Examples with Hop org.apache.sysml.hops.Hop used on opensource projects

Example 26 with Hop

use of org.apache.sysml.hops.Hop in project incubator-systemml by apache.

the class ProgramRecompiler method rFindAndRecompileIndexingHOP.

/**
 * NOTE: if force is set, we set and recompile the respective indexing hops;
 * otherwise, we release the forced exec type and recompile again. Hence,
 * any changes can be exactly reverted with the same access behavior.
 *
 * @param sb statement block
 * @param pb program block
 * @param var variable
 * @param ec execution context
 * @param force if true, set and recompile the respective indexing hops
 */
public static void rFindAndRecompileIndexingHOP(StatementBlock sb, ProgramBlock pb, String var, ExecutionContext ec, boolean force) {
    if (pb instanceof IfProgramBlock && sb instanceof IfStatementBlock) {
        IfProgramBlock ipb = (IfProgramBlock) pb;
        IfStatementBlock isb = (IfStatementBlock) sb;
        IfStatement is = (IfStatement) sb.getStatement(0);
        // process if condition
        if (isb.getPredicateHops() != null)
            ipb.setPredicate(rFindAndRecompileIndexingHOP(isb.getPredicateHops(), ipb.getPredicate(), var, ec, force));
        // process if branch
        int len = is.getIfBody().size();
        for (int i = 0; i < ipb.getChildBlocksIfBody().size() && i < len; i++) {
            ProgramBlock lpb = ipb.getChildBlocksIfBody().get(i);
            StatementBlock lsb = is.getIfBody().get(i);
            rFindAndRecompileIndexingHOP(lsb, lpb, var, ec, force);
        }
        // process else branch
        if (ipb.getChildBlocksElseBody() != null) {
            int len2 = is.getElseBody().size();
            for (int i = 0; i < ipb.getChildBlocksElseBody().size() && i < len2; i++) {
                ProgramBlock lpb = ipb.getChildBlocksElseBody().get(i);
                StatementBlock lsb = is.getElseBody().get(i);
                rFindAndRecompileIndexingHOP(lsb, lpb, var, ec, force);
            }
        }
    } else if (pb instanceof WhileProgramBlock && sb instanceof WhileStatementBlock) {
        WhileProgramBlock wpb = (WhileProgramBlock) pb;
        WhileStatementBlock wsb = (WhileStatementBlock) sb;
        WhileStatement ws = (WhileStatement) sb.getStatement(0);
        // process while condition
        if (wsb.getPredicateHops() != null)
            wpb.setPredicate(rFindAndRecompileIndexingHOP(wsb.getPredicateHops(), wpb.getPredicate(), var, ec, force));
        // process body
        // robustness for potentially added problem blocks
        int len = ws.getBody().size();
        for (int i = 0; i < wpb.getChildBlocks().size() && i < len; i++) {
            ProgramBlock lpb = wpb.getChildBlocks().get(i);
            StatementBlock lsb = ws.getBody().get(i);
            rFindAndRecompileIndexingHOP(lsb, lpb, var, ec, force);
        }
    } else if (// for or parfor
    pb instanceof ForProgramBlock && sb instanceof ForStatementBlock) {
        ForProgramBlock fpb = (ForProgramBlock) pb;
        ForStatementBlock fsb = (ForStatementBlock) sb;
        ForStatement fs = (ForStatement) fsb.getStatement(0);
        if (fsb.getFromHops() != null)
            fpb.setFromInstructions(rFindAndRecompileIndexingHOP(fsb.getFromHops(), fpb.getFromInstructions(), var, ec, force));
        if (fsb.getToHops() != null)
            fpb.setToInstructions(rFindAndRecompileIndexingHOP(fsb.getToHops(), fpb.getToInstructions(), var, ec, force));
        if (fsb.getIncrementHops() != null)
            fpb.setIncrementInstructions(rFindAndRecompileIndexingHOP(fsb.getIncrementHops(), fpb.getIncrementInstructions(), var, ec, force));
        // process body
        // robustness for potentially added problem blocks
        int len = fs.getBody().size();
        for (int i = 0; i < fpb.getChildBlocks().size() && i < len; i++) {
            ProgramBlock lpb = fpb.getChildBlocks().get(i);
            StatementBlock lsb = fs.getBody().get(i);
            rFindAndRecompileIndexingHOP(lsb, lpb, var, ec, force);
        }
    } else // last level program block
    {
        try {
            // process actual hops
            boolean ret = false;
            Hop.resetVisitStatus(sb.getHops());
            if (force) {
                // set forced execution type
                for (Hop h : sb.getHops()) ret |= rFindAndSetCPIndexingHOP(h, var);
            } else {
                // release forced execution type
                for (Hop h : sb.getHops()) ret |= rFindAndReleaseIndexingHOP(h, var);
            }
            // recompilation on-demand
            if (ret) {
                // construct new instructions
                ArrayList<Instruction> newInst = Recompiler.recompileHopsDag(sb, sb.getHops(), ec.getVariables(), null, true, false, 0);
                pb.setInstructions(newInst);
            }
        } catch (Exception ex) {
            throw new DMLRuntimeException(ex);
        }
    }
}

