Tests for simd min/max

tests/CMakeLists.txt

@@ -1,5 +1,6 @@
 include_directories( ${ENGINE_COMMON_INCLUDES} ${ENGINE_BLOCKCACHE_INCLUDE} ${ENGINE_PRIMPROC_INCLUDE} )
 cmake_policy(SET CMP0054 NEW)
+SET(CMAKE_CXX_STANDARD 20)
 
 if (WITH_UNITTESTS)
     set(EXTERNAL_INSTALL_LOCATION ${CMAKE_BINARY_DIR}/external)

tests/simd_processors.cpp

@@ -16,22 +16,40 @@
    MA 02110-1301, USA. */
 
 
+#include <cstdint>
 #include <iostream>
+#include <type_traits>
 #include <gtest/gtest.h>
 #include "datatypes/mcs_datatype.h"
 #include "datatypes/mcs_int128.h"
 
 #if defined(__x86_64__)
-#include "simd_sse.h"
+  #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:
   using IntegralType = T;
-  public:
+  #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, TypeParam>;
+  #endif
   void SetUp() override
   {
   }
@@ -42,53 +60,7 @@ TYPED_TEST_SUITE(SimdProcessorTypedTest, SimdProcessor128TypedTestTypes);
 
 TYPED_TEST(SimdProcessorTypedTest, SimdFilterProcessor_simd128)
 {
-  using Proc = typename simd::SimdFilterProcessor<simd::vi128_wr, TypeParam>;
+  using Proc = typename SimdProcessorTypedTest<TypeParam>::Proc;
-  using SimdType = typename Proc::SimdType;
-  constexpr static simd::MT allTrue = 0xFFFF;
-  constexpr static simd::MT allFalse = 0x0;
-  Proc proc;
-  SimdType lhs = proc.loadValue((TypeParam)-2);
-  SimdType rhs = proc.loadValue((TypeParam)-3);
-  EXPECT_GT((uint64_t)-2LL, (uint64_t)-3LL);
-  EXPECT_EQ(proc.cmpGe(lhs, rhs), allTrue);
-  EXPECT_EQ(proc.cmpGt(lhs, rhs), allTrue);
-  EXPECT_EQ(proc.cmpGe(rhs, lhs), allFalse);
-  EXPECT_EQ(proc.cmpGt(rhs, lhs), allFalse);
-  EXPECT_EQ(proc.cmpLe(rhs, lhs), allTrue);
-  EXPECT_EQ(proc.cmpLt(rhs, lhs), allTrue);
-  EXPECT_EQ(proc.cmpLe(lhs, rhs), allFalse);
-  EXPECT_EQ(proc.cmpLt(lhs, rhs), allFalse);
-  EXPECT_EQ(proc.cmpEq(rhs, lhs), allFalse);
-  EXPECT_EQ(proc.cmpNe(rhs, lhs), allTrue);
-  lhs = proc.loadValue((TypeParam)-3);
-  EXPECT_EQ(proc.cmpEq(lhs, rhs), allTrue);
-  EXPECT_EQ(proc.cmpNe(rhs, lhs), allFalse);
-}
-#endif
-#ifdef __aarch64__
-#include "simd_arm.h"
-
-
-using namespace std;
-
-template <typename T>
-class SimdProcessorTypedTest : public testing::Test
-{
-  using IntegralType = T;
-
- public:
-  void SetUp() override
-  {
-  }
-};
-
-using SimdProcessor128TypedTestTypes =
-    ::testing::Types<uint64_t, uint32_t, uint16_t, uint8_t, int64_t, int32_t, int16_t, int8_t>;
-TYPED_TEST_SUITE(SimdProcessorTypedTest, SimdProcessor128TypedTestTypes);
-
-TYPED_TEST(SimdProcessorTypedTest, SimdFilterProcessor_simdarm128)
-{
-  using Proc = typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<TypeParam>::WrapperType, TypeParam>;
   using SimdType = typename Proc::SimdType;
   constexpr static simd::MT allTrue = 0xFFFF;
   constexpr static simd::MT allFalse = 0x0;
@@ -122,21 +94,32 @@ TYPED_TEST(SimdProcessorTypedTest, SimdFilterProcessor_simdarm128)
   EXPECT_EQ(proc.cmpEq(rhs, lhs), allTrue);
   EXPECT_EQ(proc.cmpNe(rhs, lhs), allFalse);
 }
-TEST(SimdProcessorTest,Int8)
+
+
+TEST(SimdProcessorTest, Int8)
 {
-    using Proc = typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<int8_t>::WrapperType, int8_t>;
+    using Proc = typename SimdProcessorTypedTest<int8_t>::Proc;
     using SimdType = typename Proc::SimdType;
     Proc proc;
+    constexpr static simd::MT allTrue = 0xFFFF;
     simd::MT expect = 0x0;
     int8_t l[16]{0, 1, 2, 5, 4, 3, 8, 5, 6, 10, 58, 2, 32, 41, 2, 5};
     int8_t r[16]{0, 1, 8, 35, 24, 13, 8, 25, 16, 10, 58, 2, 32, 41, 2, 5};
+    int8_t minlr[16]{0, 1, 2, 5, 4, 3, 8, 5, 6, 10, 58, 2, 32, 41, 2, 5};
+    int8_t maxlr[16]{0, 1, 8, 35, 24, 13, 8, 25, 16, 10, 58, 2, 32, 41, 2, 5};
     SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
     SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+    SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+    SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
     for (int i = 0; i < 16; i++)
       if (l[i] > r[i])
         expect |= 1 << i;
     EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
     EXPECT_EQ(proc.cmpLe(lhs, rhs),(simd::MT) ~expect);
+    SimdType testmax = proc.max(lhs, rhs);
+    SimdType testmin = proc.min(lhs, rhs);
+    EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+    EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
