Struct CosineJoiner 

pub struct CosineJoiner { /* private fields */ }

A reusable cosine-join handle that owns the corpus and — under feature = "gpu" on macOS — the Metal device, compiled batch_cosine kernel, and the corpus CSR uploaded to unified memory, all acquired once at construction. Repeated joins at different thresholds then skip the per-call kernel compile + CSR upload that cosine_join_with pays (only the t-specific inverted index is rebuilt each call, on the CPU). Always constructible; degrades to the pure-CPU join when the gpu feature is off or no Metal device is available — mirroring Rationer. This is the right entry point for sweeping thresholds or joining repeatedly.

Implementations

impl CosineJoiner

pub fn new(corpus: Corpus) -> Self

Build a joiner over corpus, acquiring the GPU device + uploading the corpus CSR once if the gpu feature is on and a Metal device is present.

pub fn corpus(&self) -> &Corpus

The owned corpus (e.g. for len() or to run other queries against it).

pub fn has_gpu(&self) -> bool

Whether a Metal GPU backend was acquired. Always false without feature = "gpu" on macOS; when false, every join runs on the CPU regardless of mode.

pub fn join(&self, t: f64, mode: Concurrency) -> Vec<(usize, usize, f64)>

Run the join at threshold t under mode, reusing the handle’s GPU resources. Returns the same results as cosine_join_with (Cpu/GpuPlusCpu exact, Gpu f32→f64); falls back to the CPU join when the GPU is unavailable.