Beispiel #1
0
  /**
   * Create schedules for init, prime-pump and steady phases.
   *
   * @return a Scheduler from which the schedules for the phases may be extracted.
   */
  public static void scheduleSlices(SpaceTimeScheduleAndSlicer schedule) {
    Slicer slicer = schedule.getSlicer();

    // set init schedule in standard order
    schedule.setInitSchedule(DataFlowOrder.getTraversal(slicer.getSliceGraph()));

    // set the prime pump to be empty
    new GeneratePrimePump(schedule).setEmptySchedule();

    // for space multiplexing on SMP we need to use a different primepump scheduler because
    // we are space multiplexing and we need to prime the pipe more so that everything can fire
    // when ready
    if (at.dms.kjc.smp.SMPBackend.scheduler.isSMD())
      new at.dms.kjc.smp.GeneratePrimePumpScheduleSMD(schedule).schedule(slicer.getSliceGraph());
    else new GeneratePrimePump(schedule).schedule(slicer.getSliceGraph());

    // Still need to generate the steady state schedule!
    schedule.setSchedule(DataFlowOrder.getTraversal(slicer.getTopSlices()));
  }
Beispiel #2
0
  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);
  }