No extract no more

src/PIM/Dialect/Spatial/Transforms/MergeComputeNodes/Scheduling/PeftScheduler.cpp

@@ -5,6 +5,7 @@
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/FormatVariadic.h"
 
+#include <functional>
 #include <limits>
 #include <queue>
 #include <vector>
@@ -177,20 +178,25 @@ MergeScheduleResult runPeftScheduler(const ComputeGraph& graph, const PeftSchedu
     Time bestEst = 0;
     Time bestEft = 0;
     Time bestOeft = std::numeric_limits<Time>::max();
+    unsigned int bestOverlapWeight = 0;
     bool crossbarRejected = false;
 
+    auto crossbarsAreContainedInProcessor = [&processorCrossbars](mlir::Value nodeCrossbar, size_t processor) {
+      return llvm::is_contained(processorCrossbars[processor], nodeCrossbar);
+    };
+
     for (size_t processor = 0; processor < processorCount; ++processor) {
+      auto crossbarsAreContained = std::bind(crossbarsAreContainedInProcessor, std::placeholders::_1, processor);
       if (graph.nodes[task].crossbarUsage.size() != 0
-          && !llvm::all_of(graph.nodes[task].crossbarUsage,
-                           [&processorCrossbars, processor](mlir::Value nodeCrossbar) {
-                             return llvm::is_contained(processorCrossbars[processor], nodeCrossbar);
-                           })
+          && !llvm::all_of(graph.nodes[task].crossbarUsage, crossbarsAreContained)
           && addOrMax(processorCrossbars[processor].size(), graph.nodes[task].crossbarUsage.size())
                > options.crossbarCapacity) {
 
-        crossbarRejected = true;
-        continue;
-      }
+            crossbarRejected = true;
+            continue;
+          }
+      unsigned int overlapWeight =
+        llvm::count_if(graph.nodes[task].crossbarUsage, crossbarsAreContained);
 
       Time dataReady = 0;
       for (const auto& [pred, comm] : graph.predecessors[task]) {
@@ -224,12 +230,19 @@ MergeScheduleResult runPeftScheduler(const ComputeGraph& graph, const PeftSchedu
       Time oeft = addOrMax(eft, oct[task * processorCount + processor]);
 
       if (oeft < bestOeft || (oeft == bestOeft && eft < bestEft)
-          || (oeft == bestOeft && eft == bestEft && est < bestEst)
-          || (oeft == bestOeft && eft == bestEft && est == bestEst && processor < bestProcessor)) {
+          || (oeft == bestOeft && eft == bestEft && est < bestEst)) {
         bestProcessor = processor;
         bestEst = est;
         bestEft = eft;
         bestOeft = oeft;
+        bestOverlapWeight = overlapWeight;
+      }
+      else if (oeft == bestOeft && eft == bestEft && est < bestEst && overlapWeight < bestOverlapWeight) {
+        bestProcessor = processor;
+        bestEst = est;
+        bestEft = eft;
+        bestOeft = oeft;
+        bestOverlapWeight = overlapWeight;
       }
     }