Now our shape inference is a very winding. Here is an example: to get some non-duplicatable input's dim in compile time, our program has to switch repeatedly between CompileTimeInferShapeContext and InferShapeContext. The call stack is like this:
(Calling from top to bottom)

DDim CompileTimeInferShapeContext::GetInputDim(string param_name); | vector<DDim> InferShapeContext::GetInputsDim(string param_name); | vector<DDim> InferShapeContext::GetDims(vectir<string> arg_name); | DDim CompileTimeInferShapeContext::GetDim(string arg_name); | DDim VerDesc::Shape(); 

What happens at runtime is similar, just replace CompileTimeInferShapeContext with RuntimeInferShapeContext.

There are at least two issues here:

1. For a non-duplicatable input, the CompileTimeInferShapeContext::GetInputDim is invoked, while for a duplicatable input the InferShapeContext::GetInputsDim will be directly invoked. It mean's the entries of the same function dispersed in two class.
2. vector<DDim> InferShapeContext::GetInputsDim(string param_name); is intended for duplicatable inputs. However, even the input is non-duplicatable, it will still be invoked indirectly. That is inefficient.

Solution:

Move the GetInputDim from CompileTimeInferShapeContext(and RunimeInferShapeContext) to InferShapeContext. And the GetInputDim invokes GetDim directly instead of detouring via GetInputsDim.