1 /* Copyright 2015 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 #ifndef TENSORFLOW_UTIL_TENSOR_SLICE_UTIL_H_ 17 #define TENSORFLOW_UTIL_TENSOR_SLICE_UTIL_H_ 18 19 #include "tensorflow/core/framework/tensor_shape.h" 20 #include "tensorflow/core/framework/tensor_slice.h" 21 #include "tensorflow/core/platform/logging.h" 22 23 namespace tensorflow { 24 25 namespace { 26 27 // Some hackery to invoke eigen tensor to copy over tensor slices with variable 28 // dimension tensors. 29 // TODO(yangke): get rid of that once the variable dimension tensor support is 30 // in. 31 static const int kTensorSliceMaxRank = 8; 32 33 // Create a tensor map with the given shape: we support up to 8 dimensions. If 34 // the shape has less than 8 dimensions, we pad the remaining dimension with 1. 35 template <typename T> 36 Eigen::TensorMap<Eigen::Tensor<T, kTensorSliceMaxRank, Eigen::RowMajor>> 37 GetEigenTensorMapFromTensorShape(const TensorShape& shape, T* data) { 38 Eigen::DSizes<Eigen::DenseIndex, kTensorSliceMaxRank> dsizes = 39 shape.AsEigenDSizesWithPadding<kTensorSliceMaxRank>(); 40 Eigen::TensorMap<Eigen::Tensor<T, kTensorSliceMaxRank, Eigen::RowMajor>> eig( 41 data, dsizes); 42 return eig; 43 } 44 45 // For everything except string, a standard Eigen cast and assignment works 46 template <typename DstT> 47 struct CopyThatWorksWithStringPointer { 48 template <typename SrcTensor, typename DstTensor, typename Shape> 49 static void Copy(const SrcTensor& s, Shape s_start, Shape len, DstTensor& d, 50 Shape d_start) { 51 d.slice(d_start, len) = s.slice(s_start, len).template cast<DstT>(); 52 } 53 }; 54 55 // Eigen makes it extremely difficult to dereference a tensor of string* into 56 // string, so we roll our own loop instead. 57 template <> 58 struct CopyThatWorksWithStringPointer<string> { 59 template <typename SrcTensor, typename DstTensor, typename Shape> 60 static void Copy(const SrcTensor& s, Shape s_start, Shape len, DstTensor& d, 61 Shape d_start) { 62 typedef typename SrcTensor::Index Index; 63 static_assert(kTensorSliceMaxRank == 8, 64 "If kTensorSliceMaxRank changes, modify the loop below."); 65 for (Index i0 = 0; i0 < len[0]; i0++) { 66 for (Index i1 = 0; i1 < len[1]; i1++) { 67 for (Index i2 = 0; i2 < len[2]; i2++) { 68 for (Index i3 = 0; i3 < len[3]; i3++) { 69 for (Index i4 = 0; i4 < len[4]; i4++) { 70 for (Index i5 = 0; i5 < len[5]; i5++) { 71 for (Index i6 = 0; i6 < len[6]; i6++) { 72 for (Index i7 = 0; i7 < len[7]; i7++) { 73 d(d_start[0] + i0, d_start[1] + i1, d_start[2] + i2, 74 d_start[3] + i3, d_start[4] + i4, d_start[5] + i5, 75 d_start[6] + i6, d_start[7] + i7) = 76 *s(s_start[0] + i0, s_start[1] + i1, s_start[2] + i2, 77 s_start[3] + i3, s_start[4] + i4, s_start[5] + i5, 78 s_start[6] + i6, s_start[7] + i7); 79 } 80 } 81 } 82 } 83 } 84 } 85 } 86 } 87 } 88 }; 89 90 // Checkpointing of half is done by storing the raw 16 bits as a signed 32bit 91 // integer. To restore the checkpoint we need to do the reverse operation by 92 // reinterpreting the integer as a 16 bit float. This prevents us from using 93 // the default cast operation. 