PostingsDecodingUtil: interchange loops to enable better memory access and SIMD vectorisation (#15110)

Author: RamakrishnaChilaka

lucene/CHANGES.txt

@@ -154,6 +154,8 @@ New Features
 
 * GITHUB#14729: Support for Re-Ranking Queries using Late Interaction Model Multi-Vectors. (Vigya Sharma, Jim Ferenczi)
 
+* GITHUB#15110: PostingsDecodingUtil: interchange loops to enable better memory access and SIMD vectorisation. (Ramakrishna Chilaka)
+
 Improvements
 ---------------------
 * GITHUB#14458: Add an IndexDeletion policy that retains the last N commits. (Owais Kazi)

lucene/benchmark-jmh/src/java/org/apache/lucene/benchmark/jmh/PostingIndexInputBenchmark.java

@@ -105,4 +105,20 @@ public void decodeAndPrefixSum(Blackhole bh) throws IOException {
     postingIn.decodeAndPrefixSum(bpv, 100, values);
     bh.consume(values);
   }
+
+  @Benchmark
+  @Fork(jvmArgsPrepend = {"--add-modules=jdk.incubator.vector"})
+  public void decodeVector(Blackhole bh) throws IOException {
+    in.seek(3); // random unaligned offset
+    postingIn.decode(bpv, values);
+    bh.consume(values);
+  }
+
+  @Benchmark
+  @Fork(jvmArgsPrepend = {"--add-modules=jdk.incubator.vector"})
+  public void decodeAndPrefixSumVector(Blackhole bh) throws IOException {
+    in.seek(3); // random unaligned offset
+    postingIn.decodeAndPrefixSum(bpv, 100, values);
+    bh.consume(values);
+  }
 }

lucene/core/src/java/org/apache/lucene/internal/vectorization/PostingDecodingUtil.java

@@ -44,14 +44,21 @@ protected PostingDecodingUtil(IndexInput in) {
   public void splitInts(
       int count, int[] b, int bShift, int dec, int bMask, int[] c, int cIndex, int cMask)
       throws IOException {
-    // Default implementation, which takes advantage of the C2 compiler's loop unrolling and
-    // auto-vectorization.
     in.readInts(c, cIndex, count);
-    int maxIter = (bShift - 1) / dec;
-    for (int i = 0; i < count; ++i) {
-      for (int j = 0; j <= maxIter; ++j) {
-        b[count * j + i] = (c[cIndex + i] >>> (bShift - j * dec)) & bMask;
+    final int maxIter = (bShift - 1) / dec;
+
+    // Process each shift level across all elements (better for vectorization)
+    for (int j = 0; j <= maxIter; ++j) {
+      final int shift = bShift - j * dec;
+      final int bOffset = count * j;
+      // Vectorizable loop: contiguous memory access with simple operations
+      for (int i = 0; i < count; ++i) {
+        b[bOffset + i] = (c[cIndex + i] >>> shift) & bMask;
       }
+    }
+
+    // Apply mask to c array (vectorizable)
+    for (int i = 0; i < count; ++i) {
       c[cIndex + i] &= cMask;
     }
   }