Skip to content

CHANGELOG.md

Latest commit

 

History

History
138 lines (125 loc) · 8.21 KB

CHANGELOG.md

File metadata and controls

138 lines (125 loc) · 8.21 KB
Raw

XpressAPI release notes

All notable changes to the XpressAPI Julia package are documented here. Version numbers track the FICO Xpress version from which the bindings were generated.

47.1.5

  • Fixed newly added JuMP / MOI variables silently defaulting to a lower bound of 0.0. MOI variables are free by default (-Inf, +Inf), but Xpress defaults a column's lower bound to 0.0, so a variable with no explicit bound was made non-negative -- causing wrong results or spurious infeasibilities for models relying on free variables. New variables are now created with an explicit -Inf lower bound.
  • Added in-place set(::Optimizer, ::MOI.ConstraintSet, ...) to the JuMP / MOI interface. Variable bounds (GreaterThan, LessThan, EqualTo, Interval) and affine constraint right-hand sides can now be modified without rebuilding the model, routing through XPRSchgbounds / XPRSchgrhs, so warm-start and re-solve loops pick up the new set on the next optimize!.
  • Added MOI.DualObjectiveValue to the JuMP / MOI interface. It reports the objective value of the dual solution after an LP solve, derived from the constraint duals so that it follows the MOI sign convention and, at optimality, matches MOI.ObjectiveValue within tolerance.
  • Fixed the MOI.ConstraintDual sign conventions in the JuMP / MOI interface. Duals are now reported relative to a minimization problem as MOI requires (negated for maximization models), so a GreaterThan row/bound has a non-negative dual and a LessThan one a non-positive dual regardless of the objective sense. Duals of variable bounds are now returned as the column reduced cost (via XPRSgetredcosts) instead of NaN, and the dual of a parent constraint (such as a split Interval bound) is taken from its binding child; the vector getter now returns a flat Vector rather than a malformed nested one.
  • Shipped MOI_FEATURES.md inside the installed package. The README links to it, but it was previously omitted from the install, leaving a dead link. The bridge-usage examples in that document were also corrected to create the model directly with Model(XpressAPI.Optimizer) (JuMP adds the required bridges automatically) rather than wrapping it in a MOI.Bridges.full_bridge_optimizer, and the README wording no longer claims the package provides no modeling API given the JuMP / MOI extension it ships.
  • Continued the low-level wrapper clean-up following upstream review feedback: string arguments and results now use the concrete String type (and name lists are returned as Vector{String}) instead of AbstractString, except for name-array inputs which stay generic so any string vector can be passed; wrapped functions emit an explicit return; and the docstring signature line now shows the return type rather than the output parameter names.
  • Fixed the "argument too short" error raised when a pre-allocated output array is smaller than required. The generated code concatenated the required length (an integer) directly onto the message string, which threw an opaque MethodError instead of the intended XPRSexception; the length is now converted with string(...).
  • Fixed an intermittent heap-corruption crash on Windows (an access violation during garbage collection) that could occur while a callback was active during a solve. Each callback invocation previously allocated a fresh problem wrapper to pass to the user callback; because callbacks run re-entrantly while the solver holds the call stack, this per-invocation allocation could trigger garbage collection at an unsafe point. The wrapper is now created once and reused across all invocations of a given callback, so callbacks no longer allocate it per call. Note: this reuse is only safe because MIP branch-and-bound callbacks are serialised on the main thread (XPRS_CALLBACKFROMMAINTHREAD). That guarantee does not hold for the SLP / nonlinear / multistart solves, where the optimizer may invoke callbacks from worker threads; callbacks are not currently safe to use during those solves.

47.1.4

  • Added MOI.SolverVersion to the JuMP / MOI interface. It reports the version of the Xpress library actually loaded at runtime (via XPRSgetversionnumbers), which may differ from the version the bindings were built against, as a major.minor.build string.
  • Added MOI.ListOfConstraintTypesPresent to the JuMP / MOI interface. It reports each (F, S) constraint-type tuple currently present in the model exactly once.
  • Added MOI.ObjectiveFunctionType to the JuMP / MOI interface. It reports the type of the objective function currently set (MOI.VariableIndex, MOI.ScalarAffineFunction, MOI.ScalarQuadraticFunction, or MOI.VectorOfVariables), defaulting to MOI.ScalarAffineFunction when no objective has been set.
  • Fixed the low-level wrapper for deferred-reference (dref) output-array functions such as XPRSgetqrowqmatrixtriplets when called with pre-allocated arrays (or nothing). The generated code previously skipped the C call on that path, returning the untouched input arrays and a size of 0; it now issues the call and returns the correct data and size.
  • Hardened the generated bindings against a garbage-collection use-after-free that could corrupt the heap (observed as intermittent access violations on Windows). String output buffers are now kept alive with GC.@preserve across the unsafe_string(pointer(...)) conversion, and callbacks are rooted in the problem's callback list before being registered with the C library.
  • Reported callback support accurately: MOI.supports now returns true only for MOI.UserCutCallback (the one implemented callback) and false for MOI.LazyConstraintCallback and MOI.HeuristicCallback, instead of a blanket true that made JuMP offer callbacks that then failed at runtime. The feature docs were corrected to match.
  • Cleaned up the generated low-level wrapper following upstream review feedback: module globals are now const or concretely typed (avoiding the untyped-global performance penalty); Libdl/SparseArrays are brought in with import and referenced qualified; the broken default argument on Base.showerror was removed; and redundant global/export statements were dropped (all XPRS-prefixed symbols are exported by a single loop).
  • Fixed the SparseMatrixCSC overload of XPRSaddcols, which passed a malformed map.nzval instead of the matrix's non-zero values.

47.1.3

  • Added warm-start support to the JuMP / MOI interface via set_start_value (MOI.VariablePrimalStart). The starting point is forwarded to the solver according to the problem type: MIP starts through XPRSaddmipsol, non-linear starts through XPRSnlpsetinitval (both accept partial solutions), and pure LP starts through XPRSloadlpsol (which requires a complete vector, so unset variables are filled from their bounds or zero). Basis reuse across solves is left to the user to drive explicitly with XPRSgetbasis / XPRSloadbasis; the interface does not cache and reload a basis implicitly.

47.1.2

  • Fixed compute_conflict! so that an infeasible subproblem is reported as MOI.CONFLICT_FOUND even when the IIS search did not finish (for example, when it was stopped by a time limit) or when the returned infeasible set is not guaranteed to be irreducible. Previously such results were silently dropped and reported as no conflict.
  • Constrained the julia compat entry to a bounded range so the package satisfies the Julia General registry AutoMerge requirements.

47.1.1

  • Introduced the JuMP / MathOptInterface (MOI) interface, shipped as the XpressMOIExt package extension. This provides an Optimizer usable from JuMP, including linear, quadratic, nonlinear, SOS and indicator constraints, callbacks, and IIS/conflict support.
  • Added DREF-style array allocation helpers for the low-level wrapper.
  • Fixed Windows-specific callback and test issues.

44.1.1

  • First public release. Introduced the low-level Julia wrapper for the Xpress Optimizer C API: a Julia function wrapper for every supported C function, with error-code translation, output parameters returned as values, optional array allocation, resource finalizers, and callback support.