CHANGELOG.md

###v0.1 (targets)

• TODO: OMP parallelization for C vector operations
• TODO: documentation
• TODO: continuous integration
• TODO: Cython bindings
• TODO: operator typesafe equilibration
• TODO: operator IO

###v0.0.5 (next release)

• TODO: license
• TODO: cite POGS
• TODO: append version numbers to .so, check version numbers when loading libs in python
• TODO: add tests for vector_rand, transformable operators, operator export/import
• TODO: improve modularity in Makefile
• TODO: sparse POGS (entails sparse CG, equil, projection: performance vs. abstract??)

###v0.0.4 (current)

• Migrate tests to unittests
• Eliminate slow Python implementations (to be replaced with Cython implementations in future, potentially)
• Abstract linear operators
• Operator implementations: dense, sparse, diagonal
• CGLS, preconditioned CG methods for abstract operators
• Pperator equilibration for: dense, sparse
• Approximate projection (with CGLS) for abstract linear operators
• Operator POGS for: dense, sparse
• Matrix, vector reductions (min, max, indmin);
• Memory cleanup when tests fail
• K-means clustering on rows of dense matrices
• Error checking throughout C libraries (all calls that don't return a pointer return an error code)

###v0.0.3

• CPU Sparse Linear Algebra
• GPU Sparse Linear Algebra
• Python Sparse Linear Algebra bindings
• Separate FunctionVector (python implementation of POGS) from Objective (python wrapper for C implementation of POGS)
• Python backend counts allocations and calls cudaDeviceReset() when allocated blas handles and PogsSolver go out of scope (purpose: for scripts creating multiple PogsSolver instances, can delete solver objects to free GPU memory)
• Option to suppress [y, mu, nu] from POGS solver output (C and Py)
• Version number calls in C, Py

###v0.0.2

• C implementations of equilibration, projection and POGS (dense only)

###v0.0.1

• CPU/GPU dense linear algebra (32/64-bit)
• CPU/GPU (separable) prox (32/64-bit)
• Sinkhorn-Knopp matrix equilibration (see http://arxiv.org/abs/1503.08366)
• Direct (dense) graph projection
• Dense, fully-separable implementation of POGS block splitting algorithm (see http://foges.github.io/pogs/ and http://arxiv.org/abs/1503.08366)
• I/O for saving solver state (equilbrated matrix, projector factorization, iterates)
• Backend switching

#ROADMAP: ###v0.2

• CPU EVD

###v0.2.1

• GPU EVD

###v0.3

• CPU cone prox
• CPU block prox

###v0.3.1

• GPU cone prox
• GPU block prox

###v1.0

• Cone solver

###v++

• adaptive alpha (line search) for POGS
• DCP ADMM
• installation manifest (text or yaml) for clean uninstall
• operator l-BFGS preconditioner
• operator k-means