Performance and functional parity of C and Python API for the graph mode

[deadeyegoodwin]

System information

• TF 2.3
• Are you willing to contribute it (Yes/No): No

Describe the feature and the current behavior/state. Currently the TF C/C++ API is incomplete (e.g. does not have API to save savedmodel, does not expose all optimization and configuration options) compared to the Python API. It also seems that the Python API enables or performs additional optimization passes that are not available when using the C/C++ API.

Will this change the current api? How? Yes. The C/C++ API will likely need to be enhanced significantly. Ideally this complete functionality will be implemented in a C API to ensure maximum ABI compatibility and portability.

Who will benefit with this feature? Applications that want a more portable, high-performance solution that does not require using python. We are particularly interested in applications like model servers. A C API would be the most portable and performant and is relatively easy to integrate and wrap into other languages. Conceptually the python API (and any other language binding) should be a wrapper over the C API that presents the API in a language appropriate way but that does otherwise add (or remove) functionality.