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

Example 6 with AggregateBinaryInstruction

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

the class MMRJMRReducer method reduce.

//	private MatrixIndexes indexBuf=new MatrixIndexes();
@Override
public void reduce(TripleIndexes triple, Iterator<TaggedMatrixValue> values, OutputCollector<MatrixIndexes, MatrixValue> out, Reporter report) throws IOException {
    long start = System.currentTimeMillis();
    //	System.out.println("~~~~~ group: "+triple);
    commonSetup(report);
    //output previous results if needed
    if (prevIndexes.getFirstIndex() != triple.getFirstIndex() || prevIndexes.getSecondIndex() != triple.getSecondIndex()) {
        //	System.out.println("cacheValues before processReducerInstructions: \n"+cachedValues);
        //perform mixed operations
        processReducerInstructions();
        //		System.out.println("cacheValues before output: \n"+cachedValues);
        //output results
        outputResultsFromCachedValues(report);
        cachedValues.reset();
    } else {
        //clear the buffer
        for (AggregateBinaryInstruction aggBinInstruction : aggBinInstructions) {
            //				System.out.println("cacheValues before remore: \n"+cachedValues);
            cachedValues.remove(aggBinInstruction.input1);
            //		System.out.println("cacheValues after remore: "+aggBinInstruction.input1+"\n"+cachedValues);
            cachedValues.remove(aggBinInstruction.input2);
        //		System.out.println("cacheValues after remore: "+aggBinInstruction.input2+"\n"+cachedValues);
        }
    }
    //perform aggregation first
    aggIndexes.setIndexes(triple.getFirstIndex(), triple.getSecondIndex());
    processAggregateInstructions(aggIndexes, values);
    //perform aggbinary for this group
    for (AggregateBinaryInstruction aggBinInstruction : aggBinInstructions) processAggBinaryPerGroup(aggIndexes, aggBinInstruction);
    //	System.out.println("cacheValues after aggbinary: \n"+cachedValues);
    prevIndexes.setIndexes(triple);
    report.incrCounter(Counters.COMBINE_OR_REDUCE_TIME, System.currentTimeMillis() - start);
}

Example 7 with AggregateBinaryInstruction

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

the class MRJobConfiguration method computeMatrixCharacteristics.

