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
18 #include "tensorflow/core/framework/device_base.h"
19
20 #include <algorithm>
21 #include <vector>
22
23 #include "absl/container/flat_hash_set.h"
24 #include "absl/synchronization/notification.h"
25 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
26 #include "tensorflow/core/util/work_sharder.h"
27
28 namespace tensorflow {
29
~DeviceBase()30 DeviceBase::~DeviceBase() {
31 for (auto& temp : eigen_cpu_devices_) {
32 delete temp;
33 }
34 eigen_cpu_devices_.clear();
35 }
36
CopyDeviceTensorToCPUSync(const Tensor * device_tensor,StringPiece tensor_name,Device * device,Tensor * cpu_tensor)37 Status DeviceContext::CopyDeviceTensorToCPUSync(const Tensor* device_tensor,
38 StringPiece tensor_name,
39 Device* device,
40 Tensor* cpu_tensor) {
41 absl::Notification n;
42 Status status;
43 CopyDeviceTensorToCPU(device_tensor, tensor_name, device, cpu_tensor,
44 [&](const Status& s) {
45 status = s;
46 n.Notify();
47 });
48 n.WaitForNotification();
49 return status;
50 }
51
CopyCPUTensorToDeviceSync(const Tensor * cpu_tensor,Device * device,Tensor * device_tensor) const52 Status DeviceContext::CopyCPUTensorToDeviceSync(const Tensor* cpu_tensor,
53 Device* device,
54 Tensor* device_tensor) const {
55 absl::Notification n;
56 Status status;
57 CopyCPUTensorToDevice(cpu_tensor, device, device_tensor,
58 [&](const Status& s) {
59 status = s;
60 n.Notify();
61 });
62 n.WaitForNotification();
63 return status;
64 }
65
attributes() const66 const DeviceAttributes& DeviceBase::attributes() const {
67 LOG(FATAL) << "Device does not implement attributes()";
68 std::abort();
69 }
70
name() const71 const string& DeviceBase::name() const {
72 LOG(FATAL) << "Device does not implement name()";
73 std::abort();
74 }
75
set_eigen_cpu_device(Eigen::ThreadPoolDevice * d)76 void DeviceBase::set_eigen_cpu_device(Eigen::ThreadPoolDevice* d) {
77 // Eigen::ThreadPoolDevice is a very cheap struct (two pointers and
78 // an int). Therefore, we can afford a pre-allocated array of
79 // Eigen::ThreadPoolDevice. Here, we ensure that
80 // Eigen::ThreadPoolDevices in eigen_cpu_devices_ has increasingly
81 // larger numThreads.
82 for (int i = 1; i <= d->numThreads(); ++i) {
83 eigen_cpu_devices_.push_back(new Eigen::ThreadPoolDevice(
84 d->getPool(), i /* numThreads() */, d->allocator()));
85 }
86 }
87
eigen_cpu_device()88 const Eigen::ThreadPoolDevice* DeviceBase::eigen_cpu_device() {
89 // Based on GetPerThreadMaxParallelism(), we return a different
90 // pre-allocated Eigen::ThreadPoolDevice. All these ThreadPoolDevice
91 // use the same underlying threadpool. But they use different
92 // nominal numThreads() hoping that the user of the returned
93 // Eigen::ThreadPoolDevice may not aggressively occupy all the
94 // threads in the underlying threadpool.
95 const int parallelism = std::max<int>(
96 1,
97 std::min<int>(GetPerThreadMaxParallelism(), eigen_cpu_devices_.size()));
98 return eigen_cpu_devices_[parallelism - 1];
99 }
100
101 namespace {
102
GetSymbolicDeviceList()103 absl::flat_hash_set<std::string>* GetSymbolicDeviceList() {
104 static absl::flat_hash_set<std::string>* symbolic_device_list =
105 new absl::flat_hash_set<std::string>();
106 return symbolic_device_list;
107 }
108
109 } // namespace
110
AddSymbolicExecutionDevice(const absl::string_view device_name)111 void AddSymbolicExecutionDevice(const absl::string_view device_name) {
112 GetSymbolicDeviceList()->insert(std::string(device_name));
113 }
114
IsSymbolicExecutionDevice(const absl::string_view device_name)115 bool IsSymbolicExecutionDevice(const absl::string_view device_name) {
116 absl::flat_hash_set<std::string>* symbolic_devices = GetSymbolicDeviceList();
117 if (symbolic_devices->contains(device_name)) {
118 return true;
119 } else {
120 return false;
121 }
122 }
123
124 } // namespace tensorflow
125