Failing tests on i386 and ppc64el

[nileshpatra]

Hi @jpritikin

Thanks for your work on OpenMx.
The version 2.19.8 seems to fail a few tests on i386 and ppc64 - all of them stem from test-cor.R. Logs for both can be found here and here respectively
I am also pasting the problem below. Please consider fixing this.

══ Failed tests ════════════════════════════════════════════════════════════════
── Failure (test-cor.R:38:3): (code run outside of `test_that()`) ──────────────
diag(fit$expectedCovariance$values) not equal to rep(1, 5).
5/5 mismatches (average diff: 0.788)
[1] 1.55 - 1 == 0.546
[2] 1.48 - 1 == 0.479
[3] 1.45 - 1 == 0.452
[4] 1.48 - 1 == 0.477
[5] 2.99 - 1 == 1.988
── Failure (test-cor.R:39:3): (code run outside of `test_that()`) ──────────────
diag(mxGetExpected(fit, "covariance")) not equivalent to rep(1, 5).
5/5 mismatches (average diff: 0.788)
[1] 1.55 - 1 == 0.546
[2] 1.48 - 1 == 0.479
[3] 1.45 - 1 == 0.452
[4] 1.48 - 1 == 0.477
[5] 2.99 - 1 == 1.988
── Failure (test-cor.R:38:3): (code run outside of `test_that()`) ──────────────
diag(fit$expectedCovariance$values) not equal to rep(1, 5).
5/5 mismatches (average diff: 0.516)
[1] 1.53 - 1 == 0.532
[2] 1.50 - 1 == 0.501
[3] 1.58 - 1 == 0.581
[4] 1.48 - 1 == 0.477
[5] 1.49 - 1 == 0.491
── Failure (test-cor.R:39:3): (code run outside of `test_that()`) ──────────────
diag(mxGetExpected(fit, "covariance")) not equivalent to rep(1, 5).
5/5 mismatches (average diff: 0.516)
[1] 1.53 - 1 == 0.532
[2] 1.50 - 1 == 0.501
[3] 1.58 - 1 == 0.581
[4] 1.48 - 1 == 0.477
[5] 1.49 - 1 == 0.491
── Failure (test-cor.R:56:1): (code run outside of `test_that()`) ──────────────
`mgConstraintJacobian` not equivalent to `lrConstraintJacobian`.
10/200 mismatches (average diff: 0.562)
[1]   2.246 - 1.78 ==  0.463
[12]  2.355 - 1.87 ==  0.490
[23]  2.384 - 1.89 ==  0.496
[34]  2.306 - 1.92 ==  0.381
[45]  0.207 - 1.95 == -1.738
[106] 2.243 - 1.78 ==  0.460
[117] 2.338 - 1.87 ==  0.472
[128] 2.244 - 1.89 ==  0.356
[139] 2.303 - 1.92 ==  0.378
...

[jpritikin]

Hm, yeah. We've seen this before, but it is not clear how to fix. There is some math precision difference between amd64 and these architectures. Maybe we should just move the test so it isn't run by default?

[nileshpatra]

Hm, yeah. We've seen this before, but it is not clear how to fix. There is some math precision difference between amd64 and these architectures. Maybe we should just move the test so it isn't run by default?

I am not sure on how to fix those float point problems either. If these are not very very important, it could be nice to move the tests so they do not run by default. For now, I could exclude them in Debian package and upload.

Thanks for replying!