nqt o2

[Yurlungur]

PR Summary

This threads in the second order NQT method developed by Peter Hammond on the AthenaK team. More details soon as I develop them. Also there will be an archive note eventually.

PR Checklist

• Adds a test for any bugs fixed. Adds tests for new features.
• Format your changes by using the make format command after configuring with cmake.
• Document any new features, update documentation for changes made.
• Make sure the copyright notice on any files you modified is up to date.
• After creating a pull request, note it in the CHANGELOG.md file.
• LANL employees: make sure tests pass both on the github CI and on the Darwin CI

If preparing for a new release, in addition please check the following:

• Update the version in cmake.
• Move the changes in the CHANGELOG.md file under a new header for the new release, and reset the categories.
• Ensure that any when='@main' dependencies are updated to the release version in the package.py

[mauneyc-LANL]

Apologies if there's code I'm not seeing that addresses this, but do we check if the precision of the typedef Real is double before including this? Is there equivalent 'float' procedures?

[Yurlungur]

There are not equivalent float procedures at this time. Nor are there little endians. However all functions in this code are explicitly doubles.

You're worried about building for single-precision I assume. Yes, we should be worried about that. Let's extend that in a later PR, though. Getting this working was enough of a lift.

[mauneyc-LANL]

Comments for discussion rather than blocking requests.

[Yurlungur]

Timing results

x86-volta

NQT o1

                ...L2 difference = 0.0254583
                ...EOSPAC time/point (microseconds) = 0.493015
                ...spiner host time/point (microseconds) = 0.0335447
                ...spiner device time/point (microseconds) = 0.00160943

NQT o2

                ...L2 difference = 0.00807353
                ...EOSPAC time/point (microseconds) = 0.491144
                ...spiner host time/point (microseconds) = 0.033117
                ...spiner device time/point (microseconds) = 0.00142899

NQT true

                ...L2 difference = 0.0171703
                ...EOSPAC time/point (microseconds) = 0.491845
                ...spiner host time/point (microseconds) = 0.0619208
                ...spiner device time/point (microseconds) = 0.00149041

grace-hopper

NQT o1

                ...L2 difference = 0.0254583
                ...EOSPAC time/point (microseconds) = 0.327759
                ...spiner host time/point (microseconds) = 0.0110745
                ...spiner device time/point (microseconds) = 0.000517656

NQT o2

                ...L2 difference = 0.00807353
                ...EOSPAC time/point (microseconds) = 0.302807
                ...spiner host time/point (microseconds) = 0.0137009
                ...spiner device time/point (microseconds) = 0.000571444

NQT true

                ...L2 difference = 0.0171703
                ...EOSPAC time/point (microseconds) = 0.326109
                ...spiner host time/point (microseconds) = 0.0203759
                ...spiner device time/point (microseconds) = 0.000564959

[Yurlungur]

Pointwise error

[Yurlungur]

Timings for the stellar collapse reader in nanoseconds/point

jonahm@sgrb:~/simulations/spiner$ tail -n 2 timings_SFHo_o1_x86_volta.txt 
host = 22.3432
device = 0.641151
jonahm@sgrb:~/simulations/spiner$ tail -n 2 timings_SFHo_o2_x86_volta.txt 
host = 31.665
device = 0.612225
jonahm@sgrb:~/simulations/spiner$ tail -n 2 timings_SFHo_true_x86_volta.txt 
host = 77.0338
device = 0.665203
jonahm@sgrb:~/simulations/spiner$ tail -n 2 timings_SFHo_o1_grace_hopper.txt 
host = 9.79919
device = 0.443848
jonahm@sgrb:~/simulations/spiner$ tail -n 2 timings_SFHo_o2_grace_hopper.txt 
host = 15.484
device = 0.453125
jonahm@sgrb:~/simulations/spiner$ tail -n 2 timings_SFHo_true_grace_hopper.txt 
host = 22.6683
device = 0.443237

so on GPUs, no difference. On grace, o2/true speedup is ~25% and on x86, its ~2x.

[Yurlungur]

Error plots

[Yurlungur]

@jhp-lanl @dholladay00 @pdmullen @jdolence this is ready for review

[Yurlungur]

New example showing how to profile. Can be set to profile any EOS, even analytic.

[Yurlungur]

The change to this documentation is where the core of the method is. Essentially the quadratic interpolation of $\lg(mantissa)$ on the interval $[0.5, 1)$ has the advantage that the NQT function is C1 everywhere, which improves convergence of linear interpolation. The function itself is still a 100% accurate version of itself, but improving the continuity class of the method improves convergence of stencil ops.

[Yurlungur]

Tests passing on re-git. Any objections to merging this? @AstroBarker lets discuss phoebus when you get the chance.