quill-fastsort-numa

NUMA-aware parallel MSD-radix sort backend for quill-sort. Portable C++17 reimplementation of the original Windows-only compiled wheel, building natively on macOS (arm64/x86_64), Linux (x86_64/aarch64) and Windows.

Provides the entry points quill._backends.NumaBackend (backend name numa) probes for:

numa_sort_i64  numa_sort_u64  numa_sort_i32
numa_sort_u32  numa_sort_f64  numa_sort_f32

Each sorts a contiguous, writable numpy buffer of the matching dtype in place, ascending. The kernel is quillcore::parallel_radix — a top-byte MSD partition across a thread pool followed by an in-bucket LSD radix, with no merge step. On a multi-socket Linux host the backend is NUMA-aware; on single-socket machines (and off Linux) it falls back to plain parallel radix. The extra detect_topology() helper returns (nodes, cores) on multi-socket Linux and None otherwise.

• Never-lose: results are identical to np.sort on NaN-free numeric data (verified exhaustively against np.sort across dtypes, sizes and adversarial distributions). NaN is stripped by the Quill dispatcher before any kernel runs.
• Speed: measured 3–11× over single-threaded std::sort for int64 on 8 cores (scales with core count and array size).

Build

pip install .            # needs only a C++17 compiler

The sort kernels are header-only (src/quill_core.hpp, a synced copy of companions/_core/quill_core.hpp — run companions/sync_core.sh after editing the canonical core). No numpy headers or third-party libraries are required.