[javacpp-pytorch] Overhead of exception-based flow control in AbstractTensor.create

[kilink]

I believe after the fix from #1722, an exception is thrown on every call to AbstractTensor.create.

It gets caught, and it looks like it is the behavior of PointerScope.attach, however the exception-based flow control adds overhead if this is called in a hot path, so perhaps it could be avoided or reworked.

[saudet]

Right, we can remove the arguments for PointerScope, but we'll need to make sure to call Pointer.retainReference() on any other Pointer that need to escape from there.

[saudet]

Let's see, that might be something we can ask Copilot to fix. How do we use that thing...

[saudet]

Looks like it's only available as a paid service, bummer

[mullerhai]

I believe after the fix from #1722, an exception is thrown on every call to AbstractTensor.create.

It gets caught, and it looks like it is the behavior of `PointerScope.attach`, however the exception-based flow control adds overhead if this is called in a hot path, so perhaps it could be avoided or reworked.

HI，I see you work on Netflix ，does Netflix want to try javacpp-pytorch？

[saudet]

I've fixed that in commit 7a3d912. Please try again with the snapshots:
https://central.sonatype.com/service/rest/repository/browse/maven-snapshots
https://central.sonatype.org/publish/publish-portal-snapshots/#consuming-snapshot-releases-for-your-project

/cc @jxtps

[mullerhai]

@saudet by the way, could add some overwrite torch.tensor method to support double array ,like torch.tensor(double[] elements)

[mullerhai]

            // 1D primitive arrays
            case arr: Array[Float] => toFloatTensor(arr)
//            case arr: Array[Double] => toDoubleTensor(arr) // torch.tensor() not support
            case arr: Array[Long] => toLongTensor(arr)
            case arr: Array[Int] => toIntTensor(arr)
            case arr: Array[Short] => toShortTensor(arr)
            case arr: Array[Byte] => toByteTensor(arr)
//            case arr: Array[Boolean] => toBoolTensor(arr)  // torch.tensor() not support


            // 1D Double/Boolean arrays (must handle before generic fallback)
            case arr: Array[Double] => torch.tensor(arr*)
            case arr: Array[Boolean] => torch.tensor(arr.map(if (_) 1 else 0)*)

@saudet

[saudet]

@mullerhai Please try to use AbstractTensor.create() instead

[mullerhai]

@mullerhai Please try to use AbstractTensor.create() instead

HI，@saudet ，You misunderstood what I meant. If the underlying C++ LibTorch does not support double arrays and boolean arrays for torch.tensor, then we don’t need to add such support. However, if the native C++ LibTorch does support them, I would prefer to implement the corresponding functionality. I’ve noticed that double arrays are quite commonly used when creating torch.tensor.
I know AbstractTensor likely supports these types, but developers mostly use the torch.tensor*() methods to create tensors in practice.

[saudet]

I see, there are some mappings for float, int, etc, but not for double. I've added those to commit a719486 but this isn't going to work for torch.tensor() anyway. Please use AbstractTensor.create()