pymomentum-cpu 0.1.93.post0__cp312-cp312-macosx_14_0_arm64.whl

include/momentum/simd/simd.h

@@ -0,0 +1,221 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <momentum/common/exception.h>
+#include <momentum/math/transform.h>
+
+#include <drjit/array.h>
+#include <drjit/array_router.h>
+#include <drjit/packet.h>
+#include <drjit/util.h>
+#include <Eigen/Core>
+#include <Eigen/Geometry>
+
+#define DRJIT_VERSION_GE(major, minor, patch)                           \
+  ((DRJIT_VERSION_MAJOR > (major)) ||                                   \
+   (DRJIT_VERSION_MAJOR == (major) && DRJIT_VERSION_MINOR > (minor)) || \
+   (DRJIT_VERSION_MAJOR == (major) && DRJIT_VERSION_MINOR == (minor) && \
+    DRJIT_VERSION_PATCH >= (patch)))
+
+// Utilities for writing cross-platform SIMD code.
+// This currently uses the DrJit library for SIMD primitives.
+
+namespace momentum {
+
+inline constexpr size_t kAvxPacketSize = 8;
+inline constexpr size_t kAvxAlignment = kAvxPacketSize * sizeof(float);
+
+inline constexpr size_t kSimdPacketSize = drjit::DefaultSize;
+inline constexpr size_t kSimdAlignment = kSimdPacketSize * sizeof(float);
+
+template <typename T>
+using Packet = drjit::Packet<T, kSimdPacketSize>;
+
+using FloatP = Packet<float>;
+using DoubleP = Packet<double>;
+using IntP = Packet<int>;
+
+template <typename T, int Dim>
+using VectorP = drjit::Array<Packet<T>, Dim>;
+
+template <typename T>
+using Vector2P = VectorP<T, 2>;
+
+template <typename T>
+using Vector3P = VectorP<T, 3>;
+
+using Vector2fP = Vector2P<float>;
+using Vector3fP = Vector3P<float>;
+
+using Vector2dP = Vector2P<double>;
+using Vector3dP = Vector3P<double>;
+
+/// Computes the offset required for the matrix data to meet the alignment requirement.
+template <size_t Alignment>
+[[nodiscard]] size_t computeOffset(const Eigen::Ref<Eigen::MatrixXf>& mat) {
+  constexpr size_t sizeOfScalar = sizeof(typename Eigen::MatrixXf::Scalar);
+  const size_t addressOffset =
+      Alignment / sizeOfScalar - (((size_t)mat.data() % Alignment) / sizeOfScalar);
+
+  // If the current alignment already meets the requirement, no offset is needed.
+  if (addressOffset == Alignment / sizeOfScalar) {
+    return 0;
+  }
+
+  return addressOffset;
+}
+
+/// Checks if the data of the matrix is aligned correctly.
+template <size_t Alignment>
+void checkAlignment(const Eigen::Ref<Eigen::MatrixXf>& mat, size_t offset = 0) {
+  MT_THROW_IF(
+      (uintptr_t(mat.data() + offset)) % Alignment != 0,
+      "Matrix ({}x{}, ptr: {}) is not aligned ({}) correctly.",
+      mat.rows(),
+      mat.cols(),
+      uintptr_t(mat.data()),
+      Alignment);
+}
+
+/// Calculates dot product of Eigen::Vector3f and 3-vector of packets.
+///
+/// @tparam Derived A derived class of Eigen::MatrixBase, which should be compatible with
+/// Eigen::Vector3f.
+/// @param[in] v1 A 3D vector of Eigen.
+/// @param[in] v2 A 3-vector of packets.
+template <typename Derived, typename S>
+FloatP dot(const Eigen::MatrixBase<Derived>& v1, const Vector3P<S>& v2) {
+  return drjit::fmadd(v1.x(), v2.x(), drjit::fmadd(v1.y(), v2.y(), v1.z() * v2.z()));
+}
+
+/// Calculates dot product of Vector3P and Eigen::Vector3f.
+///
+/// @tparam Derived A derived class of Eigen::MatrixBase, which should be compatible with
+/// Eigen::Vector3f.
+/// @param[in] v1 A 3-vector of packets.
+/// @param[in] v2 A 3D vector of Eigen.
