Add unit-of-least-precision float comparison (#13723)

lucene/CHANGES.txt

@@ -187,8 +187,10 @@ Build
   (Uwe Schindler)
 
 Other
----------------------
-(No changes)
+--------------------
+
+* GITHUB#13720: Add float comparison based on unit of least precision and use it to stop test failures caused by float
+  summation not being associative in IEEE 754. (Alex Herbert, Stefan Vodita)
 
 ======================== Lucene 9.11.1 =======================
 

lucene/facet/src/test/org/apache/lucene/facet/taxonomy/TestTaxonomyFacetAssociations.java

@@ -677,7 +677,7 @@ private void assertFloatFacetResultsEqual(List<FacetResult> expected, List<Facet
 
       assertEquals(expectedResult.dim, actualResult.dim);
       assertArrayEquals(expectedResult.path, actualResult.path);
-      assertEquals((float) expectedResult.value, (float) actualResult.value, 2e-1);
+      assertFloatUlpEquals((float) expectedResult.value, (float) actualResult.value, (short) 2);
       assertEquals(expectedResult.childCount, actualResult.childCount);
     }
   }

lucene/test-framework/src/java/org/apache/lucene/tests/util/LuceneTestCase.java

@@ -916,6 +916,18 @@ public static void assumeNoException(String msg, Exception e) {
     RandomizedTest.assumeNoException(msg, e);
   }
 
+  public static void assertFloatUlpEquals(final float x, final float y, final short maxUlps) {
+    assertTrue(
+        x + " and " + y + " are not within " + maxUlps + " ULPs of each other",
+        TestUtil.floatUlpEquals(x, y, maxUlps));
+  }
+
+  public static void assertDoubleUlpEquals(final double x, final double y, final int maxUlps) {
+    assertTrue(
+        x + " and " + y + " are not within " + maxUlps + " ULPs of each other",
+        TestUtil.doubleUlpEquals(x, y, maxUlps));
+  }
+
   /**
    * Return <code>args</code> as a {@link Set} instance. The order of elements is not preserved in
    * iterators.

lucene/test-framework/src/java/org/apache/lucene/tests/util/TestUtil.java

@@ -471,6 +471,90 @@ private static void checkReaderSanity(LeafReader reader) throws IOException {
     }
   }
 
+  /**
+   * Returns true if the arguments are equal or within the range of allowed error (inclusive).
+   * Returns {@code false} if either of the arguments is NaN.
+   *
+   * <p>Two float numbers are considered equal if there are {@code (maxUlps - 1)} (or fewer)
+   * floating point numbers between them, i.e. two adjacent floating point numbers are considered
+   * equal.
+   *
+   * <p>Adapted from org.apache.commons.numbers.core.Precision
+   *
+   * <p>github: https://github.com/apache/commons-numbers release 1.2
+   *
+   * @param x first value
+   * @param y second value
+   * @param maxUlps {@code (maxUlps - 1)} is the number of floating point values between {@code x}
+   *     and {@code y}.
+   * @return {@code true} if there are fewer than {@code maxUlps} floating point values between
+   *     {@code x} and {@code y}.
+   */
+  public static boolean floatUlpEquals(final float x, final float y, final short maxUlps) {
+    final int xInt = Float.floatToRawIntBits(x);
+    final int yInt = Float.floatToRawIntBits(y);
+
+    if ((xInt ^ yInt) < 0) {
+      // Numbers have opposite signs, take care of overflow.
+      // Remove the sign bit to obtain the absolute ULP above zero.
+      final int deltaPlus = xInt & Integer.MAX_VALUE;
+      final int deltaMinus = yInt & Integer.MAX_VALUE;
+
+      // Note:
+      // If either value is NaN, the exponent bits are set to (255 << 23) and the
+      // distance above 0.0 is always above a short ULP error. So omit the test
+      // for NaN and return directly.
+
+      // Avoid possible overflow from adding the deltas by splitting the comparison
+      return deltaPlus <= maxUlps && deltaMinus <= (maxUlps - deltaPlus);
+    }
+
+    // Numbers have same sign, there is no risk of overflow.
+    return Math.abs(xInt - yInt) <= maxUlps && !Float.isNaN(x) && !Float.isNaN(y);
+  }
+
+  /**
+   * Returns true if the arguments are equal or within the range of allowed error (inclusive).
+   * Returns {@code false} if either of the arguments is NaN.
+   *
+   * <p>Two double numbers are considered equal if there are {@code (maxUlps - 1)} (or fewer)
+   * floating point numbers between them, i.e. two adjacent floating point numbers are considered
+   * equal.
+   *
+   * <p>Adapted from org.apache.commons.numbers.core.Precision
+   *
+   * <p>github: https://github.com/apache/commons-numbers release 1.2
+   *
+   * @param x first value
+   * @param y second value
+   * @param maxUlps {@code (maxUlps - 1)} is the number of floating point values between {@code x}
+   *     and {@code y}.
+   * @return {@code true} if there are fewer than {@code maxUlps} floating point values between
+   *     {@code x} and {@code y}.
+   */
+  public static boolean doubleUlpEquals(final double x, final double y, final int maxUlps) {
+    final long xInt = Double.doubleToRawLongBits(x);
+    final long yInt = Double.doubleToRawLongBits(y);
+
+    if ((xInt ^ yInt) < 0) {
+      // Numbers have opposite signs, take care of overflow.
+      // Remove the sign bit to obtain the absolute ULP above zero.
+      final long deltaPlus = xInt & Long.MAX_VALUE;
+      final long deltaMinus = yInt & Long.MAX_VALUE;
+
+      // Note:
+      // If either value is NaN, the exponent bits are set to (2047L << 52) and the
+      // distance above 0.0 is always above an integer ULP error. So omit the test
+      // for NaN and return directly.
+
+      // Avoid possible overflow from adding the deltas by splitting the comparison
+      return deltaPlus <= maxUlps && deltaMinus <= (maxUlps - deltaPlus);
+    }
+
+    // Numbers have same sign, there is no risk of overflow.
+    return Math.abs(xInt - yInt) <= maxUlps && !Double.isNaN(x) && !Double.isNaN(y);
+  }
+
   /** start and end are BOTH inclusive */
   public static int nextInt(Random r, int start, int end) {
     return RandomNumbers.randomIntBetween(r, start, end);

