cuDNN: "cudnnTensor4dDescriptor_t" for sigmoid component in fully connected layer

Hi,
I try to experiment with cuDNN, trying to integrate it into an already existing DNN code. To start with something simple, I try to call the formula for sigmoid, which will be wrapped in a class. How should I set the “cudnnTensor4dDescriptor_t” needed by cudnnActivationForward? My data is stored in a buffer with row-major layout, where matrix dimensions are set by rows (size of minibatch), cols (number of neurons), stride (address interval between 2 successive rows, i.e. datapoints in minibatch).

Is it like this?
cudnnSetTensor4dDescriptorEx(d, t, rows, 1, cols, 1, rows, 1, stride, 1)

of does the ‘map per image’ correspond to a number of neurons in fully connected layer?
cudnnSetTensor4dDescriptorEx(d, t, rows, 1, cols, 1, rows, 1, stride, 1)

The concept of Tensor4dDescriptor is hard to understand in context if sigmoid.
Btw. why do we need a 4d tensor descriptor for something simple as sigmoid?
The intended DNN is fully connected…

Thanks,
Karel.

Solved already,
the row-major 2d matrix with stride is expressed as 4d-tensor this way :
cudnnSetTensor4dDescriptorEx(d, t, rows, cols, 1, 1, stride, 1, 1, 1));