Examples with GroupedAggregateInstruction org.apache.sysml.runtime.instructions.mr.GroupedAggregateInstruction used on opensource projects

Example 1 with GroupedAggregateInstruction

use of org.apache.sysml.runtime.instructions.mr.GroupedAggregateInstruction in project incubator-systemml by apache.

the class CostEstimatorStaticRuntime method extractMRInstStatistics.

private Object[] extractMRInstStatistics(String inst, VarStats[] stats) {
    // stats, attrs
    Object[] ret = new Object[2];
    VarStats[] vs = new VarStats[3];
    String[] attr = null;
    String[] parts = InstructionUtils.getInstructionParts(inst);
    String opcode = parts[0];
    if (opcode.equals(DataGen.RAND_OPCODE)) {
        vs[0] = _unknownStats;
        vs[1] = _unknownStats;
        vs[2] = stats[Integer.parseInt(parts[2])];
        int type = 2;
        // awareness of instruction patching min/max
        if (!parts[7].contains(Lop.VARIABLE_NAME_PLACEHOLDER) && !parts[8].contains(Lop.VARIABLE_NAME_PLACEHOLDER)) {
            double minValue = Double.parseDouble(parts[7]);
            double maxValue = Double.parseDouble(parts[8]);
            double sparsity = Double.parseDouble(parts[9]);
            if (minValue == 0.0 && maxValue == 0.0)
                type = 0;
            else if (sparsity == 1.0 && minValue == maxValue)
                type = 1;
        }
        attr = new String[] { String.valueOf(type) };
    }
    if (opcode.equals(DataGen.SEQ_OPCODE)) {
        vs[0] = _unknownStats;
        vs[1] = _unknownStats;
        vs[2] = stats[Integer.parseInt(parts[2])];
    } else // general case
    {
        String inst2 = replaceInstructionPatch(inst);
        MRInstruction mrinst = MRInstructionParser.parseSingleInstruction(inst2);
        if (mrinst instanceof UnaryMRInstructionBase) {
            UnaryMRInstructionBase uinst = (UnaryMRInstructionBase) mrinst;
            vs[0] = uinst.input >= 0 ? stats[uinst.input] : _unknownStats;
            vs[1] = _unknownStats;
            vs[2] = stats[uinst.output];
            if (// scalar input, e.g., print
            vs[0] == null)
                vs[0] = _scalarStats;
            if (// scalar output
            vs[2] == null)
                vs[2] = _scalarStats;
            if (mrinst instanceof MMTSJMRInstruction) {
                String type = ((MMTSJMRInstruction) mrinst).getMMTSJType().toString();
                attr = new String[] { type };
            } else if (mrinst instanceof CM_N_COVInstruction) {
                if (opcode.equals("cm"))
                    attr = new String[] { parts[parts.length - 2] };
            } else if (mrinst instanceof GroupedAggregateInstruction) {
                if (opcode.equals("groupedagg")) {
                    AggregateOperationTypes type = CMOperator.getAggOpType(parts[2], parts[3]);
                    attr = new String[] { String.valueOf(type.ordinal()) };
                }
            }
        } else if (mrinst instanceof BinaryMRInstructionBase) {
            BinaryMRInstructionBase binst = (BinaryMRInstructionBase) mrinst;
