ClickHouse/dbms/include/DB/AggregateFunctions/AggregateFunctionGroupUniqArray.h

254 lines
6.7 KiB
C++
Raw Normal View History

#pragma once
#include <DB/IO/WriteHelpers.h>
#include <DB/IO/ReadHelpers.h>
#include <DB/DataTypes/DataTypeArray.h>
#include <DB/DataTypes/DataTypesNumber.h>
#include <DB/DataTypes/DataTypeString.h>
#include <DB/Columns/ColumnArray.h>
#include <DB/Common/HashTable/HashSet.h>
#include <DB/AggregateFunctions/AggregateFunctionGroupArray.h>
#define AGGREGATE_FUNCTION_GROUP_ARRAY_UNIQ_MAX_SIZE 0xFFFFFF
namespace DB
{
template <typename T>
struct AggregateFunctionGroupUniqArrayData
{
2017-03-09 04:26:17 +00:00
/// When creating, the hash table must be small.
using Set = HashSet<
T,
DefaultHash<T>,
2014-05-03 16:03:49 +00:00
HashTableGrower<4>,
HashTableAllocatorWithStackMemory<sizeof(T) * (1 << 4)>
>;
Set value;
};
/// Puts all values to the hash set. Returns an array of unique values. Implemented for numeric types.
template <typename T>
2016-07-10 17:19:35 +00:00
class AggregateFunctionGroupUniqArray
: public IUnaryAggregateFunction<AggregateFunctionGroupUniqArrayData<T>, AggregateFunctionGroupUniqArray<T>>
{
private:
using State = AggregateFunctionGroupUniqArrayData<T>;
2015-11-11 02:04:23 +00:00
public:
2015-11-11 02:04:23 +00:00
String getName() const override { return "groupUniqArray"; }
2015-11-11 02:04:23 +00:00
DataTypePtr getReturnType() const override
{
return std::make_shared<DataTypeArray>(std::make_shared<DataTypeNumber<T>>());
}
void setArgument(const DataTypePtr & argument)
{
}
void addImpl(AggregateDataPtr place, const IColumn & column, size_t row_num, Arena *) const
{
this->data(place).value.insert(static_cast<const ColumnVector<T> &>(column).getData()[row_num]);
}
void merge(AggregateDataPtr place, ConstAggregateDataPtr rhs, Arena * arena) const override
{
this->data(place).value.merge(this->data(rhs).value);
}
2015-11-11 02:04:23 +00:00
void serialize(ConstAggregateDataPtr place, WriteBuffer & buf) const override
{
2016-09-22 23:26:08 +00:00
auto & set = this->data(place).value;
size_t size = set.size();
writeVarUInt(size, buf);
2016-09-22 23:26:08 +00:00
for (auto & elem : set)
writeIntBinary(elem, buf);
}
2016-09-22 23:26:08 +00:00
void deserialize(AggregateDataPtr place, ReadBuffer & buf, Arena *) const override
{
2016-03-12 04:01:03 +00:00
this->data(place).value.read(buf);
}
2015-11-11 02:04:23 +00:00
void insertResultInto(ConstAggregateDataPtr place, IColumn & to) const override
{
ColumnArray & arr_to = static_cast<ColumnArray &>(to);
ColumnArray::Offsets_t & offsets_to = arr_to.getOffsets();
const typename State::Set & set = this->data(place).value;
size_t size = set.size();
offsets_to.push_back((offsets_to.size() == 0 ? 0 : offsets_to.back()) + size);
typename ColumnVector<T>::Container_t & data_to = static_cast<ColumnVector<T> &>(arr_to.getData()).getData();
size_t old_size = data_to.size();
data_to.resize(old_size + size);
size_t i = 0;
for (auto it = set.begin(); it != set.end(); ++it, ++i)
data_to[old_size + i] = *it;
}
};
/// Generic implementation, it uses serialized representation as object descriptor.
struct AggreagteFunctionGroupUniqArrayGenericData
{
static constexpr size_t INIT_ELEMS = 2; /// adjustable
static constexpr size_t ELEM_SIZE = sizeof(HashSetCellWithSavedHash<StringRef, StringRefHash>);
using Set = HashSetWithSavedHash<StringRef, StringRefHash, HashTableGrower<INIT_ELEMS>, HashTableAllocatorWithStackMemory<INIT_ELEMS * ELEM_SIZE>>;
Set value;
};
2016-09-22 23:26:08 +00:00
/** Template parameter with true value should be used for columns that store their elements in memory continuously.
