1 | /* Copyright 2021 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_CONTROL_FLOW_DEPS_TO_CHAINS_H_ |
17 | #define TENSORFLOW_CORE_COMMON_RUNTIME_CONTROL_FLOW_DEPS_TO_CHAINS_H_ |
18 | |
19 | #include "tensorflow/core/common_runtime/optimization_registry.h" |
20 | |
21 | namespace tensorflow { |
22 | |
23 | // Move control flow dependencies in functional control flow to chains. |
24 | // Chains are extra loop variables that serve as tokens for wiring control |
25 | // dependencies across loop iterations at a finer granularity, compared to just |
26 | // a single barrier at the end of each iteration. This enables the |
27 | // parallel_iterations feature for tf.while_loop. |
28 | // |
29 | // One separate chain is added for each of the body function's `control_ret`. |
30 | // |
31 | // For example: |
32 | // |
33 | // while i > 0: |
34 | // r = v.read_value() |
35 | // s += expensive_operation(r) |
36 | // assign = v.assign_add(1) # control: r |
37 | // i += 1 |
38 | // |
39 | // The loop above can safely compute `r` and `assign` ahead of `s`, by the |
40 | // as-if rule. The separate switch/merge nodes that the loop lowers into support |
41 | // that. |
42 | // This transformation enables that to happen by rewriting the loop as follows: |
43 | // |
44 | // chain = 0.0 |
45 | // while i > 0: |
46 | // r = v.read_value() # control: chain |
47 | // s += expensive_operation(r) |
48 | // assign = v.assign_add(1) # control: r |
49 | // i += 1 |
50 | // chain = identity(chain) # control: assign |
51 | // |
52 | // This only rewires dependencies which need to cross scope boundaries, as the |
53 | // switch/merge lowering process has no other way of dealing correctly with |
54 | // those. |
55 | // |
56 | // This pass is best-effort and conservative, requiring attributes set by |
57 | // tf.while_loop and automatic_control_dependencies. When the required |
58 | // attributes are missing for a particular While node, no change is made to |
59 | // that node. Other While nodes are still processed if they do have the needed |
60 | // annotations. |
61 | // The pass can also be toggled by omitting the `_stateful_parallelism=True` |
62 | // attribute on the While node. |
63 | // When the pass returns with error, the graph is left in an invalid state. |
64 | // If successful, this pass also clears the body function's control_ret, |
65 | // which in effect removes the hard barrier that gates each loop iteration. |
66 | // |
67 | // |
68 | // TODO(mdan): Can we define that more formally? |
69 | class ControlFlowDepsToChainsPass : public GraphOptimizationPass { |
70 | public: |
71 | Status Run(const GraphOptimizationPassOptions& options) override; |
72 | }; |
73 | |
74 | } // namespace tensorflow |
75 | |
76 | #endif // TENSORFLOW_CORE_COMMON_RUNTIME_CONTROL_FLOW_DEPS_TO_CHAINS_H_ |
77 | |