/**
	 * NOTE: this method needs to be in-sync with MRBaseForCommonInstructions.processOneInstruction,
	 * otherwise, the latter will potentially fail with missing dimension information.
	 * 
	 * @param job job configuration
	 * @param inputIndexes array of byte indexes
	 * @param dataGenInstructions data gen instructions as a string
	 * @param instructionsInMapper instruction in mapper as a string
	 * @param reblockInstructions reblock instructions as a string
	 * @param aggInstructionsInReducer aggregate instructions in reducer as a string
	 * @param aggBinInstructions binary aggregate instructions as a string
	 * @param otherInstructionsInReducer other instructions in reducer as a string
	 * @param resultIndexes array of byte result indexes
	 * @param mapOutputIndexes set of map output indexes
	 * @param forMMCJ ?
	 * @return reducer groups
	 * @throws DMLRuntimeException if DMLRuntimeException occurs
	 */
public static MatrixChar_N_ReducerGroups computeMatrixCharacteristics(JobConf job, byte[] inputIndexes, String dataGenInstructions, String instructionsInMapper, String reblockInstructions, String aggInstructionsInReducer, String aggBinInstructions, String otherInstructionsInReducer, byte[] resultIndexes, HashSet<Byte> mapOutputIndexes, boolean forMMCJ) throws DMLRuntimeException {
    HashSet<Byte> intermediateMatrixIndexes = new HashSet<Byte>();
    HashMap<Byte, MatrixCharacteristics> dims = new HashMap<Byte, MatrixCharacteristics>();
    for (byte i : inputIndexes) {
        MatrixCharacteristics dim = new MatrixCharacteristics(getNumRows(job, i), getNumColumns(job, i), getNumRowsPerBlock(job, i), getNumColumnsPerBlock(job, i), getNumNonZero(job, i));
        dims.put(i, dim);
    }
    DataGenMRInstruction[] dataGenIns = null;
    dataGenIns = MRInstructionParser.parseDataGenInstructions(dataGenInstructions);
    if (dataGenIns != null) {
        for (DataGenMRInstruction ins : dataGenIns) {
            MatrixCharacteristics.computeDimension(dims, ins);
        }
    }
    MRInstruction[] insMapper = MRInstructionParser.parseMixedInstructions(instructionsInMapper);
    if (insMapper != null) {
        for (MRInstruction ins : insMapper) {
            MatrixCharacteristics.computeDimension(dims, ins);
            if (ins instanceof UnaryMRInstructionBase) {
                UnaryMRInstructionBase tempIns = (UnaryMRInstructionBase) ins;
                setIntermediateMatrixCharactristics(job, tempIns.input, dims.get(tempIns.input));
                intermediateMatrixIndexes.add(tempIns.input);
            } else if (ins instanceof AppendMInstruction) {
                AppendMInstruction tempIns = (AppendMInstruction) ins;
                setIntermediateMatrixCharactristics(job, tempIns.input1, dims.get(tempIns.input1));
                intermediateMatrixIndexes.add(tempIns.input1);
            } else if (ins instanceof AppendGInstruction) {
                AppendGInstruction tempIns = (AppendGInstruction) ins;
                setIntermediateMatrixCharactristics(job, tempIns.input1, dims.get(tempIns.input1));
                intermediateMatrixIndexes.add(tempIns.input1);
            } else if (ins instanceof BinaryMInstruction) {
                BinaryMInstruction tempIns = (BinaryMInstruction) ins;
                setIntermediateMatrixCharactristics(job, tempIns.input1, dims.get(tempIns.input1));
                intermediateMatrixIndexes.add(tempIns.input1);
            } else if (ins instanceof AggregateBinaryInstruction) {
                AggregateBinaryInstruction tempIns = (AggregateBinaryInstruction) ins;
                setIntermediateMatrixCharactristics(job, tempIns.input1, dims.get(tempIns.input1));
                //TODO
                intermediateMatrixIndexes.add(tempIns.input1);
            } else if (ins instanceof MapMultChainInstruction) {
                MapMultChainInstruction tempIns = (MapMultChainInstruction) ins;
                setIntermediateMatrixCharactristics(job, tempIns.getInput1(), dims.get(tempIns.getInput2()));
                intermediateMatrixIndexes.add(tempIns.getInput1());
            } else if (ins instanceof PMMJMRInstruction) {
                PMMJMRInstruction tempIns = (PMMJMRInstruction) ins;
                setIntermediateMatrixCharactristics(job, tempIns.input2, dims.get(tempIns.input2));
                intermediateMatrixIndexes.add(tempIns.input2);
            }
        }
    }
    ReblockInstruction[] reblockIns = MRInstructionParser.parseReblockInstructions(reblockInstructions);
    if (reblockIns != null) {
        for (ReblockInstruction ins : reblockIns) {
            MatrixCharacteristics.computeDimension(dims, ins);
            setMatrixCharactristicsForReblock(job, ins.output, dims.get(ins.output));
        }
    }
    Instruction[] aggIns = MRInstructionParser.parseAggregateInstructions(aggInstructionsInReducer);
    if (aggIns != null) {
        for (Instruction ins : aggIns) {
            MatrixCharacteristics.computeDimension(dims, (MRInstruction) ins);
            // if instruction's output is not in resultIndexes, then add its dimensions to jobconf
            MRInstruction mrins = (MRInstruction) ins;
            boolean found = false;
            for (byte b : resultIndexes) {
                if (b == mrins.output) {
                    found = true;
                    break;
                }
            }
            if (!found) {
                setIntermediateMatrixCharactristics(job, mrins.output, dims.get(mrins.output));
                intermediateMatrixIndexes.add(mrins.output);
            }
        }
    }
    long numReduceGroups = 0;
    AggregateBinaryInstruction[] aggBinIns = getAggregateBinaryInstructions(job);
    if (aggBinIns != null) {
        for (AggregateBinaryInstruction ins : aggBinIns) {
            MatrixCharacteristics dim1 = dims.get(ins.input1);
            MatrixCharacteristics dim2 = dims.get(ins.input2);
            setMatrixCharactristicsForBinAgg(job, ins.input1, dim1);
            setMatrixCharactristicsForBinAgg(job, ins.input2, dim2);
            MatrixCharacteristics.computeDimension(dims, ins);
            if (//there will be only one aggbin operation for MMCJ
            forMMCJ)
                numReduceGroups = (long) Math.ceil((double) dim1.getCols() / (double) dim1.getColsPerBlock());
        }
    }
    if (!forMMCJ) {
        //store the skylines
        ArrayList<Long> xs = new ArrayList<Long>(mapOutputIndexes.size());
        ArrayList<Long> ys = new ArrayList<Long>(mapOutputIndexes.size());
        for (byte idx : mapOutputIndexes) {
            MatrixCharacteristics dim = dims.get(idx);
            long x = (long) Math.ceil((double) dim.getRows() / (double) dim.getRowsPerBlock());
            long y = (long) Math.ceil((double) dim.getCols() / (double) dim.getColsPerBlock());
            int i = 0;
            boolean toadd = true;
            while (i < xs.size()) {
                if ((x >= xs.get(i) && y > ys.get(i)) || (x > xs.get(i) && y >= ys.get(i))) {
                    //remove any included x's and y's
                    xs.remove(i);
                    ys.remove(i);
                } else if (//if included in others, stop
                x <= xs.get(i) && y <= ys.get(i)) {
                    toadd = false;
                    break;
                } else
                    i++;
            }
            if (toadd) {
                xs.add(x);
                ys.add(y);
            }
        }
        //sort by x
        TreeMap<Long, Long> map = new TreeMap<Long, Long>();
        for (int i = 0; i < xs.size(); i++) map.put(xs.get(i), ys.get(i));
        numReduceGroups = 0;
        //compute area
        long prev = 0;
        for (Entry<Long, Long> e : map.entrySet()) {
            numReduceGroups += (e.getKey() - prev) * e.getValue();
            prev = e.getKey();
        }
    }
    MRInstruction[] insReducer = MRInstructionParser.parseMixedInstructions(otherInstructionsInReducer);
    if (insReducer != null) {
        for (MRInstruction ins : insReducer) {
            MatrixCharacteristics.computeDimension(dims, ins);
            if (ins instanceof UnaryMRInstructionBase) {
                UnaryMRInstructionBase tempIns = (UnaryMRInstructionBase) ins;
                setIntermediateMatrixCharactristics(job, tempIns.input, dims.get(tempIns.input));
                intermediateMatrixIndexes.add(tempIns.input);
            } else if (ins instanceof RemoveEmptyMRInstruction) {
                RemoveEmptyMRInstruction tempIns = (RemoveEmptyMRInstruction) ins;
                setIntermediateMatrixCharactristics(job, tempIns.input1, dims.get(tempIns.input1));
                intermediateMatrixIndexes.add(tempIns.input1);
            }
            // if instruction's output is not in resultIndexes, then add its dimensions to jobconf
            boolean found = false;
            for (byte b : resultIndexes) {
                if (b == ins.output) {
                    found = true;
                    break;
                }
            }
            if (!found) {
                setIntermediateMatrixCharactristics(job, ins.output, dims.get(ins.output));
                intermediateMatrixIndexes.add(ins.output);
            }
        }
    }
    setIntermediateMatrixIndexes(job, intermediateMatrixIndexes);
    for (byte tag : mapOutputIndexes) setMatrixCharactristicsForMapperOutput(job, tag, dims.get(tag));
    MatrixCharacteristics[] stats = new MatrixCharacteristics[resultIndexes.length];
    MatrixCharacteristics resultDims;
    for (int i = 0; i < resultIndexes.length; i++) {
        resultDims = dims.get(resultIndexes[i]);
        stats[i] = resultDims;
        setMatrixCharactristicsForOutput(job, resultIndexes[i], stats[i]);
    }
    return new MatrixChar_N_ReducerGroups(stats, numReduceGroups);
}

