gpesquero/organicmaps
1
Fork 0
forked from organicmaps/organicmaps
Code Pull requests Activity

[search] Use indices vector to sort and select results in PreRanker.

Browse source 
Signed-off-by: Viktor Govako <viktor.govako@gmail.com>
This commit is contained in:
Viktor Govako 2022-07-31 18:21:00 +03:00
parent ecf603e06a
commit 00366cc5d8
5 changed files with 81 additions and 92 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
indexer/feature_decl.cpp
View file

@ -1,12 +1,12 @@
#include "indexer/feature_decl.hpp"
#include <boost/container_hash/hash.hpp>
#include <sstream>
using namespace std;
namespace feature
{
string DebugPrint(GeomType type)
std::string DebugPrint(GeomType type)
{
switch (type)
{
@ -19,7 +19,7 @@ string DebugPrint(GeomType type)
}
} // namespace feature
string DebugPrint(FeatureID const & id)
std::string DebugPrint(FeatureID const & id)
{
return "{ " + DebugPrint(id.m_mwmId) + ", " + std::to_string(id.m_index) + " }";
}
@ -30,7 +30,7 @@ char const * const FeatureID::kInvalidFileName = "INVALID";
int64_t const FeatureID::kInvalidMwmVersion = -1;
string FeatureID::GetMwmName() const
std::string FeatureID::GetMwmName() const
{
return IsValid() ? m_mwmId.GetInfo()->GetCountryName() : kInvalidFileName;
}
@ -39,3 +39,11 @@ int64_t FeatureID::GetMwmVersion() const
{
return IsValid() ? m_mwmId.GetInfo()->GetVersion() : kInvalidMwmVersion;
}
size_t std::hash<FeatureID>::operator()(FeatureID const & fID) const
{
size_t seed = 0;
boost::hash_combine(seed, fID.m_mwmId.GetInfo());
boost::hash_combine(seed, fID.m_index);
return seed;
}

9
indexer/feature_decl.hpp
View file

@ -50,3 +50,12 @@ struct FeatureID
};
std::string DebugPrint(FeatureID const & id);
namespace std
{
template <>
struct hash<FeatureID>
{
size_t operator()(FeatureID const & fID) const;
};
} // namespace std

