Examples with DataIdentifier org.apache.sysml.parser.DataIdentifier used on opensource projects

Example 31 with DataIdentifier

use of org.apache.sysml.parser.DataIdentifier in project incubator-systemml by apache.

the class RewriteSplitDagDataDependentOperators method handleReplicatedOperators.

private void handleReplicatedOperators(ArrayList<Hop> rootsSB1, ArrayList<Hop> rootsSB2, VariableSet sb1out, VariableSet sb2in) {
    // step 1: create probe set SB1
    HashSet<Hop> probeSet = new HashSet<>();
    Hop.resetVisitStatus(rootsSB1);
    for (Hop h : rootsSB1) rAddHopsToProbeSet(h, probeSet);
    // step 2: probe SB2 operators top-down (collect cut candidates)
    HashSet<Pair<Hop, Hop>> candSet = new HashSet<>();
    Hop.resetVisitStatus(rootsSB2);
    for (Hop h : rootsSB2) rProbeAndAddHopsToCandidateSet(h, probeSet, candSet);
    // step 3: create additional cuts with reuse for common references
    HashMap<Long, DataOp> reuseTRead = new HashMap<>();
    for (Pair<Hop, Hop> p : candSet) {
        Hop hop = p.getKey();
        Hop c = p.getValue();
        DataOp tread = reuseTRead.get(c.getHopID());
        if (tread == null) {
            String varname = createCutVarName(false);
            tread = new DataOp(varname, c.getDataType(), c.getValueType(), DataOpTypes.TRANSIENTREAD, null, c.getDim1(), c.getDim2(), c.getNnz(), c.getUpdateType(), c.getRowsInBlock(), c.getColsInBlock());
            tread.setVisited();
            HopRewriteUtils.copyLineNumbers(c, tread);
            reuseTRead.put(c.getHopID(), tread);
            DataOp twrite = new DataOp(varname, c.getDataType(), c.getValueType(), c, DataOpTypes.TRANSIENTWRITE, null);
            twrite.setVisited();
            twrite.setOutputParams(c.getDim1(), c.getDim2(), c.getNnz(), c.getUpdateType(), c.getRowsInBlock(), c.getColsInBlock());
            HopRewriteUtils.copyLineNumbers(c, twrite);
            // update live in and out of new statement block (for piggybacking)
            DataIdentifier diVar = new DataIdentifier(varname);
            diVar.setDimensions(c.getDim1(), c.getDim2());
            diVar.setBlockDimensions(c.getRowsInBlock(), c.getColsInBlock());
            diVar.setDataType(c.getDataType());
            diVar.setValueType(c.getValueType());
            sb1out.addVariable(varname, new DataIdentifier(diVar));
            sb2in.addVariable(varname, new DataIdentifier(diVar));
            rootsSB1.add(twrite);
        }
        // create additional cut by rewriting both hop dags
        int pos = HopRewriteUtils.getChildReferencePos(hop, c);
        HopRewriteUtils.removeChildReferenceByPos(hop, c, pos);
        HopRewriteUtils.addChildReference(hop, tread, pos);
    }
}

Example 32 with DataIdentifier

use of org.apache.sysml.parser.DataIdentifier in project incubator-systemml by apache.

the class RewriteSplitDagDataDependentOperators method rewriteStatementBlock.

