casadi 3.7.2__cp314-none-manylinux2014_i686.whl

casadi/include/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorCostModel.h

@@ -0,0 +1,214 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Rasmus Munk Larsen <rmlarsen@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H
+#define EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H
+
+namespace Eigen {
+
+/** \class TensorEvaluator
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief A cost model used to limit the number of threads used for evaluating
+  * tensor expression.
+  *
+  */
+
+// Class storing the cost of evaluating a tensor expression in terms of the
+// estimated number of operand bytes loads, bytes stored, and compute cycles.
+class TensorOpCost {
+ public:
+  // TODO(rmlarsen): Fix the scalar op costs in Eigen proper. Even a simple
+  // model based on minimal reciprocal throughput numbers from Intel or
+  // Agner Fog's tables would be better than what is there now.
+  template <typename ArgType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int MulCost() {
+    return internal::functor_traits<
+        internal::scalar_product_op<ArgType, ArgType> >::Cost;
+  }
+  template <typename ArgType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int AddCost() {
+    return internal::functor_traits<internal::scalar_sum_op<ArgType> >::Cost;
+  }
+  template <typename ArgType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int DivCost() {
+    return internal::functor_traits<
+        internal::scalar_quotient_op<ArgType, ArgType> >::Cost;
+  }
+  template <typename ArgType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int ModCost() {
+    return internal::functor_traits<internal::scalar_mod_op<ArgType> >::Cost;
+  }
+  template <typename SrcType, typename TargetType>
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int CastCost() {
+    return internal::functor_traits<
+        internal::scalar_cast_op<SrcType, TargetType> >::Cost;
+  }
+
+  EIGEN_DEVICE_FUNC
+  TensorOpCost() : bytes_loaded_(0), bytes_stored_(0), compute_cycles_(0) {}
+  EIGEN_DEVICE_FUNC
+  TensorOpCost(double bytes_loaded, double bytes_stored, double compute_cycles)
+      : bytes_loaded_(bytes_loaded),
+        bytes_stored_(bytes_stored),
+        compute_cycles_(compute_cycles) {}
+
+  EIGEN_DEVICE_FUNC
+  TensorOpCost(double bytes_loaded, double bytes_stored, double compute_cycles,
+               bool vectorized, double packet_size)
+      : bytes_loaded_(bytes_loaded),
+        bytes_stored_(bytes_stored),
+        compute_cycles_(vectorized ? compute_cycles / packet_size
+                                   : compute_cycles) {
+    eigen_assert(bytes_loaded >= 0 && (numext::isfinite)(bytes_loaded));
+    eigen_assert(bytes_stored >= 0 && (numext::isfinite)(bytes_stored));
+    eigen_assert(compute_cycles >= 0 && (numext::isfinite)(compute_cycles));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double bytes_loaded() const {
+    return bytes_loaded_;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double bytes_stored() const {
+    return bytes_stored_;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double compute_cycles() const {
+    return compute_cycles_;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double total_cost(
+      double load_cost, double store_cost, double compute_cost) const {
+    return load_cost * bytes_loaded_ + store_cost * bytes_stored_ +
+           compute_cost * compute_cycles_;
+  }
+
+  // Drop memory access component. Intended for cases when memory accesses are
+  // sequential or are completely masked by computations.
