Tensorrt: cudnnBatchNormalizationBackwardEx

[SvanKeulen]

When using TensorRT 7.0.0.11,
I'm getting a complaint about resolving cudnnBatchNormalizationBackwardEx from module myelin64_1.dll (see attachment)
I'm sure I'm using the correct versions according to the support-matrix.

Inspecting dependencies and entry-points shows this entry-point is from an other module.
The actual DLL containing the entry-point is cudnn64_7.dll

Any help is appreciated.

Environment:
Windows10
VisualStudio.15.Release/15.9.16+28307.858
cuda_10.2.89_441.22_win10
cudnn-10.2-windows10-x64-v7.6.5.32
TensorRT-7.0.0.11.Windows10.x86_64.cuda-10.2.cudnn7.6

[SvanKeulen]

To add some more context...
When I use functionality from TensorRT in a basic command-line application, all seems fine.
When I use it in a DLL that is used inside a managed environment (C#), the issue described above arises.
No TensorRT functionality is called from C# directly. Only raw-data is shared.

Currently TensorRT6 workes just fine, only with TensorRT7 I get this problem.
I tried a clean install of cuda, cudnn and tensorrt7 on a clean system to be sure. Same problem.

[rmccorm4]

Hi @SvanKeulen,

Sorry, I don't have any experience using TensorRT on Windows. However, I see this commit on the upstream ONNX parser which sounds related: https://github.com/onnx/onnx-tensorrt/commit/2e60cfde7941e9819dfae79cdb1d893e8191d16b

Does building the ONNX parser master branch from source (per this repo's README) solve your issue?

(Though I don't know how supported that is currently per: https://github.com/NVIDIA/TensorRT/issues/36, though it's a pretty old issue)

[SvanKeulen]

@rmccorm4
Thanks for the reply, unfortunately I don't have an opportunity to do a deeper, code level investigation at the moment.

[mnboos]

I had the same error, but it was the result of a wrong cuDNN version (7.3). After updating to 7.6.5 it worked with CUDA 10.0.

[SvanKeulen]

After testing TensorRT 7.1.3

CUDA Runtime            : 11.0 (11000)
CUDA driver             : 11.0 (11000)
cuDNN                   : 8.0.3
NVidia ONNX parser      : 0.1.0
NVidia UFF parser       : 0.6.9

I can confirm the issue is now resolved.