Search in sources :

Example 1 with WeightedSquaredLoss

use of org.apache.sysml.lops.WeightedSquaredLoss in project systemml by apache.

the class QuaternaryOp method constructSparkLopsWeightedSquaredLoss.

private void constructSparkLopsWeightedSquaredLoss(WeightsType wtype) {
    // NOTE: the common case for wsloss are factors U/V with a rank of 10s to 100s; the current runtime only
    // supports single block outer products (U/V rank <= blocksize, i.e., 1000 by default); we enforce this
    // by applying the hop rewrite for Weighted Squared Loss only if this constraint holds.
    // Notes: Any broadcast needs to fit twice in local memory because we partition the input in cp,
    // and needs to fit once in executor broadcast memory. The 2GB broadcast constraint is no longer
    // required because the max_int byte buffer constraint has been fixed in Spark 1.4
    double memBudgetExec = SparkExecutionContext.getBroadcastMemoryBudget();
    double memBudgetLocal = OptimizerUtils.getLocalMemBudget();
    Hop X = getInput().get(0);
    Hop U = getInput().get(1);
    Hop V = getInput().get(2);
    Hop W = getInput().get(3);
    // MR operator selection, part1
    // size U
    double m1Size = OptimizerUtils.estimateSize(U.getDim1(), U.getDim2());
    // size V
    double m2Size = OptimizerUtils.estimateSize(V.getDim1(), V.getDim2());
    boolean isMapWsloss = (!wtype.hasFourInputs() && m1Size + m2Size < memBudgetExec && 2 * m1Size < memBudgetLocal && 2 * m2Size < memBudgetLocal);
    if (// broadcast
    !FORCE_REPLICATION && isMapWsloss) {
        // map-side wsloss always with broadcast
        Lop wsloss = new WeightedSquaredLoss(X.constructLops(), U.constructLops(), V.constructLops(), W.constructLops(), DataType.SCALAR, ValueType.DOUBLE, wtype, ExecType.SPARK);
        setOutputDimensions(wsloss);
        setLineNumbers(wsloss);
        setLops(wsloss);
    } else // general case
    {
        // MR operator selection part 2
        boolean cacheU = !FORCE_REPLICATION && (m1Size < memBudgetExec && 2 * m1Size < memBudgetLocal);
        boolean cacheV = !FORCE_REPLICATION && ((!cacheU && m2Size < memBudgetExec) || (cacheU && m1Size + m2Size < memBudgetExec)) && 2 * m2Size < memBudgetLocal;
        // reduce-side wsloss w/ or without broadcast
        Lop wsloss = new WeightedSquaredLossR(X.constructLops(), U.constructLops(), V.constructLops(), W.constructLops(), DataType.SCALAR, ValueType.DOUBLE, wtype, cacheU, cacheV, ExecType.SPARK);
        setOutputDimensions(wsloss);
        setLineNumbers(wsloss);
        setLops(wsloss);
    }
}
Also used : MultiThreadedHop(org.apache.sysml.hops.Hop.MultiThreadedHop) WeightedSquaredLoss(org.apache.sysml.lops.WeightedSquaredLoss) Lop(org.apache.sysml.lops.Lop) WeightedSquaredLossR(org.apache.sysml.lops.WeightedSquaredLossR)

Example 2 with WeightedSquaredLoss

use of org.apache.sysml.lops.WeightedSquaredLoss in project systemml by apache.

the class QuaternaryOp method constructMRLopsWeightedSquaredLoss.

