v2024.04

[lukeshingles]

• Add VPKT_WRITE_CONTRIBS option to save all virtual packets contributions for later time and wavelength binning, with emission/absorption types by artistools.

• Improve performance by enabling fast-math compiler options, constexpr radfield bins, improving bound-free estimator accumulation, removing storage for empty cells, and using node shared memory for initmassfracstable, elem_meanweight.

  • 3D AD2 (Shingles et al. 2023) 1e8pkt 1920 core JUWELS runtime decreased from 210k core hrs to 134k core hrs (1.56x faster). Memory per core decreased from 2.8GB to 1.5GB.
  • W7 (Shingles et al. 2022) nebular 960 core 2e9pkt 150-410d JUWELS runtime decreased from 13.6k core hrs (20230811 version) to 10.4k core hrs (1.31x faster). Compared to the 2022 published version with 18.6k core hrs, this is 1.79x faster.
    • With a Simpson rule integrator in place of GSL's adaptive method (producing identical observables) runtime decreases to 6.83k core hours, 1.99x faster than 20230811 and 2.72x faster than the 2022 published version.
  • subMch Shen+2018 nebular 960 core 1e9pkt 150-410d runtime decreased from 13.14k core hrs (20210417 version) to 7.30 kCore hours (1.80x faster). With Simpson rule integrator, runtime decreased to 6.07 k core hrs (2.16x faster)
  • sim2010 classic mode 2-120 days 2e7pkt 960 core JUWELS runtime decreased from 3.81k core hrs (20230526 version) to 1.87k core hrs (2.04x faster). Compared to ARTIS classic with 11.45k core hrs, this version is 6.13x faster.

• Reduce memory per core to enable future GPU support. All threads now share a single cell cache and operate on packets within the same cell. This may later be expanded to multiple cell caches to increase thread occupancy as memory allows.

• Enable C++ standard parallelism for packet propagation (for future GPU offloading).

• Fix virtual packets for models with internal empty cells. Passing into an empty cell previously triggered instant escape.

• Fix a 1e5 factor in free-free opacity for opacity case 5 (Tanaka Ye-dependent grey mode and line-by-line non-grey mode for kilonovae). This increased the 3D AD2 (Shingles et al. 2023) early luminosity by about 20%. The spectra are almost unchanged (except for being brighter)

• Reduce noise in particle deposition rate estimators by contributing partial deposition of packet energy prior to Monte Carlo transition event (similar to gamma deposition rates from path estimators).

• Add support for calculating and using expansion opacities.

• Add option for XCOM gamma-ray photoionization #52 by Gerrit Leck

• Add option to limit number of BF estimators #47 by Josh Pollin

• Add some fixes to hybrid LTE-NLTE mode by Christine Collins