github.com/weaviate/weaviate@v1.24.6/adapters/repos/db/vector/hnsw/distancer/c/dot_arm64.c (about)

     1  //                           _       _
     2  // __      _____  __ ___   ___  __ _| |_ ___
     3  // \ \ /\ / / _ \/ _` \ \ / / |/ _` | __/ _ \
     4  //  \ V  V /  __/ (_| |\ V /| | (_| | ||  __/
     5  //   \_/\_/ \___|\__,_| \_/ |_|\__,_|\__\___|
     6  //
     7  //  Copyright © 2016 - 2024 Weaviate B.V. All rights reserved.
     8  //
     9  //  CONTACT: hello@weaviate.io
    10  //
    11  
    12  #include <arm_neon.h>
    13  
    14  // dot only works with length >= 16
    15  void dot(float *a, float *b, float *res, long *len)
    16  {
    17      int size = *len;
    18  
    19      // use the vectorized version for the first n - (n % 4) elements
    20      int l = size - (size % 4);
    21  
    22      // create 4*4 registers to store the result
    23      float32x4_t res_vec0 = vdupq_n_f32(0);
    24      float32x4_t res_vec1 = vdupq_n_f32(0);
    25      float32x4_t res_vec2 = vdupq_n_f32(0);
    26      float32x4_t res_vec3 = vdupq_n_f32(0);
    27  
    28      int i = 0;
    29  
    30      // load 4*4 floats at a time
    31      while (i + 16 <= l)
    32      {
    33          float32x4x4_t a4 = vld1q_f32_x4(a + i);
    34          float32x4x4_t b4 = vld1q_f32_x4(b + i);
    35  
    36          res_vec0 += vmulq_f32(a4.val[0], b4.val[0]);
    37          res_vec1 += vmulq_f32(a4.val[1], b4.val[1]);
    38          res_vec2 += vmulq_f32(a4.val[2], b4.val[2]);
    39          res_vec3 += vmulq_f32(a4.val[3], b4.val[3]);
    40  
    41          i += 16;
    42      }
    43  
    44      while (i < l)
    45      {
    46          float32x4_t a_vec = vld1q_f32(a + i);
    47          float32x4_t b_vec = vld1q_f32(b + i);
    48          res_vec0 += vmulq_f32(a_vec, b_vec);
    49  
    50          i += 4;
    51      }
    52  
    53      // convert to scalar
    54      float sum = vaddvq_f32(res_vec0);
    55      sum += vaddvq_f32(res_vec1);
    56      sum += vaddvq_f32(res_vec2);
    57      sum += vaddvq_f32(res_vec3);
    58  
    59      // add the remaining vectors
    60      for (int i = l; i < size; i++)
    61      {
    62          sum += a[i] * b[i];
    63      }
    64  
    65      res[0] = sum;
    66  }