Make sparse jacobian/hessian threashold configurable

Summary:
The threshold was previously hardcoded to 200, but can now be configured while maintaining the default value (no behavioral changes for downstream users unless explicitly modified).

Additionally, enforcing dense Jacobian/Hessian for retargeter benchmark improved performance from 3.22 ms to 2.93 ms by increasing the threshold configuration to 300 (where the parameter size is 204, so enforcing to dense jacobian/hessian).

Reviewed By: yutingye

Differential Revision: D69091231

fbshipit-source-id: 2fe8bb29f17e8ef224a5001e3dd19140d73fbb68

momentum/solver/gauss_newton_solver.cpp

@@ -39,6 +39,7 @@ void GaussNewtonSolverT<T>::setOptions(const SolverOptions& options) {
     doLineSearch_ = derivedOptions->doLineSearch;
     useBlockJtJ_ = derivedOptions->useBlockJtJ;
     directSparseJtJ_ = derivedOptions->directSparseJtJ;
+    sparseMatrixThreshold_ = derivedOptions->sparseMatrixThreshold;
   }
 }
 
@@ -49,7 +50,7 @@ void GaussNewtonSolverT<T>::initializeSolver() {
   this->newParameterPattern_ = true;
   this->lastError_ = std::numeric_limits<double>::max();
 
-  denseIteration_ = this->numParameters_ < 200;
+  denseIteration_ = this->numParameters_ < sparseMatrixThreshold_;
 }
 
 template <typename T>

momentum/solver/gauss_newton_solver.h

@@ -27,6 +27,10 @@ struct GaussNewtonSolverOptions : SolverOptions {
   /// to true.
   bool directSparseJtJ = false;
 
+  /// The minimum number of parameters required to switch from dense to sparse matrix
+  /// representations for Jacobian and Hessian computations.
+  size_t sparseMatrixThreshold = 200;
+
   GaussNewtonSolverOptions() = default;
 
   /* implicit */ GaussNewtonSolverOptions(const SolverOptions& baseOptions)
@@ -55,6 +59,7 @@ class GaussNewtonSolverT : public SolverT<T> {
 
   bool useBlockJtJ_{};
   bool directSparseJtJ_{};
+  size_t sparseMatrixThreshold_{200};
   bool initialized_;
   bool doLineSearch_;