[tf.data] Meta optimizer to manage tf.data optimization passes to perform optimizations in a meaningful order (instead of an arbitrary order as before).

PiperOrigin-RevId: 226944242

tensorflow/core/grappler/optimizers/data/BUILD

@@ -3,6 +3,30 @@ licenses(["notice"])  # Apache 2.0
 load("//tensorflow:tensorflow.bzl", "tf_cc_test")
 load("//tensorflow/core:platform/default/build_config.bzl", "tf_protos_all")
 
+cc_library(
+    name = "meta_optimizer",
+    srcs = ["meta_optimizer.cc"],
+    hdrs = [
+        "meta_optimizer.h",
+    ],
+    visibility = ["//visibility:public"],
+    deps = [
+        "@com_google_absl//absl/container:flat_hash_map",
+        "//tensorflow/core/grappler:grappler_item",
+        "//tensorflow/core:lib",
+        "//tensorflow/core/grappler/clusters:cluster",
+        "//tensorflow/core/grappler/optimizers:arithmetic_optimizer",
+        "//tensorflow/core/grappler/optimizers:model_pruner",
+        "//tensorflow/core/grappler/optimizers:shape_optimizer",
+        "//tensorflow/core/grappler/optimizers:dependency_optimizer",
+        "//tensorflow/core/grappler/optimizers:function_optimizer",
+        "//tensorflow/core/grappler/optimizers:custom_graph_optimizer",
+        "//tensorflow/core/grappler/optimizers:custom_graph_optimizer_registry",
+        "//tensorflow/core:lib_internal",
+        "//tensorflow/core:ptr_util",
+    ] + tf_protos_all(),
+)
+
 cc_library(
     name = "filter_fusion",
     srcs = ["filter_fusion.cc"],
@@ -561,6 +585,7 @@ cc_library(
         ":map_fusion",
         ":map_parallelization",
         ":map_vectorization",
+        ":meta_optimizer",
         ":noop_elimination",
         ":shuffle_and_repeat_fusion",
     ],

