Support split broadcast join condition into ast and non-ast [databric…

…ks] (NVIDIA#9760)

Signed-off-by: Ferdinand Xu <[email protected]>
Co-authored-by: Jason Lowe <[email protected]>
Co-authored-by: Navin Kumar <[email protected]>

integration_tests/src/main/python/join_test.py

@@ -19,8 +19,9 @@
 from asserts import assert_gpu_and_cpu_are_equal_collect, assert_gpu_fallback_collect, assert_cpu_and_gpu_are_equal_collect_with_capture
 from conftest import is_databricks_runtime, is_emr_runtime, is_not_utc
 from data_gen import *
-from marks import ignore_order, allow_non_gpu, incompat, validate_execs_in_gpu_plan
-from spark_session import with_cpu_session, is_before_spark_330, is_databricks_runtime
+from marks import ignore_order, allow_non_gpu, incompat, validate_execs_in_gpu_plan, \
+    datagen_overrides
+from spark_session import with_cpu_session, is_before_spark_330, is_databricks113_or_later, is_databricks_runtime
 
 pytestmark = [pytest.mark.nightly_resource_consuming_test]
 
@@ -436,6 +437,38 @@ def do_join(spark):
         return broadcast(left).join(right, left.a > f.log(right.r_a), join_type)
     assert_gpu_fallback_collect(do_join, 'BroadcastNestedLoopJoinExec')
 
+# Allowing non Gpu for ShuffleExchangeExec is mainly for Databricks where its exchange is CPU based ('Exchange SinglePartition, EXECUTOR_BROADCAST').
+db_113_cpu_bhj_join_allow=["ShuffleExchangeExec"] if is_databricks113_or_later() else []
+
+
+@allow_non_gpu(*db_113_cpu_bhj_join_allow)
+@ignore_order(local=True)
+@pytest.mark.parametrize('data_gen', [IntegerGen(), LongGen()], ids=idfn)
+@pytest.mark.parametrize('join_type', ['Left', 'Inner', 'LeftSemi', 'LeftAnti'], ids=idfn)
+def test_broadcast_hash_join_on_non_ast_condition_without_fallback(data_gen, join_type):
+    # This is to test BHJ with a condition not fully supported by AST. With extra project nodes wrapped, join can still run on GPU other than fallback.
+    def do_join(spark):
+        left, right = create_df(spark, data_gen, 50, 25)
+        # AST does not support cast or logarithm yet
+        return left.join(right.hint("broadcast"), ((left.b == right.r_b) & (f.round(left.a).cast('integer') > f.round(f.log(right.r_a).cast('integer')))), join_type)
+    assert_gpu_and_cpu_are_equal_collect(do_join, conf = {"spark.rapids.sql.castFloatToIntegralTypes.enabled": True})
+
+
+@allow_non_gpu('BroadcastHashJoinExec', 'BroadcastExchangeExec')
+@ignore_order(local=True)
+@pytest.mark.parametrize('data_gen', [IntegerGen(), LongGen()], ids=idfn)
+@pytest.mark.parametrize('join_type', ['Left', 'LeftSemi', 'LeftAnti'], ids=idfn)
+def test_broadcast_hash_join_on_non_ast_condition_fallback(data_gen, join_type):
+    # This is to test BHJ with a condition not fully supported by AST. Since AST doesn't support double, this query fallback to CPU.
+    # Inner join is not included since it can be supported by GPU via a post filter.
+    def do_join(spark):
+        left, right = create_df(spark, data_gen, 50, 25)
+        # AST does not support cast or logarithm yet and also it's not able to be split as project
+        # node those both sides are involved in join condition
+        return left.join(right.hint("broadcast"), ((left.b == right.r_b) & (left.a.cast('double') > right.r_a.cast('double'))), join_type)
+    assert_gpu_fallback_collect(do_join, 'BroadcastHashJoinExec')
+
+
 @ignore_order(local=True)
 @pytest.mark.parametrize('data_gen', [byte_gen, short_gen, int_gen, long_gen,
                                       float_gen, double_gen,

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

@@ -16,11 +16,15 @@
 
 package com.nvidia.spark.rapids
 
+import java.io.Serializable
+
+import com.nvidia.spark.rapids.Arm.withResource
 import scala.collection.mutable
 import scala.collection.mutable.ListBuffer
 
