案例Demo

可通过JavaVectorHelper工具对单个文件/整个项目进行扫描，本章节基于单个文件提供对应可用向量化的示例代码，示例代码片段功能是对数组间运算并赋值，代码仅供参考。

1. 扫描前源码文件如图1所示。
   图1 BpVectorReorderer.java
   
2. 运行BpVectorReorderer.java文件中的For循环。
   图2 改造前运行结果
   
3. 使用JavaVectorHelper工具，识别出代码块中加粗显示的代码，可替换为可用向量化代码。

   private static boolean centroidValid(
           float[] centroid, FloatVectorValues vectors, IntsRef ids, int count) throws IOException {
         // recompute centroid to check the incremental calculation
         float[] check = new float[centroid.length];
         computeCentroid(ids, vectors, check, VectorSimilarityFunction.EUCLIDEAN);
         for (int i = 0; i < check.length; ++i) {
           float diff = Math.abs(check[i] - centroid[i]);
           if (diff > 1e-4) {
             return false;
           }
         }
         return true;
       }
       private static void swapIdsAndCentroids(
           IntsRef ids,
           int from,
           int to,
           int midPoint,
           FloatVectorValues vectors,
           float[] leftCentroid,
           float[] rightCentroid,
           float[] scratch,
           int count,
           VectorSimilarityFunction vectorScore)
           throws IOException {
         assert from < to;
         int[] idArr = ids.ints;
         int fromId = idArr[from];
         int toId = idArr[to];
         // Now update the centroids, this makes things much faster than invalidating it and having to
         // recompute on the next iteration.  Should be faster than full recalculation if the number of
         // swaps is reasonable.
         // We want the net effect to be moving "from" left->right and "to" right->left
         // (1) scratch = to - from
         if (count <= MAX_CENTROID_UPDATES && vectorScore == VectorSimilarityFunction.EUCLIDEAN) {
           int relativeMidpoint = midPoint - ids.offset;
           vectorSubtract(vectors.vectorValue(toId), vectors.vectorValue(fromId), scratch);
           // we must normalize to the proper scale by accounting for the number of points contributing
           // to each centroid
           // left += scratch / size(left)
           vectorScalarMul(1 / (float) relativeMidpoint, scratch);
           VectorUtil.add(leftCentroid, scratch);
           // right -= scratch / size(right)
           vectorScalarMul(-relativeMidpoint / (float) (ids.length - relativeMidpoint), scratch);
           VectorUtil.add(rightCentroid, scratch);
         }
         idArr[from] = toId;
         idArr[to] = fromId;
         if (count <= MAX_CENTROID_UPDATES) {
           assert centroidValid(
               leftCentroid, vectors, new IntsRef(idArr, ids.offset, midPoint - ids.offset), count);
           assert centroidValid(
               rightCentroid,
               vectors,
               new IntsRef(ids.ints, midPoint, ids.length - midPoint + ids.offset),
               count);
         }
       }
     /**
      * Adds the second argument to the first
      *
      * @param u the destination
      * @param v the vector to add to the destination
      **/
     static void vectorSubtract(float[] u, float[] v, float[] result) {
       for (int i = 0; i < u.length; i++) {
         result[i] = u[i] - v[i];//可替换为向量化代码
       }
     }

4. 替换为向量化代码，若代码中定义了变量名称，替换后会自动引入所需要的import语句和需要声明的常量语句。

   static void vectorSubtract(float[] u, float[] v, float[] result) {
       int i = 0;//已替换为向量化代码
       for (; i < FLOAT_SPECIES.loopBound(u.length); i = i + FLOAT_SPECIES.length()) {
         FloatVector uVector = FloatVector.fromArray(FLOAT_SPECIES, u, i);
         FloatVector vVector = FloatVector.fromArray(FLOAT_SPECIES, v, i);
         uVector.sub(vVector).intoArray(result, i);
       }
       for (; i < u.length; i++) {
         result[i] = u[i] - v[i];
       }
     }

   需要声明的常量语句：

   private static final VectorSpecies<Float> FLOAT_SPECIES = FloatVector.SPECIES_PREFERRED;

   import语句：

   import jdk.incubator.vector.VectorSpecies;
   import jdk.incubator.vector.FloatVector;

5. 运行BpVectorReorderer.java文件中改造后的For循环。
   图3 改造后运行结果
   
   

   编译运行改造后的文件，运行结果保持不变，性能有所提升。