private void constructMRLopsWeightedSquaredLoss(WeightsType wtype) {
    // NOTE: the common case for wsloss are factors U/V with a rank of 10s to 100s; the current runtime only
    // supports single block outer products (U/V rank <= blocksize, i.e., 1000 by default); we enforce this
    // by applying the hop rewrite for Weighted Squared Loss only if this constraint holds.
    Hop X = getInput().get(0);
    Hop U = getInput().get(1);
    Hop V = getInput().get(2);
    Hop W = getInput().get(3);
    // MR operator selection, part1
    // size U
    double m1Size = OptimizerUtils.estimateSize(U.getDim1(), U.getDim2());
    // size V
    double m2Size = OptimizerUtils.estimateSize(V.getDim1(), V.getDim2());
    boolean isMapWsloss = (!wtype.hasFourInputs() && m1Size + m2Size < OptimizerUtils.getRemoteMemBudgetMap(true));
    if (// broadcast
    !FORCE_REPLICATION && isMapWsloss) {
        // partitioning of U
        boolean needPartU = !U.dimsKnown() || U.getDim1() * U.getDim2() > DistributedCacheInput.PARTITION_SIZE;
        Lop lU = U.constructLops();
        if (needPartU) {
            // requires partitioning
            lU = new DataPartition(lU, DataType.MATRIX, ValueType.DOUBLE, (m1Size > OptimizerUtils.getLocalMemBudget()) ? ExecType.MR : ExecType.CP, PDataPartitionFormat.ROW_BLOCK_WISE_N);
            lU.getOutputParameters().setDimensions(U.getDim1(), U.getDim2(), getRowsInBlock(), getColsInBlock(), U.getNnz());
            setLineNumbers(lU);
        }
        // partitioning of V
        boolean needPartV = !V.dimsKnown() || V.getDim1() * V.getDim2() > DistributedCacheInput.PARTITION_SIZE;
        Lop lV = V.constructLops();
        if (needPartV) {
            // requires partitioning
            lV = new DataPartition(lV, DataType.MATRIX, ValueType.DOUBLE, (m2Size > OptimizerUtils.getLocalMemBudget()) ? ExecType.MR : ExecType.CP, PDataPartitionFormat.ROW_BLOCK_WISE_N);
            lV.getOutputParameters().setDimensions(V.getDim1(), V.getDim2(), getRowsInBlock(), getColsInBlock(), V.getNnz());
            setLineNumbers(lV);
        }
        // map-side wsloss always with broadcast
        Lop wsloss = new WeightedSquaredLoss(X.constructLops(), lU, lV, W.constructLops(), DataType.MATRIX, ValueType.DOUBLE, wtype, ExecType.MR);
        wsloss.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(wsloss);
        Group grp = new Group(wsloss, Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
        grp.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(grp);
        Aggregate agg1 = new Aggregate(grp, HopsAgg2Lops.get(AggOp.SUM), DataType.MATRIX, ValueType.DOUBLE, ExecType.MR);
        // aggregation uses kahanSum
        agg1.setupCorrectionLocation(CorrectionLocationType.NONE);
        agg1.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(agg1);
        UnaryCP unary1 = new UnaryCP(agg1, HopsOpOp1LopsUS.get(OpOp1.CAST_AS_SCALAR), getDataType(), getValueType());
        unary1.getOutputParameters().setDimensions(0, 0, 0, 0, -1);
        setLineNumbers(unary1);
        setLops(unary1);
    } else // general case
    {
        // MR operator selection part 2
        boolean cacheU = !FORCE_REPLICATION && (m1Size < OptimizerUtils.getRemoteMemBudgetReduce());
        boolean cacheV = !FORCE_REPLICATION && ((!cacheU && m2Size < OptimizerUtils.getRemoteMemBudgetReduce()) || (cacheU && m1Size + m2Size < OptimizerUtils.getRemoteMemBudgetReduce()));
        Group grpX = new Group(X.constructLops(), Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
        grpX.getOutputParameters().setDimensions(X.getDim1(), X.getDim2(), X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(grpX);
        Lop grpW = W.constructLops();
        if (grpW.getDataType() == DataType.MATRIX) {
            grpW = new Group(W.constructLops(), Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
            grpW.getOutputParameters().setDimensions(W.getDim1(), W.getDim2(), W.getRowsInBlock(), W.getColsInBlock(), -1);
            setLineNumbers(grpW);
        }
        Lop lU = constructLeftFactorMRLop(U, V, cacheU, m1Size);
        Lop lV = constructRightFactorMRLop(U, V, cacheV, m2Size);
        // reduce-side wsloss w/ or without broadcast
        Lop wsloss = new WeightedSquaredLossR(grpX, lU, lV, grpW, DataType.MATRIX, ValueType.DOUBLE, wtype, cacheU, cacheV, ExecType.MR);
        wsloss.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(wsloss);
        Group grp = new Group(wsloss, Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
        grp.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(grp);
        Aggregate agg1 = new Aggregate(grp, HopsAgg2Lops.get(AggOp.SUM), DataType.MATRIX, ValueType.DOUBLE, ExecType.MR);
        // aggregation uses kahanSum
        agg1.setupCorrectionLocation(CorrectionLocationType.NONE);
        agg1.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(agg1);
        UnaryCP unary1 = new UnaryCP(agg1, HopsOpOp1LopsUS.get(OpOp1.CAST_AS_SCALAR), getDataType(), getValueType());
        unary1.getOutputParameters().setDimensions(0, 0, 0, 0, -1);
        setLineNumbers(unary1);
        setLops(unary1);
    }
}
Also used : Group(org.apache.sysml.lops.Group) MultiThreadedHop(org.apache.sysml.hops.Hop.MultiThreadedHop) WeightedSquaredLoss(org.apache.sysml.lops.WeightedSquaredLoss) Lop(org.apache.sysml.lops.Lop) Aggregate(org.apache.sysml.lops.Aggregate) DataPartition(org.apache.sysml.lops.DataPartition) WeightedSquaredLossR(org.apache.sysml.lops.WeightedSquaredLossR) UnaryCP(org.apache.sysml.lops.UnaryCP)

