cudnnPoolingForward() tensor format support

cuDNN 2.0 on CUDA 6.5 (Jetson TK1 / L4T)

This function only appears to support the NCHW tensor format, but not the NHWC format. Clearly the NCHW format is going to offer speed advantages, but is NHWC completely unsupported?


It’s recommended to use TX1 or TX2 for deep learning use case.

NHWC is supported in our newer cuDNN version but is not available for TK1.