Skip to content

TODO.md

Latest commit

 

History

History
95 lines (93 loc) · 3.73 KB

TODO.md

File metadata and controls

95 lines (93 loc) · 3.73 KB

Primary

  • Legacy arma::field vs typed structs
    • dual return system across frontends
    • inconsistent return types
  • Document C++ interface
  • GPU backend (see context/GPU_BACKEND_PLAN.md)

Secondary

  • Compile to Windows x86
  • Add formal test infrastructure (test/ directory)

Done

  • Automate link SuiteSparse on unix/macos
  • Upgrade Armadillo
  • Remove obsolete and broken igraph.
    • Fails to compile on ARM and newer x86.
  • Remove obsolete libleidenalg
    • Fails to compile because of above. Unnecessarily complicated dependency.
  • Compile to Unix x86
  • Compile to Apple arm64 native
  • Compile to Apple x86 via Rosetta2
  • Remove Harmony
  • Create "utils" module
  • Duplicate JSD functions (wtf?) in "network_construction" and hnsw (space_js.h)
    • Moved to "network_construction_ext"
  • Replace and remove inlined "ParallelFor" in build_network.
  • Update StatsLib
  • Fix duplicate PCG headers
  • Fix threading (RcppThread, mini_thread, inline, OpenMP)
    • OpenMP is now the threading model.
  • Rename .cc/.cpp and .h/.hpp
  • Fix defaults (Source -> header)
  • Fix namespace usage in headers
  • Consistent header guards
  • Rcpp wrappers
    • Condensed and modularized.
  • Update hnswlib
    • Fixed redefinition bug
  • Condense redundant SVD functions
    • New interface for SVD
  • Automatically link R BLAS/LAPACK
  • Removed packaged cblas.h
    • Automatically finds headers used by installed BLAS/LAPACK
  • Completely automate build system.
    • Fully portable and cross-platform
    • Automatically detects if built by R or stand-alone
    • Uses R build system in R package mode (99% CRAN compliant)
  • Select armadillo based on build mode
    • Compiles with RcppArmadillo in R build mode and packaged arma in stand-alone.
  • Automatically find BLAS/LAPACK
    • System BLAS (Linux), OpenBLAS, Accelerate, MKL supported.
    • Prefers BLAS used by R if in R build mode.
  • Restructured R wrappers
    • Added config for wrappers in R build mode
    • Conversion from arma::vec to Rcpp::NumricVector now automatic
  • Template abstraction of functions accepting both dense and sparse matrix input.
  • Add hnsw parameters to interfaces
  • Completely remade UMAP/uwot integration.
    • Updated uwot to 0.2.4
    • Modular graph optimization
  • Stand-alone color mapping i.e. de novo colors.
  • Consolidate and abstract network_diffusion
  • Fix various normalizations
    • Fast matrix normalization/manipulation interface for R/Python preprocessing
  • Detect Rosetta build (x86) outside R "CMAKE_OSX_ARCHITECTURES
  • Modular cmake
  • Standardize 'norm_type' vs. 'norm_method'
  • pybind 11 wrappers
  • Consolidate svdIRLB()
  • Backed SVD
    • Operator-based PRIMME, Halko, Feng, and IRLB paths
    • HDF5-backed sparse and dense matrix operators
    • Chunked I/O with configurable byte budget
  • Backed network construction
  • Backed specificity and annotation
    • Parallel backed specificity with thread hint
    • Overloads for BackedSparseMatrixOperator and BackedDenseMatrixOperator
    • Precomputed marker stats path for faster Python workflows
  • Speed up AA and SPA
    • Stacked C/H matrices during ACTION decomposition to reduce memory footprint
    • Release memory and move archetype matrices in runACTION
    • Conditionally parallel simplex regression
    • Optimized archetype merging
  • Reduce memory footprint
    • Eliminated unnecessary matrix copies across decomposition and annotation
    • Chebyshev diffusion with pointer rotation to avoid matrix moves
    • Removed excess graph copies in diffusion
    • Memory corruption fixes
  • Optimized kNN implementation
    • New graph builder with improved k*nn
  • Batch orthogonalization (operator-backed variant)
  • C++ified IRLB to fix memory leaks
  • Optimized pseudobulk/aggregation functions
  • Optimized specificity (analytical min-shift and single-pass label scatter)
  • Approximate fastlog transform for backed operators