Xnnpack execution provider Resize::IsOnnxNodeSupported causes crash for models where Resize layer scales tensor is an empty tensor

[pl121lp]

Describe the issue

onnxruntime/core/providers/xnnpack/tensor/resize.cc:69 (const auto scales = scale_val.DataAsSpan();)
results in crash if scales tensor is an empty tensor.
This is the case from some onnx models where both scales and sizes tensor is provided. In this case scales is an empty tensor and size tensor has data. See attached netron visualization clip from the model.

Applying a condition (e.g. "if(scale_val.size() > 1) {" ) to check size of scales resolves this issue. I suggest this fix is introduced.

To reproduce

I cannot not share proprietary model.

Urgency

No response

Platform

Android

OS Version

Any recent version

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1ffe793

ONNX Runtime API

C++

Architecture

ARM64

Execution Provider

Other / Unknown

Execution Provider Library Version

xnnpack

[skottmckay]

Your model does not conform to the ONNX spec for Resize. You cannot provide both scales and sizes. Pass an empty string (which indicates a missing optional input) and not an empty tensor for scales.

https://github.com/onnx/onnx/blob/main/docs/Operators.md#inputs-1---4

scales (optional, non-differentiable) : tensor(float)
The scale array along each dimension. It takes value greater than 0. If it's less than 1, it's sampling down, otherwise, it's upsampling. The number of elements of 'scales' should be the same as the rank of input 'X' or the length of 'axes', if provided. One of 'scales' and 'sizes' MUST be specified and it is an error if both are specified. If 'sizes' is needed, the user can use an empty string as the name of 'scales' in this operator's input list.
sizes (optional, non-differentiable) : tensor(int64)
Target size of the output tensor. Its interpretation depends on the 'keep_aspect_ratio_policy' value.The number of elements of 'sizes' should be the same as the rank of input 'X', or the length of 'axes', if provided. Only one of 'scales' and 'sizes' can be specified.

[pl121lp]

Right, I concur that the model does not adhere to spec.
However, note that the model can run just just fine using the default CPUExecutionProvider.
This means that the default code path does have the ability to handle this case (even though IR is deviation from spec).