[QNN EP] Add QNN EP UDO support

[chenweng-quic]

Description

Add UDO support in QNN

Example usage:
./onnx_test_runner -v -e qnn -j 1 -i "backend_path|./libQnnCpu.so op_packages|<op_type>:<op_package_path>:<interface_symbol_name>[:<target>],<op_type2>:<op_package_path2>:<interface_symbol_nam2e>[:<target2>]" <models>

Motivation and Context

Add QNN EP UDO support.
For more information, ref op packages

[HectorSVC]

/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline,Windows x64 QNN CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 5 pipeline(s).

[Copilot]

Pull Request Overview

Adds support for loading and registering user-defined op packages (UDO) in the QNN execution provider via a new op_packages configuration option.

• Extended test runners and CLI tools to parse and validate an op_packages argument.
• Implemented ParseOpPackages, integrated it into QNNExecutionProvider and QnnBackendManager, and invoked a new LoadOpPackage() step.
• Introduced UDOBuilder and updated the builder factory to register custom UDO-based operators.

Reviewed Changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
onnxruntime/test/perftest/ort_test_session.cc	Added op_packages to the session config keys and validated its value.
onnxruntime/test/perftest/command_args_parser.cc	Updated help text to document the op_packages CLI option.
onnxruntime/test/onnx/main.cc	Enhanced usage output to include op_packages description.
onnxruntime/core/providers/qnn/qnn_execution_provider.cc	Added ParseOpPackages helper and integrated it into provider initialization.
onnxruntime/core/providers/qnn/builder/qnn_utils.cc	Clarified printing of 64-bit scalar types to note UDO support.
onnxruntime/core/providers/qnn/builder/qnn_backend_manager.h	Extended QnnBackendManagerConfig and manager class to store op_packages.
onnxruntime/core/providers/qnn/builder/qnn_backend_manager.cc	Implemented LoadOpPackage() logic and invoked it in SetupBackend().
onnxruntime/core/providers/qnn/builder/opbuilder/udo_builder.cc	Added UDOBuilder for UDO operator construction.
onnxruntime/core/providers/qnn/builder/opbuilder/base_op_builder.h	Enabled scalar handling for int64_t (64-bit) types for UDO.
onnxruntime/core/providers/qnn/builder/op_builder_factory.h	Introduced APIs to register UDO builders in the factory.
onnxruntime/core/providers/qnn/builder/op_builder_factory.cc	Registered the global OpBuilderRegistrations instance for UDO.

Comments suppressed due to low confidence (3)

onnxruntime/core/providers/qnn/builder/qnn_backend_manager.cc:324

• Missing #include at the top of the file; std::filesystem::path requires the header.

std::filesystem::path op_package_path = op_package[1];

onnxruntime/core/providers/qnn/builder/qnn_backend_manager.cc:382

• Typo in variable name 'op_package_for_registeration' and related log messages; should be 'op_package_for_registration'.

std::string op_package_for_registeration = op_package_path.stem().string();

onnxruntime/core/providers/qnn/qnn_execution_provider.cc:182

• No unit tests currently cover ParseOpPackages; adding tests for valid and invalid package strings would help prevent regressions.

static void ParseOpPackages(const std::string& op_packages_string, std::vector<std::vector<std::string>>& op_packages) {

[HectorSVC]

Need to add new provider options "op_packages" here:

onnxruntime/include/onnxruntime/core/session/onnxruntime_c_api.h

Line 3765 in 53f3c6c

* QNN supported keys:

Please also list all the valid value for target.

Also need to update the doc under gh-pages branch once the PR get merged:
https://github.com/microsoft/onnxruntime/blob/gh-pages/docs/execution-providers/QNN-ExecutionProvider.md

[chenweng-quic]

Need to add new provider options "op_packages" here:

onnxruntime/include/onnxruntime/core/session/onnxruntime_c_api.h

Line 3765 in 53f3c6c

* QNN supported keys:

Please also list all the valid value for target.

done

Also need to update the doc under gh-pages branch once the PR get merged: https://github.com/microsoft/onnxruntime/blob/gh-pages/docs/execution-providers/QNN-ExecutionProvider.md

This might need discussion since on Windows it has signing issue, and most of users may not be able to use this feature.

[HectorSVC]

/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline,Windows x64 QNN CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 5 pipeline(s).