faild to load .so lib dynamically as tensorflow env

After building a shared lib in external project, I'd like to make it work with tensorflow, as a platform backend storage env.

My own package as follows.

/path/to/tensorflow/tensorflow/core/platform/xxx $ tree
.
├── BUILD
├── xxx.cc
├── xxx.h
├── xxx_test.cc
└── libxxx.so

Building and running the target xxx_test, it failed as follows.

Seems something wrong while initializing, but it works fine with dlopen and dlsym in local test.

g++ version: 8.3.0

Need some advise, thanks in advance.