/**
* Clones all fields of this into
*
* <pre>other</pre>
*/
protected void deepCloneInto(at.dms.kjc.slicegraph.FilterSliceNode other) {
super.deepCloneInto(other);
other.filter =
(at.dms.kjc.slicegraph.FilterContent) at.dms.kjc.AutoCloner.cloneToplevel(this.filter);
other.predefined = this.predefined;
other.laidout = this.laidout;
}
public static void run(
SIRStream str,
JInterfaceDeclaration[] interfaces,
SIRInterfaceTable[] interfaceTables,
SIRStructure[] structs,
SIRHelper[] helpers,
SIRGlobal global) {
System.out.println("Entry to SMP Backend...");
checkArguments();
setScheduler();
if (KjcOptions.smp > 16) {
setupLargeConfig();
}
// create cores in desired amount and order
int[] cores = new int[KjcOptions.smp];
for (int x = 0; x < KjcOptions.smp; x++) cores[x] = coreOrder[x];
chip = new SMPMachine(cores);
// create a new structs.h file for typedefs etc.
structs_h = new Structs_h(structs);
// The usual optimizations and transformation to slice graph
CommonPasses commonPasses = new CommonPasses();
// perform standard optimizations, use the number of cores the user wants to target
commonPasses.run(str, interfaces, interfaceTables, structs, helpers, global, chip.size());
// perform some standard cleanup on the slice graph.
commonPasses.simplifySlices();
// dump slice graph to dot file
commonPasses.getSlicer().dumpGraph("traces.dot", null);
// partition the slice graph based on the scheduling policy
SpaceTimeScheduleAndSlicer graphSchedule =
new SpaceTimeScheduleAndSlicer(commonPasses.getSlicer());
scheduler.setGraphSchedule(graphSchedule);
scheduler.run(chip.size());
FilterInfo.reset();
// generate schedules for initialization, primepump and steady-state
scheduleSlices(graphSchedule);
// generate layout for filters
scheduler.runLayout();
// dump final slice graph to dot file
graphSchedule.getSlicer().dumpGraph("after_slice_partition.dot", scheduler);
graphSchedule.getSlicer().dumpGraph("slice_graph.dot", scheduler, false);
// if load balancing, find candidiate fission groups to load balance
if (KjcOptions.loadbalance) {
LoadBalancer.findCandidates();
LoadBalancer.instrumentMainMethods();
}
// create all buffers and set the rotation lengths
RotatingBuffer.createBuffers(graphSchedule);
// now convert to Kopi code plus communication commands
backEndBits = new SMPBackEndFactory(chip, scheduler);
backEndBits.getBackEndMain().run(graphSchedule, backEndBits);
// generate code for file writer
CoreCodeStore.generatePrintOutputCode();
if (KjcOptions.numbers > 0) chip.getNthComputeNode(0).getComputeCode().generateNumbersCode();
// emit c code for all cores
EmitSMPCode.doit(backEndBits);
// dump structs.h file
structs_h.writeToFile();
// display final assignment of filters to cores
System.out.println("Final filter assignments:");
System.out.println("========================================");
for (int x = 0; x < KjcOptions.smp; x++) {
Core core = chip.getNthComputeNode(x);
Set<FilterSliceNode> filters = core.getComputeCode().getFilters();
long totalWork = 0;
System.out.println("Core " + core.getCoreID() + ": ");
for (FilterSliceNode filter : filters) {
long work = SliceWorkEstimate.getWork(filter.getParent());
System.out.format("%16d | " + filter + "\n", work);
totalWork += work;
}
System.out.format("%16d | Total\n", totalWork);
}
// calculate computation to communication ratio
if (KjcOptions.sharedbufs) {
LinkedList<Slice> slices =
DataFlowOrder.getTraversal(graphSchedule.getSlicer().getTopSlices());
HashSet<Slice> compProcessed = new HashSet<Slice>();
HashSet<Slice> commProcessed = new HashSet<Slice>();
long comp = 0;
long comm = 0;
for (Slice slice : slices) {
if (compProcessed.contains(slice)) continue;
comp += SliceWorkEstimate.getWork(slice);
compProcessed.add(slice);
}
/*
for(Slice slice : slices) {
if(commProcessed.contains(slice))
continue;
FilterInfo info = FilterInfo.getFilterInfo(slice.getFirstFilter());
int totalItemsReceived = info.totalItemsReceived(SchedulingPhase.STEADY);
if(totalItemsReceived == 0)
continue;
InputSliceNode input = slice.getHead();
Set<InterSliceEdge> sources = input.getSourceSet(SchedulingPhase.STEADY);
int numInputRots = totalItemsReceived / input.totalWeights(SchedulingPhase.STEADY);
if(!FissionGroupStore.isFizzed(slice)) {
for(InterSliceEdge source : sources) {
Slice srcSlice = source.getSrc().getParent();
// if(srcSlice.getFirstFilter().isFileInput())
