Examples with SqlCountAggFunction org.apache.calcite.sql.fun.SqlCountAggFunction used on opensource projects

Example 1 with SqlCountAggFunction

use of org.apache.calcite.sql.fun.SqlCountAggFunction in project flink by apache.

the class FlinkAggregateExpandDistinctAggregatesRule method onMatch.

//~ Methods ----------------------------------------------------------------
public void onMatch(RelOptRuleCall call) {
    final Aggregate aggregate = call.rel(0);
    if (!aggregate.containsDistinctCall()) {
        return;
    }
    // Find all of the agg expressions. We use a LinkedHashSet to ensure
    // determinism.
    int nonDistinctCount = 0;
    int distinctCount = 0;
    int filterCount = 0;
    int unsupportedAggCount = 0;
    final Set<Pair<List<Integer>, Integer>> argLists = new LinkedHashSet<>();
    for (AggregateCall aggCall : aggregate.getAggCallList()) {
        if (aggCall.filterArg >= 0) {
            ++filterCount;
        }
        if (!aggCall.isDistinct()) {
            ++nonDistinctCount;
            if (!(aggCall.getAggregation() instanceof SqlCountAggFunction || aggCall.getAggregation() instanceof SqlSumAggFunction || aggCall.getAggregation() instanceof SqlMinMaxAggFunction)) {
                ++unsupportedAggCount;
            }
            continue;
        }
        ++distinctCount;
        argLists.add(Pair.of(aggCall.getArgList(), aggCall.filterArg));
    }
    Preconditions.checkState(argLists.size() > 0, "containsDistinctCall lied");
    // arguments then we can use a more efficient form.
    if (nonDistinctCount == 0 && argLists.size() == 1) {
        final Pair<List<Integer>, Integer> pair = Iterables.getOnlyElement(argLists);
        final RelBuilder relBuilder = call.builder();
        convertMonopole(relBuilder, aggregate, pair.left, pair.right);
        call.transformTo(relBuilder.build());
        return;
    }
    if (useGroupingSets) {
        rewriteUsingGroupingSets(call, aggregate, argLists);
        return;
    }
    // we can generate multi-phase aggregates
    if (// one distinct aggregate
    distinctCount == 1 && // no filter
    filterCount == 0 && // sum/min/max/count in non-distinct aggregate
    unsupportedAggCount == 0 && nonDistinctCount > 0) {
        // one or more non-distinct aggregates
        final RelBuilder relBuilder = call.builder();
        convertSingletonDistinct(relBuilder, aggregate, argLists);
        call.transformTo(relBuilder.build());
        return;
    }
    // Create a list of the expressions which will yield the final result.
    // Initially, the expressions point to the input field.
    final List<RelDataTypeField> aggFields = aggregate.getRowType().getFieldList();
    final List<RexInputRef> refs = new ArrayList<>();
    final List<String> fieldNames = aggregate.getRowType().getFieldNames();
    final ImmutableBitSet groupSet = aggregate.getGroupSet();
    final int groupAndIndicatorCount = aggregate.getGroupCount() + aggregate.getIndicatorCount();
    for (int i : Util.range(groupAndIndicatorCount)) {
        refs.add(RexInputRef.of(i, aggFields));
    }
    // Aggregate the original relation, including any non-distinct aggregates.
    final List<AggregateCall> newAggCallList = new ArrayList<>();
    int i = -1;
    for (AggregateCall aggCall : aggregate.getAggCallList()) {
        ++i;
        if (aggCall.isDistinct()) {
            refs.add(null);
            continue;
        }
        refs.add(new RexInputRef(groupAndIndicatorCount + newAggCallList.size(), aggFields.get(groupAndIndicatorCount + i).getType()));
        newAggCallList.add(aggCall);
    }
    // In the case where there are no non-distinct aggregates (regardless of
    // whether there are group bys), there's no need to generate the
    // extra aggregate and join.
    final RelBuilder relBuilder = call.builder();
    relBuilder.push(aggregate.getInput());
    int n = 0;
    if (!newAggCallList.isEmpty()) {
        final RelBuilder.GroupKey groupKey = relBuilder.groupKey(groupSet, aggregate.indicator, aggregate.getGroupSets());
        relBuilder.aggregate(groupKey, newAggCallList);
        ++n;
    }
    // set of operands.
    for (Pair<List<Integer>, Integer> argList : argLists) {
        doRewrite(relBuilder, aggregate, n++, argList.left, argList.right, refs);
    }
    relBuilder.project(refs, fieldNames);
    call.transformTo(relBuilder.build());
}

