Better float/double cases for casting tests (#1781)

* enhance float/double cases for casting tests Signed-off-by: sperlingxx <lovedreamf@gmail.com> * continue * code clean * code clean * fix typo * fix typo * some updates * fix typo
NVIDIA · Feb 23, 2021 · 52d95b1 · 52d95b1
1 parent f45a3b7
commit 52d95b1
Show file tree

Hide file tree

Showing 4 changed files with 38 additions and 205 deletions.
diff --git a/tests/src/test/scala/com/nvidia/spark/rapids/AnsiCastOpSuite.scala b/tests/src/test/scala/com/nvidia/spark/rapids/AnsiCastOpSuite.scala
@@ -338,31 +338,31 @@ class AnsiCastOpSuite extends GpuExpressionTestSuite {
     frame => testCastTo(DataTypes.DoubleType)(frame)
   }
 
-  testCastFailsForBadInputs("Test bad cast 1 from strings to floats", badFloatStringsDf,
+  testCastFailsForBadInputs("Test bad cast 1 from strings to floats", invalidFloatStringsDf,
       msg = GpuCast.INVALID_FLOAT_CAST_MSG) {
     frame =>frame.select(col("c0").cast(FloatType))
   }
 
-  testCastFailsForBadInputs("Test bad cast 2 from strings to floats", badFloatStringsDf,
+  testCastFailsForBadInputs("Test bad cast 2 from strings to floats", invalidFloatStringsDf,
       msg = GpuCast.INVALID_FLOAT_CAST_MSG) {
     frame =>frame.select(col("c1").cast(FloatType))
   }
 
-  testCastFailsForBadInputs("Test bad cast 1 from strings to double", badFloatStringsDf,
+  testCastFailsForBadInputs("Test bad cast 1 from strings to double", invalidFloatStringsDf,
       msg = GpuCast.INVALID_FLOAT_CAST_MSG) {
     frame =>frame.select(col("c0").cast(DoubleType))
   }
 
-  testCastFailsForBadInputs("Test bad cast 2 from strings to double", badFloatStringsDf,
+  testCastFailsForBadInputs("Test bad cast 2 from strings to double", invalidFloatStringsDf,
       msg = GpuCast.INVALID_FLOAT_CAST_MSG) {
     frame =>frame.select(col("c1").cast(DoubleType))
   }
 
-  //Currently there is a bug in cudf which doesn't convert one value correctly
+  // Currently there is a bug in cudf which doesn't convert some corner cases correctly
   // The bug is documented here https://github.com/rapidsai/cudf/issues/5225
   ignore("Test cast from strings to double that doesn't match") {
     testSparkResultsAreEqual("Test cast from strings to double that doesn't match",
-        badDoubleStringsDf) {
+        badDoubleStringsDf, conf = sparkConf, maxFloatDiff = 0.0001) {
       frame =>frame.select(
         col("c0").cast(DoubleType))
     }
@@ -372,12 +372,12 @@ class AnsiCastOpSuite extends GpuExpressionTestSuite {
   // Ansi cast from floating point to string
   ///////////////////////////////////////////////////////////////////////////
 
-  ignore("ansi_cast float to string") {
+  test("ansi_cast float to string") {
     testCastToString[Float](DataTypes.FloatType, ansiMode = true,
       comparisonFunc = Some(compareStringifiedFloats))
   }
 
-  ignore("ansi_cast double to string") {
+  test("ansi_cast double to string") {
     testCastToString[Double](DataTypes.DoubleType, ansiMode = true,
       comparisonFunc = Some(compareStringifiedFloats))
   }

diff --git a/tests/src/test/scala/com/nvidia/spark/rapids/CastOpSuite.scala b/tests/src/test/scala/com/nvidia/spark/rapids/CastOpSuite.scala
@@ -203,16 +203,6 @@ class CastOpSuite extends GpuExpressionTestSuite {
     }
   }
 
-  private def testCastTo(castTo: DataType)(frame: DataFrame): DataFrame ={
-    frame.withColumn("c1", col("c0").cast(castTo))
-  }
-
-  private def stringDf(str: String)(session: SparkSession): DataFrame = {
-    import session.sqlContext.implicits._
-    // use more than one value otherwise spark optimizes it out as a literal
-    Seq(str, str).toDF("c0")
-  }
-
   private def castToStringExpectedFun[T]: T => Option[String] = (d: T) => Some(String.valueOf(d))
 
   test("cast byte to string") {
@@ -231,11 +221,11 @@ class CastOpSuite extends GpuExpressionTestSuite {
     testCastToString[Long](DataTypes.LongType)
   }
 
-  ignore("cast float to string") {
+  test("cast float to string") {
     testCastToString[Float](DataTypes.FloatType, comparisonFunc = Some(compareStringifiedFloats))
   }
 
-  ignore("cast double to string") {
+  test("cast double to string") {
     testCastToString[Double](DataTypes.DoubleType, comparisonFunc = Some(compareStringifiedFloats))
   }
 
@@ -298,30 +288,6 @@ class CastOpSuite extends GpuExpressionTestSuite {
       col("doubles").cast(TimestampType))
   }
 