     expect = 0x0;
     for (int i = 0; i < 16; i++)
@@ -154,20 +137,28 @@ TEST(SimdProcessorTest,Int8)
 }
 TEST(SimdProcessorTest, Uint8)
 {
-  using Proc =
+  using Proc = typename SimdProcessorTypedTest<uint8_t>::Proc;
-      typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<uint8_t>::WrapperType, uint8_t>;
   using SimdType = typename Proc::SimdType;
   Proc proc;
+  constexpr static simd::MT allTrue = 0xFFFF;
   simd::MT expect = 0x0;
   uint8_t l[16]{0, 1, 2, 5, 4, 3, 8, 5, 6, 10, 5, 2, 32, 41, 2, 5};
   uint8_t r[16]{0, 1, 8, 35, 24, 13, 8, 25, 16, 10, 58, 2, 32, 41, 2, 5};
+  uint8_t minlr[16]{0, 1, 2, 5, 4, 3, 8, 5, 6, 10, 5, 2, 32, 41, 2, 5};
+  uint8_t maxlr[16]{0, 1, 8, 35, 24, 13, 8, 25, 16, 10, 58, 2, 32, 41, 2, 5};
   SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
   SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+  SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+  SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
   for (int i = 0; i < 16; i++)
     if (l[i] > r[i])
       expect |= 1 << i;
   EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpLe(lhs, rhs),(simd::MT) ~expect);
+  SimdType testmax = proc.max(lhs, rhs);
+  SimdType testmin = proc.min(lhs, rhs);
+  EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+  EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
   expect = 0x0;
   for (int i = 0; i < 16; i++)
@@ -183,22 +174,31 @@ TEST(SimdProcessorTest, Uint8)
   EXPECT_EQ(proc.cmpLt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpGe(lhs, rhs),(simd::MT) ~expect);
 }
+
 TEST(SimdProcessorTest, Int16)
 {
-  using Proc =
+  using Proc = typename SimdProcessorTypedTest<int16_t>::Proc;
-      typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<int16_t>::WrapperType, int16_t>;
   using SimdType = typename Proc::SimdType;
   Proc proc;
+  constexpr static simd::MT allTrue = 0xFFFF;
   simd::MT expect = 0x0;
   int16_t l[8]{0, 1, 2, -5, 4, 3, -8, 200};
-  int16_t r[8]{0, 105, -8, 35, 24, 13, 8};
+  int16_t r[8]{0, 105, -8, 35, 24, 13, 8, 100};
+  int16_t minlr[8]{0, 1, -8, -5, 4, 3, -8, 100};
+  int16_t maxlr[8]{0, 105, 2, 35, 24, 13, 8, 200};
   SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
   SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+  SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+  SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
   for (int i = 0; i < 8; i++)
     if (l[i] > r[i])
       expect |= 3 << i * 2;
   EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpLe(lhs, rhs), (simd::MT)~expect);
+  SimdType testmax = proc.max(lhs, rhs);
+  SimdType testmin = proc.min(lhs, rhs);
+  EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+  EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
   expect = 0x0;
   for (int i = 0; i < 8; i++)
@@ -216,20 +216,28 @@ TEST(SimdProcessorTest, Int16)
 }
 TEST(SimdProcessorTest, Uint16)
 {
-  using Proc = typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<uint16_t>::WrapperType,
+  using Proc = typename SimdProcessorTypedTest<uint16_t>::Proc;
-                                                  uint16_t>;
   using SimdType = typename Proc::SimdType;
   Proc proc;
+  constexpr static simd::MT allTrue = 0xFFFF;
   simd::MT expect = 0x0;
   uint16_t l[8]{0, 1, 2, 5, 4, 3, 8, 5};
-  uint16_t r[8]{0, 1, 8, 35, 24, 13, 8};
+  uint16_t r[8]{0, 1, 8, 35, 24, 13, 8, 17};
+  uint16_t minlr[8]{0, 1, 2, 5, 4, 3, 8, 5};
+  uint16_t maxlr[8]{0, 1, 8, 35, 24, 13, 8, 17};
   SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
   SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+  SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+  SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
   for (int i = 0; i < 8; i++)
     if (l[i] > r[i])
       expect |= 3 << i*2;
   EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpLe(lhs, rhs), (simd::MT)~expect);
+  SimdType testmax = proc.max(lhs, rhs);
+  SimdType testmin = proc.min(lhs, rhs);
+  EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+  EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
   expect = 0x0;
   for (int i = 0; i < 8; i++)
@@ -248,20 +256,28 @@ TEST(SimdProcessorTest, Uint16)
 
