[LoopPipelining] recursively fully unroll all inner loops until no failure are met

README.md

@@ -40,7 +40,7 @@ $ # Loop and pragma-level optimizations, performance estimation, and C++ code ge
 $ scalehls-opt samples/polybench/syrk.mlir \
     -affine-loop-perfection -remove-variable-bound -affine-loop-normalize \
     -partial-affine-loop-tile="tile-level=1 tile-size=4" \
-    -legalize-to-hlscpp="top-function=test_syrk" -loop-pipelining="pipeline-level=1" \
+    -legalize-to-hlscpp="top-func=test_syrk" -loop-pipelining="pipeline-level=1" \
     -affine-store-forward -simplify-memref-access -array-partition -cse -canonicalize \
     -qor-estimation="target-spec=config/target-spec.ini" \
     | scalehls-translate -emit-hlscpp
@@ -55,7 +55,7 @@ $ benchmark-gen -type "cnn" -config "config/cnn-config.ini" -number 1 \
     | scalehls-opt -legalize-dataflow -split-function \
     -hlskernel-bufferize -hlskernel-to-affine -func-bufferize \
     -affine-loop-perfection -affine-loop-normalize \
-    -legalize-to-hlscpp="top-function=auto_gen_cnn" \
+    -legalize-to-hlscpp="top-func=auto_gen_cnn" \
     -affine-store-forward -simplify-memref-access -cse -canonicalize \
     -qor-estimation="target-spec=config/target-spec.ini" \
     | scalehls-translate -emit-hlscpp
@@ -85,7 +85,7 @@ $ # Legalize the output of ONNX-MLIR, optimize and emit C++ code.
 $ scalehls-opt resnet18.mlir -legalize-onnx -affine-loop-normalize -canonicalize \
     -legalize-dataflow="min-gran=2 insert-copy=false" -split-function \
     -convert-linalg-to-affine-loops -affine-loop-fusion \
-    -legalize-to-hlscpp="top-function=main_graph" \
+    -legalize-to-hlscpp="top-func=main_graph" \
     | scalehls-translate -emit-hlscpp
 ```
 

include/Conversion/Passes.td

@@ -22,7 +22,7 @@ def LegalizeToHLSCpp : Pass<"legalize-to-hlscpp", "FuncOp"> {
   let constructor = "mlir::scalehls::createLegalizeToHLSCppPass()";
 
   let options = [
-    Option<"topFunction", "top-function", "std::string", /*default=*/"\"main\"", 
+    Option<"topFunc", "top-func", "std::string", /*default=*/"\"main\"", 
            "The top function for HLS synthesis">,
   ];
 }

include/Transforms/Passes.td

@@ -66,7 +66,7 @@ def FuncPipelining : Pass<"func-pipelining", "FuncOp"> {
   let constructor = "mlir::scalehls::createFuncPipeliningPass()";
 
   let options = [
-    Option<"targetFunction", "target-function", "std::string",
+    Option<"targetFunc", "target-func", "std::string",
            /*default=*/"\"main\"", "The target function to be pipelined">
   ];
 }

lib/Conversion/LegalizeToHLSCpp.cpp

@@ -29,7 +29,7 @@ void LegalizeToHLSCpp::runOnOperation() {
   if (!func->getAttr("dataflow"))
     func->setAttr("dataflow", builder.getBoolAttr(false));
 
-  if (func.getName() == topFunction)
+  if (func.getName() == topFunc)
     func->setAttr("top_function", builder.getBoolAttr(true));
   else
     func->setAttr("top_function", builder.getBoolAttr(false));

lib/Transforms/FuncPipelining.cpp

@@ -18,7 +18,7 @@ struct FuncPipelining : public FuncPipeliningBase<FuncPipelining> {
     auto func = getOperation();
     auto builder = OpBuilder(func);
 
-    if (func.getName() == targetFunction) {
+    if (func.getName() == targetFunc) {
       applyFuncPipelining(func, builder);
 
       // Canonicalize the IR after function pipelining.
@@ -35,7 +35,14 @@ struct FuncPipelining : public FuncPipeliningBase<FuncPipelining> {
 /// Apply function pipelining to the input function, all contained loops are
 /// automatically fully unrolled.
 bool scalehls::applyFuncPipelining(FuncOp func, OpBuilder &builder) {
-  func.walk([&](AffineForOp loop) { loopUnrollFull(loop); });
+  bool hasFullyUnrolled = false;
+  while (hasFullyUnrolled == false) {
+    hasFullyUnrolled = true;
+    func.walk([&](AffineForOp loop) {
+      if (failed(loopUnrollFull(loop)))
+        hasFullyUnrolled = false;
+    });
+  }
 
   func->setAttr("pipeline", builder.getBoolAttr(true));
   func->setAttr("dataflow", builder.getBoolAttr(false));

lib/Transforms/LoopPipelining.cpp

@@ -63,10 +63,15 @@ bool scalehls::applyLoopPipelining(AffineForOp targetLoop, OpBuilder &builder) {
   targetLoop->setAttr("pipeline", builder.getBoolAttr(true));
 
   // All inner loops of the pipelined loop are automatically unrolled.
-  targetLoop.walk([&](AffineForOp loop) {
-    if (loop != targetLoop)
-      loopUnrollFull(loop);
-  });
+  bool hasFullyUnrolled = false;
+  while (hasFullyUnrolled == false) {
+    hasFullyUnrolled = true;
+    targetLoop.walk([&](AffineForOp loop) {
+      if (loop != targetLoop)
+        if (failed(loopUnrollFull(loop)))
+          hasFullyUnrolled = false;
+    });
+  }
 
   // All outer loops that perfect nest the pipelined loop can be flattened.
   SmallVector<AffineForOp, 4> flattenedLoops;

test/Conversion/LegalizeToHLSCpp/test_assign.mlir

@@ -1,4 +1,4 @@
-// RUN: scalehls-opt -legalize-to-hlscpp="top-function=test_assign" %s | FileCheck %s
+// RUN: scalehls-opt -legalize-to-hlscpp="top-func=test_assign" %s | FileCheck %s
 
 // CHECK-LABEL: func @test_assign(
 // CHECK-SAME:  %arg0: f32, %arg1: memref<16xf32, 1>) -> (f32, memref<16xf32, 1>, i32, memref<2x2xi32, 1>) attributes {dataflow = false, top_function = true} {