+  EIGEN_DEVICE_FUNC void dropMemoryCost() {
+    bytes_loaded_ = 0;
+    bytes_stored_ = 0;
+  }
+
+  // TODO(rmlarsen): Define min in terms of total cost, not elementwise.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost cwiseMin(
+      const TensorOpCost& rhs) const {
+    double bytes_loaded = numext::mini(bytes_loaded_, rhs.bytes_loaded());
+    double bytes_stored = numext::mini(bytes_stored_, rhs.bytes_stored());
+    double compute_cycles = numext::mini(compute_cycles_, rhs.compute_cycles());
+    return TensorOpCost(bytes_loaded, bytes_stored, compute_cycles);
+  }
+
+  // TODO(rmlarsen): Define max in terms of total cost, not elementwise.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost cwiseMax(
+      const TensorOpCost& rhs) const {
+    double bytes_loaded = numext::maxi(bytes_loaded_, rhs.bytes_loaded());
+    double bytes_stored = numext::maxi(bytes_stored_, rhs.bytes_stored());
+    double compute_cycles = numext::maxi(compute_cycles_, rhs.compute_cycles());
+    return TensorOpCost(bytes_loaded, bytes_stored, compute_cycles);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost& operator+=(
+      const TensorOpCost& rhs) {
+    bytes_loaded_ += rhs.bytes_loaded();
+    bytes_stored_ += rhs.bytes_stored();
+    compute_cycles_ += rhs.compute_cycles();
+    return *this;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost& operator*=(double rhs) {
+    bytes_loaded_ *= rhs;
+    bytes_stored_ *= rhs;
+    compute_cycles_ *= rhs;
+    return *this;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE friend TensorOpCost operator+(
+      TensorOpCost lhs, const TensorOpCost& rhs) {
+    lhs += rhs;
+    return lhs;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE friend TensorOpCost operator*(
+      TensorOpCost lhs, double rhs) {
+    lhs *= rhs;
+    return lhs;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE friend TensorOpCost operator*(
+      double lhs, TensorOpCost rhs) {
+    rhs *= lhs;
+    return rhs;
+  }
+
+  friend std::ostream& operator<<(std::ostream& os, const TensorOpCost& tc) {
+    return os << "[bytes_loaded = " << tc.bytes_loaded()
+              << ", bytes_stored = " << tc.bytes_stored()
+              << ", compute_cycles = " << tc.compute_cycles() << "]";
+  }
+
+ private:
+  double bytes_loaded_;
+  double bytes_stored_;
+  double compute_cycles_;
+};
+
+// TODO(rmlarsen): Implement a policy that chooses an "optimal" number of theads
+// in [1:max_threads] instead of just switching multi-threading off for small
+// work units.
+template <typename Device>
+class TensorCostModel {
+ public:
+  // Scaling from Eigen compute cost to device cycles.
+  static const int kDeviceCyclesPerComputeCycle = 1;
+
+ // Costs in device cycles.
+  static const int kStartupCycles = 100000;
+  static const int kPerThreadCycles = 100000;
+  static const int kTaskSize = 40000;
+
+  // Returns the number of threads in [1:max_threads] to use for
+  // evaluating an expression with the given output size and cost per
+  // coefficient.
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int numThreads(
+      double output_size, const TensorOpCost& cost_per_coeff, int max_threads) {
+    double cost = totalCost(output_size, cost_per_coeff);
+    double threads = (cost - kStartupCycles) / kPerThreadCycles + 0.9;
+    // Make sure we don't invoke undefined behavior when we convert to an int.
+    threads = numext::mini<double>(threads, GenericNumTraits<int>::highest());
+    return numext::mini(max_threads,
+                        numext::maxi<int>(1, static_cast<int>(threads)));
+  }
+
+  // taskSize assesses parallel task size.
+  // Value of 1.0 means ideal parallel task size. Values < 1.0 mean that task
+  // granularity needs to be increased to mitigate parallelization overheads.
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double taskSize(
+      double output_size, const TensorOpCost& cost_per_coeff) {
+    return totalCost(output_size, cost_per_coeff) / kTaskSize;
+  }
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double totalCost(
+      double output_size, const TensorOpCost& cost_per_coeff) {
+    // Cost of memory fetches from L2 cache. 64 is typical cache line size.
+    // 11 is L2 cache latency on Haswell.
+    // We don't know whether data is in L1, L2 or L3. But we are most interested
+    // in single-threaded computational time around 100us-10ms (smaller time
+    // is too small for parallelization, larger time is not interesting
+    // either because we are probably using all available threads already).
+    // And for the target time range, L2 seems to be what matters. Data set
+    // fitting into L1 is too small to take noticeable time. Data set fitting
+    // only into L3 presumably will take more than 10ms to load and process.
+    const double kLoadCycles = 1.0 / 64 * 11;
+    const double kStoreCycles = 1.0 / 64 * 11;
+    // Scaling from Eigen compute cost to device cycles.
+    return output_size *
+        cost_per_coeff.total_cost(kLoadCycles, kStoreCycles,
+                                  kDeviceCyclesPerComputeCycle);
+  }
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_TENSOR_TENSOR_COST_MODEL_H

casadi/include/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorCustomOp.h

@@ -0,0 +1,347 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_CUSTOM_OP_H
+#define EIGEN_CXX11_TENSOR_TENSOR_CUSTOM_OP_H
+
+namespace Eigen {
+
+/** \class TensorCustomUnaryOp
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor custom class.
