Merge pull request #58334 from Algunenano/minmax_non_numeric

Speedup MIN/MAX for non numeric types

src/AggregateFunctions/AggregateFunctionMax.cpp

@@ -1,7 +1,8 @@
 #include <AggregateFunctions/AggregateFunctionFactory.h>
 #include <AggregateFunctions/FactoryHelpers.h>
 #include <AggregateFunctions/HelpersMinMaxAny.h>
-#include <AggregateFunctions/findNumeric.h>
+#include <Common/Concepts.h>
+#include <Common/findExtreme.h>
 
 namespace DB
 {
@@ -19,7 +20,7 @@ public:
     explicit AggregateFunctionsSingleValueMax(const DataTypePtr & type) : Parent(type) { }
 
     /// Specializations for native numeric types
-    ALWAYS_INLINE inline void addBatchSinglePlace(
+    void addBatchSinglePlace(
         size_t row_begin,
         size_t row_end,
         AggregateDataPtr __restrict place,
@@ -27,7 +28,7 @@ public:
         Arena * arena,
         ssize_t if_argument_pos) const override;
 
-    ALWAYS_INLINE inline void addBatchSinglePlaceNotNull(
+    void addBatchSinglePlaceNotNull(
         size_t row_begin,
         size_t row_end,
         AggregateDataPtr __restrict place,
@@ -53,10 +54,10 @@ void AggregateFunctionsSingleValueMax<typename DB::AggregateFunctionMaxData<Sing
     if (if_argument_pos >= 0) \
     { \
         const auto & flags = assert_cast<const ColumnUInt8 &>(*columns[if_argument_pos]).getData(); \
-        opt = findNumericMaxIf(column.getData().data(), flags.data(), row_begin, row_end); \
+        opt = findExtremeMaxIf(column.getData().data(), flags.data(), row_begin, row_end); \
     } \
     else \
-        opt = findNumericMax(column.getData().data(), row_begin, row_end); \
+        opt = findExtremeMax(column.getData().data(), row_begin, row_end); \
     if (opt.has_value()) \
         this->data(place).changeIfGreater(opt.value()); \
 }
@@ -74,7 +75,57 @@ void AggregateFunctionsSingleValueMax<Data>::addBatchSinglePlace(
     Arena * arena,
     ssize_t if_argument_pos) const
 {
+    if constexpr (!is_any_of<typename Data::Impl, SingleValueDataString, SingleValueDataGeneric>)
+    {
+        /// Leave other numeric types (large integers, decimals, etc) to keep doing the comparison as it's
+        /// faster than doing a permutation
         return Parent::addBatchSinglePlace(row_begin, row_end, place, columns, arena, if_argument_pos);
+    }
+
+    constexpr int nan_direction_hint = 1;
+    auto const & column = *columns[0];
+    if (if_argument_pos >= 0)
+    {
+        size_t index = row_begin;
+        const auto & if_flags = assert_cast<const ColumnUInt8 &>(*columns[if_argument_pos]).getData();
+        while (if_flags[index] == 0 && index < row_end)
+            index++;
+        if (index >= row_end)
+            return;
+
+        for (size_t i = index + 1; i < row_end; i++)
+        {
+            if ((if_flags[i] != 0) && (column.compareAt(i, index, column, nan_direction_hint) > 0))
+                index = i;
+        }
+        this->data(place).changeIfGreater(column, index, arena);
+    }
+    else
+    {
+        if (row_begin >= row_end)
+            return;
+
+        /// TODO: Introduce row_begin and row_end to getPermutation
+        if (row_begin != 0 || row_end != column.size())
+        {
+            size_t index = row_begin;
+            for (size_t i = index + 1; i < row_end; i++)
+            {
+                if (column.compareAt(i, index, column, nan_direction_hint) > 0)
+                    index = i;
+            }
+            this->data(place).changeIfGreater(column, index, arena);
+        }
+        else
+        {
+            constexpr IColumn::PermutationSortDirection direction = IColumn::PermutationSortDirection::Descending;
+            constexpr IColumn::PermutationSortStability stability = IColumn::PermutationSortStability::Unstable;
+            IColumn::Permutation permutation;
+            constexpr UInt64 limit = 1;
+            column.getPermutation(direction, stability, limit, nan_direction_hint, permutation);
+            this->data(place).changeIfGreater(column, permutation[0], arena);
+        }
+    }
 }
 
