add avx512 support for mergetreereader

src/Common/TargetSpecific.cpp

@@ -16,6 +16,8 @@ UInt32 getSupportedArchs()
         result |= static_cast<UInt32>(TargetArch::AVX2);
     if (Cpu::CpuFlagsCache::have_AVX512F)
         result |= static_cast<UInt32>(TargetArch::AVX512F);
+    if (Cpu::CpuFlagsCache::have_AVX512BW)
+        result |= static_cast<UInt32>(TargetArch::AVX512BW);        
     return result;
 }
 
@@ -34,6 +36,7 @@ String toString(TargetArch arch)
         case TargetArch::AVX:     return "avx";
         case TargetArch::AVX2:    return "avx2";
         case TargetArch::AVX512F: return "avx512f";
+        case TargetArch::AVX512BW: return "avx512bw";
     }
 
     __builtin_unreachable();

src/Common/TargetSpecific.h

@@ -80,6 +80,7 @@ enum class TargetArch : UInt32
     AVX      = (1 << 1),
     AVX2     = (1 << 2),
     AVX512F  = (1 << 3),
+    AVX512BW  = (1 << 4),
 };
 
 /// Runtime detection.

src/Storages/MergeTree/MergeTreeRangeReader.cpp

@@ -2,6 +2,7 @@
 #include <Columns/FilterDescription.h>
 #include <Columns/ColumnConst.h>
 #include <Columns/ColumnsCommon.h>
+#include <Common/TargetSpecific.h>
 #include <base/range.h>
 #include <Interpreters/castColumn.h>
 #include <DataTypes/DataTypeNothing.h>
@@ -10,6 +11,7 @@
 #include <emmintrin.h>
 #endif
 
+
 namespace DB
 {
 namespace ErrorCodes
@@ -449,6 +451,33 @@ size_t MergeTreeRangeReader::ReadResult::numZerosInTail(const UInt8 * begin, con
 {
     size_t count = 0;
 
+#if defined(__AVX512F__) && defined(__AVX512BW__) /// check if avx512 instructions are compiled
+    if (isArchSupported(TargetArch::AVX512BW)) {
+        /// check if cpu support avx512 dynamically, haveAVX512BW contains check of haveAVX512F
+        const __m512i zero64 = _mm512_setzero_epi32();
+        while (end - begin >= 64)
+        {
+            end -= 64;
+            const auto * pos = end;
+            UInt64 val = static_cast<UInt64>(_mm512_cmp_epi8_mask(_mm512_loadu_si512(reinterpret_cast<const __m512i *>(pos)), zero64, _MM_CMPINT_EQ));
+            val = ~val;
+            if (val == 0) 
+            {
+                count += 64;
+            } else 
+            {
+                count += __builtin_clzll(val);
+                return count;
+            }
+        }
+        while (end > begin && *(--end) == 0)
+        {
+            ++count;
+        }
+        return count;
+    }
+#endif
+
 #if defined(__SSE2__) && defined(__POPCNT__)
     const __m128i zero16 = _mm_setzero_si128();
     while (end - begin >= 64)