+template <typename Derived, typename S>
+FloatP dot(const Vector3P<S>& v1, const Eigen::MatrixBase<Derived>& v2) {
+  return drjit::fmadd(v1.x(), v2.x(), drjit::fmadd(v1.y(), v2.y(), v1.z() * v2.z()));
+}
+
+/// Calculates summation of Eigen::Vector3f and Vector3P
+///
+/// @tparam Derived A derived class of Eigen::MatrixBase, which should be compatible with
+/// Eigen::Vector3f.
+/// @param[in] v1 A 3D vector of Eigen.
+/// @param[in] v2 A 3-vector of packets.
+template <typename S, typename Derived>
+Vector3P<S> operator+(const Eigen::MatrixBase<Derived>& v1, const Vector3P<S>& v2) {
+  return {v1.x() + v2.x(), v1.y() + v2.y(), v1.z() + v2.z()};
+}
+
+/// Calculates summation of Vector3P and Eigen::Vector3f
+///
+/// @tparam Derived A derived class of Eigen::MatrixBase, which should be compatible with
+/// Eigen::Vector3f.
+/// @param[in] v1 A 3-vector of packets.
+/// @param[in] v2 A 3D vector of Eigen.
+template <typename S, typename Derived>
+Vector3P<S> operator+(const Vector3P<S>& v1, const Eigen::MatrixBase<Derived>& v2) {
+  return {v1.x() + v2.x(), v1.y() + v2.y(), v1.z() + v2.z()};
+}
+
+/// Calculates subtraction of Eigen::Vector3f and Vector3P
+///
+/// @tparam Derived A derived class of Eigen::MatrixBase, which should be compatible with
+/// Eigen::Vector3f.
+/// @param[in] v1 A 3D vector of Eigen.
+/// @param[in] v2 A 3-vector of packets.
+template <typename S, typename Derived>
+Vector3P<S> operator-(const Eigen::MatrixBase<Derived>& v1, const Vector3P<S>& v2) {
+  return {v1.x() - v2.x(), v1.y() - v2.y(), v1.z() - v2.z()};
+}
+
+/// Calculates subtraction of Vector3P and Eigen::Vector3f
+///
+/// @tparam Derived A derived class of Eigen::MatrixBase, which should be compatible with
+/// Eigen::Vector3f.
+/// @param[in] v1 A 3-vector of packets.
+/// @param[in] v2 A 3D vector of Eigen.
+template <typename S, typename Derived>
+Vector3P<S> operator-(const Vector3P<S>& v1, const Eigen::MatrixBase<Derived>& v2) {
+  return {v1.x() - v2.x(), v1.y() - v2.y(), v1.z() - v2.z()};
+}
+
+/// Calculates multiplication of quaternion and each 3x1 vector in packet
+///
+/// @param[in] q A quaternion.
+/// @param[in] vec A 3-vector of packets.
+template <typename S>
+Vector3P<S> operator*(const Eigen::Quaternion<S>& q, const Vector3P<S>& vec) {
+  return q * vec;
+}
+
+/// Calculates multiplication of 3x3 matrix and each 3x1 vector in packet
+///
+/// @tparam Derived A derived class of Eigen::MatrixBase, which should be compatible with
+/// Eigen::Matrix3f.
+/// @param[in] xf A 3x3 matrix.
+/// @param[in] vec A 3-vector of packets.
+template <typename S, typename Derived>
+Vector3P<S> operator*(const Eigen::MatrixBase<Derived>& xf, const Vector3P<S>& vec) {
+  return Vector3P<S>{
+      momentum::dot(xf.row(0), vec), momentum::dot(xf.row(1), vec), momentum::dot(xf.row(2), vec)};
+}
+
+/// Calculates affine transformation on each 3x1 vector in packet
+///
+/// @param[in] xf An affine transformation matrix.
+/// @param[in] vec A 3-vector of packets.
+template <typename S>
+Vector3P<S> operator*(const Eigen::Transform<S, 3, Eigen::Affine>& xf, const Vector3P<S>& vec) {
+  return momentum::operator+(momentum::operator*(xf.linear(), vec), xf.translation());
+}
+
+/// Calculates transformation on each 3x1 vector in packet using momentum::Transform
+///
+/// @param[in] xf A momentum transform.
+/// @param[in] vec A 3-vector of packets.
+template <typename S>
+Vector3P<S> operator*(const TransformT<S>& xf, const Vector3P<S>& vec) {
+  return momentum::operator+(momentum::operator*(xf.toLinear(), vec), xf.translation);
+}
+
+/// Calculates cross product of Eigen::Vector3f and Vector3P
+///
+/// @tparam Derived A derived class of Eigen::MatrixBase, which should be compatible with
+/// Eigen::Vector3f.
+/// @param[in] v1 A 3D vector of Eigen.
+/// @param[in] v2 A 3-vector of packets.
+template <typename S, typename Derived>
+Vector3P<S> cross(const Eigen::MatrixBase<Derived>& v1, const Vector3P<S>& v2) {
+  return {
+      drjit::fmsub(v1.y(), v2.z(), v1.z() * v2.y()),
+      drjit::fmsub(v1.z(), v2.x(), v1.x() * v2.z()),
+      drjit::fmsub(v1.x(), v2.y(), v1.y() * v2.x()),
+  };
+}
+
+/// Calculates cross product of Vector3P and Eigen::Vector3f
+///
+/// @tparam Derived A derived class of Eigen::MatrixBase, which should be compatible with
+/// Eigen::Vector3f.
+/// @param[in] v1 A 3-vector of packets.
+/// @param[in] v2 A 3D vector of Eigen.
+template <typename S, typename Derived>
+Vector3P<S> cross(const Vector3P<S>& v1, const Eigen::MatrixBase<Derived>& v2) {
+  return {
+      drjit::fmsub(v1.y(), v2.z(), v1.z() * v2.y()),
+      drjit::fmsub(v1.z(), v2.x(), v1.x() * v2.z()),
+      drjit::fmsub(v1.x(), v2.y(), v1.y() * v2.x()),
+  };
+}
+
+} // namespace momentum

include/momentum/solver/fwd.h

@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+// This file is @generated by gen_fwd.py
+// Do not edit this file directly, instead update gen_fwd_input.toml and run:
+// buck run @arvr/mode/platform010/opt //arvr/libraries/momentum:gen_fwd
+
+#pragma once
+
+#include <memory>
+
+namespace momentum {
+
+struct SolverOptions;
+
+using SolverOptions_p = ::std::shared_ptr<SolverOptions>;
+using SolverOptions_u = ::std::unique_ptr<SolverOptions>;
+using SolverOptions_w = ::std::weak_ptr<SolverOptions>;
+using SolverOptions_const_p = ::std::shared_ptr<const SolverOptions>;
+using SolverOptions_const_u = ::std::unique_ptr<const SolverOptions>;
+using SolverOptions_const_w = ::std::weak_ptr<const SolverOptions>;
+
+struct GaussNewtonSolverOptions;
+
+using GaussNewtonSolverOptions_p = ::std::shared_ptr<GaussNewtonSolverOptions>;
+using GaussNewtonSolverOptions_u = ::std::unique_ptr<GaussNewtonSolverOptions>;
+using GaussNewtonSolverOptions_w = ::std::weak_ptr<GaussNewtonSolverOptions>;
+using GaussNewtonSolverOptions_const_p = ::std::shared_ptr<const GaussNewtonSolverOptions>;
+using GaussNewtonSolverOptions_const_u = ::std::unique_ptr<const GaussNewtonSolverOptions>;
+using GaussNewtonSolverOptions_const_w = ::std::weak_ptr<const GaussNewtonSolverOptions>;
+
+template <typename T>
+class SolverT;
+using Solver = SolverT<float>;
+using Solverd = SolverT<double>;
+
+using Solver_p = ::std::shared_ptr<Solver>;
+using Solver_u = ::std::unique_ptr<Solver>;
+using Solver_w = ::std::weak_ptr<Solver>;
+using Solver_const_p = ::std::shared_ptr<const Solver>;
+using Solver_const_u = ::std::unique_ptr<const Solver>;
+using Solver_const_w = ::std::weak_ptr<const Solver>;
+
+using Solverd_p = ::std::shared_ptr<Solverd>;
+using Solverd_u = ::std::unique_ptr<Solverd>;
+using Solverd_w = ::std::weak_ptr<Solverd>;
+using Solverd_const_p = ::std::shared_ptr<const Solverd>;
+using Solverd_const_u = ::std::unique_ptr<const Solverd>;
+using Solverd_const_w = ::std::weak_ptr<const Solverd>;
+
+template <typename T>
+class SolverFunctionT;
+using SolverFunction = SolverFunctionT<float>;
+using SolverFunctiond = SolverFunctionT<double>;
+
+using SolverFunction_p = ::std::shared_ptr<SolverFunction>;
+using SolverFunction_u = ::std::unique_ptr<SolverFunction>;
+using SolverFunction_w = ::std::weak_ptr<SolverFunction>;
+using SolverFunction_const_p = ::std::shared_ptr<const SolverFunction>;
+using SolverFunction_const_u = ::std::unique_ptr<const SolverFunction>;
+using SolverFunction_const_w = ::std::weak_ptr<const SolverFunction>;
+
+using SolverFunctiond_p = ::std::shared_ptr<SolverFunctiond>;
+using SolverFunctiond_u = ::std::unique_ptr<SolverFunctiond>;
+using SolverFunctiond_w = ::std::weak_ptr<SolverFunctiond>;
+using SolverFunctiond_const_p = ::std::shared_ptr<const SolverFunctiond>;
+using SolverFunctiond_const_u = ::std::unique_ptr<const SolverFunctiond>;
+using SolverFunctiond_const_w = ::std::weak_ptr<const SolverFunctiond>;
+
+template <typename T>
+class GaussNewtonSolverT;
+using GaussNewtonSolver = GaussNewtonSolverT<float>;
+using GaussNewtonSolverd = GaussNewtonSolverT<double>;
+
+using GaussNewtonSolver_p = ::std::shared_ptr<GaussNewtonSolver>;
+using GaussNewtonSolver_u = ::std::unique_ptr<GaussNewtonSolver>;
+using GaussNewtonSolver_w = ::std::weak_ptr<GaussNewtonSolver>;
+using GaussNewtonSolver_const_p = ::std::shared_ptr<const GaussNewtonSolver>;
+using GaussNewtonSolver_const_u = ::std::unique_ptr<const GaussNewtonSolver>;
+using GaussNewtonSolver_const_w = ::std::weak_ptr<const GaussNewtonSolver>;
+
+using GaussNewtonSolverd_p = ::std::shared_ptr<GaussNewtonSolverd>;
+using GaussNewtonSolverd_u = ::std::unique_ptr<GaussNewtonSolverd>;
+using GaussNewtonSolverd_w = ::std::weak_ptr<GaussNewtonSolverd>;
+using GaussNewtonSolverd_const_p = ::std::shared_ptr<const GaussNewtonSolverd>;
+using GaussNewtonSolverd_const_u = ::std::unique_ptr<const GaussNewtonSolverd>;
+using GaussNewtonSolverd_const_w = ::std::weak_ptr<const GaussNewtonSolverd>;
+
+template <typename T>
+class GradientDescentSolverT;
+using GradientDescentSolver = GradientDescentSolverT<float>;
+using GradientDescentSolverd = GradientDescentSolverT<double>;
+
+using GradientDescentSolver_p = ::std::shared_ptr<GradientDescentSolver>;
+using GradientDescentSolver_u = ::std::unique_ptr<GradientDescentSolver>;
+using GradientDescentSolver_w = ::std::weak_ptr<GradientDescentSolver>;
+using GradientDescentSolver_const_p = ::std::shared_ptr<const GradientDescentSolver>;
+using GradientDescentSolver_const_u = ::std::unique_ptr<const GradientDescentSolver>;
+using GradientDescentSolver_const_w = ::std::weak_ptr<const GradientDescentSolver>;
+
+using GradientDescentSolverd_p = ::std::shared_ptr<GradientDescentSolverd>;
+using GradientDescentSolverd_u = ::std::unique_ptr<GradientDescentSolverd>;
+using GradientDescentSolverd_w = ::std::weak_ptr<GradientDescentSolverd>;
+using GradientDescentSolverd_const_p = ::std::shared_ptr<const GradientDescentSolverd>;
