faiss 0.3.1 → 0.3.3

data/vendor/faiss/faiss/utils/distances_simd.cpp

@@ -1,5 +1,5 @@
-/**
- * Copyright (c) Facebook, Inc. and its affiliates.
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
  *
  * This source code is licensed under the MIT license found in the
  * LICENSE file in the root directory of this source tree.
@@ -23,10 +23,16 @@
 #include <immintrin.h>
 #endif
 
-#ifdef __AVX2__
+#if defined(__AVX512F__)
+#include <faiss/utils/transpose/transpose-avx512-inl.h>
+#elif defined(__AVX2__)
 #include <faiss/utils/transpose/transpose-avx2-inl.h>
 #endif
 
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif
+
 #ifdef __aarch64__
 #include <arm_neon.h>
 #endif
@@ -346,6 +352,14 @@ inline float horizontal_sum(const __m256 v) {
 }
 #endif
 
+#ifdef __AVX512F__
+/// helper function for AVX512
+inline float horizontal_sum(const __m512 v) {
+    // performs better than adding the high and low parts
+    return _mm512_reduce_add_ps(v);
+}
+#endif
+
 /// Function that does a component-wise operation between x and y
 /// to compute L2 distances. ElementOp can then be used in the fvec_op_ny
 /// functions below
@@ -366,6 +380,13 @@ struct ElementOpL2 {
         return _mm256_mul_ps(tmp, tmp);
     }
 #endif
+
+#ifdef __AVX512F__
+    static __m512 op(__m512 x, __m512 y) {
+        __m512 tmp = _mm512_sub_ps(x, y);
+        return _mm512_mul_ps(tmp, tmp);
+    }
+#endif
 };
 
 /// Function that does a component-wise operation between x and y
@@ -384,6 +405,12 @@ struct ElementOpIP {
         return _mm256_mul_ps(x, y);
     }
 #endif
+
+#ifdef __AVX512F__
+    static __m512 op(__m512 x, __m512 y) {
+        return _mm512_mul_ps(x, y);
+    }
+#endif
 };
 
 template <class ElementOp>
@@ -426,7 +453,130 @@ void fvec_op_ny_D2(float* dis, const float* x, const float* y, size_t ny) {
     }
 }
 
-#ifdef __AVX2__
+#if defined(__AVX512F__)
+
+template <>
+void fvec_op_ny_D2<ElementOpIP>(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t ny16 = ny / 16;
+    size_t i = 0;
+
+    if (ny16 > 0) {
+        // process 16 D2-vectors per loop.
+        _mm_prefetch((const char*)y, _MM_HINT_T0);
+        _mm_prefetch((const char*)(y + 32), _MM_HINT_T0);
+
+        const __m512 m0 = _mm512_set1_ps(x[0]);
+        const __m512 m1 = _mm512_set1_ps(x[1]);
+
+        for (i = 0; i < ny16 * 16; i += 16) {
+            _mm_prefetch((const char*)(y + 64), _MM_HINT_T0);
+
+            // load 16x2 matrix and transpose it in registers.
+            // the typical bottleneck is memory access, so
+            // let's trade instructions for the bandwidth.
+
+            __m512 v0;
+            __m512 v1;
+
+            transpose_16x2(
+                    _mm512_loadu_ps(y + 0 * 16),
+                    _mm512_loadu_ps(y + 1 * 16),
+                    v0,
+                    v1);
+
+            // compute distances (dot product)
+            __m512 distances = _mm512_mul_ps(m0, v0);
+            distances = _mm512_fmadd_ps(m1, v1, distances);
+
+            // store
+            _mm512_storeu_ps(dis + i, distances);
+
+            y += 32; // move to the next set of 16x2 elements
+        }
+    }
+
+    if (i < ny) {
+        // process leftovers
+        float x0 = x[0];
+        float x1 = x[1];
+
+        for (; i < ny; i++) {
+            float distance = x0 * y[0] + x1 * y[1];
+            y += 2;
+            dis[i] = distance;
+        }
+    }
+}
+
+template <>
+void fvec_op_ny_D2<ElementOpL2>(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t ny16 = ny / 16;
+    size_t i = 0;
+
+    if (ny16 > 0) {
+        // process 16 D2-vectors per loop.
+        _mm_prefetch((const char*)y, _MM_HINT_T0);
+        _mm_prefetch((const char*)(y + 32), _MM_HINT_T0);
+
+        const __m512 m0 = _mm512_set1_ps(x[0]);
+        const __m512 m1 = _mm512_set1_ps(x[1]);
+
+        for (i = 0; i < ny16 * 16; i += 16) {
+            _mm_prefetch((const char*)(y + 64), _MM_HINT_T0);
+
+            // load 16x2 matrix and transpose it in registers.
+            // the typical bottleneck is memory access, so
+            // let's trade instructions for the bandwidth.
+
+            __m512 v0;
+            __m512 v1;
+
+            transpose_16x2(
+                    _mm512_loadu_ps(y + 0 * 16),
+                    _mm512_loadu_ps(y + 1 * 16),
+                    v0,
+                    v1);
+
+            // compute differences
+            const __m512 d0 = _mm512_sub_ps(m0, v0);
+            const __m512 d1 = _mm512_sub_ps(m1, v1);
+
+            // compute squares of differences
+            __m512 distances = _mm512_mul_ps(d0, d0);
+            distances = _mm512_fmadd_ps(d1, d1, distances);
+
+            // store
+            _mm512_storeu_ps(dis + i, distances);
+
+            y += 32; // move to the next set of 16x2 elements
+        }
+    }
+
+    if (i < ny) {
+        // process leftovers
+        float x0 = x[0];
+        float x1 = x[1];
+
+        for (; i < ny; i++) {
+            float sub0 = x0 - y[0];
+            float sub1 = x1 - y[1];
+            float distance = sub0 * sub0 + sub1 * sub1;
+
+            y += 2;
+            dis[i] = distance;
+        }
+    }
+}
+
+#elif defined(__AVX2__)
 
 template <>
 void fvec_op_ny_D2<ElementOpIP>(
@@ -562,7 +712,137 @@ void fvec_op_ny_D4(float* dis, const float* x, const float* y, size_t ny) {
     }
 }
 
-#ifdef __AVX2__
+#if defined(__AVX512F__)
+
+template <>
+void fvec_op_ny_D4<ElementOpIP>(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t ny16 = ny / 16;
+    size_t i = 0;
+
+    if (ny16 > 0) {
+        // process 16 D4-vectors per loop.
+        const __m512 m0 = _mm512_set1_ps(x[0]);
+        const __m512 m1 = _mm512_set1_ps(x[1]);
+        const __m512 m2 = _mm512_set1_ps(x[2]);
+        const __m512 m3 = _mm512_set1_ps(x[3]);
+
+        for (i = 0; i < ny16 * 16; i += 16) {
+            // load 16x4 matrix and transpose it in registers.
+            // the typical bottleneck is memory access, so
+            // let's trade instructions for the bandwidth.
+
+            __m512 v0;
+            __m512 v1;
+            __m512 v2;
+            __m512 v3;
+
+            transpose_16x4(
+                    _mm512_loadu_ps(y + 0 * 16),
+                    _mm512_loadu_ps(y + 1 * 16),
+                    _mm512_loadu_ps(y + 2 * 16),
+                    _mm512_loadu_ps(y + 3 * 16),
+                    v0,
+                    v1,
+                    v2,
+                    v3);
+
+            // compute distances
+            __m512 distances = _mm512_mul_ps(m0, v0);
+            distances = _mm512_fmadd_ps(m1, v1, distances);
+            distances = _mm512_fmadd_ps(m2, v2, distances);
+            distances = _mm512_fmadd_ps(m3, v3, distances);
+
+            // store
+            _mm512_storeu_ps(dis + i, distances);
+
+            y += 64; // move to the next set of 16x4 elements
+        }
+    }
+
+    if (i < ny) {
+        // process leftovers
+        __m128 x0 = _mm_loadu_ps(x);
+
+        for (; i < ny; i++) {
+            __m128 accu = ElementOpIP::op(x0, _mm_loadu_ps(y));
+            y += 4;
+            dis[i] = horizontal_sum(accu);
+        }
+    }
+}
+
+template <>
+void fvec_op_ny_D4<ElementOpL2>(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t ny16 = ny / 16;
+    size_t i = 0;
+
+    if (ny16 > 0) {
+        // process 16 D4-vectors per loop.
+        const __m512 m0 = _mm512_set1_ps(x[0]);
+        const __m512 m1 = _mm512_set1_ps(x[1]);
+        const __m512 m2 = _mm512_set1_ps(x[2]);
+        const __m512 m3 = _mm512_set1_ps(x[3]);
+
+        for (i = 0; i < ny16 * 16; i += 16) {
+            // load 16x4 matrix and transpose it in registers.
+            // the typical bottleneck is memory access, so
+            // let's trade instructions for the bandwidth.
+
+            __m512 v0;
+            __m512 v1;
+            __m512 v2;
+            __m512 v3;
+
+            transpose_16x4(
+                    _mm512_loadu_ps(y + 0 * 16),
+                    _mm512_loadu_ps(y + 1 * 16),
+                    _mm512_loadu_ps(y + 2 * 16),
+                    _mm512_loadu_ps(y + 3 * 16),
+                    v0,
+                    v1,
+                    v2,
+                    v3);
+
+            // compute differences
+            const __m512 d0 = _mm512_sub_ps(m0, v0);
+            const __m512 d1 = _mm512_sub_ps(m1, v1);
+            const __m512 d2 = _mm512_sub_ps(m2, v2);
+            const __m512 d3 = _mm512_sub_ps(m3, v3);
+
+            // compute squares of differences
+            __m512 distances = _mm512_mul_ps(d0, d0);
+            distances = _mm512_fmadd_ps(d1, d1, distances);
+            distances = _mm512_fmadd_ps(d2, d2, distances);
+            distances = _mm512_fmadd_ps(d3, d3, distances);
+
+            // store
+            _mm512_storeu_ps(dis + i, distances);
+
+            y += 64; // move to the next set of 16x4 elements
+        }
+    }
+
+    if (i < ny) {
+        // process leftovers
+        __m128 x0 = _mm_loadu_ps(x);
+
+        for (; i < ny; i++) {
+            __m128 accu = ElementOpL2::op(x0, _mm_loadu_ps(y));
+            y += 4;
+            dis[i] = horizontal_sum(accu);
+        }
+    }
+}
+
+#elif defined(__AVX2__)
 
