• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2018 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 #include "tensorflow/core/grappler/optimizers/evaluation_utils.h"
17 
18 #include "tensorflow/core/framework/tensor.pb.h"
19 #include "tensorflow/core/lib/core/threadpool.h"
20 #include "tensorflow/core/platform/cpu_info.h"
21 #include "tensorflow/core/platform/denormal.h"
22 #include "tensorflow/core/platform/setround.h"
23 #include "tensorflow/core/public/version.h"
24 
25 namespace tensorflow {
26 namespace grappler {
27 using TensorVector = gtl::InlinedVector<TensorValue, 4>;
28 
29 namespace {
30 class EigenThreadPoolWrapper : public Eigen::ThreadPoolInterface {
31  public:
EigenThreadPoolWrapper(thread::ThreadPool * pool)32   explicit EigenThreadPoolWrapper(thread::ThreadPool* pool) : pool_(pool) {}
~EigenThreadPoolWrapper()33   ~EigenThreadPoolWrapper() override {}
Schedule(std::function<void ()> fn)34   void Schedule(std::function<void()> fn) override {
35     auto wrapped = [=]() {
36       // TensorFlow flushes denormals to zero and rounds to nearest, so we do
37       // the same here.
38       port::ScopedFlushDenormal flush;
39       port::ScopedSetRound round(FE_TONEAREST);
40       fn();
41     };
42     pool_->Schedule(std::move(wrapped));
43   }
NumThreads() const44   int NumThreads() const override { return pool_->NumThreads(); }
CurrentThreadId() const45   int CurrentThreadId() const override { return pool_->CurrentThreadId(); }
46 
47  private:
48   thread::ThreadPool* pool_ = nullptr;
49 };
50 
51 }  // namespace
52 
DeviceSimple()53 DeviceSimple::DeviceSimple() : DeviceBase(Env::Default()) {
54   eigen_worker_threads_.num_threads = port::NumSchedulableCPUs();
55   eigen_worker_threads_.workers = new thread::ThreadPool(
56       Env::Default(), "evaluation_utils", eigen_worker_threads_.num_threads);
57   eigen_threadpool_wrapper_.reset(
58       new EigenThreadPoolWrapper(eigen_worker_threads_.workers));
59   eigen_device_.reset(new Eigen::ThreadPoolDevice(
60       eigen_threadpool_wrapper_.get(), eigen_worker_threads_.num_threads));
61   set_tensorflow_cpu_worker_threads(&eigen_worker_threads_);
62   set_eigen_cpu_device(eigen_device_.get());
63 }
64 
~DeviceSimple()65 DeviceSimple::~DeviceSimple() {
66   eigen_threadpool_wrapper_.reset();
67   eigen_device_.reset();
68   delete eigen_worker_threads_.workers;
69 }
70 
MakeTensorFromProto(const TensorProto & tensor_proto,const AllocatorAttributes alloc_attrs,Tensor * tensor)71 Status DeviceSimple::MakeTensorFromProto(const TensorProto& tensor_proto,
72                                          const AllocatorAttributes alloc_attrs,
73                                          Tensor* tensor) {
74   Tensor parsed(tensor_proto.dtype());
75   if (!parsed.FromProto(cpu_allocator(), tensor_proto)) {
76     return errors::InvalidArgument("Cannot parse tensor from tensor_proto.");
77   }
78   *tensor = parsed;
79   return Status::OK();
80 }
81 
EvaluateNode(const NodeDef & node,const TensorVector & inputs,DeviceBase * cpu_device,ResourceMgr * resource_mgr,TensorVector * output)82 Status EvaluateNode(const NodeDef& node, const TensorVector& inputs,
83                     DeviceBase* cpu_device, ResourceMgr* resource_mgr,
84                     TensorVector* output) {
85   Status status;
86   std::unique_ptr<DeviceBase> device;
87   if (cpu_device == nullptr) {
88     device.reset(new DeviceSimple());
89     cpu_device = device.get();
90   }
91 
92   std::unique_ptr<OpKernel> op_kernel(
93       CreateOpKernel("CPU", cpu_device, cpu_device->GetAllocator({}), node,
94                      TF_GRAPH_DEF_VERSION, &status));
95   TF_RETURN_IF_ERROR(status);
96   OpKernelContext::Params params;
97   params.device = cpu_device;
98   params.frame_iter = FrameAndIter(0, 0);
99   params.inputs = &inputs;
100   params.op_kernel = op_kernel.get();
101   params.resource_manager = resource_mgr;
102 
103   gtl::InlinedVector<AllocatorAttributes, 4> output_attrs;
104   const int num_outputs = op_kernel->num_outputs();
105   for (int i = 0; i < num_outputs; i++) {
106     AllocatorAttributes attr;
107     attr.set_on_host(true);
108     output_attrs.push_back(attr);
109   }
110   params.output_attr_array = output_attrs.data();
111 
112   OpKernelContext op_context(&params);
113   op_kernel->Compute(&op_context);
114   for (int i = 0; i < num_outputs; i++) {
115     output->push_back(op_context.release_output(i));
116   }
117   return op_context.status();
118 }
119 
120 }  // end namespace grappler
121 }  // end namespace tensorflow
122