cudnn: input tensor smaller than filter size

Accelerated Computing GPU-Accelerated Libraries
1

I use cudnn v2 to do CNN.
An error occured during cudnnGetConvolutionForwardWorkspaceSize as:
CUDNN failure: CUDNN_STATUS_BAD_PARAM

The reason I found is that the input tensor is 1x128x3x12 while the filter is 128x128x5x5.

The convolution descriptor is pad 2, strid 1. So even an input tensor of size 1x1 can be convoluted with a kernel 5x5 actually.
Is cudnn finding this as an error?
Is this cudnn v2 specific or later versions do the same?