 template <>
 void fvec_op_ny_D4<ElementOpIP>(
@@ -710,7 +990,181 @@ void fvec_op_ny_D8(float* dis, const float* x, const float* y, size_t ny) {
     }
 }
 
-#ifdef __AVX2__
+#if defined(__AVX512F__)
+
+template <>
+void fvec_op_ny_D8<ElementOpIP>(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t ny16 = ny / 16;
+    size_t i = 0;
+
+    if (ny16 > 0) {
+        // process 16 D16-vectors per loop.
+        const __m512 m0 = _mm512_set1_ps(x[0]);
+        const __m512 m1 = _mm512_set1_ps(x[1]);
+        const __m512 m2 = _mm512_set1_ps(x[2]);
+        const __m512 m3 = _mm512_set1_ps(x[3]);
+        const __m512 m4 = _mm512_set1_ps(x[4]);
+        const __m512 m5 = _mm512_set1_ps(x[5]);
+        const __m512 m6 = _mm512_set1_ps(x[6]);
+        const __m512 m7 = _mm512_set1_ps(x[7]);
+
+        for (i = 0; i < ny16 * 16; i += 16) {
+            // load 16x8 matrix and transpose it in registers.
+            // the typical bottleneck is memory access, so
+            // let's trade instructions for the bandwidth.
+
+            __m512 v0;
+            __m512 v1;
+            __m512 v2;
+            __m512 v3;
+            __m512 v4;
+            __m512 v5;
+            __m512 v6;
+            __m512 v7;
+
+            transpose_16x8(
+                    _mm512_loadu_ps(y + 0 * 16),
+                    _mm512_loadu_ps(y + 1 * 16),
+                    _mm512_loadu_ps(y + 2 * 16),
+                    _mm512_loadu_ps(y + 3 * 16),
+                    _mm512_loadu_ps(y + 4 * 16),
+                    _mm512_loadu_ps(y + 5 * 16),
+                    _mm512_loadu_ps(y + 6 * 16),
+                    _mm512_loadu_ps(y + 7 * 16),
+                    v0,
+                    v1,
+                    v2,
+                    v3,
+                    v4,
+                    v5,
+                    v6,
+                    v7);
+
+            // compute distances
+            __m512 distances = _mm512_mul_ps(m0, v0);
+            distances = _mm512_fmadd_ps(m1, v1, distances);
+            distances = _mm512_fmadd_ps(m2, v2, distances);
+            distances = _mm512_fmadd_ps(m3, v3, distances);
+            distances = _mm512_fmadd_ps(m4, v4, distances);
+            distances = _mm512_fmadd_ps(m5, v5, distances);
+            distances = _mm512_fmadd_ps(m6, v6, distances);
+            distances = _mm512_fmadd_ps(m7, v7, distances);
+
+            // store
+            _mm512_storeu_ps(dis + i, distances);
+
+            y += 128; // 16 floats * 8 rows
+        }
+    }
+
+    if (i < ny) {
+        // process leftovers
+        __m256 x0 = _mm256_loadu_ps(x);
+
+        for (; i < ny; i++) {
+            __m256 accu = ElementOpIP::op(x0, _mm256_loadu_ps(y));
+            y += 8;
+            dis[i] = horizontal_sum(accu);
+        }
+    }
+}
+
+template <>
+void fvec_op_ny_D8<ElementOpL2>(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t ny16 = ny / 16;
+    size_t i = 0;
+
+    if (ny16 > 0) {
+        // process 16 D16-vectors per loop.
+        const __m512 m0 = _mm512_set1_ps(x[0]);
+        const __m512 m1 = _mm512_set1_ps(x[1]);
+        const __m512 m2 = _mm512_set1_ps(x[2]);
+        const __m512 m3 = _mm512_set1_ps(x[3]);
+        const __m512 m4 = _mm512_set1_ps(x[4]);
+        const __m512 m5 = _mm512_set1_ps(x[5]);
+        const __m512 m6 = _mm512_set1_ps(x[6]);
+        const __m512 m7 = _mm512_set1_ps(x[7]);
+
+        for (i = 0; i < ny16 * 16; i += 16) {
+            // load 16x8 matrix and transpose it in registers.
+            // the typical bottleneck is memory access, so
+            // let's trade instructions for the bandwidth.
+
+            __m512 v0;
+            __m512 v1;
+            __m512 v2;
+            __m512 v3;
+            __m512 v4;
+            __m512 v5;
+            __m512 v6;
+            __m512 v7;
+
+            transpose_16x8(
+                    _mm512_loadu_ps(y + 0 * 16),
+                    _mm512_loadu_ps(y + 1 * 16),
+                    _mm512_loadu_ps(y + 2 * 16),
+                    _mm512_loadu_ps(y + 3 * 16),
+                    _mm512_loadu_ps(y + 4 * 16),
+                    _mm512_loadu_ps(y + 5 * 16),
+                    _mm512_loadu_ps(y + 6 * 16),
+                    _mm512_loadu_ps(y + 7 * 16),
+                    v0,
+                    v1,
+                    v2,
+                    v3,
+                    v4,
+                    v5,
+                    v6,
+                    v7);
+
+            // compute differences
+            const __m512 d0 = _mm512_sub_ps(m0, v0);
+            const __m512 d1 = _mm512_sub_ps(m1, v1);
+            const __m512 d2 = _mm512_sub_ps(m2, v2);
+            const __m512 d3 = _mm512_sub_ps(m3, v3);
+            const __m512 d4 = _mm512_sub_ps(m4, v4);
+            const __m512 d5 = _mm512_sub_ps(m5, v5);
+            const __m512 d6 = _mm512_sub_ps(m6, v6);
+            const __m512 d7 = _mm512_sub_ps(m7, v7);
+
+            // compute squares of differences
+            __m512 distances = _mm512_mul_ps(d0, d0);
+            distances = _mm512_fmadd_ps(d1, d1, distances);
+            distances = _mm512_fmadd_ps(d2, d2, distances);
+            distances = _mm512_fmadd_ps(d3, d3, distances);
+            distances = _mm512_fmadd_ps(d4, d4, distances);
+            distances = _mm512_fmadd_ps(d5, d5, distances);
+            distances = _mm512_fmadd_ps(d6, d6, distances);
+            distances = _mm512_fmadd_ps(d7, d7, distances);
+
+            // store
+            _mm512_storeu_ps(dis + i, distances);
+
+            y += 128; // 16 floats * 8 rows
+        }
+    }
+
+    if (i < ny) {
+        // process leftovers
+        __m256 x0 = _mm256_loadu_ps(x);
+
+        for (; i < ny; i++) {
+            __m256 accu = ElementOpL2::op(x0, _mm256_loadu_ps(y));
+            y += 8;
+            dis[i] = horizontal_sum(accu);
+        }
+    }
+}
+
+#elif defined(__AVX2__)
 
 template <>
 void fvec_op_ny_D8<ElementOpIP>(
@@ -955,7 +1409,83 @@ void fvec_inner_products_ny(
 #undef DISPATCH
 }
 
