[LegalizeDataflow] update impl structure and fix bugs, support resnet18 legalization; [EmitHLSCpp] emit bool rather than ap_int<1>

.gitignore

@@ -2,6 +2,7 @@
 .ipynb_checkpoints
 
 build
+tmp
 samples/hls_proj
 samples/cpp_src
 samples/test_results

include/Analysis/Utils.h

@@ -116,17 +116,12 @@ hlscpp::ArrayOp getArrayOp(Value memref);
 
 hlscpp::ArrayOp getArrayOp(Operation *op);
 
-// For storing all accessed memrefs indexed by an operation (e.g. AffineForOp).
-using MemRefs = SmallVector<Value, 4>;
-using MemRefsMap = DenseMap<Operation *, MemRefs>;
+// For storing the intermediate memory and successor loops indexed by the
+// predecessor loop.
+using Successors = SmallVector<std::pair<Value, Operation *>, 2>;
+using SuccessorsMap = DenseMap<Operation *, Successors>;
 
-/// With the generated MemRefsMap, given a specific loop, we can easily find all
-/// memories which are consumed by the loop.
-void getLoopLoadMemsMap(Block &block, MemRefsMap &map);
-
-/// With the generated MemAccessesMap, given a specific memory, we can easily
-/// find the loops which produce data to the memory.
-void getLoopMemStoresMap(Block &block, MemAccessesMap &map);
+void getSuccessorsMap(Block &block, SuccessorsMap &map);
 
 } // namespace scalehls
 } // namespace mlir

lib/Analysis/Utils.cpp

@@ -4,6 +4,7 @@
 
 #include "Analysis/Utils.h"
 #include "mlir/Analysis/AffineAnalysis.h"
+#include "llvm/ADT/SmallPtrSet.h"
 
 using namespace mlir;
 using namespace scalehls;
@@ -140,41 +141,34 @@ hlscpp::ArrayOp scalehls::getArrayOp(Operation *op) {
   return getArrayOp(MemRefAccess(op).memref);
 }
 
-/// With the generated MemRefsMap, given a specific loop, we can easily find all
-/// memories which are consumed by the loop.
-void scalehls::getLoopLoadMemsMap(Block &block, MemRefsMap &map) {
-  for (auto loop : block.getOps<AffineForOp>()) {
-    loop.walk([&](Operation *op) {
-      if (auto affineLoad = dyn_cast<AffineLoadOp>(op)) {
-        auto &mems = map[loop];
-        if (std::find(mems.begin(), mems.end(), affineLoad.getMemRef()) ==
-            mems.end())
-          mems.push_back(affineLoad.getMemRef());
+void scalehls::getSuccessorsMap(Block &block, SuccessorsMap &map) {
+  DenseMap<Operation *, SmallPtrSet<Value, 2>> memsMap;
+  DenseMap<Value, SmallPtrSet<Operation *, 2>> loopsMap;
 
-      } else if (auto load = dyn_cast<LoadOp>(op)) {
-        auto &mems = map[loop];
-        if (std::find(mems.begin(), mems.end(), load.getMemRef()) == mems.end())
-          mems.push_back(load.getMemRef());
-      }
-    });
-  }
-}
-
-/// With the generated MemAccessesMap, given a specific memory, we can easily
-/// find the loops which produce data to the memory.
-void scalehls::getLoopMemStoresMap(Block &block, MemAccessesMap &map) {
-  for (auto loop : block.getOps<AffineForOp>()) {
+  for (auto loop : block.getOps<AffineForOp>())
     loop.walk([&](Operation *op) {
       if (auto affineStore = dyn_cast<AffineStoreOp>(op)) {
-        auto &loops = map[affineStore.getMemRef()];
-        if (std::find(loops.begin(), loops.end(), loop) == loops.end())
-          loops.push_back(loop);
+        memsMap[loop].insert(affineStore.getMemRef());
 
       } else if (auto store = dyn_cast<StoreOp>(op)) {
-        auto &loops = map[store.getMemRef()];
-        if (std::find(loops.begin(), loops.end(), loop) == loops.end())
-          loops.push_back(loop);
+        memsMap[loop].insert(store.getMemRef());
+
+      } else if (auto affineLoad = dyn_cast<AffineLoadOp>(op)) {
+        loopsMap[affineLoad.getMemRef()].insert(loop);
+
+      } else if (auto load = dyn_cast<LoadOp>(op)) {
+        loopsMap[load.getMemRef()].insert(loop);
       }
     });
-  }
+
+  for (auto loop : block.getOps<AffineForOp>())
+    for (auto mem : memsMap[loop])
+      for (auto successor : loopsMap[mem]) {
+        // If the successor loop not only loads from the memory, but also store
+        // to the memory, it will not be considered as a legal successor.
+        if (successor == loop || memsMap[successor].count(mem))
+          continue;
+
+        map[loop].push_back(std::pair<Value, Operation *>(mem, successor));
+      }
 }

