non-SIMD fallback

sql/bloom_filters.h

@@ -30,6 +30,13 @@ SOFTWARE.
 #include <algorithm>
 #ifdef HAVE_IMMINTRIN_H
 #include <immintrin.h>
+#if __GNUC__ > 7 && defined __x86_64__
+#define DEFAULT_IMPLEMENTATION    __attribute__ ((target ("default")))
+#define AVX2_IMPLEMENTATION __attribute__ ((target ("avx2,avx,fma")))
+#endif
+#endif
+#ifndef DEFAULT_IMPLEMENTATION
+#define DEFAULT_IMPLEMENTATION
 #endif
 
 template <typename T>
@@ -49,8 +56,8 @@ struct PatternedSimdBloomFilter
     return std::max<uint32_t>(512, static_cast<uint32_t>(bits_per_val * n + 0.5));
   }
 
-#if __GNUC__ > 7 && defined(HAVE_IMMINTRIN_H)
-  __attribute__ ((target ("avx2,avx,fma")))
+#ifdef AVX2_IMPLEMENTATION
+  AVX2_IMPLEMENTATION
   __m256i CalcHash(__m256i vecData)
   {
     // (almost) xxHash parallel version, 64bit input, 64bit output, seed=0
@@ -76,7 +83,7 @@ struct PatternedSimdBloomFilter
     return _mm256_xor_si256(step9, _mm256_srli_epi64(step9, 28));
   }
 
-  __attribute__ ((target ("avx2,avx,fma")))
+  AVX2_IMPLEMENTATION
   __m256i GetBlockIdx(__m256i vecHash)
   {
     __m256i vecNumBlocksMask = _mm256_set1_epi64x(num_blocks - 1);
@@ -84,7 +91,7 @@ struct PatternedSimdBloomFilter
     return _mm256_and_si256(vecBlockIdx, vecNumBlocksMask);
   }
 
-  __attribute__ ((target ("avx2,avx,fma")))
+  AVX2_IMPLEMENTATION
   __m256i ConstructMask(__m256i vecHash)
   {
     __m256i vecMaskIdxMask = _mm256_set1_epi64x((1 << mask_idx_bits) - 1);
@@ -103,7 +110,7 @@ struct PatternedSimdBloomFilter
     return _mm256_or_si256(vecShiftDown, vecShiftUp);
   }
 
-  __attribute__ ((target ("avx2,avx,fma")))
+  AVX2_IMPLEMENTATION
   void Insert(const T **data)
   {
     __m256i vecDataA = _mm256_loadu_si256(reinterpret_cast<__m256i *>(data + 0));
@@ -137,7 +144,7 @@ struct PatternedSimdBloomFilter
     bv[block7] |= _mm256_extract_epi64(vecMaskB, 3);
   }
 
-  __attribute__ ((target ("avx2,avx,fma")))
+  AVX2_IMPLEMENTATION
   uint8_t Query(T **data)
   {
     __m256i vecDataA = _mm256_loadu_si256(reinterpret_cast<__m256i *>(data + 0));
@@ -164,6 +171,70 @@ struct PatternedSimdBloomFilter
   }
 #endif
 
