class gpr::linalg::ProbeExecutor¶

Overview¶

#include <ProbeExecutor.h>
 
class ProbeExecutor {
public:
    // construction
 
    ProbeExecutor();
 
    // methods
 
    Eigen::VectorXd run(int n_probes, const std::function<double(int)>& probe_fn, KernelInverse::Kind kind, const MpiContext& ctx) const;
};

Detailed Documentation¶

Methods¶

Eigen::VectorXd run(int n_probes, const std::function<double(int)>& probe_fn, KernelInverse::Kind kind, const MpiContext& ctx) const

Run n_probes probe evaluations under the dispatch policy implied by kind, against the MPI context ctx, and return the global result vector. The result vector has length n_probes on every rank.

Performance contract: the std::function<double(int)> parameter is type-erased; lambdas larger than the SBO threshold (~24-32 bytes on libstdc++) heap-allocate on construction. Callers should construct the std::function ONCE outside the loop and reuse it, not re-bind on every probe. The probe lambda itself is invoked n_probes / nprocs times under RankLocal kind, n_probes times under Distributed kind.

Parameters:

n_probes	Total number of probes to evaluate. Must be >= 0.
probe_fn	Callable with signature `double(int)` returning the scalar contribution of probe `idx` (in the range ```` [0, n_probes)). May return `NaN` to signal failure; the executor will then return a vector of `NaN` on every rank.
kind	Dispatch policy. `RankLocal` triggers round-robin partitioning + Allreduce-SUM. `Distributed` runs every probe on every rank with no Allreduce.
ctx	MPI context. Use `MpiContext::local()` for the single-process path (always available) or `MpiContext::world()` under the ScaLAPACK build.

Returns:

The n_probes -element vector of probe values, identical on every rank.