Define Custom Operator based on CPP program

[rwbfd]

🚀 Feature

Define Custom Operator based on a CPP program

Motivation

It is possible to define a custom-based CPP program or CUDA program as an operator. However, I am not entirely sure whether this can be done in PyTorch XLA.

If so, is it possible to create the code example with it? Or is it the same as defining a custom operator in PyTorch?

[JackCaoG]

Can you be a bit more specified on what are you trying to do? There is a way to define a c++ op like

xla/torch_xla/csrc/aten_xla_type.cpp

Lines 896 to 902 in 53e5d1d

 at::Tensor XLANativeFunctions::clamp(const at::Tensor& self, 
                                      const c10::optional<at::Scalar>& min, 
                                      const c10::optional<at::Scalar>& max) { 
   TORCH_LAZY_FN_COUNTER("xla::"); 
   return bridge::AtenFromXlaTensor( 
       tensor_methods::clamp(bridge::GetXlaTensor(self), min, max)); 
 } 

but you will either need a python binding to expose this op or you want to call this op directly from C++ somewhere.