+  /********************************************************
+  ********* non-SIMD fallback version ********************/
+
+  uint64_t CalcHash_1(const T* data)
+  {
+    static constexpr uint64_t prime_mx2= 0x9FB21C651E98DF25ULL;
+    static constexpr uint64_t bitflip= 0xC73AB174C5ECD5A2ULL;
+    uint64_t step1= ((intptr)data) ^ bitflip;
+    uint64_t step2= (step1 >> 48) ^ (step1 << 16);
+    uint64_t step3= (step1 >> 24) ^ (step1 << 40);
+    uint64_t step4= step1 ^ step2 ^ step3;
+    uint64_t step5= step4 * prime_mx2;
+    uint64_t step6= step5 >> 35;
+    uint64_t step7= step6 + 8;
+    uint64_t step8= step5 ^ step7;
+    uint64_t step9= step8 * prime_mx2;
+    return step9 ^ (step9 >> 28);
+  }
+
+  uint64_t GetBlockIdx_1(uint64_t hash)
+  {
+    uint64_t blockIdx = hash >> (mask_idx_bits + rotate_bits);
+    return blockIdx & (num_blocks - 1);
+  }
+
+  uint64_t ConstructMask_1(uint64_t hash)
+  {
+    uint64_t maskIdxMask = (1 << mask_idx_bits) - 1;
+    uint64_t maskMask = (1ULL << bits_per_mask) - 1;
+    uint64_t maskIdx = hash & maskIdxMask;
+    uint64_t maskByteIdx = maskIdx >> 3;
+    uint64_t maskBitIdx = maskIdx & 7;
+    uint64_t rawMask = *(uint64_t *)(masks + maskByteIdx);
+    uint64_t unrotated = (rawMask >> maskBitIdx) & maskMask;
+    uint64_t rotation = (hash >> mask_idx_bits) & ((1 << rotate_bits) - 1);
+    return rotation ? (unrotated << rotation) | (unrotated >> (64 - rotation))
+                    : unrotated;
+  }
+
+  DEFAULT_IMPLEMENTATION
+  void Insert(const T **data)
+  {
+    for (size_t i = 0; i < 8; i++)
+    {
+      uint64_t hash = CalcHash_1(data[i]);
+      uint64_t mask = ConstructMask_1(hash);
+      bv[GetBlockIdx_1(hash)] |= mask;
+    }
+  }
+
+  DEFAULT_IMPLEMENTATION
+  uint8_t Query(T **data)
+  {
+    uint8_t res_bits = 0;
+    for (size_t i = 0; i < 8; i++)
+    {
+      uint64_t hash = CalcHash_1(data[i]);
+      uint64_t mask = ConstructMask_1(hash);
+      if ((bv[GetBlockIdx_1(hash)] & mask) == mask)
+        res_bits |= 1 << i;
+    }
+    return res_bits;
+  }
+
   int n;
   float epsilon;
 

sql/vector_mhnsw.cc

@@ -85,6 +85,30 @@ struct Neighborhood: public Sql_alloc
 };
 
 
+#ifdef AVX2_IMPLEMENTATION
+AVX2_IMPLEMENTATION
+float vec_distance(float *v1, float *v2, size_t len)
+{
+  typedef float v8f __attribute__((vector_size(SIMD_word)));
+  v8f *p1= (v8f*)v1;
+  v8f *p2= (v8f*)v2;
+  v8f d= {0};
+  for (size_t i= 0; i < len/SIMD_floats; p1++, p2++, i++)
+  {
+    v8f dist= *p1 - *p2;
+    d+= dist * dist;
+  }
+  return d[0] + d[1] + d[2] + d[3] + d[4] + d[5] + d[6] + d[7];
+}
+#endif
+
+DEFAULT_IMPLEMENTATION
+float vec_distance(float *v1, float *v2, size_t len)
+{
+  return euclidean_vec_distance(v1, v2, len);
+}
+
+
 /*
   One node in a graph = one row in the graph table
 
@@ -526,20 +550,7 @@ FVectorNode::FVectorNode(MHNSW_Context *ctx_, const void *tref_, uint8_t layer,
 
 float FVectorNode::distance_to(const FVector &other) const
 {
-#if __GNUC__ > 7
-  typedef float v8f __attribute__((vector_size(SIMD_word)));
-  v8f *p1= (v8f*)vec;
-  v8f *p2= (v8f*)other.vec;
-  v8f d= {0,0,0,0,0,0,0,0};
-  for (size_t i= 0; i < ctx->vec_len/SIMD_floats; p1++, p2++, i++)
-  {
-    v8f dist= *p1 - *p2;
-    d+= dist * dist;
-  }
-  return d[0] + d[1] + d[2] + d[3] + d[4] + d[5] + d[6] + d[7];
-#else
-  return euclidean_vec_distance(vec, other.vec, ctx->vec_len);
-#endif
+  return vec_distance(vec, other.vec, ctx->vec_len);
 }
 
 int FVectorNode::alloc_neighborhood(uint8_t layer)