Examples with WeightedUnaryMMR org.apache.sysml.lops.WeightedUnaryMMR used on opensource projects

Example 1 with WeightedUnaryMMR

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

the class QuaternaryOp method constructSparkLopsWeightedUMM.

private void constructSparkLopsWeightedUMM(WUMMType wtype) throws HopsException, LopsException {
    //NOTE: the common case for wumm 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 UnaryMM only if this constraint holds. 
    Unary.OperationTypes uop = _uop != null ? HopsOpOp1LopsU.get(_uop) : _sop == OpOp2.POW ? Unary.OperationTypes.POW2 : Unary.OperationTypes.MULTIPLY2;
    //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);
    //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 = (m1Size + m2Size < memBudgetExec && 2 * m1Size < memBudgetLocal && 2 * m2Size < memBudgetLocal);
    if (//broadcast
    !FORCE_REPLICATION && isMapWsloss) {
        //map-side wumm always with broadcast
        Lop wumm = new WeightedUnaryMM(X.constructLops(), U.constructLops(), V.constructLops(), DataType.MATRIX, ValueType.DOUBLE, wtype, uop, ExecType.SPARK);
        setOutputDimensions(wumm);
        setLineNumbers(wumm);
        setLops(wumm);
    } 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 wumm w/ or without broadcast
        Lop wumm = new WeightedUnaryMMR(X.constructLops(), U.constructLops(), V.constructLops(), DataType.MATRIX, ValueType.DOUBLE, wtype, uop, cacheU, cacheV, ExecType.SPARK);
        setOutputDimensions(wumm);
        setLineNumbers(wumm);
        setLops(wumm);
    }
}

Example 2 with WeightedUnaryMMR

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

the class QuaternaryOp method constructMRLopsWeightedUMM.

private void constructMRLopsWeightedUMM(WUMMType wtype) throws HopsException, LopsException {
    //NOTE: the common case for wumm 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 UnaryMM  only if this constraint holds. 
    Unary.OperationTypes uop = _uop != null ? HopsOpOp1LopsU.get(_uop) : _sop == OpOp2.POW ? Unary.OperationTypes.POW2 : Unary.OperationTypes.MULTIPLY2;
    Hop X = getInput().get(0);
    Hop U = getInput().get(1);
    Hop V = getInput().get(2);
    //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 isMapWumm = (m1Size + m2Size < OptimizerUtils.getRemoteMemBudgetMap(true));
    if (//broadcast
    !FORCE_REPLICATION && isMapWumm) {
        //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 wumm always with broadcast
        Lop wumm = new WeightedUnaryMM(X.constructLops(), lU, lV, DataType.MATRIX, ValueType.DOUBLE, wtype, uop, ExecType.MR);
        setOutputDimensions(wumm);
        setLineNumbers(wumm);
        setLops(wumm);
    //in contrast to wsloss no aggregation required 
    } 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(), X.getNnz());
        setLineNumbers(grpX);
        Lop lU = null;
        if (cacheU) {
            //partitioning of U for read through distributed cache
            boolean needPartU = !U.dimsKnown() || U.getDim1() * U.getDim2() > DistributedCacheInput.PARTITION_SIZE;
            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);
            }
        } else {
            //replication of U for shuffle to target block
            //ncol of t(V) -> nrow of V determines num replicates
            Lop offset = createOffsetLop(V, false);
            lU = new RepMat(U.constructLops(), offset, true, V.getDataType(), V.getValueType());
            lU.getOutputParameters().setDimensions(U.getDim1(), U.getDim2(), U.getRowsInBlock(), U.getColsInBlock(), U.getNnz());
            setLineNumbers(lU);
            Group grpU = new Group(lU, Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
            grpU.getOutputParameters().setDimensions(U.getDim1(), U.getDim2(), U.getRowsInBlock(), U.getColsInBlock(), -1);
            setLineNumbers(grpU);
            lU = grpU;
        }
        Lop lV = null;
        if (cacheV) {
            //partitioning of V for read through distributed cache
            boolean needPartV = !V.dimsKnown() || V.getDim1() * V.getDim2() > DistributedCacheInput.PARTITION_SIZE;
            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);
            }
        } else {
            //replication of t(V) for shuffle to target block
            Transform ltV = new Transform(V.constructLops(), HopsTransf2Lops.get(ReOrgOp.TRANSPOSE), getDataType(), getValueType(), ExecType.MR);
            ltV.getOutputParameters().setDimensions(V.getDim2(), V.getDim1(), V.getColsInBlock(), V.getRowsInBlock(), V.getNnz());
            setLineNumbers(ltV);
            //nrow of U determines num replicates
            Lop offset = createOffsetLop(U, false);
            lV = new RepMat(ltV, offset, false, V.getDataType(), V.getValueType());
            lV.getOutputParameters().setDimensions(V.getDim2(), V.getDim1(), V.getColsInBlock(), V.getRowsInBlock(), V.getNnz());
            setLineNumbers(lV);
            Group grpV = new Group(lV, Group.OperationTypes.Sort, DataType.MATRIX, ValueType.DOUBLE);
            grpV.getOutputParameters().setDimensions(V.getDim2(), V.getDim1(), V.getColsInBlock(), V.getRowsInBlock(), -1);
            setLineNumbers(grpV);
            lV = grpV;
        }
        //reduce-side wumm w/ or without broadcast
        Lop wumm = new WeightedUnaryMMR(grpX, lU, lV, DataType.MATRIX, ValueType.DOUBLE, wtype, uop, cacheU, cacheV, ExecType.MR);
        setOutputDimensions(wumm);
        setLineNumbers(wumm);
        setLops(wumm);
    //in contrast to wsloss no aggregation required 	
    }
}