Exposure of _ExecutionContext into public API

[s1ddok]

Currently I'm trying to implement custom CUDA kernel for S4TF. It seems mostly straightforward (given that C API contains TF_LibraryLoad and all other necessary stuff), but operation launching is what I'm struggling with. It is possible to create a new TFE_Context, but I would love my ops to share context with built-in ops.

Can ExectuionContext be exposed into public API? I know it may change with time, but as well can do things like _Raw, so maybe making it publicly available for development time will be useful for advanced experimenting.

Thanks!

[dan-zheng]

_ExecutionContext is already public:

import TensorFlow
print(_ExecutionContext.global.deviceNames)
// ["/job:localhost/replica:0/task:0/device:CPU:0", "/job:localhost/replica:0/task:0/device:XLA_CPU:0"]

It sounds like you'd like specific internal _ExecutionContext members to be public. Could you please clarify which ones?

Making members public sounds pretty good to me since _ExecutionContext is an underscored API.

[s1ddok]

Yes, initially I wanted to obtain a pointer to eagerContext so that I could create my own ops in the same context and execute them. I still think it could be useful, but after a closer look I see that TFTensorOperation don't expect a context as an injection and instead use singletone context in initializer, so now I think maybe it is more reasonable to just make

func makeOp(_ name: String, _ nOutputs: Int) -> TFTensorOperation {
    _ExecutionContext.makeOp(name, nOutputs)
}

a public function to enable users to create their own ops?

[s1ddok]

this is now partially addressed in #697