Hydro_InitExcision sphere_pugh_ppm test fails

Issue #1911 open
Barry Wardell created an issue

The Hydro_InitExcision sphere_pugh_ppm test fails for me when run on 1 process on an Ubuntu 16.04 machine. The diffs are attached.

Keyword:

Comments (20)

  1. Frank Löffler
    • removed comment

    Do you use the standard Ubuntu configuration out of Simfactory? Can you try the debian configuration? The main difference I see could be that the Ubuntu configuration uses the -ffast-math option, while the Debian doesn't.

  2. Barry Wardell reporter
    • removed comment

    Yes, this is with the ubuntu.cfg currently in Simfactory. I will try the Debian configuration or with -ffast-math removed.

  3. Barry Wardell reporter
    • changed status to resolved
    • removed comment

    I can confirm that the test passes (as do other failures in NaNChecker.nancount) when I remove -ffast-math from ubuntu.cfg.

  4. Erik Schnetter
    • changed status to open
    • removed comment

    It should not be necessary to avoid -ffast-math to make the tests pass. In Cactus, we have always taken the approach that the order in which floating-point expressions are evaluated should not matter, and that we do not want to rely on IEEE semantics when it comes to NaN (or similar exceptional values).

    While it would be nice to use this part of the standard, it (a) slows things down considerably, and (b) the standard doesn't guarantee bitwise identical results anyway.

  5. Frank Löffler
    • removed comment

    Test suites containing nans have to be avoided then (a good idea anyway), because -ffast-math is not guaranteed to produce them. The only test that I could find using a quick search that contains nans is CT_MultiLevel/test/boostedpuncture.

  6. Frank Löffler
    • removed comment

    Non concerning nans: the absolute differences in the velocity here are larger than 10e-7. This is usually considered too large for double-precision hydro simulations. Only looking at the Cactus-diff-output I cannot judge why that is, but I suspect these values are supposed to be zero, or close to it, but -ffast-math seems to produce a considerably larger error. This is a problem.

    I don't think we are likely to find a workaround for this in time for the release (code that works better even with fast math). I would suggest to remove --fast-math from the Ubuntu option list for the time being.

  7. Frank Löffler
    • removed comment

    Replying to [comment:5 knarf]:

    The only test that I could find using a quick search that contains nans is CT_MultiLevel/test/boostedpuncture.

    #1909

  8. Erik Schnetter
    • removed comment

    Compiler behaviour is not OS-dependent -- this is probably triggered by a combination of compiler version, compiler flags, and CPU capabilities. I argue against remove -ffast-math from the options; instead, we should either have suitably stable tests, or sufficiently large tolerances.

  9. Barry Wardell reporter
    • removed comment

    Replying to [comment:5 knarf]:

    Test suites containing nans have to be avoided then (a good idea anyway), because -ffast-math is not guaranteed to produce them. The only test that I could find using a quick search that contains nans is CT_MultiLevel/test/boostedpuncture.

    The NaNChecker.nancount test is also expected to produce NaNs, and fails with ubuntu.cfg when -ffast-math is enabled. This is a somewhat different case, however, as the whole point of the test is to check for NaNs.

  10. Ian Hinder
    • removed comment

    This test might fail with -march=native and depend on the CPU being used. We might see failures when we test on multiple machines for the release. Leaving this ticket open for now.

  11. Log in to comment