 TEST(SimdProcessorTest, Int32)
 {
-  using Proc =
+  using Proc = typename SimdProcessorTypedTest<int32_t>::Proc;
-      typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<int32_t>::WrapperType, int32_t>;
   using SimdType = typename Proc::SimdType;
   Proc proc;
+  constexpr static simd::MT allTrue = 0xFFFF;
   simd::MT expect = 0x0;
   int32_t l[8]{0, 1, 2, -5};
   int32_t r[8]{0, 105, -8,54333};
+  int32_t minlr[8]{0, 1, -8, -5};
+  int32_t maxlr[8]{0, 105, 2, 54333};
   SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
   SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+  SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+  SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
   for (int i = 0; i < 4; i++)
     if (l[i] > r[i])
       expect |= 15 << i * 4;
   EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpLe(lhs, rhs), (simd::MT)~expect);
+  SimdType testmax = proc.max(lhs, rhs);
+  SimdType testmin = proc.min(lhs, rhs);
+  EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+  EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
   expect = 0x0;
   for (int i = 0; i < 4; i++)
@@ -279,20 +295,28 @@ TEST(SimdProcessorTest, Int32)
 }
 TEST(SimdProcessorTest, Uint32)
 {
-  using Proc = typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<uint32_t>::WrapperType,
+  using Proc = typename SimdProcessorTypedTest<uint32_t>::Proc;
-                                                  uint32_t>;
   using SimdType = typename Proc::SimdType;
   Proc proc;
+  constexpr static simd::MT allTrue = 0xFFFF;
   simd::MT expect = 0x0;
   uint32_t l[4]{0, 1002, 2, 514};
   uint32_t r[4]{2, 1, 80555, 35};
+  uint32_t minlr[8]{0, 1, 2, 35};
+  uint32_t maxlr[8]{2, 1002, 80555, 514};
   SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
   SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+  SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+  SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
   for (int i = 0; i < 4; i++)
     if (l[i] > r[i])
       expect |= 15 << i * 4;
   EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpLe(lhs, rhs), (simd::MT)~expect);
+  SimdType testmax = proc.max(lhs, rhs);
+  SimdType testmin = proc.min(lhs, rhs);
+  EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+  EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
   expect = 0x0;
   for (int i = 0; i < 4; i++)
@@ -310,20 +334,28 @@ TEST(SimdProcessorTest, Uint32)
 }
 TEST(SimdProcessorTest, Int64)
 {
-  using Proc =
+  using Proc = typename SimdProcessorTypedTest<int64_t>::Proc;
-      typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<int64_t>::WrapperType, int64_t>;
   using SimdType = typename Proc::SimdType;
   Proc proc;
+  constexpr static simd::MT allTrue = 0xFFFF;
   simd::MT expect = 0x0;
   int64_t l[2]{-5, 122020};
   int64_t r[2]{0, 105};
+  int64_t minlr[8]{-5, 105};
+  int64_t maxlr[8]{0, 122020};
   SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
   SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+  SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+  SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
   for (int i = 0; i < 2; i++)
     if (l[i] > r[i])
       expect |= 0xFF << i * 8;
   EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpLe(lhs, rhs), (simd::MT)~expect);
+  SimdType testmax = proc.max(lhs, rhs);
+  SimdType testmin = proc.min(lhs, rhs);
+  EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+  EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
   expect = 0x0;
   for (int i = 0; i < 2; i++)
