min max by plugin change (#39)

wip

Signed-off-by: Haoyang Li <[email protected]>

sql-plugin/src/main/java/com/nvidia/spark/rapids/GpuColumnVector.java

@@ -561,6 +561,7 @@ public static ColumnarBatch from(Table table, DataType[] colTypes) {
    */
   static boolean typeConversionAllowed(ColumnView cv, DataType colType) {
     DType dt = cv.getType();
+    System.out.println("Checking " + dt + " vs " + colType);
     if (!dt.isNestedType()) {
       return getNonNestedRapidsType(colType).equals(dt);
     }

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

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2023, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2024, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -359,8 +359,11 @@ class AggHelper(
         val cols = GpuColumnVector.extractColumns(preProcessed)
         val reductionCol = cols(aggOrdinals(ix))
         withResource(aggFn(reductionCol.getBase)) { res =>
+          println("!!!res: " + res)
+          println("!!!cudfAgg.dataType: " + cudfAgg.dataType)
           cvs += GpuColumnVector.from(
             cudf.ColumnVector.fromScalar(res, 1), cudfAgg.dataType)
+            // cudf.ColumnVector.fromScalar(res, 1), IntegerType)
         }
       }
       new ColumnarBatch(cvs.toArray, 1)
@@ -381,6 +384,8 @@ class AggHelper(
             .withKeysSorted(doSortAgg)
             .build()
 
+        println("!!!preProcessedTbl: " + preProcessedTbl)
+
         val cudfAggsOnColumn = cudfAggregates.zip(aggOrdinals).map {
           case (cudfAgg, ord) => cudfAgg.groupByAggregate.onColumn(ord)
         }
@@ -595,19 +600,30 @@ object GpuAggFinalPassIterator {
     val aggBufferAttributes = groupingAttributes ++
         aggregateExpressions.flatMap(_.aggregateFunction.aggBufferAttributes)
 
+    println("groupingAttributes: " + groupingAttributes)
+    println("aggBufferAttributes: " + aggBufferAttributes)
+
+    println("modeInfo: " + modeInfo)
+
     val boundFinalProjections = if (modeInfo.hasFinalMode || modeInfo.hasCompleteMode) {
       val finalProjections = groupingAttributes ++
           aggregateExpressions.map(_.aggregateFunction.evaluateExpression)
+      println("finalProjections: " + finalProjections)
       Some(GpuBindReferences.bindGpuReferences(finalProjections, aggBufferAttributes))
     } else {
       None
     }
-
+    
     // allAttributes can be different things, depending on aggregation mode:
     // - Partial mode: grouping key + cudf aggregates (e.g. no avg, intead sum::count
     // - Final mode: grouping key + spark aggregates (e.g. avg)
     val finalAttributes = groupingAttributes ++ aggregateAttributes
 
+    println("resultExpressions: " + resultExpressions)
+    println("resultExpressions.map(_.toAttribute): " + resultExpressions.map(_.toAttribute))
+    println("finalAttributes: " + finalAttributes)
+    println("groupingAttributes: " + groupingAttributes)
+
     // boundResultReferences is used to project the aggregated input batch(es) for the result.
     // - Partial mode: it's a pass through. We take whatever was aggregated and let it come
     //   out of the node as is.
@@ -636,6 +652,12 @@ object GpuAggFinalPassIterator {
       metrics: GpuHashAggregateMetrics): ColumnarBatch = {
     // Perform the last project to get the correct shape that Spark expects. Note this may
     // add things like literals that were not part of the aggregate into the batch.
+    println("!!!boundExpressions.boundResultReferences: " + boundExpressions.boundResultReferences)
+    // val a = boundExpressions.boundResultReferences(0)
+    // val minBy: AttributeReference = AttributeReference("CudfMinBy", StructType(Seq(
+    //     StructField("_key_value", IntegerType, nullable = true),
+    //     StructField("_key_ordering", LongType, nullable = true))))()
+    // val hardcodeNewResultReferences = Seq(a, minBy)
     closeOnExcept(GpuProjectExec.projectAndClose(finalBatch,
       boundExpressions.boundResultReferences, NoopMetric)) { ret =>
       metrics.numOutputRows += ret.numRows()
@@ -649,13 +671,24 @@ object GpuAggFinalPassIterator {
       metrics: GpuHashAggregateMetrics): Iterator[ColumnarBatch] = {
     val aggTime = metrics.computeAggTime
     val opTime = metrics.opTime
+    println("!!makeIter")
+    println("!!boundExpressions: " + boundExpressions)
+    println("!!boundExpressions.boundFinalProjections: " + boundExpressions.boundFinalProjections)
     cbIter.map { batch =>
       withResource(new NvtxWithMetrics("finalize agg", NvtxColor.DARK_GREEN, aggTime,
         opTime)) { _ =>
         val finalBatch = boundExpressions.boundFinalProjections.map { exprs =>
-          GpuProjectExec.projectAndClose(batch, exprs, NoopMetric)
+          println("## flag 1")
+          val temp = GpuProjectExec.projectAndClose(batch, exprs, NoopMetric)
+          temp
         }.getOrElse(batch)
-        reorderFinalBatch(finalBatch, boundExpressions, metrics)
+        println("!!finalBatch: " + finalBatch)
+        println("!!finalBatch column num: " + finalBatch.numCols())
+        for (i <- 0 until finalBatch.numCols()) {
+          println("!!finalBatch column " + i + ": " + finalBatch.column(i).dataType())
+        }
+        // reorderFinalBatch(finalBatch, boundExpressions, metrics)
+        finalBatch
       }
     }
   }
@@ -1227,6 +1260,8 @@ abstract class GpuBaseAggregateMeta[INPUT <: SparkPlan](
     // the agg-reduction in each task and make a choice there what to do.
 
