Extend ort with custom input/output types for custom op

[maaft]

I can extend onnxruntime with custom ops.

Example:

 void KernelOne(const Ort::Custom::Tensor<float> &X,
                   const Ort::Custom::Tensor<float> &Y,
                   Ort::Custom::Tensor<float> &Z)
    {
        auto input_shape = X.Shape();
        auto x_raw = X.Data();
        auto y_raw = Y.Data();
        auto z_raw = Z.Allocate(input_shape);
        for (int64_t i = 0; i < Z.NumberOfElement(); ++i)
        {
            z_raw[i] = x_raw[i] + y_raw[i];
        }
    }
    
 static const std::unique_ptr<OrtLiteCustomOp> c_CustomOpOne{Ort::Custom::CreateLiteCustomOp("CustomOpOne", "CPUExecutionProvider", KernelOne)};
 
 domain.Add(c_CustomOpOne.get());

Although this works, I'm currently limited to use Ort::Custom::Tensor<T> for input/output types.

Can I also extend onnxruntime with custom data-types?

E.g.

struct FooParam {};

void KernelOne(const FooParam &X)
   {

   }

Use-case: I want to move custom data through the graph.

I want to reuse onnx graph scheduling capabilites to add pre- and post-processing custom ops. In particular post-processing nodes will produce non-tensor objects (e.g. cv::Mat). Therefore I need a way to extend onnxruntime with custom input/output types or a way to extend ort::Custom::Tensor.