compact syntax for spatial tensor ops

better IR compaction after dcp merge
remove pim.mvm op
better memory report

src/PIM/Compiler/PimCodeGen.hpp

@@ -33,6 +33,18 @@ struct MemoryReportRow {
   }
 };
 
+struct MemoryReportEntry {
+  enum class Kind {
+    Core,
+    Batch
+  };
+
+  Kind kind = Kind::Core;
+  uint64_t id = 0;
+  llvm::SmallVector<int32_t, 8> coreIds;
+  MemoryReportRow row;
+};
+
 class PimMemory {
   llvm::SmallVector<std::pair<MemEntry, mlir::Value>, 32> memEntries;
   llvm::SmallDenseMap<mlir::Value, MemEntry, 32>& globalMemEntriesMap;
@@ -66,7 +78,7 @@ private:
   llvm::SmallDenseMap<size_t, PimMemory> deviceMem;
   std::fstream fileReport;
   std::optional<MemoryReportRow> hostReportRow;
-  llvm::SmallVector<std::pair<size_t, MemoryReportRow>, 32> coreReportRows;
+  llvm::SmallVector<MemoryReportEntry, 32> reportEntries;
 
 public:
   PimAcceleratorMemory()
@@ -86,7 +98,8 @@ public:
 
   size_t getValueAddress(mlir::Value value, const StaticValueKnowledge& knowledge = {}) const;
   void reportHost();
-  void reportCore(size_t coreId);
+  void recordCoreReport(size_t coreId, const MemoryReportRow& row);
+  void recordBatchReport(uint64_t batchId, llvm::ArrayRef<int32_t> coreIds, const MemoryReportRow& row);
   void flushReport();
   void clean(mlir::Operation* op);
 };