Example 2 with SqlCountAggFunction

use of org.apache.calcite.sql.fun.SqlCountAggFunction in project drill by axbaretto.

the class DrillReduceAggregatesRule method reduceStddev.

private RexNode reduceStddev(Aggregate oldAggRel, AggregateCall oldCall, boolean biased, boolean sqrt, List<AggregateCall> newCalls, Map<AggregateCall, RexNode> aggCallMapping, List<RexNode> inputExprs) {
    // stddev_pop(x) ==>
    // power(
    // (sum(x * x) - sum(x) * sum(x) / count(x))
    // / count(x),
    // .5)
    // 
    // stddev_samp(x) ==>
    // power(
    // (sum(x * x) - sum(x) * sum(x) / count(x))
    // / nullif(count(x) - 1, 0),
    // .5)
    final PlannerSettings plannerSettings = (PlannerSettings) oldAggRel.getCluster().getPlanner().getContext();
    final boolean isInferenceEnabled = plannerSettings.isTypeInferenceEnabled();
    final int nGroups = oldAggRel.getGroupCount();
    RelDataTypeFactory typeFactory = oldAggRel.getCluster().getTypeFactory();
    final RexBuilder rexBuilder = oldAggRel.getCluster().getRexBuilder();
    assert oldCall.getArgList().size() == 1 : oldCall.getArgList();
    final int argOrdinal = oldCall.getArgList().get(0);
    final RelDataType argType = getFieldType(oldAggRel.getInput(), argOrdinal);
    // final RexNode argRef = inputExprs.get(argOrdinal);
    RexNode argRef = rexBuilder.makeCall(CastHighOp, inputExprs.get(argOrdinal));
    inputExprs.set(argOrdinal, argRef);
    final RexNode argSquared = rexBuilder.makeCall(SqlStdOperatorTable.MULTIPLY, argRef, argRef);
    final int argSquaredOrdinal = lookupOrAdd(inputExprs, argSquared);
    RelDataType sumType = TypeInferenceUtils.getDrillSqlReturnTypeInference(SqlKind.SUM.name(), ImmutableList.<DrillFuncHolder>of()).inferReturnType(oldCall.createBinding(oldAggRel));
    sumType = typeFactory.createTypeWithNullability(sumType, true);
    final AggregateCall sumArgSquaredAggCall = AggregateCall.create(new DrillCalciteSqlAggFunctionWrapper(new SqlSumAggFunction(sumType), sumType), oldCall.isDistinct(), oldCall.isApproximate(), ImmutableIntList.of(argSquaredOrdinal), -1, sumType, null);
    final RexNode sumArgSquared = rexBuilder.addAggCall(sumArgSquaredAggCall, nGroups, oldAggRel.indicator, newCalls, aggCallMapping, ImmutableList.of(argType));
    final AggregateCall sumArgAggCall = AggregateCall.create(new DrillCalciteSqlAggFunctionWrapper(new SqlSumAggFunction(sumType), sumType), oldCall.isDistinct(), oldCall.isApproximate(), ImmutableIntList.of(argOrdinal), -1, sumType, null);
    final RexNode sumArg = rexBuilder.addAggCall(sumArgAggCall, nGroups, oldAggRel.indicator, newCalls, aggCallMapping, ImmutableList.of(argType));
    final RexNode sumSquaredArg = rexBuilder.makeCall(SqlStdOperatorTable.MULTIPLY, sumArg, sumArg);
    final SqlCountAggFunction countAgg = (SqlCountAggFunction) SqlStdOperatorTable.COUNT;
    final RelDataType countType = countAgg.getReturnType(typeFactory);
    final AggregateCall countArgAggCall = AggregateCall.create(countAgg, oldCall.isDistinct(), oldCall.isApproximate(), oldCall.getArgList(), -1, countType, null);
    final RexNode countArg = rexBuilder.addAggCall(countArgAggCall, nGroups, oldAggRel.indicator, newCalls, aggCallMapping, ImmutableList.of(argType));
    final RexNode avgSumSquaredArg = rexBuilder.makeCall(SqlStdOperatorTable.DIVIDE, sumSquaredArg, countArg);
    final RexNode diff = rexBuilder.makeCall(SqlStdOperatorTable.MINUS, sumArgSquared, avgSumSquaredArg);
    final RexNode denominator;
    if (biased) {
        denominator = countArg;
    } else {
        final RexLiteral one = rexBuilder.makeExactLiteral(BigDecimal.ONE);
        final RexNode nul = rexBuilder.makeNullLiteral(countArg.getType());
        final RexNode countMinusOne = rexBuilder.makeCall(SqlStdOperatorTable.MINUS, countArg, one);
        final RexNode countEqOne = rexBuilder.makeCall(SqlStdOperatorTable.EQUALS, countArg, one);
        denominator = rexBuilder.makeCall(SqlStdOperatorTable.CASE, countEqOne, nul, countMinusOne);
    }
    final SqlOperator divide;
    if (isInferenceEnabled) {
        divide = new DrillSqlOperator("divide", 2, true, oldCall.getType(), false);
    } else {
        divide = SqlStdOperatorTable.DIVIDE;
    }
    final RexNode div = rexBuilder.makeCall(divide, diff, denominator);
    RexNode result = div;
    if (sqrt) {
        final RexNode half = rexBuilder.makeExactLiteral(new BigDecimal("0.5"));
        result = rexBuilder.makeCall(SqlStdOperatorTable.POWER, div, half);
    }
    if (isInferenceEnabled) {
        return result;
    } else {
        /*
      * Currently calcite's strategy to infer the return type of aggregate functions
      * is wrong because it uses the first known argument to determine output type. For
      * instance if we are performing stddev on an integer column then it interprets the
      * output type to be integer which is incorrect as it should be double. So based on
      * this if we add cast after rewriting the aggregate we add an additional cast which
      * would cause wrong results. So we simply add a cast to ANY.
      */
        return rexBuilder.makeCast(typeFactory.createSqlType(SqlTypeName.ANY), result);
    }
}

