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_GRAPPLER_OPTIMIZERS_DATA_MAKE_DETERMINISTIC_H_ 17 #define TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_DATA_MAKE_DETERMINISTIC_H_ 18 19 #include "tensorflow/core/grappler/optimizers/data/optimizer_base.h" 20 21 namespace tensorflow { 22 namespace grappler { 23 24 // Removes sources on nondeterminism from dataset ops. Nondeterminism can occur 25 // in the follow ways, each which this pass addresses: 26 // 27 // 1. The datasets ParallelInterleave, ParallelMap, and MapAndBatch can 28 // introduce nondeterminism by running a function multiple times in parallel. 29 // Specifically, if the function can mutate state, it is potentially 30 // nondeterministic. In such cases, this pass converts such dataset ops to a 31 // non-parallel version. As a performance optimization, in certain cases this 32 // pass will instead move nondeterministic ops to a separate non-parallel Map 33 // op, so that most of the ops can still run in parallel. 34 // 35 // 2. Certain datasets, such as Prefetch, can introduce asynchrony by running a 36 // dataset iterator in a background thread while ops outside the dataset are 37 // also running. This can introduce nondeterminism if the input pipeline has 38 // certain stateful ops. Other than Prefetch, datasets with a 39 // `num_parallel_calls` argument also introduce asynchrony, which includes 40 // the parallel datasets mentioned in (1) above. 41 // 42 // This pass modifies nodes to remove asynchrony when there are any datasets 43 // in the graph with problematic stateful ops. This is done by converting 44 // parallel ops into non-parallel versions, as in (1), and by removing 45 // Prefetch nodes. Unlike (1), legacy random ops such as RandomUniform are 46 // not problematic despite being stateful, as if the op is within a dataset's 47 // function, ops outside the dataset cannot access the state. Also unlike 48 // (1), nondeterministic ops are never moved to a separate Map op, since 49 // doing so would not remove asynchrony. 50 // 51 // 3. Nondeterminism occurs if an op has a "deterministic" attribute that is 52 // false or a "sloppy" attribute that is true. This pass changes such 53 // attributes to be deterministic. 54 class MakeDeterministic : public TFDataOptimizerBase { 55 public: 56 MakeDeterministic() = default; 57 ~MakeDeterministic() override = default; 58 name()59 string name() const override { return "make_deterministic"; }; 60 UsesFunctionLibrary()61 bool UsesFunctionLibrary() const override { return false; } 62 Init(const tensorflow::RewriterConfig_CustomGraphOptimizer * config)63 Status Init( 64 const tensorflow::RewriterConfig_CustomGraphOptimizer* config) override { 65 return OkStatus(); 66 } 67 68 Status OptimizeAndCollectStats(Cluster* cluster, const GrapplerItem& item, 69 GraphDef* output, 70 OptimizationStats* stats) override; 71 }; 72 73 } // namespace grappler 74 } // namespace tensorflow 75 76 #endif // TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_DATA_MAKE_DETERMINISTIC_H_ 77