I tired to do the same operation in cuDNN and in Tensorflow and the “SAME” mode in cuDNN and Tensorflow might different.
For example, I want do follow convolution
input_tensor 300 x 300 x 3
output_tensor 150 x 150 x 3
filter 3 x 3 x 3 x 3
In tensorflow, top and left pad are both 0. But it cause error in cuDNN, when I set pad as 0. When pad is 1, it seems to be right, but the shape of output_tensor is 149 x 149 x 3.
Can I set the pad of top smaller than that of bottem?