 // NOLINTBEGIN(bugprone-macro-parentheses)
@@ -97,10 +148,10 @@ void AggregateFunctionsSingleValueMax<typename DB::AggregateFunctionMaxData<Sing
         auto final_flags = std::make_unique<UInt8[]>(row_end); \
         for (size_t i = row_begin; i < row_end; ++i) \
             final_flags[i] = (!null_map[i]) & !!if_flags[i]; \
-        opt = findNumericMaxIf(column.getData().data(), final_flags.get(), row_begin, row_end); \
+        opt = findExtremeMaxIf(column.getData().data(), final_flags.get(), row_begin, row_end); \
     } \
     else \
-        opt = findNumericMaxNotNull(column.getData().data(), null_map, row_begin, row_end); \
+        opt = findExtremeMaxNotNull(column.getData().data(), null_map, row_begin, row_end); \
     if (opt.has_value()) \
         this->data(place).changeIfGreater(opt.value()); \
 }
@@ -119,7 +170,46 @@ void AggregateFunctionsSingleValueMax<Data>::addBatchSinglePlaceNotNull(
     Arena * arena,
     ssize_t if_argument_pos) const
 {
+    if constexpr (!is_any_of<typename Data::Impl, SingleValueDataString, SingleValueDataGeneric>)
+    {
+        /// Leave other numeric types (large integers, decimals, etc) to keep doing the comparison as it's
+        /// faster than doing a permutation
         return Parent::addBatchSinglePlaceNotNull(row_begin, row_end, place, columns, null_map, arena, if_argument_pos);
+    }
+
+    constexpr int nan_direction_hint = 1;
+    auto const & column = *columns[0];
+    if (if_argument_pos >= 0)
+    {
+        size_t index = row_begin;
+        const auto & if_flags = assert_cast<const ColumnUInt8 &>(*columns[if_argument_pos]).getData();
+        while ((if_flags[index] == 0 || null_map[index] != 0) && (index < row_end))
+            index++;
+        if (index >= row_end)
+            return;
+
+        for (size_t i = index + 1; i < row_end; i++)
+        {
+            if ((if_flags[i] != 0) && (null_map[i] == 0) && (column.compareAt(i, index, column, nan_direction_hint) > 0))
+                index = i;
+        }
+        this->data(place).changeIfGreater(column, index, arena);
+    }
+    else
+    {
+        size_t index = row_begin;
+        while ((null_map[index] != 0) && (index < row_end))
+            index++;
+        if (index >= row_end)
+            return;
+
+        for (size_t i = index + 1; i < row_end; i++)
+        {
+            if ((null_map[i] == 0) && (column.compareAt(i, index, column, nan_direction_hint) > 0))
+                index = i;
+        }
+        this->data(place).changeIfGreater(column, index, arena);
+    }
 }
 
 AggregateFunctionPtr createAggregateFunctionMax(

src/AggregateFunctions/AggregateFunctionMin.cpp

@@ -1,7 +1,8 @@
 #include <AggregateFunctions/AggregateFunctionFactory.h>
 #include <AggregateFunctions/FactoryHelpers.h>
 #include <AggregateFunctions/HelpersMinMaxAny.h>
-#include <AggregateFunctions/findNumeric.h>
+#include <Common/Concepts.h>
+#include <Common/findExtreme.h>
 
 
 namespace DB
@@ -20,7 +21,7 @@ public:
     explicit AggregateFunctionsSingleValueMin(const DataTypePtr & type) : Parent(type) { }
 
     /// Specializations for native numeric types
-    ALWAYS_INLINE inline void addBatchSinglePlace(
+    void addBatchSinglePlace(
         size_t row_begin,
         size_t row_end,
         AggregateDataPtr __restrict place,
@@ -28,7 +29,7 @@ public:
         Arena * arena,
         ssize_t if_argument_pos) const override;
 
