fix dcp merge bug

src/PIM/Dialect/Spatial/Transforms/MergeComputeNodes/DCPGraph/DCPAnalysis.cpp

@@ -56,12 +56,10 @@ struct WindowScheduleResult {
   size_t maxMergeGroupSize = 0;
 };
 
-constexpr CPU kDefaultMaxCpuCount = 1000;
-
 size_t getSchedulingCpuBudget() {
   if (coresCount.getValue() > 0)
     return static_cast<size_t>(coresCount.getValue());
-  return static_cast<size_t>(kDefaultMaxCpuCount);
+  return std::numeric_limits<size_t>::max();
 }
 
 size_t getBatchChunkTargetCount(int32_t laneCount) {

src/PIM/Dialect/Spatial/Transforms/MergeComputeNodes/MergeComputeNodesPass.cpp

@@ -54,10 +54,9 @@ struct ProducerValueRef {
 std::optional<ProducerValueRef> getProducerValueRef(Value value);
 
 static size_t getFastPathCpuBudget() {
-  constexpr size_t kDefaultMaxCpuCount = 1000;
   if (coresCount.getValue() > 0)
     return static_cast<size_t>(coresCount.getValue());
-  return kDefaultMaxCpuCount;
+  return std::numeric_limits<size_t>::max();
 }
 
 static size_t getBatchChunkTargetCount(int32_t laneCount) {
@@ -670,6 +669,9 @@ void rebatchEquivalentComputes(func::FuncOp funcOp, int64_t& nextChannelId) {
       continue;
 
     SmallVector<SpatCompute> group {anchor};
+    llvm::SmallDenseSet<int32_t, 8> usedCoreIds;
+    if (auto coreId = getComputeCoreId(anchor))
+      usedCoreIds.insert(*coreId);
     if (!anchor.getResults().empty())
       continue;
     for (size_t candidateIndex = index + 1; candidateIndex < computes.size(); ++candidateIndex) {
@@ -680,8 +682,14 @@ void rebatchEquivalentComputes(func::FuncOp funcOp, int64_t& nextChannelId) {
         continue;
       if (!candidate.getResults().empty())
         continue;
-      if (areEquivalentForRebatch(anchor, candidate))
-        group.push_back(candidate);
+      if (!areEquivalentForRebatch(anchor, candidate))
+        continue;
+
+      if (auto coreId = getComputeCoreId(candidate))
+        if (!usedCoreIds.insert(*coreId).second)
+          continue;
+
+      group.push_back(candidate);
     }
 
     if (group.size() <= 1)