     lazy val estimatedPreProcessGrowth = {
+      println("!!gpuChild: " + gpuChild)
+      println("!!gpuChild.output: " + gpuChild.output)
       val inputAggBufferAttributes =
         GpuHashAggregateExecBase.calcInputAggBufferAttributes(gpuAggregateExpressions)
       val inputAttrs = GpuHashAggregateExecBase.calcInputAttributes(gpuAggregateExpressions,
@@ -1663,6 +1698,10 @@ object GpuHashAggregateExecBase {
   def calcInputAttributes(aggregateExpressions: Seq[GpuAggregateExpression],
                           childOutput: Seq[Attribute],
                           inputAggBufferAttributes: Seq[Attribute]): Seq[Attribute] = {
+    println("!!calcInputAttributes")
+    println("aggregateExpressions: " + aggregateExpressions)
+    println("childOutput: " + childOutput)
+    println("inputAggBufferAttributes: " + inputAggBufferAttributes)
     val modes = aggregateExpressions.map(_.mode).distinct
     if (modes.contains(Final) || modes.contains(PartialMerge)) {
       // SPARK-31620: when planning aggregates, the partial aggregate uses aggregate function's
@@ -1797,6 +1836,7 @@ case class GpuHashAggregateExec(
     val boundGroupExprs = GpuBindReferences.bindGpuReferencesTiered(groupingExprs, inputAttrs, true)
 
     rdd.mapPartitions { cbIter =>
+      println("resultExprs: " + resultExprs)
       val postBoundReferences = GpuAggFinalPassIterator.setupReferences(groupingExprs,
         aggregateExprs, aggregateAttrs, resultExprs, modeInfo)
 
@@ -1974,6 +2014,9 @@ class DynamicGpuPartialSortAggregateIterator(
         outputRows = NoopMetric)
     }
 
+    println("!!inputAttrs" + inputAttrs)
+    println("!!aggregateAttrs" + aggregateAttrs)
+
     // After sorting we want to split the input for the project so that
     // we don't get ourselves in trouble.
     val sortedSplitIter = new PreProjectSplitIterator(sortedIter,
@@ -1992,6 +2035,10 @@ class DynamicGpuPartialSortAggregateIterator(
   private[this] def fullHashAggWithMerge(
       inputIter: Iterator[ColumnarBatch],
       preProcessAggHelper: AggHelper): Iterator[ColumnarBatch] = {
+
+    println("!!inputAttrs: " + inputAttrs)
+    println("!!aggregateAttrs: " + aggregateAttrs)
+
     // We still want to split the input, because the heuristic may not be perfect and
     //  this is relatively light weight
     val splitInputIter = new PreProjectSplitIterator(inputIter,

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

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2023, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2024, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -176,7 +176,17 @@ case class GpuBoundReference(ordinal: Int, dataType: DataType, nullable: Boolean
     s"input[$ordinal, ${dataType.simpleString}, $nullable]($name#${exprId.id})"
 
   override def columnarEval(batch: ColumnarBatch): GpuColumnVector = {
-    batch.column(ordinal) match {
+    println("!!!GpuBoundReference" + toString)
+    println("!!!ordinal: " + ordinal)
+    val xx = if (ordinal >= batch.numCols()) {
+      println("!!!GpuBoundReference" + toString + "ordinal: " + ordinal + "batch.numCols(): " + batch.numCols())
+      batch.numCols() - 1
+    } else {
+      println("!!!ordinal: " + ordinal)
+      ordinal
+    }
+    println("!!!xx: " + xx + "batch.numCols(): " + batch.numCols() + "ordinal: " + ordinal)
+    batch.column(xx) match {
       case fb: GpuColumnVectorFromBuffer =>
         // When doing a project we might re-order columns or do other things that make it
         // so this no longer looks like the original contiguous buffer it came from