-    ALWAYS_INLINE inline void addBatchSinglePlaceNotNull(
+    void addBatchSinglePlaceNotNull(
         size_t row_begin,
         size_t row_end,
         AggregateDataPtr __restrict place,
@@ -54,10 +55,10 @@ public:
         if (if_argument_pos >= 0) \
         { \
             const auto & flags = assert_cast<const ColumnUInt8 &>(*columns[if_argument_pos]).getData(); \
-            opt = findNumericMinIf(column.getData().data(), flags.data(), row_begin, row_end); \
+            opt = findExtremeMinIf(column.getData().data(), flags.data(), row_begin, row_end); \
         } \
         else \
-            opt = findNumericMin(column.getData().data(), row_begin, row_end); \
+            opt = findExtremeMin(column.getData().data(), row_begin, row_end); \
         if (opt.has_value()) \
             this->data(place).changeIfLess(opt.value()); \
     }
@@ -75,7 +76,57 @@ void AggregateFunctionsSingleValueMin<Data>::addBatchSinglePlace(
     Arena * arena,
     ssize_t if_argument_pos) const
 {
+    if constexpr (!is_any_of<typename Data::Impl, SingleValueDataString, SingleValueDataGeneric>)
+    {
+        /// Leave other numeric types (large integers, decimals, etc) to keep doing the comparison as it's
+        /// faster than doing a permutation
         return Parent::addBatchSinglePlace(row_begin, row_end, place, columns, arena, if_argument_pos);
+    }
+
+    constexpr int nan_direction_hint = 1;
+    auto const & column = *columns[0];
+    if (if_argument_pos >= 0)
+    {
+        size_t index = row_begin;
+        const auto & if_flags = assert_cast<const ColumnUInt8 &>(*columns[if_argument_pos]).getData();
+        while (if_flags[index] == 0 && index < row_end)
+            index++;
+        if (index >= row_end)
+            return;
+
+        for (size_t i = index + 1; i < row_end; i++)
+        {
+            if ((if_flags[i] != 0) && (column.compareAt(i, index, column, nan_direction_hint) < 0))
+                index = i;
+        }
+        this->data(place).changeIfLess(column, index, arena);
+    }
+    else
+    {
+        if (row_begin >= row_end)
+            return;
+
+        /// TODO: Introduce row_begin and row_end to getPermutation
+        if (row_begin != 0 || row_end != column.size())
+        {
+            size_t index = row_begin;
+            for (size_t i = index + 1; i < row_end; i++)
+            {
+                if (column.compareAt(i, index, column, nan_direction_hint) < 0)
+                    index = i;
+            }
+            this->data(place).changeIfLess(column, index, arena);
+        }
+        else
+        {
+            constexpr IColumn::PermutationSortDirection direction = IColumn::PermutationSortDirection::Ascending;
+            constexpr IColumn::PermutationSortStability stability = IColumn::PermutationSortStability::Unstable;
+            IColumn::Permutation permutation;
+            constexpr UInt64 limit = 1;
+            column.getPermutation(direction, stability, limit, nan_direction_hint, permutation);
+            this->data(place).changeIfLess(column, permutation[0], arena);
+        }
+    }
 }
 
 // NOLINTBEGIN(bugprone-macro-parentheses)
@@ -98,10 +149,10 @@ void AggregateFunctionsSingleValueMin<Data>::addBatchSinglePlace(
             auto final_flags = std::make_unique<UInt8[]>(row_end); \
             for (size_t i = row_begin; i < row_end; ++i) \
                 final_flags[i] = (!null_map[i]) & !!if_flags[i]; \
-            opt = findNumericMinIf(column.getData().data(), final_flags.get(), row_begin, row_end); \
+            opt = findExtremeMinIf(column.getData().data(), final_flags.get(), row_begin, row_end); \
         } \
         else \
-            opt = findNumericMinNotNull(column.getData().data(), null_map, row_begin, row_end); \
+            opt = findExtremeMinNotNull(column.getData().data(), null_map, row_begin, row_end); \
         if (opt.has_value()) \
             this->data(place).changeIfLess(opt.value()); \
     }
@@ -120,7 +171,46 @@ void AggregateFunctionsSingleValueMin<Data>::addBatchSinglePlaceNotNull(
     Arena * arena,
     ssize_t if_argument_pos) const
 {
+    if constexpr (!is_any_of<typename Data::Impl, SingleValueDataString, SingleValueDataGeneric>)
+    {
+        /// Leave other numeric types (large integers, decimals, etc) to keep doing the comparison as it's
+        /// faster than doing a permutation
         return Parent::addBatchSinglePlaceNotNull(row_begin, row_end, place, columns, null_map, arena, if_argument_pos);
+    }
+
+    constexpr int nan_direction_hint = 1;
+    auto const & column = *columns[0];
+    if (if_argument_pos >= 0)
+    {
+        size_t index = row_begin;
+        const auto & if_flags = assert_cast<const ColumnUInt8 &>(*columns[if_argument_pos]).getData();
+        while ((if_flags[index] == 0 || null_map[index] != 0) && (index < row_end))
+            index++;
+        if (index >= row_end)
+            return;
+
+        for (size_t i = index + 1; i < row_end; i++)
+        {
+            if ((if_flags[i] != 0) && (null_map[index] == 0) && (column.compareAt(i, index, column, nan_direction_hint) < 0))
+                index = i;
+        }
+        this->data(place).changeIfLess(column, index, arena);
+    }
+    else
+    {
+        size_t index = row_begin;
+        while ((null_map[index] != 0) && (index < row_end))
+            index++;
+        if (index >= row_end)
+            return;
+
+        for (size_t i = index + 1; i < row_end; i++)
+        {
+            if ((null_map[i] == 0) && (column.compareAt(i, index, column, nan_direction_hint) < 0))
+                index = i;
+        }
+        this->data(place).changeIfLess(column, index, arena);
+    }
 }
 
