Can the input and output tensors in cuDNN be the same? For instance, in cudnnOpTensor (API Reference :: NVIDIA Deep Learning cuDNN Documentation), can
C have the same address as
B so I can do
A = op(A, B)?
Also, is it possible in general for other cuDNN functions as well?
I think C can have the same address as A, but not B (unless A==B==C). we would encourage you to try it yourself, if it passes then you are allowed to do so, otherwise, you would see CUDNN_STATUS_BAD_PARAM.
Okay. Thanks. I tried it at that time and it did work for C=A but not B like you said but I was not sure if it’s still gonna cause some issue or not.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.