Example 27 with Hop

use of org.apache.sysml.hops.Hop in project incubator-systemml by apache.

the class ProgramConverter method createIfStatementBlockCopy.

public static IfStatementBlock createIfStatementBlockCopy(IfStatementBlock sb, long pid, boolean plain, boolean forceDeepCopy) {
    IfStatementBlock ret = null;
    try {
        if (ConfigurationManager.getCompilerConfigFlag(ConfigType.ALLOW_PARALLEL_DYN_RECOMPILATION) && // forced deep copy for function recompile
        sb != null && (Recompiler.requiresRecompilation(sb.getPredicateHops()) || forceDeepCopy)) {
            // create new statement (shallow copy livein/liveout for recompile, line numbers for explain)
            ret = new IfStatementBlock();
            ret.setDMLProg(sb.getDMLProg());
            ret.setParseInfo(sb);
            ret.setLiveIn(sb.liveIn());
            ret.setLiveOut(sb.liveOut());
            ret.setUpdatedVariables(sb.variablesUpdated());
            ret.setReadVariables(sb.variablesRead());
            // shallow copy child statements
            ret.setStatements(sb.getStatements());
            // deep copy predicate hops dag for concurrent recompile
            Hop hops = Recompiler.deepCopyHopsDag(sb.getPredicateHops());
            ret.setPredicateHops(hops);
            ret.updatePredicateRecompilationFlag();
        } else {
            ret = sb;
        }
    } catch (Exception ex) {
        throw new DMLRuntimeException(ex);
    }
    return ret;
}

Example 28 with Hop

use of org.apache.sysml.hops.Hop in project incubator-systemml by apache.

the class PlanAnalyzer method rCollectPartitionInputNodes.

private static void rCollectPartitionInputNodes(Hop current, HashSet<Long> visited, HashSet<Long> partition, HashSet<Long> I) {
    // memoization (not via hops because in middle of dag)
    if (visited.contains(current.getHopID()))
        return;
    // process children recursively
    for (Hop c : current.getInput()) if (partition.contains(c.getHopID()))
        rCollectPartitionInputNodes(c, visited, partition, I);
    else
        I.add(c.getHopID());
    visited.add(current.getHopID());
}

Example 29 with Hop

use of org.apache.sysml.hops.Hop in project incubator-systemml by apache.

the class PlanSelectionFuseCostBased method isValidMultiAggregate.

private static boolean isValidMultiAggregate(CPlanMemoTable memo, MemoTableEntry me) {
    // ensure input consistent sizes (otherwise potential for incorrect results)
    boolean ret = true;
    Hop refSize = memo.getHopRefs().get(me.input1).getInput().get(0);
    for (int i = 1; ret && i < 3; i++) {
        if (me.isPlanRef(i))
            ret &= HopRewriteUtils.isEqualSize(refSize, memo.getHopRefs().get(me.input(i)).getInput().get(0));
    }
    // they to not have potentially transitive parent child references
    for (int i = 0; ret && i < 3; i++) if (me.isPlanRef(i)) {
        HashSet<Long> probe = new HashSet<>();
        for (int j = 0; j < 3; j++) if (i != j)
            probe.add(me.input(j));
        ret &= rCheckMultiAggregate(memo.getHopRefs().get(me.input(i)), probe);
    }
    return ret;
}

