This is a question regarding the API for the function cudnnConvolutionBackwardFilter
The API reference states that the function takes a param of const cudnnFilterDescriptor_t dwDesc
,
where,
dwDesc
: Input . Handle to a previously initialized filter gradient descriptor
Granted that I have instantiated and used a cudnnFilterDescriptor_t filter_desc_
in cudnnConvolutionForward
- should I create a new tensor descriptor dwDesc
- or should I pass the previously existing (from forward propagation) filter_desc_
as the param for dwDesc
in cudnnConvolutionBackwardFilter
.