GalaxyGlasses.jl

GalaxyGlasses.jl implements the numerical forward model used to generate SIDES-style millimeter-source catalogs. Given galaxy properties, model parameters, and explicit random draws, it computes:

• quenched and starburst classifications and star-formation rates;
• gravitational magnification;
• dust SED quantities, monochromatic fluxes, and optional filter fluxes;
• CO, [CII], and [CI] line luminosities and fluxes.

The core entry point is BayesMMfwd.forward_model(inputs, params, noise). It is pure numerical Julia: file loading, random-number generation, backend selection, compilation, and catalog output remain outside the function. The same forward-model source therefore runs as ordinary Julia or compiles with Reactant for CPU, NVIDIA GPU, and TPU execution.

Setup

BayesMMfwd currently targets Julia 1.11. From the repository root:

julia --project=. -e 'using Pkg; Pkg.instantiate()'

The paths in data/SIDES_from_original.par must resolve to the model tables:

• SED_finegrid_dict.h5
• LFIR_LIR_ratio.h5
• data/Psupmu_table_Bethermin17.txt
• data/Daddi15_SLED.txt
• a SIDES CSV catalog

Input catalogs must contain redshift, Mstar, Mhalo, ra, and dec. Filter-grid HDF5 files are required only when filter fluxes are enabled.

Forward-model API

using BayesMMfwd

configuration = BayesMMfwd.load_par_file("data/SIDES_from_original.par")
catalog = BayesMMfwd.load_sides_csv("data/SIDES_Bethermin2017_short2.csv")

inputs = BayesMMfwd.build_forward_inputs(catalog)
parameters = BayesMMfwd.build_forward_parameters(configuration; filters=false)
noise = BayesMMfwd.make_simulation_noise(length(inputs.redshift))

output = BayesMMfwd.forward_model(inputs, parameters.numeric, noise)

The explicit noise argument makes a run reproducible and lets all backends consume identical random draws. The result is a named tuple of arrays; use BayesMMfwd.add_output_columns! to attach those arrays to a catalog.

The Reactant path changes data placement and compilation, not the model:

using Reactant

reactant_inputs = Reactant.to_rarray(inputs)
reactant_parameters = Reactant.to_rarray(parameters.numeric)
reactant_noise = Reactant.to_rarray(noise)

compiled_forward = Reactant.@compile sync=true BayesMMfwd.forward_model(
    reactant_inputs,
    reactant_parameters,
    reactant_noise,
)
reactant_output = compiled_forward(
    reactant_inputs,
    reactant_parameters,
    reactant_noise,
)

Production driver

gen_pysides_from_original.jl is a small Reactant driver around the forward model. CPU is the default backend:

julia --project=. gen_pysides_from_original.jl
julia --project=. gen_pysides_from_original.jl --backend CUDA --large
julia --project=. gen_pysides_from_original.jl --backend TPU --large

It reports transfer, compilation, execution, materialization, and total time. Add --write-output to write the configured FITS catalog, --rows N to limit a run, or --filters to include configured filter fluxes. Run with --help for all options.

Performance and reproducibility

The Julia-versus-Reactant harness is isolated under benchmark/. It records machine-readable timing, allocation, overhead, provenance, and correctness data.

Reproducing the paper figures

Run every command below from the BayesMMfwd.jl repository root. Raw benchmark JSON is read from benchmark/results/, and generated paper figures, LaTeX table fragments, summaries, and handoff notes are written to benchmark/plots/generated/. Figure 1 is the implementation architecture authored in the manuscript's LaTeX/TikZ source; this repository does not generate a fig1 file.

Instantiate the isolated benchmark environment once:

julia --project=benchmark -e 'using Pkg; Pkg.instantiate()'

Generate every paper artifact from the canonical raw directory with:

julia --startup-file=no --history-file=no --project=benchmark \
  benchmark/plots/plot_results.jl \
  --results-dir benchmark/results \
  --output-dir benchmark/plots/generated

Paper item	Generated artifact	Description
Figure 1	Not generated (LaTeX/TikZ)	BayesMMfwd implementation architecture
Figure 2	benchmark/plots/generated/fig2_runtime_speedup.pdf	Warm runtime and speedup
Figure 3	benchmark/plots/generated/fig3_overheads_amortization.pdf	Measured overheads and derived amortization
Figure 4	benchmark/plots/generated/fig4_memory_scaling.pdf	GPU peak memory and host allocation traffic
Table II	benchmark/plots/generated/table2_largest_common.tex	1,000,000-source comparison
Table III	benchmark/plots/generated/table3_correctness.tex	Julia/Reactant numerical agreement

The complete paper matrix consists of these eight commands. Every command writes uniquely named sibling JSON files directly into the same canonical raw directory:

julia --startup-file=no --history-file=no --project=benchmark benchmark/benchmark_reactant.jl --large --rows 1024 --backend CPU --results-dir benchmark/results
julia --startup-file=no --history-file=no --project=benchmark benchmark/benchmark_reactant.jl --large --rows 100000 --backend CPU --results-dir benchmark/results
julia --startup-file=no --history-file=no --project=benchmark benchmark/benchmark_reactant.jl --large --rows 1000000 --backend CPU --results-dir benchmark/results
julia --startup-file=no --history-file=no --project=benchmark benchmark/benchmark_reactant.jl --large --rows 5584998 --backend CPU --results-dir benchmark/results
julia --startup-file=no --history-file=no --project=benchmark benchmark/benchmark_reactant.jl --large --rows 1024 --backend CUDA --results-dir benchmark/results
julia --startup-file=no --history-file=no --project=benchmark benchmark/benchmark_reactant.jl --large --rows 100000 --backend CUDA --results-dir benchmark/results
julia --startup-file=no --history-file=no --project=benchmark benchmark/benchmark_reactant.jl --large --rows 1000000 --backend CUDA --results-dir benchmark/results
julia --startup-file=no --history-file=no --project=benchmark benchmark/benchmark_reactant.jl --large --rows 5584998 --backend CUDA --results-dir benchmark/results

Run each command once. Each invocation relies on the harness's internal Chairmarks sampling for Julia and Reactant CPU or its synchronized 25-execution XProf sampling for Reactant GPU; no external replication loop is needed. Filters remain disabled, explicit noise uses seed 1234, ordinary Julia uses one Julia thread, and Reactant CPU retains XLA's default CPU parallelization.

The 5,584,998-source GPU case may exhaust the usable memory of the selected RTX 5090. A recorded OOM is a valid reproduced outcome and is shown explicitly in the generated artifacts; it is not an artifact-generation failure.

Expected results: Reactant CPU should cross over after the smallest catalog, and the single GPU should be fastest at the larger successfully measured sizes. These are qualitative checks only. Consult benchmark/plots/generated/results_summary.json for the exact recorded values, status of every point, correctness extrema, and break-even calculations.

Tests

julia --startup-file=no --history-file=no --project=. test/runtests.jl

The regression suite checks Julia/Reactant agreement for the complete forward model, input-table boundary cases, output assembly, and FITS writing.

Source layout

• src/reactant_pipeline.jl: backend-independent numerical forward model.
• src/pipeline_host.jl: input preparation and table loading.
• src/io.jl: parameter, catalog, and FITS I/O.
• data/: tracked model inputs and local, ignored catalogs.
• gen_pysides_from_original.jl: Reactant execution and optional FITS output.
• benchmark/: benchmark runner and result utilities.
• benchmark/plots/: plotting scripts.