• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_CONTRACTION_H
11 #define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_H
12 
13 namespace Eigen {
14 
15 /** \class TensorContraction
16   * \ingroup CXX11_Tensor_Module
17   *
18   * \brief Tensor contraction class.
19   *
20   *
21   */
22 namespace internal {
23 
24 template<typename Dimensions, typename LhsXprType, typename RhsXprType>
25 struct traits<TensorContractionOp<Dimensions, LhsXprType, RhsXprType> >
26 {
27   // Type promotion to handle the case where the types of the lhs and the rhs are different.
28   typedef typename gebp_traits<typename remove_const<typename LhsXprType::Scalar>::type,
29                                typename remove_const<typename RhsXprType::Scalar>::type>::ResScalar Scalar;
30 
31   typedef typename promote_storage_type<typename traits<LhsXprType>::StorageKind,
32                                         typename traits<RhsXprType>::StorageKind>::ret StorageKind;
33   typedef typename promote_index_type<typename traits<LhsXprType>::Index,
34                                       typename traits<RhsXprType>::Index>::type Index;
35   typedef typename LhsXprType::Nested LhsNested;
36   typedef typename RhsXprType::Nested RhsNested;
37   typedef typename remove_reference<LhsNested>::type _LhsNested;
38   typedef typename remove_reference<RhsNested>::type _RhsNested;
39 
40   // From NumDims below.
41   static const int NumDimensions = traits<RhsXprType>::NumDimensions + traits<RhsXprType>::NumDimensions - 2 * array_size<Dimensions>::value;
42   static const int Layout = traits<LhsXprType>::Layout;
43 
44   enum {
45     Flags = 0
46   };
47 };
48 
49 template<typename Dimensions, typename LhsXprType, typename RhsXprType>
50 struct eval<TensorContractionOp<Dimensions, LhsXprType, RhsXprType>, Eigen::Dense>
51 {
52   typedef const TensorContractionOp<Dimensions, LhsXprType, RhsXprType>& type;
53 };
54 
55 template<typename Dimensions, typename LhsXprType, typename RhsXprType>
56 struct nested<TensorContractionOp<Dimensions, LhsXprType, RhsXprType>, 1, typename eval<TensorContractionOp<Dimensions, LhsXprType, RhsXprType> >::type>
57 {
58   typedef TensorContractionOp<Dimensions, LhsXprType, RhsXprType> type;
59 };
60 
61 template<typename Indices_, typename LeftArgType_, typename RightArgType_, typename Device_>
62 struct traits<TensorEvaluator<const TensorContractionOp<Indices_, LeftArgType_, RightArgType_>, Device_> > {
63   typedef Indices_ Indices;
64   typedef LeftArgType_ LeftArgType;
65   typedef RightArgType_ RightArgType;
66   typedef Device_ Device;
67 
68   // From NumDims below.
