ClickHouse/src/AggregateFunctions/AggregateFunctionUniq.h

296 lines
8.2 KiB
C++
Raw Normal View History

2011-09-26 04:00:46 +00:00
#pragma once
#include <city.h>
#include <type_traits>
2011-09-26 04:00:46 +00:00
#include <ext/bit_cast.h>
#include <IO/WriteHelpers.h>
#include <IO/ReadHelpers.h>
2011-09-26 04:00:46 +00:00
#include <DataTypes/DataTypesNumber.h>
#include <DataTypes/DataTypeTuple.h>
2011-09-26 04:00:46 +00:00
#include <Interpreters/AggregationCommon.h>
2019-02-10 17:40:52 +00:00
#include <Common/HashTable/Hash.h>
#include <Common/HashTable/HashSet.h>
#include <Common/HyperLogLogWithSmallSetOptimization.h>
#include <Common/CombinedCardinalityEstimator.h>
2017-07-13 20:58:19 +00:00
#include <Common/typeid_cast.h>
#include <Common/assert_cast.h>
2019-02-10 17:40:52 +00:00
#include <AggregateFunctions/UniquesHashSet.h>
#include <AggregateFunctions/IAggregateFunction.h>
#include <AggregateFunctions/UniqVariadicHash.h>
2011-09-26 04:00:46 +00:00
namespace DB
{
/// uniq
2011-09-26 04:00:46 +00:00
struct AggregateFunctionUniqUniquesHashSetData
{
using Set = UniquesHashSet<DefaultHash<UInt64>>;
Set set;
static String getName() { return "uniq"; }
};
2017-03-09 00:56:38 +00:00
/// For a function that takes multiple arguments. Such a function pre-hashes them in advance, so TrivialHash is used here.
struct AggregateFunctionUniqUniquesHashSetDataForVariadic
{
using Set = UniquesHashSet<TrivialHash>;
Set set;
static String getName() { return "uniq"; }
};
/// uniqHLL12
template <typename T>
struct AggregateFunctionUniqHLL12Data
{
using Set = HyperLogLogWithSmallSetOptimization<T, 16, 12>;
Set set;
static String getName() { return "uniqHLL12"; }
};
template <>
struct AggregateFunctionUniqHLL12Data<String>
{
using Set = HyperLogLogWithSmallSetOptimization<UInt64, 16, 12>;
Set set;
static String getName() { return "uniqHLL12"; }
};
template <>
struct AggregateFunctionUniqHLL12Data<UInt128>
{
using Set = HyperLogLogWithSmallSetOptimization<UInt64, 16, 12>;
Set set;
static String getName() { return "uniqHLL12"; }
};
2015-10-29 04:13:18 +00:00
struct AggregateFunctionUniqHLL12DataForVariadic
{
using Set = HyperLogLogWithSmallSetOptimization<UInt64, 16, 12, TrivialHash>;
Set set;
2015-10-29 04:13:18 +00:00
static String getName() { return "uniqHLL12"; }
2015-10-29 04:13:18 +00:00
};
/// uniqExact
template <typename T>
struct AggregateFunctionUniqExactData
{
using Key = T;
/// When creating, the hash table must be small.
2017-05-11 22:21:06 +00:00
using Set = HashSet<
Key,
HashCRC32<Key>,
HashTableGrower<4>,
2017-05-11 22:21:06 +00:00
HashTableAllocatorWithStackMemory<sizeof(Key) * (1 << 4)>>;
Set set;
static String getName() { return "uniqExact"; }
};
/// For rows, we put the SipHash values (128 bits) into the hash table.
template <>
struct AggregateFunctionUniqExactData<String>
{
using Key = UInt128;
/// When creating, the hash table must be small.
2017-05-11 22:21:06 +00:00
using Set = HashSet<
Key,
UInt128TrivialHash,
HashTableGrower<3>,
2017-05-11 22:21:06 +00:00
HashTableAllocatorWithStackMemory<sizeof(Key) * (1 << 3)>>;
Set set;
static String getName() { return "uniqExact"; }
};
2018-03-17 18:02:19 +00:00
2015-08-31 13:52:17 +00:00
namespace detail
{
2017-03-09 00:56:38 +00:00
/** Hash function for uniq.
*/
template <typename T> struct AggregateFunctionUniqTraits
{
static UInt64 hash(T x)
{
if constexpr (std::is_same_v<T, UInt128>)
{
return sipHash64(x);
}
else if constexpr (std::is_same_v<T, Float32> || std::is_same_v<T, Float64>)
{
return ext::bit_cast<UInt64>(x);
}
else if constexpr (sizeof(T) <= sizeof(UInt64))
return x;
else
return DefaultHash64<T>(x);
}
};
2017-12-02 03:42:21 +00:00
2017-03-09 00:56:38 +00:00
/** The structure for the delegation work to add one element to the `uniq` aggregate functions.
* Used for partial specialization to add strings.
