Preliminary support for keeping broadcast exchanges on GPU when AQE i…

…s enabled (NVIDIA#448)

shims/spark300/src/main/scala/com/nvidia/spark/rapids/shims/spark300/Spark300Shims.scala

@@ -22,10 +22,12 @@ import com.nvidia.spark.rapids._
 import com.nvidia.spark.rapids.spark300.RapidsShuffleManager
 
 import org.apache.spark.SparkEnv
+import org.apache.spark.sql.{SparkSession, SparkSessionExtensions}
 import org.apache.spark.sql.catalyst.encoders.ExpressionEncoder
 import org.apache.spark.sql.catalyst.expressions._
 import org.apache.spark.sql.catalyst.expressions.aggregate.{First, Last}
 import org.apache.spark.sql.catalyst.plans.JoinType
+import org.apache.spark.sql.catalyst.rules.Rule
 import org.apache.spark.sql.execution._
 import org.apache.spark.sql.execution.datasources.HadoopFsRelation
 import org.apache.spark.sql.execution.joins.{BroadcastHashJoinExec, BroadcastNestedLoopJoinExec, HashJoin, SortMergeJoinExec}
@@ -193,4 +195,11 @@ class Spark300Shims extends SparkShims {
   override def getRapidsShuffleManagerClass: String = {
     classOf[RapidsShuffleManager].getCanonicalName
   }
+
+  override def injectQueryStagePrepRule(
+      extensions: SparkSessionExtensions,
+      ruleBuilder: SparkSession => Rule[SparkPlan]): Unit = {
+    // not supported in 3.0.0 but it doesn't matter because AdaptiveSparkPlanExec in 3.0.0 will
+    // never allow us to replace an Exchange node, so they just stay on CPU
+  }
 }

shims/spark301/src/main/scala/com/nvidia/spark/rapids/shims/spark301/Spark301Shims.scala

@@ -20,8 +20,11 @@ import com.nvidia.spark.rapids._
 import com.nvidia.spark.rapids.shims.spark300.Spark300Shims
 import com.nvidia.spark.rapids.spark301.RapidsShuffleManager
 
+import org.apache.spark.sql.{SparkSession, SparkSessionExtensions}
 import org.apache.spark.sql.catalyst.expressions._
 import org.apache.spark.sql.catalyst.expressions.aggregate.{First, Last}
+import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.execution.SparkPlan
 
 class Spark301Shims extends Spark300Shims {
 
@@ -47,4 +50,10 @@ class Spark301Shims extends Spark300Shims {
   override def getRapidsShuffleManagerClass: String = {
     classOf[RapidsShuffleManager].getCanonicalName
   }
+
+  override def injectQueryStagePrepRule(
+      extensions: SparkSessionExtensions,
+      ruleBuilder: SparkSession => Rule[SparkPlan]): Unit = {
+    extensions.injectQueryStagePrepRule(ruleBuilder)
+  }
 }

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

@@ -1744,6 +1744,16 @@ object GpuOverrides {
   val execs: Map[Class[_ <: SparkPlan], ExecRule[_ <: SparkPlan]] =
     commonExecs ++ ShimLoader.getSparkShims.getExecs
 }
+/** Tag the initial plan when AQE is enabled */
+case class GpuQueryStagePrepOverrides() extends Rule[SparkPlan] with Logging {
+  override def apply(plan: SparkPlan) :SparkPlan = {
+    // Note that we disregard the GPU plan returned here and instead rely on side effects of
+    // tagging the underlying SparkPlan.
+    GpuOverrides().apply(plan)
+    // return the original plan which is now modified as a side-effect of invoking GpuOverrides
+    plan
+  }
+}
 
 case class GpuOverrides() extends Rule[SparkPlan] with Logging {
   override def apply(plan: SparkPlan) :SparkPlan = {

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

@@ -50,6 +50,7 @@ class SQLExecPlugin extends (SparkSessionExtensions => Unit) with Logging {
     logWarning("Installing extensions to enable rapids GPU SQL support." +
       s" To disable GPU support set `${RapidsConf.SQL_ENABLED}` to false")
     extensions.injectColumnar(_ => ColumnarOverrideRules())
+    ShimLoader.getSparkShims.injectQueryStagePrepRule(extensions, _ => GpuQueryStagePrepOverrides())
   }
 }
 

