Yep, as nnunn says, PyFR caches the results of curtain function calls so when called again with the same input the function doesn’t have to be evaluated again, but rather the previous result can be used.
In the kernel pipeline of the cuda backend, the key line for this caching is here: PyFR/provider.py at 37479d405cf5e35cc42bce206d395e68274968bb · PyFR/PyFR · GitHub