Examples with AggregateCall org.apache.calcite.rel.core.AggregateCall used on opensource projects

Example 6 with AggregateCall

use of org.apache.calcite.rel.core.AggregateCall in project hive by apache.

the class PlanModifierForASTConv method replaceEmptyGroupAggr.

private static void replaceEmptyGroupAggr(final RelNode rel, RelNode parent) {
    // If this function is called, the parent should only include constant
    List<RexNode> exps = parent.getChildExps();
    for (RexNode rexNode : exps) {
        if (!rexNode.accept(new HiveCalciteUtil.ConstantFinder())) {
            throw new RuntimeException("We expect " + parent.toString() + " to contain only constants. However, " + rexNode.toString() + " is " + rexNode.getKind());
        }
    }
    HiveAggregate oldAggRel = (HiveAggregate) rel;
    RelDataTypeFactory typeFactory = oldAggRel.getCluster().getTypeFactory();
    RelDataType longType = TypeConverter.convert(TypeInfoFactory.longTypeInfo, typeFactory);
    RelDataType intType = TypeConverter.convert(TypeInfoFactory.intTypeInfo, typeFactory);
    // Create the dummy aggregation.
    SqlAggFunction countFn = SqlFunctionConverter.getCalciteAggFn("count", false, ImmutableList.of(intType), longType);
    // TODO: Using 0 might be wrong; might need to walk down to find the
    // proper index of a dummy.
    List<Integer> argList = ImmutableList.of(0);
    AggregateCall dummyCall = new AggregateCall(countFn, false, argList, longType, null);
    Aggregate newAggRel = oldAggRel.copy(oldAggRel.getTraitSet(), oldAggRel.getInput(), oldAggRel.indicator, oldAggRel.getGroupSet(), oldAggRel.getGroupSets(), ImmutableList.of(dummyCall));
    RelNode select = introduceDerivedTable(newAggRel);
    parent.replaceInput(0, select);
}

Example 7 with AggregateCall

use of org.apache.calcite.rel.core.AggregateCall in project hive by apache.

the class HiveAggregateJoinTransposeRule method onMatch.