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

@@ -23,6 +23,7 @@ import com.nvidia.spark.rapids.GpuOverrides.isStringLit
 import org.apache.spark.sql.catalyst.expressions.{BinaryExpression, ComplexTypeMergingExpression, Expression, String2TrimExpression, TernaryExpression, UnaryExpression}
 import org.apache.spark.sql.catalyst.expressions.aggregate.AggregateFunction
 import org.apache.spark.sql.catalyst.plans.physical.Partitioning
+import org.apache.spark.sql.catalyst.trees.TreeNodeTag
 import org.apache.spark.sql.connector.read.Scan
 import org.apache.spark.sql.execution.SparkPlan
 import org.apache.spark.sql.execution.command.DataWritingCommand
@@ -116,12 +117,21 @@ abstract class RapidsMeta[INPUT <: BASE, BASE, OUTPUT <: BASE](
 
   private var shouldBeRemovedReasons: Option[mutable.Set[String]] = None
 
+  val gpuSupportedTag = TreeNodeTag[String]("rapids.gpu.supported")
+
   /**
    * Call this to indicate that this should not be replaced with a GPU enabled version
    * @param because why it should not be replaced.
    */
-  final def willNotWorkOnGpu(because: String): Unit =
+  final def willNotWorkOnGpu(because: String): Unit = {
     cannotBeReplacedReasons.get.add(because)
+    // annotate the real spark plan with the reason as well so that the information is available
+    // during query stage planning when AQE is on
+    wrapped match {
+      case p: SparkPlan => p.setTagValue(gpuSupportedTag, because)
+      case _ =>
+    }
+  }
 
   final def shouldBeRemoved(because: String): Unit =
     shouldBeRemovedReasons.get.add(because)

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

@@ -16,9 +16,11 @@
 
 package com.nvidia.spark.rapids
 
+import org.apache.spark.sql.{SparkSession, SparkSessionExtensions}
 import org.apache.spark.sql.catalyst.encoders.ExpressionEncoder
 import org.apache.spark.sql.catalyst.expressions.Expression
 import org.apache.spark.sql.catalyst.plans.JoinType
+import org.apache.spark.sql.catalyst.rules.Rule
 import org.apache.spark.sql.execution.SparkPlan
 import org.apache.spark.sql.execution.joins._
 import org.apache.spark.sql.rapids.execution.GpuBroadcastNestedLoopJoinExecBase
@@ -68,4 +70,8 @@ trait SparkShims {
     endMapIndex: Int,
     startPartition: Int,
     endPartition: Int): Iterator[(BlockManagerId, Seq[(BlockId, Long, Int)])]
+
+  def injectQueryStagePrepRule(
+      extensions: SparkSessionExtensions,
+      rule: SparkSession => Rule[SparkPlan])
 }

sql-plugin/src/main/scala/org/apache/spark/sql/rapids/execution/GpuBroadcastExchangeExec.scala

@@ -217,9 +217,18 @@ class GpuBroadcastMeta(
       case _: BroadcastNestedLoopJoinExec => true
       case _ => false
     }
-    if (!parent.exists(isSupported)) {
-      willNotWorkOnGpu("BroadcastExchange only works on the GPU if being used " +
-        "with a GPU version of BroadcastHashJoinExec or BroadcastNestedLoopJoinExec")
+    if (parent.isDefined) {
+      if (!parent.exists(isSupported)) {
+        willNotWorkOnGpu("BroadcastExchange only works on the GPU if being used " +
+            "with a GPU version of BroadcastHashJoinExec or BroadcastNestedLoopJoinExec")
+      }
+    } else {
+      // when AQE is enabled and we are planning a new query stage, parent will be None so
+      // we need to look at meta-data previously stored on the spark plan
+      wrapped.getTagValue(gpuSupportedTag) match {
+        case Some(reason) => willNotWorkOnGpu(reason)
+        case None => // this broadcast is supported on GPU
+      }
     }
   }