-import org.apache.spark.sql.catalyst.expressions.{AttributeSeq, Expression, ExprId, NamedExpression}
+import org.apache.spark.sql.catalyst.expressions.{Attribute, AttributeSeq, Expression, ExprId, NamedExpression}
 import org.apache.spark.sql.rapids.catalyst.expressions.{GpuEquivalentExpressions, GpuExpressionEquals}
+import org.apache.spark.sql.vectorized.ColumnarBatch
 
 
 object AstUtil {
@@ -119,4 +123,164 @@ object AstUtil {
       }
     }
   }
+
+  /**
+   * Transforms the original join condition into extra filter/project when necessary.
+   * It's targeted for some cases join condition is not fully evaluated by ast.
+   * Based on join condition, it can be transformed into three major strategies:
+   * (1) [NoopJoinCondSplit]: noop when join condition can be fully evaluated with ast.
+   * (2) [JoinCondSplitAsPostFilter]: entire join condition is pulled out as a post filter
+   * after join condition.
+   * (3) [JoinCondSplitAsProject]: extract not supported join condition into pre-project nodes
+   * on each join child. One extra project node is introduced to remove intermediate attributes.
+   */
+  abstract class JoinCondSplitStrategy(left: Seq[NamedExpression],
+      right: Seq[NamedExpression], buildSide: GpuBuildSide) extends Serializable {
+
+    // Actual output of build/stream side project due to join condition split
+    private[this] val (buildOutputAttr, streamOutputAttr) = buildSide match {
+      case GpuBuildLeft => (joinLeftOutput, joinRightOutput)
+      case GpuBuildRight => (joinRightOutput, joinLeftOutput)
+    }
+
+    // This is the left side child of join. In `split as project` strategy, it may be different
+    // from original left child with extracted join condition attribute.
+    def leftOutput(): Seq[NamedExpression] = left
+
+    // This is the right side child of join. In `split as project` strategy, it may be different
+    // from original right child with extracted join condition attribute.
+    def rightOutput(): Seq[NamedExpression] = right
+
+    def astCondition(): Option[Expression]
+
+    def processBuildSideAndClose(input: ColumnarBatch): ColumnarBatch = input
+
+    def processStreamSideAndClose(input: ColumnarBatch): ColumnarBatch = input
+
+    def processPostJoin(iter: Iterator[ColumnarBatch]): Iterator[ColumnarBatch] = iter
+
+    // This is the left side child of join. In `split as project` strategy, it may be different
+    // from original left child with extracted join condition attribute.
+    def joinLeftOutput(): Seq[Attribute] = leftOutput.map(expr => expr.toAttribute)
+
+    // This is the right side child of join. In `split as project` strategy, it may be different
+    // from original right child with extracted join condition attribute.
+    def joinRightOutput(): Seq[Attribute] = rightOutput.map(expr => expr.toAttribute)
+
+    // Updated build attribute list after join condition split as project node.
+    // It may include extra attributes from split join condition.
+    def buildSideOutput(): Seq[Attribute] = buildOutputAttr
+
+    // Updated stream attribute list after join condition split as project node.
+    // It may include extra attributes from split join condition.
+    def streamedSideOutput(): Seq[Attribute] = streamOutputAttr
+  }
+
+  // For the case entire join condition can be evaluated as ast.
+  case class NoopJoinCondSplit(condition: Option[Expression], left: Seq[NamedExpression],
+      right: Seq[NamedExpression], buildSide: GpuBuildSide)
+    extends JoinCondSplitStrategy(left, right, buildSide) {
+    override def astCondition(): Option[Expression] = condition
+  }
+
+  // For inner joins we can apply a post-join condition for any conditions that cannot be
+  // evaluated directly in a mixed join that leverages a cudf AST expression.
+  case class JoinCondSplitAsPostFilter(expr: Option[Expression],
+      attributeSeq: Seq[Attribute], left: Seq[NamedExpression],
+      right: Seq[NamedExpression], buildSide: GpuBuildSide)
+    extends JoinCondSplitStrategy(left, right, buildSide) {
+    private[this] val postFilter = expr.map { e =>
+      GpuBindReferences.bindGpuReferencesTiered(
+        Seq(e), attributeSeq, false)
+    }
+
+    override def astCondition(): Option[Expression] = None
+
+    override def processPostJoin(iter: Iterator[ColumnarBatch]): Iterator[ColumnarBatch] = {
+      postFilter.map { filter =>
+        iter.flatMap { cb =>
+          GpuFilter.filterAndClose(cb, filter, NoopMetric, NoopMetric, NoopMetric)
+        }
+      }.getOrElse(iter)
+    }
+  }
+
+  /**
+   * This is the split strategy targeting on the case where ast not supported join condition can be
+   * extracted and wrapped into extra project node(s).
+   *
+   * @param astCond   remained join condition after extracting ast not supported parts
+   * @param left      original expressions from join's left child. It's left project input
+   *                  attribute.
+   * @param leftProj  extra expressions extracted from original join condition which is not
+   *                  supported by ast. It will be evaluated as a project on left side batch.
+   * @param right     original expressions from join's right child. It's left project input
+   *                  attribute.
+   * @param rightProj extra expressions extracted from original join condition which is not
+   *                  supported by ast. It will be evaluated as a project on right side batch.
+   * @param post      eliminate the extra columns introduced by join condition split
+   * @param buildSide indicates which side is build
+   */
+  case class JoinCondSplitAsProject(
+      astCond: Option[Expression],
+      left: Seq[NamedExpression], leftProj: Seq[NamedExpression],
+      right: Seq[NamedExpression], rightProj: Seq[NamedExpression],
+      post: Seq[NamedExpression], buildSide: GpuBuildSide
+  ) extends JoinCondSplitStrategy(left ++ leftProj, right ++ rightProj, buildSide) {
+    private[this] val leftInput = left.map(_.toAttribute)
+    private[this] val rightInput = right.map(_.toAttribute)
+
+    // Used to build build/stream side project
+    private[this] val (buildOutput, streamOutput, buildInput, streamInput) = buildSide match {
+      case GpuBuildLeft =>
+        (leftOutput, rightOutput, leftInput, rightInput)
+      case GpuBuildRight =>
+        (rightOutput, leftOutput, rightInput, leftInput)
+    }
+
+    private[this] val buildProj = if (!buildOutput.isEmpty) {
+      Some(GpuBindReferences.bindGpuReferencesTiered(buildOutput, buildInput, false))
+    } else None
+
+    private[this] val streamProj = if (!streamOutput.isEmpty) {
+      Some(GpuBindReferences.bindGpuReferencesTiered(streamOutput, streamInput, false))
+    } else None
+
+    // Remove the intermediate attributes from left and right side project nodes. Output attributes
+    // need to be updated based on join type. And its attributes covers both original plan and
+    // extra project node.
+    private[this] val postProj = if (!post.isEmpty) {
+      Some(
+        GpuBindReferences.bindGpuReferencesTiered(
+          post, (leftOutput ++ rightOutput).map(_.toAttribute), false))
+    } else None
+
+    override def astCondition(): Option[Expression] = astCond
+
+    override def processBuildSideAndClose(input: ColumnarBatch): ColumnarBatch = {
+      buildProj.map { pj =>
+        withResource(input) { cb =>
+          pj.project(cb)
+        }
+      }.getOrElse(input)
+    }
+
+    override def processStreamSideAndClose(input: ColumnarBatch): ColumnarBatch = {
+      streamProj.map { pj =>
+        withResource(input) { cb =>
+          pj.project(cb)
+        }
+      }.getOrElse(input)
+    }
+
+    override def processPostJoin(iter: Iterator[ColumnarBatch]): Iterator[ColumnarBatch] = {
+      postProj.map { proj =>
+        iter.map { cb =>
+          withResource(cb) { b =>
+            proj.project(b)
+          }
+        }
+      }.getOrElse(iter)
+    }
+  }
 }