-#ifdef __AVX2__
+#if defined(__AVX512F__)
+
+template <size_t DIM>
+void fvec_L2sqr_ny_y_transposed_D(
+        float* distances,
+        const float* x,
+        const float* y,
+        const float* y_sqlen,
+        const size_t d_offset,
+        size_t ny) {
+    // current index being processed
+    size_t i = 0;
+
+    // squared length of x
+    float x_sqlen = 0;
+    for (size_t j = 0; j < DIM; j++) {
+        x_sqlen += x[j] * x[j];
+    }
+
+    // process 16 vectors per loop
+    const size_t ny16 = ny / 16;
+
+    if (ny16 > 0) {
+        // m[i] = (2 * x[i], ... 2 * x[i])
+        __m512 m[DIM];
+        for (size_t j = 0; j < DIM; j++) {
+            m[j] = _mm512_set1_ps(x[j]);
+            m[j] = _mm512_add_ps(m[j], m[j]); // m[j] = 2 * x[j]
+        }
+
+        __m512 x_sqlen_ymm = _mm512_set1_ps(x_sqlen);
+
+        for (; i < ny16 * 16; i += 16) {
+            // Load vectors for 16 dimensions
+            __m512 v[DIM];
+            for (size_t j = 0; j < DIM; j++) {
+                v[j] = _mm512_loadu_ps(y + j * d_offset);
+            }
+
+            // Compute dot products
+            __m512 dp = _mm512_fnmadd_ps(m[0], v[0], x_sqlen_ymm);
+            for (size_t j = 1; j < DIM; j++) {
+                dp = _mm512_fnmadd_ps(m[j], v[j], dp);
+            }
+
+            // Compute y^2 - (2 * x, y) + x^2
+            __m512 distances_v = _mm512_add_ps(_mm512_loadu_ps(y_sqlen), dp);
+
+            _mm512_storeu_ps(distances + i, distances_v);
+
+            // Scroll y and y_sqlen forward
+            y += 16;
+            y_sqlen += 16;
+        }
+    }
+
+    if (i < ny) {
+        // Process leftovers
+        for (; i < ny; i++) {
+            float dp = 0;
+            for (size_t j = 0; j < DIM; j++) {
+                dp += x[j] * y[j * d_offset];
+            }
+
+            // Compute y^2 - 2 * (x, y), which is sufficient for looking for the
+            // lowest distance.
+            const float distance = y_sqlen[0] - 2 * dp + x_sqlen;
+            distances[i] = distance;
+
+            y += 1;
+            y_sqlen += 1;
+        }
+    }
+}
+
+#elif defined(__AVX2__)
+
 template <size_t DIM>
 void fvec_L2sqr_ny_y_transposed_D(
         float* distances,
@@ -1014,58 +1544,368 @@ void fvec_L2sqr_ny_y_transposed_D(
     }
 