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

@@ -244,6 +244,7 @@ abstract class AbstractProjectSplitIterator(iter: Iterator[ColumnarBatch],
         }
       }
     } else {
+      println("!!!AbstractProjectSplitIterator schema: " + schema.mkString(", "))
       val cb = iter.next()
       opTime.ns {
         val numSplits = closeOnExcept(cb) { cb =>
@@ -286,16 +287,19 @@ abstract class AbstractProjectSplitIterator(iter: Iterator[ColumnarBatch],
 object PreProjectSplitIterator {
   def calcMinOutputSize(cb: ColumnarBatch, boundExprs: GpuTieredProject): Long = {
     val numRows = cb.numRows()
+    println("boundExprs.outputTypes: " + boundExprs.outputTypes)
+    println("boundExprs.outputTypes: " + boundExprs.outputTypes.size)
     boundExprs.outputTypes.zipWithIndex.map {
-      case (dataType, index) =>
+      case (dataType, _) => // index
         if (GpuBatchUtils.isFixedWidth(dataType)) {
           GpuBatchUtils.minGpuMemory(dataType, true, numRows)
         } else {
-          boundExprs.getPassThroughIndex(index).map { inputIndex =>
-            cb.column(inputIndex).asInstanceOf[GpuColumnVector].getBase.getDeviceMemorySize
-          }.getOrElse {
-            GpuBatchUtils.minGpuMemory(dataType, true, numRows)
-          }
+          // java.lang.ArrayIndexOutOfBoundsException: 2 related to getPassThroughIndex
+          // boundExprs.getPassThroughIndex(index).map { inputIndex =>
+          //   cb.column(inputIndex).asInstanceOf[GpuColumnVector].getBase.getDeviceMemorySize
+          // }.getOrElse {
+          GpuBatchUtils.minGpuMemory(dataType, true, numRows)
+          // }
         }
     }.sum
   }

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

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022-2023, NVIDIA CORPORATION.
+ * Copyright (c) 2022-2024, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -21,7 +21,7 @@ import ai.rapids.cudf.JCudfSerialization.HostConcatResult
 import com.nvidia.spark.rapids.{GpuColumnVectorFromBuffer, RmmRapidsRetryIterator}
 import com.nvidia.spark.rapids.Arm.withResource
 
-import org.apache.spark.sql.types.DataType
+import org.apache.spark.sql.types._
 import org.apache.spark.sql.vectorized.ColumnarBatch
 
 object HostConcatResultUtil {
@@ -50,6 +50,10 @@ object HostConcatResultUtil {
       // that it is acquired in the coalesce code.
       new ColumnarBatch(Array.empty, hostConcatResult.getTableHeader.getNumRows)
     } else {
+      println("!!!sparkSchema: " + sparkSchema.mkString(", "))
+      // val hardCodeSparkSchema = Array(IntegerType, StructType(Seq(
+      //   StructField("_key_value", IntegerType, nullable = true),
+      //   StructField("_key_ordering", LongType, nullable = true))).asInstanceOf[DataType])
       RmmRapidsRetryIterator.withRetryNoSplit {
         withResource(hostConcatResult.toContiguousTable) { ct =>
           GpuColumnVectorFromBuffer.from(ct, sparkSchema)

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

@@ -2036,15 +2036,14 @@ abstract class CudfMaxMinByAggregate(
   protected val sortOrder: Int => cudf.OrderByArg
 
   // This is a short term solution. and better to have a dedicate reduction for this.
-  override val reductionAggregate: cudf.ColumnVector => cudf.Scalar = col => {
-    val tmpTable = withResource(col) { _ =>
-      val children = Seq(0, 1).safeMap { idx =>
-        col.getChildColumnView(idx).copyToColumnVector()
-      }
-      withResource(children) { _ =>
-        new cudf.Table(children: _*)
-      }
+  override lazy val reductionAggregate: cudf.ColumnVector => cudf.Scalar = col => {
+    val children = Seq(0, 1).safeMap { idx =>
+      col.getChildColumnView(idx).copyToColumnVector()
+    }
+    val tmpTable = withResource(children) { _ =>
+      new cudf.Table(children: _*)
     }
+    // }
     val sorted = withResource(tmpTable) { _ =>
       // columns in table [value column, ordering column]
       tmpTable.orderBy(sortOrder(1))
@@ -2056,50 +2055,85 @@ abstract class CudfMaxMinByAggregate(
   }
 
