Currently we have some comments in tests saying "due to precision loss we can only use a larger atol/rtol". When error accumulates it might be hard to distinguish numerical issues and true bugs.
We can use float64 instead of float32 in tests to provide a higher precision. The speed loss should not be important for tests.
Currently we have some comments in tests saying "due to precision loss we can only use a larger
atol
/rtol
". When error accumulates it might be hard to distinguish numerical issues and true bugs.We can use float64 instead of float32 in tests to provide a higher precision. The speed loss should not be important for tests.