tensorflow/core/grappler/optimizers/data/meta_optimizer.cc

@@ -0,0 +1,125 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/core/grappler/optimizers/data/meta_optimizer.h"
+
+#include "tensorflow/core/grappler/clusters/cluster.h"
+#include "tensorflow/core/grappler/grappler_item.h"
+#include "tensorflow/core/grappler/optimizers/arithmetic_optimizer.h"
+#include "tensorflow/core/grappler/optimizers/custom_graph_optimizer_registry.h"
+#include "tensorflow/core/grappler/optimizers/dependency_optimizer.h"
+#include "tensorflow/core/grappler/optimizers/function_optimizer.h"
+#include "tensorflow/core/grappler/optimizers/model_pruner.h"
+#include "tensorflow/core/grappler/optimizers/shape_optimizer.h"
+#include "tensorflow/core/lib/gtl/map_util.h"
+#include "tensorflow/core/util/ptr_util.h"
+
+namespace tensorflow {
+namespace grappler {
+
+Status TFDataMetaOptimizer::Optimize(Cluster* cluster, const GrapplerItem& item,
+                                     GraphDef* output) {
+  // Stores the optimized item so far.
+  GrapplerItem optimized_item = item;
+
+  // Perform optimizations in a meaningful order.
+  for (const auto& optimization :
+       {"noop_elimination",
+        "shuffle_and_repeat_fusion",
+        "map_fusion",
+        "filter_fusion",
+        "map_and_filter_fusion",
+        "hoist_random_uniform",
+        "map_parallelization",
+        "map_and_batch_fusion",
+        "map_vectorization",
+        "make_numa_aware",
+        "latency_all_edges",
+        "make_sloppy",
+        "pruning",
+        "function",
+        "shape",
+        "arithmetic",
+        "dependency"}) {
+    TF_RETURN_IF_ERROR(
+        ApplyOptimization(optimization, cluster, &optimized_item));
+  }
+
+  // Store the final result of all the optimizations in `output`.
+  output->Swap(&optimized_item.graph);
+  return Status::OK();
+}
+
+Status TFDataMetaOptimizer::ApplyOptimization(const string& name,
+                                              Cluster* cluster,
+                                              GrapplerItem* item) const {
+  GRAPPLER_RETURN_IF_DEADLINE_EXCEEDED();
+
+  const auto* optimizer = gtl::FindOrNull(enabled_optimizers_, name);
+  if (!optimizer) {
+    return Status::OK();
+  }
+
+  GraphDef result;
+  (*optimizer)->set_deadline_usec(this->deadline_usec());
+  TF_RETURN_IF_ERROR((*optimizer)->Optimize(cluster, *item, &result));
+  item->graph.Swap(&result);
+
+  return Status::OK();
+}
+
+Status TFDataMetaOptimizer::Init(
+    const tensorflow::RewriterConfig_CustomGraphOptimizer* config) {
+  if (!config) return Status::OK();
+
+  // Initialize custom tf.data optimizers based on config.
+  auto& optimizers = config->parameter_map().at("optimizers").list().s();
+  for (const auto& optimizer_name : optimizers) {
+    auto optimizer =
+        CustomGraphOptimizerRegistry::CreateByNameOrNull(optimizer_name);
+    if (optimizer) {
+      // None of our data optimizers implement a meaningful Init function.
+      // This returns an error in case any of them does.
+      TF_RETURN_IF_ERROR(optimizer->Init());
+      enabled_optimizers_[optimizer_name] = std::move(optimizer);
+    } else {
+      // This should never happen.
+      return errors::Internal(
+          "Tried to register a dataset optimizer that doesn't exist: ",
+          optimizer_name);
+    }
+  }
+
+  // Initialize standard grappler optimizers.
+  enabled_optimizers_["pruning"] = MakeUnique<ModelPruner>();
+  enabled_optimizers_["function"] =
+      MakeUnique<FunctionOptimizer>(RewriterConfig::ON);
+  enabled_optimizers_["shape"] = MakeUnique<ShapeOptimizer>();
+  enabled_optimizers_["arithmetic"] = MakeUnique<ArithmeticOptimizer>();
+  enabled_optimizers_["dependency"] = MakeUnique<DependencyOptimizer>();
+
+  return Status::OK();
+}
+
+void TFDataMetaOptimizer::Feedback(Cluster* cluster, const GrapplerItem& item,
+                                   const GraphDef& optimize_output,
+                                   double result) {
+  // no-op
+}
+
+REGISTER_GRAPH_OPTIMIZER_AS(TFDataMetaOptimizer, "tf_data_meta_optimizer");
+
+}  // end namespace grappler
+}  // end namespace tensorflow

tensorflow/core/grappler/optimizers/data/meta_optimizer.h

@@ -0,0 +1,56 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_DATA_META_OPTIMIZER_H_
+#define TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_DATA_META_OPTIMIZER_H_
+
+#include "absl/container/flat_hash_map.h"
+#include "tensorflow/core/grappler/optimizers/custom_graph_optimizer.h"
+
+namespace tensorflow {
+namespace grappler {
+
+// This optimizer performs tf.data-specific optimizations by invoking
+// other optimizers.
+class TFDataMetaOptimizer : public CustomGraphOptimizer {
+ public:
+  TFDataMetaOptimizer() = default;
+  ~TFDataMetaOptimizer() override = default;
+
+  string name() const override { return "tf_data_meta_optimizer"; };
+
+  Status Init(
+      const tensorflow::RewriterConfig_CustomGraphOptimizer* config) override;
+
+  Status Optimize(Cluster* cluster, const GrapplerItem& item,
+                  GraphDef* output) override;
+
+  void Feedback(Cluster* cluster, const GrapplerItem& item,
+                const GraphDef& optimize_output, double result) override;
+
+ private:
+  absl::flat_hash_map<string, std::unique_ptr<GraphOptimizer>>
+      enabled_optimizers_;
+
+  // Applies an optimization with the specified name on `item`, and stores
+  // the result in `item.graph`
+  Status ApplyOptimization(const string& name, Cluster* cluster,
+                           GrapplerItem* item) const;
+};
+
+}  // end namespace grappler
+}  // end namespace tensorflow
+
+#endif  // TENSORFLOW_CORE_GRAPPLER_OPTIMIZERS_DATA_META_OPTIMIZER_H_