            vs[0] = stats[binst.input1];
            vs[1] = stats[binst.input2];
            vs[2] = stats[binst.output];
            if (// scalar input,
            vs[0] == null)
                vs[0] = _scalarStats;
            if (// scalar input,
            vs[1] == null)
                vs[1] = _scalarStats;
            if (// scalar output
            vs[2] == null)
                vs[2] = _scalarStats;
            if (opcode.equals("rmempty")) {
                RemoveEmptyMRInstruction rbinst = (RemoveEmptyMRInstruction) mrinst;
                attr = new String[] { rbinst.isRemoveRows() ? "0" : "1" };
            }
        } else if (mrinst instanceof TernaryInstruction) {
            TernaryInstruction tinst = (TernaryInstruction) mrinst;
            byte[] ix = tinst.getAllIndexes();
            for (int i = 0; i < ix.length - 1; i++) vs[0] = stats[ix[i]];
            vs[2] = stats[ix[ix.length - 1]];
            if (// scalar input,
            vs[0] == null)
                vs[0] = _scalarStats;
            if (// scalar input,
            vs[1] == null)
                vs[1] = _scalarStats;
            if (// scalar output
            vs[2] == null)
                vs[2] = _scalarStats;
        } else if (mrinst instanceof CtableInstruction) {
            CtableInstruction tinst = (CtableInstruction) mrinst;
            vs[0] = stats[tinst.input1];
            vs[1] = stats[tinst.input2];
            vs[2] = stats[tinst.input3];
            if (// scalar input,
            vs[0] == null)
                vs[0] = _scalarStats;
            if (// scalar input,
            vs[1] == null)
                vs[1] = _scalarStats;
            if (// scalar input
            vs[2] == null)
                vs[2] = _scalarStats;
        } else if (mrinst instanceof PickByCountInstruction) {
            PickByCountInstruction pinst = (PickByCountInstruction) mrinst;
            vs[0] = stats[pinst.input1];
            vs[2] = stats[pinst.output];
            if (// scalar input,
            vs[0] == null)
                vs[0] = _scalarStats;
            if (// scalar input,
            vs[1] == null)
                vs[1] = _scalarStats;
            if (// scalar input
            vs[2] == null)
                vs[2] = _scalarStats;
        } else if (mrinst instanceof MapMultChainInstruction) {
            MapMultChainInstruction minst = (MapMultChainInstruction) mrinst;
            vs[0] = stats[minst.getInput1()];
            vs[1] = stats[minst.getInput2()];
            if (minst.getInput3() >= 0)
                vs[2] = stats[minst.getInput3()];
            if (// scalar input,
            vs[0] == null)
                vs[0] = _scalarStats;
            if (// scalar input,
            vs[1] == null)
                vs[1] = _scalarStats;
            if (// scalar input
            vs[2] == null)
                vs[2] = _scalarStats;
        }
    }
    // maintain var status (CP output always inmem)
    vs[2]._inmem = true;
    ret[0] = vs;
    ret[1] = attr;
    return ret;
}

