why does TFRT implement C++ style's kernel, instead of writing all kernels through MLIR format?

[Kaiven Long]

This section(https://github.com/tensorflow/runtime/blob/master/documents/tfrt_host_runtime_design.md#eager-execution-sketch) said that TFRT also use c++ fast path for op. I want to understand the design philosophy. For example, if there is a fast path, maybe no one will pursue the general path, for the consideration on performance.  thanks in advance if someone can help provide usage scenarios on general path and fast path.  

[Jing Dong]

As stated in the doc, the idea was to use the general path to handle "composite ops". The general path will use the compiler process to decompose the composite ops into a graph of more lower level ops. This allows the end-user to add ops by composing lower-level ops without writing the kernels for them.