MPI Design Discussion

Developers
38

Is ...static data... going to be a sequence of data or transformed data variable names?

I take it the idea is to then generate a class like the model class in which those variable names are instantiated with actual data from the member variables rather than being passed as arguments. And then f wouldn’t actually be passed those arguments, it’d just access them locally.

That probably wouldn’t be too hard to generate. So if we had something like

map(foo, thetas, theta_ends, x, y, N);

then we’d generate a functor class like

template <typename T>
Matrix<T, -1, 1> foo(const Matrix<T, -1, 1>& theta, const MatrixXd& x,
      const VectorXd& y, int N) {
   ... code generated for body of foo ...
}

struct mapped_functor {
  MatrixXd x;
  VectorXd y;
  int N;

  template <typename T>
  Matrix<T, -1, 1> operator()(const Matrix<T, -1, 1>& theta) const {
    return foo(theta, x, y, N);
  }
}

And then the mapped functor is the thing that actually gets called in the child main processes. So it would require the main program for the child to create the mapped functor.

P.S. Stan math already is stateful—autodiff is stateful and it’s state is held in a global singleton. Also, any uses of the LDLT types are stateful in that they hold onto data and/or solutions.

1 Like