Example 3 with SqlCountAggFunction

use of org.apache.calcite.sql.fun.SqlCountAggFunction in project drill by axbaretto.

the class DrillReduceAggregatesRule method reduceAvg.

private RexNode reduceAvg(Aggregate oldAggRel, AggregateCall oldCall, List<AggregateCall> newCalls, Map<AggregateCall, RexNode> aggCallMapping) {
    final PlannerSettings plannerSettings = (PlannerSettings) oldAggRel.getCluster().getPlanner().getContext();
    final boolean isInferenceEnabled = plannerSettings.isTypeInferenceEnabled();
    final int nGroups = oldAggRel.getGroupCount();
    RelDataTypeFactory typeFactory = oldAggRel.getCluster().getTypeFactory();
    RexBuilder rexBuilder = oldAggRel.getCluster().getRexBuilder();
    int iAvgInput = oldCall.getArgList().get(0);
    RelDataType avgInputType = getFieldType(oldAggRel.getInput(), iAvgInput);
    RelDataType sumType = TypeInferenceUtils.getDrillSqlReturnTypeInference(SqlKind.SUM.name(), ImmutableList.<DrillFuncHolder>of()).inferReturnType(oldCall.createBinding(oldAggRel));
    sumType = typeFactory.createTypeWithNullability(sumType, sumType.isNullable() || nGroups == 0);
    SqlAggFunction sumAgg = new DrillCalciteSqlAggFunctionWrapper(new SqlSumEmptyIsZeroAggFunction(), sumType);
    AggregateCall sumCall = AggregateCall.create(sumAgg, oldCall.isDistinct(), oldCall.isApproximate(), oldCall.getArgList(), -1, sumType, null);
    final SqlCountAggFunction countAgg = (SqlCountAggFunction) SqlStdOperatorTable.COUNT;
    final RelDataType countType = countAgg.getReturnType(typeFactory);
    AggregateCall countCall = AggregateCall.create(countAgg, oldCall.isDistinct(), oldCall.isApproximate(), oldCall.getArgList(), -1, countType, null);
    RexNode tmpsumRef = rexBuilder.addAggCall(sumCall, nGroups, oldAggRel.indicator, newCalls, aggCallMapping, ImmutableList.of(avgInputType));
    RexNode tmpcountRef = rexBuilder.addAggCall(countCall, nGroups, oldAggRel.indicator, newCalls, aggCallMapping, ImmutableList.of(avgInputType));
    RexNode n = rexBuilder.makeCall(SqlStdOperatorTable.CASE, rexBuilder.makeCall(SqlStdOperatorTable.EQUALS, tmpcountRef, rexBuilder.makeExactLiteral(BigDecimal.ZERO)), rexBuilder.constantNull(), tmpsumRef);
    // NOTE:  these references are with respect to the output
    // of newAggRel
    /*
    RexNode numeratorRef =
        rexBuilder.makeCall(CastHighOp,
          rexBuilder.addAggCall(
              sumCall,
              nGroups,
              newCalls,
              aggCallMapping,
              ImmutableList.of(avgInputType))
        );
    */
    RexNode numeratorRef = rexBuilder.makeCall(CastHighOp, n);
    RexNode denominatorRef = rexBuilder.addAggCall(countCall, nGroups, oldAggRel.indicator, newCalls, aggCallMapping, ImmutableList.of(avgInputType));
    if (isInferenceEnabled) {
        return rexBuilder.makeCall(new DrillSqlOperator("divide", 2, true, oldCall.getType(), false), numeratorRef, denominatorRef);
    } else {
        final RexNode divideRef = rexBuilder.makeCall(SqlStdOperatorTable.DIVIDE, numeratorRef, denominatorRef);
        return rexBuilder.makeCast(typeFactory.createSqlType(SqlTypeName.ANY), divideRef);
    }
}

