I’m trying to implement Nd convolution and Nd pooling layers in Caffe using the cuDNN primitives,
and some of them returns error code that are not clearly documented in the case of Nd tensors.
First the cudnnAddTensor() function used to add biases to the convolution results returns CUDNN_STATUS_NOT_SUPPORTED which is not listed in the possible return values for this function…
mode is CUDNN_ADD_SAME_C
biasDesc is a 1 x C x 1x…x1 tensor with C the number of output features,
and as many ones as there are dimensions.
Then the cudnnPoolingForward() function also returns CUDNN_STATUS_NOT_SUPPORTED,
which is supposed to mean that “the wStride of the input or output tensor is not 1”.
But in the Nd case, it doesn’t make sense, as the width or height stride may not even be defined.
Thanks in advance for helping me understand what I do wrong!