Example 8 with AggregateBinaryInstruction

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

the class MMCJMRCombinerReducerBase method configure.

//	protected int elementSize=8;
public void configure(JobConf job) {
    super.configure(job);
    AggregateBinaryInstruction[] ins;
    try {
        ins = MRJobConfiguration.getAggregateBinaryInstructions(job);
    } catch (DMLRuntimeException e) {
        throw new RuntimeException(e);
    }
    if (ins.length != 1)
        throw new RuntimeException("MMCJ only perform one aggregate binary instruction");
    aggBinInstruction = ins[0];
    //decide which matrix need to be cached for cross product
    dim1 = MRJobConfiguration.getMatrixCharactristicsForBinAgg(job, aggBinInstruction.input1);
    dim2 = MRJobConfiguration.getMatrixCharactristicsForBinAgg(job, aggBinInstruction.input2);
    if (dim1.getRows() > dim2.getCols()) {
        tagForLeft = 1;
        tagForRight = 0;
    }
    //allocate space for the temporary variable
    try {
        buffer = valueClass.newInstance();
    } catch (Exception e) {
        throw new RuntimeException(e);
    }
//		if(valueClass.equals(MatrixCell.class))
//			elementSize=90;
}

Example 9 with AggregateBinaryInstruction

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

the class MMCJMRMapper method configure.

public void configure(JobConf job) {
    super.configure(job);
    AggregateBinaryInstruction[] ins;
    try {
        ins = MRJobConfiguration.getAggregateBinaryInstructions(job);
    } catch (DMLRuntimeException e) {
        throw new RuntimeException(e);
    }
    if (ins.length != 1)
        throw new RuntimeException("MMCJ only perform one aggregate binary instruction");
    aggBinInstruction = ins[0];
    //decide which matrix need to be cached for cross product
    MatrixCharacteristics dim1 = MRJobConfiguration.getMatrixCharactristicsForBinAgg(job, aggBinInstruction.input1);
    MatrixCharacteristics dim2 = MRJobConfiguration.getMatrixCharactristicsForBinAgg(job, aggBinInstruction.input2);
    if (dim1.getRows() > dim2.getCols()) {
        tagForLeft = 1;
        tagForRight = 0;
    }
}