@Override
public List<StatementBlock> rewriteStatementBlock(StatementBlock sb, ProgramRewriteStatus state) {
    // DAG splits not required for forced single node
    if (DMLScript.rtplatform == RUNTIME_PLATFORM.SINGLE_NODE || !HopRewriteUtils.isLastLevelStatementBlock(sb))
        return Arrays.asList(sb);
    ArrayList<StatementBlock> ret = new ArrayList<>();
    // collect all unknown csv reads hops
    ArrayList<Hop> cand = new ArrayList<>();
    collectDataDependentOperators(sb.getHops(), cand);
    Hop.resetVisitStatus(sb.getHops());
    // split hop dag on demand
    if (!cand.isEmpty()) {
        // collect child operators of candidates (to prevent rewrite anomalies)
        HashSet<Hop> candChilds = new HashSet<>();
        collectCandidateChildOperators(cand, candChilds);
        try {
            // duplicate sb incl live variable sets
            StatementBlock sb1 = new StatementBlock();
            sb1.setDMLProg(sb.getDMLProg());
            sb1.setParseInfo(sb);
            sb1.setLiveIn(new VariableSet());
            sb1.setLiveOut(new VariableSet());
            // move data-dependent ops incl transient writes to new statement block
            // (and replace original persistent read with transient read)
            ArrayList<Hop> sb1hops = new ArrayList<>();
            for (Hop c : cand) {
                // if there are already transient writes use them and don't introduce artificial variables;
                // unless there are transient reads w/ the same variable name in the current dag which can
                // lead to invalid reordering if variable consumers are not feeding into the candidate op.
                boolean hasTWrites = hasTransientWriteParents(c);
                boolean moveTWrite = hasTWrites ? HopRewriteUtils.rHasSimpleReadChain(c, getFirstTransientWriteParent(c).getName()) : false;
                String varname = null;
                long rlen = c.getDim1();
                long clen = c.getDim2();
                long nnz = c.getNnz();
                UpdateType update = c.getUpdateType();
                int brlen = c.getRowsInBlock();
                int bclen = c.getColsInBlock();
                if (// reuse existing transient_write
                hasTWrites && moveTWrite) {
                    Hop twrite = getFirstTransientWriteParent(c);
                    varname = twrite.getName();
                    // create new transient read
                    DataOp tread = new DataOp(varname, c.getDataType(), c.getValueType(), DataOpTypes.TRANSIENTREAD, null, rlen, clen, nnz, update, brlen, bclen);
                    tread.setVisited();
                    HopRewriteUtils.copyLineNumbers(c, tread);
                    // replace data-dependent operator with transient read
                    ArrayList<Hop> parents = new ArrayList<>(c.getParent());
                    for (int i = 0; i < parents.size(); i++) {
                        // prevent concurrent modification by index access
                        Hop parent = parents.get(i);
                        if (!candChilds.contains(parent)) {
                            // anomaly filter
                            if (parent != twrite)
                                HopRewriteUtils.replaceChildReference(parent, c, tread);
                            else
                                sb.getHops().remove(parent);
                        }
                    }
                    // add data-dependent operator sub dag to first statement block
                    sb1hops.add(twrite);
                } else // create transient write to artificial variables
                {
                    varname = createCutVarName(false);
                    // create new transient read
                    DataOp tread = new DataOp(varname, c.getDataType(), c.getValueType(), DataOpTypes.TRANSIENTREAD, null, rlen, clen, nnz, update, brlen, bclen);
                    tread.setVisited();
                    HopRewriteUtils.copyLineNumbers(c, tread);
                    // replace data-dependent operator with transient read
                    ArrayList<Hop> parents = new ArrayList<>(c.getParent());
                    for (int i = 0; i < parents.size(); i++) {
                        // prevent concurrent modification by index access
                        Hop parent = parents.get(i);
                        if (// anomaly filter
                        !candChilds.contains(parent))
                            HopRewriteUtils.replaceChildReference(parent, c, tread);
                    }
                    // add data-dependent operator sub dag to first statement block
                    DataOp twrite = new DataOp(varname, c.getDataType(), c.getValueType(), c, DataOpTypes.TRANSIENTWRITE, null);
                    twrite.setVisited();
                    twrite.setOutputParams(rlen, clen, nnz, update, brlen, bclen);
                    HopRewriteUtils.copyLineNumbers(c, twrite);
                    sb1hops.add(twrite);
                }
                // update live in and out of new statement block (for piggybacking)
                DataIdentifier diVar = new DataIdentifier(varname);
                diVar.setDimensions(rlen, clen);
                diVar.setBlockDimensions(brlen, bclen);
                diVar.setDataType(c.getDataType());
                diVar.setValueType(c.getValueType());
                sb1.liveOut().addVariable(varname, new DataIdentifier(diVar));
                sb.liveIn().addVariable(varname, new DataIdentifier(diVar));
            }
            // ensure disjoint operators across DAGs (prevent replicated operations)
            handleReplicatedOperators(sb1hops, sb.getHops(), sb1.liveOut(), sb.liveIn());
            // deep copy new dag (in order to prevent any dangling references)
            sb1.setHops(Recompiler.deepCopyHopsDag(sb1hops));
            sb1.updateRecompilationFlag();
            // avoid later merge by other rewrites
            sb1.setSplitDag(true);
            // recursive application of rewrite rule (in case of multiple data dependent operators
            // with data dependencies in between each other)
            List<StatementBlock> tmp = rewriteStatementBlock(sb1, state);
            // add new statement blocks to output
            // statement block with data dependent hops
            ret.addAll(tmp);
            // statement block with remaining hops
            ret.add(sb);
            // avoid later merge by other rewrites
            sb.setSplitDag(true);
        } catch (Exception ex) {
            throw new HopsException("Failed to split hops dag for data dependent operators with unknown size.", ex);
        }
        LOG.debug("Applied splitDagDataDependentOperators (lines " + sb.getBeginLine() + "-" + sb.getEndLine() + ").");
    } else // keep original hop dag
    {
        ret.add(sb);
    }
    return ret;
}

