Loosen numerically iffy test tolerance

[esseivaju]

The InvoluteSolver test fails on Perlmutter because the precision is too strict.

/global/cfs/cdirs/atlas/esseivaj/devel/celeritas/test/orange/surf/detail/InvoluteSolver.test.cc:392: Failure
Value of: dist_on[0]
  Actual: 0.0017713715293239856
Expected: 0.0017713715293786088
Which is: 0.0017713715293786088
(Relative error -3.0836664554021502e-11 exceeds tolerance 9.9999999999999998e-13)

/global/cfs/cdirs/atlas/esseivaj/devel/celeritas/test/orange/surf/detail/InvoluteSolver.test.cc:409: Failure
Value of: dist_off[0]
  Actual: 0.0017713715293239856
Expected: 0.0017713715293786088
Which is: 0.0017713715293786088
(Relative error -3.0836664554021502e-11 exceeds tolerance 9.9999999999999998e-13)

/global/cfs/cdirs/atlas/esseivaj/devel/celeritas/test/orange/surf/detail/InvoluteSolver.test.cc:432: Failure
Value of: dist_on[0]
  Actual: 0.00053216327768802003
Expected: 0.00053216327674743909
Which is: 0.00053216327674743909
(Relative error 1.7674668425285303e-09 exceeds tolerance 9.9999999999999998e-13)

/global/cfs/cdirs/atlas/esseivaj/devel/celeritas/test/orange/surf/detail/InvoluteSolver.test.cc:449: Failure
Value of: dist_off[0]
  Actual: 0.00053216327768802003
Expected: 0.00053216327674743909
Which is: 0.00053216327674743909
(Relative error 1.7674668425285303e-09 exceeds tolerance 9.9999999999999998e-13)

/global/cfs/cdirs/atlas/esseivaj/devel/celeritas/test/orange/surf/detail/InvoluteSolver.test.cc:573: Failure
Value of: dist_on[0]
  Actual: 0.00195043766381975
Expected: 0.0019504376639951655
Which is: 0.0019504376639951655
(Relative error -8.9936457836800689e-11 exceeds tolerance 9.9999999999999998e-13)

/global/cfs/cdirs/atlas/esseivaj/devel/celeritas/test/orange/surf/detail/InvoluteSolver.test.cc:590: Failure
Value of: dist_off[0]
  Actual: 0.00195043766381975
Expected: 0.0019504376639951655
Which is: 0.0019504376639951655
(Relative error -8.9936457836800689e-11 exceeds tolerance 9.9999999999999998e-13)

I have to significantly reduce the precision for the test to pass. Maybe we can refactor the computation so that we don't lose that much precision?

[sethrj]

Oh dear this is far too wide ranging a change. We need to drill down to this one test. It looks like the failing tests may be in a part of the solver that's numerically unstable or suffers from catastrophic rounding or something, because it's producing very different answers for single precision. I saw something similar in the nuclear form factor PR: 664ac38 .

The best solution here might just be to change EXPECT_SOFT_EQ to EXPECT_SOFT_NEAR to add the third tolerance option (and make that one 1e-8).

[esseivaju]

I agree the tolerance would be too wide. For now, let's use EXPECT_SOFT_NEAR.

[github-actions]

Test summary

 3 249 files   5 047 suites   3m 42s ⏱️
 1 520 tests  1 493 ✅ 27 💤 0 ❌
16 838 runs  16 776 ✅ 62 💤 0 ❌

Results for commit c8e7404.

♻️ This comment has been updated with latest results.