 AggregateFunctionPtr createAggregateFunctionMin(

src/AggregateFunctions/AggregateFunctionMinMaxAny.h

@@ -965,6 +965,7 @@ template <typename Data>
 struct AggregateFunctionMinData : Data
 {
     using Self = AggregateFunctionMinData;
+    using Impl = Data;
 
     bool changeIfBetter(const IColumn & column, size_t row_num, Arena * arena)     { return this->changeIfLess(column, row_num, arena); }
     bool changeIfBetter(const Self & to, Arena * arena)                            { return this->changeIfLess(to, arena); }
@@ -993,6 +994,7 @@ template <typename Data>
 struct AggregateFunctionMaxData : Data
 {
     using Self = AggregateFunctionMaxData;
+    using Impl = Data;
 
     bool changeIfBetter(const IColumn & column, size_t row_num, Arena * arena)     { return this->changeIfGreater(column, row_num, arena); }
     bool changeIfBetter(const Self & to, Arena * arena)                            { return this->changeIfGreater(to, arena); }

src/AggregateFunctions/findNumeric.cpp

@@ -1,15 +0,0 @@
-#include <AggregateFunctions/findNumeric.h>
-
-namespace DB
-{
-#define INSTANTIATION(T) \
-    template std::optional<T> findNumericMin(const T * __restrict ptr, size_t start, size_t end); \
-    template std::optional<T> findNumericMinNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
-    template std::optional<T> findNumericMinIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
-    template std::optional<T> findNumericMax(const T * __restrict ptr, size_t start, size_t end); \
-    template std::optional<T> findNumericMaxNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
-    template std::optional<T> findNumericMaxIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end);
-
-FOR_BASIC_NUMERIC_TYPES(INSTANTIATION)
-#undef INSTANTIATION
-}

src/Common/findExtreme.cpp

@@ -1,18 +1,9 @@
-#pragma once
-
 #include <DataTypes/IDataType.h>
-#include <base/defines.h>
-#include <base/types.h>
-#include <Common/Concepts.h>
 #include <Common/TargetSpecific.h>
-
-#include <algorithm>
-#include <optional>
+#include <Common/findExtreme.h>
 
 namespace DB
 {
-template <typename T>
-concept is_any_native_number = (is_any_of<T, Int8, Int16, Int32, Int64, UInt8, UInt16, UInt32, UInt64, Float32, Float64>);
 
 template <is_any_native_number T>
 struct MinComparator
@@ -28,8 +19,8 @@ struct MaxComparator
 
