1// This file is part of Eigen, a lightweight C++ template library
2// for linear algebra.
3//
4// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
5//
6// This Source Code Form is subject to the terms of the Mozilla
7// Public License v. 2.0. If a copy of the MPL was not distributed
8// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9
10#ifndef EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H
11#define EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H
12
13namespace Eigen {
14
15/** \class TensorForcedEval
16  * \ingroup CXX11_Tensor_Module
17  *
18  * \brief Tensor reshaping class.
19  *
20  *
21  */
22/// template <class> class MakePointer_ is added to convert the host pointer to the device pointer.
23/// It is added due to the fact that for our device compiler T* is not allowed.
24/// If we wanted to use the same Evaluator functions we have to convert that type to our pointer T.
25/// This is done through our MakePointer_ class. By default the Type in the MakePointer_<T> is T* .
26/// Therefore, by adding the default value, we managed to convert the type and it does not break any
27/// existing code as its default value is T*.
28namespace internal {
29template<typename XprType, template <class> class MakePointer_>
30struct traits<TensorForcedEvalOp<XprType, MakePointer_> >
31{
32  // Type promotion to handle the case where the types of the lhs and the rhs are different.
33  typedef typename XprType::Scalar Scalar;
34  typedef traits<XprType> XprTraits;
35  typedef typename traits<XprType>::StorageKind StorageKind;
36  typedef typename traits<XprType>::Index Index;
37  typedef typename XprType::Nested Nested;
38  typedef typename remove_reference<Nested>::type _Nested;
39  static const int NumDimensions = XprTraits::NumDimensions;
40  static const int Layout = XprTraits::Layout;
41
42  enum {
43    Flags = 0
44  };
45  template <class T> struct MakePointer {
46    // Intermediate typedef to workaround MSVC issue.
47    typedef MakePointer_<T> MakePointerT;
48    typedef typename MakePointerT::Type Type;
49  };
50};
51
52template<typename XprType, template <class> class MakePointer_>
53struct eval<TensorForcedEvalOp<XprType, MakePointer_>, Eigen::Dense>
54{
55  typedef const TensorForcedEvalOp<XprType, MakePointer_>& type;
56};
57
58template<typename XprType, template <class> class MakePointer_>
59struct nested<TensorForcedEvalOp<XprType, MakePointer_>, 1, typename eval<TensorForcedEvalOp<XprType, MakePointer_> >::type>
60{
61  typedef TensorForcedEvalOp<XprType, MakePointer_> type;
62};
63
64}  // end namespace internal
65
66
67
68template<typename XprType, template <class> class MakePointer_>
69class TensorForcedEvalOp : public TensorBase<TensorForcedEvalOp<XprType, MakePointer_>, ReadOnlyAccessors>
70{
71  public:
72  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::Scalar Scalar;
73  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar;
74  typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
75  typedef typename Eigen::internal::nested<TensorForcedEvalOp>::type Nested;
76  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::StorageKind StorageKind;
77  typedef typename Eigen::internal::traits<TensorForcedEvalOp>::Index Index;
78
79  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorForcedEvalOp(const XprType& expr)
80      : m_xpr(expr) {}
81
82    EIGEN_DEVICE_FUNC
83    const typename internal::remove_all<typename XprType::Nested>::type&
84    expression() const { return m_xpr; }
85
86  protected:
87    typename XprType::Nested m_xpr;
88};
89
90
91template<typename ArgType, typename Device, template <class> class MakePointer_>
92struct TensorEvaluator<const TensorForcedEvalOp<ArgType, MakePointer_>, Device>
93{
94  typedef TensorForcedEvalOp<ArgType, MakePointer_> XprType;
95  typedef typename ArgType::Scalar Scalar;
96  typedef typename TensorEvaluator<ArgType, Device>::Dimensions Dimensions;
97  typedef typename XprType::Index Index;
98  typedef typename XprType::CoeffReturnType CoeffReturnType;
99  typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
100  static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
101
102  enum {
103    IsAligned = true,
104    PacketAccess = (PacketSize > 1),
105    Layout = TensorEvaluator<ArgType, Device>::Layout,
106    RawAccess = true
107  };
108
109  EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device)
110	/// op_ is used for sycl
111      : m_impl(op.expression(), device), m_op(op.expression()), m_device(device), m_buffer(NULL)
112  { }
113
114  EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); }
115
116  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType*) {
117    const Index numValues =  internal::array_prod(m_impl.dimensions());
118    m_buffer = (CoeffReturnType*)m_device.allocate(numValues * sizeof(CoeffReturnType));
119    // Should initialize the memory in case we're dealing with non POD types.
120    if (NumTraits<CoeffReturnType>::RequireInitialization) {
121      for (Index i = 0; i < numValues; ++i) {
122        new(m_buffer+i) CoeffReturnType();
123      }
124    }
125    typedef TensorEvalToOp< const typename internal::remove_const<ArgType>::type > EvalTo;
126    EvalTo evalToTmp(m_buffer, m_op);
127    const bool PacketAccess = internal::IsVectorizable<Device, const ArgType>::value;
128    internal::TensorExecutor<const EvalTo, typename internal::remove_const<Device>::type, PacketAccess>::run(evalToTmp, m_device);
129    return true;
130  }
131  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
132    m_device.deallocate(m_buffer);
133    m_buffer = NULL;
134  }
135
136  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
137  {
138    return m_buffer[index];
139  }
140
141  template<int LoadMode>
142  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
143  {
144    return internal::ploadt<PacketReturnType, LoadMode>(m_buffer + index);
145  }
146
147  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
148    return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
149  }
150
151  EIGEN_DEVICE_FUNC typename MakePointer<Scalar>::Type data() const { return m_buffer; }
152
153  /// required by sycl in order to extract the sycl accessor
154  const TensorEvaluator<ArgType, Device>& impl() { return m_impl; }
155  /// used by sycl in order to build the sycl buffer
156  const Device& device() const{return m_device;}
157 private:
158  TensorEvaluator<ArgType, Device> m_impl;
159  const ArgType m_op;
160  const Device& m_device;
161  typename MakePointer<CoeffReturnType>::Type m_buffer;
162};
163
164
165} // end namespace Eigen
166
167#endif // EIGEN_CXX11_TENSOR_TENSOR_FORCED_EVAL_H
168