Example 4 with SqlCountAggFunction

use of org.apache.calcite.sql.fun.SqlCountAggFunction in project drill by apache.

the class DrillReduceAggregatesRule method reduceSum.

private RexNode reduceSum(Aggregate oldAggRel, AggregateCall oldCall, List<AggregateCall> newCalls, Map<AggregateCall, RexNode> aggCallMapping) {
    final PlannerSettings plannerSettings = (PlannerSettings) oldAggRel.getCluster().getPlanner().getContext();
    final boolean isInferenceEnabled = plannerSettings.isTypeInferenceEnabled();
    final int nGroups = oldAggRel.getGroupCount();
    RelDataTypeFactory typeFactory = oldAggRel.getCluster().getTypeFactory();
    RexBuilder rexBuilder = oldAggRel.getCluster().getRexBuilder();
    int arg = oldCall.getArgList().get(0);
    RelDataType argType = getFieldType(oldAggRel.getInput(), arg);
    final RelDataType sumType;
    final SqlAggFunction sumZeroAgg;
    if (isInferenceEnabled) {
        sumType = oldCall.getType();
    } else {
        sumType = typeFactory.createTypeWithNullability(oldCall.getType(), argType.isNullable());
    }
    sumZeroAgg = new DrillCalciteSqlAggFunctionWrapper(new SqlSumEmptyIsZeroAggFunction(), sumType);
    AggregateCall sumZeroCall = AggregateCall.create(sumZeroAgg, oldCall.isDistinct(), oldCall.isApproximate(), oldCall.getArgList(), -1, sumType, null);
    final SqlCountAggFunction countAgg = (SqlCountAggFunction) SqlStdOperatorTable.COUNT;
    final RelDataType countType = countAgg.getReturnType(typeFactory);
    AggregateCall countCall = AggregateCall.create(countAgg, oldCall.isDistinct(), oldCall.isApproximate(), oldCall.getArgList(), -1, countType, null);
    // NOTE:  these references are with respect to the output
    // of newAggRel
    RexNode sumZeroRef = rexBuilder.addAggCall(sumZeroCall, nGroups, newCalls, aggCallMapping, ImmutableList.of(argType));
    if (!oldCall.getType().isNullable()) {
        // null). Therefore we translate to SUM0(x).
        return sumZeroRef;
    }
    RexNode countRef = rexBuilder.addAggCall(countCall, nGroups, newCalls, aggCallMapping, ImmutableList.of(argType));
    return rexBuilder.makeCall(SqlStdOperatorTable.CASE, rexBuilder.makeCall(SqlStdOperatorTable.EQUALS, countRef, rexBuilder.makeExactLiteral(BigDecimal.ZERO)), rexBuilder.constantNull(), sumZeroRef);
}

Example 5 with SqlCountAggFunction

use of org.apache.calcite.sql.fun.SqlCountAggFunction in project drill by apache.

the class DrillReduceAggregatesRule method reduceStddev.