2015-08-31 13:52:17 +00:00
*/
template <typename T, typename Data>
2017-12-02 03:42:21 +00:00
struct OneAdder
{
static void ALWAYS_INLINE add(Data & data, const IColumn & column, size_t row_num)
{
2017-12-25 04:01:46 +00:00
if constexpr (std::is_same_v<Data, AggregateFunctionUniqUniquesHashSetData>
|| std::is_same_v<Data, AggregateFunctionUniqHLL12Data<T>>)
2017-12-02 03:42:21 +00:00
{
2017-12-25 04:01:46 +00:00
if constexpr (!std::is_same_v<T, String>)
2017-12-02 03:42:21 +00:00
{
2019-09-30 15:45:53 +00:00
const auto & value = assert_cast<const ColumnVector<T> &>(column).getElement(row_num);
2017-12-02 03:42:21 +00:00
data.set.insert(AggregateFunctionUniqTraits<T>::hash(value));
}
else
{
StringRef value = column.getDataAt(row_num);
data.set.insert(CityHash_v1_0_2::CityHash64(value.data, value.size));
}
}
2017-12-25 04:01:46 +00:00
else if constexpr (std::is_same_v<Data, AggregateFunctionUniqExactData<T>>)
2017-12-02 03:42:21 +00:00
{
2017-12-25 04:01:46 +00:00
if constexpr (!std::is_same_v<T, String>)
2017-12-02 03:42:21 +00:00
{
data.set.insert(assert_cast<const ColumnVector<T> &>(column).getData()[row_num]);
2017-12-02 03:42:21 +00:00
}
else
{
StringRef value = column.getDataAt(row_num);
UInt128 key;
SipHash hash;
hash.update(value.data, value.size);
hash.get128(key.low, key.high);
data.set.insert(key);
}
}
}
};
}
/// Calculates the number of different values approximately or exactly.
template <typename T, typename Data>
class AggregateFunctionUniq final : public IAggregateFunctionDataHelper<Data, AggregateFunctionUniq<T, Data>>
2011-09-26 04:00:46 +00:00
{
public:
AggregateFunctionUniq(const DataTypes & argument_types_)
: IAggregateFunctionDataHelper<Data, AggregateFunctionUniq<T, Data>>(argument_types_, {}) {}
2019-02-11 19:26:32 +00:00
String getName() const override { return Data::getName(); }
DataTypePtr getReturnType() const override
{
return std::make_shared<DataTypeUInt64>();
}
2019-11-11 08:36:19 +00:00
/// ALWAYS_INLINE is required to have better code layout for uniqHLL12 function
void ALWAYS_INLINE add(AggregateDataPtr place, const IColumn ** columns, size_t row_num, Arena *) const override
{
detail::OneAdder<T, Data>::add(this->data(place), *columns[0], row_num);
}
2017-12-01 21:51:50 +00:00
void merge(AggregateDataPtr place, ConstAggregateDataPtr rhs, Arena *) const override
{
this->data(place).set.merge(this->data(rhs).set);
}
void serialize(ConstAggregateDataPtr place, WriteBuffer & buf) const override
{
this->data(place).set.write(buf);
}
void deserialize(AggregateDataPtr place, ReadBuffer & buf, Arena *) const override
{
this->data(place).set.read(buf);
}
2020-06-17 19:36:27 +00:00
void insertResultInto(AggregateDataPtr place, IColumn & to, Arena *) const override
{
assert_cast<ColumnUInt64 &>(to).getData().push_back(this->data(place).set.size());
}
};
2017-03-09 00:56:38 +00:00
/** For multiple arguments. To compute, hashes them.
* You can pass multiple arguments as is; You can also pass one argument - a tuple.
2018-03-17 18:02:19 +00:00
* But (for the possibility of efficient implementation), you can not pass several arguments, among which there are tuples.
*/
template <typename Data, bool is_exact, bool argument_is_tuple>
class AggregateFunctionUniqVariadic final : public IAggregateFunctionDataHelper<Data, AggregateFunctionUniqVariadic<Data, is_exact, argument_is_tuple>>
{
private:
size_t num_args = 0;
public:
AggregateFunctionUniqVariadic(const DataTypes & arguments)
2019-02-12 09:31:20 +00:00
: IAggregateFunctionDataHelper<Data, AggregateFunctionUniqVariadic<Data, is_exact, argument_is_tuple>>(arguments, {})
{
if (argument_is_tuple)
num_args = typeid_cast<const DataTypeTuple &>(*arguments[0]).getElements().size();
else
num_args = arguments.size();
}
String getName() const override { return Data::getName(); }
DataTypePtr getReturnType() const override
{
return std::make_shared<DataTypeUInt64>();
}
void add(AggregateDataPtr place, const IColumn ** columns, size_t row_num, Arena *) const override
{
2020-08-03 00:01:23 +00:00
this->data(place).set.insert(typename Data::Set::value_type(
UniqVariadicHash<is_exact, argument_is_tuple>::apply(num_args, columns, row_num)));
}
2017-12-01 21:51:50 +00:00
void merge(AggregateDataPtr place, ConstAggregateDataPtr rhs, Arena *) const override
{
this->data(place).set.merge(this->data(rhs).set);
}
void serialize(ConstAggregateDataPtr place, WriteBuffer & buf) const override
{
this->data(place).set.write(buf);
}
void deserialize(AggregateDataPtr place, ReadBuffer & buf, Arena *) const override
{
this->data(place).set.read(buf);
}
2020-06-17 19:36:27 +00:00
void insertResultInto(AggregateDataPtr place, IColumn & to, Arena *) const override
{
assert_cast<ColumnUInt64 &>(to).getData().push_back(this->data(place).set.size());
}
};
2011-09-26 04:00:46 +00:00
}