LookaheadContractor.h

Go to the documentation of this file.

 
#pragma once
 
#ifndef __LOOKAHEAD_CONTRACTOR_H_
#define __LOOKAHEAD_CONTRACTOR_H_ 1
 
#include "BaseContractor.h"
 
#include <boost/container_hash/hash.hpp>
 
namespace TensorNetworks {

class LookaheadContractor : public BaseContractor {
public:
  using TensorPair = std::pair<Eigen::Index, Eigen::Index>;
  using VisitedPairType =
      std::unordered_set<TensorPair, boost::hash<TensorPair>>;
  using DummyTensorsMap = std::map<Eigen::Index, std::shared_ptr<TensorNode>>;

  double Contract(const TensorNetwork &network, Types::qubit_t qubit) override {
    std::vector<Eigen::Index> keys;
    std::unordered_map<Eigen::Index, Eigen::Index> keysKeys;
 
    TensorsMap tensors =
        InitializeTensors(network, qubit, keys, keysKeys, false);
    DummyTensorsMap dummyTensors = CloneTensorsToDummy(tensors);
 
    // while there is more than one tensor...
    while (tensors.size() > 1) {
      // just to have them set to something
      auto it = tensors.begin();
      Eigen::Index tensor1Id = it->first;
      ++it;
      Eigen::Index tensor2Id = it->first;
 
      const auto saveMaxTensorRank = maxTensorRank;
 
      Eigen::Index maxCost = std::numeric_limits<Eigen::Index>::max();
      size_t bestDepth = numberOfLevels;
      Lookahead(qubit, tensor1Id, tensor2Id, bestDepth, 0, 0, dummyTensors, 1,
                maxCost, 0);
 
      const auto &tensor = tensors[tensor1Id];
      const auto resultRank = GetResultRank(tensor, tensors[tensor2Id]);
 
      // also contract the dummies for the next step
      ContractNodes(qubit, dummyTensors, tensor1Id, tensor2Id, resultRank);
 
      maxTensorRank = saveMaxTensorRank;
 
      ContractNodes(qubit, tensors, tensor1Id, tensor2Id, resultRank);
 
      if (resultRank == 0) {
        if (tensors.size() == 1 || tensor->contractsTheNeededQubit)
          return std::real(tensor->tensor->atOffset(0));
        // erasing this tensor happens because (not the case anymore, it's
        // avoided) the tensor network might be a disjoint one and a subnetwork
        // is contracted that does not contain the needed qubit
        tensors.erase(tensor1Id);
        dummyTensors.erase(tensor1Id);
      }
    }
 
    return std::real(tensors.begin()->second->tensor->atOffset(0));
  }
 
  size_t GetNumberOfLevels() const { return numberOfLevels; }
 
  void SetNumberOfLevels(size_t levels) { numberOfLevels = levels; }
 
  bool GetUseMaxRankCost() const { return useMaxRankCost; }
 
  void SetUseMaxRankCost(bool val = true) { useMaxRankCost = val; }

  std::shared_ptr<ITensorContractor> Clone() const override {
    auto cloned = std::make_shared<LookaheadContractor>();
 
    cloned->maxTensorRank = maxTensorRank;
    cloned->enableMultithreading = enableMultithreading;
    cloned->numberOfLevels = numberOfLevels;
    cloned->useMaxRankCost = useMaxRankCost;
 
    return cloned;
  }
 
private:
  void Lookahead(Types::qubit_t qubit, Eigen::Index &tensor1Id,
                 Eigen::Index &tensor2Id, size_t &bestDepth, Eigen::Index curt1,
                 Eigen::Index curt2, DummyTensorsMap &dummyTensors,
                 size_t curLevel, Eigen::Index &maxCost,
                 Eigen::Index CostOnPath) {
    if (dummyTensors.size() > 1 && curLevel < numberOfLevels) {
      std::unordered_set<TensorPair, boost::hash<TensorPair>> visitedPairs;
 
      for (auto tensorIt = dummyTensors.begin(); tensorIt != dummyTensors.end();
           ++tensorIt) {
        auto tensor = tensorIt->second;
        const auto curTensorId = tensor->GetId();
 
        assert(tensor->tensor->IsDummy());
 
