experimental
Folders and files
Name | Name | Last commit date | ||
---|---|---|---|---|
parent directory.. | ||||
-------------------------------------------------------------------------------- LAGraph/experimental/README.txt -------------------------------------------------------------------------------- This experimental folder includes algorithms and utilities in various stages of completion. Their prototypes appear in LAGraph/include/LAGraphX.h. When new methods methods are developed for LAGraph, they are first placed in this folder, and then when they are finished and polised, they are moved into the src/ folder. -------------------------------------------------------------------------------- FUTURE WORK: FUTURE: LAGraph_SingleSourceShortestPath: Write a Basic method that computes G->emin (that's easy). And G->emax so it can decide on Delta on its own (hard to do). FUTURE: LAGraph_BreadthFirstSearch basic method that computes G->AT and G->out_degree first. FUTURE: all algorithms can use Basic and Advanced variants. FUTURE: file I/O with a common, binary, non-opaque format that all GraphBLAS libraries can read. Ideally using compression (see SuiteSparse:GraphBLAS uses LZ4 for its GrB_Matrix_serialize/ deserialize and it would not be hard to add LZ4 to LAGraph). We could add the format to the *.lagraph file format now used by the experimental/utility/LAGraph_S*.c methods. That is, the same file format *.lagraph could include matrices in either the universal format, or a library-specific format. The format would be specified by the json header in the *.lagraph file. See the doc/lagraph_format.txt file for details. It would require the LAGraph_SLoad, *SSave, etc to move from experimental to src, and those functions need work (API design, text compression). Adding this feature also requires an LAGraph_serialize/deserialize, which requies LZ4 compression. We could do this for v1.1. FUTURE: can we exploit asynch algorithms? C += A*C for example? Gauss-Seidel, Afforest, etc? I can do it in GraphBLAS; can LAGraph use it? FUTURE: for LG_CC_FastSV6.c: * need new GxB methods in GraphBLAS for CC, GxB_select with GxB_RankUnaryOp, and GxB_extract with GrB_Vectors as inputs instead of (GrB_Index *) arrays. FUTURE: add interfaces to external packages. GTgraph: (Madduri and Bader) for generating synthetic graphs CSparse or CXSparse (for depth-first search, scc, dmperm, amd, in the sequential case) graph partitioning: METIS, Mongoose, etc SuiteSparse solvers (UMFPACK, CHOLMOD, KLU, SPQR, ...) graph drawing methods others? FUTURE: interfaces to MATLAB, Python, Julia, etc. FUTURE: need more algorithms and utilities