86
search/pre_ranker.cpp
View file

@ -24,7 +24,7 @@ using namespace std;
namespace
{
void SweepNearbyResults(m2::PointD const & eps, set<FeatureID> const & prevEmit,
void SweepNearbyResults(m2::PointD const & eps, unordered_set<FeatureID> const & prevEmit,
vector<PreRankerResult> & results)
{
m2::NearbyPointsSweeper sweeper(eps.x, eps.y);
@ -143,6 +143,22 @@ void PreRanker::FillMissingFieldsInPreResults()
});
}
namespace
{
template <class CompT, class ContT> class CompareIndices
{
CompT m_cmp;
ContT const & m_cont;
public:
CompareIndices(CompT const & cmp, ContT const & cont) : m_cmp(cmp), m_cont(cont) {}
bool operator()(size_t l, size_t r) const
{
return m_cmp(m_cont[l], m_cont[r]);
}
};
} // namespace
void PreRanker::Filter(bool viewportSearch)
{
auto const lessForUnique = [](PreRankerResult const & lhs, PreRankerResult const & rhs)
@ -167,64 +183,22 @@ void PreRanker::Filter(bool viewportSearch)
if (m_results.size() <= BatchSize())
return;
sort(m_results.begin(), m_results.end(), &PreRankerResult::LessDistance);
vector<size_t> indices(m_results.size());
generate(indices.begin(), indices.end(), [n = 0] () mutable { return n++; });
unordered_set<size_t> filtered;
/// @todo To have any benefit from the next sort-shuffle code block, we should have at least 2 *strictly* equal
/// (distance in meters) results in the middle of m_results vector. The probability of that is -> 0.
/// This code had sence, when we had some approximated viewport distance before centers table.
/*{
// Priority is some kind of distance from the viewport or
// position, therefore if we have a bunch of results with the same
// priority, we have no idea here which results are relevant. To
// prevent bias from previous search routines (like sorting by
// feature id) this code randomly selects tail of the
// sorted-by-priority list of pre-results.
double const last = m_results[BatchSize()].GetDistance();
auto b = m_results.begin() + BatchSize();
for (; b != m_results.begin() && b->GetDistance() == last; --b)
;
if (b->GetDistance() != last)
++b;
auto e = m_results.begin() + BatchSize();
for (; e != m_results.end() && e->GetDistance() == last; ++e)
;
// The main reason of shuffling here is to select a random subset
// from the low-priority results. We're using a linear
// congruential method with default seed because it is fast,
// simple and doesn't need an external entropy source.
//
// TODO (@y, @m, @vng): consider to take some kind of hash from
// features and then select a subset with smallest values of this
// hash. In this case this subset of results will be persistent
// to small changes in the original set.
shuffle(b, e, m_rng);
}*/
struct LessFeatureID
{
inline bool operator()(PreRankerResult const & lhs, PreRankerResult const & rhs) const
{
return lhs.GetId() < rhs.GetId();
}
};
set<PreRankerResult, LessFeatureID> filtered;
auto const numResults = min(m_results.size(), BatchSize());
auto const iBeg = m_results.begin();
auto const iMiddle = iBeg + numResults;
auto const iEnd = m_results.end();
auto const iBeg = indices.begin();
auto const iMiddle = iBeg + BatchSize();
auto const iEnd = indices.end();
nth_element(iBeg, iMiddle, iEnd, CompareIndices(&PreRankerResult::LessDistance, m_results));
filtered.insert(iBeg, iMiddle);
if (!m_params.m_categorialRequest)
{
nth_element(iBeg, iMiddle, iEnd, &PreRankerResult::LessRankAndPopularity);
nth_element(iBeg, iMiddle, iEnd, CompareIndices(&PreRankerResult::LessRankAndPopularity, m_results));
filtered.insert(iBeg, iMiddle);
nth_element(iBeg, iMiddle, iEnd, &PreRankerResult::LessByExactMatch);
nth_element(iBeg, iMiddle, iEnd, CompareIndices(&PreRankerResult::LessByExactMatch, m_results));
filtered.insert(iBeg, iMiddle);
}
else
@ -238,11 +212,15 @@ void PreRanker::Filter(bool viewportSearch)
2 * kPedestrianRadiusMeters;
comparator.m_viewport = m_params.m_viewport;
nth_element(iBeg, iMiddle, iEnd, comparator);
nth_element(iBeg, iMiddle, iEnd, CompareIndices(comparator, m_results));
filtered.insert(iBeg, iMiddle);
}
m_results.assign(make_move_iterator(filtered.begin()), make_move_iterator(filtered.end()));
PreResultsContainerT newResults;
newResults.reserve(filtered.size());
for (size_t idx : filtered)
newResults.push_back(m_results[idx]);
m_results.swap(newResults);
}
void PreRanker::UpdateResults(bool lastUpdate)

15
search/pre_ranker.hpp
View file

@ -12,10 +12,9 @@
#include <algorithm>
#include <limits>
#include <optional>
#include <random>
#include <set>
#include <string>
#include <utility>
#include <unordered_set>
#include <vector>
class DataSource;
@ -144,8 +143,10 @@ private:
DataSource const & m_dataSource;
Ranker & m_ranker;
std::vector<PreRankerResult> m_results;
std::vector<PreRankerResult> m_relaxedResults;
using PreResultsContainerT = std::vector<PreRankerResult>;
PreResultsContainerT m_results, m_relaxedResults;
Params m_params;
// Amount of results sent up the pipeline.
@ -161,12 +162,10 @@ private:
/// search session. They're used for filtering of current search, because we need to give more priority
/// to results that were on map previously, to avoid result's annoying blinking/flickering on map.
/// @{
std::set<FeatureID> m_currEmit;
std::set<FeatureID> m_prevEmit;
std::unordered_set<FeatureID> m_currEmit;
std::unordered_set<FeatureID> m_prevEmit;
/// @}
std::minstd_rand m_rng;
DISALLOW_COPY_AND_MOVE(PreRanker);
};
} // namespace search

