Use protobuf-lite to reduce onnxruntime.dll size.

[pranavsharma]

More details about the LITE_RUNTIME can be found here https://developers.google.com/protocol-buffers/docs/proto. The reduction is significant. For commit id: 4873b452151bafe49da332aaeab639ef0318fc1ca28d728, the size reduced by 700K; from 4873728 to 4172800 bytes.

• Introduce a new build option "use_full_protobuf" so that onnx-tensorrt converter doesn't need to accommodate the protobuf-lite related changes. This option is OFF by default for non-tensorrt builds.
• Introduced 2 new files protobuf_parsing_utils.h/cc that is a copy of zero_copy_stream_impl.h/cc. These are required when using protobuf-lite to retain the same efficiency in loading the model files. This is based on the suggestion by the protobuf author here: https://groups.google.com/forum/#!topic/protobuf/J3H0GSlrfIo
• Note that I had to disable 2 test cases since they relied on parsing protobuf represented in text format; protobuf-lite doesn't support this functionality. We can convert the text protobuf to binary if required. As such in our case this functionality is needed for unittests only.

More details on the binary analysis can be found here: https://microsoft-my.sharepoint.com/:w:/p/prs/ERGfQ3IIjWxFtznJCM_nLbwBVrziUgzcvNN9T3a6ETDIQw?e=OydgWL

[snnn]

Any performance change?

[pranavsharma]

Any performance change?

There should be zero change in performance since the model loading functionality has not changed. Previously we were able to take advantage of single APIs: ParseFromIStream and ParseFromFileDescriptor. Now we use the exact same functionality in 2 steps. See inference_session.cc and model.cc.

[snnn]

Please take a look at the build log.
It has a lot of error messages like:
"Parsing config failed."

Apparently, your new config parsing code doesn't work.

[pranavsharma]

Please take a look at the build log.
It has a lot of error messages like:
"Parsing config failed."

Apparently, your new config parsing code doesn't work.

There was no problem with the parsing. The file just doesn't exist for many tests. The error message printed was misleading. I'll remove that line since it made sense only when we did protobuf parsing.