94 template <> 95 struct CopyThatWorksWithStringPointer<Eigen::half> { 96 template <typename SrcTensor, typename DstTensor, typename Shape> 97 static void Copy(const SrcTensor& s, Shape s_start, Shape len, DstTensor& d, 98 Shape d_start) { 99 typedef typename SrcTensor::Index Index; 100 static_assert(kTensorSliceMaxRank == 8, 101 "If kTensorSliceMaxRank changes, modify the loop below."); 102 for (Index i0 = 0; i0 < len[0]; i0++) { 103 for (Index i1 = 0; i1 < len[1]; i1++) { 104 for (Index i2 = 0; i2 < len[2]; i2++) { 105 for (Index i3 = 0; i3 < len[3]; i3++) { 106 for (Index i4 = 0; i4 < len[4]; i4++) { 107 for (Index i5 = 0; i5 < len[5]; i5++) { 108 for (Index i6 = 0; i6 < len[6]; i6++) { 109 for (Index i7 = 0; i7 < len[7]; i7++) { 110 d(d_start[0] + i0, d_start[1] + i1, d_start[2] + i2, 111 d_start[3] + i3, d_start[4] + i4, d_start[5] + i5, 112 d_start[6] + i6, d_start[7] + i7) = 113 Eigen::half_impl::raw_uint16_to_half( 114 s(s_start[0] + i0, s_start[1] + i1, s_start[2] + i2, 115 s_start[3] + i3, s_start[4] + i4, s_start[5] + i5, 116 s_start[6] + i6, s_start[7] + i7)); 117 } 118 } 119 } 120 } 121 } 122 } 123 } 124 } 125 } 126 }; 127 128 // Given a tensor described by "shape", two slices "slice_s" and "slice_d", 129 // and two pointers "ptr_s" and "ptr_d", where "ptr_s" points to a chunk of 130 // memory that stores the data for "slice_s" and "ptr_d" points to a chunk of 131 // memory that stores the data for "slice_d". This function copies the data 132 // that belongs to the intersection of the two slices from slice_s to 133 // slice_d. Uses Tensor cast<DstT>() to convert from SrcT to DstT. Returns true 134 // iff the two slices share any intersection (and thus some data is copied). 135 // TODO(yangke): figure out if we can make it private. 136 template <typename SrcT, typename DstT> 137 static bool CopyDataFromTensorSliceToTensorSlice(const TensorShape& shape, 138 const TensorSlice& slice_s, 139 const TensorSlice& slice_d, 140 const SrcT* ptr_s, 141 DstT* ptr_d) { 142 CHECK_LE(shape.dims(), kTensorSliceMaxRank) 143 << "Only tensors of size up to " << kTensorSliceMaxRank 144 << " are supported"; 145 // We need to compute the intersection of the two slices. 146 TensorSlice inter; 147 if (!slice_s.Intersect(slice_d, &inter)) { 148 // There is no intersection: returns false. 149 return false; 150 } else { 151 // We need to compute the applied shapes after applying slice_s and 152 // slice_d. 153 TensorShape shp_s, shp_d; 154 Status s; 155 s = slice_s.SliceTensorShape(shape, &shp_s); 156 if (!s.ok()) { 157 LOG(WARNING) << s; 158 return false; 159 } 160 s = slice_d.SliceTensorShape(shape, &shp_d); 161 if (!s.ok()) { 162 LOG(WARNING) << s; 163 return false; 164 } 165 166 // We need to compute the relative slice of "inter" w.r.t. both slice_s and 167 // slice_d. 168 TensorSlice rel_s, rel_d; 169 slice_s.ComputeRelative(inter, &rel_s); 170 slice_d.ComputeRelative(inter, &rel_d); 171 172 // Get the eigen tensor maps to the data. 173 auto t_s = GetEigenTensorMapFromTensorShape(shp_s, ptr_s); 174 auto t_d = GetEigenTensorMapFromTensorShape(shp_d, ptr_d); 175 176 Eigen::DSizes<Eigen::DenseIndex, kTensorSliceMaxRank> s_start, s_len, 177 d_start, d_len; 178 179 rel_s.FillIndicesAndSizes<kTensorSliceMaxRank>(shp_s, &s_start, &s_len); 180 rel_d.FillIndicesAndSizes<kTensorSliceMaxRank>(shp_d, &d_start, &d_len); 181 CopyThatWorksWithStringPointer<DstT>::Copy(t_s, s_start, s_len, t_d, 182 d_start); 183 return true; 184 } 185 } 186 187 } // namespace 188 189 } // namespace tensorflow 190 191 #endif // TENSORFLOW_UTIL_TENSOR_SLICE_UTIL_H_ 192