     if (i < ny) {
-        // process leftovers
-        for (; i < ny; i++) {
-            float dp = 0;
-            for (size_t j = 0; j < DIM; j++) {
-                dp += x[j] * y[j * d_offset];
-            }
+        // process leftovers
+        for (; i < ny; i++) {
+            float dp = 0;
+            for (size_t j = 0; j < DIM; j++) {
+                dp += x[j] * y[j * d_offset];
+            }
+
+            // compute y^2 - 2 * (x, y), which is sufficient for looking for the
+            //   lowest distance.
+            const float distance = y_sqlen[0] - 2 * dp + x_sqlen;
+            distances[i] = distance;
+
+            y += 1;
+            y_sqlen += 1;
+        }
+    }
+}
+
+#endif
+
+void fvec_L2sqr_ny_transposed(
+        float* dis,
+        const float* x,
+        const float* y,
+        const float* y_sqlen,
+        size_t d,
+        size_t d_offset,
+        size_t ny) {
+    // optimized for a few special cases
+
+#ifdef __AVX2__
+#define DISPATCH(dval)                             \
+    case dval:                                     \
+        return fvec_L2sqr_ny_y_transposed_D<dval>( \
+                dis, x, y, y_sqlen, d_offset, ny);
+
+    switch (d) {
+        DISPATCH(1)
+        DISPATCH(2)
+        DISPATCH(4)
+        DISPATCH(8)
+        default:
+            return fvec_L2sqr_ny_y_transposed_ref(
+                    dis, x, y, y_sqlen, d, d_offset, ny);
+    }
+#undef DISPATCH
+#else
+    // non-AVX2 case
+    return fvec_L2sqr_ny_y_transposed_ref(dis, x, y, y_sqlen, d, d_offset, ny);
+#endif
+}
+
+#if defined(__AVX512F__)
+
+size_t fvec_L2sqr_ny_nearest_D2(
+        float* distances_tmp_buffer,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    // this implementation does not use distances_tmp_buffer.
+
+    size_t i = 0;
+    float current_min_distance = HUGE_VALF;
+    size_t current_min_index = 0;
+
+    const size_t ny16 = ny / 16;
+    if (ny16 > 0) {
+        _mm_prefetch((const char*)y, _MM_HINT_T0);
+        _mm_prefetch((const char*)(y + 32), _MM_HINT_T0);
+
+        __m512 min_distances = _mm512_set1_ps(HUGE_VALF);
+        __m512i min_indices = _mm512_set1_epi32(0);
+
+        __m512i current_indices = _mm512_setr_epi32(
+                0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+        const __m512i indices_increment = _mm512_set1_epi32(16);
+
+        const __m512 m0 = _mm512_set1_ps(x[0]);
+        const __m512 m1 = _mm512_set1_ps(x[1]);
+
+        for (; i < ny16 * 16; i += 16) {
+            _mm_prefetch((const char*)(y + 64), _MM_HINT_T0);
+
+            __m512 v0;
+            __m512 v1;
+
+            transpose_16x2(
+                    _mm512_loadu_ps(y + 0 * 16),
+                    _mm512_loadu_ps(y + 1 * 16),
+                    v0,
+                    v1);
+
+            const __m512 d0 = _mm512_sub_ps(m0, v0);
+            const __m512 d1 = _mm512_sub_ps(m1, v1);
+
+            __m512 distances = _mm512_mul_ps(d0, d0);
+            distances = _mm512_fmadd_ps(d1, d1, distances);
+
+            __mmask16 comparison =
+                    _mm512_cmp_ps_mask(distances, min_distances, _CMP_LT_OS);
+
+            min_distances = _mm512_min_ps(distances, min_distances);
+            min_indices = _mm512_mask_blend_epi32(
+                    comparison, min_indices, current_indices);
+
+            current_indices =
+                    _mm512_add_epi32(current_indices, indices_increment);
+
+            y += 32;
+        }
+
+        alignas(64) float min_distances_scalar[16];
+        alignas(64) uint32_t min_indices_scalar[16];
+        _mm512_store_ps(min_distances_scalar, min_distances);
+        _mm512_store_epi32(min_indices_scalar, min_indices);
+
+        for (size_t j = 0; j < 16; j++) {
+            if (current_min_distance > min_distances_scalar[j]) {
+                current_min_distance = min_distances_scalar[j];
+                current_min_index = min_indices_scalar[j];
+            }
+        }
+    }
+
+    if (i < ny) {
+        float x0 = x[0];
+        float x1 = x[1];
+
+        for (; i < ny; i++) {
+            float sub0 = x0 - y[0];
+            float sub1 = x1 - y[1];
+            float distance = sub0 * sub0 + sub1 * sub1;
+
+            y += 2;
+
+            if (current_min_distance > distance) {
+                current_min_distance = distance;
+                current_min_index = i;
+            }
+        }
+    }
+
+    return current_min_index;
+}
+
+size_t fvec_L2sqr_ny_nearest_D4(
+        float* distances_tmp_buffer,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    // this implementation does not use distances_tmp_buffer.
+
+    size_t i = 0;
+    float current_min_distance = HUGE_VALF;
+    size_t current_min_index = 0;
+
+    const size_t ny16 = ny / 16;
+
+    if (ny16 > 0) {
+        __m512 min_distances = _mm512_set1_ps(HUGE_VALF);
+        __m512i min_indices = _mm512_set1_epi32(0);
+
+        __m512i current_indices = _mm512_setr_epi32(
+                0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+        const __m512i indices_increment = _mm512_set1_epi32(16);
+
+        const __m512 m0 = _mm512_set1_ps(x[0]);
+        const __m512 m1 = _mm512_set1_ps(x[1]);
+        const __m512 m2 = _mm512_set1_ps(x[2]);
+        const __m512 m3 = _mm512_set1_ps(x[3]);
+
+        for (; i < ny16 * 16; i += 16) {
+            __m512 v0;
+            __m512 v1;
+            __m512 v2;
+            __m512 v3;
+
+            transpose_16x4(
+                    _mm512_loadu_ps(y + 0 * 16),
+                    _mm512_loadu_ps(y + 1 * 16),
+                    _mm512_loadu_ps(y + 2 * 16),
+                    _mm512_loadu_ps(y + 3 * 16),
+                    v0,
+                    v1,
+                    v2,
+                    v3);
+
+            const __m512 d0 = _mm512_sub_ps(m0, v0);
+            const __m512 d1 = _mm512_sub_ps(m1, v1);
+            const __m512 d2 = _mm512_sub_ps(m2, v2);
+            const __m512 d3 = _mm512_sub_ps(m3, v3);
+
+            __m512 distances = _mm512_mul_ps(d0, d0);
+            distances = _mm512_fmadd_ps(d1, d1, distances);
+            distances = _mm512_fmadd_ps(d2, d2, distances);
+            distances = _mm512_fmadd_ps(d3, d3, distances);
+
+            __mmask16 comparison =
+                    _mm512_cmp_ps_mask(distances, min_distances, _CMP_LT_OS);
+
+            min_distances = _mm512_min_ps(distances, min_distances);
+            min_indices = _mm512_mask_blend_epi32(
+                    comparison, min_indices, current_indices);
+
+            current_indices =
+                    _mm512_add_epi32(current_indices, indices_increment);
+
+            y += 64;
+        }
+
+        alignas(64) float min_distances_scalar[16];
+        alignas(64) uint32_t min_indices_scalar[16];
+        _mm512_store_ps(min_distances_scalar, min_distances);
+        _mm512_store_epi32(min_indices_scalar, min_indices);
+
+        for (size_t j = 0; j < 16; j++) {
+            if (current_min_distance > min_distances_scalar[j]) {
+                current_min_distance = min_distances_scalar[j];
+                current_min_index = min_indices_scalar[j];
+            }
+        }
+    }
+
+    if (i < ny) {
+        __m128 x0 = _mm_loadu_ps(x);
+
+        for (; i < ny; i++) {
+            __m128 accu = ElementOpL2::op(x0, _mm_loadu_ps(y));
+            y += 4;
+            const float distance = horizontal_sum(accu);
+
+            if (current_min_distance > distance) {
+                current_min_distance = distance;
+                current_min_index = i;
+            }
+        }
+    }
+
+    return current_min_index;
+}
+
+size_t fvec_L2sqr_ny_nearest_D8(
+        float* distances_tmp_buffer,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    // this implementation does not use distances_tmp_buffer.
+
+    size_t i = 0;
+    float current_min_distance = HUGE_VALF;
+    size_t current_min_index = 0;
+
+    const size_t ny16 = ny / 16;
+    if (ny16 > 0) {
+        __m512 min_distances = _mm512_set1_ps(HUGE_VALF);
+        __m512i min_indices = _mm512_set1_epi32(0);
+
+        __m512i current_indices = _mm512_setr_epi32(
+                0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+        const __m512i indices_increment = _mm512_set1_epi32(16);
+
+        const __m512 m0 = _mm512_set1_ps(x[0]);
+        const __m512 m1 = _mm512_set1_ps(x[1]);
+        const __m512 m2 = _mm512_set1_ps(x[2]);
+        const __m512 m3 = _mm512_set1_ps(x[3]);
+
+        const __m512 m4 = _mm512_set1_ps(x[4]);
+        const __m512 m5 = _mm512_set1_ps(x[5]);
+        const __m512 m6 = _mm512_set1_ps(x[6]);
+        const __m512 m7 = _mm512_set1_ps(x[7]);
+
+        for (; i < ny16 * 16; i += 16) {
+            __m512 v0;
+            __m512 v1;
+            __m512 v2;
+            __m512 v3;
+            __m512 v4;
+            __m512 v5;
+            __m512 v6;
+            __m512 v7;
+
+            transpose_16x8(
+                    _mm512_loadu_ps(y + 0 * 16),
+                    _mm512_loadu_ps(y + 1 * 16),
+                    _mm512_loadu_ps(y + 2 * 16),
+                    _mm512_loadu_ps(y + 3 * 16),
+                    _mm512_loadu_ps(y + 4 * 16),
+                    _mm512_loadu_ps(y + 5 * 16),
+                    _mm512_loadu_ps(y + 6 * 16),
+                    _mm512_loadu_ps(y + 7 * 16),
+                    v0,
+                    v1,
+                    v2,
+                    v3,
+                    v4,
+                    v5,
+                    v6,
+                    v7);
+
+            const __m512 d0 = _mm512_sub_ps(m0, v0);
+            const __m512 d1 = _mm512_sub_ps(m1, v1);
+            const __m512 d2 = _mm512_sub_ps(m2, v2);
+            const __m512 d3 = _mm512_sub_ps(m3, v3);
+            const __m512 d4 = _mm512_sub_ps(m4, v4);
+            const __m512 d5 = _mm512_sub_ps(m5, v5);
+            const __m512 d6 = _mm512_sub_ps(m6, v6);
+            const __m512 d7 = _mm512_sub_ps(m7, v7);
+
+            __m512 distances = _mm512_mul_ps(d0, d0);
+            distances = _mm512_fmadd_ps(d1, d1, distances);
+            distances = _mm512_fmadd_ps(d2, d2, distances);
+            distances = _mm512_fmadd_ps(d3, d3, distances);
+            distances = _mm512_fmadd_ps(d4, d4, distances);
+            distances = _mm512_fmadd_ps(d5, d5, distances);
+            distances = _mm512_fmadd_ps(d6, d6, distances);
+            distances = _mm512_fmadd_ps(d7, d7, distances);
+
+            __mmask16 comparison =
+                    _mm512_cmp_ps_mask(distances, min_distances, _CMP_LT_OS);
+
+            min_distances = _mm512_min_ps(distances, min_distances);
+            min_indices = _mm512_mask_blend_epi32(
+                    comparison, min_indices, current_indices);
+
+            current_indices =
+                    _mm512_add_epi32(current_indices, indices_increment);
+
+            y += 128;
+        }
+
+        alignas(64) float min_distances_scalar[16];
+        alignas(64) uint32_t min_indices_scalar[16];
+        _mm512_store_ps(min_distances_scalar, min_distances);
+        _mm512_store_epi32(min_indices_scalar, min_indices);
+
+        for (size_t j = 0; j < 16; j++) {
+            if (current_min_distance > min_distances_scalar[j]) {
+                current_min_distance = min_distances_scalar[j];
+                current_min_index = min_indices_scalar[j];
+            }
+        }
+    }
+
+    if (i < ny) {
+        __m256 x0 = _mm256_loadu_ps(x);
 
-            // compute y^2 - 2 * (x, y), which is sufficient for looking for the
-            //   lowest distance.
-            const float distance = y_sqlen[0] - 2 * dp + x_sqlen;
-            distances[i] = distance;
+        for (; i < ny; i++) {
+            __m256 accu = ElementOpL2::op(x0, _mm256_loadu_ps(y));
+            y += 8;
+            const float distance = horizontal_sum(accu);
 
-            y += 1;
-            y_sqlen += 1;
+            if (current_min_distance > distance) {
+                current_min_distance = distance;
+                current_min_index = i;
+            }
         }
     }
-}
-#endif
-
-void fvec_L2sqr_ny_transposed(
-        float* dis,
-        const float* x,
-        const float* y,
-        const float* y_sqlen,
-        size_t d,
-        size_t d_offset,
-        size_t ny) {
-    // optimized for a few special cases
-
-#ifdef __AVX2__
-#define DISPATCH(dval)                             \
-    case dval:                                     \
-        return fvec_L2sqr_ny_y_transposed_D<dval>( \
-                dis, x, y, y_sqlen, d_offset, ny);
 
-    switch (d) {
-        DISPATCH(1)
-        DISPATCH(2)
-        DISPATCH(4)
-        DISPATCH(8)
-        default:
-            return fvec_L2sqr_ny_y_transposed_ref(
-                    dis, x, y, y_sqlen, d, d_offset, ny);
-    }
-#undef DISPATCH
-#else
-    // non-AVX2 case
-    return fvec_L2sqr_ny_y_transposed_ref(dis, x, y, y_sqlen, d, d_offset, ny);
-#endif
+    return current_min_index;
 }
 
