Windows tensorrt 7.2.1 version give wrong result

Description

Tensorrt windows version give wrong convolution result.

I have tested linux version cudnn, linux tensorrt 7.2.1, windows cudnn, they all give consistent results, but windows tensorrt 7.2.1 give differenct result.

linux cudnn: -2.06 1.87
linux trt: -2.06 1.87 0.000
windows cudnn: -2.06 1.87 0.000
windows trt: -2.06 1.87 0.429

-2.06 is the minimum value of output tensor, 1.87 is the maximum value of output tensor, 0.429 is the max abs difference compared to linux cudnn.

Environment

TensorRT Version: 7.2.1
GPU Type: RTX2080Ti
Nvidia Driver Version: 456.81
CUDA Version: 11.1
CUDNN Version: 8.0.4
Operating System + Version: windows 10
Python Version (if applicable):
TensorFlow Version (if applicable):
PyTorch Version (if applicable):
Baremetal or Container (if container which image + tag):

Relevant Files

link:
https://drive.google.com/file/d/1KCxvblYp0K_ITwd8z2L7iCFAyAfKGpLa/view?usp=sharing

Steps To Reproduce

just compile the test code, input and output tensors are flattened to 1-d array, and stored to binary file!

Hi @liweihua,
Please allow us some time, as we are checking on this.

Thanks!

May I ask now do you support Tensorflow-based tensorRT on Win 10? thank you very much

We did’t install tensorflow on windows 10.