+  *
+  *
+  */
+namespace internal {
+template<typename CustomUnaryFunc, typename XprType>
+struct traits<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
+{
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::StorageKind StorageKind;
+  typedef typename XprType::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+  static const int NumDimensions = traits<XprType>::NumDimensions;
+  static const int Layout = traits<XprType>::Layout;
+  typedef typename traits<XprType>::PointerType PointerType;
+};
+
+template<typename CustomUnaryFunc, typename XprType>
+struct eval<TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Eigen::Dense>
+{
+  typedef const TensorCustomUnaryOp<CustomUnaryFunc, XprType>EIGEN_DEVICE_REF type;
+};
+
+template<typename CustomUnaryFunc, typename XprType>
+struct nested<TensorCustomUnaryOp<CustomUnaryFunc, XprType> >
+{
+  typedef TensorCustomUnaryOp<CustomUnaryFunc, XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename CustomUnaryFunc, typename XprType>
+class TensorCustomUnaryOp : public TensorBase<TensorCustomUnaryOp<CustomUnaryFunc, XprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename internal::traits<TensorCustomUnaryOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename internal::nested<TensorCustomUnaryOp>::type Nested;
+  typedef typename internal::traits<TensorCustomUnaryOp>::StorageKind StorageKind;
+  typedef typename internal::traits<TensorCustomUnaryOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCustomUnaryOp(const XprType& expr, const CustomUnaryFunc& func)
+      : m_expr(expr), m_func(func) {}
+
+  EIGEN_DEVICE_FUNC
+  const CustomUnaryFunc& func() const { return m_func; }
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename XprType::Nested>::type&
+  expression() const { return m_expr; }
+
+  protected:
+    typename XprType::Nested m_expr;
+    const CustomUnaryFunc m_func;
+};
+
+
+// Eval as rvalue
+template<typename CustomUnaryFunc, typename XprType, typename Device>
+struct TensorEvaluator<const TensorCustomUnaryOp<CustomUnaryFunc, XprType>, Device>
+{
+  typedef TensorCustomUnaryOp<CustomUnaryFunc, XprType> ArgType;
+  typedef typename internal::traits<ArgType>::Index Index;
+  static const int NumDims = internal::traits<ArgType>::NumDimensions;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename internal::remove_const<typename ArgType::Scalar>::type Scalar;
+  typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef typename Eigen::internal::traits<XprType>::PointerType TensorPointerType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<XprType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const ArgType& op, const Device& device)
+      : m_op(op), m_device(device), m_result(NULL)
+  {
+    m_dimensions = op.func().dimensions(op.expression());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    if (data) {
+      evalTo(data);
+      return false;
+    } else {
+      m_result = static_cast<EvaluatorPointerType>(m_device.get( (CoeffReturnType*)
+          m_device.allocate_temp(dimensions().TotalSize() * sizeof(Scalar))));
+      evalTo(m_result);
+      return true;
+    }
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    if (m_result) {
+      m_device.deallocate_temp(m_result);
+      m_result = NULL;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    return m_result[index];
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_result + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    // TODO(rmlarsen): Extend CustomOp API to return its cost estimate.
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_result; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_result.bind(cgh);
+  }
+#endif
+
+ protected:
+  void evalTo(EvaluatorPointerType data) {
+    TensorMap<Tensor<CoeffReturnType, NumDims, Layout, Index> > result(m_device.get(data), m_dimensions);
+    m_op.func().eval(m_op.expression(), result, m_device);
+  }
+
+  Dimensions m_dimensions;
+  const ArgType m_op;
+  const Device EIGEN_DEVICE_REF m_device;
+  EvaluatorPointerType m_result;
+};
+
+
+
+/** \class TensorCustomBinaryOp
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor custom class.
