Open geofjamg opened 1 month ago
Based on the tolerances you have set, the differences seem reasonable considering the CPU architectures are quite different (presumably x86_64 on Windows/Linux and then Apple ARM64). Small floating point roundoff differences due to architecture and/or compiler differences are to be expected, and with a looser tolerance, over the integration interval these can add up. Try the same tests with smaller tolerances and you should see the results get closer together.
Hello,
Current Behavior:
As an example, when simulating this simple model using SUNDIALS 7.0.0
At the last time step (5s), I get on Linux and Windows exactly the same values for yret and ypret:
While on MacOS, I get quite different values (and difference is growing when running more and more steps):
Expected Behavior:
I would expect to get consistent results on Linux, Windows and MacOS.
Steps To Reproduce:
Just compile this code and link with IDA and KLU linear solver libraries.
Environment:
I got this results using SUNDIALS 7.0.0 and GitHub action Linux, MacOS and Windows latest workers (ubuntu-latest, windows-latest, macos-latest). I also observe the same issue on my MacBook Pro (Darwin mbp-de-xxx 23.5.0 Darwin Kernel Version 23.5.0: Wed May 1 20:12:58 PDT 2024; root:xnu-10063.121.3~5/RELEASE_ARM64_T6000 arm64).
Anything else: