[gar] Change LQRKnotTpl class to work with single block of allocate…

…d memory per block

+ using `MemReq`

Give no_init_t tag type a constexpr explicit ctor

CHANGELOG.md

@@ -26,6 +26,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - [gar] Rework `RiccatiSolverDense` to not use inner struct `FactorData`
 - Various changes to `gar` tests and `test_util`, add `LQRKnot::isApprox()`
 - Add `MemReq` struct to handle requests for single blocks of memory for multiple arrays
+- Change `LQRKnotTpl` to work with a single block of allocated memory using `MemReq`
 
 ### Removed
 

bindings/python/src/gar/expose-gar.cpp

@@ -59,8 +59,9 @@ void exposeGAR() {
   exposeBlockMatrices();
 
   bp::class_<knot_t>("LQRKnot", bp::no_init)
-      .def(bp::init<uint, uint, uint>(("nx"_a, "nu", "nc")))
-      .def(bp::init<uint, uint, uint, uint>(("nx"_a, "nu", "nc", "nx2")))
+      .def(bp::init<uint, uint, uint>(("self"_a, "nx", "nu", "nc")))
+      .def(bp::init<uint, uint, uint, uint, uint>(
+          ("self"_a, "nx"_a, "nu", "nc", "nx2", "nth"_a = 0)))
       .def_readonly("nx", &knot_t::nx)
       .def_readonly("nu", &knot_t::nu)
       .def_readonly("nc", &knot_t::nc)

gar/include/aligator/gar/lqr-problem.hpp

@@ -2,6 +2,7 @@
 #pragma once
 
 #include "aligator/math.hpp"
+#include "mem-req.hpp"
 #include <fmt/format.h>
 
 #include <optional>
@@ -27,35 +28,81 @@ namespace gar {
 ///
 template <typename Scalar> struct LQRKnotTpl {
   ALIGATOR_DYNAMIC_TYPEDEFS(Scalar);
+  enum { Alignment = Eigen::AlignedMax };
+  using VectorMap = Eigen::Map<VectorXs, Alignment>;
+  using MatrixMap = Eigen::Map<MatrixXs, Alignment>;
+
+  // tag type
+  struct no_alloc_t {
+    explicit constexpr no_alloc_t() {}
+  };
+  static constexpr no_alloc_t no_alloc{};
 
   uint nx, nu, nc, nx2, nth;
-  MatrixXs Q, S, R;
-  VectorXs q, r;
-  MatrixXs A, B, E;
-  VectorXs f;
-  MatrixXs C, D;
-  VectorXs d;
-
-  MatrixXs Gth;
-  MatrixXs Gx;
-  MatrixXs Gu;
-  MatrixXs Gv;
-  VectorXs gamma;
+
+  MatrixMap Q, S, R;
+  VectorMap q, r;
+  MatrixMap A, B, E;
+  VectorMap f;
+  MatrixMap C, D;
+  VectorMap d;
+
+  MatrixMap Gth;
+  MatrixMap Gx;
+  MatrixMap Gu;
+  MatrixMap Gv;
+  VectorMap gamma;
 
   LQRKnotTpl(uint nx, uint nu, uint nc, uint nx2, uint nth = 0);
 
   LQRKnotTpl(uint nx, uint nu, uint nc) : LQRKnotTpl(nx, nu, nc, nx) {}
 
+  void allocate();
+  // initialize the matrices.
+  void initialize();
   // reallocates entire buffer for contigousness
   void addParameterization(uint nth);
+
+  LQRKnotTpl(const LQRKnotTpl &other);
+  LQRKnotTpl(LQRKnotTpl &&other);
+  LQRKnotTpl &operator=(const LQRKnotTpl &other);
+  LQRKnotTpl &operator=(LQRKnotTpl &&other);
+
+  ~LQRKnotTpl();
+
+  friend void swap(LQRKnotTpl &lhs, LQRKnotTpl &rhs) {
+    using std::swap;
+    swap(lhs.nx, rhs.nx);
+    swap(lhs.nu, rhs.nu);
+    swap(lhs.nc, rhs.nc);
+    swap(lhs.nx2, rhs.nx2);
+    swap(lhs.nth, rhs.nth);
+    // only swap the memory ptr, do not swap the Eigen::Map objects.
+    swap(lhs.memory, rhs.memory);
+    swap(lhs.req, rhs.req);
+
+    lhs.initialize();
+    rhs.initialize();
+  }
+
   bool isApprox(const LQRKnotTpl &other,
                 Scalar prec = std::numeric_limits<Scalar>::epsilon()) const;
 
   friend bool operator==(const LQRKnotTpl &lhs, const LQRKnotTpl &rhs) {
     return lhs.isApprox(rhs);
   }
+
+private:
+  LQRKnotTpl(no_alloc_t, uint nx, uint nu, uint nc, uint nx2, uint nth);
+  Scalar *memory;
+  MemReq req;
 };
 
+template <typename Scalar> LQRKnotTpl<Scalar>::~LQRKnotTpl() {
+  if (memory)
+    std::free(memory);
+}
+
 template <typename Scalar> struct LQRProblemTpl {
   ALIGATOR_DYNAMIC_TYPEDEFS(Scalar);
   using KnotType = LQRKnotTpl<Scalar>;

gar/include/aligator/gar/lqr-problem.hxx

@@ -4,43 +4,163 @@
 
 namespace aligator::gar {
 
+namespace detail {
+template <typename EigenType, int MapOptions,
+          typename Scalar = typename EigenType::Scalar>
+void emplaceMap(Eigen::Map<EigenType, MapOptions> &map, long size,
+                Scalar *ptr) {
+  using MapType = Eigen::Map<EigenType, MapOptions>;
+  new (&map) MapType{ptr, size};
+}
+
+/// \brief Placement-new a map type using the provided memory pointer.
+template <typename EigenType, int MapOptions,
+          typename Scalar = typename EigenType::Scalar>
+void emplaceMap(Eigen::Map<EigenType, MapOptions> &map, long rows, long cols,
+                Scalar *ptr) {
+  using MapType = Eigen::Map<EigenType, MapOptions>;
+  new (&map) MapType{ptr, rows, cols};
+}
+} // namespace detail
+
+template <typename Scalar>
+LQRKnotTpl<Scalar>::LQRKnotTpl(no_alloc_t, uint nx, uint nu, uint nc, uint nx2,
+                               uint nth)
+    : nx(nx), nu(nu), nc(nc), nx2(nx2), nth(nth),                          //
+      Q(NULL, 0, 0), S(NULL, 0, 0), R(NULL, 0, 0), q(NULL, 0), r(NULL, 0), //
+      A(NULL, 0, 0), B(NULL, 0, 0), E(NULL, 0, 0), f(NULL, 0),             //
+      C(NULL, 0, 0), D(NULL, 0, 0), d(NULL, 0),                            //
+      Gth(NULL, 0, 0), Gx(NULL, 0, 0), Gu(NULL, 0, 0), Gv(NULL, 0, 0),
+      gamma(NULL, 0), //
+      memory(NULL), req(Alignment) {}
+
 template <typename Scalar>
 LQRKnotTpl<Scalar>::LQRKnotTpl(uint nx, uint nu, uint nc, uint nx2, uint nth)
-    : nx(nx), nu(nu), nc(nc), nx2(nx2), nth(nth),    //
-      Q(nx, nx), S(nx, nu), R(nu, nu), q(nx), r(nu), //
-      A(nx2, nx), B(nx2, nu), E(nx2, nx), f(nx2),    //
-      C(nc, nx), D(nc, nu), d(nc), Gth(nth, nth), Gx(nx, nth), Gu(nu, nth),
-      Gv(nc, nth), gamma(nth) {
-  Q.setZero();
-  S.setZero();
-  R.setZero();
-  q.setZero();
-  r.setZero();
-
-  A.setZero();
-  B.setZero();
-  E.setZero();
-  f.setZero();
-
-  C.setZero();
-  D.setZero();
-  d.setZero();
-
-  Gth.setZero();
-  Gx.setZero();
-  Gu.setZero();
-  Gv.setZero();
-  gamma.setZero();
+    : LQRKnotTpl(no_alloc, nx, nu, nc, nx2, nth) {
+
+  this->allocate();
+  this->initialize();
+}
+
+template <typename Scalar> void LQRKnotTpl<Scalar>::allocate() {
+  req.addArray<double>(nx, nx)    // Q
+      .addArray<double>(nx, nu)   // S
+      .addArray<double>(nu, nu)   // R
+      .addArray<double>(nx)       // q
+      .addArray<double>(nu)       // r
+      .addArray<double>(nx2, nx)  // A
+      .addArray<double>(nx2, nu)  // B
+      .addArray<double>(nx2, nx2) // E
+      .addArray<double>(nx2)      // f
+      .addArray<double>(nc, nx)   // C
+      .addArray<double>(nc, nu)   // D
+      .addArray<double>(nc)       // d
+      .addArray<double>(nth, nth) // Gth
+      .addArray<double>(nx, nth)  // Gx
+      .addArray<double>(nu, nth)  // Gu
+      .addArray<double>(nc, nth)  // Gv
+      .addArray<double>(nth);     // gamma
+
+  this->memory = static_cast<Scalar *>(req.allocate());
+  std::memset(memory, 0, req.totalBytes());
+}
+
+template <typename Scalar> void LQRKnotTpl<Scalar>::initialize() {
+  Scalar *ptr = memory;
+  detail::emplaceMap(Q, nx, nx, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(S, nx, nu, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(R, nu, nu, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(q, nx, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(r, nu, ptr);
+  req.advance(ptr);
+
+  detail::emplaceMap(A, nx2, nx, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(B, nx2, nu, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(E, nx2, nx2, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(f, nx2, ptr);
+  req.advance(ptr);
+
+  detail::emplaceMap(C, nc, nx, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(D, nc, nu, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(d, nc, ptr);
+  req.advance(ptr);
+
+  detail::emplaceMap(Gth, nth, nth, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(Gx, nx, nth, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(Gu, nu, nth, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(Gv, nc, nth, ptr);
+  req.advance(ptr);
+  detail::emplaceMap(gamma, nth, ptr);
+  req.advance(ptr);
+
+  req.reset();
 }
 
 template <typename Scalar>
 void LQRKnotTpl<Scalar>::addParameterization(uint nth) {
-  this->nth = nth;
-  Gth.setZero(nth, nth);
-  Gx.setZero(nx, nth);
-  Gu.setZero(nu, nth);
-  Gv.setZero(nc, nth);
-  gamma.setZero(nth);
+  LQRKnotTpl copy(nx, nu, nc, nx2, nth);
+  copy.Q = Q;
+  copy.S = S;
+  copy.R = R;
+  copy.q = q;
+  copy.r = r;
+
+  copy.A = A;
+  copy.B = B;
+  copy.E = E;
+  copy.f = f;
+
+  copy.C = C;
+  copy.D = D;
+  copy.d = d;
+
+  *this = LQRKnotTpl{copy};
+}
+
+template <typename Scalar>
+LQRKnotTpl<Scalar>::LQRKnotTpl(const LQRKnotTpl &other)
+    : LQRKnotTpl(no_alloc, other.nx, other.nu, other.nc, other.nx2, other.nth) {
+  this->allocate();
+  assert(req.totalBytes() == other.req.totalBytes());
+  // copy memory over from other
+  std::memcpy(memory, other.memory, other.req.totalBytes());
+  this->initialize();
+}
+
+template <typename Scalar>
+LQRKnotTpl<Scalar>::LQRKnotTpl(LQRKnotTpl &&other)
+    : LQRKnotTpl(no_alloc, other.nx, other.nu, other.nc, other.nx2, other.nth) {
+  // no need to allocate, just bring in the other
+  // memory buffer
+  memory = other.memory;
+  other.memory = NULL;
+  req = other.req;
+  this->initialize();
+}
+
+template <typename Scalar>
+LQRKnotTpl<Scalar> &LQRKnotTpl<Scalar>::operator=(const LQRKnotTpl &other) {
+  this->~LQRKnotTpl();
+  new (this) LQRKnotTpl{other};
+  return *this;
+}
+
+template <typename Scalar>
+LQRKnotTpl<Scalar> &LQRKnotTpl<Scalar>::operator=(LQRKnotTpl &&other) {
+  swap(*this, other);
+  return *this;
 }
 
 template <typename Scalar>