* For such columns groupUniqArray() can be implemented more efficently (especially for small numeric arrays).
*/
template <bool is_plain_column = false>
class AggreagteFunctionGroupUniqArrayGeneric : public IUnaryAggregateFunction<AggreagteFunctionGroupUniqArrayGenericData, AggreagteFunctionGroupUniqArrayGeneric<is_plain_column>>
{
2016-09-22 23:26:08 +00:00
DataTypePtr input_data_type;
using State = AggreagteFunctionGroupUniqArrayGenericData;
static StringRef getSerialization(const IColumn & column, size_t row_num, Arena & arena);
static void deserializeAndInsert(StringRef str, IColumn & data_to);
public:
String getName() const override { return "groupUniqArray"; }
void setArgument(const DataTypePtr & argument)
{
input_data_type = argument;
}
DataTypePtr getReturnType() const override
{
return std::make_shared<DataTypeArray>(input_data_type->clone());
}
bool allocatesMemoryInArena() const override
{
return true;
}
void serialize(ConstAggregateDataPtr place, WriteBuffer & buf) const override
{
auto & set = this->data(place).value;
writeVarUInt(set.size(), buf);
2016-09-22 23:26:08 +00:00
for (const auto & elem : set)
{
writeStringBinary(elem, buf);
}
}
2016-09-22 23:26:08 +00:00
void deserialize(AggregateDataPtr place, ReadBuffer & buf, Arena * arena) const override
{
2016-09-22 23:26:08 +00:00
auto & set = this->data(place).value;
size_t size;
readVarUInt(size, buf);
//TODO: set.reserve(size);
for (size_t i = 0; i < size; i++)
{
2016-09-22 23:26:08 +00:00
set.insert(readStringBinaryInto(*arena, buf));
}
}
void addImpl(AggregateDataPtr place, const IColumn & column, size_t row_num, Arena * arena) const
{
auto & set = this->data(place).value;
bool inserted;
State::Set::iterator it;
StringRef str_serialized = getSerialization(column, row_num, *arena);
set.emplace(str_serialized, it, inserted);
if (!is_plain_column)
{
2016-09-22 23:26:08 +00:00
if (!inserted)
arena->rollback(str_serialized.size);
}
else
{
2016-09-22 23:26:08 +00:00
if (inserted)
it->data = arena->insert(str_serialized.data, str_serialized.size);
}
}
void merge(AggregateDataPtr place, ConstAggregateDataPtr rhs, Arena * arena) const override
{
auto & cur_set = this->data(place).value;
auto & rhs_set = this->data(rhs).value;
bool inserted;
State::Set::iterator it;
for (auto & rhs_elem : rhs_set)
{
cur_set.emplace(rhs_elem, it, inserted);
if (inserted)
it->data = arena->insert(it->data, it->size);
}
}
void insertResultInto(ConstAggregateDataPtr place, IColumn & to) const override
{
ColumnArray & arr_to = static_cast<ColumnArray &>(to);
ColumnArray::Offsets_t & offsets_to = arr_to.getOffsets();
IColumn & data_to = arr_to.getData();
auto & set = this->data(place).value;
offsets_to.push_back((offsets_to.size() == 0 ? 0 : offsets_to.back()) + set.size());
for (auto & elem : set)
{
deserializeAndInsert(elem, data_to);
}
}
};
template <>
inline StringRef AggreagteFunctionGroupUniqArrayGeneric<false>::getSerialization(const IColumn & column, size_t row_num, Arena & arena)
{
const char * begin = nullptr;
return column.serializeValueIntoArena(row_num, arena, begin);
}
template <>
inline StringRef AggreagteFunctionGroupUniqArrayGeneric<true>::getSerialization(const IColumn & column, size_t row_num, Arena &)
{
return column.getDataAt(row_num);
}
template <>
inline void AggreagteFunctionGroupUniqArrayGeneric<false>::deserializeAndInsert(StringRef str, IColumn & data_to)
{
data_to.deserializeAndInsertFromArena(str.data);
}
template <>
inline void AggreagteFunctionGroupUniqArrayGeneric<true>::deserializeAndInsert(StringRef str, IColumn & data_to)
{
data_to.insertData(str.data, str.size);
}
#undef AGGREGATE_FUNCTION_GROUP_ARRAY_UNIQ_MAX_SIZE
}