• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #define EIGEN_USE_THREADS
17 
18 #include <complex>
19 
20 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
21 #include "tensorflow/core/framework/attr_value.pb.h"
22 #include "tensorflow/core/framework/tensor.h"
23 #include "tensorflow/core/kernels/ops_util.h"
24 #include "tensorflow/core/kernels/transpose_functor.h"
25 #include "tensorflow/core/lib/core/status.h"
26 #include "tensorflow/core/lib/gtl/array_slice.h"
27 #include "tensorflow/core/lib/gtl/inlined_vector.h"
28 
29 typedef Eigen::ThreadPoolDevice CPUDevice;
30 
31 namespace tensorflow {
32 namespace {
33 
34 template <typename T, bool conjugate>
TransposeSimple(const CPUDevice & device,const Tensor & in,const gtl::ArraySlice<int32> perm,Tensor * out)35 void TransposeSimple(const CPUDevice& device, const Tensor& in,
36                      const gtl::ArraySlice<int32> perm, Tensor* out) {
37   const int ndims = in.dims();
38   gtl::InlinedVector<int64_t, 8> in_strides =
39       ComputeStride<int64_t>(in.shape());
40   gtl::InlinedVector<int64_t, 8> out_strides =
41       ComputeStride<int64_t>(out->shape());
42   const T* p = reinterpret_cast<const T*>(in.tensor_data().data());
43   T* q = reinterpret_cast<T*>(const_cast<char*>((out->tensor_data().data())));
44   auto transpose_fn = [=, &in_strides, &out_strides, &perm](int64_t begin,
45                                                             int64_t end) {
46     for (int64_t o_idx = begin; o_idx < end; ++o_idx) {
47       int64_t i_idx = 0;
48       int64_t t = o_idx;
49       for (int i = 0; i < ndims; ++i) {
50         const int64_t ratio = t / out_strides[i];
51         t -= ratio * out_strides[i];
52         i_idx += ratio * in_strides[perm[i]];
53       }
54       if (conjugate) {
55         q[o_idx] = Eigen::numext::conj(p[i_idx]);
56       } else {
57         q[o_idx] = p[i_idx];
58       }
59     }
60   };
61   double cycles_per_element =
62       (conjugate ? 1 : 0) +
63       ndims * (Eigen::TensorOpCost::DivCost<int64_t>() +
64                2 * Eigen::TensorOpCost::MulCost<int64_t>() +
65                2 * Eigen::TensorOpCost::AddCost<int64_t>());
66   Eigen::TensorOpCost cost(/*bytes_loaded=*/sizeof(T),
67                            /*bytes_stored=*/sizeof(T), cycles_per_element);
68   device.parallelFor(in.NumElements(), cost, std::move(transpose_fn));
69 }
70 
71 }  // namespace
72 
73 template <typename T, bool conjugate>
74 struct Transpose<CPUDevice, T, conjugate> {
runtensorflow::Transpose75   static void run(const CPUDevice& d, const Tensor& in,
76                   const gtl::ArraySlice<int32> perm, Tensor* out) {
77     switch (in.dims()) {
78       case 2:
79         internal::TransposeUsingEigen<CPUDevice, T, 2>(d, in, perm, conjugate,
80                                                        out);
81         break;
82       case 3:
83         internal::TransposeUsingEigen<CPUDevice, T, 3>(d, in, perm, conjugate,
84                                                        out);
85         break;
86       case 4:
87         internal::TransposeUsingEigen<CPUDevice, T, 4>(d, in, perm, conjugate,
88                                                        out);
89         break;
90       case 5:
91         internal::TransposeUsingEigen<CPUDevice, T, 5>(d, in, perm, conjugate,
92                                                        out);
93         break;
94       case 6:
95         internal::TransposeUsingEigen<CPUDevice, T, 6>(d, in, perm, conjugate,
96                                                        out);
97         break;
98       case 7:
99         internal::TransposeUsingEigen<CPUDevice, T, 7>(d, in, perm, conjugate,
100                                                        out);
101         break;
102       case 8:
103         internal::TransposeUsingEigen<CPUDevice, T, 8>(d, in, perm, conjugate,
104                                                        out);
105         break;
106       default:
107         TransposeSimple<T, conjugate>(d, in, perm, out);
108         break;
109     }
110   }
111 };
112 
113 #define INSTANTIATE(DEVICE)                                                 \
114   template <>                                                               \
115   Status DoTranspose(const DEVICE& device, const Tensor& in,                \
116                      const gtl::ArraySlice<int32> perm, Tensor* out) {      \
117     return internal::DoTransposeImpl(device, in, perm, /*conjugate=*/false, \
118                                      out);                                  \
119   }                                                                         \
120   template <>                                                               \
121   Status DoConjugateTranspose(const DEVICE& device, const Tensor& in,       \
122                               const gtl::ArraySlice<int32> perm,            \
123                               Tensor* out) {                                \
124     return internal::DoTransposeImpl(device, in, perm, /*conjugate=*/true,  \
125                                      out);                                  \
126   }                                                                         \
127   template <>                                                               \
128   Status DoMatrixTranspose(const DEVICE& device, const Tensor& in,          \
129                            Tensor* out) {                                   \
130     return internal::DoMatrixTransposeImpl(device, in, /*conjugate=*/false, \
131                                            out);                            \
132   }                                                                         \
133   template <>                                                               \
134   Status DoConjugateMatrixTranspose(const DEVICE& device, const Tensor& in, \
135                                     Tensor* out) {                          \
136     return internal::DoMatrixTransposeImpl(device, in, /*conjugate=*/true,  \
137                                            out);                            \
138   }
139 
140 INSTANTIATE(CPUDevice)
141 
142 
143 }  // namespace tensorflow
144