Example 2 with GroupedAggregateInstruction

use of org.apache.sysml.runtime.instructions.mr.GroupedAggregateInstruction in project incubator-systemml by apache.

the class MatrixCharacteristics method computeDimension.

public static void computeDimension(HashMap<Byte, MatrixCharacteristics> dims, MRInstruction ins) {
    MatrixCharacteristics dimOut = dims.get(ins.output);
    if (dimOut == null) {
        dimOut = new MatrixCharacteristics();
        dims.put(ins.output, dimOut);
    }
    if (ins instanceof ReorgInstruction) {
        ReorgInstruction realIns = (ReorgInstruction) ins;
        reorg(dims.get(realIns.input), (ReorgOperator) realIns.getOperator(), dimOut);
    } else if (ins instanceof AppendInstruction) {
        AppendInstruction realIns = (AppendInstruction) ins;
        MatrixCharacteristics in_dim1 = dims.get(realIns.input1);
        MatrixCharacteristics in_dim2 = dims.get(realIns.input2);
        if (realIns.isCBind())
            dimOut.set(in_dim1.numRows, in_dim1.numColumns + in_dim2.numColumns, in_dim1.numRowsPerBlock, in_dim2.numColumnsPerBlock);
        else
            dimOut.set(in_dim1.numRows + in_dim2.numRows, in_dim1.numColumns, in_dim1.numRowsPerBlock, in_dim2.numColumnsPerBlock);
    } else if (ins instanceof CumulativeAggregateInstruction) {
        AggregateUnaryInstruction realIns = (AggregateUnaryInstruction) ins;
        MatrixCharacteristics in = dims.get(realIns.input);
        dimOut.set((long) Math.ceil((double) in.getRows() / in.getRowsPerBlock()), in.getCols(), in.getRowsPerBlock(), in.getColsPerBlock());
    } else if (ins instanceof AggregateUnaryInstruction) {
        AggregateUnaryInstruction realIns = (AggregateUnaryInstruction) ins;
        aggregateUnary(dims.get(realIns.input), (AggregateUnaryOperator) realIns.getOperator(), dimOut);
    } else if (ins instanceof AggregateBinaryInstruction) {
        AggregateBinaryInstruction realIns = (AggregateBinaryInstruction) ins;
        aggregateBinary(dims.get(realIns.input1), dims.get(realIns.input2), (AggregateBinaryOperator) realIns.getOperator(), dimOut);
    } else if (ins instanceof MapMultChainInstruction) {
        // output size independent of chain type
        MapMultChainInstruction realIns = (MapMultChainInstruction) ins;
        MatrixCharacteristics mc1 = dims.get(realIns.getInput1());
        MatrixCharacteristics mc2 = dims.get(realIns.getInput2());
        dimOut.set(mc1.numColumns, mc2.numColumns, mc1.numRowsPerBlock, mc1.numColumnsPerBlock);
    } else if (ins instanceof QuaternaryInstruction) {
        QuaternaryInstruction realIns = (QuaternaryInstruction) ins;
        MatrixCharacteristics mc1 = dims.get(realIns.getInput1());
        MatrixCharacteristics mc2 = dims.get(realIns.getInput2());
        MatrixCharacteristics mc3 = dims.get(realIns.getInput3());
        realIns.computeMatrixCharacteristics(mc1, mc2, mc3, dimOut);
    } else if (ins instanceof ReblockInstruction) {
        ReblockInstruction realIns = (ReblockInstruction) ins;
        MatrixCharacteristics in_dim = dims.get(realIns.input);
        dimOut.set(in_dim.numRows, in_dim.numColumns, realIns.brlen, realIns.bclen, in_dim.nonZero);
    } else if (ins instanceof MatrixReshapeMRInstruction) {
        MatrixReshapeMRInstruction mrinst = (MatrixReshapeMRInstruction) ins;
        MatrixCharacteristics in_dim = dims.get(mrinst.input);
        dimOut.set(mrinst.getNumRows(), mrinst.getNumColunms(), in_dim.getRowsPerBlock(), in_dim.getColsPerBlock(), in_dim.getNonZeros());
    } else if (ins instanceof RandInstruction || ins instanceof SeqInstruction) {
        DataGenMRInstruction dataIns = (DataGenMRInstruction) ins;
        dimOut.set(dims.get(dataIns.getInput()));
    } else if (ins instanceof ReplicateInstruction) {
        ReplicateInstruction realIns = (ReplicateInstruction) ins;
        realIns.computeOutputDimension(dims.get(realIns.input), dimOut);
    } else if (// before unary
    ins instanceof ParameterizedBuiltinMRInstruction) {
        ParameterizedBuiltinMRInstruction realIns = (ParameterizedBuiltinMRInstruction) ins;
        realIns.computeOutputCharacteristics(dims.get(realIns.input), dimOut);
