support_max_min

primitives/linux-port/column.cpp

@@ -41,6 +41,7 @@ using namespace boost;
 #include "dataconvert.h"
 #include "mcs_decimal.h"
 #include "simd_sse.h"
+#include "simd_arm.h"
 #include "utils/common/columnwidth.h"
 
 #include "exceptclasses.h"
@@ -118,24 +119,6 @@ inline int compareBlock(const void* a, const void* b)
   return ((*(T*)a) - (*(T*)b));
 }
 
-template <class To, class From>
-std::enable_if_t<
-    sizeof(To) == sizeof(From) &&
-    std::is_trivially_copyable_v<From> &&
-    std::is_trivially_copyable_v<To>,
-    To>
-// constexpr support needs compiler magic
-bitCast(const From& src) noexcept
-{
-    static_assert(std::is_trivially_constructible_v<To>,
-        "This implementation additionally requires "
-        "destination type to be trivially constructible");
-
-    To dst;
-    std::memcpy(&dst, &src, sizeof(To));
-    return dst;
-}
-
 // this function is out-of-band, we don't need to inline it
 void logIt(int mid, int arg1, const string& arg2 = string())
 {
@@ -942,7 +925,7 @@ inline void writeColValue(uint8_t OutputType, ColResultHeader* out, uint16_t rid
   }
 }
 
-#if defined(__x86_64__)
+#if defined(__x86_64__) || defined(__aarch64__)
 template <typename T, ENUM_KIND KIND, bool HAS_INPUT_RIDS,
           typename std::enable_if<HAS_INPUT_RIDS == false, T>::type* = nullptr>
 inline void vectUpdateMinMax(const bool validMinMax, const bool isNonNullOrEmpty, T& Min, T& Max, T curValue,
@@ -1239,7 +1222,7 @@ void scalarFiltering(
   }
 }
 
-#if defined(__x86_64__)
+#if defined(__x86_64__) || defined(__aarch64__)
 template <typename VT, typename SIMD_WRAPPER_TYPE, bool HAS_INPUT_RIDS, typename T,
           typename std::enable_if<HAS_INPUT_RIDS == false, T>::type* = nullptr>
 inline SIMD_WRAPPER_TYPE simdDataLoad(VT& processor, const T* srcArray, const T* origSrcArray,
@@ -1295,19 +1278,13 @@ inline SIMD_WRAPPER_TYPE simdSwapedOrderDataLoad(const ColRequestHeaderDataType
 }
 
 template <typename VT, typename SimdType>
-void vectorizedUpdateMinMax(const bool validMinMax, const MT nonNullOrEmptyMask, VT simdProcessor,
+void vectorizedUpdateMinMax(const bool validMinMax, const MT nonNullOrEmptyMask, VT& simdProcessor,
                             SimdType& dataVec, SimdType& simdMin, SimdType& simdMax)
 {
-  if (validMinMax)
+  if (validMinMax && nonNullOrEmptyMask)
   {
-    simdMin = simdProcessor.blend(
-        simdMin, dataVec,
-        simdProcessor.bwAnd(simdProcessor.cmpGt2(simdMin, dataVec),
-                            bitCast<SimdType>(simd::bitMaskToByteMask16(nonNullOrEmptyMask))));
-    simdMax = simdProcessor.blend(
-        simdMax, dataVec,
-        simdProcessor.bwAnd(simdProcessor.cmpGt2(dataVec, simdMax),
-                            bitCast<SimdType>(simd::bitMaskToByteMask16(nonNullOrEmptyMask))));
+    simdMin = simdProcessor.min(simdMin, dataVec);
+    simdMax = simdProcessor.max(simdMax, dataVec);
   }
 }
 