lucene/test-framework/src/test/org/apache/lucene/tests/util/TestFloatingPointUlpEquality.java

@@ -0,0 +1,136 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package org.apache.lucene.tests.util;
+
+import static org.apache.lucene.tests.util.TestUtil.doubleUlpEquals;
+import static org.apache.lucene.tests.util.TestUtil.floatUlpEquals;
+
+/**
+ * Tests for floating point equality utility methods.
+ *
+ * <p>Adapted from org.apache.commons.numbers.core.PrecisionTest
+ *
+ * <p>github: https://github.com/apache/commons-numbers release 1.2
+ */
+public class TestFloatingPointUlpEquality extends LuceneTestCase {
+  public static void testDoubleEqualsWithAllowedUlps() {
+    assertTrue(doubleUlpEquals(0.0, -0.0, 1));
+    assertTrue(doubleUlpEquals(Double.MIN_VALUE, -0.0, 1));
+    assertFalse(doubleUlpEquals(Double.MIN_VALUE, -Double.MIN_VALUE, 1));
+
+    assertTrue(doubleUlpEquals(1.0, 1 + Math.ulp(1d), 1));
+    assertFalse(doubleUlpEquals(1.0, 1 + 2 * Math.ulp(1d), 1));
+
+    for (double value : new double[] {153.0, -128.0, 0.0, 1.0}) {
+      assertTrue(doubleUlpEquals(value, value, 1));
+      assertTrue(doubleUlpEquals(value, Math.nextUp(value), 1));
+      assertFalse(doubleUlpEquals(value, Math.nextUp(Math.nextUp(value)), 1));
+      assertTrue(doubleUlpEquals(value, Math.nextDown(value), 1));
+      assertFalse(doubleUlpEquals(value, Math.nextDown(Math.nextDown(value)), 1));
+      assertFalse(doubleUlpEquals(value, value, -1));
+      assertFalse(doubleUlpEquals(value, Math.nextUp(value), 0));
+      assertTrue(doubleUlpEquals(value, Math.nextUp(Math.nextUp(value)), 2));
+      assertTrue(doubleUlpEquals(value, Math.nextDown(Math.nextDown(value)), 2));
+    }
+
+    assertTrue(doubleUlpEquals(Double.POSITIVE_INFINITY, Double.POSITIVE_INFINITY, 1));
+    assertTrue(doubleUlpEquals(Double.MAX_VALUE, Double.POSITIVE_INFINITY, 1));
+
+    assertTrue(doubleUlpEquals(Double.NEGATIVE_INFINITY, Double.NEGATIVE_INFINITY, 1));
+    assertTrue(doubleUlpEquals(-Double.MAX_VALUE, Double.NEGATIVE_INFINITY, 1));
+
+    assertFalse(doubleUlpEquals(Double.NaN, Double.NaN, 1));
+    assertFalse(doubleUlpEquals(Double.NaN, Double.NaN, 0));
+    assertFalse(doubleUlpEquals(Double.NaN, 0, 0));
+    assertFalse(doubleUlpEquals(0, Double.NaN, 0));
+    assertFalse(doubleUlpEquals(Double.NaN, Double.POSITIVE_INFINITY, 0));
+    assertFalse(doubleUlpEquals(Double.NaN, Double.NEGATIVE_INFINITY, 0));
+
+    // Create a NaN representation 1 ulp above infinity.
+    // This hits not equal coverage for binary representations within the ulp but using NaN.
+    final double nan =
+        Double.longBitsToDouble(Double.doubleToRawLongBits(Double.POSITIVE_INFINITY) + 1);
+    assertFalse(doubleUlpEquals(nan, Double.POSITIVE_INFINITY, 1));
+    assertFalse(doubleUlpEquals(Double.POSITIVE_INFINITY, nan, 1));
+
+    assertFalse(
+        doubleUlpEquals(Double.NEGATIVE_INFINITY, Double.POSITIVE_INFINITY, Integer.MAX_VALUE));
+    assertFalse(doubleUlpEquals(0, Double.MAX_VALUE, Integer.MAX_VALUE));
+    // Here: f == 5.304989477E-315;
+    // it is used to test the maximum ULP distance between two opposite sign numbers.
+    final double f = Double.longBitsToDouble(1L << 30);
+    assertFalse(doubleUlpEquals(-f, f, Integer.MAX_VALUE));
+    assertTrue(doubleUlpEquals(-f, Math.nextDown(f), Integer.MAX_VALUE));
+    assertTrue(doubleUlpEquals(Math.nextUp(-f), f, Integer.MAX_VALUE));