69   static const int NumDimensions = traits<LeftArgType_>::NumDimensions + traits<RightArgType_>::NumDimensions - 2 * array_size<Indices_>::value;
70 };
71 
72 }  // end namespace internal
73 
74 template<typename Indices, typename LhsXprType, typename RhsXprType>
75 class TensorContractionOp : public TensorBase<TensorContractionOp<Indices, LhsXprType, RhsXprType>, ReadOnlyAccessors>
76 {
77   public:
78   typedef typename Eigen::internal::traits<TensorContractionOp>::Scalar Scalar;
79   typedef typename internal::gebp_traits<typename LhsXprType::CoeffReturnType,
80                                                    typename RhsXprType::CoeffReturnType>::ResScalar CoeffReturnType;
81   typedef typename Eigen::internal::nested<TensorContractionOp>::type Nested;
82   typedef typename Eigen::internal::traits<TensorContractionOp>::StorageKind StorageKind;
83   typedef typename Eigen::internal::traits<TensorContractionOp>::Index Index;
84 
85   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorContractionOp(
86       const LhsXprType& lhs, const RhsXprType& rhs, const Indices& dims)
87       : m_lhs_xpr(lhs), m_rhs_xpr(rhs), m_indices(dims) {}
88 
89   EIGEN_DEVICE_FUNC
90   const Indices& indices() const { return m_indices; }
91 
92   /** \returns the nested expressions */
93   EIGEN_DEVICE_FUNC
94   const typename internal::remove_all<typename LhsXprType::Nested>::type&
95   lhsExpression() const { return m_lhs_xpr; }
96 
97   EIGEN_DEVICE_FUNC
98   const typename internal::remove_all<typename RhsXprType::Nested>::type&
99   rhsExpression() const { return m_rhs_xpr; }
100 
101   protected:
102     typename LhsXprType::Nested m_lhs_xpr;
103     typename RhsXprType::Nested m_rhs_xpr;
104     const Indices m_indices;
105 };
106 
107 
108 template<typename Derived>
109 struct TensorContractionEvaluatorBase
110 {
111   typedef typename internal::traits<Derived>::Indices Indices;
112   typedef typename internal::traits<Derived>::LeftArgType LeftArgType;
113   typedef typename internal::traits<Derived>::RightArgType RightArgType;
114   typedef typename internal::traits<Derived>::Device Device;
115 
116   typedef TensorContractionOp<Indices, LeftArgType, RightArgType> XprType;
117   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
118   typedef typename XprType::Index Index;
119   typedef typename XprType::CoeffReturnType CoeffReturnType;
120   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
121 
122   enum {
123     IsAligned = true,
124     PacketAccess = (internal::unpacket_traits<PacketReturnType>::size > 1),
125     Layout = TensorEvaluator<LeftArgType, Device>::Layout,
126     CoordAccess = false,  // to be implemented
127     RawAccess = true
128   };
129 
130   // Most of the code is assuming that both input tensors are ColMajor. If the
131   // inputs are RowMajor, we will "cheat" by swapping the LHS and RHS:
132   // If we want to compute A * B = C, where A is LHS and B is RHS, the code
133   // will pretend B is LHS and A is RHS.
134   typedef typename internal::conditional<
135     static_cast<int>(Layout) == static_cast<int>(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType;
136   typedef typename internal::conditional<
137     static_cast<int>(Layout) == static_cast<int>(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType;
138 
139   static const int LDims =
140       internal::array_size<typename TensorEvaluator<EvalLeftArgType, Device>::Dimensions>::value;
141   static const int RDims =
142       internal::array_size<typename TensorEvaluator<EvalRightArgType, Device>::Dimensions>::value;
143   static const int ContractDims = internal::array_size<Indices>::value;
144   static const int NumDims = LDims + RDims - 2 * ContractDims;
145 
146   typedef array<Index, ContractDims> contract_t;
147   typedef array<Index, LDims - ContractDims> left_nocontract_t;
148   typedef array<Index, RDims - ContractDims> right_nocontract_t;
149 
150   typedef DSizes<Index, NumDims> Dimensions;
151 
152   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
153   TensorContractionEvaluatorBase(const XprType& op, const Device& device)
154     : m_leftImpl(choose(Cond<static_cast<int>(Layout) == static_cast<int>(ColMajor)>(),
155                           op.lhsExpression(), op.rhsExpression()), device),
156     m_rightImpl(choose(Cond<static_cast<int>(Layout) == static_cast<int>(ColMajor)>(),
157                           op.rhsExpression(), op.lhsExpression()), device),
158         m_device(device),
159         m_result(NULL) {
160     EIGEN_STATIC_ASSERT((static_cast<int>(TensorEvaluator<LeftArgType, Device>::Layout) ==
161 			   static_cast<int>(TensorEvaluator<RightArgType, Device>::Layout)),
162                         YOU_MADE_A_PROGRAMMING_MISTAKE);
163 
164 
165     DSizes<Index, LDims> eval_left_dims;
166     DSizes<Index, RDims> eval_right_dims;
167     array<IndexPair<Index>, ContractDims> eval_op_indices;
168     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
169       // For ColMajor, we keep using the existing dimensions
170       for (int i = 0; i < LDims; i++) {
171         eval_left_dims[i] = m_leftImpl.dimensions()[i];
172       }
173       for (int i = 0; i < RDims; i++) {
174         eval_right_dims[i] = m_rightImpl.dimensions()[i];
175       }
176       // We keep the pairs of contracting indices.
