tensorflow/core/grappler/optimizers/shape_optimizer.cc

/* 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/shape_optimizer.h"

#include "tensorflow/core/framework/tensor.pb.h"
#include "tensorflow/core/framework/tensor_shape.pb.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/grappler/graph_view.h"
#include "tensorflow/core/grappler/grappler_item.h"
#include "tensorflow/core/grappler/op_types.h"
#include "tensorflow/core/grappler/utils.h"
#include "tensorflow/core/grappler/utils/symbolic_shapes.h"
#include "tensorflow/core/lib/core/errors.h"

namespace tensorflow {
namespace grappler {

Status ShapeOptimizer::Optimize(Cluster* cluster, const GrapplerItem& item,
                                GraphDef* optimized_graph) {
  *optimized_graph = item.graph;

  GraphProperties properties(item);
  TF_RETURN_IF_ERROR(properties.InferStatically(false));
  GraphView graph(optimized_graph);

  // The product of all the dimensions in a tensor shape can be expressed more
  // simply as the size of the tensor.
  for (auto& node : *optimized_graph->mutable_node()) {
    if (!IsShape(node)) {
      continue;
    }
    for (GraphView::InputPort fanout :
         graph.GetFanout(GraphView::OutputPort(&node, 0))) {
      if (fanout.node->op() != "Prod") {
        continue;
      }
      if (fanout.node->attr().count("keep_dims") != 0 &&
          fanout.node->attr().at("keep_dims").b()) {
        // Keeping the reduced dimensions won't result in a scalar, so we can't
        // rewrite the whole expression directly as a Size operation.
        continue;
      }
      const GraphView::OutputPort reduce_indices =
          graph.GetRegularFanin(GraphView::InputPort(fanout.node, 1));
      const auto& prop =
          properties.GetOutputProperties(reduce_indices.node->name());
      if (prop.size() < reduce_indices.port_id) {
        continue;
      }
      const TensorShapeProto& reduction_indices_shape =
          prop[reduce_indices.port_id].shape();
      if (NumCoefficients(reduction_indices_shape) == 1) {
        const auto& input_props = properties.GetInputProperties(node.name());
        if (input_props.size() != 1) {
          continue;
        }
        // Rewrite the reduction of the shape dimensions as a Size operation.
        const DataType type = input_props[0].dtype();
        fanout.node->set_op("Size");
        fanout.node->set_input(0, node.input(0));
        fanout.node->set_input(1, AsControlDependency(node));
        fanout.node->mutable_attr()->erase("Tidx");
        fanout.node->mutable_attr()->erase("keep_dims");
        (*fanout.node->mutable_attr())["out_type"] =
            fanout.node->attr().at("T");
        (*fanout.node->mutable_attr())["T"].set_type(type);
      }
    }
  }
  for (auto& node : *optimized_graph->mutable_node()) {
    // Try to convert the ratio of 2 symbolic tensor sizes into a constant. This
    // is possible whenever the symbolic dimensions in the numerator and
    // denominator cancel each other.
    if (node.op() == "Div") {
      const GraphView::OutputPort input1 =
          graph.GetRegularFanin(GraphView::InputPort(&node, 0));
      const GraphView::OutputPort input2 =
          graph.GetRegularFanin(GraphView::InputPort(&node, 1));
      if (!IsSize(*input1.node) || !IsSize(*input2.node)) {
        continue;
      }
      const auto& prop1 = properties.GetInputProperties(input1.node->name());
      const auto& prop2 = properties.GetInputProperties(input2.node->name());
      if (prop1.size() != 1 || prop2.size() != 1) {
        continue;
      }
      const TensorShapeProto& shape1 = prop1[0].shape();
      const TensorShapeProto& shape2 = prop2[0].shape();
      int64 result = ComputeSizeRatio(shape1, shape2);
      if (result >= 0) {
        // Replace div with constant.
        node.set_op("Const");
        DataType dtype = node.attr().at("T").type();
        node.mutable_attr()->erase("T");
        (*node.mutable_attr())["dtype"].set_type(dtype);
        TensorProto* t = (*node.mutable_attr())["value"].mutable_tensor();
        t->set_dtype(dtype);
        *t->mutable_tensor_shape() = TensorShapeProto();
        if (dtype == DT_INT32) {
          t->add_int_val(result);
        } else {
          t->add_int64_val(result);
        }
        node.set_input(0, AsControlDependency(node.input(0)));
        node.set_input(1, AsControlDependency(node.input(1)));
      }
    }
  }
  return Status::OK();
}

void ShapeOptimizer::Feedback(Cluster* /*cluster*/,
                              const GrapplerItem& /*item*/,
                              const GraphDef& /*optimized_graph*/,
                              double /*result*/) {
  // Nothing to do for LoopOptimizer.
}

}  // end namespace grappler
}  // namespace tensorflow