Determinism: iOS/ARM vs. Win32/x86

[Spaddlewit]

My program runs on both Win32/x86 and iOS/ARM. I am getting different collision results from both - the Win32/x86 seems to be more reliable.

On iOS/ARM, Sometimes my sphere will stop on the edge of a cliff on a btBvhTriangleMesh, and the wrong m_index for the contact point's polygon index will be returned.

Is there some reason ARM would be less accurate or produce occasional bogus results compared to x86?

[codetiger]

My program runs on both Win32/x86 and iOS/ARM. I am getting different collision results from both - the Win32/x86 seems to be more reliable.

On iOS/ARM, Sometimes my sphere will stop on the edge of a cliff on a btBvhTriangleMesh, and the wrong m_index for the contact point's polygon index will be returned.

Is there some reason ARM would be less accurate or produce occasional bogus results compared to x86?

Are you testing with same stepsimulations and CollisionConfig settings? We have a game developed for Win32, iOS, Android and Mac. Determinism mostly depends on step simulation parameters.

[Spaddlewit]

Yes, exactly the same timestep. A lot of times the m_partId* and m_index* will return 0 instead of the uninitialized -8655302 (whatever). Even if I change the bullet code to initialize these variables to -1, it will still return bogus 0 values. I have had to hardcode to ignore index 0 and pad out my meshes with a fake first triangle.