Galaxy Angular Correlation (CUDA)

CUDA kernels for the 2‑point angular correlation function, comparing real and random galaxy catalogs.

Overview

• Computes DD, DR, and RR histograms of pair angles, then evaluates the Landy–Szalay estimator w(θ) = (DD - 2·DR + RR) / RR.
• Multiple CUDA implementations: shared-memory optimized, no-cache baseline, and a manual/experimental variant.
• Organized layout with Makefile builds, sample data, and output folders for reproducible runs.

Repository Layout

• src/galaxy_shared_mem.cu — shared-memory kernel (90° range, 1024 threads/block).
• src/galaxy_no_cache.cu — global-memory baseline (180° range).
• src/galaxy_manual.cu — experimental/manual tuning.
• src/template.cu — scaffold for new kernels.
• data/real.dat, data/synthetic.dat — sample catalogs; first line = count, subsequent lines: ra dec in arcminutes.
• bin/ — compiled binaries.
• outputs/ — place run artifacts and plots.
• docs/ — long-form notes, profiling reports, figures.

Requirements

• NVIDIA CUDA toolkit (nvcc on PATH).
• NVIDIA GPU; set ARCH in the Makefile to your compute capability (e.g., sm_86).
• Bash-compatible shell on Linux/macOS (adjust commands for other platforms).

Build

make                  # builds all targets into bin/
# override defaults if needed:
# make ARCH=sm_86 NVCCFLAGS="-O3 -lineinfo"

Targets produced: galaxy_shared_mem, galaxy_no_cache, galaxy_manual.

Run

./bin/galaxy_shared_mem data/real.dat data/synthetic.dat outputs/shared.out
./bin/galaxy_no_cache   data/real.dat data/synthetic.dat outputs/no_cache.out
./bin/galaxy_manual     data/real.dat data/synthetic.dat outputs/manual.out

• Inputs are arcminutes. In code: ra_rad = ra_arcmin * (π / 10800), dec_rad = dec_arcmin * (π / 10800).
• Outputs are per-bin counts; post-process to derive w(θ) or plot histograms.

Data Format

N
ra_0 dec_0
ra_1 dec_1
...

• First line N = number of entries.
• Right Ascension and Declination are arcminutes; conversion to radians occurs in the code.

Implementation Notes

• Bin width is 0.25° (binsperdegree = 4). totaldegrees controls the angular extent (90° in shared-memory variant, 180° in baseline).
• galaxy_shared_mem.cu: tiles coordinates into shared memory, accumulates per-block histograms, then atomically merges.
• galaxy_no_cache.cu: simple global-memory approach using unified memory for histograms.
• galaxy_manual.cu: sandbox for manual tuning or alternative strategies.

Profiling & Best Practices

• Use nvprof, nsys profile, or ncu to inspect occupancy, shared-memory usage, and atomic contention.
• Tune threadsperblock for your GPU; ensure ARCH matches the device.
• For larger catalogs, consider hierarchical binning, pair subsampling, or widening bins to reduce atomic pressure.
• Keep outputs and plots under outputs/; store profiling notes in docs/.

Contributing

• Add new kernels in src/ and hook them into the Makefile pattern rule.
• Favor deterministic reductions when comparing variants.
• Submit changes under MIT License (see LICENSE).

License

MIT License. See LICENSE for details.