[Web] Fail to run some WPT WebNN argMin/argMax conformance tests of uint32/uint64 types by default CPU EP

[BruceDai]

Describe the issue

When running WPT WebNN argMin/argMax conformance tests by default CPU EP on latest Chrome Canary (version: 140.0.7260.0), below 4 tests of uint32/uint64 types run Fail.

argMin uint32 4D tensor, axis=1, all options
argMin uint64 4D tensor, axis=1, all options
argMax uint32 4D tensor, axis=1, all options
argMax uint64 4D tensor, axis=1, all options

Error message on the tab of navigating to chrome://gpu.
[24856:30700:0626/134813.515:ERROR:services\webnn\ort\graph_impl_ort.cc:178] : [WebNN] Failed to call ort_model_editor_api->CreateSessionFromModel( env.get(), model_info->model.get(), session_options->get(), ScopedOrtSession::Receiver(session).get()): [WebNN] ORT status error code: 9 error message: Could not find an implementation for ArgMin(13) node with name '_0'
[24856:30700:0626/134813.545:ERROR:services\webnn\ort\graph_impl_ort.cc:178] : [WebNN] Failed to call ort_model_editor_api->CreateSessionFromModel( env.get(), model_info->model.get(), session_options->get(), ScopedOrtSession::Receiver(session).get()): [WebNN] ORT status error code: 9 error message: Could not find an implementation for ArgMin(13) node with name '_0'
[24856:30700:0626/134814.361:ERROR:services\webnn\ort\graph_impl_ort.cc:178] : [WebNN] Failed to call ort_model_editor_api->CreateSessionFromModel( env.get(), model_info->model.get(), session_options->get(), ScopedOrtSession::Receiver(session).get()): [WebNN] ORT status error code: 9 error message: Could not find an implementation for ArgMax(13) node with name '_0'
[24856:30700:0626/134814.384:ERROR:services\webnn\ort\graph_impl_ort.cc:178] : [WebNN] Failed to call ort_model_editor_api->CreateSessionFromModel( env.get(), model_info->model.get(), session_options->get(), ScopedOrtSession::Receiver(session).get()): [WebNN] ORT status error code: 9 error message: Could not find an implementation for ArgMax(13) node with name '_0'

@fdwr @shiyi9801 PTAL, thanks!

To reproduce

1. Download and install latest Chrome Canary for Windows 11/10 64-bit
2. Open Command Prompt, execute blow command to launch Chrome Canary

"%LOCALAPPDATA%\Google\Chrome SxS\Application\chrome.exe" --enable-features=WebMachineLearningNeuralNetwork,WebNNOnnxRuntime

3. Navigate to test URL https://wpt.live/webnn/conformance_tests/arg_min_max.https.any.html?cpu

Urgency

No response

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

f217402

Execution Provider

Other / Unknown

[shiyi9801]

I think the cause might be that the default CPU EP didn't register uint32 uint64 ArgMin/Max kernel:
https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/cpu/cpu_execution_provider.cc#L685

but the docs list them in the type constraints:
https://onnx.ai/onnx/operators/onnx__ArgMin.html#type-constraints

[BruceDai]

Similar issues on Fail conformance tests of int64 type for prelu and relu operators.

Error message
9564:24872:0626/152425.521:ERROR:services\webnn\ort\graph_impl_ort.cc:178] : [WebNN] Failed to call ort_model_editor_api->CreateSessionFromModel( env.get(), model_info->model.get(), session_options->get(), ScopedOrtSession::Receiver(session).get()): [WebNN] ORT status error code: 9 error message: Could not find an implementation for Relu(14) node with name '_0'

[19564:24872:0626/152455.384:ERROR:services\webnn\ort\graph_impl_ort.cc:178] : [WebNN] Failed to call ort_model_editor_api->CreateSessionFromModel( env.get(), model_info->model.get(), session_options->get(), ScopedOrtSession::Receiver(session).get()): [WebNN] ORT status error code: 9 error message: Could not find an implementation for PRelu(16) node with name '_0'

prelu test URL: test URL https://wpt.live/webnn/conformance_tests/prelu.https.any.html?cpu
relu test URL: test URL https://wpt.live/webnn/conformance_tests/relu.https.any.html?cpu

[skottmckay]

ORT doesn't implement all types for an operator. The types implemented are typically based on what real models have required. Based on the lack of registration it doesn't appear that any production models have used uint32 or uint64 for ArgMin/ArgMax operations in the history of ORT. This avoids unnecessary binary size and development and maintenance cost.

If you really really need this (which I'd question given no production model has previously and this sounds like an arbitrary conformance test)...

It may be a case of simply adding registrations in reduction_ops.cc

onnxruntime/onnxruntime/core/providers/cpu/reduction/reduction_ops.cc

Lines 316 to 320 in 7a6cef6

	REGISTER_UNARY_ELEMENTWISE_KERNEL(ArgMin, 13);
	REGISTER_UNARY_ELEMENTWISE_KERNEL_DOUBLE_ONLY(ArgMin, 13);
	REGISTER_UNARY_ELEMENTWISE_KERNEL_INT64_ONLY(ArgMin, 13);
	REGISTER_UNARY_ELEMENTWISE_KERNEL_INT8_ONLY(ArgMin, 13);
	REGISTER_UNARY_ELEMENTWISE_KERNEL_UINT8_ONLY(ArgMin, 13);

And updating here and here to add new entries.

I wouldn't bother doing it for older opsets.

Finally a couple of unit tests should be added for the data types. Should be fine to copy a trivial one like this and change the data type.

But we'd need to carefully check the binary size cost of the changes.