Use Panama Vector API to SIMD-evaluate fixed-cardinality sorted numeric range queries in rangeIntoBitSet()

lucene/CHANGES.txt

@@ -291,7 +291,8 @@ Improvements
 
 Optimizations
 ---------------------
-(No changes)
+* GITHUB#16283: Use Panama Vector API to SIMD-evaluate fixed-cardinality sorted numeric range
+  queries in rangeIntoBitSet. (Costin Leau)
 
 Bug Fixes
 ---------------------

lucene/core/src/java/org/apache/lucene/codecs/lucene90/Lucene90DocValuesProducer.java

@@ -495,29 +495,6 @@ private static int fixedCardinality(
     return (int) cardinality;
   }
 
-  private static void sortedNumericScalarRangeIntoBitSet(
-      LongValues values,
-      int fromDoc,
-      int toDoc,
-      int cardinality,
-      long minValue,
-      long maxValue,
-      FixedBitSet bitSet,
-      int offset) {
-    for (int doc = fromDoc; doc < toDoc; doc++) {
-      long valueOffset = (long) doc * cardinality;
-      for (int i = 0; i < cardinality; i++) {
-        long value = values.get(valueOffset + i);
-        if (value >= minValue) {
-          if (value <= maxValue) {
-            bitSet.set(doc - offset);
-          }
-          break;
-        }
-      }
-    }
-  }
-
   private static boolean sortedNumericMatchesRange(
       LongValues values, long start, long end, long minValue, long maxValue) {
     for (long valueOffset = start; valueOffset < end; valueOffset++) {
@@ -1906,7 +1883,7 @@ public void rangeIntoBitSet(
           }
           int cardinality = denseFixedCardinality;
           if (cardinality > 1) {
-            sortedNumericScalarRangeIntoBitSet(
+            DOC_VALUES_RANGE_SUPPORT.sortedNumericRangeIntoBitSet(
                 values, fromDoc, endDoc, cardinality, minValue, maxValue, bitSet, offset);
             return;
           }

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

@@ -50,4 +50,31 @@ void rangeIntoBitSet(
       long maxValue,
       FixedBitSet bitSet,
       int offset);
+
+  /**
+   * Fills {@code bitSet} with docs in {@code [fromDoc, toDoc)} whose {@code cardinality} sorted
+   * values contain at least one value in {@code [minValue, maxValue]}.
+   */
+  default void sortedNumericRangeIntoBitSet(
+      LongValues values,
+      int fromDoc,
+      int toDoc,
+      int cardinality,
+      long minValue,
+      long maxValue,
+      FixedBitSet bitSet,
+      int offset) {
+    for (int doc = fromDoc; doc < toDoc; doc++) {
+      long valueOffset = (long) doc * cardinality;
+      for (int i = 0; i < cardinality; i++) {
+        long value = values.get(valueOffset + i);
+        if (value >= minValue) {
+          if (value <= maxValue) {
+            bitSet.set(doc - offset);
+          }
+          break;
+        }
+      }
+    }
+  }
 }

lucene/core/src/java25/org/apache/lucene/internal/vectorization/PanamaDocValuesRangeSupport.java

@@ -69,4 +69,60 @@ public void rangeIntoBitSet(
       }
     }
   }
+
+  @Override
+  public void sortedNumericRangeIntoBitSet(
+      LongValues values,
+      int fromDoc,
+      int toDoc,
+      int cardinality,
+      long minValue,
+      long maxValue,
+      FixedBitSet bitSet,
+      int offset) {
+    final int vectorLen = LONG_SPECIES.length();
+    final int docsPerVector = vectorLen / cardinality;
+    if (docsPerVector == 0 || vectorLen % cardinality != 0) {
+      DocValuesRangeSupport.super.sortedNumericRangeIntoBitSet(
+          values, fromDoc, toDoc, cardinality, minValue, maxValue, bitSet, offset);
+      return;
+    }
+
+    final long[] scratch = new long[vectorLen];
+    final int vectorDocEnd = fromDoc + (toDoc - fromDoc) / docsPerVector * docsPerVector;
+    int doc = fromDoc;
+    for (; doc < vectorDocEnd; doc += docsPerVector) {
+      long valueOffset = (long) doc * cardinality;
+      for (int lane = 0; lane < vectorLen; lane++) {
+        scratch[lane] = values.get(valueOffset + lane);
+      }
+      LongVector vector = LongVector.fromArray(LONG_SPECIES, scratch, 0);
+      // Flat range check on all lanes. Equivalent to the scalar early-break on sorted values:
+      // a value outside [min,max] that is >= min must be > max, so its lane stays unset, and
+      // collapseToDocMask OR-reduces lanes per doc to match "any value in range" semantics.
+      VectorMask<Long> inRange =
+          vector
+              .compare(VectorOperators.GE, minValue)
+              .and(vector.compare(VectorOperators.LE, maxValue));
+      long docMask = collapseToDocMask(inRange.toLong(), cardinality, docsPerVector);
+      if (docMask != 0) {
+        bitSet.orMask(doc - offset, docMask, docsPerVector);
+      }
+    }
+
+    DocValuesRangeSupport.super.sortedNumericRangeIntoBitSet(
+        values, doc, toDoc, cardinality, minValue, maxValue, bitSet, offset);
+  }
+
+  private static long collapseToDocMask(long valueMask, int cardinality, int docsPerVector) {
+    final long valueMaskPerDoc = (1L << cardinality) - 1L;
+    long docMask = 0L;
+    for (int d = 0; d < docsPerVector; d++) {
+      long perDoc = (valueMask >>> (d * cardinality)) & valueMaskPerDoc;
+      if (perDoc != 0) {
+        docMask |= 1L << d;
+      }
+    }
+    return docMask;
+  }
 }