+  *
+  *
+  */
+namespace internal {
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
+struct traits<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
+{
+  typedef typename internal::promote_storage_type<typename LhsXprType::Scalar,
+                                                  typename RhsXprType::Scalar>::ret Scalar;
+  typedef typename internal::promote_storage_type<typename LhsXprType::CoeffReturnType,
+                                                  typename RhsXprType::CoeffReturnType>::ret CoeffReturnType;
+  typedef typename promote_storage_type<typename traits<LhsXprType>::StorageKind,
+                                        typename traits<RhsXprType>::StorageKind>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<LhsXprType>::Index,
+                                      typename traits<RhsXprType>::Index>::type Index;
+  typedef typename LhsXprType::Nested LhsNested;
+  typedef typename RhsXprType::Nested RhsNested;
+  typedef typename remove_reference<LhsNested>::type _LhsNested;
+  typedef typename remove_reference<RhsNested>::type _RhsNested;
+  static const int NumDimensions = traits<LhsXprType>::NumDimensions;
+  static const int Layout = traits<LhsXprType>::Layout;
+  typedef typename conditional<Pointer_type_promotion<typename LhsXprType::Scalar, Scalar>::val,
+                                typename traits<LhsXprType>::PointerType, typename traits<RhsXprType>::PointerType>::type PointerType;
+};
+
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
+struct eval<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, Eigen::Dense>
+{
+  typedef const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>& type;
+};
+
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
+struct nested<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> >
+{
+  typedef TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType>
+class TensorCustomBinaryOp : public TensorBase<TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, ReadOnlyAccessors>
+{
+  public:
+  typedef typename internal::traits<TensorCustomBinaryOp>::Scalar Scalar;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename internal::traits<TensorCustomBinaryOp>::CoeffReturnType CoeffReturnType;
+  typedef typename internal::nested<TensorCustomBinaryOp>::type Nested;
+  typedef typename internal::traits<TensorCustomBinaryOp>::StorageKind StorageKind;
+  typedef typename internal::traits<TensorCustomBinaryOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorCustomBinaryOp(const LhsXprType& lhs, const RhsXprType& rhs, const CustomBinaryFunc& func)
+
+      : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_func(func) {}
+
+  EIGEN_DEVICE_FUNC
+  const CustomBinaryFunc& func() const { return m_func; }
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename LhsXprType::Nested>::type&
+  lhsExpression() const { return m_lhs_xpr; }
+
+  EIGEN_DEVICE_FUNC
+  const typename internal::remove_all<typename RhsXprType::Nested>::type&
+  rhsExpression() const { return m_rhs_xpr; }
+
+  protected:
+    typename LhsXprType::Nested m_lhs_xpr;
+    typename RhsXprType::Nested m_rhs_xpr;
+    const CustomBinaryFunc m_func;
+};
+
+
+// Eval as rvalue
+template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType, typename Device>
+struct TensorEvaluator<const TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType>, Device>
+{
+  typedef TensorCustomBinaryOp<CustomBinaryFunc, LhsXprType, RhsXprType> XprType;
+  typedef typename internal::traits<XprType>::Index Index;
+  static const int NumDims = internal::traits<XprType>::NumDimensions;
+  typedef DSizes<Index, NumDims> Dimensions;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
+  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
+  static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+
+  typedef typename Eigen::internal::traits<XprType>::PointerType TensorPointerType;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
+
+  enum {
+    IsAligned = false,
+    PacketAccess = (PacketType<CoeffReturnType, Device>::size > 1),
+    BlockAccess = false,
+    PreferBlockAccess = false,
+    Layout = TensorEvaluator<LhsXprType, Device>::Layout,
+    CoordAccess = false,  // to be implemented
+    RawAccess = false
+  };
+
+  //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
+  typedef internal::TensorBlockNotImplemented TensorBlock;
+  //===--------------------------------------------------------------------===//
+
+  EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
+      : m_op(op), m_device(device), m_result(NULL)
+  {
+    m_dimensions = op.func().dimensions(op.lhsExpression(), op.rhsExpression());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
+
+  EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType data) {
+    if (data) {
+      evalTo(data);
+      return false;
+    } else {
+      m_result = static_cast<EvaluatorPointerType>(m_device.get( (CoeffReturnType*)
+        m_device.allocate_temp(dimensions().TotalSize() * sizeof(CoeffReturnType))));
+      evalTo(m_result);
+      return true;
+    }
+  }
+
+  EIGEN_STRONG_INLINE void cleanup() {
+    if (m_result != NULL) {
+      m_device.deallocate_temp(m_result);
+      m_result = NULL;
+    }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
+    return m_result[index];
+  }
+
+  template<int LoadMode>
+  EIGEN_DEVICE_FUNC PacketReturnType packet(Index index) const {
+    return internal::ploadt<PacketReturnType, LoadMode>(m_result + index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
+    // TODO(rmlarsen): Extend CustomOp API to return its cost estimate.
+    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
+  }
+
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return m_result; }
+
+#ifdef EIGEN_USE_SYCL
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_result.bind(cgh);
+  }
+#endif
+
+ protected:
+  void evalTo(EvaluatorPointerType data) {
+    TensorMap<Tensor<CoeffReturnType, NumDims, Layout> > result(m_device.get(data), m_dimensions);
+    m_op.func().eval(m_op.lhsExpression(), m_op.rhsExpression(), result, m_device);
+  }
+
+  Dimensions m_dimensions;
+  const XprType m_op;
+  const Device EIGEN_DEVICE_REF m_device;
+  EvaluatorPointerType m_result;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_CUSTOM_OP_H

casadi/include/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h

@@ -0,0 +1,137 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CXX11_TENSOR_TENSOR_DEVICE_H
+#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_H
+
+namespace Eigen {
+
+/** \class TensorDevice
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Pseudo expression providing an operator = that will evaluate its argument
+  * on the specified computing 'device' (GPU, thread pool, ...)
