ggml_flip or ggml_pad_reflect?

[PABannier]

Hello,

I have implemented a custom operation ggml_pad_reflec_1d on my ggml fork. This is required for Encodec.cpp.
Should I upstream this operation? I would write a ggml_pad_reflect supporting 1D and 2D input, as in PyTorch's nn.ReflectionPad2d

Alternatively, I was considering implementing a ggml_flip operation (PyTorch's flip). This would allow us to implement a pad_reflect operation directly in 'user space' using flipping instead of adding a quite niche operation (pad_reflect) to ggml.

What are your thoughts? @slaren @ggerganov

[ggerganov]

Alternatively, I was considering implementing a ggml_flip operation (PyTorch's flip). This would allow us to implement a pad_reflect operation directly in 'user space' using flipping instead of adding a quite niche operation (pad_reflect) to ggml.

I don't have a good idea which option would be more useful, so whatever you think makes more sense - PR welcome