177       for (int i = 0; i < ContractDims; i++) {
178         eval_op_indices[i].first = op.indices()[i].first;
179         eval_op_indices[i].second = op.indices()[i].second;
180       }
181     } else {
182       // For RowMajor, we need to reverse the existing dimensions
183       for (int i = 0; i < LDims; i++) {
184         eval_left_dims[i] = m_leftImpl.dimensions()[LDims - i - 1];
185       }
186       for (int i = 0; i < RDims; i++) {
187         eval_right_dims[i] = m_rightImpl.dimensions()[RDims - i - 1];
188       }
189       // We need to flip all the pairs of contracting indices as well as
190       // reversing the dimensions.
191       for (int i = 0; i < ContractDims; i++) {
192         eval_op_indices[i].first = LDims - 1 - op.indices()[ContractDims - 1 - i].second;
193         eval_op_indices[i].second = RDims - 1 - op.indices()[ContractDims - 1 - i].first;
194       }
195     }
196 
197     // Check for duplicate axes and make sure the first index in eval_op_indices
198     // is increasing. Using O(n^2) sorting is OK since ContractDims is small
199     for (int i = 0; i < ContractDims; i++) {
200       for (int j = i + 1; j < ContractDims; j++) {
201         eigen_assert(eval_op_indices[j].first != eval_op_indices[i].first &&
202                      eval_op_indices[j].second != eval_op_indices[i].second &&
203                      "contraction axes should be unique");
204         if (eval_op_indices[j].first < eval_op_indices[i].first) {
205           numext::swap(eval_op_indices[j], eval_op_indices[i]);
206         }
207       }
208     }
209 
210     array<Index, LDims> lhs_strides;
211     lhs_strides[0] = 1;
212     for (int i = 0; i < LDims-1; ++i) {
213       lhs_strides[i+1] = lhs_strides[i] * eval_left_dims[i];
214     }
215 
216     array<Index, RDims> rhs_strides;
217     rhs_strides[0] = 1;
218     for (int i = 0; i < RDims-1; ++i) {
219       rhs_strides[i+1] = rhs_strides[i] * eval_right_dims[i];
220     }
221 
222     if (m_i_strides.size() > 0) m_i_strides[0] = 1;
223     if (m_j_strides.size() > 0) m_j_strides[0] = 1;
224     if (m_k_strides.size() > 0) m_k_strides[0] = 1;
225 
226     m_i_size = 1;
227     m_j_size = 1;
228     m_k_size = 1;
229 
230     // To compute the dimension, we simply concatenate the non-contracting
231     // dimensions of the left and then the right tensor. Additionally, we also
232     // compute the strides corresponding to the left non-contracting
233     // dimensions and right non-contracting dimensions.