-  ignore("Test cast from double to string") {
-
-    //NOTE that the testSparkResultsAreEqual method isn't adequate in this case because we
-    // need to use a specialized comparison function
-
-    val conf = new SparkConf()
-      .set(RapidsConf.ENABLE_CAST_FLOAT_TO_STRING.key, "true")
-
-    val (cpu, gpu) = runOnCpuAndGpu(doubleDf, frame => frame.select(
-      col("doubles").cast(StringType))
-      .orderBy(col("doubles")), conf)
-
-    val fromCpu = cpu.map(row => row.getAs[String](0))
-    val fromGpu = gpu.map(row => row.getAs[String](0))
-
-    fromCpu.zip(fromGpu).foreach {
-      case (c, g) =>
-        if (!compareStringifiedFloats(c, g)) {
-          fail(s"Running on the GPU and on the CPU did not match: CPU value: $c. " +
-            s"GPU value: $g.")
-        }
-    }
-  }
-
   testSparkResultsAreEqual("Test cast from boolean", booleanDf) {
     frame => frame.select(
       col("bools").cast(IntegerType),
@@ -396,14 +362,6 @@ class CastOpSuite extends GpuExpressionTestSuite {
       col("doubles").cast(TimestampType))
   }
 
-  ignore("Test cast from strings to double that doesn't match") {
-        testSparkResultsAreEqual("Test cast from strings to double that doesn't match",
-          badDoubleStringsDf) {
-          frame =>frame.select(
-              col("doubles").cast(DoubleType))
-        }
-  }
-
   testSparkResultsAreEqual("Test cast from strings to doubles", doublesAsStrings,
     conf = sparkConf, maxFloatDiff = 0.0001) {
     frame => frame.select(
@@ -416,7 +374,7 @@ class CastOpSuite extends GpuExpressionTestSuite {
       col("c0").cast(FloatType))
   }
 
-  testSparkResultsAreEqual("Test bad cast from strings to floats", badFloatStringsDf,
+  testSparkResultsAreEqual("Test bad cast from strings to floats", invalidFloatStringsDf,
     conf = sparkConf, maxFloatDiff = 0.0001) {
     frame =>frame.select(
       col("c0").cast(DoubleType),
@@ -425,6 +383,16 @@ class CastOpSuite extends GpuExpressionTestSuite {
       col("c1").cast(FloatType))
   }
 
+  // Currently there is a bug in cudf which doesn't convert some corner cases correctly
+  // The bug is documented here https://github.com/rapidsai/cudf/issues/5225
+  ignore("Test cast from strings to double that doesn't match") {
+    testSparkResultsAreEqual("Test cast from strings to double that doesn't match",
+        badDoubleStringsDf, conf = sparkConf, maxFloatDiff = 0.0001) {
+      frame => frame.select(
+        col("c0").cast(DoubleType))
+    }
+  }
+
   testSparkResultsAreEqual("ansi_cast string to double exp", exponentsAsStringsDf,
     conf = sparkConf, maxFloatDiff = 0.0001) {
     frame => frame.select(
@@ -720,13 +688,13 @@ object CastOpSuite {
 
   def doublesAsStrings(session: SparkSession): DataFrame = {
     val schema = FuzzerUtils.createSchema(Seq(DoubleType), false)
-    val df = FuzzerUtils.generateDataFrame(session, schema, 100)
+    val df = FuzzerUtils.generateDataFrame(session, schema, 2048)
     df.withColumn("c0", col("c0").cast(StringType))
   }
 
   def floatsAsStrings(session: SparkSession): DataFrame = {
     val schema = FuzzerUtils.createSchema(Seq(FloatType), false)
-    val df = FuzzerUtils.generateDataFrame(session, schema, 100)
+    val df = FuzzerUtils.generateDataFrame(session, schema, 2048)
     df.withColumn("c0", col("c0").cast(StringType))
   }
 

diff --git a/tests/src/test/scala/com/nvidia/spark/rapids/FuzzerUtils.scala b/tests/src/test/scala/com/nvidia/spark/rapids/FuzzerUtils.scala
@@ -301,30 +301,34 @@ class EnhancedRandom(protected val r: Random, protected val options: FuzzerOptio
   }
 
   def nextFloat(): Float = {
-    r.nextInt(9) match {
+    r.nextInt(11) match {
       case 0 => Float.NaN
       case 1 => Float.PositiveInfinity
       case 2 => Float.NegativeInfinity
-      case 3 => r.nextFloat() * Float.MinValue
-      case 4 => r.nextFloat() * Float.MaxValue
+      case 3 => Float.MinValue
+      case 4 => Float.MaxValue
       case 5 => 0 - r.nextFloat()
       case 6 => r.nextFloat()
       case 7 => 0f
       case 8 => -0f
+      case 9 => r.nextFloat() * Float.MinValue
+      case 10 => r.nextFloat() * Float.MaxValue
     }
   }
 
   def nextDouble(): Double = {
-    r.nextInt(9) match {
+    r.nextInt(11) match {
       case 0 => Double.NaN
       case 1 => Double.PositiveInfinity
       case 2 => Double.NegativeInfinity
-      case 3 => r.nextDouble() * Double.MinValue
-      case 4 => r.nextDouble() * Double.MaxValue
+      case 3 => Double.MaxValue
+      case 4 => Double.MinValue
       case 5 => 0 - r.nextDouble()
       case 6 => r.nextDouble()
       case 7 => 0d
       case 8 => -0d
+      case 9 => r.nextDouble() * Double.MinValue
+      case 10 => r.nextDouble() * Double.MaxValue
     }
   }