Lower threshold even more for SIFTFast to make the test pass on raoul

Summary

Update threshold for SIFTFast test to make this last failing test green: https://dash.orfeo-toolbox.org/testDetails.php?test=99045451&build=350210

Copyright

The copyright owner is CNES and has signed the ORFEO ToolBox Contributor License Agreement.

---

Check before merging:

• All discussions are resolved
• At least 2 votes from core developers, no vote.
• The feature branch is (reasonably) up-to-date with the base branch
• Dashboard is green
• Copyright owner has signed the ORFEO ToolBox Contributor License Agreement

+1, but I don't understand why this is needed. I mean, IEEE 754 is the same everywhere and SIFTFast is not a random algorithm, right? Why is it failing only on one platform if precision is the issue?

I was wrong... this is an interesting read: https://www.appinf.com/download/FPIssues.pdf

Well, the point is SIFT is not a random algorithm, but SIFTFast has a random behavior on windows, probably because of bugs in the SSE / memalign instructions (take a look at the code for a good laugh).

But it is not our responsibility to fix SIFTFast (or even test it). We only need to test the client code we wrote for SIFTFast (which is the purpose of the new test). We can decide SIFTFast is not good enough for us in term of stability, yet, it is still our best implementation of sift in terms of performances, you can see it in the output log of the new test.

The random behavior appears with the activation of openMP...

merged

mentioned in commit b8695a74

added bug label

changed milestone to %7.0.0