StructureΒΆ

The API is written in C++ (code in /cpp) and is then wrapped to other target languages by using SWIG. Therefore, any modification, fix, or the code to add a solver API must be written in such language and follow its structure.

The interfaces are intented to be very simple, to encourage people contributions and to ease debugging. In a nutshell, they are organised as a (core) set of interfaces and some derived implementations; specifically, each solver must implement three interfaces, that are represented in C++ by three abstract classes.

It has to be noted that, since the AMPLModel-derived object instance contains a reference to the solver-specific in memory representation of the model, there is no possible functionality to create a model instance linked to one solver and solve it with another one.

The following are the base classes implemented for each solver:

digraph hierarchy {
size="6,6"
node[shape=record,style=filled,fillcolor=gray95]
edge[dir=back, arrowtail=empty]

SolverDriver[label = "{SolverDriver|Constructs an instance\lof a model in solver\lmemory}"]

AMPLModel[label = "{AMPLModel|Holds a reference to the model in\lsolver memory and wraps the\lmost common\lfunctionalities}"]

BaseCB[label = "{BaseCallback|Represents a callback to be used\lduring the solution process}"]

}
  • AMPLModel (cpp, python) is the main interface, it provides functionalities like:

    • keep a reference to the in-memory representation of the model

    • solve the model

    • get solution status, solution vector and objective value

    • access common properties like number of variables, of constraints

    • assign callback object

  • BaseCallback when derived for a specific solver (see for example cpp and python), provides:

    • access to various properties accessible during the solution process

    • functions to add cuts or lazy constraints

    • ability to terminate the solution process

So, taking gurobi as an example, we have:

digraph hierarchy {
size="8,8"
node[shape=record,style=filled,fillcolor=gray95]
edge[dir=back, arrowtail=empty]

SolverDriver[label = "{SolverDriver|}"]
GurobiDrv[label = "{GurobiDrv|Construct an instance of GurobiModel}"]

AMPLModel[label = "{AMPLModel|}"]
GurobiModel[label = "{GurobiModel|Holds a reference to the native\lGRBModel object and implements\lcommon functionalities by means\lof the Gurobi C API}"]

BaseCallback[label = "{BaseCallback|Represents a callback to be used\lduring the solution process}"]
GurobiCallback[label = "{GurobiCallback|Provides the interface functionalities\lby means of the Gurobi C API}"]

AMPLModel->GurobiModel
SolverDriver->GurobiDrv
BaseCallback->GurobiCallback
}

The derived classes implement the interfaces specified in their base classes and extend them with some solver-specific functionalities. This leads to three possible usages:

  • Generic usage: using only functions available in the base classes, the same code can be reused to implement solver-independent logic. To support this type of usage, the class GenericCallback is provided (cpp, python), deriving which a solver-independent callback can be defined.

  • Solver specific usage: using the additional functions provided by each solver implementation, the code can be used only with one solver

  • Solver native usage: a subcase of the case above: each solver-specific model class provides access to the native C model object, on which all the C solver libraries functionalities can be used