 MULTITARGET_FUNCTION_AVX2_SSE42(
     MULTITARGET_FUNCTION_HEADER(template <is_any_native_number T, typename ComparatorClass, bool add_all_elements, bool add_if_cond_zero> static std::optional<T> NO_INLINE),
-    findNumericExtremeImpl,
-    MULTITARGET_FUNCTION_BODY((const T * __restrict ptr, const UInt8 * __restrict condition_map [[maybe_unused]], size_t row_begin, size_t row_end)
+    findExtremeImpl,
+    MULTITARGET_FUNCTION_BODY((const T * __restrict ptr, const UInt8 * __restrict condition_map [[maybe_unused]], size_t row_begin, size_t row_end) /// NOLINT
     {
         size_t count = row_end - row_begin;
         ptr += row_begin;
@@ -86,69 +77,67 @@ MULTITARGET_FUNCTION_AVX2_SSE42(
     }
 ))
 
-
 /// Given a vector of T finds the extreme (MIN or MAX) value
 template <is_any_native_number T, class ComparatorClass, bool add_all_elements, bool add_if_cond_zero>
 static std::optional<T>
-findNumericExtreme(const T * __restrict ptr, const UInt8 * __restrict condition_map [[maybe_unused]], size_t start, size_t end)
+findExtreme(const T * __restrict ptr, const UInt8 * __restrict condition_map [[maybe_unused]], size_t start, size_t end)
 {
 #if USE_MULTITARGET_CODE
     /// We see no benefit from using AVX512BW or AVX512F (over AVX2), so we only declare SSE and AVX2
     if (isArchSupported(TargetArch::AVX2))
-        return findNumericExtremeImplAVX2<T, ComparatorClass, add_all_elements, add_if_cond_zero>(ptr, condition_map, start, end);
+        return findExtremeImplAVX2<T, ComparatorClass, add_all_elements, add_if_cond_zero>(ptr, condition_map, start, end);
 
     if (isArchSupported(TargetArch::SSE42))
-        return findNumericExtremeImplSSE42<T, ComparatorClass, add_all_elements, add_if_cond_zero>(ptr, condition_map, start, end);
+        return findExtremeImplSSE42<T, ComparatorClass, add_all_elements, add_if_cond_zero>(ptr, condition_map, start, end);
 #endif
-    return findNumericExtremeImpl<T, ComparatorClass, add_all_elements, add_if_cond_zero>(ptr, condition_map, start, end);
+    return findExtremeImpl<T, ComparatorClass, add_all_elements, add_if_cond_zero>(ptr, condition_map, start, end);
 }
 
 template <is_any_native_number T>
-std::optional<T> findNumericMin(const T * __restrict ptr, size_t start, size_t end)
+std::optional<T> findExtremeMin(const T * __restrict ptr, size_t start, size_t end)
 {
-    return findNumericExtreme<T, MinComparator<T>, true, false>(ptr, nullptr, start, end);
+    return findExtreme<T, MinComparator<T>, true, false>(ptr, nullptr, start, end);
 }
 
 template <is_any_native_number T>
-std::optional<T> findNumericMinNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end)
+std::optional<T> findExtremeMinNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end)
 {
-    return findNumericExtreme<T, MinComparator<T>, false, true>(ptr, condition_map, start, end);
+    return findExtreme<T, MinComparator<T>, false, true>(ptr, condition_map, start, end);
 }
 
 template <is_any_native_number T>
-std::optional<T> findNumericMinIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end)
+std::optional<T> findExtremeMinIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end)
 {
-    return findNumericExtreme<T, MinComparator<T>, false, false>(ptr, condition_map, start, end);
+    return findExtreme<T, MinComparator<T>, false, false>(ptr, condition_map, start, end);
 }
 
 template <is_any_native_number T>
-std::optional<T> findNumericMax(const T * __restrict ptr, size_t start, size_t end)
+std::optional<T> findExtremeMax(const T * __restrict ptr, size_t start, size_t end)
 {
-    return findNumericExtreme<T, MaxComparator<T>, true, false>(ptr, nullptr, start, end);
+    return findExtreme<T, MaxComparator<T>, true, false>(ptr, nullptr, start, end);
 }
 
 template <is_any_native_number T>
-std::optional<T> findNumericMaxNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end)
+std::optional<T> findExtremeMaxNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end)
 {
-    return findNumericExtreme<T, MaxComparator<T>, false, true>(ptr, condition_map, start, end);
+    return findExtreme<T, MaxComparator<T>, false, true>(ptr, condition_map, start, end);
 }
 
 template <is_any_native_number T>
-std::optional<T> findNumericMaxIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end)
+std::optional<T> findExtremeMaxIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end)
 {
-    return findNumericExtreme<T, MaxComparator<T>, false, false>(ptr, condition_map, start, end);
+    return findExtreme<T, MaxComparator<T>, false, false>(ptr, condition_map, start, end);
 }
 
