Reworked the expression evaluation mechanism in order to make it possible to efficiently compute convolutions and contractions in the future:

 * The scheduling of computation is moved out the the assignment code and into a new TensorExecutor class
 * The assignment itself is now a regular node on the expression tree
 * The expression evaluators start by recursively evaluating all their subexpressions if needed

1 files changed, 142 insertions, 0 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
new file mode 100644
index 000000000..6f6641de6
--- /dev/null
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -0,0 +1,142 @@
+// 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_FORCED_EVAL_H
+#define EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H
+
+namespace Eigen {
+
+/** \class TensorForcedEval
+  * \ingroup CXX11_Tensor_Module
+  *
+  * \brief Tensor reshaping class.
+  *
+  *
+  */
+namespace internal {
+template<typename XprType>
+struct traits<TensorForcedEvalOp<XprType> >
+{
+  // Type promotion to handle the case where the types of the lhs and the rhs are different.
+  typedef typename XprType::Scalar Scalar;
+  typedef typename internal::packet_traits<Scalar>::type Packet;
+  typedef typename traits<XprType>::StorageKind StorageKind;
+  typedef typename traits<XprType>::Index Index;
+  typedef typename XprType::Nested Nested;
+  typedef typename remove_reference<Nested>::type _Nested;
+
+  enum {
+    Flags = 0,
+  };
+};
+
+template<typename XprType>
+struct eval<TensorForcedEvalOp<XprType>, Eigen::Dense>
+{
+  typedef const TensorForcedEvalOp<XprType>& type;
+};
+
+template<typename XprType>
+struct nested<TensorForcedEvalOp<XprType>, 1, typename eval<TensorForcedEvalOp<XprType> >::type>
+{
+  typedef TensorForcedEvalOp<XprType> type;
+};
+
+}  // end namespace internal
+
+
+
+template<typename XprType>
+class TensorForcedEvalOp : public TensorBase<TensorForcedEvalOp<XprType> >
+{
+  public:
+  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::Scalar Scalar;
+  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::Packet Packet;
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+  typedef typename Eigen::internal::nested<TensorForcedEvalOp>::type Nested;
+  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::StorageKind StorageKind;
+  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::Index Index;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorForcedEvalOp(const XprType& expr)
+      : m_xpr(expr) {}
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<typename XprType::Nested>::type&
+    expression() const { return m_xpr; }
+
+  protected:
+    typename XprType::Nested m_xpr;
+};
+
+
+template<typename ArgType, typename Device>
+struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
+{
+  typedef TensorForcedEvalOp<ArgType> XprType;
+  typedef typename ArgType::Scalar Scalar;
+  typedef typename ArgType::Packet Packet;
+  typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
+
+  enum {
+    IsAligned = true,
+    PacketAccess = true,
+  };
+
+  EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device)
+      : m_impl(op.expression(), device), m_op(op.expression()), m_device(device), m_buffer(NULL)
+  { }
+
+  EIGEN_DEVICE_FUNC ~TensorEvaluator() {
+    eigen_assert(!m_buffer);
+  }
+
+  typedef typename XprType::Index Index;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+
+  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); }
+
+  EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() {
+    m_impl.evalSubExprsIfNeeded();
+    m_buffer = (Scalar*)m_device.allocate(m_impl.dimensions().TotalSize() * sizeof(Scalar));
+
+    typedef TensorEvalToOp<const ArgType> EvalTo;
+    EvalTo evalToTmp(m_buffer, m_op);
+    internal::TensorExecutor<const EvalTo, Device, TensorEvaluator<ArgType, Device>::PacketAccess>::run(evalToTmp, m_device);
+    m_impl.cleanup();
+  }
+  EIGEN_STRONG_INLINE void cleanup() {
+    m_device.deallocate(m_buffer);
+    m_buffer = NULL;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
+  {
+    return m_buffer[index];
+  }
+
+  template<int LoadMode>
+  EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  {
+    return internal::ploadt<Packet, LoadMode>(m_buffer + index);
+  }
+
+ private:
+  TensorEvaluator<ArgType, Device> m_impl;
+  const ArgType m_op;
+  const Device& m_device;
+  Scalar* m_buffer;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H