Example 33 with DataIdentifier

use of org.apache.sysml.parser.DataIdentifier in project incubator-systemml by apache.

the class InterProceduralAnalysis method extractFunctionCallReturnStatistics.

/**
 * Extract return variable statistics from this function into the
 * calling program.
 *
 * @param fstmt  The function statement.
 * @param fop  The function op.
 * @param tmpVars  Function's map of variables eligible for
 *                    extraction.
 * @param callVars  Calling program's map of variables.
 * @param overwrite  Whether or not to overwrite variables in the
 *                      calling program's variable map.
 */
private static void extractFunctionCallReturnStatistics(FunctionStatement fstmt, FunctionOp fop, LocalVariableMap tmpVars, LocalVariableMap callVars, boolean overwrite) {
    ArrayList<DataIdentifier> foutputOps = fstmt.getOutputParams();
    String[] outputVars = fop.getOutputVariableNames();
    String fkey = fop.getFunctionKey();
    try {
        for (int i = 0; i < foutputOps.size(); i++) {
            DataIdentifier di = foutputOps.get(i);
            // name in function signature
            String fvarname = di.getName();
            // name in calling program
            String pvarname = outputVars[i];
            // output, remove that variable from the calling program's variable map.
            if (callVars.keySet().contains(pvarname)) {
                DataType fdataType = di.getDataType();
                DataType pdataType = callVars.get(pvarname).getDataType();
                if (fdataType != pdataType) {
                    // datatype has changed, and the calling program is reassigning the
                    // the variable, so remove it from the calling variable map
                    callVars.remove(pvarname);
                }
            }
            // Update or add to the calling program's variable map.
            if (di.getDataType() == DataType.MATRIX && tmpVars.keySet().contains(fvarname)) {
                MatrixObject moIn = (MatrixObject) tmpVars.get(fvarname);
                if (// not existing so far
                !callVars.keySet().contains(pvarname) || overwrite) {
                    MatrixObject moOut = createOutputMatrix(moIn.getNumRows(), moIn.getNumColumns(), moIn.getNnz());
                    callVars.put(pvarname, moOut);
                } else // already existing: take largest
                {
                    Data dat = callVars.get(pvarname);
                    if (dat instanceof MatrixObject) {
                        MatrixObject moOut = (MatrixObject) dat;
                        MatrixCharacteristics mc = moOut.getMatrixCharacteristics();
                        if (OptimizerUtils.estimateSizeExactSparsity(mc.getRows(), mc.getCols(), (mc.getNonZeros() > 0) ? OptimizerUtils.getSparsity(mc) : 1.0) < OptimizerUtils.estimateSize(moIn.getNumRows(), moIn.getNumColumns())) {
                            // update statistics if necessary
                            mc.setDimension(moIn.getNumRows(), moIn.getNumColumns());
                            mc.setNonZeros(moIn.getNnz());
                        }
                    }
                }
            }
        }
    } catch (Exception ex) {
        throw new HopsException("Failed to extract output statistics of function " + fkey + ".", ex);
    }
}

Example 34 with DataIdentifier

use of org.apache.sysml.parser.DataIdentifier in project incubator-systemml by apache.

the class InterProceduralAnalysis method populateLocalVariableMapForFunctionCall.

private static void populateLocalVariableMapForFunctionCall(FunctionStatement fstmt, FunctionOp fop, LocalVariableMap callvars, LocalVariableMap vars, FunctionCallSizeInfo fcallSizes) {
    ArrayList<DataIdentifier> inputVars = fstmt.getInputParams();
    ArrayList<Hop> inputOps = fop.getInput();
    String fkey = fop.getFunctionKey();
    for (int i = 0; i < inputVars.size(); i++) {
        // create mapping between input hops and vars
        DataIdentifier dat = inputVars.get(i);
        Hop input = inputOps.get(i);
        if (input.getDataType() == DataType.MATRIX) {
            // propagate matrix characteristics
            MatrixObject mo = new MatrixObject(ValueType.DOUBLE, null);
            MatrixCharacteristics mc = new MatrixCharacteristics(input.getDim1(), input.getDim2(), ConfigurationManager.getBlocksize(), ConfigurationManager.getBlocksize(), fcallSizes.isSafeNnz(fkey, i) ? input.getNnz() : -1);
            MetaDataFormat meta = new MetaDataFormat(mc, null, null);
            mo.setMetaData(meta);
            vars.put(dat.getName(), mo);
        } else if (input.getDataType() == DataType.SCALAR) {
            // (for multiple calls, literal equivalence already checked)
            if (input instanceof LiteralOp) {
                vars.put(dat.getName(), ScalarObjectFactory.createScalarObject(input.getValueType(), (LiteralOp) input));
            } else // and input scalar is existing variable in symbol table
            if (PROPAGATE_SCALAR_VARS_INTO_FUN && fcallSizes.getFunctionCallCount(fkey) == 1 && input instanceof DataOp) {
                Data scalar = callvars.get(input.getName());
                if (scalar != null && scalar instanceof ScalarObject) {
                    vars.put(dat.getName(), scalar);
                }
            }
        }
    }
}

Example 35 with DataIdentifier

use of org.apache.sysml.parser.DataIdentifier in project incubator-systemml by apache.

the class HopRewriteUtils method createDataGenOp.

public static Hop createDataGenOp(Hop input, double value) {
    Hop rows = input.rowsKnown() ? new LiteralOp(input.getDim1()) : new UnaryOp("tmprows", DataType.SCALAR, ValueType.INT, OpOp1.NROW, input);
    Hop cols = input.colsKnown() ? new LiteralOp(input.getDim2()) : new UnaryOp("tmpcols", DataType.SCALAR, ValueType.INT, OpOp1.NCOL, input);
    Hop val = new LiteralOp(value);
    HashMap<String, Hop> params = new HashMap<>();
    params.put(DataExpression.RAND_ROWS, rows);
    params.put(DataExpression.RAND_COLS, cols);
    params.put(DataExpression.RAND_MIN, val);
    params.put(DataExpression.RAND_MAX, val);
    params.put(DataExpression.RAND_PDF, new LiteralOp(DataExpression.RAND_PDF_UNIFORM));
    params.put(DataExpression.RAND_LAMBDA, new LiteralOp(-1.0));
    params.put(DataExpression.RAND_SPARSITY, new LiteralOp(1.0));
    params.put(DataExpression.RAND_SEED, new LiteralOp(DataGenOp.UNSPECIFIED_SEED));
    // note internal refresh size information
    Hop datagen = new DataGenOp(DataGenMethod.RAND, new DataIdentifier("tmp"), params);
    datagen.setOutputBlocksizes(input.getRowsInBlock(), input.getColsInBlock());
    copyLineNumbers(input, datagen);
    if (value == 0)
        datagen.setNnz(0);
    return datagen;
}