 
-#define EXTERN_INSTANTIATION(T) \
-    extern template std::optional<T> findNumericMin(const T * __restrict ptr, size_t start, size_t end); \
-    extern template std::optional<T> findNumericMinNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
-    extern template std::optional<T> findNumericMinIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
-    extern template std::optional<T> findNumericMax(const T * __restrict ptr, size_t start, size_t end); \
-    extern template std::optional<T> findNumericMaxNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
-    extern template std::optional<T> findNumericMaxIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end);
-
-    FOR_BASIC_NUMERIC_TYPES(EXTERN_INSTANTIATION)
-#undef EXTERN_INSTANTIATION
+#define INSTANTIATION(T) \
+    template std::optional<T> findExtremeMin(const T * __restrict ptr, size_t start, size_t end); \
+    template std::optional<T> findExtremeMinNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
+    template std::optional<T> findExtremeMinIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
+    template std::optional<T> findExtremeMax(const T * __restrict ptr, size_t start, size_t end); \
+    template std::optional<T> findExtremeMaxNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
+    template std::optional<T> findExtremeMaxIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end);
 
+FOR_BASIC_NUMERIC_TYPES(INSTANTIATION)
+#undef INSTANTIATION
 }

src/Common/findExtreme.h

@@ -0,0 +1,45 @@
+#pragma once
+
+#include <DataTypes/IDataType.h>
+#include <base/defines.h>
+#include <base/types.h>
+#include <Common/Concepts.h>
+
+#include <algorithm>
+#include <optional>
+
+namespace DB
+{
+template <typename T>
+concept is_any_native_number = (is_any_of<T, Int8, Int16, Int32, Int64, UInt8, UInt16, UInt32, UInt64, Float32, Float64>);
+
+template <is_any_native_number T>
+std::optional<T> findExtremeMin(const T * __restrict ptr, size_t start, size_t end);
+
+template <is_any_native_number T>
+std::optional<T> findExtremeMinNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end);
+
+template <is_any_native_number T>
+std::optional<T> findExtremeMinIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end);
+
+template <is_any_native_number T>
+std::optional<T> findExtremeMax(const T * __restrict ptr, size_t start, size_t end);
+
+template <is_any_native_number T>
+std::optional<T> findExtremeMaxNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end);
+
+template <is_any_native_number T>
+std::optional<T> findExtremeMaxIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end);
+
+#define EXTERN_INSTANTIATION(T) \
+    extern template std::optional<T> findExtremeMin(const T * __restrict ptr, size_t start, size_t end); \
+    extern template std::optional<T> findExtremeMinNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
+    extern template std::optional<T> findExtremeMinIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
+    extern template std::optional<T> findExtremeMax(const T * __restrict ptr, size_t start, size_t end); \
+    extern template std::optional<T> findExtremeMaxNotNull(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end); \
+    extern template std::optional<T> findExtremeMaxIf(const T * __restrict ptr, const UInt8 * __restrict condition_map, size_t start, size_t end);
+
+    FOR_BASIC_NUMERIC_TYPES(EXTERN_INSTANTIATION)
+#undef EXTERN_INSTANTIATION
+
+}

tests/performance/agg_functions_min_max_any.xml

@@ -87,4 +87,9 @@
 <query>select any(FromTag) from hits_100m_single where FromTag != '' group by intHash32(UserID) % {group_scale} FORMAT Null</query>
 <query>select anyHeavy(FromTag) from hits_100m_single where FromTag != '' group by intHash32(UserID) % {group_scale} FORMAT Null</query>
 
+<!-- Test with tuples (useful when you want to keep 2 columns of the same row) -->
+<query>select min((WatchID, CounterID)) from hits_100m_single FORMAT Null</query>
+<query>select max((WatchID, CounterID)) from hits_100m_single FORMAT Null</query>
+<query>select any((WatchID, CounterID)) from hits_100m_single FORMAT Null</query>
+<query>select anyHeavy((WatchID, CounterID)) from hits_100m_single FORMAT Null</query>
 </test>