@Override
public void onMatch(RelOptRuleCall call) {
    final Aggregate aggregate = call.rel(0);
    final Join join = call.rel(1);
    final RexBuilder rexBuilder = aggregate.getCluster().getRexBuilder();
    // If any aggregate call has a filter, bail out
    for (AggregateCall aggregateCall : aggregate.getAggCallList()) {
        if (aggregateCall.getAggregation().unwrap(SqlSplittableAggFunction.class) == null) {
            return;
        }
        if (aggregateCall.filterArg >= 0) {
            return;
        }
    }
    // aggregate operator
    if (join.getJoinType() != JoinRelType.INNER) {
        return;
    }
    if (!allowFunctions && !aggregate.getAggCallList().isEmpty()) {
        return;
    }
    // Do the columns used by the join appear in the output of the aggregate?
    RelMetadataQuery mq = RelMetadataQuery.instance();
    final ImmutableBitSet aggregateColumns = aggregate.getGroupSet();
    final ImmutableBitSet keyColumns = keyColumns(aggregateColumns, mq.getPulledUpPredicates(join).pulledUpPredicates);
    final ImmutableBitSet joinColumns = RelOptUtil.InputFinder.bits(join.getCondition());
    final boolean allColumnsInAggregate = keyColumns.contains(joinColumns);
    final ImmutableBitSet belowAggregateColumns = aggregateColumns.union(joinColumns);
    // Split join condition
    final List<Integer> leftKeys = Lists.newArrayList();
    final List<Integer> rightKeys = Lists.newArrayList();
    final List<Boolean> filterNulls = Lists.newArrayList();
    RexNode nonEquiConj = RelOptUtil.splitJoinCondition(join.getLeft(), join.getRight(), join.getCondition(), leftKeys, rightKeys, filterNulls);
    // If it contains non-equi join conditions, we bail out
    if (!nonEquiConj.isAlwaysTrue()) {
        return;
    }
    // Push each aggregate function down to each side that contains all of its
    // arguments. Note that COUNT(*), because it has no arguments, can go to
    // both sides.
    final Map<Integer, Integer> map = new HashMap<>();
    final List<Side> sides = new ArrayList<>();
    int uniqueCount = 0;
    int offset = 0;
    int belowOffset = 0;
    for (int s = 0; s < 2; s++) {
        final Side side = new Side();
        final RelNode joinInput = join.getInput(s);
        int fieldCount = joinInput.getRowType().getFieldCount();
        final ImmutableBitSet fieldSet = ImmutableBitSet.range(offset, offset + fieldCount);
        final ImmutableBitSet belowAggregateKeyNotShifted = belowAggregateColumns.intersect(fieldSet);
        for (Ord<Integer> c : Ord.zip(belowAggregateKeyNotShifted)) {
            map.put(c.e, belowOffset + c.i);
        }
        final ImmutableBitSet belowAggregateKey = belowAggregateKeyNotShifted.shift(-offset);
        final boolean unique;
        if (!allowFunctions) {
            assert aggregate.getAggCallList().isEmpty();
            // If there are no functions, it doesn't matter as much whether we
            // aggregate the inputs before the join, because there will not be
            // any functions experiencing a cartesian product effect.
            //
            // But finding out whether the input is already unique requires a call
            // to areColumnsUnique that currently (until [CALCITE-794] "Detect
            // cycles when computing statistics" is fixed) places a heavy load on
            // the metadata system.
            //
            // So we choose to imagine the the input is already unique, which is
            // untrue but harmless.
            //
            unique = true;
        } else {
            final Boolean unique0 = mq.areColumnsUnique(joinInput, belowAggregateKey);
            unique = unique0 != null && unique0;
        }
        if (unique) {
            ++uniqueCount;
            side.newInput = joinInput;
        } else {
            List<AggregateCall> belowAggCalls = new ArrayList<>();
            final SqlSplittableAggFunction.Registry<AggregateCall> belowAggCallRegistry = registry(belowAggCalls);
            final Mappings.TargetMapping mapping = s == 0 ? Mappings.createIdentity(fieldCount) : Mappings.createShiftMapping(fieldCount + offset, 0, offset, fieldCount);
            for (Ord<AggregateCall> aggCall : Ord.zip(aggregate.getAggCallList())) {
                final SqlAggFunction aggregation = aggCall.e.getAggregation();
                final SqlSplittableAggFunction splitter = Preconditions.checkNotNull(aggregation.unwrap(SqlSplittableAggFunction.class));
                final AggregateCall call1;
                if (fieldSet.contains(ImmutableBitSet.of(aggCall.e.getArgList()))) {
                    call1 = splitter.split(aggCall.e, mapping);
                } else {
                    call1 = splitter.other(rexBuilder.getTypeFactory(), aggCall.e);
                }
                if (call1 != null) {
                    side.split.put(aggCall.i, belowAggregateKey.cardinality() + belowAggCallRegistry.register(call1));
                }
            }
            side.newInput = aggregateFactory.createAggregate(joinInput, false, belowAggregateKey, null, belowAggCalls);
        }
        offset += fieldCount;
        belowOffset += side.newInput.getRowType().getFieldCount();
        sides.add(side);
    }
    if (uniqueCount == 2) {
        // invocation of this rule; if we continue we might loop forever.
        return;
    }
    // Update condition
    final Mapping mapping = (Mapping) Mappings.target(new Function<Integer, Integer>() {

