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 | #ifndef TENSORFLOW_CORE_COMMON_RUNTIME_SINGLE_THREADED_EXECUTOR_H_ |
17 | #define TENSORFLOW_CORE_COMMON_RUNTIME_SINGLE_THREADED_EXECUTOR_H_ |
18 | |
19 | #include "tensorflow/core/common_runtime/executor.h" |
20 | |
21 | namespace tensorflow { |
22 | |
23 | // Creates a new `Executor` for executing `graph` synchronously on the caller |
24 | // thread. |
25 | // |
26 | // NOTE(mrry): The returned executor is optimized to impose low overhead on |
27 | // graphs that perform a small amount of work (e.g. <15us of work per graph on |
28 | // present architectures). It eschews concurrency, because issuing work to |
29 | // multiple threads can dominate the cost of executing small ops synchronously, |
30 | // and because contention in the executor data structures can reduce throughput |
31 | // (in terms of ops executed per unit time). |
32 | // |
33 | // However, the current implementation has the following limitations: |
34 | // |
35 | // 1. Reference-typed tensors are not supported and will not be supported in |
36 | // future. |
37 | // 2. Graphs with control flow (containing "Switch" and "Merge" nodes) are not |
38 | // currently supported. The current plan is to extend support to "functional" |
39 | // control flow after the TensorFlow APIs transition to building graphs in |
40 | // that form (e.g. `tf.cond_v2()`). |
41 | // 3. Partitioned graphs (containing "_Recv" nodes) are not currently supported. |
42 | // The present implementation executes kernels one at a time in topological |
43 | // order, and cannot currently distinguish between disconnected subgraphs |
44 | // that are logically connected by subgraphs on a different device. |
45 | // 4. Memory logging is not currently supported. |
46 | // 5. Allocation forwarding is not currently supported. |
47 | // 6. Non-default device contexts are not currently supported. In effect, this |
48 | // limits the executor to CPU devices. |
49 | // 7. Ops that rely on `OpKernelContext::slice_reader_cache()` being non-null |
50 | // are not currently supported. |
51 | // |
52 | // The single-threaded executor is primarily suitable for executing simple |
53 | // TensorFlow functions, such as one might find in a `tf.data` pipeline. |
54 | Status NewSingleThreadedExecutor(const LocalExecutorParams& params, |
55 | const Graph& graph, Executor** executor); |
56 | |
57 | // Returns OkStatus() for ops which are compatible with synchronous execution, |
58 | // and otherwise returns an error message appropriate for propagation if needed. |
59 | // If `allow_control_flow_sync_execution` is set to `true` control |
60 | // nodes are marked as safe for execution on the SingleThreadedExecutor. |
61 | Status ValidateOpIsSafeForSyncExecution(const Node& n, |
62 | bool allow_control_flow_sync_execution); |
63 | |
64 | } // namespace tensorflow |
65 | |
66 | #endif // TENSORFLOW_CORE_COMMON_RUNTIME_SINGLE_THREADED_EXECUTOR_H_ |
67 | |