    } else if (ins instanceof ScalarInstruction || ins instanceof AggregateInstruction || (ins instanceof UnaryInstruction && !(ins instanceof MMTSJMRInstruction)) || ins instanceof ZeroOutInstruction) {
        UnaryMRInstructionBase realIns = (UnaryMRInstructionBase) ins;
        dimOut.set(dims.get(realIns.input));
    } else if (ins instanceof MMTSJMRInstruction) {
        MMTSJMRInstruction mmtsj = (MMTSJMRInstruction) ins;
        MMTSJType tstype = mmtsj.getMMTSJType();
        MatrixCharacteristics mc = dims.get(mmtsj.input);
        dimOut.set(tstype.isLeft() ? mc.numColumns : mc.numRows, tstype.isLeft() ? mc.numColumns : mc.numRows, mc.numRowsPerBlock, mc.numColumnsPerBlock);
    } else if (ins instanceof PMMJMRInstruction) {
        PMMJMRInstruction pmmins = (PMMJMRInstruction) ins;
        MatrixCharacteristics mc = dims.get(pmmins.input2);
        dimOut.set(pmmins.getNumRows(), mc.numColumns, mc.numRowsPerBlock, mc.numColumnsPerBlock);
    } else if (ins instanceof RemoveEmptyMRInstruction) {
        RemoveEmptyMRInstruction realIns = (RemoveEmptyMRInstruction) ins;
        MatrixCharacteristics mc = dims.get(realIns.input1);
        long min = realIns.isEmptyReturn() ? 1 : 0;
        if (realIns.isRemoveRows())
            dimOut.set(Math.max(realIns.getOutputLen(), min), mc.getCols(), mc.numRowsPerBlock, mc.numColumnsPerBlock);
        else
            dimOut.set(mc.getRows(), Math.max(realIns.getOutputLen(), min), mc.numRowsPerBlock, mc.numColumnsPerBlock);
    } else if (// needs to be checked before binary
    ins instanceof UaggOuterChainInstruction) {
        UaggOuterChainInstruction realIns = (UaggOuterChainInstruction) ins;
        MatrixCharacteristics mc1 = dims.get(realIns.input1);
        MatrixCharacteristics mc2 = dims.get(realIns.input2);
        realIns.computeOutputCharacteristics(mc1, mc2, dimOut);
    } else if (ins instanceof GroupedAggregateMInstruction) {
        GroupedAggregateMInstruction realIns = (GroupedAggregateMInstruction) ins;
        MatrixCharacteristics mc1 = dims.get(realIns.input1);
        realIns.computeOutputCharacteristics(mc1, dimOut);
    } else if (ins instanceof BinaryInstruction || ins instanceof BinaryMInstruction || ins instanceof CombineBinaryInstruction) {
        BinaryMRInstructionBase realIns = (BinaryMRInstructionBase) ins;
        MatrixCharacteristics mc1 = dims.get(realIns.input1);
        MatrixCharacteristics mc2 = dims.get(realIns.input2);
        if (mc1.getRows() > 1 && mc1.getCols() == 1 && mc2.getRows() == 1 && // outer
        mc2.getCols() > 1) {
            dimOut.set(mc1.getRows(), mc2.getCols(), mc1.getRowsPerBlock(), mc2.getColsPerBlock());
        } else {
            // default case
            dimOut.set(mc1);
        }
    } else if (ins instanceof TernaryInstruction) {
        dimOut.set(dims.get(ins.getInputIndexes()[0]));
    } else if (ins instanceof CombineTernaryInstruction) {
        CtableInstruction realIns = (CtableInstruction) ins;
        dimOut.set(dims.get(realIns.input1));
    } else if (ins instanceof CombineUnaryInstruction) {
        dimOut.set(dims.get(((CombineUnaryInstruction) ins).input));
    } else if (ins instanceof CM_N_COVInstruction || ins instanceof GroupedAggregateInstruction) {
        dimOut.set(1, 1, 1, 1);
    } else if (ins instanceof RangeBasedReIndexInstruction) {
        RangeBasedReIndexInstruction realIns = (RangeBasedReIndexInstruction) ins;
        MatrixCharacteristics dimIn = dims.get(realIns.input);
        realIns.computeOutputCharacteristics(dimIn, dimOut);
    } else if (ins instanceof CtableInstruction) {
        CtableInstruction realIns = (CtableInstruction) ins;
        MatrixCharacteristics in_dim = dims.get(realIns.input1);
        dimOut.set(realIns.getOutputDim1(), realIns.getOutputDim2(), in_dim.numRowsPerBlock, in_dim.numColumnsPerBlock);
    } else {
        /*
			 * if ins is none of the above cases then we assume that dim_out dimensions are unknown
			 */
        dimOut.numRows = -1;
        dimOut.numColumns = -1;
        dimOut.numRowsPerBlock = 1;
        dimOut.numColumnsPerBlock = 1;
    }
}

