RobustStatisticsTest failing on Alpine aarch64, ppc64le, s390x

[Hello71]

Describe the bug
I am packaging libjxl for Alpine Linux. Some tests failing, but I am unsure of the cause.

To Reproduce
Build libjxl on Alpine Linux and run tests.

Expected behavior
Successful build and tests.

Environment

• OS: Alpine Linux
• Compiler version: gcc 10.3.1_git20210921-r1 (see https://git.alpinelinux.org/aports/tree/main/gcc for details), musl 1.2.2
• CPU type: aarch64, ppc64le, s390x, x86 (32-bit, 64-bit is ok)
• libjxl version: 0.6

Additional context
For x86, the test suite requires -msse -mfpmath=sse in order to pass. For aarch64, ppc64le, and s390x, three tests fail in identical ways: RobustStatisticsTest.TestMode, RobustStatisticsTest.CleanLine, and DecodeTest.PixelTestWithICCProfileLossy. The last one affects all archs and is tracked as #500 already.

aarch64, ppc64le, and s390x failing test outputs (identical except for duration)

[ RUN      ] RobustStatisticsTest.TestMode
===========================frac 1
hsm 10.0 hrm 9 argmax 9.0 median 9.999786
1510 1761 1805 1810 1611
===========================frac 2
hsm 20.0 hrm 19 argmax 19.0 median 19.999693
1206 1227 1335 1310 1222
===========================frac 3
hsm 30.0 hrm 29 argmax 29.0 median 29.999731
1079 1092 1109 1110 1004
===========================frac 4
hsm 40.0 hrm 40 argmax 40.0 median 39.999756
895 937 956 916 850
===========================frac 5
hsm 49.0 hrm 49 argmax 49.0 median 49.999603
771 884 870 844 791
===========================frac 6
hsm 60.0 hrm 60 argmax 60.0 median 59.999840
768 756 785 768 723
===========================frac 7
/builds/alxu/aports/testing/libjxl/src/libjxl-0.6/lib/jxl/robust_statistics_test.cc:75: Failure
Value of: is_near(hsm, argmax) || is_near(hsm, median)
  Actual: false
Expected: true
hsm 69.0 hrm 69 argmax 70.0 median 69.999565
708 703 736 691 670
===========================frac 8
hsm 79.0 hrm 79 argmax 79.0 median 79.999634
655 692 700 700 616
===========================frac 9
hsm 90.0 hrm 90 argmax 90.0 median 90.000053
617 656 637 662 614
[  FAILED  ] RobustStatisticsTest.TestMode (41 ms)

[ RUN      ] RobustStatisticsTest.CleanLine
x 2319.072266  slope=0.60 b=2.00  max_l1 0.000002 mad 0.000000
/builds/alxu/aports/testing/libjxl/src/libjxl-0.6/lib/jxl/robust_statistics_test.cc:132: Failure
Expected: (max_l1) <= (max_l1_limit), actual: 1.90735e-06 vs 1e-06
[  FAILED  ] RobustStatisticsTest.CleanLine (3 ms)

I looked into this, but couldn't find any obvious culprits. I know that many floating-point functions return different results with different implementations, but it seems that these tests use only basic arithmetic, comparisons, and implicit comparisons (std::max, std::nth_element, etc), which should return identical results on all implementations.

Let me know what additional information is needed.

[veluca93]

If I had to bet, I'd say this is rng-related.

Fortunately, RobustStatistics is not used a lot - I think we might get away with removing the test and the tested code...

I'll look into it.