use of org.apache.sysml.hops.codegen.template.TemplateBase.TemplateType 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 -> isValid(p, current)).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.CellTpl).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;
}
use of org.apache.sysml.hops.codegen.template.TemplateBase.TemplateType in project incubator-systemml by apache.
the class PlanSelectionFuseCostBased method rSelectPlansFuseAll.
private void rSelectPlansFuseAll(CPlanMemoTable memo, Hop current, TemplateType currentType, HashSet<Long> partition) {
if (isVisited(current.getHopID(), currentType) || !partition.contains(current.getHopID()))
return;
//step 1: prune subsumed plans of same type
if (memo.contains(current.getHopID())) {
HashSet<MemoTableEntry> rmSet = new HashSet<MemoTableEntry>();
List<MemoTableEntry> hopP = memo.get(current.getHopID());
for (MemoTableEntry e1 : hopP) for (MemoTableEntry e2 : hopP) if (e1 != e2 && e1.subsumes(e2))
rmSet.add(e2);
memo.remove(current, rmSet);
}
//step 2: select plan for current path
MemoTableEntry best = null;
if (memo.contains(current.getHopID())) {
if (currentType == null) {
best = memo.get(current.getHopID()).stream().filter(p -> isValid(p, current)).min(new BasicPlanComparator()).orElse(null);
} else {
best = memo.get(current.getHopID()).stream().filter(p -> p.type == currentType || p.type == TemplateType.CellTpl).min(Comparator.comparing(p -> 7 - ((p.type == currentType) ? 4 : 0) - p.countPlanRefs())).orElse(null);
}
addBestPlan(current.getHopID(), best);
}
//step 3: recursively process children
for (int i = 0; i < current.getInput().size(); i++) {
TemplateType pref = (best != null && best.isPlanRef(i)) ? best.type : null;
rSelectPlansFuseAll(memo, current.getInput().get(i), pref, partition);
}
setVisited(current.getHopID(), currentType);
}
use of org.apache.sysml.hops.codegen.template.TemplateBase.TemplateType in project incubator-systemml by apache.
the class PlanSelectionFuseNoRedundancy method rSelectPlans.
private void rSelectPlans(CPlanMemoTable memo, Hop current, TemplateType currentType) {
if (isVisited(current.getHopID(), currentType))
return;
//step 0: remove plans that refer to a common partial plan
if (memo.contains(current.getHopID())) {
HashSet<MemoTableEntry> rmSet = new HashSet<MemoTableEntry>();
List<MemoTableEntry> hopP = memo.get(current.getHopID());
for (MemoTableEntry e1 : hopP) for (int i = 0; i < 3; i++) if (e1.isPlanRef(i) && current.getInput().get(i).getParent().size() > 1)
//remove references to hops w/ multiple consumers
rmSet.add(e1);
memo.remove(current, rmSet);
}
//step 1: prune subsumed plans of same type
if (memo.contains(current.getHopID())) {
HashSet<MemoTableEntry> rmSet = new HashSet<MemoTableEntry>();
List<MemoTableEntry> hopP = memo.get(current.getHopID());
for (MemoTableEntry e1 : hopP) for (MemoTableEntry e2 : hopP) if (e1 != e2 && e1.subsumes(e2))
rmSet.add(e2);
memo.remove(current, rmSet);
}
//step 2: select plan for current path
MemoTableEntry best = null;
if (memo.contains(current.getHopID())) {
if (currentType == null) {
best = memo.get(current.getHopID()).stream().filter(p -> isValid(p, current)).min(new BasicPlanComparator()).orElse(null);
} else {
best = memo.get(current.getHopID()).stream().filter(p -> p.type == currentType || p.type == TemplateType.CellTpl).min(Comparator.comparing(p -> 7 - ((p.type == currentType) ? 4 : 0) - p.countPlanRefs())).orElse(null);
}
addBestPlan(current.getHopID(), best);
}
//step 3: recursively process children
for (int i = 0; i < current.getInput().size(); i++) {
TemplateType pref = (best != null && best.isPlanRef(i)) ? best.type : null;
rSelectPlans(memo, current.getInput().get(i), pref);
}
setVisited(current.getHopID(), currentType);
}
use of org.apache.sysml.hops.codegen.template.TemplateBase.TemplateType in project incubator-systemml by apache.
the class PlanSelectionFuseAll method rSelectPlans.
private void rSelectPlans(CPlanMemoTable memo, Hop current, TemplateType currentType) {
if (isVisited(current.getHopID(), currentType))
return;
//step 1: prune subsumed plans of same type
if (memo.contains(current.getHopID())) {
HashSet<MemoTableEntry> rmSet = new HashSet<MemoTableEntry>();
List<MemoTableEntry> hopP = memo.get(current.getHopID());
for (MemoTableEntry e1 : hopP) for (MemoTableEntry e2 : hopP) if (e1 != e2 && e1.subsumes(e2))
rmSet.add(e2);
memo.remove(current, rmSet);
}
//step 2: select plan for current path
MemoTableEntry best = null;
if (memo.contains(current.getHopID())) {
if (currentType == null) {
best = memo.get(current.getHopID()).stream().filter(p -> isValid(p, current)).min(new BasicPlanComparator()).orElse(null);
} else {
best = memo.get(current.getHopID()).stream().filter(p -> p.type == currentType || p.type == TemplateType.CellTpl).min(Comparator.comparing(p -> 7 - ((p.type == currentType) ? 4 : 0) - p.countPlanRefs())).orElse(null);
}
addBestPlan(current.getHopID(), best);
}
//step 3: recursively process children
for (int i = 0; i < current.getInput().size(); i++) {
TemplateType pref = (best != null && best.isPlanRef(i)) ? best.type : null;
rSelectPlans(memo, current.getInput().get(i), pref);
}
setVisited(current.getHopID(), currentType);
}
Aggregations