        @Override
        public Integer apply(Integer a0) {
            return map.get(a0);
        }
    }, join.getRowType().getFieldCount(), belowOffset);
    final RexNode newCondition = RexUtil.apply(mapping, join.getCondition());
    // Create new join
    RelNode newJoin = joinFactory.createJoin(sides.get(0).newInput, sides.get(1).newInput, newCondition, join.getJoinType(), join.getVariablesStopped(), join.isSemiJoinDone());
    // Aggregate above to sum up the sub-totals
    final List<AggregateCall> newAggCalls = new ArrayList<>();
    final int groupIndicatorCount = aggregate.getGroupCount() + aggregate.getIndicatorCount();
    final int newLeftWidth = sides.get(0).newInput.getRowType().getFieldCount();
    final List<RexNode> projects = new ArrayList<>(rexBuilder.identityProjects(newJoin.getRowType()));
    for (Ord<AggregateCall> aggCall : Ord.zip(aggregate.getAggCallList())) {
        final SqlAggFunction aggregation = aggCall.e.getAggregation();
        final SqlSplittableAggFunction splitter = Preconditions.checkNotNull(aggregation.unwrap(SqlSplittableAggFunction.class));
        final Integer leftSubTotal = sides.get(0).split.get(aggCall.i);
        final Integer rightSubTotal = sides.get(1).split.get(aggCall.i);
        newAggCalls.add(splitter.topSplit(rexBuilder, registry(projects), groupIndicatorCount, newJoin.getRowType(), aggCall.e, leftSubTotal == null ? -1 : leftSubTotal, rightSubTotal == null ? -1 : rightSubTotal + newLeftWidth));
    }
    RelNode r = newJoin;
    b: if (allColumnsInAggregate && newAggCalls.isEmpty() && RelOptUtil.areRowTypesEqual(r.getRowType(), aggregate.getRowType(), false)) {
    // no need to aggregate
    } else {
        r = RelOptUtil.createProject(r, projects, null, true, relBuilderFactory.create(aggregate.getCluster(), null));
        if (allColumnsInAggregate) {
            // let's see if we can convert
            List<RexNode> projects2 = new ArrayList<>();
            for (int key : Mappings.apply(mapping, aggregate.getGroupSet())) {
                projects2.add(rexBuilder.makeInputRef(r, key));
            }
            for (AggregateCall newAggCall : newAggCalls) {
                final SqlSplittableAggFunction splitter = newAggCall.getAggregation().unwrap(SqlSplittableAggFunction.class);
                if (splitter != null) {
                    projects2.add(splitter.singleton(rexBuilder, r.getRowType(), newAggCall));
                }
            }
            if (projects2.size() == aggregate.getGroupSet().cardinality() + newAggCalls.size()) {
                // We successfully converted agg calls into projects.
                r = RelOptUtil.createProject(r, projects2, null, true, relBuilderFactory.create(aggregate.getCluster(), null));
                break b;
            }
        }
        r = aggregateFactory.createAggregate(r, aggregate.indicator, Mappings.apply(mapping, aggregate.getGroupSet()), Mappings.apply2(mapping, aggregate.getGroupSets()), newAggCalls);
    }
    // Make a cost based decision to pick cheaper plan
    RelOptCost afterCost = mq.getCumulativeCost(r);
    RelOptCost beforeCost = mq.getCumulativeCost(aggregate);
    if (afterCost.isLt(beforeCost)) {
        call.transformTo(r);
    }
}

Example 8 with AggregateCall

use of org.apache.calcite.rel.core.AggregateCall in project hive by apache.

the class HiveExpandDistinctAggregatesRule method convertMonopole.

/**
   * Converts an aggregate relational expression that contains just one
   * distinct aggregate function (or perhaps several over the same arguments)
   * and no non-distinct aggregate functions.
   */
private RelNode convertMonopole(Aggregate aggregate, List<Integer> argList) {
    // For example,
    //    SELECT deptno, COUNT(DISTINCT sal), SUM(DISTINCT sal)
    //    FROM emp
    //    GROUP BY deptno
    //
    // becomes
    //
    //    SELECT deptno, COUNT(distinct_sal), SUM(distinct_sal)
    //    FROM (
    //      SELECT DISTINCT deptno, sal AS distinct_sal
    //      FROM EMP GROUP BY deptno)
    //    GROUP BY deptno
    // Project the columns of the GROUP BY plus the arguments
    // to the agg function.
    Map<Integer, Integer> sourceOf = new HashMap<Integer, Integer>();
    final Aggregate distinct = createSelectDistinct(aggregate, argList, sourceOf);
    // Create an aggregate on top, with the new aggregate list.
    final List<AggregateCall> newAggCalls = Lists.newArrayList(aggregate.getAggCallList());
    rewriteAggCalls(newAggCalls, argList, sourceOf);
    final int cardinality = aggregate.getGroupSet().cardinality();
    return aggregate.copy(aggregate.getTraitSet(), distinct, aggregate.indicator, ImmutableBitSet.range(cardinality), null, newAggCalls);
}