45
search/search_integration_tests/pre_ranker_test.cpp
View file

@ -27,31 +27,19 @@
#include "base/assert.hpp"
#include "base/cancellable.hpp"
#include "base/limited_priority_queue.hpp"
#include "base/math.hpp"
#include "base/stl_helpers.hpp"
#include <algorithm>
#include <cstddef>
#include <cstdint>
#include <iterator>
#include <vector>
namespace pre_ranker_test
{
using namespace generator::tests_support;
using namespace search::tests_support;
using namespace search;
using namespace std;
class DataSource;
namespace storage
{
class CountryInfoGetter;
}
namespace search
{
namespace
{
class TestRanker : public Ranker
{
public:
@ -105,12 +93,15 @@ UNIT_CLASS_TEST(PreRankerTest, Smoke)
// emits results nearest to the pivot.
m2::PointD const kPivot(0, 0);
m2::RectD const kViewport(m2::PointD(-5, -5), m2::PointD(5, 5));
m2::RectD const kViewport(-5, -5, 5, 5);
size_t const kBatchSize = 50;
/// @todo Well, I'm not sure that 50 results will have unique distances to pivot.
/// 7x7 grid is 49, so potentially it can be 51 (north and south) or (east and west).
/// But we should consider circle (ellipse) around pivot and I can't say,
/// how it goes in meters radius on integer mercator grid.
size_t constexpr kBatchSize = 50;
vector<TestPOI> pois;
for (int x = -5; x <= 5; ++x)
{
for (int y = -5; y <= 5; ++y)
@ -122,7 +113,8 @@ UNIT_CLASS_TEST(PreRankerTest, Smoke)
TEST_LESS(kBatchSize, pois.size(), ());
auto mwmId = BuildCountry("Cafeland", [&](TestMwmBuilder & builder) {
auto mwmId = BuildCountry("Cafeland", [&](TestMwmBuilder & builder)
{
for (auto const & poi : pois)
builder.Add(poi);
});
@ -150,7 +142,8 @@ UNIT_CLASS_TEST(PreRankerTest, Smoke)
vector<bool> emit(pois.size());
FeaturesVectorTest fv(mwmId.GetInfo()->GetLocalFile().GetPath(MapFileType::Map));
fv.GetVector().ForEach([&](FeatureType & ft, uint32_t index) {
fv.GetVector().ForEach([&](FeatureType & ft, uint32_t index)
{
FeatureID id(mwmId, index);
ResultTracer::Provenance provenance;
preRanker.Emplace(id, PreRankingInfo(Model::TYPE_SUBPOI, TokenRange(0, 1)), provenance);
@ -164,19 +157,21 @@ UNIT_CLASS_TEST(PreRankerTest, Smoke)
TEST(all_of(emit.begin(), emit.end(), base::IdFunctor()), (emit));
TEST(ranker.Finished(), ());
TEST_EQUAL(results.size(), kBatchSize, ());
size_t const count = results.size();
// See todo comment above for details.
TEST(count == kBatchSize || count == kBatchSize + 1, (count));
vector<bool> checked(pois.size());
for (size_t i = 0; i < results.size(); ++i)
for (size_t i = 0; i < count; ++i)
{
size_t const index = results[i].GetId().m_index;
TEST_LESS(index, pois.size(), ());
TEST(!checked[index], (index));
TEST(base::AlmostEqualAbs(distances[index], results[i].GetDistance(), 1.0),
TEST(base::AlmostEqualAbs(distances[index], results[i].GetDistance(), 1.0 /* 1 meter epsilon */),
(distances[index], results[i].GetDistance()));
checked[index] = true;
}
}
} // namespace
} // namespace search
} // namespace pre_ranker_test