Add FunctionSchema based Operator Registry

[bwasti]

Summary: This enables creation of operators with FunctionSchema and IValue

Differential Revision: D13008791

[dzhulgakov]

Can you add more tests, especially if the op uses the new API to change output size?

[dzhulgakov]

this is not sufficient:

1. you assume the output is already tensor - what if it's not?
2. this is not the api we'd be calling - instead majority of places call OutputTensor(int idx, at::IntList dims, at::TensorOptions options) below which avoids reinitialization/dtype change in-place and instead creates new tensor input

in both cases we'd need to change ivalue in-place, i.e. we need a pointer to ivalue for this case

[bwasti]

makes sense -- updated

[dzhulgakov]

what about non-Tensor Output() version above? at least add an option to claim that it's not a "legacy operator"

[dzhulgakov]

Ping?

[dzhulgakov]

.get is redundant?

[dzhulgakov]

for a better test - don't initialize the IValue at all

[dzhulgakov]

if IValue wasn't holding tensor - should we reset it to the tensor type? as semantically it's "assignment"

[dzhulgakov]

?

[dzhulgakov]

if you move it inside the if and use tensor variable - you can save one refcount bump

[jerryzh168]

to keep the original behavior, we are changing this to tensor.raw_mutable_data(options.dtype()); see Line 48 of blob.h

[dzhulgakov]

Please fix mine and Jerry's comments. Otherwise looks great!

[dzhulgakov]

Ping?

[dzhulgakov]

?

[dzhulgakov]

I'd move this wrapping inside the if still to save refcount (otherwise you still bump one when you look up tensor above (or you can get device from at::Tensor directly