@@ -341,20 +373,28 @@ TEST(SimdProcessorTest, Int64)
 }
 TEST(SimdProcessorTest, Uint64)
 {
-  using Proc = typename simd::SimdFilterProcessor<typename simd::TypeToVecWrapperType<uint64_t>::WrapperType,
+  using Proc = typename SimdProcessorTypedTest<uint64_t>::Proc;
-                                                  uint64_t>;
   using SimdType = typename Proc::SimdType;
   Proc proc;
+  constexpr static simd::MT allTrue = 0xFFFF;
   simd::MT expect = 0x0;
   uint64_t l[2]{822, 1002};
   uint64_t r[2]{2, 1};
+  uint64_t minlr[8]{2, 1};
+  uint64_t maxlr[8]{822, 1002};
   SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
   SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+  SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+  SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
   for (int i = 0; i < 2; i++)
     if (l[i] > r[i])
       expect |= 0xFF << i * 8;
   EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpLe(lhs, rhs), (simd::MT)~expect);
+  SimdType testmax = proc.max(lhs, rhs);
+  SimdType testmin = proc.min(lhs, rhs);
+  EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+  EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
   expect = 0x0;
   for (int i = 0; i < 2; i++)
@@ -372,19 +412,28 @@ TEST(SimdProcessorTest, Uint64)
 }
 TEST(SimdProcessorTest, Float64)
 {
-  using Proc = typename simd::SimdFilterProcessor<simd::vi128d_wr, double>;
+  using Proc = typename SimdProcessorTypedTest<double>::Proc;
   using SimdType = typename Proc::SimdType;
   Proc proc;
+  constexpr static simd::MT allTrue = 0xFFFF;
   simd::MT expect = 0x0;
   float64_t l[2]{-5.0, 12.5620};
   float64_t r[2]{2.9, 1};
+  float64_t minlr[8]{-5.0, 1};
+  float64_t maxlr[8]{2.9, 12.5620};
   SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
   SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+  SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+  SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
   for (int i = 0; i < 2; i++)
     if (l[i] > r[i])
       expect |= 0xFF << i * 8;
   EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpLe(lhs, rhs), (simd::MT)~expect);
+  SimdType testmax = proc.max(lhs, rhs);
+  SimdType testmin = proc.min(lhs, rhs);
+  EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+  EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
   expect = 0x0;
   for (int i = 0; i < 2; i++)
@@ -402,19 +451,28 @@ TEST(SimdProcessorTest, Float64)
 }
 TEST(SimdProcessorTest, Float32)
 {
-  using Proc = typename simd::SimdFilterProcessor<simd::vi128f_wr, float>;
+  using Proc = typename SimdProcessorTypedTest<float>::Proc;
   using SimdType = typename Proc::SimdType;
   Proc proc;
+  constexpr static simd::MT allTrue = 0xFFFF;
   simd::MT expect = 0x0;
   float32_t l[4]{82, 102,-5.6,9.5};
   float32_t r[4]{2.0, 1,-5.7,6};
+  float32_t minlr[8]{2.0, 1, -5.7, 6};
+  float32_t maxlr[8]{82, 102, -5.6, 9.5};
   SimdType lhs = proc.loadFrom(reinterpret_cast<char*>(l));
   SimdType rhs = proc.loadFrom(reinterpret_cast<char*>(r));
+  SimdType min = proc.loadFrom(reinterpret_cast<char*>(minlr));
+  SimdType max = proc.loadFrom(reinterpret_cast<char*>(maxlr));
   for (int i = 0; i < 4; i++)
     if (l[i] > r[i])
       expect |= 15 << i * 4;
   EXPECT_EQ(proc.cmpGt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpLe(lhs, rhs), (simd::MT)~expect);
+  SimdType testmax = proc.max(lhs, rhs);
+  SimdType testmin = proc.min(lhs, rhs);
+  EXPECT_EQ(proc.cmpEq(testmax, max), allTrue);
+  EXPECT_EQ(proc.cmpEq(testmin, min), allTrue);
 
