Support HyperLogLog++

Signed-off-by: Chong Gao <[email protected]>

integration_tests/src/main/python/arithmetic_ops_test.py

@@ -98,6 +98,15 @@ def _get_overflow_df(spark, data, data_type, expr):
         StructType([StructField('a', data_type)])
     ).selectExpr(expr)
 
+def test_hll():
+    assert_gpu_and_cpu_are_equal_sql(
+        lambda spark : spark.read.parquet("/home/chongg/a"),
+        "tab",
+        "select c1, APPROX_COUNT_DISTINCT(c2) from tab group by c1",
+        {"spark.rapids.sql.agg.forceSinglePassPartialSort": False,
+         "spark.rapids.sql.agg.singlePassPartialSortEnabled": False}
+    )
+
 @pytest.mark.parametrize('data_gen', _arith_data_gens, ids=idfn)
 @disable_ansi_mode
 def test_addition(data_gen):

sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuOverrides.scala

@@ -3918,6 +3918,16 @@ object GpuOverrides extends Logging {
           GpuDynamicPruningExpression(child)
         }
       }),
+    expr[HyperLogLogPlusPlus](
+      "Aggregation approximate count distinct",
+      ExprChecks.reductionAndGroupByAgg(TypeSig.LONG, TypeSig.LONG,
+        Seq(ParamCheck("input", TypeSig.all, TypeSig.all))),
+      (a, conf, p, r) => new UnaryExprMeta[HyperLogLogPlusPlus](a, conf, p, r) {
+        override def convertToGpu(child: Expression): GpuExpression = {
+          GpuHyperLogLogPlusPlus(child, a.relativeSD)
+        }
+      }
+    ),
     SparkShimImpl.ansiCastRule
   ).collect { case r if r != null => (r.getClassFor.asSubclass(classOf[Expression]), r)}.toMap
 

sql-plugin/src/main/scala/org/apache/spark/sql/rapids/aggregate/GpuHyperLogLogPlusPlus.scala