-#ifdef __AVX2__
+#elif defined(__AVX2__)
 
 size_t fvec_L2sqr_ny_nearest_D2(
         float* distances_tmp_buffer,
@@ -1476,7 +2316,123 @@ size_t fvec_L2sqr_ny_nearest(
 #undef DISPATCH
 }
 
-#ifdef __AVX2__
+#if defined(__AVX512F__)
+
+template <size_t DIM>
+size_t fvec_L2sqr_ny_nearest_y_transposed_D(
+        float* distances_tmp_buffer,
+        const float* x,
+        const float* y,
+        const float* y_sqlen,
+        const size_t d_offset,
+        size_t ny) {
+    // This implementation does not use distances_tmp_buffer.
+
+    // Current index being processed
+    size_t i = 0;
+
+    // Min distance and the index of the closest vector so far
+    float current_min_distance = HUGE_VALF;
+    size_t current_min_index = 0;
+
+    // Process 16 vectors per loop
+    const size_t ny16 = ny / 16;
+
+    if (ny16 > 0) {
+        // Track min distance and the closest vector independently
+        // for each of 16 AVX-512 components.
+        __m512 min_distances = _mm512_set1_ps(HUGE_VALF);
+        __m512i min_indices = _mm512_set1_epi32(0);
+
+        __m512i current_indices = _mm512_setr_epi32(
+                0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+        const __m512i indices_increment = _mm512_set1_epi32(16);
+
+        // m[i] = (2 * x[i], ... 2 * x[i])
+        __m512 m[DIM];
+        for (size_t j = 0; j < DIM; j++) {
+            m[j] = _mm512_set1_ps(x[j]);
+            m[j] = _mm512_add_ps(m[j], m[j]);
+        }
+
+        for (; i < ny16 * 16; i += 16) {
+            // Compute dot products
+            const __m512 v0 = _mm512_loadu_ps(y + 0 * d_offset);
+            __m512 dp = _mm512_mul_ps(m[0], v0);
+            for (size_t j = 1; j < DIM; j++) {
+                const __m512 vj = _mm512_loadu_ps(y + j * d_offset);
+                dp = _mm512_fmadd_ps(m[j], vj, dp);
+            }
+
+            // Compute y^2 - (2 * x, y), which is sufficient for looking for the
+            // lowest distance.
+            // x^2 is the constant that can be avoided.
+            const __m512 distances =
+                    _mm512_sub_ps(_mm512_loadu_ps(y_sqlen), dp);
+
+            // Compare the new distances to the min distances
+            __mmask16 comparison =
+                    _mm512_cmp_ps_mask(min_distances, distances, _CMP_LT_OS);
+
+            // Update min distances and indices with closest vectors if needed
+            min_distances =
+                    _mm512_mask_blend_ps(comparison, distances, min_distances);
+            min_indices = _mm512_castps_si512(_mm512_mask_blend_ps(
+                    comparison,
+                    _mm512_castsi512_ps(current_indices),
+                    _mm512_castsi512_ps(min_indices)));
+
+            // Update current indices values. Basically, +16 to each of the 16
+            // AVX-512 components.
+            current_indices =
+                    _mm512_add_epi32(current_indices, indices_increment);
+
+            // Scroll y and y_sqlen forward.
+            y += 16;
+            y_sqlen += 16;
+        }
+
+        // Dump values and find the minimum distance / minimum index
+        float min_distances_scalar[16];
+        uint32_t min_indices_scalar[16];
+        _mm512_storeu_ps(min_distances_scalar, min_distances);
+        _mm512_storeu_si512((__m512i*)(min_indices_scalar), min_indices);
+
+        for (size_t j = 0; j < 16; j++) {
+            if (current_min_distance > min_distances_scalar[j]) {
+                current_min_distance = min_distances_scalar[j];
+                current_min_index = min_indices_scalar[j];
+            }
+        }
+    }
+
+    if (i < ny) {
+        // Process leftovers
+        for (; i < ny; i++) {
+            float dp = 0;
+            for (size_t j = 0; j < DIM; j++) {
+                dp += x[j] * y[j * d_offset];
+            }
+
+            // Compute y^2 - 2 * (x, y), which is sufficient for looking for the
+            // lowest distance.
+            const float distance = y_sqlen[0] - 2 * dp;
+
+            if (current_min_distance > distance) {
+                current_min_distance = distance;
+                current_min_index = i;
+            }
+
+            y += 1;
+            y_sqlen += 1;
+        }
+    }
+
+    return current_min_index;
+}
+
+#elif defined(__AVX2__)
+
 template <size_t DIM>
 size_t fvec_L2sqr_ny_nearest_y_transposed_D(
         float* distances_tmp_buffer,
@@ -1592,6 +2548,7 @@ size_t fvec_L2sqr_ny_nearest_y_transposed_D(
 
     return current_min_index;
 }
+
 #endif
 
 size_t fvec_L2sqr_ny_nearest_y_transposed(
@@ -1632,6 +2589,7 @@ size_t fvec_L2sqr_ny_nearest_y_transposed(
 
 float fvec_L1(const float* x, const float* y, size_t d) {
     __m256 msum1 = _mm256_setzero_ps();
+    // signmask used for absolute value
     __m256 signmask = _mm256_castsi256_ps(_mm256_set1_epi32(0x7fffffffUL));
 
     while (d >= 8) {
@@ -1639,7 +2597,9 @@ float fvec_L1(const float* x, const float* y, size_t d) {
         x += 8;
         __m256 my = _mm256_loadu_ps(y);
         y += 8;
+        // subtract
         const __m256 a_m_b = _mm256_sub_ps(mx, my);
+        // find sum of absolute value of distances (manhattan distance)
         msum1 = _mm256_add_ps(msum1, _mm256_and_ps(signmask, a_m_b));
         d -= 8;
     }
@@ -1672,6 +2632,7 @@ float fvec_L1(const float* x, const float* y, size_t d) {
 
 float fvec_Linf(const float* x, const float* y, size_t d) {
     __m256 msum1 = _mm256_setzero_ps();
+    // signmask used for absolute value
     __m256 signmask = _mm256_castsi256_ps(_mm256_set1_epi32(0x7fffffffUL));
 
     while (d >= 8) {
@@ -1679,7 +2640,9 @@ float fvec_Linf(const float* x, const float* y, size_t d) {
         x += 8;
         __m256 my = _mm256_loadu_ps(y);
         y += 8;
+        // subtract
         const __m256 a_m_b = _mm256_sub_ps(mx, my);
+        // find max of absolute value of distances (chebyshev distance)
         msum1 = _mm256_max_ps(msum1, _mm256_and_ps(signmask, a_m_b));
         d -= 8;
     }
@@ -1720,6 +2683,441 @@ float fvec_Linf(const float* x, const float* y, size_t d) {
     return fvec_Linf_ref(x, y, d);
 }
 