// continue;
if(FissionGroupStore.isFizzed(srcSlice)) {
// Filter is not fizzed, source is fizzed
// Filter must receive (N-1)/N of inputs from different cores
comm += numInputRots *
input.getWeight(source, SchedulingPhase.STEADY) /
KjcOptions.smp * (KjcOptions.smp - 1);
}
else {
// Filter is not fizzed, source is not fizzed
// Check to see if on same core
// If not, must communicate all elements
if(!scheduler.getComputeNode(slice.getFirstFilter()).equals(
scheduler.getComputeNode(srcSlice.getFirstFilter()))) {
comm += numInputRots *
input.getWeight(source, SchedulingPhase.STEADY);
}
}
}
}
else {
for(InterSliceEdge source : sources) {
Slice srcSlice = source.getSrc().getParent();
// if(srcSlice.getFirstFilter().isFileInput())
// continue;
if(FissionGroupStore.isFizzed(srcSlice)) {
// Filter is fizzed, source is also fizzed
int totalItemsReceivedPerFizzed = totalItemsReceived /
FissionGroupStore.getFissionGroup(slice).fizzedSlices.length;
int numInputRotsPerFizzed = numInputRots /
FissionGroupStore.getFissionGroup(slice).fizzedSlices.length;
System.out.println("totalItemsReceivedPerFizzed: " + totalItemsReceivedPerFizzed);
System.out.println("numInputRotsPerFizzed: " + numInputRotsPerFizzed);
int inputWeightBeforeSrc =
input.weightBefore(source, SchedulingPhase.STEADY);
int inputWeightSrc = input.getWeight(source, SchedulingPhase.STEADY);
int inputTotalWeight = input.totalWeights(SchedulingPhase.STEADY);
System.out.println("inputWeightBeforeSrc: " + inputWeightBeforeSrc);
System.out.println("inputWeightSrc: " + inputWeightSrc);
System.out.println("copyDown: " + info.copyDown);
int numXmit = 0;
for(int rot = 0 ; rot < numInputRotsPerFizzed ; rot++) {
numXmit += Math.min(inputWeightSrc,
Math.max(0,
info.copyDown +
rot * inputTotalWeight +
inputWeightBeforeSrc + inputWeightSrc -
totalItemsReceivedPerFizzed));
}
System.out.println("numXmit: " + numXmit);
comm += KjcOptions.smp * numXmit;
}
else {
// Filter is fizzed, source is not fizzed
// Source must send (N-1)/N of outputs to different cores
comm += numInputRots *
input.getWeight(source, SchedulingPhase.STEADY) /
KjcOptions.smp * (KjcOptions.smp - 1);
}
}
}
commProcessed.add(slice);
}
*/
// Simple communication estimation
for (Slice slice : slices) {
if (commProcessed.contains(slice)) continue;
FilterInfo info = FilterInfo.getFilterInfo(slice.getFirstFilter());
int totalItemsReceived = info.totalItemsReceived(SchedulingPhase.STEADY);
if (totalItemsReceived == 0) continue;
comm += totalItemsReceived;
if (FissionGroupStore.isFizzed(slice)) {
assert info.peek >= info.pop;
comm += (info.peek - info.pop) * KjcOptions.smp;
}
commProcessed.add(slice);
}
// Simple communication estimation 2
/*
for(Slice slice : slices) {
if(commProcessed.contains(slice))
continue;
FilterInfo info = FilterInfo.getFilterInfo(slice.getFirstFilter());
int totalItemsReceived = info.totalItemsReceived(SchedulingPhase.STEADY);
int totalItemsSent = info.totalItemsSent(SchedulingPhase.STEADY);
comm += totalItemsReceived;
comm += totalItemsSent;
if(totalItemsReceived == 0)
continue;
if(FissionGroupStore.isFizzed(slice)) {
assert info.peek >= info.pop;
comm += (info.peek - info.pop) * KjcOptions.smp;
}
commProcessed.add(slice);
}
*/
System.out.println("Final Computation: " + comp);
System.out.println("Final Communication: " + comm);
System.out.println("Final Comp/Comm Ratio: " + (float) comp / (float) comm);
}
System.exit(0);
}
private void flattenInternal(FlatNode top) {
Iterator<FlatNode> dataFlow = DataFlowTraversal.getTraversal(top).iterator();
while (dataFlow.hasNext()) {
FlatNode node = dataFlow.next();
System.out.println(node);
InputSliceNode input = sliceNodes.inputNodes.get(node.contents);
OutputSliceNode output = sliceNodes.outputNodes.get(node.contents);
FilterSliceNode filterNode = sliceNodes.filterNodes.get(node.contents);
assert input != null && output != null && filterNode != null;
// set up the slice
Slice slice = new Slice(input);
input.setNext(filterNode);
filterNode.setPrevious(input);
filterNode.setNext(output);
output.setPrevious(filterNode);
input.setParent(slice);
output.setParent(slice);
filterNode.setParent(slice);
System.out.println(" outputs: " + node.ways);
if (node.ways != 0) {
assert node.ways == node.getEdges().length && node.ways == node.weights.length;
// set up the i/o arcs
// set up the splitting...