+using GradientDescentSolverd_const_u = ::std::unique_ptr<const GradientDescentSolverd>;
+using GradientDescentSolverd_const_w = ::std::weak_ptr<const GradientDescentSolverd>;
+
+template <typename T>
+class SubsetGaussNewtonSolverT;
+using SubsetGaussNewtonSolver = SubsetGaussNewtonSolverT<float>;
+using SubsetGaussNewtonSolverd = SubsetGaussNewtonSolverT<double>;
+
+using SubsetGaussNewtonSolver_p = ::std::shared_ptr<SubsetGaussNewtonSolver>;
+using SubsetGaussNewtonSolver_u = ::std::unique_ptr<SubsetGaussNewtonSolver>;
+using SubsetGaussNewtonSolver_w = ::std::weak_ptr<SubsetGaussNewtonSolver>;
+using SubsetGaussNewtonSolver_const_p = ::std::shared_ptr<const SubsetGaussNewtonSolver>;
+using SubsetGaussNewtonSolver_const_u = ::std::unique_ptr<const SubsetGaussNewtonSolver>;
+using SubsetGaussNewtonSolver_const_w = ::std::weak_ptr<const SubsetGaussNewtonSolver>;
+
+using SubsetGaussNewtonSolverd_p = ::std::shared_ptr<SubsetGaussNewtonSolverd>;
+using SubsetGaussNewtonSolverd_u = ::std::unique_ptr<SubsetGaussNewtonSolverd>;
+using SubsetGaussNewtonSolverd_w = ::std::weak_ptr<SubsetGaussNewtonSolverd>;
+using SubsetGaussNewtonSolverd_const_p = ::std::shared_ptr<const SubsetGaussNewtonSolverd>;
+using SubsetGaussNewtonSolverd_const_u = ::std::unique_ptr<const SubsetGaussNewtonSolverd>;
+using SubsetGaussNewtonSolverd_const_w = ::std::weak_ptr<const SubsetGaussNewtonSolverd>;
+
+} // namespace momentum

include/momentum/solver/gauss_newton_solver.h

@@ -0,0 +1,136 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <momentum/solver/fwd.h>
+#include <momentum/solver/solver.h>
+
+namespace momentum {
+
+/// Extended options specific to the Gauss-Newton optimization algorithm
+struct GaussNewtonSolverOptions : SolverOptions {
+  /// Damping parameter added to Hessian diagonal for numerical stability; see
+  /// https://en.wikipedia.org/wiki/Levenberg%E2%80%93Marquardt_algorithm
+  ///
+  /// Higher values improve stability but may slow convergence
+  float regularization = 0.05f;
+
+  /// Enables backtracking line search to ensure error reduction at each step
+  bool doLineSearch = false;
+
+  /// Uses pre-computed JᵀJ and JᵀR from the solver function
+  ///
+  /// Can improve performance for problems with specialized structure
+  bool useBlockJtJ = false;
+
+  /// Directly computes sparse JᵀJ without dense intermediate representation
+  ///
+  /// Only effective when useBlockJtJ is true
+  bool directSparseJtJ = false;
+
+  /// Parameter count threshold for switching to sparse matrix operations
+  ///
+  /// Problems with more parameters than this threshold will use sparse solvers
+  size_t sparseMatrixThreshold = 200;
+
+  /// Default constructor
+  GaussNewtonSolverOptions() = default;
+
+  /// Construct from base solver options while preserving Gauss-Newton defaults
+  /* implicit */ GaussNewtonSolverOptions(const SolverOptions& baseOptions)
+      : SolverOptions(baseOptions) {}
+};
+
+/// Implementation of the Gauss-Newton optimization algorithm
+///
+/// Minimizes non-linear least squares problems by iteratively approximating
+/// the objective function with a quadratic model based on first derivatives.
+/// Supports both dense and sparse matrix operations depending on problem size.