234     m_lhs_inner_dim_contiguous = true;
235     int dim_idx = 0;
236     unsigned int nocontract_idx = 0;
237 
238     for (int i = 0; i < LDims; i++) {
239       // find if we are contracting on index i of left tensor
240       bool contracting = false;
241       for (int j = 0; j < ContractDims; j++) {
242         if (eval_op_indices[j].first == i) {
243           contracting = true;
244           break;
245         }
246       }
247       if (!contracting) {
248         // add dimension size to output dimensions
249         m_dimensions[dim_idx] = eval_left_dims[i];
250         m_left_nocontract_strides[nocontract_idx] = lhs_strides[i];
251         if (dim_idx != i) {
252           m_lhs_inner_dim_contiguous = false;
253         }
254         if (nocontract_idx+1 < internal::array_size<left_nocontract_t>::value) {
255           m_i_strides[nocontract_idx+1] =
256               m_i_strides[nocontract_idx] * eval_left_dims[i];
257         } else {
258           m_i_size = m_i_strides[nocontract_idx] * eval_left_dims[i];
259         }
260         dim_idx++;
261         nocontract_idx++;
262       }
263     }
264 
265     nocontract_idx = 0;
266     for (int i = 0; i < RDims; i++) {
267       bool contracting = false;
268       // find if we are contracting on index i of right tensor
269       for (int j = 0; j < ContractDims; j++) {
270         if (eval_op_indices[j].second == i) {
271           contracting = true;
272           break;
273         }
274       }
275       if (!contracting) {
276         m_dimensions[dim_idx] = eval_right_dims[i];
277         if (nocontract_idx+1 < internal::array_size<right_nocontract_t>::value) {
278           m_j_strides[nocontract_idx+1] =
279               m_j_strides[nocontract_idx] * eval_right_dims[i];
280         } else {
281           m_j_size = m_j_strides[nocontract_idx] * eval_right_dims[i];
282         }
283         m_right_nocontract_strides[nocontract_idx] = rhs_strides[i];
284         dim_idx++;
285         nocontract_idx++;
286       }
287     }
288 
289     // Now compute the strides corresponding to the contracting dimensions. We
290     // assumed above that non-contracting axes are represented in the same order
291     // in the matrix as they are in the tensor. This is not the case for
292     // contracting axes. As the contracting axes must be of the same size in
293     // each tensor, we'll only look at the first tensor here.
294     m_rhs_inner_dim_contiguous = true;
295     m_rhs_inner_dim_reordered = false;
296     for (int i = 0; i < ContractDims; i++) {
297       Index left = eval_op_indices[i].first;
298       Index right = eval_op_indices[i].second;
299 
300       Index size = eval_left_dims[left];
301       eigen_assert(size == eval_right_dims[right] &&
302                    "Contraction axes must be same size");
303 
304       if (i+1 < static_cast<int>(internal::array_size<contract_t>::value)) {
305         m_k_strides[i+1] = m_k_strides[i] * size;
306       } else {
307         m_k_size = m_k_strides[i] * size;
308       }
309       m_left_contracting_strides[i] = lhs_strides[left];
310       m_right_contracting_strides[i] = rhs_strides[right];
311 
312       if (i > 0 && right < eval_op_indices[i-1].second) {
313         m_rhs_inner_dim_reordered = true;
314       }
315       if (right != i) {
316         m_rhs_inner_dim_contiguous = false;
317       }
318     }
319 
320     // If the layout is RowMajor, we need to reverse the m_dimensions
321     if (static_cast<int>(Layout) == static_cast<int>(RowMajor)) {
322       for (int i = 0, j = NumDims - 1; i < j; i++, j--) {
323         numext::swap(m_dimensions[i], m_dimensions[j]);
324       }
325     }
326   }
327 
328   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
329 
330   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
331     m_leftImpl.evalSubExprsIfNeeded(NULL);
332     m_rightImpl.evalSubExprsIfNeeded(NULL);
333     if (data) {
334       evalTo(data);
335       return false;
336     } else {
337       m_result = static_cast<Scalar *>(m_device.allocate(dimensions().TotalSize() * sizeof(Scalar)));
338       evalTo(m_result);
339       return true;
340     }
341   }
342 
343   EIGEN_DEVICE_FUNC void evalTo(Scalar* buffer) const {
344     if (this->m_lhs_inner_dim_contiguous) {
345       if (this->m_rhs_inner_dim_contiguous) {
346         if (this->m_rhs_inner_dim_reordered) {
