How to fill a PyTorch string_view input

[sbrunk]

The PyTorch bindings have a bunch of functions taking a c10:string_view, which is currently mapped to a Pointer.

Example:

@Namespace("at") public static native @ByVal Tensor einsum(@ByVal @Cast("c10::string_view*") Pointer equation, @ByVal @Cast("at::TensorList*") TensorArrayRef tensors);

I'm trying to call einsum with a string. My naive approach was to just initialize a BytePointer

import org.bytedeco.pytorch.*;
var rand = global.torch.torch_randn(4,4);
global.torch.einsum(BytePointer("ii"), new TensorArrayRef(rand.position(0), 1));

But that doesn't work:

Caused by: java.lang.RuntimeException: einsum(): the number of subscripts in the equation (0) does not match the number of dimensions (2) for operand 0 and no ellipsis was given

The error is the same when calling the Python API with an empty string.

Any idea how to correctly add the string_view value or does that need mapping support (i.e. a cast)?

Tagging @HGuillemet because you're deep into the PyTorch mappings right now.

[HGuillemet]

I'll look into this.
I think we could add a mapping.

[HGuillemet]

@saudet, please advise about this:

I first tried to create a custom adapter StringViewAdapter implementing proper conversions, but it causes problem, that I could explicit if needed, when instantiating containers like c10::optional<c10::string_view>.

Then I realized that the @StdString adapter is almost sufficient.
For functions returning a string_view, we should generate:

@Cast({"", "", "std::basic_string<char>"}) @StdString String f()

because string_view knows how to be cast to std::basic_string only.

For functions taking a string_view, we simply need:

void f(@StdString String)

We must not cast to std::basic_string. string_view has an implicit constructor taking a char *.

Is there an existing trick to generate the @Cast only for returned values ?
Else should we consider adding this option to Parser ?

A third option that should work, but less friendly, is to map string_view to a specific Java class.

[saudet]

It's probably easier to come up with something like a StringViewAdapter, possibly based on StringAdapter like this:
https://github.com/bytedeco/javacpp-presets/blob/master/tensorflow/src/main/resources/org/bytedeco/tensorflow/include/tensorflow_adapters.h#L53

[HGuillemet]

Ok. I have came up with a solution with an adapter working. New PR coming.
Like in StringAdapter, I tested the size passed to the (ptr, size, owner) constructors and did a strlen if it's 0.
But I wonder if it won't be better to have the generator produce a JNI call to getStringUTFLength, along with getStringUTFChars, and pass this length to the adapter.

[sbrunk]

For reference: The PR is #1392 and since it's merged, this is now working fine.