+#elif defined(__ARM_FEATURE_SVE)
+
+struct ElementOpIP {
+    static svfloat32_t op(svbool_t pg, svfloat32_t x, svfloat32_t y) {
+        return svmul_f32_x(pg, x, y);
+    }
+    static svfloat32_t merge(
+            svbool_t pg,
+            svfloat32_t z,
+            svfloat32_t x,
+            svfloat32_t y) {
+        return svmla_f32_x(pg, z, x, y);
+    }
+};
+
+template <typename ElementOp>
+void fvec_op_ny_sve_d1(float* dis, const float* x, const float* y, size_t ny) {
+    const size_t lanes = svcntw();
+    const size_t lanes2 = lanes * 2;
+    const size_t lanes3 = lanes * 3;
+    const size_t lanes4 = lanes * 4;
+    const svbool_t pg = svptrue_b32();
+    const svfloat32_t x0 = svdup_n_f32(x[0]);
+    size_t i = 0;
+    for (; i + lanes4 < ny; i += lanes4) {
+        svfloat32_t y0 = svld1_f32(pg, y);
+        svfloat32_t y1 = svld1_f32(pg, y + lanes);
+        svfloat32_t y2 = svld1_f32(pg, y + lanes2);
+        svfloat32_t y3 = svld1_f32(pg, y + lanes3);
+        y0 = ElementOp::op(pg, x0, y0);
+        y1 = ElementOp::op(pg, x0, y1);
+        y2 = ElementOp::op(pg, x0, y2);
+        y3 = ElementOp::op(pg, x0, y3);
+        svst1_f32(pg, dis, y0);
+        svst1_f32(pg, dis + lanes, y1);
+        svst1_f32(pg, dis + lanes2, y2);
+        svst1_f32(pg, dis + lanes3, y3);
+        y += lanes4;
+        dis += lanes4;
+    }
+    const svbool_t pg0 = svwhilelt_b32_u64(i, ny);
+    const svbool_t pg1 = svwhilelt_b32_u64(i + lanes, ny);
+    const svbool_t pg2 = svwhilelt_b32_u64(i + lanes2, ny);
+    const svbool_t pg3 = svwhilelt_b32_u64(i + lanes3, ny);
+    svfloat32_t y0 = svld1_f32(pg0, y);
+    svfloat32_t y1 = svld1_f32(pg1, y + lanes);
+    svfloat32_t y2 = svld1_f32(pg2, y + lanes2);
+    svfloat32_t y3 = svld1_f32(pg3, y + lanes3);
+    y0 = ElementOp::op(pg0, x0, y0);
+    y1 = ElementOp::op(pg1, x0, y1);
+    y2 = ElementOp::op(pg2, x0, y2);
+    y3 = ElementOp::op(pg3, x0, y3);
+    svst1_f32(pg0, dis, y0);
+    svst1_f32(pg1, dis + lanes, y1);
+    svst1_f32(pg2, dis + lanes2, y2);
+    svst1_f32(pg3, dis + lanes3, y3);
+}
+
+template <typename ElementOp>
+void fvec_op_ny_sve_d2(float* dis, const float* x, const float* y, size_t ny) {
+    const size_t lanes = svcntw();
+    const size_t lanes2 = lanes * 2;
+    const size_t lanes4 = lanes * 4;
+    const svbool_t pg = svptrue_b32();
+    const svfloat32_t x0 = svdup_n_f32(x[0]);
+    const svfloat32_t x1 = svdup_n_f32(x[1]);
+    size_t i = 0;
+    for (; i + lanes2 < ny; i += lanes2) {
+        const svfloat32x2_t y0 = svld2_f32(pg, y);
+        const svfloat32x2_t y1 = svld2_f32(pg, y + lanes2);
+        svfloat32_t y00 = svget2_f32(y0, 0);
+        const svfloat32_t y01 = svget2_f32(y0, 1);
+        svfloat32_t y10 = svget2_f32(y1, 0);
+        const svfloat32_t y11 = svget2_f32(y1, 1);
+        y00 = ElementOp::op(pg, x0, y00);
+        y10 = ElementOp::op(pg, x0, y10);
+        y00 = ElementOp::merge(pg, y00, x1, y01);
+        y10 = ElementOp::merge(pg, y10, x1, y11);
+        svst1_f32(pg, dis, y00);
+        svst1_f32(pg, dis + lanes, y10);
+        y += lanes4;
+        dis += lanes2;
+    }
+    const svbool_t pg0 = svwhilelt_b32_u64(i, ny);
+    const svbool_t pg1 = svwhilelt_b32_u64(i + lanes, ny);
+    const svfloat32x2_t y0 = svld2_f32(pg0, y);
+    const svfloat32x2_t y1 = svld2_f32(pg1, y + lanes2);
+    svfloat32_t y00 = svget2_f32(y0, 0);
+    const svfloat32_t y01 = svget2_f32(y0, 1);
+    svfloat32_t y10 = svget2_f32(y1, 0);
+    const svfloat32_t y11 = svget2_f32(y1, 1);
+    y00 = ElementOp::op(pg0, x0, y00);
+    y10 = ElementOp::op(pg1, x0, y10);
+    y00 = ElementOp::merge(pg0, y00, x1, y01);
+    y10 = ElementOp::merge(pg1, y10, x1, y11);
+    svst1_f32(pg0, dis, y00);
+    svst1_f32(pg1, dis + lanes, y10);
+}
+
+template <typename ElementOp>
+void fvec_op_ny_sve_d4(float* dis, const float* x, const float* y, size_t ny) {
+    const size_t lanes = svcntw();
+    const size_t lanes4 = lanes * 4;
+    const svbool_t pg = svptrue_b32();
+    const svfloat32_t x0 = svdup_n_f32(x[0]);
+    const svfloat32_t x1 = svdup_n_f32(x[1]);
+    const svfloat32_t x2 = svdup_n_f32(x[2]);
+    const svfloat32_t x3 = svdup_n_f32(x[3]);
+    size_t i = 0;
+    for (; i + lanes < ny; i += lanes) {
+        const svfloat32x4_t y0 = svld4_f32(pg, y);
+        svfloat32_t y00 = svget4_f32(y0, 0);
+        const svfloat32_t y01 = svget4_f32(y0, 1);
+        svfloat32_t y02 = svget4_f32(y0, 2);
+        const svfloat32_t y03 = svget4_f32(y0, 3);
+        y00 = ElementOp::op(pg, x0, y00);
+        y02 = ElementOp::op(pg, x2, y02);
+        y00 = ElementOp::merge(pg, y00, x1, y01);
+        y02 = ElementOp::merge(pg, y02, x3, y03);
+        y00 = svadd_f32_x(pg, y00, y02);
+        svst1_f32(pg, dis, y00);
+        y += lanes4;
+        dis += lanes;
+    }
+    const svbool_t pg0 = svwhilelt_b32_u64(i, ny);
+    const svfloat32x4_t y0 = svld4_f32(pg0, y);
+    svfloat32_t y00 = svget4_f32(y0, 0);
+    const svfloat32_t y01 = svget4_f32(y0, 1);
+    svfloat32_t y02 = svget4_f32(y0, 2);
+    const svfloat32_t y03 = svget4_f32(y0, 3);
+    y00 = ElementOp::op(pg0, x0, y00);
+    y02 = ElementOp::op(pg0, x2, y02);
+    y00 = ElementOp::merge(pg0, y00, x1, y01);
+    y02 = ElementOp::merge(pg0, y02, x3, y03);
+    y00 = svadd_f32_x(pg0, y00, y02);
+    svst1_f32(pg0, dis, y00);
+}
+
+template <typename ElementOp>
+void fvec_op_ny_sve_d8(float* dis, const float* x, const float* y, size_t ny) {
+    const size_t lanes = svcntw();
+    const size_t lanes4 = lanes * 4;
+    const size_t lanes8 = lanes * 8;
+    const svbool_t pg = svptrue_b32();
+    const svfloat32_t x0 = svdup_n_f32(x[0]);
+    const svfloat32_t x1 = svdup_n_f32(x[1]);
+    const svfloat32_t x2 = svdup_n_f32(x[2]);
+    const svfloat32_t x3 = svdup_n_f32(x[3]);
+    const svfloat32_t x4 = svdup_n_f32(x[4]);
+    const svfloat32_t x5 = svdup_n_f32(x[5]);
+    const svfloat32_t x6 = svdup_n_f32(x[6]);
+    const svfloat32_t x7 = svdup_n_f32(x[7]);
+    size_t i = 0;
+    for (; i + lanes < ny; i += lanes) {
+        const svfloat32x4_t ya = svld4_f32(pg, y);
+        const svfloat32x4_t yb = svld4_f32(pg, y + lanes4);
+        const svfloat32_t ya0 = svget4_f32(ya, 0);
+        const svfloat32_t ya1 = svget4_f32(ya, 1);
+        const svfloat32_t ya2 = svget4_f32(ya, 2);
+        const svfloat32_t ya3 = svget4_f32(ya, 3);
+        const svfloat32_t yb0 = svget4_f32(yb, 0);
+        const svfloat32_t yb1 = svget4_f32(yb, 1);
+        const svfloat32_t yb2 = svget4_f32(yb, 2);
+        const svfloat32_t yb3 = svget4_f32(yb, 3);
+        svfloat32_t y0 = svuzp1(ya0, yb0);
+        const svfloat32_t y1 = svuzp1(ya1, yb1);
+        svfloat32_t y2 = svuzp1(ya2, yb2);
+        const svfloat32_t y3 = svuzp1(ya3, yb3);
+        svfloat32_t y4 = svuzp2(ya0, yb0);
+        const svfloat32_t y5 = svuzp2(ya1, yb1);
+        svfloat32_t y6 = svuzp2(ya2, yb2);
+        const svfloat32_t y7 = svuzp2(ya3, yb3);
+        y0 = ElementOp::op(pg, x0, y0);
+        y2 = ElementOp::op(pg, x2, y2);
+        y4 = ElementOp::op(pg, x4, y4);
+        y6 = ElementOp::op(pg, x6, y6);
+        y0 = ElementOp::merge(pg, y0, x1, y1);
+        y2 = ElementOp::merge(pg, y2, x3, y3);
+        y4 = ElementOp::merge(pg, y4, x5, y5);
+        y6 = ElementOp::merge(pg, y6, x7, y7);
+        y0 = svadd_f32_x(pg, y0, y2);
+        y4 = svadd_f32_x(pg, y4, y6);
+        y0 = svadd_f32_x(pg, y0, y4);
+        svst1_f32(pg, dis, y0);
+        y += lanes8;
+        dis += lanes;
+    }
+    const svbool_t pg0 = svwhilelt_b32_u64(i, ny);
+    const svbool_t pga = svwhilelt_b32_u64(i * 2, ny * 2);
+    const svbool_t pgb = svwhilelt_b32_u64(i * 2 + lanes, ny * 2);
+    const svfloat32x4_t ya = svld4_f32(pga, y);
+    const svfloat32x4_t yb = svld4_f32(pgb, y + lanes4);