347           static_cast<const Derived*>(this)->template evalProduct<true, true, true, Unaligned>(buffer);
348         }
349         else {
350           static_cast<const Derived*>(this)->template evalProduct<true, true, false, Unaligned>(buffer);
351         }
352       }
353       else {
354        if (this->m_rhs_inner_dim_reordered) {
355           static_cast<const Derived*>(this)->template evalProduct<true, false, true, Unaligned>(buffer);
356         }
357         else {
358           static_cast<const Derived*>(this)->template evalProduct<true, false, false, Unaligned>(buffer);
359         }
360       }
361     }
362     else {
363       if (this->m_rhs_inner_dim_contiguous) {
364         if (this->m_rhs_inner_dim_reordered) {
365           static_cast<const Derived*>(this)->template evalProduct<false, true, true, Unaligned>(buffer);
366         }
367         else {
368           static_cast<const Derived*>(this)->template evalProduct<false, true, false, Unaligned>(buffer);
369         }
370       }
371       else {
372        if (this->m_rhs_inner_dim_reordered) {
373           static_cast<const Derived*>(this)->template evalProduct<false, false, true, Unaligned>(buffer);
374         }
375         else {
376           static_cast<const Derived*>(this)->template evalProduct<false, false, false, Unaligned>(buffer);
377         }
378       }
379     }
380   }
381 
382   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
383   EIGEN_DEVICE_FUNC void evalGemv(Scalar* buffer) const {
384     const Index rows = m_i_size;
385     const Index cols = m_k_size;
386 
387     typedef typename internal::remove_const<typename EvalLeftArgType::Scalar>::type LhsScalar;
388     typedef typename internal::remove_const<typename EvalRightArgType::Scalar>::type RhsScalar;
389     typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
390     typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
391     const Index lhs_packet_size = internal::unpacket_traits<typename LeftEvaluator::PacketReturnType>::size;
392     const Index rhs_packet_size = internal::unpacket_traits<typename RightEvaluator::PacketReturnType>::size;
393     const int lhs_alignment = LeftEvaluator::IsAligned ? Aligned : Unaligned;
394     const int rhs_alignment = RightEvaluator::IsAligned ? Aligned : Unaligned;
395     typedef internal::TensorContractionInputMapper<LhsScalar, Index, internal::Lhs,
396                                                    LeftEvaluator, left_nocontract_t,
397                                                    contract_t, lhs_packet_size,
398                                                    lhs_inner_dim_contiguous,
399                                                    false, lhs_alignment> LhsMapper;
400 
401     typedef internal::TensorContractionInputMapper<RhsScalar, Index, internal::Rhs,
402                                                    RightEvaluator, right_nocontract_t,
403                                                    contract_t, rhs_packet_size,
404                                                    rhs_inner_dim_contiguous,
405                                                    rhs_inner_dim_reordered, rhs_alignment> RhsMapper;
406 
407     LhsMapper lhs(m_leftImpl, m_left_nocontract_strides, m_i_strides,
408                   m_left_contracting_strides, m_k_strides);
409     RhsMapper rhs(m_rightImpl, m_right_nocontract_strides, m_j_strides,
410                   m_right_contracting_strides, m_k_strides);
411 
412     const Scalar alpha(1);
413     const Index resIncr(1);
414 
415     // zero out the result buffer (which must be of size at least rows * sizeof(Scalar)
416     m_device.memset(buffer, 0, rows * sizeof(Scalar));
417 
418     internal::general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,false,RhsScalar,RhsMapper,false>::run(
419         rows, cols, lhs, rhs,
420         buffer, resIncr, alpha);
421   }
422 
423   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
424   EIGEN_DEVICE_FUNC void evalGemm(Scalar* buffer) const {
425     // columns in left side, rows in right side
426     const Index k = this->m_k_size;
427 
428     // rows in left side
429     const Index m = this->m_i_size;
430 
431     // columns in right side
432     const Index n = this->m_j_size;
433 
434     // zero out the result buffer (which must be of size at least m * n * sizeof(Scalar)
435     this->m_device.memset(buffer, 0, m * n * sizeof(Scalar));