   override val dataType: DataType = StructType(Seq(
-    StructField(CudfMaxMinBy.KEY_VALUE, valueType),
-    StructField(CudfMaxMinBy.KEY_ORDERING, orderingType)))
+    StructField(CudfMaxMinBy.KEY_VALUE, valueType, nullable = true),
+    StructField(CudfMaxMinBy.KEY_ORDERING, orderingType, nullable = true)))
 }
 
 class CudfMaxBy(valueType: DataType, orderingType: DataType)
   extends CudfMaxMinByAggregate(valueType, orderingType) {
 
   override val name: String = "CudfMaxBy"
   override val sortOrder: Int => cudf.OrderByArg = cudf.OrderByArg.desc
-  // TODO
-  override val groupByAggregate: GroupByAggregation = null
+  override lazy val groupByAggregate: GroupByAggregation = null
 }
 
 class CudfMinBy(valueType: DataType, orderingType: DataType)
   extends CudfMaxMinByAggregate(valueType, orderingType) {
 
   override val name: String = "CudfMinBy"
   override val sortOrder: Int => cudf.OrderByArg = cudf.OrderByArg.asc
-  // TODO
-  override val groupByAggregate: GroupByAggregation = null
+  override lazy val groupByAggregate: GroupByAggregation = GroupByAggregation.minBy()
 }
 
 abstract class GpuMaxMinByBase(valueExpr: Expression, orderingExpr: Expression)
   extends GpuAggregateFunction with Serializable {
 
   protected val cudfMaxMinByAggregate: CudfAggregate
 
-  override val initialValues: Seq[Expression] = Seq(
+  override lazy val initialValues: Seq[Expression] = Seq(
     GpuLiteral.default(cudfMaxMinByAggregate.dataType))
 
-  override val inputProjection: Seq[Expression] = Seq(
+  protected lazy val bufferValue: AttributeReference =
+    AttributeReference("value", valueExpr.dataType)()
+
+  protected lazy val bufferOrdering: AttributeReference =
+    AttributeReference("ordering", orderingExpr.dataType)()
+
+  protected lazy val bufferStruct: AttributeReference =
+    AttributeReference(cudfMaxMinByAggregate.name, StructType(Seq(
+      StructField(CudfMaxMinBy.KEY_VALUE, valueExpr.dataType, nullable = true),
+      StructField(CudfMaxMinBy.KEY_ORDERING, orderingExpr.dataType, nullable = true))))()
+
+  override lazy val inputProjection: Seq[Expression] = Seq(
     GpuCreateNamedStruct(Seq(
-      GpuLiteral(CudfMaxMinBy.KEY_VALUE, StringType), valueExpr,
-      GpuLiteral(CudfMaxMinBy.KEY_ORDERING, StringType), orderingExpr
+      GpuLiteral(CudfMaxMinBy.KEY_VALUE, StringType), bufferValue,
+      GpuLiteral(CudfMaxMinBy.KEY_ORDERING, StringType), bufferOrdering
     )))
+  // override lazy val inputProjection: Seq[Expression] = Seq(
+  //   valueExpr, orderingExpr)
 
-  override val updateAggregates: Seq[CudfAggregate] = Seq(cudfMaxMinByAggregate)
-  override val mergeAggregates: Seq[CudfAggregate] = Seq(cudfMaxMinByAggregate)
-  override val evaluateExpression: Expression = GpuGetStructField(
+  override def aggBufferAttributes: Seq[AttributeReference] = bufferValue :: Nil
+
+  override lazy val postUpdate: Seq[Expression] = {
+    Seq(
+      GpuGetStructField(cudfMaxMinByAggregate.attr, ordinal = 0, name = Some(CudfMaxMinBy.KEY_VALUE)),
+      GpuGetStructField(cudfMaxMinByAggregate.attr, ordinal = 1, name = Some(CudfMaxMinBy.KEY_ORDERING))
+    )
+  }
+
+  override lazy val updateAggregates: Seq[CudfAggregate] = Seq(cudfMaxMinByAggregate)
+  override lazy val mergeAggregates: Seq[CudfAggregate] = Seq(cudfMaxMinByAggregate)
+  override lazy val evaluateExpression: Expression = GpuGetStructField(
     cudfMaxMinByAggregate.attr, ordinal = 0, name = Some(CudfMaxMinBy.KEY_VALUE))
 
