xref: /aosp_15_r20/external/tensorflow/tensorflow/core/kernels/tensor_array.cc (revision b6fb3261f9314811a0f4371741dbb8839866f948)
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 #define EIGEN_USE_THREADS
17 #include "tensorflow/core/kernels/tensor_array.h"
18 
19 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
20 #include "tensorflow/core/framework/register_types.h"
21 #include "tensorflow/core/framework/tensor_util.h"
22 #include "tensorflow/core/kernels/aggregate_ops_cpu.h"
23 
24 namespace tensorflow {
25 
26 typedef Eigen::ThreadPoolDevice CPUDevice;
27 typedef Eigen::GpuDevice GPUDevice;
28 
29 namespace tensor_array {
30 
31 #define TENSOR_ARRAY_WRITE_OR_ADD(Device, T)                                \
32   template <>                                                               \
33   Status AddToTensor<Device, T>(OpKernelContext * ctx, Tensor * sum,        \
34                                 const Tensor* current, const Tensor* add) { \
35     functor::Add2Functor<Device, T> add_functor;                            \
36     add_functor(ctx->template eigen_device<Device>(), sum->flat<T>(),       \
37                 current->flat<T>(), add->flat<T>());                        \
38     return OkStatus();                                                      \
39   }
40 
41 #define TENSOR_ARRAY_WRITE_OR_ADD_CPU(T) TENSOR_ARRAY_WRITE_OR_ADD(CPUDevice, T)
42 TF_CALL_NUMBER_TYPES(TENSOR_ARRAY_WRITE_OR_ADD_CPU)
43 #undef TENSOR_ARRAY_WRITE_OR_ADD_CPU
44 
45 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
46 
47 #define TENSOR_ARRAY_WRITE_OR_ADD_GPU(T) TENSOR_ARRAY_WRITE_OR_ADD(GPUDevice, T)
48 TF_CALL_GPU_NUMBER_TYPES(TENSOR_ARRAY_WRITE_OR_ADD_GPU);
49 TF_CALL_COMPLEX_TYPES(TENSOR_ARRAY_WRITE_OR_ADD_GPU);
50 #undef TENSOR_ARRAY_WRITE_OR_ADD_GPU
51 
52 #endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
53 
54 #undef TENSOR_ARRAY_WRITE_OR_ADD
55 
56 #define TENSOR_ARRAY_SET_ZERO(Device, T)                                      \
57   template <>                                                                 \
58   Status TensorSetZero<Device, T>(OpKernelContext * ctx, Tensor * value) {    \
59     functor::SetZeroFunctor<Device, T> set_zero_functor;                      \
60     set_zero_functor(ctx->template eigen_device<Device>(), value->flat<T>()); \
61     return OkStatus();                                                        \
62   }
63 
64 #define TENSOR_ARRAY_SET_ZERO_CPU(T) TENSOR_ARRAY_SET_ZERO(CPUDevice, T)
65 TF_CALL_NUMBER_TYPES(TENSOR_ARRAY_SET_ZERO_CPU);
66 TF_CALL_bool(TENSOR_ARRAY_SET_ZERO_CPU);
67 #undef TENSOR_ARRAY_SET_ZERO_CPU
68 
69 #if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
70 
71 #define TENSOR_ARRAY_SET_ZERO_GPU(T) TENSOR_ARRAY_SET_ZERO(GPUDevice, T)
72 TF_CALL_GPU_NUMBER_TYPES(TENSOR_ARRAY_SET_ZERO_GPU);
73 TF_CALL_COMPLEX_TYPES(TENSOR_ARRAY_SET_ZERO_GPU);
74 #undef TENSOR_ARRAY_SET_ZERO_GPU
75 
76 #endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
77 
78 #undef TENSOR_ARRAY_SET_ZERO
79 
80 }  // namespace tensor_array
81 
82 std::atomic<int64_t> TensorArray::tensor_array_counter{0};
83 
CopyShapesFrom(TensorArray * rhs,const TensorShape * shape_to_prepend)84 Status TensorArray::CopyShapesFrom(TensorArray* rhs,
85                                    const TensorShape* shape_to_prepend) {
86   mutex_lock l(mu_);
87   mutex_lock l_rhs(rhs->mu_);
88   TF_RETURN_IF_ERROR(LockedReturnIfClosed());
89   TF_RETURN_IF_ERROR(rhs->LockedReturnIfClosed());
90   if (tensors_.size() != rhs->tensors_.size()) {
91     return errors::InvalidArgument(
92         "TensorArray sizes do not match during CopyShapesFrom: ",
93         handle_.vec<tstring>()(1), " has size ", tensors_.size(), " but rhs ",
94         rhs->handle_.vec<tstring>()(1), " has size ", rhs->tensors_.size());
95   }
96   for (std::size_t i = 0; i < tensors_.size(); ++i) {
97     // Skip "soft copy" of indices which have not been written.
98     if (!rhs->tensors_[i].written) continue;
99 
100     // Copy the shape over.
101     if (shape_to_prepend) {
102       tensors_[i].shape = *shape_to_prepend;
103       tensors_[i].shape.AppendShape(rhs->tensors_[i].shape);
104     } else {
105       tensors_[i].shape = rhs->tensors_[i].shape;
106     }
107     // Mark as written.  Reads will know that if written is true and
108     // read is false, and cleared is false, to return zeros of the
109     // appropriate shape.  Future aggregating writes will only use the shape
110     // for validation.
111     tensors_[i].written = true;
112   }
113 
114   return OkStatus();
115 }
116 
117 }  // namespace tensorflow
118