[BYOC][TensorRT] Make TRT runtime robust to empty or weird subgraphs

[trevor-m]

TensorRT doesn't allow a graph to use the same nodes as both an input and output, nor does it allow an output to be marked as an output multiple times.

It seems there are still some corners cases where the partitioning can make these subgraphs which would cause TRT to crash at runtime during graph build. This PR will prevent the crash and allow the model to execute properly. I think it's good to have this in place in case there are more issues with partitioning we aren't aware of yet.

I will make a separate PR to fix the partitioning bug I have found.

[comaniac]

Thanks for the fix. Is it possbile to add a test case?

[trevor-m]

Thanks @comaniac just added one

Test output before fix

ubuntu@ip-172-31-15-219:/data/neo-ai-tvm/build$ python3 -m pytest ../tests/python/contrib/test_tensorrt.py -sk test_empty
enabled targets: llvm -device=arm_cpu; nvptx; llvm; cuda
pytest marker: 
====================================================================== test session starts =======================================================================
platform linux -- Python 3.6.6, pytest-6.1.2, py-1.9.0, pluggy-0.13.1
rootdir: /data/neo-ai-tvm, configfile: pytest.ini
plugins: arraydiff-0.2, cov-2.10.1, openfiles-0.3.0, remotedata-0.3.2, doctestplus-0.1.3
collected 49 items / 48 deselected / 1 selected                                                                                                                  

../tests/python/contrib/test_tensorrt.py [21:52:03] /data/neo-ai-tvm/src/runtime/contrib/tensorrt/tensorrt_logger.h:51: ERROR: Tensor tensorrt_output_0 cannot be both input and output
[21:52:03] /data/neo-ai-tvm/src/runtime/contrib/tensorrt/tensorrt_logger.h:51: ERROR: Network must have at least one output
[21:52:03] /data/neo-ai-tvm/src/runtime/contrib/tensorrt/tensorrt_logger.h:51: ERROR: Network validation failed.
Segmentation fault (core dumped)

With fix:

ubuntu@ip-172-31-15-219:/data/neo-ai-tvm/build$ python3 -m pytest ../tests/python/contrib/test_tensorrt.py -sk test_empty
enabled targets: cuda; nvptx; llvm -device=arm_cpu; llvm
pytest marker: 
====================================================================== test session starts =======================================================================
platform linux -- Python 3.6.6, pytest-6.1.2, py-1.9.0, pluggy-0.13.1
rootdir: /data/neo-ai-tvm, configfile: pytest.ini
plugins: arraydiff-0.2, cov-2.10.1, openfiles-0.3.0, remotedata-0.3.2, doctestplus-0.1.3
collected 49 items / 48 deselected / 1 selected                                                                                                                  

../tests/python/contrib/test_tensorrt.py [21:52:25] /data/neo-ai-tvm/src/runtime/contrib/tensorrt/tensorrt_builder.cc:106: tensorrt_output_0 is both an input and an output.
[21:52:26] /data/neo-ai-tvm/src/runtime/contrib/tensorrt/tensorrt_logger.h:54: WARNING: Current optimization profile is: 0. Please ensure there are no enqueued operations pending in this context prior to switching profiles
.

================================================================ 1 passed, 48 deselected in 3.05s ================================================================

[comaniac]

LGTM

[comaniac]

Thanks @trevor-m