Example 30 with Hop

use of org.apache.sysml.hops.Hop in project incubator-systemml by apache.

the class PlanSelectionFuseCostBased method rGetPlanCosts.

private static double rGetPlanCosts(CPlanMemoTable memo, Hop current, HashSet<Pair<Long, Long>> visited, HashSet<Long> partition, ArrayList<Long> M, boolean[] plan, HashMap<Long, Double> computeCosts, CostVector costsCurrent, TemplateType currentType) {
    // memoization per hop id and cost vector to account for redundant
    // computation without double counting materialized results or compute
    // costs of complex operation DAGs within a single fused operator
    Pair<Long, Long> tag = Pair.of(current.getHopID(), (costsCurrent == null) ? 0 : costsCurrent.ID);
    if (visited.contains(tag))
        return 0;
    visited.add(tag);
    // open template if necessary, including memoization
    // under awareness of current plan choice
    MemoTableEntry best = null;
    boolean opened = false;
    if (memo.contains(current.getHopID())) {
        if (currentType == null) {
            best = memo.get(current.getHopID()).stream().filter(p -> p.isValid()).filter(p -> hasNoRefToMaterialization(p, M, plan)).min(new BasicPlanComparator()).orElse(null);
            opened = true;
        } else {
            best = memo.get(current.getHopID()).stream().filter(p -> p.type == currentType || p.type == TemplateType.CELL).filter(p -> hasNoRefToMaterialization(p, M, plan)).min(Comparator.comparing(p -> 7 - ((p.type == currentType) ? 4 : 0) - p.countPlanRefs())).orElse(null);
        }
    }
    // create new cost vector if opened, initialized with write costs
    CostVector costVect = !opened ? costsCurrent : new CostVector(Math.max(current.getDim1(), 1) * Math.max(current.getDim2(), 1));
    // add compute costs of current operator to costs vector
    if (partition.contains(current.getHopID()))
        costVect.computeCosts += computeCosts.get(current.getHopID());
    // process children recursively
    double costs = 0;
    for (int i = 0; i < current.getInput().size(); i++) {
        Hop c = current.getInput().get(i);
        if (best != null && best.isPlanRef(i))
            costs += rGetPlanCosts(memo, c, visited, partition, M, plan, computeCosts, costVect, best.type);
        else if (best != null && isImplicitlyFused(current, i, best.type))
            costVect.addInputSize(c.getInput().get(0).getHopID(), Math.max(c.getDim1(), 1) * Math.max(c.getDim2(), 1));
        else {
            // include children and I/O costs
            costs += rGetPlanCosts(memo, c, visited, partition, M, plan, computeCosts, null, null);
            if (costVect != null && c.getDataType().isMatrix())
                costVect.addInputSize(c.getHopID(), Math.max(c.getDim1(), 1) * Math.max(c.getDim2(), 1));
        }
    }
    // add costs for opened fused operator
    if (partition.contains(current.getHopID())) {
        if (opened) {
            if (LOG.isTraceEnabled())
                LOG.trace("Cost vector for fused operator (hop " + current.getHopID() + "): " + costVect);
            // time for output write
            costs += costVect.outSize * 8 / WRITE_BANDWIDTH;
            costs += Math.max(costVect.computeCosts * costVect.getMaxInputSize() / COMPUTE_BANDWIDTH, costVect.getSumInputSizes() * 8 / READ_BANDWIDTH);
        } else // add costs for non-partition read in the middle of fused operator
        if (hasNonPartitionConsumer(current, partition)) {
            costs += rGetPlanCosts(memo, current, visited, partition, M, plan, computeCosts, null, null);
        }
    }
    // sanity check non-negative costs
    if (costs < 0 || Double.isNaN(costs) || Double.isInfinite(costs))
        throw new RuntimeException("Wrong cost estimate: " + costs);
    return costs;
}