lib/EmitHLSCpp/EmitHLSCpp.cpp

@@ -1347,10 +1347,14 @@ void ModuleEmitter::emitValue(Value val, unsigned rank, bool isPtr) {
   else if (valType.isa<IndexType>())
     os << "int ";
   else if (auto intType = valType.dyn_cast<IntegerType>()) {
-    os << "ap_";
-    if (intType.getSignedness() == IntegerType::SignednessSemantics::Unsigned)
-      os << "u";
-    os << "int<" << intType.getWidth() << "> ";
+    if (intType.getWidth() == 1)
+      os << "bool ";
+    else {
+      os << "ap_";
+      if (intType.getSignedness() == IntegerType::SignednessSemantics::Unsigned)
+        os << "u";
+      os << "int<" << intType.getWidth() << "> ";
+    }
   } else
     emitError(val.getDefiningOp(), "has unsupported type.");
 

lib/Transforms/LegalizeDataflow.cpp

@@ -41,7 +41,7 @@ void LegalizeDataflow::runOnOperation() {
           dataflowLevel = max(dataflowLevel, attr.getInt());
         else
           op->emitError(
-              "HLSKernelOp has unexpected predecessor, legalization failed");
+              "HLSKernelOp has unexpected successor, legalization failed");
       }
     }
 
@@ -86,69 +86,59 @@ void LegalizeDataflow::runOnOperation() {
   // this point. Therefore, HLSKernel ops and loops will never have dependencies
   // with each other in this pass.
   // TODO: analyze live ins.
-  MemRefsMap loadMemsMap;
-  MemAccessesMap memStoresMap;
-  getLoopLoadMemsMap(func.front(), loadMemsMap);
-  getLoopMemStoresMap(func.front(), memStoresMap);
+  SuccessorsMap successorsMap;
+  getSuccessorsMap(func.front(), successorsMap);
 
-  for (auto loop : func.front().getOps<mlir::AffineForOp>()) {
-    int64_t dataflowLevel = 0;
-    for (auto mem : loadMemsMap[loop]) {
-      for (auto predLoop : memStoresMap[mem]) {
-        if (predLoop == loop)
-          continue;
+  for (auto it = func.front().rbegin(); it != func.front().rend(); ++it) {
+    if (auto loop = dyn_cast<mlir::AffineForOp>(*it)) {
+      int64_t dataflowLevel = 0;
 
-        // Establish an ASAP dataflow schedule.
-        if (auto attr = predLoop->getAttrOfType<IntegerAttr>("dataflow_level"))
+      // Walk through all successor loops.
+      for (auto pair : successorsMap[loop]) {
+        auto successor = pair.second;
+        if (auto attr = successor->getAttrOfType<IntegerAttr>("dataflow_level"))
           dataflowLevel = max(dataflowLevel, attr.getInt());
-        else
-          loop.emitError(
-              "loop has unexpected predecessor, legalization failed");
+        else {
+          loop.emitError("loop has unexpected successor, legalization failed");
+          return;
+        }
       }
-    }
 
-    // Set an attribute for indicating the scheduled dataflow level.
-    loop.setAttr("dataflow_level", builder.getIntegerAttr(builder.getI64Type(),
-                                                          dataflowLevel + 1));
+      // Set an attribute for indicating the scheduled dataflow level.
+      loop.setAttr(
+          "dataflow_level",
+          builder.getIntegerAttr(builder.getI64Type(), dataflowLevel + 1));
 
-    // Eliminate bypass paths.
-    for (auto mem : loadMemsMap[loop]) {
-      for (auto predLoop : memStoresMap[mem]) {
-        if (predLoop == loop)
-          continue;
+      // Eliminate bypass paths.
+      for (auto pair : successorsMap[loop]) {
+        auto mem = pair.first;
+        auto successor = pair.second;
+        auto successorDataflowLevel =
+            successor->getAttrOfType<IntegerAttr>("dataflow_level").getInt();
 
-        auto predDataflowLevel =
-            predLoop->getAttrOfType<IntegerAttr>("dataflow_level").getInt();
+        // Insert CopyOps if required.
+        SmallVector<Value, 4> mems;
+        mems.push_back(mem);
+        builder.setInsertionPoint(successor);
 
