Use smart pivot in key-value

Author: Raghuveer Devulapalli

src/avx512-64bit-common.h

@@ -34,6 +34,7 @@ struct ymm_vector<float> {
     using opmask_t = __mmask8;
     static const uint8_t numlanes = 8;
     static constexpr simd_type vec_type = simd_type::AVX512;
+    using swizzle_ops = avx512_ymm_64bit_swizzle_ops;
 
     using swizzle_ops = avx512_ymm_64bit_swizzle_ops;
 
@@ -210,6 +211,9 @@ struct ymm_vector<float> {
     {
         return _mm256_castps_si256(v);
     }
+    static bool all_false(opmask_t k){
+        return k == 0;
+    }
     static reg_t reverse(reg_t ymm)
     {
         const __m256i rev_index = _mm256_set_epi32(NETWORK_32BIT_AVX2_2);
@@ -232,6 +236,7 @@ struct ymm_vector<uint32_t> {
     using opmask_t = __mmask8;
     static const uint8_t numlanes = 8;
     static constexpr simd_type vec_type = simd_type::AVX512;
+    using swizzle_ops = avx512_ymm_64bit_swizzle_ops;
 
     using swizzle_ops = avx512_ymm_64bit_swizzle_ops;
 
@@ -394,6 +399,9 @@ struct ymm_vector<uint32_t> {
     {
         return v;
     }
+    static bool all_false(opmask_t k){
+        return k == 0;
+    }
     static reg_t reverse(reg_t ymm)
     {
         const __m256i rev_index = _mm256_set_epi32(NETWORK_32BIT_AVX2_2);
@@ -416,6 +424,7 @@ struct ymm_vector<int32_t> {
     using opmask_t = __mmask8;
     static const uint8_t numlanes = 8;
     static constexpr simd_type vec_type = simd_type::AVX512;
+    using swizzle_ops = avx512_ymm_64bit_swizzle_ops;
 
     using swizzle_ops = avx512_ymm_64bit_swizzle_ops;
 
@@ -578,6 +587,9 @@ struct ymm_vector<int32_t> {
     {
         return v;
     }
+    static bool all_false(opmask_t k){
+        return k == 0;
+    }
     static reg_t reverse(reg_t ymm)
     {
         const __m256i rev_index = _mm256_set_epi32(NETWORK_32BIT_AVX2_2);

src/xss-common-keyvaluesort.hpp

@@ -392,7 +392,10 @@ X86_SIMD_SORT_INLINE void kvsort_(type1_t *keys,
         return;
     }
 
-    type1_t pivot = get_pivot_blocks<vtype1>(keys, left, right);
+    type1_t pivot;
+    auto pivot_result = get_pivot_smart<vtype1, type1_t>(keys, left, right);
+    pivot = pivot_result.pivot;
+
     type1_t smallest = vtype1::type_max();
     type1_t biggest = vtype1::type_min();
     arrsize_t pivot_index = kvpartition_unrolled<vtype1, vtype2, 4>(