When filling a 4D input tensor without initializing the pad elements, memcheck may complain about the NNPA instruction like this: ==3180208== Syscall param NNPA(in_tensor_1) points to uninitialised byte(s) This is a false positive, because the architecture declares the pad elements to be ignored.
Fixed with commit bba9d299ca5ed0d6220fd47cf0a5eed5a7ce3b55.