kodama-cpp API

kodama-cpp is a standalone C++17 library for float32 KODAMA optimization, cross-validated KNN and PLS-LDA, graph construction, PCA, UMAP, and openTSNE. The same typed core is called by the R and Python wrappers.

Start with the C++/R/Python walkthrough, then use the namespaced declarations in kodama/kodama.hpp as the authoritative C++ API. The R API map links wrapper names to their C++ counterparts, and the CPU coverage report states the measured scope and gaps.

Backend contract

• Backend::CPU runs package-owned CPU code.
• Backend::CUDA requires a CUDA-enabled build and reports an error if CUDA is unavailable.
• Backend::Metal requires a Metal-enabled macOS build and reports an error if Metal is unavailable.

Accelerator requests never silently fall back to CPU. Analysis matrices and intermediate numerical buffers are float32 inside the core. Reference labels are not inputs to KODAMA optimization; they belong only in downstream diagnostics.

Main entry points

• KNNCV and PLSLDACV: classifier-level cross-validation.
• CoreKNN and CorePLSLDA: one label-vector optimization.
• KODAMAMatrix: complete independent-run ensemble and graph correction.
• KODAMAMatrixFromGraph: alternative input when only a KNN graph is available.
• BuildResidentIVFIndex and SearchResidentIVFIndex: explicit CUDA/Metal ownership for repeated searches without rebuilding the IVF index.
• KODAMAKNNGraph, PCA, KODAMAUMAP_CPU/CUDA/METAL, and KODAMAOpenTSNE_CPU/CUDA/METAL: graph and visualization primitives.