@@ -1328,7 +1305,7 @@ void scalarUpdateMinMax(const bool validMinMax, const MT nonNullOrEmptyMask, VT&
 }
 
 template<typename T, typename VT, typename SimdType>
-void extractMinMax(VT& simdProcessor, SimdType simdMin, SimdType simdMax, T& min, T& max)
+void extractMinMax(VT& simdProcessor, SimdType& simdMin, SimdType& simdMax, T& min, T& max)
 {
   constexpr const uint16_t size = VT::vecByteSize / sizeof(T);
   T* simdMinVec = reinterpret_cast<T*>(&simdMin);
@@ -1336,13 +1313,6 @@ void extractMinMax(VT& simdProcessor, SimdType simdMin, SimdType simdMax, T& min
   max = *std::max_element(simdMaxVec, simdMaxVec + size);
   min = *std::min_element(simdMinVec, simdMinVec + size);
 }
-
-template <typename T, typename VT, typename SimdType>
-void getInitialSimdMinMax(VT& simdProcessor, SimdType& simdMin, SimdType& simdMax, T min, T max)
-{
-  simdMin = simdProcessor.loadValue(min);
-  simdMax = simdProcessor.loadValue(max);
-}
 // This routine filters input block in a vectorized manner.
 // It supports all output types, all input types.
 // It doesn't support KIND==TEXT so upper layers filters this KIND out beforehand.
@@ -1478,12 +1448,9 @@ void vectorizedFiltering(NewColRequestHeader* in, ColResultHeader* out, const T*
       }
     }
   }
-  [[maybe_unused]] SimdType simdMin;
-  [[maybe_unused]] SimdType simdMax;
-  if constexpr (KIND != KIND_TEXT)
-  {
-    getInitialSimdMinMax(simdProcessor, simdMin, simdMax, min, max);
-  }
+  [[maybe_unused]] SimdType simdMin = simdDataLoad<VT, SimdWrapperType, HAS_INPUT_RIDS, T>(simdProcessor, srcArray,
+      origSrcArray, ridArray, 0).v;;
+  [[maybe_unused]] SimdType simdMax = simdMin;
   // main loop
   // writeMask tells which values must get into the result. Includes values that matches filters. Can have
   // NULLs. nonEmptyMask tells which vector coords are not EMPTY magics. nonNullMask tells which vector coords
@@ -1704,7 +1671,7 @@ void filterColumnData(NewColRequestHeader* in, ColResultHeader* out, uint16_t* r
   // Syscat queries mustn't follow vectorized processing path b/c PP must return
   // all values w/o any filter(even empty values filter) applied.
 
-#if defined(__x86_64__)
+#if defined(__x86_64__) || defined(__aarch64__)
   // Don't use vectorized filtering for text based data types.
   if (WIDTH < 16 &&
     (KIND != KIND_TEXT || (KIND == KIND_TEXT && in->colType.strnxfrmIsValid()) ))

tests/simd_processors.cpp

@@ -22,37 +22,33 @@
 #include <gtest/gtest.h>
 #include "datatypes/mcs_datatype.h"
 #include "datatypes/mcs_int128.h"
-
+#include "simd_sse.h"
+#include "simd_arm.h"
 #if defined(__x86_64__)
-  #include "simd_sse.h"
   #define TESTS_USING_SSE 1
   using float64_t = double;
   using float32_t = float;
 #endif
 #ifdef __aarch64__
-  #include "simd_arm.h"
   #define TESTS_USING_ARM 1
 #endif
 
 using namespace std;
 
 template <typename T>
-class SimdProcessorTypedTest : public testing::Test
-{
- public:
+class SimdProcessorTypedTest : public testing::Test {
+public:
   using IntegralType = T;
-#if TESTS_USING_SSE
-  using SimdType =
-      std::conditional_t<std::is_same<T, float>::value, simd::vi128f_wr,
-                         std::conditional_t<std::is_same<T, double>::value, simd::vi128d_wr, simd::vi128_wr>>;
-  using Proc = typename simd::SimdFilterProcessor<SimdType, T>;
-#else
-  using SimdType =
-      std::conditional_t<std::is_same<T, float>::value, simd::vi128f_wr,
-                         std::conditional_t<std::is_same<T, double>::value, simd::vi128d_wr,
-                                            typename simd::TypeToVecWrapperType<T>::WrapperType>>;
-  using Proc = typename simd::SimdFilterProcessor<SimdType, T>;
-#endif
+  #if TESTS_USING_SSE
+    using SimdType = std::conditional_t<std::is_same<T, float>::value,
+                                        simd::vi128f_wr,
+                                        std::conditional_t<std::is_same<T, double>::value,
+                                                           simd::vi128d_wr,
+                                                           simd::vi128_wr>>;
+    using Proc = typename simd::SimdFilterProcessor<SimdType, T>;
+    #else
+    using Proc = typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<T>::WrapperType, T>;
+  #endif
   void SetUp() override
   {
   }