436 
437     // define mr, nr, and all of my data mapper types
438     typedef typename internal::remove_const<typename EvalLeftArgType::Scalar>::type LhsScalar;
439     typedef typename internal::remove_const<typename EvalRightArgType::Scalar>::type RhsScalar;
440     typedef typename internal::gebp_traits<LhsScalar, RhsScalar> Traits;
441 
442     const Index nr = Traits::nr;
443     const Index mr = Traits::mr;
444 
445     typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
446     typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
447 
448     const Index lhs_packet_size = internal::unpacket_traits<typename LeftEvaluator::PacketReturnType>::size;
449     const Index rhs_packet_size = internal::unpacket_traits<typename RightEvaluator::PacketReturnType>::size;
450 
451     typedef internal::TensorContractionInputMapper<LhsScalar, Index, internal::Lhs,
452                                                    LeftEvaluator, left_nocontract_t,
453                                                    contract_t, lhs_packet_size,
454                                                    lhs_inner_dim_contiguous,
455                                                    false, Unaligned> LhsMapper;
456 
457     typedef internal::TensorContractionInputMapper<RhsScalar, Index, internal::Rhs,
458                                                    RightEvaluator, right_nocontract_t,
459                                                    contract_t, rhs_packet_size,
460                                                    rhs_inner_dim_contiguous,
461                                                    rhs_inner_dim_reordered, Unaligned> RhsMapper;
462 
463     typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
464 
465     // Declare GEBP packing and kernel structs
466     internal::gemm_pack_lhs<LhsScalar, Index, typename LhsMapper::SubMapper, mr, Traits::LhsProgress, ColMajor> pack_lhs;
467     internal::gemm_pack_rhs<RhsScalar, Index, typename RhsMapper::SubMapper, nr, ColMajor> pack_rhs;
468 
469     internal::gebp_kernel<LhsScalar, RhsScalar, Index, OutputMapper, mr, nr, false, false> gebp;
470 
471     // initialize data mappers
472     LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides,
473                   this->m_left_contracting_strides, this->m_k_strides);
474 
475     RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides,
476                   this->m_right_contracting_strides, this->m_k_strides);
477 
478     OutputMapper output(buffer, m);
479 
480     // Sizes of the blocks to load in cache. See the Goto paper for details.
481     internal::TensorContractionBlocking<LhsMapper, RhsMapper, Index, internal::ShardByCol> blocking(k, m, n, 1);
482     const Index kc = blocking.kc();
483     const Index mc = numext::mini(m, blocking.mc());
484     const Index nc = numext::mini(n, blocking.nc());
485     const Index sizeA = mc * kc;
486     const Index sizeB = kc * nc;
487 
488     LhsScalar* blockA = static_cast<LhsScalar *>(this->m_device.allocate(sizeA * sizeof(LhsScalar)));
489     RhsScalar* blockB = static_cast<RhsScalar *>(this->m_device.allocate(sizeB * sizeof(RhsScalar)));
490 
491     for(Index i2=0; i2<m; i2+=mc)
492     {
493       const Index actual_mc = numext::mini(i2+mc,m)-i2;
494       for (Index k2 = 0; k2 < k; k2 += kc) {
495         // make sure we don't overshoot right edge of left matrix, then pack vertical panel
496         const Index actual_kc = numext::mini(k2 + kc, k) - k2;
497         pack_lhs(blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc, 0, 0);
498 
499         // series of horizontal blocks
500         for (Index j2 = 0; j2 < n; j2 += nc) {
501           // make sure we don't overshoot right edge of right matrix, then pack block
502           const Index actual_nc = numext::mini(j2 + nc, n) - j2;
503           pack_rhs(blockB, rhs.getSubMapper(k2, j2), actual_kc, actual_nc, 0, 0);
504 
505           // call gebp (matrix kernel)
506           // The parameters here are copied from Eigen's GEMM implementation
507           gebp(output.getSubMapper(i2, j2), blockA, blockB, actual_mc, actual_kc, actual_nc, Scalar(1), -1, -1, 0, 0);
508         }
509       }
510     }
511 
512     this->m_device.deallocate(blockA);
513     this->m_device.deallocate(blockB);
514   }
515 
516   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
517     m_leftImpl.cleanup();
