulIntegerOpTest.MatMulInteger_int8_uint8_PerColumn_ND
• All tests pass locally.

[rivkastroh]

@yuslepukhin, @yufenglee : could you please review the PR? Thanks.

[Copilot]

Pull request overview

This PR adds support for the MatMulInteger operator when input A is int8 and input B is uint8, completing the ONNX specification compliance. The ONNX spec allows all four combinations of (int8, uint8) types for the two inputs, but the (int8, uint8) combination was previously missing, causing runtime errors.

• Updated the kernel registration to allow uint8 or int8 for T2 when T1 is int8
• Added comprehensive unit tests covering both 2D and N-D cases with per-column zero points

Reviewed changes

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

File	Description
onnxruntime/core/providers/cpu/quantization/matmul_integer.cc	Modified the int8_t typed kernel registration to accept both uint8_t and int8_t for the T2 type constraint, enabling the missing int8 x uint8 input combination
onnxruntime/test/providers/cpu/math/matmul_integer_test.cc	Added two new test cases (MatMulInteger_int8_uint8_2D and MatMulInteger_int8_uint8_PerColumn_ND) to verify the int8 x uint8 combination works correctly in both simple 2D and complex N-D scenarios with per-column zero points

---

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

[yuslepukhin]

/azp run Linux QNN CI Pipeline, Win_TRT_Minimal_CUDA_Test_CI, Windows ARM64 QNN CI Pipeline, Windows GPU CUDA CI Pipeline, Windows GPU DML CI Pipeline, Windows GPU Doc Gen CI Pipeline, Windows GPU TensorRT CI Pipeline, Windows OpenVINO CI Pipeline, Windows x64 QNN CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).

[yuslepukhin]

Well, there is a test failure:

[ RUN ] MatmulIntegerOpTest.MatMulInteger_int8_uint8_PerColumn_ND
2: 2025-12-08 22:02:43.0817483 [E:onnxruntime:MatMulInteger, inference_session.cc:2542 onnxruntime::InferenceSession::Initialize::<lambda_fe9fb62842ddccf7ffc0bf349066f2d4>::operator ()] Exception during initialization: D:\a_work\onnxruntime\onnxruntime\onnxruntime\core\providers\dml\DmlExecutionProvider\src\AbiCustomRegistry.cpp(519)\onnxruntime_provider_test.exe!00007FF6F5AF1D09: (caller: 00007FF6F5A65620) Exception(38) tid(2510) 80070057 The parameter is incorrect.
2:
2: D:\a_work\onnxruntime\onnxruntime\onnxruntime\test\unittest_util\base_tester.cc(319): error: Expected equality of these values:
2: expect_result
2: Which is: 4-byte object <00-00 00-00>
2: ExpectResult::kExpectFailure
2: Which is: 4-byte object <01-00 00-00>
2: Initialize failed but expected success: Exception during initialization: D:\a_work\onnxruntime\onnxruntime\onnxruntime\core\providers\dml\DmlExecutionProvider\src\AbiCustomRegistry.cpp(519)\onnxruntime_provider_test.exe!00007FF6F5AF1D09: (caller: 00007FF6F5A65620) Exception(38) tid(2510) 80070057 The parameter is incorrect.
2:
2: Google Test trace:
2: D:\a_work\onnxruntime\onnxruntime\onnxruntime\test\unittest_util\base_tester.cc(877): registered execution providers: DmlExecutionProvider
2:
2: [ FAILED ] MatmulIntegerOpTest.MatMulInteger_int8_uint8_PerColumn_ND (237 ms)

[rivkastroh]

@yuslepukhin, I updated the new int8 x uint8 per-column test to be skipped when the DmlExecutionProvider is present, since it hits the same existing DirectML AbiCustomRegistry issue as the existing MatMulInteger_PerColumn_ND test (see the TODO comment there).

The test still runs on other execution providers, so it continues to validate the new int8 x uint8 path without being blocked by the known DML issue.