Example 3 with GroupedAggregateInstruction

use of org.apache.sysml.runtime.instructions.mr.GroupedAggregateInstruction in project incubator-systemml by apache.

the class GroupedAggMRCombiner method reduce.

@Override
public void reduce(TaggedMatrixIndexes key, Iterator<WeightedCell> values, OutputCollector<TaggedMatrixIndexes, WeightedCell> out, Reporter reporter) throws IOException {
    long start = System.currentTimeMillis();
    // get aggregate operator
    GroupedAggregateInstruction ins = grpaggInstructions.get(key.getTag());
    Operator op = ins.getOperator();
    boolean isPartialAgg = true;
    // combine iterator to single value
    try {
        if (// everything except sum
        op instanceof CMOperator) {
            if (((CMOperator) op).isPartialAggregateOperator()) {
                cmObj.reset();
                CM lcmFn = cmFn.get(key.getTag());
                // partial aggregate cm operator
                while (values.hasNext()) {
                    WeightedCell value = values.next();
                    lcmFn.execute(cmObj, value.getValue(), value.getWeight());
                }
                outCell.setValue(cmObj.getRequiredPartialResult(op));
                outCell.setWeight(cmObj.getWeight());
            } else // forward tuples to reducer
            {
                isPartialAgg = false;
                while (values.hasNext()) out.collect(key, values.next());
            }
        } else if (// sum
        op instanceof AggregateOperator) {
            AggregateOperator aggop = (AggregateOperator) op;
            if (aggop.correctionExists) {
                KahanObject buffer = new KahanObject(aggop.initialValue, 0);
                KahanPlus.getKahanPlusFnObject();
                // partial aggregate with correction
                while (values.hasNext()) {
                    WeightedCell value = values.next();
                    aggop.increOp.fn.execute(buffer, value.getValue() * value.getWeight());
                }
                outCell.setValue(buffer._sum);
                outCell.setWeight(1);
            } else // no correction
            {
                double v = aggop.initialValue;
                // partial aggregate without correction
                while (values.hasNext()) {
                    WeightedCell value = values.next();
                    v = aggop.increOp.fn.execute(v, value.getValue() * value.getWeight());
                }
                outCell.setValue(v);
                outCell.setWeight(1);
            }
        } else
            throw new IOException("Unsupported operator in instruction: " + ins);
    } catch (Exception ex) {
        throw new IOException(ex);
    }
    // collect the output (to reducer)
    if (isPartialAgg)
        out.collect(key, outCell);
    reporter.incrCounter(Counters.COMBINE_OR_REDUCE_TIME, System.currentTimeMillis() - start);
}

Example 4 with GroupedAggregateInstruction

use of org.apache.sysml.runtime.instructions.mr.GroupedAggregateInstruction in project incubator-systemml by apache.

the class GroupedAggMRMapper method configure.

@Override
public void configure(JobConf job) {
    super.configure(job);
    try {
        GroupedAggregateInstruction[] grpaggIns = MRJobConfiguration.getGroupedAggregateInstructions(job);
        if (grpaggIns == null)
            throw new RuntimeException("no GroupAggregate Instructions found!");
        ArrayList<GroupedAggregateInstruction> vec = new ArrayList<>();
        for (int i = 0; i < representativeMatrixes.size(); i++) {
            byte input = representativeMatrixes.get(i);
            for (GroupedAggregateInstruction ins : grpaggIns) if (ins.input == input)
                vec.add(ins);
            groupAgg_instructions.add(vec);
        }
    } catch (Exception e) {
        throw new RuntimeException(e);
    }
}

Example 5 with GroupedAggregateInstruction

use of org.apache.sysml.runtime.instructions.mr.GroupedAggregateInstruction in project incubator-systemml by apache.

the class GroupedAggMRReducer method reduce.

@Override
public void reduce(TaggedMatrixIndexes key, Iterator<WeightedCell> values, OutputCollector<MatrixIndexes, MatrixCell> out, Reporter report) throws IOException {
    commonSetup(report);
    // get operator
    GroupedAggregateInstruction ins = grpaggInstructions.get(key.getTag());
    Operator op = ins.getOperator();
    try {
        if (// all, but sum
        op instanceof CMOperator) {
            cmObj.reset();
            CM lcmFn = cmFn.get(key.getTag());
            while (values.hasNext()) {
                WeightedCell value = values.next();
                lcmFn.execute(cmObj, value.getValue(), value.getWeight());
            }
            outCell.setValue(cmObj.getRequiredResult(op));
        } else if (// sum
        op instanceof AggregateOperator) {
            AggregateOperator aggop = (AggregateOperator) op;
            if (aggop.correctionExists) {
                KahanObject buffer = new KahanObject(aggop.initialValue, 0);
                while (values.hasNext()) {
                    WeightedCell value = values.next();
                    aggop.increOp.fn.execute(buffer, value.getValue() * value.getWeight());
                }
                outCell.setValue(buffer._sum);
            } else {
                double v = aggop.initialValue;
                while (values.hasNext()) {
                    WeightedCell value = values.next();
                    v = aggop.increOp.fn.execute(v, value.getValue() * value.getWeight());
                }
                outCell.setValue(v);
            }
        } else
            throw new IOException("Unsupported operator in instruction: " + ins);
    } catch (Exception ex) {
        throw new IOException(ex);
    }
    outIndex.setIndexes(key.getBaseObject());
    cachedValues.reset();
    cachedValues.set(key.getTag(), outIndex, outCell);
    processReducerInstructions();
    // output the final result matrices
    outputResultsFromCachedValues(report);
}