tensorflow/core/kernels/data/optimize_dataset_op.cc

@@ -40,6 +40,8 @@ namespace tensorflow {
 namespace data {
 namespace {
 
+static const char* const kOptimizerName = "tf_data_meta_optimizer";
+
 // See documentation in ../../ops/dataset_ops.cc for a high-level
 // description of the following op.
 class OptimizeDatasetOp : public UnaryDatasetOpKernel {
@@ -286,31 +288,6 @@ class OptimizeDatasetOp : public UnaryDatasetOpKernel {
       (*meta_graph_def.mutable_collection_def())["train_op"] = collection_def;
 
       // Create Grappler item.
-      tensorflow::ConfigProto config;
-      RewriterConfig& rewriter_config =
-          *config.mutable_graph_options()->mutable_rewrite_options();
-      for (const string& optimization : optimizations_) {
-        rewriter_config.add_optimizers(optimization);
-      }
-      // If no optimizations were specified, supply a non-existent
-      // optimization to prevent Grappler from applying the default set of
-      // optimizations as some of them do not work out of the box at the
-      // moment (e.g. because we have no cost model for dataset ops).
-      if (optimizations_.empty()) {
-        rewriter_config.add_optimizers("non-existent");
-      } else {
-        // If we apply custom dataset optimizers, explicitly trigger a subset of
-        // standard grappler optimizations to further optimize modified dataset
-        // graphs (e.g. performing constant folding on merged functions,
-        // removing unused graph nodes)
-        // TODO(b/118175421): This should be part of the tf.data optimization
-        // pass manager.
-        // TODO(b/120437209): Apply `constfold` optimization when it is fixed.
-        for (const auto& optimizer :
-             {"pruning", "function", "shape", "arithmetic", "dependency"}) {
-          rewriter_config.add_optimizers(optimizer);
-        }
-      }
       tensorflow::grappler::ItemConfig item_config;
       item_config.apply_optimizations = true;
       std::unique_ptr<tensorflow::grappler::GrapplerItem> grappler_item =
@@ -319,13 +296,21 @@ class OptimizeDatasetOp : public UnaryDatasetOpKernel {
       std::unordered_map<string, tensorflow::DeviceProperties> device_map;
       tensorflow::grappler::VirtualCluster cluster(device_map);
 
-      // Run optimizer.
+      // Run data optimizer using grappler's meta optimizer.
-      if (VLOG_IS_ON(2)) {
+      tensorflow::ConfigProto config;
-        LOG(INFO) << "Performing the following optimizations:";
+      RewriterConfig& rewriter_config =
-        for (const string& optimization : optimizations_) {
+          *config.mutable_graph_options()->mutable_rewrite_options();
-          LOG(INFO) << "  " << optimization;
+      rewriter_config.add_optimizers(kOptimizerName);
-        }
+
+      auto custom_optimizer = rewriter_config.add_custom_optimizers();
+      custom_optimizer->set_name(kOptimizerName);
+      auto* custom_optimizations_list =
+          (*custom_optimizer->mutable_parameter_map())["optimizers"]
+              .mutable_list();
+      for (const auto& opt : optimizations_) {
+        custom_optimizations_list->add_s(opt);
       }
+
       TF_RETURN_IF_ERROR(tensorflow::grappler::RunMetaOptimizer(
           *grappler_item, config, ctx->device(), &cluster, graph_def));