Example 9 with AggregateCall

use of org.apache.calcite.rel.core.AggregateCall in project hive by apache.

the class HiveIntersectRewriteRule method onMatch.

// ~ Methods ----------------------------------------------------------------
public void onMatch(RelOptRuleCall call) {
    final HiveIntersect hiveIntersect = call.rel(0);
    final RelOptCluster cluster = hiveIntersect.getCluster();
    final RexBuilder rexBuilder = cluster.getRexBuilder();
    int numOfBranch = hiveIntersect.getInputs().size();
    Builder<RelNode> bldr = new ImmutableList.Builder<RelNode>();
    // 1st level GB: create a GB (col0, col1, count(1) as c) for each branch
    for (int index = 0; index < numOfBranch; index++) {
        RelNode input = hiveIntersect.getInputs().get(index);
        final List<RexNode> gbChildProjLst = Lists.newArrayList();
        final List<Integer> groupSetPositions = Lists.newArrayList();
        for (int cInd = 0; cInd < input.getRowType().getFieldList().size(); cInd++) {
            gbChildProjLst.add(rexBuilder.makeInputRef(input, cInd));
            groupSetPositions.add(cInd);
        }
        gbChildProjLst.add(rexBuilder.makeBigintLiteral(new BigDecimal(1)));
        // create the project before GB because we need a new project with extra column '1'.
        RelNode gbInputRel = null;
        try {
            gbInputRel = HiveProject.create(input, gbChildProjLst, null);
        } catch (CalciteSemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
        // groupSetPosition includes all the positions
        final ImmutableBitSet groupSet = ImmutableBitSet.of(groupSetPositions);
        List<AggregateCall> aggregateCalls = Lists.newArrayList();
        RelDataType aggFnRetType = TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory());
        // count(1), 1's position is input.getRowType().getFieldList().size()
        AggregateCall aggregateCall = HiveCalciteUtil.createSingleArgAggCall("count", cluster, TypeInfoFactory.longTypeInfo, input.getRowType().getFieldList().size(), aggFnRetType);
        aggregateCalls.add(aggregateCall);
        HiveRelNode aggregateRel = new HiveAggregate(cluster, cluster.traitSetOf(HiveRelNode.CONVENTION), gbInputRel, false, groupSet, null, aggregateCalls);
        bldr.add(aggregateRel);
    }
    // create a union above all the branches
    HiveRelNode union = new HiveUnion(cluster, TraitsUtil.getDefaultTraitSet(cluster), bldr.build());
    // 2nd level GB: create a GB (col0, col1, count(c)) for each branch
    final List<Integer> groupSetPositions = Lists.newArrayList();
    // the index of c
    int cInd = union.getRowType().getFieldList().size() - 1;
    for (int index = 0; index < union.getRowType().getFieldList().size(); index++) {
        if (index != cInd) {
            groupSetPositions.add(index);
        }
    }
    List<AggregateCall> aggregateCalls = Lists.newArrayList();
    RelDataType aggFnRetType = TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory());
    AggregateCall aggregateCall = HiveCalciteUtil.createSingleArgAggCall("count", cluster, TypeInfoFactory.longTypeInfo, cInd, aggFnRetType);
    aggregateCalls.add(aggregateCall);
    if (hiveIntersect.all) {
        aggregateCall = HiveCalciteUtil.createSingleArgAggCall("min", cluster, TypeInfoFactory.longTypeInfo, cInd, aggFnRetType);
        aggregateCalls.add(aggregateCall);
    }
    final ImmutableBitSet groupSet = ImmutableBitSet.of(groupSetPositions);
    HiveRelNode aggregateRel = new HiveAggregate(cluster, cluster.traitSetOf(HiveRelNode.CONVENTION), union, false, groupSet, null, aggregateCalls);
    // add a filter count(c) = #branches
    int countInd = cInd;
    List<RexNode> childRexNodeLst = new ArrayList<RexNode>();
    RexInputRef ref = rexBuilder.makeInputRef(aggregateRel, countInd);
    RexLiteral literal = rexBuilder.makeBigintLiteral(new BigDecimal(numOfBranch));
    childRexNodeLst.add(ref);
    childRexNodeLst.add(literal);
    ImmutableList.Builder<RelDataType> calciteArgTypesBldr = new ImmutableList.Builder<RelDataType>();
    calciteArgTypesBldr.add(TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory()));
    calciteArgTypesBldr.add(TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory()));
    RexNode factoredFilterExpr = null;
    try {
        factoredFilterExpr = rexBuilder.makeCall(SqlFunctionConverter.getCalciteFn("=", calciteArgTypesBldr.build(), TypeConverter.convert(TypeInfoFactory.longTypeInfo, cluster.getTypeFactory()), true), childRexNodeLst);
    } catch (CalciteSemanticException e) {
        LOG.debug(e.toString());
        throw new RuntimeException(e);
    }
    RelNode filterRel = new HiveFilter(cluster, cluster.traitSetOf(HiveRelNode.CONVENTION), aggregateRel, factoredFilterExpr);
    if (!hiveIntersect.all) {
        // the schema for intersect distinct is like this
        // R3 on all attributes + count(c) as cnt
        // finally add a project to project out the last column
        Set<Integer> projectOutColumnPositions = new HashSet<>();
        projectOutColumnPositions.add(filterRel.getRowType().getFieldList().size() - 1);
        try {
            call.transformTo(HiveCalciteUtil.createProjectWithoutColumn(filterRel, projectOutColumnPositions));
        } catch (CalciteSemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
    } else {
        // the schema for intersect all is like this
        // R3 + count(c) as cnt + min(c) as m
        // we create a input project for udtf whose schema is like this
        // min(c) as m + R3
        List<RexNode> originalInputRefs = Lists.transform(filterRel.getRowType().getFieldList(), new Function<RelDataTypeField, RexNode>() {

