Hybrid 2022.03.20.1: No module named 'vsdpir'

[Dan64]

the dlls I need for GTX cards are:

• cublas64_11.dll
  
• cublasLt64_11.dll
  
• cudart64_110.dll
  
• cudnn64_8.dll
  
• cudnn_cnn_infer64_8.dll
  
• cudnn_ops_infer64_8.dll
  
• cufft64_10.dll
  
• cufftw64_10.dll
  

see: https://onnxruntime.ai/docs/execution-pr...vider.html
and I thought that

• nvinfer.dll
  
• nvinfer_plugin.dll
  
• nvonnxparser.dll
  
• nvparsers.dll
  

should provide TensorRT support. (Just checked, thise re the only dlls which come with TensorRT-8.0.3.4)

Cu Selur

I was finally able to get TensorRT working with vsDPIR.

First it is necessary to perform a full installation of CUDA 11.4 Developer components.
Then after having extracted the zip file of TensorRT-8.0.3.4 it is necessary to copy  "lib\*dll" in  "c:\program files\nvidia gpu computing toolkit\cuda\v11.4\bin"
Finally it is necessary to install via "pip" the following python modules that are available in the archive of  TensorRT-8.0.3.4:

graphsurgeon-0.4.5-py2.py3-none-any.whl
uff-0.6.9-py2.py3-none-any.whl
onnx_graphsurgeon-0.3.10-py2.py3-none-any.whl

on my RTX 3060 the TensorRT version of vsDIPIR is only 5% faster than the CUDa version.