Example 3 with WeightedSquaredLoss

use of org.apache.sysml.lops.WeightedSquaredLoss in project incubator-systemml by apache.

the class QuaternaryOp method constructMRLopsWeightedSquaredLoss.

private void constructMRLopsWeightedSquaredLoss(WeightsType wtype) {
    // NOTE: the common case for wsloss are factors U/V with a rank of 10s to 100s; the current runtime only
    // supports single block outer products (U/V rank <= blocksize, i.e., 1000 by default); we enforce this
    // by applying the hop rewrite for Weighted Squared Loss only if this constraint holds.
    Hop X = getInput().get(0);
    Hop U = getInput().get(1);
    Hop V = getInput().get(2);
    Hop W = getInput().get(3);
    // MR operator selection, part1
    // size U
    double m1Size = OptimizerUtils.estimateSize(U.getDim1(), U.getDim2());
    // size V
    double m2Size = OptimizerUtils.estimateSize(V.getDim1(), V.getDim2());
    boolean isMapWsloss = (!wtype.hasFourInputs() && m1Size + m2Size < OptimizerUtils.getRemoteMemBudgetMap(true));
    if (// broadcast
    !FORCE_REPLICATION && isMapWsloss) {
        // partitioning of U
        boolean needPartU = !U.dimsKnown() || U.getDim1() * U.getDim2() > DistributedCacheInput.PARTITION_SIZE;
        Lop lU = U.constructLops();
        if (needPartU) {
            // requires partitioning
            lU = new DataPartition(lU, DataType.MATRIX, ValueType.DOUBLE, (m1Size > OptimizerUtils.getLocalMemBudget()) ? ExecType.MR : ExecType.CP, PDataPartitionFormat.ROW_BLOCK_WISE_N);
            lU.getOutputParameters().setDimensions(U.getDim1(), U.getDim2(), getRowsInBlock(), getColsInBlock(), U.getNnz());
            setLineNumbers(lU);
        }
        // partitioning of V
        boolean needPartV = !V.dimsKnown() || V.getDim1() * V.getDim2() > DistributedCacheInput.PARTITION_SIZE;
        Lop lV = V.constructLops();
        if (needPartV) {
            // requires partitioning
            lV = new DataPartition(lV, DataType.MATRIX, ValueType.DOUBLE, (m2Size > OptimizerUtils.getLocalMemBudget()) ? ExecType.MR : ExecType.CP, PDataPartitionFormat.ROW_BLOCK_WISE_N);
            lV.getOutputParameters().setDimensions(V.getDim1(), V.getDim2(), getRowsInBlock(), getColsInBlock(), V.getNnz());
            setLineNumbers(lV);
        }
        // map-side wsloss always with broadcast
        Lop wsloss = new WeightedSquaredLoss(X.constructLops(), lU, lV, W.constructLops(), DataType.MATRIX, ValueType.DOUBLE, wtype, ExecType.MR);
        wsloss.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(wsloss);
        Group grp = new Group(wsloss, Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
        grp.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(grp);
        Aggregate agg1 = new Aggregate(grp, HopsAgg2Lops.get(AggOp.SUM), DataType.MATRIX, ValueType.DOUBLE, ExecType.MR);
        // aggregation uses kahanSum
        agg1.setupCorrectionLocation(CorrectionLocationType.NONE);
        agg1.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(agg1);
        UnaryCP unary1 = new UnaryCP(agg1, HopsOpOp1LopsUS.get(OpOp1.CAST_AS_SCALAR), getDataType(), getValueType());
        unary1.getOutputParameters().setDimensions(0, 0, 0, 0, -1);
        setLineNumbers(unary1);
        setLops(unary1);
    } else // general case
    {
