I spotted a bug in TensorRT 3 / 4 conv algo 14 tactic -2535759802710599445.
The results are diverging from caffe if you set non null padding and this tactic is chosen, leading to invalid results.
The other tactics are OK AFAIK.
Is it possible to disable this very tactic for further testing?
Tested with :
Ubuntu 16.04 GTX 1080 GTX 1080 TI TensorRT 3.0.4 TensorRT 4 RC CuDNN 7.1.3 Cuda 8 Driver v. 367.48 Driver v. 381.22