pysanepic 0.2.1__tar.gz

pysanepic-0.2.1/.gitignore

@@ -0,0 +1,4 @@
+__pycache__/
+*.egg-info/
+/sims/
+docs/_build/

pysanepic-0.2.1/CHANGELOG.md

@@ -0,0 +1,49 @@
+# Changelog
+
+## 0.2.1
+
+Packaging for PyPI: metadata, source distribution without the C++ code, and
+publication from GitHub releases. No code changes.
+
+## 0.2.0 (2026-10-14)
+
+First release: a standalone Python port of the SANEPIC GLS map-maker for
+correlated (1/f) noise, used by litebird_sim (`make_sanepic_maps`).
+
+### Map-making
+
+- GLS solution of $(P^T N^{-1} P)\,m = P^T N^{-1} d$ with a preconditioned
+  conjugate gradient; $N^{-1}$ applied per chunk as a Fourier filter (ducc0
+  FFTs), I or I/Q/U maps, optional HWP angle.
+- Block 3×3 preconditioner (Jacobi still available) and the Q/U
+  angle-coverage criterion `min_pol_rcond` (#2).
+- Chunk padding, as SANEPIC's `-m`/`-q` options (#4).
+- `GLSParameters` collects the solver options (#6).
+
+### Interface
+
+- `Pointing`: either angles (θ, φ, ψ in some coordinates), rotated and
+  pixelized by `make_maps` at the output nside and coordinates, or HEALPix
+  pixels already computed (RING or NESTED), used as they are; the arrays are
+  checked when the object is created. `DetectorData` takes a `Pointing` (#6,
+  #7, #12).
+- Low-level `GLS` class: `apply_A` and `rhs` work on the pixels of the MPI
+  rank, `GLS.full` gathers a per-pixel array (#14).
+
+### Parallelism and memory
+
+- MPI: the PCG state is distributed by pixel; the $P^T$ maps are summed with a
+  reduce-scatter and the search direction is rebuilt with an allgather. Maps
+  up to nside 4096 with MPI (#14).
+- Threads (`nthreads`, default `OMP_NUM_THREADS`) for the FFTs and the numba
+  kernels; $P^T$ shares one map among the threads, each owning a range of
+  pixels (#3, #5, #15).
+- 8 bytes per sample besides the TODs: int32 pixel and 2a in float32;
+  γ cos 2a and γ sin 2a recomputed for each chunk (#6, #8, #11).
+- Faster setup: counting sort by pixel and one irfft per weights array (#13);
+  cos and sin from a table, 10 times faster than libm without SVML (#16).
+
+### Documentation
+
+- Sphinx documentation on Read the Docs, with notebooks, MPI and memory
+  notes and benchmarks on G100 (#9, #17).