maropu commented on a change in pull request #20965: [SPARK-21870][SQL] Split
aggregation code into small functions
URL: https://github.com/apache/spark/pull/20965#discussion_r318837668
##########
File path:
sql/core/src/main/scala/org/apache/spark/sql/execution/aggregate/HashAggregateExec.scala
##########
@@ -255,41 +260,148 @@ case class HashAggregateExec(
""".stripMargin
}
+ // Splits aggregate code into small functions because the most of JVM
implementations
+ // can not compile too long functions.
+ //
+ // Note: The difference from `CodeGenerator.splitExpressions` is that we
define an individual
+ // function for each aggregation function (e.g., SUM and AVG). For example,
in a query
+ // `SELECT SUM(a), AVG(a) FROM VALUES(1) t(a)`, we define two functions
+ // for `SUM(a)` and `AVG(a)`.
+ private def splitAggregateExpressions(
+ ctx: CodegenContext,
+ aggNames: Seq[String],
+ aggExprs: Seq[Seq[Expression]],
+ makeSplitAggFunctions: => Seq[String],
+ subExprs: Map[Expression, SubExprEliminationState]): Option[String] = {
+ val inputVars = aggExprs.map { aggExprsInAgg =>
+ val inputVarsInAgg = aggExprsInAgg.map(
+ CodeGenerator.getLocalInputVariableValues(ctx, _, subExprs)).reduce(_
++ _).toSeq
+ val paramLength =
CodeGenerator.calculateParamLengthFromExprValues(inputVarsInAgg)
+
+ // Checks if a parameter length for the `aggExprsInAgg` does not go over
the JVM limit
+ if (CodeGenerator.isValidParamLength(paramLength)) {
+ Some(inputVarsInAgg)
+ } else {
+ None
+ }
+ }
+
+ // Checks if all the aggregate code can be split into pieces.
+ // If the parameter length of at lease one `aggExprsInAgg` goes over the
limit,
+ // we totally give up splitting aggregate code.
+ if (inputVars.forall(_.isDefined)) {
+ val splitAggEvalCodes = makeSplitAggFunctions
+ val splitCodes = inputVars.flatten.zipWithIndex.map { case (args, i) =>
+ val doAggVal = ctx.freshName(s"doAggregateVal_${aggNames(i)}")
+ val argList = args.map(v => s"${v.javaType.getName}
${v.variableName}").mkString(", ")
+ val doAggValFuncName = ctx.addNewFunction(doAggVal,
+ s"""
+ | private void $doAggVal($argList) throws java.io.IOException {
+ | ${splitAggEvalCodes(i)}
+ | }
+ """.stripMargin)
+
+ val inputVariables = args.map(_.variableName).mkString(", ")
+ s"$doAggValFuncName($inputVariables);"
+ }
+ Some(splitCodes.mkString("\n").trim)
+ } else {
+ val errMsg = "Failed to split aggregate code into small functions
because the parameter " +
+ "length of at least one split function went over the JVM limit: " +
+ CodeGenerator.MAX_JVM_METHOD_PARAMS_LENGTH
+ if (Utils.isTesting) {
+ throw new IllegalStateException(errMsg)
+ } else {
+ logInfo(errMsg)
+ None
+ }
+ }
+ }
+
private def doConsumeWithoutKeys(ctx: CodegenContext, input: Seq[ExprCode]):
String = {
// only have DeclarativeAggregate
val functions =
aggregateExpressions.map(_.aggregateFunction.asInstanceOf[DeclarativeAggregate])
val inputAttrs = functions.flatMap(_.aggBufferAttributes) ++ child.output
- val updateExpr = aggregateExpressions.flatMap { e =>
+ val updateExprs = aggregateExpressions.map { e =>
e.mode match {
case Partial | Complete =>
e.aggregateFunction.asInstanceOf[DeclarativeAggregate].updateExpressions
case PartialMerge | Final =>
e.aggregateFunction.asInstanceOf[DeclarativeAggregate].mergeExpressions
}
}
- ctx.currentVars = bufVars ++ input
- val boundUpdateExpr = bindReferences(updateExpr, inputAttrs)
- val subExprs =
ctx.subexpressionEliminationForWholeStageCodegen(boundUpdateExpr)
+ ctx.currentVars = bufVars.flatten ++ input
+ val boundUpdateExprs = updateExprs.map { updateExprsInAgg =>
+ updateExprsInAgg.map(BindReferences.bindReference(_, inputAttrs))
+ }
+ val subExprs =
ctx.subexpressionEliminationForWholeStageCodegen(boundUpdateExprs.flatten)
val effectiveCodes = subExprs.codes.mkString("\n")
- val aggVals = ctx.withSubExprEliminationExprs(subExprs.states) {
- boundUpdateExpr.map(_.genCode(ctx))
+ val aggVals = boundUpdateExprs.map { boundUpdateExprsInAgg =>
+ ctx.withSubExprEliminationExprs(subExprs.states) {
+ boundUpdateExprsInAgg.map(_.genCode(ctx))
+ }
}
- // aggregate buffer should be updated atomic
- val updates = aggVals.zipWithIndex.map { case (ev, i) =>
+
+ lazy val nonSplitAggCode = {
+ // aggregate buffer should be updated atomically
+ val updates = aggVals.flatten.zip(bufVars.flatten).map { case (ev,
bufVar) =>
+ s"""
+ | ${bufVar.isNull} = ${ev.isNull};
+ | ${bufVar.value} = ${ev.value};
+ """.stripMargin
+ }
s"""
- | ${bufVars(i).isNull} = ${ev.isNull};
- | ${bufVars(i).value} = ${ev.value};
+ | // do aggregate
+ | // common sub-expressions
+ | $effectiveCodes
+ | // evaluate aggregate functions
+ | ${evaluateVariables(aggVals.flatten)}
+ | // update aggregation buffers
+ | ${updates.mkString("\n").trim}
""".stripMargin
}
- s"""
- | // do aggregate
- | // common sub-expressions
- | $effectiveCodes
- | // evaluate aggregate function
- | ${evaluateVariables(aggVals)}
- | // update aggregation buffer
- | ${updates.mkString("\n").trim}
- """.stripMargin
+
+ if (conf.codegenSplitAggregateFunc) {
+ val splitAggCode = splitAggregateExpressions(
+ ctx = ctx,
+ aggNames = functions.map(_.prettyName),
+ aggExprs = boundUpdateExprs,
+ makeSplitAggFunctions = {
+ aggVals.zip(bufVars).map { case (aggValsInAgg, bufVarsInAgg) =>
+ // All the update code for aggregation buffers should be placed in
the end
+ // of each aggregation function code.
+ val updates = aggValsInAgg.zip(bufVarsInAgg).map { case (ev,
bufVar) =>
+ s"""
+ | ${bufVar.isNull} = ${ev.isNull};
Review comment:
removed
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
