[#179] Make compatible with IterativeSolvers v0.9+ (#180)

* [#179] Make compatible with IterativeSolvers v0.9+

- Fix a breaking change in IterativeSolvers API
- Add information to docs about which version of IterativeSolvers to use
- Add version constraint to Project.toml

* Move CI test from Julia v1.0 to next LTS v1.6

.github/workflows/ci.yml

@@ -13,7 +13,7 @@ jobs:
           - version: '1'
             os: ubuntu-latest
             arch: x64
-          - version: '1.0'
+          - version: '1.6'
             os: ubuntu-latest
             arch: x64
           - version: '1'

Project.toml

@@ -27,6 +27,7 @@ AMD = "0.4, 0.5"
 COSMOAccelerators = "^0.1.0"
 DataStructures = "^0.17.0, ^0.18.0"
 IterTools = "^1"
+IterativeSolvers = "^0.9"
 MathOptInterface = "~0.10.7, 1"
 QDLDL = "0.4.1"
 Reexport = "0.2, ^1"

docs/src/lin_solver.md

@@ -17,7 +17,7 @@ CGIndirectKKTSolver | IterativeSolvers.jl | Conjugate Gradients on the reduced K
 MINRESIndirectKKTSolver | IterativeSolvers.jl | MINRES on the (full) KKT linear system.
 
 !!! note
-    To use the Pardiso and Intel MKL Pardiso solver, you have to install the respective libraries and the corresponding Julia wrapper. For more information about installing these, visit the [Pardiso.jl](https://github.com/JuliaSparse/Pardiso.jl) repository page. Likewise in order to use Indirect(Reduced)KKTSolver you have to install [IterativeSolvers.jl](https://github.com/JuliaMath/IterativeSolvers.jl) and [LinearMaps.jl](https://github.com/Jutho/LinearMaps.jl). We are using the `Requires` package for lazy loading of code related to `Pardiso` and `IterativeSolvers`. This means in order to use `Pardiso` / `IterativeSolvers`, you'll have to load these packages alongside `COSMO`, i.e. `using Pardiso` and `using IterativeSolvers, LinearMaps`.
+    To use the Pardiso and Intel MKL Pardiso solver, you have to install the respective libraries and the corresponding Julia wrapper. For more information about installing these, visit the [Pardiso.jl](https://github.com/JuliaSparse/Pardiso.jl) repository page. Likewise in order to use Indirect(Reduced)KKTSolver you have to install [IterativeSolvers.jl](https://github.com/JuliaMath/IterativeSolvers.jl) (v0.9+) and [LinearMaps.jl](https://github.com/Jutho/LinearMaps.jl). We are using the `Requires` package for lazy loading of code related to `Pardiso` and `IterativeSolvers`. This means in order to use `Pardiso` / `IterativeSolvers`, you'll have to load these packages alongside `COSMO`, i.e. `using Pardiso` and `using IterativeSolvers, LinearMaps`.
 
 COSMO uses the `Cholmod` linear system solver by default. You can specify a different solver in the settings by using the `kkt_solver` keyword and the respective type:
 

src/linear_solver/kktsolver_indirect.jl

@@ -67,10 +67,10 @@ function solve!(S::IndirectReducedKKTSolver, y::AbstractVector{T}, x::AbstractVe
     end
     L = LinearMap(reduced_mul!, S.n; ismutating=true, issymmetric=true)
     if S.solver_type == :CG
-        cg!(S.previous_solution, L, y1, tol=get_tolerance(S)/norm(y1))
+        cg!(S.previous_solution, L, y1, abstol=get_tolerance(S)/norm(y1), reltol = 0.)
     elseif S.solver_type == :MINRES
         init_residual = norm(L*S.previous_solution - y1)
-        minres!(S.previous_solution, L, y1, tol=get_tolerance(S)/init_residual)
+        minres!(S.previous_solution, L, y1, abstol=get_tolerance(S)/init_residual, reltol = 0.)
     end
     # Sanity check for tolerance
     # might not always hold for MINRES, as its termination criterion is approximate, (see its documentation)
@@ -149,7 +149,7 @@ function solve!(S::IndirectKKTSolver, y::AbstractVector{T}, x::AbstractVector{T}
     L = LinearMap(kkt_mul!, S.n + S.m; ismutating=true, issymmetric=true)
     if S.solver_type == :MINRES
         init_residual = norm(L*S.previous_solution - x)
-        minres!(S.previous_solution, L, x, tol=get_tolerance(S)/init_residual)
+        minres!(S.previous_solution, L, x, abstol=get_tolerance(S)/init_residual, reltol = 0.)
     end
     # Sanity check for tolerance
     # might not always hold for MINRES, as its termination criterion is approximate, (see its documentation)