+    const svfloat32_t ya0 = svget4_f32(ya, 0);
+    const svfloat32_t ya1 = svget4_f32(ya, 1);
+    const svfloat32_t ya2 = svget4_f32(ya, 2);
+    const svfloat32_t ya3 = svget4_f32(ya, 3);
+    const svfloat32_t yb0 = svget4_f32(yb, 0);
+    const svfloat32_t yb1 = svget4_f32(yb, 1);
+    const svfloat32_t yb2 = svget4_f32(yb, 2);
+    const svfloat32_t yb3 = svget4_f32(yb, 3);
+    svfloat32_t y0 = svuzp1(ya0, yb0);
+    const svfloat32_t y1 = svuzp1(ya1, yb1);
+    svfloat32_t y2 = svuzp1(ya2, yb2);
+    const svfloat32_t y3 = svuzp1(ya3, yb3);
+    svfloat32_t y4 = svuzp2(ya0, yb0);
+    const svfloat32_t y5 = svuzp2(ya1, yb1);
+    svfloat32_t y6 = svuzp2(ya2, yb2);
+    const svfloat32_t y7 = svuzp2(ya3, yb3);
+    y0 = ElementOp::op(pg0, x0, y0);
+    y2 = ElementOp::op(pg0, x2, y2);
+    y4 = ElementOp::op(pg0, x4, y4);
+    y6 = ElementOp::op(pg0, x6, y6);
+    y0 = ElementOp::merge(pg0, y0, x1, y1);
+    y2 = ElementOp::merge(pg0, y2, x3, y3);
+    y4 = ElementOp::merge(pg0, y4, x5, y5);
+    y6 = ElementOp::merge(pg0, y6, x7, y7);
+    y0 = svadd_f32_x(pg0, y0, y2);
+    y4 = svadd_f32_x(pg0, y4, y6);
+    y0 = svadd_f32_x(pg0, y0, y4);
+    svst1_f32(pg0, dis, y0);
+    y += lanes8;
+    dis += lanes;
+}
+
+template <typename ElementOp>
+void fvec_op_ny_sve_lanes1(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t lanes = svcntw();
+    const size_t lanes2 = lanes * 2;
+    const size_t lanes3 = lanes * 3;
+    const size_t lanes4 = lanes * 4;
+    const svbool_t pg = svptrue_b32();
+    const svfloat32_t x0 = svld1_f32(pg, x);
+    size_t i = 0;
+    for (; i + 3 < ny; i += 4) {
+        svfloat32_t y0 = svld1_f32(pg, y);
+        svfloat32_t y1 = svld1_f32(pg, y + lanes);
+        svfloat32_t y2 = svld1_f32(pg, y + lanes2);
+        svfloat32_t y3 = svld1_f32(pg, y + lanes3);
+        y += lanes4;
+        y0 = ElementOp::op(pg, x0, y0);
+        y1 = ElementOp::op(pg, x0, y1);
+        y2 = ElementOp::op(pg, x0, y2);
+        y3 = ElementOp::op(pg, x0, y3);
+        dis[i] = svaddv_f32(pg, y0);
+        dis[i + 1] = svaddv_f32(pg, y1);
+        dis[i + 2] = svaddv_f32(pg, y2);
+        dis[i + 3] = svaddv_f32(pg, y3);
+    }
+    for (; i < ny; ++i) {
+        svfloat32_t y0 = svld1_f32(pg, y);
+        y += lanes;
+        y0 = ElementOp::op(pg, x0, y0);
+        dis[i] = svaddv_f32(pg, y0);
+    }
+}
+
+template <typename ElementOp>
+void fvec_op_ny_sve_lanes2(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t lanes = svcntw();
+    const size_t lanes2 = lanes * 2;
+    const size_t lanes3 = lanes * 3;
+    const size_t lanes4 = lanes * 4;
+    const svbool_t pg = svptrue_b32();
+    const svfloat32_t x0 = svld1_f32(pg, x);
+    const svfloat32_t x1 = svld1_f32(pg, x + lanes);
+    size_t i = 0;
+    for (; i + 1 < ny; i += 2) {
+        svfloat32_t y00 = svld1_f32(pg, y);
+        const svfloat32_t y01 = svld1_f32(pg, y + lanes);
+        svfloat32_t y10 = svld1_f32(pg, y + lanes2);
+        const svfloat32_t y11 = svld1_f32(pg, y + lanes3);
+        y += lanes4;
+        y00 = ElementOp::op(pg, x0, y00);
+        y10 = ElementOp::op(pg, x0, y10);
+        y00 = ElementOp::merge(pg, y00, x1, y01);
+        y10 = ElementOp::merge(pg, y10, x1, y11);
+        dis[i] = svaddv_f32(pg, y00);
+        dis[i + 1] = svaddv_f32(pg, y10);
+    }
+    if (i < ny) {
+        svfloat32_t y0 = svld1_f32(pg, y);
+        const svfloat32_t y1 = svld1_f32(pg, y + lanes);
+        y0 = ElementOp::op(pg, x0, y0);
+        y0 = ElementOp::merge(pg, y0, x1, y1);
+        dis[i] = svaddv_f32(pg, y0);
+    }
+}
+
+template <typename ElementOp>
+void fvec_op_ny_sve_lanes3(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t lanes = svcntw();
+    const size_t lanes2 = lanes * 2;
+    const size_t lanes3 = lanes * 3;
+    const svbool_t pg = svptrue_b32();
+    const svfloat32_t x0 = svld1_f32(pg, x);
+    const svfloat32_t x1 = svld1_f32(pg, x + lanes);
+    const svfloat32_t x2 = svld1_f32(pg, x + lanes2);
+    for (size_t i = 0; i < ny; ++i) {
+        svfloat32_t y0 = svld1_f32(pg, y);
+        const svfloat32_t y1 = svld1_f32(pg, y + lanes);
+        svfloat32_t y2 = svld1_f32(pg, y + lanes2);
+        y += lanes3;
+        y0 = ElementOp::op(pg, x0, y0);
+        y0 = ElementOp::merge(pg, y0, x1, y1);
+        y0 = ElementOp::merge(pg, y0, x2, y2);
+        dis[i] = svaddv_f32(pg, y0);
+    }
+}
+
+template <typename ElementOp>
+void fvec_op_ny_sve_lanes4(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t ny) {
+    const size_t lanes = svcntw();
+    const size_t lanes2 = lanes * 2;
+    const size_t lanes3 = lanes * 3;
+    const size_t lanes4 = lanes * 4;
+    const svbool_t pg = svptrue_b32();
+    const svfloat32_t x0 = svld1_f32(pg, x);
+    const svfloat32_t x1 = svld1_f32(pg, x + lanes);
+    const svfloat32_t x2 = svld1_f32(pg, x + lanes2);
+    const svfloat32_t x3 = svld1_f32(pg, x + lanes3);
+    for (size_t i = 0; i < ny; ++i) {
+        svfloat32_t y0 = svld1_f32(pg, y);
+        const svfloat32_t y1 = svld1_f32(pg, y + lanes);
+        svfloat32_t y2 = svld1_f32(pg, y + lanes2);
+        const svfloat32_t y3 = svld1_f32(pg, y + lanes3);
+        y += lanes4;
+        y0 = ElementOp::op(pg, x0, y0);
+        y2 = ElementOp::op(pg, x2, y2);
+        y0 = ElementOp::merge(pg, y0, x1, y1);
+        y2 = ElementOp::merge(pg, y2, x3, y3);
+        y0 = svadd_f32_x(pg, y0, y2);
+        dis[i] = svaddv_f32(pg, y0);
+    }
+}
+
+void fvec_L2sqr_ny(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t d,
+        size_t ny) {
+    fvec_L2sqr_ny_ref(dis, x, y, d, ny);
+}
+
+void fvec_L2sqr_ny_transposed(
+        float* dis,
+        const float* x,
+        const float* y,
+        const float* y_sqlen,
+        size_t d,
+        size_t d_offset,
+        size_t ny) {
+    return fvec_L2sqr_ny_y_transposed_ref(dis, x, y, y_sqlen, d, d_offset, ny);
+}
+
+size_t fvec_L2sqr_ny_nearest(
+        float* distances_tmp_buffer,
+        const float* x,
+        const float* y,
+        size_t d,
+        size_t ny) {
+    return fvec_L2sqr_ny_nearest_ref(distances_tmp_buffer, x, y, d, ny);
+}
+
+size_t fvec_L2sqr_ny_nearest_y_transposed(
+        float* distances_tmp_buffer,
+        const float* x,
+        const float* y,
+        const float* y_sqlen,
+        size_t d,
+        size_t d_offset,
+        size_t ny) {
+    return fvec_L2sqr_ny_nearest_y_transposed_ref(
+            distances_tmp_buffer, x, y, y_sqlen, d, d_offset, ny);
+}
+
+float fvec_L1(const float* x, const float* y, size_t d) {
+    return fvec_L1_ref(x, y, d);
+}
+
+float fvec_Linf(const float* x, const float* y, size_t d) {
+    return fvec_Linf_ref(x, y, d);
+}
+
+void fvec_inner_products_ny(
+        float* dis,
+        const float* x,
+        const float* y,
+        size_t d,
+        size_t ny) {
+    const size_t lanes = svcntw();
+    switch (d) {
+        case 1:
+            fvec_op_ny_sve_d1<ElementOpIP>(dis, x, y, ny);
+            break;
+        case 2:
+            fvec_op_ny_sve_d2<ElementOpIP>(dis, x, y, ny);
+            break;
+        case 4:
+            fvec_op_ny_sve_d4<ElementOpIP>(dis, x, y, ny);
+            break;
+        case 8:
+            fvec_op_ny_sve_d8<ElementOpIP>(dis, x, y, ny);
+            break;
+        default:
+            if (d == lanes)
+                fvec_op_ny_sve_lanes1<ElementOpIP>(dis, x, y, ny);
+            else if (d == lanes * 2)
+                fvec_op_ny_sve_lanes2<ElementOpIP>(dis, x, y, ny);
+            else if (d == lanes * 3)
+                fvec_op_ny_sve_lanes3<ElementOpIP>(dis, x, y, ny);
+            else if (d == lanes * 4)
+                fvec_op_ny_sve_lanes4<ElementOpIP>(dis, x, y, ny);
+            else
+                fvec_inner_products_ny_ref(dis, x, y, d, ny);
+            break;
+    }
+}
+
 #elif defined(__aarch64__)
 