            @Override
            public RexNode apply(RelDataTypeField input) {
                return new RexInputRef(input.getIndex(), input.getType());
            }
        });
        List<RexNode> copyInputRefs = new ArrayList<>();
        copyInputRefs.add(originalInputRefs.get(originalInputRefs.size() - 1));
        for (int i = 0; i < originalInputRefs.size() - 2; i++) {
            copyInputRefs.add(originalInputRefs.get(i));
        }
        RelNode srcRel = null;
        try {
            srcRel = HiveProject.create(filterRel, copyInputRefs, null);
            HiveTableFunctionScan udtf = HiveCalciteUtil.createUDTFForSetOp(cluster, srcRel);
            // finally add a project to project out the 1st column
            Set<Integer> projectOutColumnPositions = new HashSet<>();
            projectOutColumnPositions.add(0);
            call.transformTo(HiveCalciteUtil.createProjectWithoutColumn(udtf, projectOutColumnPositions));
        } catch (SemanticException e) {
            LOG.debug(e.toString());
            throw new RuntimeException(e);
        }
    }
}

Example 10 with AggregateCall

use of org.apache.calcite.rel.core.AggregateCall in project hive by apache.

the class SubstitutionVisitor method permute.

public static MutableAggregate permute(MutableAggregate aggregate, MutableRel input, Mapping mapping) {
    ImmutableBitSet groupSet = Mappings.apply(mapping, aggregate.getGroupSet());
    ImmutableList<ImmutableBitSet> groupSets = Mappings.apply2(mapping, aggregate.getGroupSets());
    List<AggregateCall> aggregateCalls = apply(mapping, aggregate.getAggCallList());
    return MutableAggregate.of(input, aggregate.indicator, groupSet, groupSets, aggregateCalls);
}