Skip to content

README.md

Latest commit

 

History

History
80 lines (52 loc) · 5.75 KB

README.md

File metadata and controls

80 lines (52 loc) · 5.75 KB
Raw

simulators

Run-time profiling for the PyAutoLens simulators — the scripts that produce mock lensed datasets (imaging, interferometer, point-source, cluster-scale, group-scale, multi-wavelength) used as inputs to fitting pipelines. These scripts profile simulation cost, not likelihood cost (which lives under likelihood_runtime/ / likelihood_breakdown/). They time grid construction, ray-tracing, PSF convolution, point-solver work, FITS / JSON output, and whatever else dominates each science case.

Why profile the simulators?

Slow simulators block iteration. The user reported the cluster-scale simulator as the original headliner: JIT-wrapped solver.solve was supposed to make it fast, but didn't, so the per-phase breakdown surfaced exactly which step was the bottleneck. Each script here applies the same per-phase timer pattern.

Scripts

Script What it profiles Distinguishing notes
imaging.py Single-band imaging simulator (HST / Euclid / JWST / AO presets). Grid + over-sampling setup, tracer + galaxies, tracer.image_2d_from (eager + JIT), simulator.via_tracer_from (NumPy convolution path), output. Cheapest — galaxy-scale, single band. Useful baseline.
interferometer.py Visibility-space simulator with synthetic uv-wavelengths (100 baselines inline, no disk file). DFT transformer setup, then tracer.image_2d_from eager + JIT. No PSF-convolution-on-image path; the FFT is in the transformer.
point_source.py Point-source simulator: lens + source tracer, eager solver.solve, JIT-compiled solver.solve (xp=jnp, remove_infinities=False), tracer.time_delays_from. Internal eager vs JIT solver assertion — surfaces the same upstream PointSolver drift tracked in PyAutoLens#514. Smallest grid (200×200 @ 0.05″/px).
cluster.py Cluster-scale simulator — the original headliner. 800×800 grid @ 0.1″/px, multiple source galaxies, per-source JIT-compiled solver.solve with compile cost reported per source. Heaviest — multiple JIT compiles, large grid.
group.py Group-scale imaging simulator. 250×250 grid with adaptive over-sampling at 3 galaxy centres, 1 main lens + 2 extra galaxies + source. Mid-weight; tests adaptive over-sampling cost.
multi.py Multi-wavelength (two-band: g + r) imaging simulator. Per-band grid setup with adaptive over-sampling, two tracers, per-band convolution. Tests parallel-band orchestration cost.

What each script reports

Per-phase wall-time table printed to stdout, plus a structured JSON written to results/simulators/<script>_summary_v<al.__version__>.json carrying:

  • Phase timings (one entry per timer.section("...") block).
  • Key derived timings (e.g. solver_solve_eager_s, via_tracer_from_s).
  • Configuration metadata (grid shape, pixel scale, model summary).
  • al.__version__ tag for cross-release comparison.

A matching .png bar chart of the per-phase costs lands at the same path. Phase 4's dashboard reads these JSONs to refresh the run-times table in this README.

Versioned artifacts

Filename convention:

results/simulators/<script>_summary_v<al.__version__>.{json,png}

Old versions are retained alongside new ones so cross-release trends stay visible.

Headline run-times (latest per script)

Auto-generated by scripts/build_readme.py from the latest *_summary_v<version>.json artifacts under results/simulators/. Cell value is the total wall time of the simulator run-to-completion (sum of all timer.section phases), not per-likelihood. Hardware tier is CPU only today.

Script Total wall time PyAutoLens version
imaging_euclid.py 2.28 s v2026.7.6.649

Running a script

From the repo root:

cd autolens_profiling
python simulators/imaging.py

Each script defaults to dataset_name = "simple" and writes the produced dataset to dataset/<type>/<name>/.

Note: dataset/point_source/simple/point_dataset_positions_only.json and dataset/point_source/simple/tracer.json are tracked in this repo as Phase 1 likelihood inputs; running simulators/point_source.py without changing the default dataset_name will overwrite them. Use a non-conflicting dataset_name (e.g. "smoke") when smoking these scripts, or commit the regenerated bytes if the regeneration is intentional.

Other dataset/<type>/{cluster,group,multi}/ and */simple/ variants are gitignored, so writing to them does not pollute the repo.

Codex / sandboxed runs:

NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib python simulators/imaging.py

JIT readiness

The prompt for Phase 2 flagged that "JAX jitting for these are not fully implemented everywhere so we may have placeholders". Status today:

  • point_source.py — JIT path for solver.solve exercised and asserted against eager. (The upstream drift in this assertion is tracked under PyAutoLens#514.)
  • imaging.py, interferometer.py, group.py, multi.py — tracer.image_2d_from JIT path exercised; simulator.via_tracer_from (PSF convolution) is the NumPy path and is the expected bottleneck.
  • cluster.py — per-source JIT-compiled solver.solve exercised; per-source compile cost reported.

Where the JIT path isn't yet implemented for a given simulator step, the script measures the NumPy path only; future upstream JIT support will be reported in this section as it lands.