gt4py v1 removes the Storage class and allow any __array_interface__ describing object to be bound. Unfortunately, the default cupy allocation used in our model has a bad stride (should have unit stride) leading to performance decrease in the backend.
Potential solution:
use gt4py provided allocator (and optimized for the backend)
gt4pyv1 removes theStorageclass and allow any__array_interface__describing object to be bound. Unfortunately, the defaultcupyallocation used in our model has a bad stride (should have unit stride) leading to performance decrease in the backend.Potential solution: