Memory Access Streaming Fusion Pass

include/TaskflowDialect/TaskflowPasses.h

@@ -28,6 +28,7 @@ std::unique_ptr<mlir::Pass> createMapTaskOnCgraPass();
 //=========================================================//
 std::unique_ptr<mlir::Pass> createAffineLoopTreeSerializationPass();
 std::unique_ptr<mlir::Pass> createAffineLoopPerfectionPass();
+std::unique_ptr<mlir::Pass> createMemoryAccessStreamingFusionPass();
 
 #define GEN_PASS_REGISTRATION
 #include "TaskflowDialect/TaskflowPasses.h.inc"

include/TaskflowDialect/TaskflowPasses.td

@@ -72,4 +72,18 @@ def MapTaskOnCgra : Pass<"map-task-on-cgra", "func::FuncOp"> {
   }];
   let constructor = "taskflow::createMapTaskOnCgraPass()";
 }
+
+def MemoryAccessStreamingFusion : Pass<"memory-access-streaming-fusion", "func::FuncOp"> {
+  let summary = "Fuses tasks connected by memory dependencies for streaming execution";
+  let description = [{
+    Identifies and fuses taskflow.task operations that have memory access
+    dependencies (one task writes a memref, another task reads it).
+    Eliminates intermediate memref allocations and converts memory access
+    dependencies into direct SSA value dependencies.
+
+    Uses greedy fusion strategy to iteratively fuse the most beneficial
+    task pairs until no more fusion opportunities exist.
+  }];
+  let constructor = "taskflow::createMemoryAccessStreamingFusionPass()";
+}
 #endif // TASKFLOW_PASSES_TD

lib/TaskflowDialect/Transforms/Optimizations/CMakeLists.txt

@@ -3,6 +3,7 @@ include_directories(${CMAKE_CURRENT_BINARY_DIR})
 add_mlir_conversion_library(MLIRTaskflowOptimization
   AffineLoopTreeSerializationPass.cpp
   AffineLoopPerfectionPass.cpp
+  MemoryAccessStreamingFusion.cpp
 
   DEPENDS
   MLIRTaskflowTransformsIncGen