+template <typename T>
+class GaussNewtonSolverT : public SolverT<T> {
+ public:
+  /// Creates a solver with the specified options and function to optimize
+  GaussNewtonSolverT(const SolverOptions& options, SolverFunctionT<T>* solver);
+
+  /// Returns "GaussNewton" as the solver name
+  [[nodiscard]] std::string_view getName() const override;
+
+  /// Updates solver configuration, handling both base and Gauss-Newton specific options
+  void setOptions(const SolverOptions& options) final;
+
+ protected:
+  /// Performs one iteration of the Gauss-Newton algorithm
+  ///
+  /// Delegates to either doIterationDense or doIterationSparse based on problem size
+  void doIteration() final;
+
+  /// Initializes solver state before optimization begins
+  void initializeSolver() final;
+
+ private:
+  /// Performs one iteration using dense matrix operations
+  void doIterationDense();
+
+  /// Performs one iteration using sparse matrix operations
+  void doIterationSparse();
+
+  /// Updates parameters using the computed step direction
+  ///
+  /// Optionally performs line search if enabled
+  void updateParameters(Eigen::VectorX<T>& delta);
+
+  /// Whether to use pre-computed JᵀJ and JᵀR from solver function
+  bool useBlockJtJ_{};
+
+  /// Whether to directly compute sparse JᵀJ without dense intermediate
+  bool directSparseJtJ_{};
+
+  /// Parameter count threshold for switching to sparse operations
+  size_t sparseMatrixThreshold_{200};
+
+  /// Whether the solver has been initialized
+  bool initialized_;
+
+  /// Whether to perform line search during parameter updates
+  bool doLineSearch_;
+
+  /// Sparse Cholesky factorization solver
+  Eigen::SimplicialLLT<Eigen::SparseMatrix<T>, Eigen::Lower> lltSolver_;
+
+  /// Sparse approximation of Hessian matrix (JᵀJ)
+  Eigen::SparseMatrix<T> JtJ_;
+
+  /// Sparse identity matrix for regularization
+  Eigen::SparseMatrix<T> D_;
+
+  /// Jacobian matrix for dense operations
+  Eigen::MatrixX<T> jacobian_;
+
+  /// Dense approximation of Hessian matrix (JᵀJ)
+  Eigen::MatrixX<T> hessianApprox_;
+
+  /// Gradient vector (JᵀR)
+  Eigen::VectorX<T> JtR_;
+
+  /// Residual vector
+  Eigen::VectorX<T> residual_;
+
+  /// Dense Cholesky factorization solver
+  Eigen::LLT<Eigen::MatrixX<T>> llt_;
+
+  /// Base regularization parameter
+  T regularization_;
+
+  /// Current regularization value (may be adjusted during iterations)
+  T alpha_;
+
+  /// Whether the current iteration uses dense operations
+  bool denseIteration_{};
+};
+
+} // namespace momentum

include/momentum/solver/gradient_descent_solver.h

@@ -0,0 +1,65 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <momentum/solver/fwd.h>
+#include <momentum/solver/solver.h>
+
+namespace momentum {
+
+/// Extended options for the Gradient Descent optimization algorithm
+///
+/// Provides configuration specific to the first-order gradient descent method
+struct GradientDescentSolverOptions : SolverOptions {
+  /// Step size for parameter updates during optimization
+  ///
+  /// Controls how far to move in the direction of the negative gradient
+  float learningRate = 0.01f;
+
+  /// Default constructor
+  GradientDescentSolverOptions() = default;
+
+  /// Construct from base solver options while preserving Gradient Descent defaults
+  /* implicit */ GradientDescentSolverOptions(const SolverOptions& baseOptions)
+      : SolverOptions(baseOptions) {}
+};
+
+/// First-order optimization algorithm that follows the negative gradient
+///
+/// Implements the standard gradient descent method which iteratively
+/// updates parameters by moving in the direction of steepest descent
+template <typename T>
+class GradientDescentSolverT : public SolverT<T> {
+ public:
+  /// Creates a solver with the specified options and function to optimize
+  GradientDescentSolverT(const SolverOptions& options, SolverFunctionT<T>* solver);
+
+  /// Returns "GradientDescent" as the solver name
+  [[nodiscard]] std::string_view getName() const override;
+
+  /// Updates solver configuration, handling both base and Gradient Descent specific options
+  void setOptions(const SolverOptions& options) final;
+
+ protected:
+  /// Performs one iteration of the gradient descent algorithm
+  ///
+  /// Computes the gradient and updates parameters by moving in the negative gradient direction
+  void doIteration() final;
+
+  /// Initializes solver state before optimization begins
+  void initializeSolver() final;
+
+ private:
+  /// Gradient vector at current parameter values
+  Eigen::VectorX<T> gradient_;
+
+  /// Step size for parameter updates
+  float learningRate_;
+};
+
+} // namespace momentum