-  override def aggBufferAttributes: Seq[AttributeReference] = Seq(cudfMaxMinByAggregate.attr)
+  override lazy val preMerge: Seq[Expression] = {
+    println("!!!!!!!!preMerge")
+    val childrenWithNames = Seq(
+      GpuLiteral(CudfMaxMinBy.KEY_VALUE, StringType), bufferValue,
+      GpuLiteral(CudfMaxMinBy.KEY_ORDERING, StringType), bufferOrdering
+    )
+    GpuCreateNamedStruct(childrenWithNames) :: Nil
+  }
+
+  override lazy val postMerge: Seq[Expression] = {
+    println("!!!!!!!!postMerge")
+    Seq(
+    GpuGetStructField(cudfMaxMinByAggregate.attr, ordinal = 0, name = Some(CudfMaxMinBy.KEY_VALUE)),
+    GpuGetStructField(cudfMaxMinByAggregate.attr, ordinal = 1, name = Some(CudfMaxMinBy.KEY_ORDERING))
+  )
+  }
 
-  override def children: Seq[Expression] = IndexedSeq(valueExpr, orderingExpr)
+  override def children: Seq[Expression] = Seq(valueExpr, orderingExpr)
 
   override def nullable: Boolean = true
 

tools/generated_files/330/operatorsScore.csv

@@ -182,6 +182,7 @@ Md5,4
 MicrosToTimestamp,4
 MillisToTimestamp,4
 Min,4
+MinBy,4
 Minute,4
 MonotonicallyIncreasingID,4
 Month,4

tools/generated_files/330/supportedExprs.csv

@@ -712,6 +712,15 @@ Min,S,`min`,None,reduction,input,S,S,S,S,S,S,S,S,PS,S,S,S,NS,NS,PS,NA,PS,NS,NA,N
 Min,S,`min`,None,reduction,result,S,S,S,S,S,S,S,S,PS,S,S,S,NS,NS,PS,NA,PS,NS,NA,NA
 Min,S,`min`,None,window,input,S,S,S,S,S,S,S,S,PS,S,S,S,NS,NS,NS,NA,NS,NS,NA,NA
 Min,S,`min`,None,window,result,S,S,S,S,S,S,S,S,PS,S,S,S,NS,NS,NS,NA,NS,NS,NA,NA
+MinBy,S,`min_by`,None,aggregation,value,S,S,S,S,S,S,S,S,PS,S,S,S,S,NS,PS,PS,PS,NS,NS,NS
+MinBy,S,`min_by`,None,aggregation,ordering,S,S,S,S,S,S,S,S,PS,S,PS,S,NS,NS,NS,NA,PS,NS,NA,NA
+MinBy,S,`min_by`,None,aggregation,result,S,S,S,S,S,S,S,S,PS,S,S,S,S,NS,PS,PS,PS,NS,NS,NS
+MinBy,S,`min_by`,None,reduction,value,S,S,S,S,S,S,S,S,PS,S,S,S,S,NS,PS,PS,PS,NS,NS,NS
+MinBy,S,`min_by`,None,reduction,ordering,S,S,S,S,S,S,S,S,PS,S,PS,S,NS,NS,NS,NA,PS,NS,NA,NA
+MinBy,S,`min_by`,None,reduction,result,S,S,S,S,S,S,S,S,PS,S,S,S,S,NS,PS,PS,PS,NS,NS,NS
+MinBy,S,`min_by`,None,window,value,S,S,S,S,S,S,S,S,PS,S,S,S,S,NS,PS,PS,PS,NS,NS,NS
+MinBy,S,`min_by`,None,window,ordering,S,S,S,S,S,S,S,S,PS,S,PS,S,NS,NS,NS,NA,PS,NS,NA,NA
+MinBy,S,`min_by`,None,window,result,S,S,S,S,S,S,S,S,PS,S,S,S,S,NS,PS,PS,PS,NS,NS,NS
 Percentile,S,`percentile`,None,aggregation,input,NA,S,S,S,S,S,S,NA,NA,NA,NA,NA,NA,NA,NA,NA,NA,NA,NA,NA
 Percentile,S,`percentile`,None,aggregation,percentage,NA,NA,NA,NA,NA,NA,PS,NA,NA,NA,NA,NA,NA,NA,S,NA,NA,NA,NA,NA
 Percentile,S,`percentile`,None,aggregation,frequency,NA,NA,NA,NA,S,NA,NA,NA,NA,NA,NA,NA,NA,NA,S,NA,NA,NA,NA,NA