gjkRaycast of a capsule against convex taking a really long time to terminate when PhysX is built with clang and /fp:fast

[Zeblote]

Found this mysterious issue recently. I am building PhysX with clang as it makes much of my code quite a bit faster. However, there is one problem, and that's this happening sometimes:

A sweep query taking almost a second to complete rather than the usual microseconds makes the game rather unplayable.

Neither the standard VC17 build nor a clang build with /fp:precise have this issue. What could be going on there? I'm not sure if this should be reported as a bug since using clang on windows is not the default, though /fp:fast is. Maybe the linux build would show the same problem since it is also using clang?

I've tried to put #pragma float_control(precise, off) in all the gjk functions but this had no effect at all, only setting /fp:precise for the entire library seems to fix the issue.

[preist-nvidia]

Hi @Zeblote - we are not aware of this issue on Linux - any chance you tried to repro with our defaults on Linux?

[Zeblote]

Finally got to check this again. The correct floating point arguments for clang-cl on windows are /fp:fast /clang:-fhonor-infinities /clang:-fhonor-nans /clang:-fno-associative-math. This solves the issue.

[preist-nvidia]

Thank you! We'll check and integrate!

[preist-nvidia]

Do you have a repro by any chance for the sweep lag?

[Zeblote]

Not off hand in a way I can share easily, it happens seemingly randomly on queries against our custom compound geometry, I would need to investigate a bit more and try to extract the specific shape that freezes it. Can potentially do that later this month.

[preist-nvidia]

Ok that would be very helpful to analyze the issue better, thank you.