        // MR operator selection part 2
        boolean cacheU = !FORCE_REPLICATION && (m1Size < OptimizerUtils.getRemoteMemBudgetReduce());
        boolean cacheV = !FORCE_REPLICATION && ((!cacheU && m2Size < OptimizerUtils.getRemoteMemBudgetReduce()) || (cacheU && m1Size + m2Size < OptimizerUtils.getRemoteMemBudgetReduce()));
        Group grpX = new Group(X.constructLops(), Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
        grpX.getOutputParameters().setDimensions(X.getDim1(), X.getDim2(), X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(grpX);
        Lop grpW = W.constructLops();
        if (grpW.getDataType() == DataType.MATRIX) {
            grpW = new Group(W.constructLops(), Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
            grpW.getOutputParameters().setDimensions(W.getDim1(), W.getDim2(), W.getRowsInBlock(), W.getColsInBlock(), -1);
            setLineNumbers(grpW);
        }
        Lop lU = constructLeftFactorMRLop(U, V, cacheU, m1Size);
        Lop lV = constructRightFactorMRLop(U, V, cacheV, m2Size);
        // reduce-side wsloss w/ or without broadcast
        Lop wsloss = new WeightedSquaredLossR(grpX, lU, lV, grpW, DataType.MATRIX, ValueType.DOUBLE, wtype, cacheU, cacheV, ExecType.MR);
        wsloss.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(wsloss);
        Group grp = new Group(wsloss, Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
        grp.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(grp);
        Aggregate agg1 = new Aggregate(grp, HopsAgg2Lops.get(AggOp.SUM), DataType.MATRIX, ValueType.DOUBLE, ExecType.MR);
        // aggregation uses kahanSum
        agg1.setupCorrectionLocation(CorrectionLocationType.NONE);
        agg1.getOutputParameters().setDimensions(1, 1, X.getRowsInBlock(), X.getColsInBlock(), -1);
        setLineNumbers(agg1);
        UnaryCP unary1 = new UnaryCP(agg1, HopsOpOp1LopsUS.get(OpOp1.CAST_AS_SCALAR), getDataType(), getValueType());
        unary1.getOutputParameters().setDimensions(0, 0, 0, 0, -1);
        setLineNumbers(unary1);
        setLops(unary1);
    }
}
Also used : Group(org.apache.sysml.lops.Group) MultiThreadedHop(org.apache.sysml.hops.Hop.MultiThreadedHop) WeightedSquaredLoss(org.apache.sysml.lops.WeightedSquaredLoss) Lop(org.apache.sysml.lops.Lop) Aggregate(org.apache.sysml.lops.Aggregate) DataPartition(org.apache.sysml.lops.DataPartition) WeightedSquaredLossR(org.apache.sysml.lops.WeightedSquaredLossR) UnaryCP(org.apache.sysml.lops.UnaryCP)

Example 4 with WeightedSquaredLoss

use of org.apache.sysml.lops.WeightedSquaredLoss in project incubator-systemml by apache.

the class QuaternaryOp method constructCPLopsWeightedSquaredLoss.

private void constructCPLopsWeightedSquaredLoss(WeightsType wtype) {
    WeightedSquaredLoss wsloss = new WeightedSquaredLoss(getInput().get(0).constructLops(), getInput().get(1).constructLops(), getInput().get(2).constructLops(), getInput().get(3).constructLops(), getDataType(), getValueType(), wtype, ExecType.CP);
    // set degree of parallelism
    int k = OptimizerUtils.getConstrainedNumThreads(_maxNumThreads);
    wsloss.setNumThreads(k);
    setOutputDimensions(wsloss);
    setLineNumbers(wsloss);
    setLops(wsloss);
}
Also used : WeightedSquaredLoss(org.apache.sysml.lops.WeightedSquaredLoss)