+  *
+  * Example:
+  *    C.device(EIGEN_GPU) = A + B;
+  *
+  * Todo: operator *= and /=.
+  */
+
+template <typename ExpressionType, typename DeviceType> class TensorDevice {
+  public:
+    TensorDevice(const DeviceType& device, ExpressionType& expression) : m_device(device), m_expression(expression) {}
+
+    EIGEN_DEFAULT_COPY_CONSTRUCTOR(TensorDevice)
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE TensorDevice& operator=(const OtherDerived& other) {
+      typedef TensorAssignOp<ExpressionType, const OtherDerived> Assign;
+      Assign assign(m_expression, other);
+      internal::TensorExecutor<const Assign, DeviceType>::run(assign, m_device);
+      return *this;
+    }
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE TensorDevice& operator+=(const OtherDerived& other) {
+      typedef typename OtherDerived::Scalar Scalar;
+      typedef TensorCwiseBinaryOp<internal::scalar_sum_op<Scalar>, const ExpressionType, const OtherDerived> Sum;
+      Sum sum(m_expression, other);
+      typedef TensorAssignOp<ExpressionType, const Sum> Assign;
+      Assign assign(m_expression, sum);
+      internal::TensorExecutor<const Assign, DeviceType>::run(assign, m_device);
+      return *this;
+    }
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE TensorDevice& operator-=(const OtherDerived& other) {
+      typedef typename OtherDerived::Scalar Scalar;
+      typedef TensorCwiseBinaryOp<internal::scalar_difference_op<Scalar>, const ExpressionType, const OtherDerived> Difference;
+      Difference difference(m_expression, other);
+      typedef TensorAssignOp<ExpressionType, const Difference> Assign;
+      Assign assign(m_expression, difference);
+      internal::TensorExecutor<const Assign, DeviceType>::run(assign, m_device);
+      return *this;
+    }
+
+  protected:
+    const DeviceType& m_device;
+    ExpressionType& m_expression;
+};
+
+/** \class TensorAsyncDevice
+ * \ingroup CXX11_Tensor_Module
+ *
+ * \brief Pseudo expression providing an operator = that will evaluate its
+ * argument asynchronously on the specified device. Currently only
+ * ThreadPoolDevice implements proper asynchronous execution, while the default
+ * and GPU devices just run the expression synchronously and call m_done() on
+ * completion..
+ *
+ * Example:
+ *    auto done = []() { ... expression evaluation done ... };
+ *    C.device(thread_pool_device, std::move(done)) = A + B;
+ */
+
+template <typename ExpressionType, typename DeviceType, typename DoneCallback>
+class TensorAsyncDevice {
+ public:
+  TensorAsyncDevice(const DeviceType& device, ExpressionType& expression,
+                    DoneCallback done)
+      : m_device(device), m_expression(expression), m_done(std::move(done)) {}
+
+  template <typename OtherDerived>
+  EIGEN_STRONG_INLINE TensorAsyncDevice& operator=(const OtherDerived& other) {
+    typedef TensorAssignOp<ExpressionType, const OtherDerived> Assign;
+    typedef internal::TensorExecutor<const Assign, DeviceType> Executor;
+
+    Assign assign(m_expression, other);
+    Executor::run(assign, m_device);
+    m_done();
+
+    return *this;
+  }
+
+ protected:
+  const DeviceType& m_device;
+  ExpressionType& m_expression;
+  DoneCallback m_done;
+};
+
+
+#ifdef EIGEN_USE_THREADS
+template <typename ExpressionType, typename DoneCallback>
+class TensorAsyncDevice<ExpressionType, ThreadPoolDevice, DoneCallback> {
+ public:
+  TensorAsyncDevice(const ThreadPoolDevice& device, ExpressionType& expression,
+                    DoneCallback done)
+      : m_device(device), m_expression(expression), m_done(std::move(done)) {}
+
+  template <typename OtherDerived>
+  EIGEN_STRONG_INLINE TensorAsyncDevice& operator=(const OtherDerived& other) {
+    typedef TensorAssignOp<ExpressionType, const OtherDerived> Assign;
+    typedef internal::TensorAsyncExecutor<const Assign, ThreadPoolDevice, DoneCallback> Executor;
+
+    // WARNING: After assignment 'm_done' callback will be in undefined state.
+    Assign assign(m_expression, other);
+    Executor::runAsync(assign, m_device, std::move(m_done));
+
+    return *this;
+  }
+
+ protected:
+  const ThreadPoolDevice& m_device;
+  ExpressionType& m_expression;
+  DoneCallback m_done;
+};
+#endif
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_H