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

@@ -19,6 +19,7 @@ package org.apache.spark.sql.rapids.execution
 import ai.rapids.cudf.{NvtxColor, NvtxRange}
 import com.nvidia.spark.rapids._
 import com.nvidia.spark.rapids.Arm.{closeOnExcept, withResource}
+import com.nvidia.spark.rapids.AstUtil.JoinCondSplitStrategy
 import com.nvidia.spark.rapids.shims.{GpuBroadcastJoinMeta, ShimBinaryExecNode}
 
 import org.apache.spark.TaskContext
@@ -55,6 +56,28 @@ abstract class GpuBroadcastHashJoinMetaBase(
 
   override val childExprs: Seq[BaseExprMeta[_]] = leftKeys ++ rightKeys ++ conditionMeta
 
+  private var taggedForAstCheck = false
+
+  // Avoid checking multiple times
+  private var isAstCond = false
+
+  /**
+   * Check whether condition can be ast-able. It includes two cases: 1) all join conditions are
+   * ast-able; 2) join conditions are ast-able after split and push down to child plans.
+   */
+  def canJoinCondAstAble(): Boolean = {
+    if (!taggedForAstCheck) {
+      val Seq(leftPlan, rightPlan) = childPlans
+      isAstCond = conditionMeta match {
+        case Some(e) => AstUtil.canExtractNonAstConditionIfNeed(
+          e, leftPlan.outputAttributes.map(_.exprId), rightPlan.outputAttributes.map(_.exprId))
+        case None => true
+      }
+      taggedForAstCheck = true
+    }
+    isAstCond
+  }
+
   override def tagPlanForGpu(): Unit = {
     GpuHashJoin.tagJoin(this, join.joinType, buildSide, join.leftKeys, join.rightKeys,
       conditionMeta)
@@ -103,6 +126,7 @@ abstract class GpuBroadcastHashJoinExecBase(
     joinType: JoinType,
     buildSide: GpuBuildSide,
     override val condition: Option[Expression],
+    override val joinCondSplitStrategy: JoinCondSplitStrategy,
     left: SparkPlan,
     right: SparkPlan) extends ShimBinaryExecNode with GpuHashJoin {
   import GpuMetric._