Example 5 with WeightedSquaredLoss

use of org.apache.sysml.lops.WeightedSquaredLoss in project incubator-systemml by apache.

the class QuaternaryOp method constructSparkLopsWeightedSquaredLoss.

private void constructSparkLopsWeightedSquaredLoss(WeightsType wtype) {
    // NOTE: the common case for wsloss are factors U/V with a rank of 10s to 100s; the current runtime only
    // supports single block outer products (U/V rank <= blocksize, i.e., 1000 by default); we enforce this
    // by applying the hop rewrite for Weighted Squared Loss only if this constraint holds.
    // Notes: Any broadcast needs to fit twice in local memory because we partition the input in cp,
    // and needs to fit once in executor broadcast memory. The 2GB broadcast constraint is no longer
    // required because the max_int byte buffer constraint has been fixed in Spark 1.4
    double memBudgetExec = SparkExecutionContext.getBroadcastMemoryBudget();
    double memBudgetLocal = OptimizerUtils.getLocalMemBudget();
    Hop X = getInput().get(0);
    Hop U = getInput().get(1);
    Hop V = getInput().get(2);
    Hop W = getInput().get(3);
    // MR operator selection, part1
    // size U
    double m1Size = OptimizerUtils.estimateSize(U.getDim1(), U.getDim2());
    // size V
    double m2Size = OptimizerUtils.estimateSize(V.getDim1(), V.getDim2());
    boolean isMapWsloss = (!wtype.hasFourInputs() && m1Size + m2Size < memBudgetExec && 2 * m1Size < memBudgetLocal && 2 * m2Size < memBudgetLocal);
    if (// broadcast
    !FORCE_REPLICATION && isMapWsloss) {
        // map-side wsloss always with broadcast
        Lop wsloss = new WeightedSquaredLoss(X.constructLops(), U.constructLops(), V.constructLops(), W.constructLops(), DataType.SCALAR, ValueType.DOUBLE, wtype, ExecType.SPARK);
        setOutputDimensions(wsloss);
        setLineNumbers(wsloss);
        setLops(wsloss);
    } else // general case
    {
        // MR operator selection part 2
        boolean cacheU = !FORCE_REPLICATION && (m1Size < memBudgetExec && 2 * m1Size < memBudgetLocal);
        boolean cacheV = !FORCE_REPLICATION && ((!cacheU && m2Size < memBudgetExec) || (cacheU && m1Size + m2Size < memBudgetExec)) && 2 * m2Size < memBudgetLocal;
        // reduce-side wsloss w/ or without broadcast
        Lop wsloss = new WeightedSquaredLossR(X.constructLops(), U.constructLops(), V.constructLops(), W.constructLops(), DataType.SCALAR, ValueType.DOUBLE, wtype, cacheU, cacheV, ExecType.SPARK);
        setOutputDimensions(wsloss);
        setLineNumbers(wsloss);
        setLops(wsloss);
    }
}
Also used : MultiThreadedHop(org.apache.sysml.hops.Hop.MultiThreadedHop) WeightedSquaredLoss(org.apache.sysml.lops.WeightedSquaredLoss) Lop(org.apache.sysml.lops.Lop) WeightedSquaredLossR(org.apache.sysml.lops.WeightedSquaredLossR)

Aggregations

WeightedSquaredLoss (org.apache.sysml.lops.WeightedSquaredLoss)6 MultiThreadedHop (org.apache.sysml.hops.Hop.MultiThreadedHop)4 Lop (org.apache.sysml.lops.Lop)4 WeightedSquaredLossR (org.apache.sysml.lops.WeightedSquaredLossR)4 Aggregate (org.apache.sysml.lops.Aggregate)2 DataPartition (org.apache.sysml.lops.DataPartition)2 Group (org.apache.sysml.lops.Group)2 UnaryCP (org.apache.sysml.lops.UnaryCP)2