   expect = 0x0;
   for (int i = 0; i < 4; i++)
@@ -430,4 +488,3 @@ TEST(SimdProcessorTest, Float32)
   EXPECT_EQ(proc.cmpLt(lhs, rhs), expect);
   EXPECT_EQ(proc.cmpGe(lhs, rhs), (simd::MT)~expect);
 }
-#endif

utils/common/simd_arm.h

@@ -165,13 +165,13 @@ template <typename T, typename ENABLE=void>
 struct TypeToVecWrapperType;
 
 template <typename T>
-struct TypeToVecWrapperType<T, typename std::enable_if<std::is_unsigned<T>::value>::type> 
+struct TypeToVecWrapperType<T, typename std::enable_if<std::is_unsigned<T>::value>::type>
     : WidthToVecWrapperType<sizeof(T)>
 {
 };
 
 template <typename T>
-struct TypeToVecWrapperType<T, typename std::enable_if<std::is_signed<T>::value>::type> 
+struct TypeToVecWrapperType<T, typename std::enable_if<std::is_signed<T>::value>::type>
     : WidthToSVecWrapperType<sizeof(T)>
 {
 };
@@ -419,6 +419,16 @@ class SimdFilterProcessor<
   {
     vst1q_s32(reinterpret_cast<int32_t*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return reinterpret_cast<SimdType>(std::min(reinterpret_cast<int128_t>(x), reinterpret_cast<int128_t>(y)));
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return reinterpret_cast<SimdType>(std::max(reinterpret_cast<int128_t>(x), reinterpret_cast<int128_t>(y)));
+  }
 };
 
 template <typename VT, typename T>
@@ -532,6 +542,16 @@ class SimdFilterProcessor<
   {
     vst1q_f64(reinterpret_cast<T*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vminq_f64(x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vmaxq_f64(x, y);
+  }
 };
 
 template <typename VT, typename T>
@@ -645,6 +665,16 @@ class SimdFilterProcessor<
   {
     vst1q_f32(reinterpret_cast<T*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vminq_f32(x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vmaxq_f32(x, y);
+  }
 };
 
 template <typename VT, typename CHECK_T>
@@ -752,6 +782,16 @@ class SimdFilterProcessor<
   {
     vst1q_s64(reinterpret_cast<int64_t*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vbslq_s64(vcgtq_s64(x,y), x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vbslq_s64(vcgtq_s64(y,x), x, y);
+  }
 };
 
 template <typename VT, typename CHECK_T>
@@ -859,6 +899,16 @@ class SimdFilterProcessor<
   {
     vst1q_u64(reinterpret_cast<uint64_t*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vbslq_u64(vcgtq_u64(x,y), x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vbslq_u64(vcgtq_u64(y,x), x, y);
+  }
 };
 template <typename VT, typename CHECK_T>
 class SimdFilterProcessor<
@@ -965,6 +1015,16 @@ class SimdFilterProcessor<
   {
     vst1q_s32(reinterpret_cast<int32_t*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vminq_s32(x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vmaxq_s32(x, y);
+  }
 };
 
 template <typename VT, typename CHECK_T>
@@ -1072,6 +1132,16 @@ class SimdFilterProcessor<
   {
     vst1q_u32(reinterpret_cast<uint32_t*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vminq_u32(x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vmaxq_u32(x, y);
+  }
 };
 
 template <typename VT, typename CHECK_T>
@@ -1178,6 +1248,16 @@ class SimdFilterProcessor<
   {
     vst1q_s16(reinterpret_cast<int16_t*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vminq_s16(x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vmaxq_s16(x, y);
+  }
 };
 
 template <typename VT, typename CHECK_T>
@@ -1284,6 +1364,16 @@ class SimdFilterProcessor<VT, CHECK_T,
   {
     vst1q_u16(reinterpret_cast<uint16_t*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vminq_u16(x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vmaxq_u16(x, y);
+  }
 };
 
 template <typename VT, typename CHECK_T>
@@ -1389,6 +1479,16 @@ class SimdFilterProcessor<
   {
     vst1q_s8(reinterpret_cast<int8_t*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vminq_s8(x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vmaxq_s8(x, y);
+  }
 };
 
 template <typename VT, typename CHECK_T>
@@ -1496,6 +1596,16 @@ class SimdFilterProcessor<
   {
     vst1q_u8(reinterpret_cast<uint8_t*>(dst), x);
   }
+
+  MCS_FORCE_INLINE SimdType min(SimdType& x, SimdType& y)
+  {
+    return vminq_u8(x, y);
+  }
+
+  MCS_FORCE_INLINE SimdType max(SimdType& x, SimdType& y)
+  {
+    return vmaxq_u8(x, y);
+  }
 };
 
 };      // namespace simd