@@ -0,0 +1,167 @@
+/*
+ * Copyright (c) 2023, NVIDIA CORPORATION.
+ *
+ * Licensed 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.spark.sql.rapids.aggregate
+
+import scala.collection.immutable.Seq
+
+import ai.rapids.cudf
+import ai.rapids.cudf.{DType, GroupByAggregation, ReductionAggregation}
+import com.nvidia.spark.rapids._
+import com.nvidia.spark.rapids.Arm.withResourceIfAllowed
+import com.nvidia.spark.rapids.RapidsPluginImplicits.ReallyAGpuExpression
+import com.nvidia.spark.rapids.jni.HLL
+import com.nvidia.spark.rapids.shims.ShimExpression
+
+import org.apache.spark.sql.catalyst.expressions.{AttributeReference, Expression}
+import org.apache.spark.sql.catalyst.util.HyperLogLogPlusPlusHelper
+import org.apache.spark.sql.rapids.{GpuCreateNamedStruct, GpuGetStructField}
+import org.apache.spark.sql.types._
+import org.apache.spark.sql.vectorized.ColumnarBatch
+
+case class CudfHLLPP(override val dataType: DataType,
+    numRegistersPerSketch: Int) extends CudfAggregate {
+  override lazy val reductionAggregate: cudf.ColumnVector => cudf.Scalar =
+    (input: cudf.ColumnVector) => input.reduce(
+      ReductionAggregation.HLL(numRegistersPerSketch), DType.STRUCT)
+  override lazy val groupByAggregate: GroupByAggregation =
+    GroupByAggregation.HLL(numRegistersPerSketch)
+  override val name: String = "CudfHLLPP"
+}
+
+case class CudfMergeHLLPP(override val dataType: DataType,
+    numRegistersPerSketch: Int)
+    extends CudfAggregate {
+  override lazy val reductionAggregate: cudf.ColumnVector => cudf.Scalar =
+    (input: cudf.ColumnVector) =>
+      input.reduce(ReductionAggregation.mergeHLL(numRegistersPerSketch), DType.STRUCT)
+  override lazy val groupByAggregate: GroupByAggregation =
+    GroupByAggregation.mergeHLL(numRegistersPerSketch)
+  override val name: String = "CudfMergeHLLPP"
+}
+
+/**
+ * Perform the final evaluation step to compute approximate count distinct from sketches.
+ * Input is long columns, first construct struct of long then feed to cuDF
+ */
+case class GpuHyperLogLogPlusPlusEvaluation(childExpr: Expression,
+    numRegistersPerSketch: Int)
+    extends GpuExpression with ShimExpression {
+  override def dataType: DataType = LongType
+
+  override def prettyName: String = "HLLPP_evaluation"
+
+  override def nullable: Boolean = false
+
+  override def children: scala.Seq[Expression] = Seq(childExpr)
+
+  override def columnarEval(batch: ColumnarBatch): GpuColumnVector = {
+    withResourceIfAllowed(childExpr.columnarEval(batch)) { sketches =>
+      val distinctValues = HLL.estimateDistinctValueFromSketches(
+      sketches.getBase, numRegistersPerSketch)
+      GpuColumnVector.from(distinctValues, LongType)
+    }
+  }
+}
+
+/**
+ * Gpu version of HyperLogLogPlusPlus
+ * Spark APPROX_COUNT_DISTINCT on NULLs returns zero
+ */
+case class GpuHyperLogLogPlusPlus(childExpr: Expression, relativeSD: Double)
+    extends GpuAggregateFunction with Serializable {
+
+  // Refer to Spark utility
+  private val hllppHelper = new HyperLogLogPlusPlusHelper(relativeSD)
+
+  // Consistent with Spark
+  private lazy val numRegistersPerSketch: Int =
+    1 << Math.ceil(2.0d * Math.log(1.106d / relativeSD) / Math.log(2.0d)).toInt
+
+  // Spark agg buffer type: long array
+  private lazy val sparkAggBufferAttributes: Seq[AttributeReference] = {
+    Seq.tabulate(hllppHelper.numWords) { i =>
+      AttributeReference(s"MS[$i]", LongType)()
+    }
+  }
+
+  /**
+   * Spark uses long columns to save agg buffer, e.g.: long[52]
+   * Each long compacts multiple registers to save memory
+   */
+  override val aggBufferAttributes: Seq[AttributeReference] = sparkAggBufferAttributes
+
+  /**
+   * init long array with all zero
+   */
+  override lazy val initialValues: Seq[Expression] = Seq.tabulate(hllppHelper.numWords) { _ =>
+    GpuLiteral(0L, LongType)
+  }
+
+  override lazy val inputProjection: Seq[Expression] = Seq(childExpr)
+
+  /**
+   * cuDF HLLPP sketch type: struct<long, ..., long>
+   */
+  private lazy val cuDFBufferType: DataType = StructType.fromAttributes(aggBufferAttributes)
+
+  /**
+   * cuDF uses Struct<long, ..., long> column to do aggregate
+   */
+  override lazy val updateAggregates: Seq[CudfAggregate] =
+    Seq(CudfHLLPP(cuDFBufferType, numRegistersPerSketch))
+
+  /**
+   * Convert long columns to Struct<long, ..., long> column
+   */
+  private def genStruct: Seq[Expression] = {
+    val names = Seq.tabulate(hllppHelper.numWords) { i => GpuLiteral(s"MS[$i]", StringType) }
+    Seq(GpuCreateNamedStruct(names.zip(aggBufferAttributes).flatten { case (a, b) => List(a, b) }))
+  }
+
+  /**
+   * Convert Struct<long, ..., long> column to long columns
+   */
+  override lazy val postUpdate: Seq[Expression] = Seq.tabulate(hllppHelper.numWords) {
+    i => GpuGetStructField(postUpdateAttr.head, i)
+  }
+
+  /**
+   * convert to Struct<long, ..., long>
+   */
+  override lazy val preMerge: Seq[Expression] = genStruct
+
+  override lazy val mergeAggregates: Seq[CudfAggregate] =
+    Seq(CudfMergeHLLPP(cuDFBufferType, numRegistersPerSketch))
+
+  /**
+   * Convert Struct<long, ..., long> column to long columns
+   */
+  override lazy val postMerge: Seq[Expression] = Seq.tabulate(hllppHelper.numWords) {
+    i => GpuGetStructField(postMergeAttr.head, i)
+  }
+
+  override lazy val evaluateExpression: Expression =
+    GpuHyperLogLogPlusPlusEvaluation(genStruct.head, numRegistersPerSketch)
+
+  override def dataType: DataType = LongType
+
+  override def prettyName: String = "approx_count_distinct"
+
+  override def nullable: Boolean = true
+
+  override def children: Seq[Expression] = Seq(childExpr)
+}