518     m_rightImpl.cleanup();
519 
520     if (m_result != NULL) {
521       m_device.deallocate(m_result);
522       m_result = NULL;
523     }
524   }
525 
526   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const {
527     return m_result[index];
528   }
529 
530   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool) const {
531     return TensorOpCost(sizeof(CoeffReturnType), 0, 0);
532   }
533 
534   template<int LoadMode>
535   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const {
536     return internal::ploadt<PacketReturnType, LoadMode>(m_result + index);
537   }
538 
539   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar* data() const { return m_result; }
540 
541   protected:
542   // Prevent assignment
543   TensorContractionEvaluatorBase& operator = (const TensorContractionEvaluatorBase&);
544   Dimensions m_dimensions;
545 
546   contract_t m_k_strides;
547   contract_t m_left_contracting_strides;
548   contract_t m_right_contracting_strides;
549 
550   bool m_lhs_inner_dim_contiguous;
551   bool m_rhs_inner_dim_contiguous;
552   bool m_rhs_inner_dim_reordered;
553 
554   left_nocontract_t m_i_strides;
555   right_nocontract_t m_j_strides;
556   left_nocontract_t m_left_nocontract_strides;
557   right_nocontract_t m_right_nocontract_strides;
558 
559   Index m_i_size;
560   Index m_j_size;
561   Index m_k_size;
562 
563   TensorEvaluator<EvalLeftArgType, Device> m_leftImpl;
564   TensorEvaluator<EvalRightArgType, Device> m_rightImpl;
565   const Device& m_device;
566   Scalar* m_result;
567 };
568 
569 
570 // evaluator for default device
571 template<typename Indices, typename LeftArgType, typename RightArgType, typename Device>
572 struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> :
573     public TensorContractionEvaluatorBase<
574       TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> > {
575   typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType>, Device> Self;
576   typedef TensorContractionEvaluatorBase<Self> Base;
577 
578   typedef TensorContractionOp<Indices, LeftArgType, RightArgType> XprType;
579   typedef typename internal::remove_const<typename XprType::Scalar>::type Scalar;
580   typedef typename XprType::Index Index;
581   typedef typename XprType::CoeffReturnType CoeffReturnType;
582   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
583 
584   enum {
585     Layout = TensorEvaluator<LeftArgType, Device>::Layout
586   };
587 
588   // Most of the code is assuming that both input tensors are ColMajor. If the
589   // inputs are RowMajor, we will "cheat" by swapping the LHS and RHS:
590   // If we want to compute A * B = C, where A is LHS and B is RHS, the code
591   // will pretend B is LHS and A is RHS.
592   typedef typename internal::conditional<
593     static_cast<int>(Layout) == static_cast<int>(ColMajor), LeftArgType, RightArgType>::type EvalLeftArgType;
594   typedef typename internal::conditional<
595     static_cast<int>(Layout) == static_cast<int>(ColMajor), RightArgType, LeftArgType>::type EvalRightArgType;
596 
597   static const int LDims =
598       internal::array_size<typename TensorEvaluator<EvalLeftArgType, Device>::Dimensions>::value;
599   static const int RDims =
600       internal::array_size<typename TensorEvaluator<EvalRightArgType, Device>::Dimensions>::value;
601   static const int ContractDims = internal::array_size<Indices>::value;
602 
603   typedef array<Index, ContractDims> contract_t;
604   typedef array<Index, LDims - ContractDims> left_nocontract_t;
605   typedef array<Index, RDims - ContractDims> right_nocontract_t;
606 
607   static const int NumDims = LDims + RDims - 2 * ContractDims;
608 
609   // Could we use NumDimensions here?
610   typedef DSizes<Index, NumDims> Dimensions;
611 
612   EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device) :
613       Base(op, device) { }
614 
615   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
616   EIGEN_DEVICE_FUNC void evalProduct(Scalar* buffer) const {
617     if (this->m_j_size == 1) {
618       this->template evalGemv<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Alignment>(buffer);
619       return;
620     }
621 
622     this->template evalGemm<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Alignment>(buffer);
623   }
624 };
625 
626 } // end namespace Eigen
627 
628 #endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_H
629