private RexNode reduceStddev(Aggregate oldAggRel, AggregateCall oldCall, boolean biased, boolean sqrt, List<AggregateCall> newCalls, Map<AggregateCall, RexNode> aggCallMapping, List<RexNode> inputExprs) {
    // stddev_pop(x) ==>
    // power(
    // (sum(x * x) - sum(x) * sum(x) / count(x))
    // / count(x),
    // .5)
    // 
    // stddev_samp(x) ==>
    // power(
    // (sum(x * x) - sum(x) * sum(x) / count(x))
    // / nullif(count(x) - 1, 0),
    // .5)
    final PlannerSettings plannerSettings = (PlannerSettings) oldAggRel.getCluster().getPlanner().getContext();
    final boolean isInferenceEnabled = plannerSettings.isTypeInferenceEnabled();
    final int nGroups = oldAggRel.getGroupCount();
    RelDataTypeFactory typeFactory = oldAggRel.getCluster().getTypeFactory();
    final RexBuilder rexBuilder = oldAggRel.getCluster().getRexBuilder();
    assert oldCall.getArgList().size() == 1 : oldCall.getArgList();
    final int argOrdinal = oldCall.getArgList().get(0);
    final RelDataType argType = getFieldType(oldAggRel.getInput(), argOrdinal);
    // final RexNode argRef = inputExprs.get(argOrdinal);
    RexNode argRef = rexBuilder.makeCall(CastHighOp, inputExprs.get(argOrdinal));
    inputExprs.set(argOrdinal, argRef);
    final RexNode argSquared = rexBuilder.makeCall(SqlStdOperatorTable.MULTIPLY, argRef, argRef);
    final int argSquaredOrdinal = lookupOrAdd(inputExprs, argSquared);
    RelDataType sumType = TypeInferenceUtils.getDrillSqlReturnTypeInference(SqlKind.SUM.name(), ImmutableList.of()).inferReturnType(oldCall.createBinding(oldAggRel));
    sumType = typeFactory.createTypeWithNullability(sumType, true);
    final AggregateCall sumArgSquaredAggCall = AggregateCall.create(new DrillCalciteSqlAggFunctionWrapper(new SqlSumAggFunction(sumType), sumType), oldCall.isDistinct(), oldCall.isApproximate(), ImmutableIntList.of(argSquaredOrdinal), -1, sumType, null);
    final RexNode sumArgSquared = rexBuilder.addAggCall(sumArgSquaredAggCall, nGroups, newCalls, aggCallMapping, ImmutableList.of(argType));
    final AggregateCall sumArgAggCall = AggregateCall.create(new DrillCalciteSqlAggFunctionWrapper(new SqlSumAggFunction(sumType), sumType), oldCall.isDistinct(), oldCall.isApproximate(), ImmutableIntList.of(argOrdinal), -1, sumType, null);
    final RexNode sumArg = rexBuilder.addAggCall(sumArgAggCall, nGroups, newCalls, aggCallMapping, ImmutableList.of(argType));
    final RexNode sumSquaredArg = rexBuilder.makeCall(SqlStdOperatorTable.MULTIPLY, sumArg, sumArg);
    final SqlCountAggFunction countAgg = (SqlCountAggFunction) SqlStdOperatorTable.COUNT;
    final RelDataType countType = countAgg.getReturnType(typeFactory);
    final AggregateCall countArgAggCall = AggregateCall.create(countAgg, oldCall.isDistinct(), oldCall.isApproximate(), oldCall.getArgList(), -1, countType, null);
    final RexNode countArg = rexBuilder.addAggCall(countArgAggCall, nGroups, newCalls, aggCallMapping, ImmutableList.of(argType));
    final RexNode avgSumSquaredArg = rexBuilder.makeCall(SqlStdOperatorTable.DIVIDE, sumSquaredArg, countArg);
    final RexNode diff = rexBuilder.makeCall(SqlStdOperatorTable.MINUS, sumArgSquared, avgSumSquaredArg);
    final RexNode denominator;
    if (biased) {
        denominator = countArg;
    } else {
        final RexLiteral one = rexBuilder.makeExactLiteral(BigDecimal.ONE);
        final RexNode nul = rexBuilder.makeNullLiteral(countArg.getType());
        final RexNode countMinusOne = rexBuilder.makeCall(SqlStdOperatorTable.MINUS, countArg, one);
        final RexNode countEqOne = rexBuilder.makeCall(SqlStdOperatorTable.EQUALS, countArg, one);
        denominator = rexBuilder.makeCall(SqlStdOperatorTable.CASE, countEqOne, nul, countMinusOne);
    }
    final SqlOperator divide;
    if (isInferenceEnabled) {
        divide = new DrillSqlOperator("divide", 2, true, oldCall.getType(), false);
    } else {
        divide = SqlStdOperatorTable.DIVIDE;
    }
    final RexNode div = rexBuilder.makeCall(divide, diff, denominator);
    RexNode result = div;
    if (sqrt) {
        final RexNode half = rexBuilder.makeExactLiteral(new BigDecimal("0.5"));
        result = rexBuilder.makeCall(SqlStdOperatorTable.POWER, div, half);
    }
    if (isInferenceEnabled) {
        return result;
    } else {
        /*
      * Currently calcite's strategy to infer the return type of aggregate functions
      * is wrong because it uses the first known argument to determine output type. For
      * instance if we are performing stddev on an integer column then it interprets the
      * output type to be integer which is incorrect as it should be double. So based on
      * this if we add cast after rewriting the aggregate we add an additional cast which
      * would cause wrong results. So we simply add a cast to ANY.
      */
        return rexBuilder.makeCast(typeFactory.createSqlType(SqlTypeName.ANY), result);
    }
}