LinkedList<InterSliceEdge> outEdges = new LinkedList<InterSliceEdge>();
LinkedList<Integer> outWeights = new LinkedList<Integer>();
HashMap<InputSliceNode, InterSliceEdge> newEdges =
new HashMap<InputSliceNode, InterSliceEdge>();
for (int i = 0; i < node.ways; i++) {
if (node.weights[i] == 0) continue;
InterSliceEdge edge =
new InterSliceEdge(output, sliceNodes.inputNodes.get(node.getEdges()[i].contents));
newEdges.put(sliceNodes.inputNodes.get(node.getEdges()[i].contents), edge);
outEdges.add(edge);
outWeights.add(node.weights[i]);
}
edges.put(output, newEdges);
LinkedList<LinkedList<InterSliceEdge>> translatedEdges =
new LinkedList<LinkedList<InterSliceEdge>>();
if (node.isDuplicateSplitter()) {
outWeights = new LinkedList<Integer>();
outWeights.add(new Integer(1));
translatedEdges.add(outEdges);
} else {
for (int i = 0; i < outEdges.size(); i++) {
LinkedList<InterSliceEdge> link = new LinkedList<InterSliceEdge>();
link.add(outEdges.get(i));
translatedEdges.add(link);
}
}
output.set(outWeights, translatedEdges);
} else {
// no outputs
output.setWeights(new int[0]);
output.setDests(new InterSliceEdge[0][0]);
}
if (node.isFilter()) {
if (node.getFilter().getPushInt() == 0) {
output.setWeights(new int[0]);
output.setDests(new InterSliceEdge[0][0]);
}
}
// set up the joining, the edges should exist already from upstream
System.out.println(" inputs: " + node.inputs);
if (node.inputs != 0) {
assert node.inputs == node.incoming.length && node.inputs == node.incomingWeights.length;
LinkedList<Integer> inWeights = new LinkedList<Integer>();
LinkedList<InterSliceEdge> inEdges = new LinkedList<InterSliceEdge>();
for (int i = 0; i < node.inputs; i++) {
if (node.incomingWeights[i] == 0) continue;
inEdges.add(edges.get(sliceNodes.outputNodes.get(node.incoming[i].contents)).get(input));
inWeights.add(node.incomingWeights[i]);
}
input.set(inWeights, inEdges);
} else {
input.setWeights(new int[0]);
input.setSources(new InterSliceEdge[0]);
}
if (node.isFilter() && node.getFilter().getPopInt() == 0) {
input.setWeights(new int[0]);
input.setSources(new InterSliceEdge[0]);
}
// set up the work hashmaps
int workEst = 0;
if (sliceNodes.generatedIds.contains(filterNode)) {
workEst = 3 * filterNode.getFilter().getSteadyMult();
} else {
assert node.isFilter();
workEst = work.getWork((SIRFilter) node.contents);
}
bottleNeckFilter.put(slice, filterNode);
sliceBNWork.put(slice, workEst);
workEstimation.put(filterNode.getFilter(), workEst);
slice.finish();
if (node.contents instanceof SIRFileReader || node.contents instanceof SIRFileWriter) {
System.out.println("Found io " + node.contents);
ioList.add(slice);
}
if (topSlice == null) topSlice = slice;
sliceList.add(slice);
}
}