 // not optimized for ARM
@@ -1858,7 +3256,39 @@ void fvec_inner_products_ny(
         c[i] = a[i] + bf * b[i];
 }
 
-#ifdef __AVX2__
+#if defined(__AVX512F__)
+
+static inline void fvec_madd_avx512(
+        const size_t n,
+        const float* __restrict a,
+        const float bf,
+        const float* __restrict b,
+        float* __restrict c) {
+    const size_t n16 = n / 16;
+    const size_t n_for_masking = n % 16;
+
+    const __m512 bfmm = _mm512_set1_ps(bf);
+
+    size_t idx = 0;
+    for (idx = 0; idx < n16 * 16; idx += 16) {
+        const __m512 ax = _mm512_loadu_ps(a + idx);
+        const __m512 bx = _mm512_loadu_ps(b + idx);
+        const __m512 abmul = _mm512_fmadd_ps(bfmm, bx, ax);
+        _mm512_storeu_ps(c + idx, abmul);
+    }
+
+    if (n_for_masking > 0) {
+        const __mmask16 mask = (1 << n_for_masking) - 1;
+
+        const __m512 ax = _mm512_maskz_loadu_ps(mask, a + idx);
+        const __m512 bx = _mm512_maskz_loadu_ps(mask, b + idx);
+        const __m512 abmul = _mm512_fmadd_ps(bfmm, bx, ax);
+        _mm512_mask_storeu_ps(c + idx, mask, abmul);
+    }
+}
+
+#elif defined(__AVX2__)
+
 static inline void fvec_madd_avx2(
         const size_t n,
         const float* __restrict a,
@@ -1911,6 +3341,7 @@ static inline void fvec_madd_avx2(
         _mm256_maskstore_ps(c + idx, mask, abmul);
     }
 }
+
 #endif
 
 #ifdef __SSE3__
@@ -1936,7 +3367,9 @@ static inline void fvec_madd_avx2(
 }
 
 void fvec_madd(size_t n, const float* a, float bf, const float* b, float* c) {
-#ifdef __AVX2__
+#ifdef __AVX512F__
+    fvec_madd_avx512(n, a, bf, b, c);
+#elif __AVX2__
     fvec_madd_avx2(n, a, bf, b, c);
 #else
     if ((n & 3) == 0 && ((((long)a) | ((long)b) | ((long)c)) & 15) == 0)
@@ -1946,6 +3379,60 @@ void fvec_madd(size_t n, const float* a, float bf, const float* b, float* c) {
 #endif
 }
 
+#elif defined(__ARM_FEATURE_SVE)
+
+void fvec_madd(
+        const size_t n,
+        const float* __restrict a,
+        const float bf,
+        const float* __restrict b,
+        float* __restrict c) {
+    const size_t lanes = static_cast<size_t>(svcntw());
+    const size_t lanes2 = lanes * 2;
+    const size_t lanes3 = lanes * 3;
+    const size_t lanes4 = lanes * 4;
+    size_t i = 0;
+    for (; i + lanes4 < n; i += lanes4) {
+        const auto mask = svptrue_b32();
+        const auto ai0 = svld1_f32(mask, a + i);
+        const auto ai1 = svld1_f32(mask, a + i + lanes);
+        const auto ai2 = svld1_f32(mask, a + i + lanes2);
+        const auto ai3 = svld1_f32(mask, a + i + lanes3);
+        const auto bi0 = svld1_f32(mask, b + i);
+        const auto bi1 = svld1_f32(mask, b + i + lanes);
+        const auto bi2 = svld1_f32(mask, b + i + lanes2);
+        const auto bi3 = svld1_f32(mask, b + i + lanes3);
+        const auto ci0 = svmla_n_f32_x(mask, ai0, bi0, bf);
+        const auto ci1 = svmla_n_f32_x(mask, ai1, bi1, bf);
+        const auto ci2 = svmla_n_f32_x(mask, ai2, bi2, bf);
+        const auto ci3 = svmla_n_f32_x(mask, ai3, bi3, bf);
+        svst1_f32(mask, c + i, ci0);
+        svst1_f32(mask, c + i + lanes, ci1);
+        svst1_f32(mask, c + i + lanes2, ci2);
+        svst1_f32(mask, c + i + lanes3, ci3);
+    }
+    const auto mask0 = svwhilelt_b32_u64(i, n);
+    const auto mask1 = svwhilelt_b32_u64(i + lanes, n);
+    const auto mask2 = svwhilelt_b32_u64(i + lanes2, n);
+    const auto mask3 = svwhilelt_b32_u64(i + lanes3, n);
+    const auto ai0 = svld1_f32(mask0, a + i);
+    const auto ai1 = svld1_f32(mask1, a + i + lanes);
+    const auto ai2 = svld1_f32(mask2, a + i + lanes2);
+    const auto ai3 = svld1_f32(mask3, a + i + lanes3);
+    const auto bi0 = svld1_f32(mask0, b + i);
+    const auto bi1 = svld1_f32(mask1, b + i + lanes);
+    const auto bi2 = svld1_f32(mask2, b + i + lanes2);
+    const auto bi3 = svld1_f32(mask3, b + i + lanes3);
+    const auto ci0 = svmla_n_f32_x(mask0, ai0, bi0, bf);
+    const auto ci1 = svmla_n_f32_x(mask1, ai1, bi1, bf);
+    const auto ci2 = svmla_n_f32_x(mask2, ai2, bi2, bf);
+    const auto ci3 = svmla_n_f32_x(mask3, ai3, bi3, bf);
+    svst1_f32(mask0, c + i, ci0);
+    svst1_f32(mask1, c + i + lanes, ci1);
+    svst1_f32(mask2, c + i + lanes2, ci2);
+    svst1_f32(mask3, c + i + lanes3, ci3);
+}
+
 #elif defined(__aarch64__)
 
 void fvec_madd(size_t n, const float* a, float bf, const float* b, float* c) {
@@ -2278,7 +3765,7 @@ void fvec_add(size_t d, const float* a, float b, float* c) {
     size_t i;
     simd8float32 bv(b);
     for (i = 0; i + 7 < d; i += 8) {
-        simd8float32 ci, ai, bi;
+        simd8float32 ci, ai;
         ai.loadu(a + i);
         ci = ai + bv;
         ci.storeu(c + i);