+    // Maximum distance between same sign numbers.
+    final double f2 = Double.longBitsToDouble((1L << 30) + Integer.MAX_VALUE);
+    assertTrue(doubleUlpEquals(f, f2, Integer.MAX_VALUE));
+    assertFalse(doubleUlpEquals(f, Math.nextUp(f2), Integer.MAX_VALUE));
+    assertFalse(doubleUlpEquals(Math.nextDown(f), f2, Integer.MAX_VALUE));
+  }
+
+  public static void testFloatEqualsWithAllowedUlps() {
+    assertTrue(floatUlpEquals(0.0f, -0.0f, (short) 1));
+    assertTrue(floatUlpEquals(Float.MIN_VALUE, -0.0f, (short) 1));
+    assertFalse(floatUlpEquals(Float.MIN_VALUE, -Float.MIN_VALUE, (short) 1));
+
+    assertTrue(floatUlpEquals(1.0f, 1f + Math.ulp(1f), (short) 1));
+    assertFalse(floatUlpEquals(1.0f, 1f + 2 * Math.ulp(1f), (short) 1));
+
+    for (float value : new float[] {153.0f, -128.0f, 0.0f, 1.0f}) {
+      assertTrue(floatUlpEquals(value, value, (short) 1));
+      assertTrue(floatUlpEquals(value, Math.nextUp(value), (short) 1));
+      assertFalse(floatUlpEquals(value, Math.nextUp(Math.nextUp(value)), (short) 1));
+      assertTrue(floatUlpEquals(value, Math.nextDown(value), (short) 1));
+      assertFalse(floatUlpEquals(value, Math.nextDown(Math.nextDown(value)), (short) 1));
+      assertFalse(floatUlpEquals(value, value, (short) -1));
+      assertFalse(floatUlpEquals(value, Math.nextUp(value), (short) 0));
+      assertTrue(floatUlpEquals(value, Math.nextUp(Math.nextUp(value)), (short) 2));
+      assertTrue(floatUlpEquals(value, Math.nextDown(Math.nextDown(value)), (short) 2));
+    }
+
+    assertTrue(floatUlpEquals(Float.POSITIVE_INFINITY, Float.POSITIVE_INFINITY, (short) 1));
+    assertTrue(floatUlpEquals(Float.MAX_VALUE, Float.POSITIVE_INFINITY, (short) 1));
+
+    assertTrue(floatUlpEquals(Float.NEGATIVE_INFINITY, Float.NEGATIVE_INFINITY, (short) 1));
+    assertTrue(floatUlpEquals(-Float.MAX_VALUE, Float.NEGATIVE_INFINITY, (short) 1));
+
+    assertFalse(floatUlpEquals(Float.NaN, Float.NaN, (short) 1));
+    assertFalse(floatUlpEquals(Float.NaN, Float.NaN, (short) 0));
+    assertFalse(floatUlpEquals(Float.NaN, 0, (short) 0));
+    assertFalse(floatUlpEquals(0, Float.NaN, (short) 0));
+    assertFalse(floatUlpEquals(Float.NaN, Float.POSITIVE_INFINITY, (short) 0));
+    assertFalse(floatUlpEquals(Float.NaN, Float.NEGATIVE_INFINITY, (short) 0));
+
+    assertFalse(floatUlpEquals(Float.NEGATIVE_INFINITY, Float.POSITIVE_INFINITY, (short) 32767));
+    // The 31-bit integer specification of the max positive ULP allows an extremely
+    // large range of a 23-bit mantissa and 8-bit exponent
+    assertTrue(floatUlpEquals(0, Float.MAX_VALUE, (short) 32767));
+    // Here: f == 2;
+    // it is used to test the maximum ULP distance between two opposite sign numbers.
+    final float f = Float.intBitsToFloat(1 << 30);
+    assertFalse(floatUlpEquals(-f, f, (short) 32767));
+    assertTrue(floatUlpEquals(-f, Math.nextDown(f), (short) 32767));
+    assertTrue(floatUlpEquals(Math.nextUp(-f), f, (short) 32767));
+    // Maximum distance between same sign finite numbers is not possible as the upper
+    // limit is NaN. Check that it is not equal to anything.
+    final float f2 = Float.intBitsToFloat(Integer.MAX_VALUE);
+    assertEquals(Double.NaN, f2, 0);
+    assertFalse(floatUlpEquals(f2, Float.MAX_VALUE, (short) 32767));
+    assertFalse(floatUlpEquals(f2, 0, (short) 32767));
+  }
+}