        // TODO: should I really try outer products here as well?
        for (Eigen::Index ti = 0;
             ti < static_cast<Eigen::Index>(tensor->connections.size()); ++ti) {
          const auto nextTensorId = tensor->connections[ti];
 
          if (nextTensorId != TensorNode::NotConnected) {
            auto t1 = curTensorId;
            auto t2 = nextTensorId;
            if (t1 > t2)
              std::swap(t1, t2);
            if (visitedPairs.find(std::make_pair(t1, t2)) != visitedPairs.end())
              continue;
 
            visitedPairs.insert(std::make_pair(t1, t2));
 
            const Eigen::Index newRank =
                GetResultRank(tensor, dummyTensors[nextTensorId]);
            // const Eigen::Index Cost = newRank - (tensor->GetRank() +
            // dummyTensors[nextTensorId]->GetRank()) / 2; const Eigen::Index
            // newCostOnPath = CostOnPath + Cost;
            const Eigen::Index newCostOnPath =
                useMaxRankCost ? std::max(newRank, CostOnPath)
                               : CostOnPath + newRank;
            if (newCostOnPath < maxCost) {
              // 1. if the curLevel is not the last level
              // contract the dummy tensors
              // then recursively call Lookahead (with CostOnPath = max(newRank,
              // CostOnPath)
 
              // TODO: improve this, this is slow - actually save only the
              // contracted tensors and restore them (and the linked ones as
              // well) after the recursive call
              // auto newTensors = CloneTensorsToDummy(dummyTensors);
 
              auto saveTensor2 = dummyTensors[nextTensorId];
 
              ContractNodes(qubit, dummyTensors, curTensorId, nextTensorId,
                            newRank);
 
              Lookahead(qubit, tensor1Id, tensor2Id, bestDepth,
                        curLevel == 1 ? curTensorId : curt1,
                        curLevel == 1 ? nextTensorId : curt2, dummyTensors,
                        curLevel + 1, maxCost, newCostOnPath);
 
              // then restore the contraction for the next iteration
 
              dummyTensors[curTensorId] = tensor;
              dummyTensors[nextTensorId] = saveTensor2;
 
              // need to restore the connections back from the other connected
              // tensors, too, because the contracting affected them
 
              for (Eigen::Index i = 0;
                   i < static_cast<Eigen::Index>(tensor->connections.size());
                   ++i) {
                if (tensor->connections[i] != TensorNode::NotConnected &&
                    tensor->connections[i] != nextTensorId) {
                  const auto connectedTensorId = tensor->connections[i];
                  const auto otherTensorIndex = tensor->connectionsIndices[i];
                  // dummyTensors[otherTensorId]->connections[otherTensorIndex]
                  // = curTensorId;
                  dummyTensors[connectedTensorId]
                      ->connectionsIndices[otherTensorIndex] = i;
                }
              }
 
              for (Eigen::Index i = 0; i < static_cast<Eigen::Index>(
                                               saveTensor2->connections.size());
                   ++i) {
                if (saveTensor2->connections[i] != TensorNode::NotConnected &&
                    saveTensor2->connections[i] != curTensorId) {
                  const auto connectedTensorId = saveTensor2->connections[i];
                  const auto otherTensorIndex =
                      saveTensor2->connectionsIndices[i];
                  dummyTensors[connectedTensorId]
                      ->connections[otherTensorIndex] = nextTensorId;
                  dummyTensors[connectedTensorId]
                      ->connectionsIndices[otherTensorIndex] = i;
                }
              }
 
              // otherwise ugly things will happen
              // tensorIt =
              // std::make_reverse_iterator(dummyTensors.find(curTensorId));
              tensorIt = dummyTensors.find(curTensorId);
            }
          }
        }
      }
    }
 
    if ((curLevel == numberOfLevels && CostOnPath < maxCost) ||
        (dummyTensors.size() <= 1 &&
         (CostOnPath < maxCost ||
          (maxCost == CostOnPath && curLevel < bestDepth)))) {
      maxCost = CostOnPath;
      bestDepth = curLevel;
      // set the tensor1Id and tensor2Id to the best ones (so far)
      tensor1Id = curt1;
      tensor2Id = curt2;
    }
  }
 
  static DummyTensorsMap CloneTensorsToDummy(const TensorsMap &tensors) {
    DummyTensorsMap clonedTensors;
    for (const auto &[tensorId, tensor] : tensors)
      clonedTensors[tensorId] = tensor->CloneWithADummyTensor();
 
    return clonedTensors;
  }
 
  size_t numberOfLevels = 3;
  bool useMaxRankCost = true;
};
 
} // namespace TensorNetworks
 
#endif // __LOOKAHEAD_CONTRACTOR_H_