-        // Insert dummy CopyOps if required.
-        SmallVector<Operation *, 4> dummyOps;
-        dummyOps.push_back(loop);
-        for (auto i = dataflowLevel; i > predDataflowLevel; --i) {
+        for (auto i = dataflowLevel; i > successorDataflowLevel; --i) {
           // Create CopyOp.
-          builder.setInsertionPoint(dummyOps.back());
-          auto interMem = builder.create<mlir::AllocOp>(
+          auto newMem = builder.create<mlir::AllocOp>(
               loop.getLoc(), mem.getType().cast<MemRefType>());
-          auto dummyOp =
-              builder.create<linalg::CopyOp>(loop.getLoc(), mem, interMem);
-          dummyOp.setAttr("dataflow_level",
-                          builder.getIntegerAttr(builder.getI64Type(), i));
+          auto copyOp = builder.create<linalg::CopyOp>(loop.getLoc(),
+                                                       mems.back(), newMem);
+
+          // Set CopyOp dataflow level.
+          copyOp.setAttr("dataflow_level",
+                         builder.getIntegerAttr(builder.getI64Type(), i));
 
           // Chain created CopyOps.
-          if (i == dataflowLevel) {
-            loop.walk([&](Operation *op) {
-              if (auto affineLoad = dyn_cast<mlir::AffineLoadOp>(op)) {
-                if (affineLoad.getMemRef() == mem)
-                  affineLoad.setMemRef(interMem);
-
-              } else if (auto load = dyn_cast<mlir::LoadOp>(op)) {
-                if (load.getMemRef() == mem)
-                  load.setMemRef(interMem);
-              }
+          if (i == successorDataflowLevel + 1)
+            mem.replaceUsesWithIf(newMem, [&](mlir::OpOperand &use) {
+              return successor->isProperAncestor(use.getOwner());
             });
-          } else
-            dummyOps.back()->setOperand(0, interMem);
-
-          dummyOps.push_back(dummyOp);
+          else
+            mems.push_back(newMem);
         }
       }
     }

lib/Transforms/SplitFunction.cpp

@@ -107,7 +107,7 @@ void SplitFunction::runOnOperation() {
         for (unsigned i = 0, e = inputValues.size(); i < e; ++i)
           inputValues[i].replaceUsesWithIf(
               entry->getArgument(i), [&](mlir::OpOperand &use) {
-                return getSameLevelDstOp(returnOp, use.getOwner());
+                return func.getOperation()->isProperAncestor(use.getOwner());
               });
         opIndex += 1;
       }

test/EmitHLSCpp/test_compare.mlir

@@ -2,7 +2,7 @@
 
 func @test_integer_compare(%arg0: i32, %arg1: i32) -> i1 {
 
-  // CHECK: ap_int<1> [[VAL_0:.*]] = [[ARG_0:.*]] == [[ARG_1:.*]];
+  // CHECK: bool [[VAL_0:.*]] = [[ARG_0:.*]] == [[ARG_1:.*]];
   %0 = cmpi "eq", %arg0, %arg1 : i32
   // CHECK: !=
   %1 = cmpi "ne", %arg0, %arg1 : i32
@@ -27,7 +27,7 @@ func @test_integer_compare(%arg0: i32, %arg1: i32) -> i1 {
 
 func @test_float_compare(%arg0: f32, %arg1: f32) -> i1 {
 
-  // CHECK: ap_int<1> [[VAL_0:.*]] = [[ARG_0:.*]] == [[ARG_1:.*]];
+  // CHECK: bool [[VAL_0:.*]] = [[ARG_0:.*]] == [[ARG_1:.*]];
   %0 = cmpf "oeq", %arg0, %arg1 : f32
   // CHECK: ==
   %1 = cmpf "ueq", %arg0, %arg1 : f32

test/EmitHLSCpp/test_constant.mlir

@@ -8,7 +8,7 @@ func @test_constant(%arg0: i32) -> (i32, tensor<2x2xi32>, vector<2xi32>, i32) {
   // CHECK: float [[VAL_1:.*]][2][2] = {1.100000e+01, 0.000000e+00, 0.000000e+00, -4.200000e+01};
   %1 = constant dense<[[11.0, 0.0], [0.0, -42.0]]> : tensor<2x2xf32>
 
-  // CHECK: ap_int<1> [[VAL_2:.*]][2][2] = {1, 0, 0, 1};
+  // CHECK: bool [[VAL_2:.*]][2][2] = {1, 0, 0, 1};
   %2 = constant dense<[[1, 0], [0, 1]]> : tensor<2x2xi1>
 
   // CHECK: ap_int<32> [[VAL_3:.*]][2] = {0, -42};
@@ -17,7 +17,7 @@ func @test_constant(%arg0: i32) -> (i32, tensor<2x2xi32>, vector<2xi32>, i32) {
   // CHECK: float [[VAL_4:.*]][2] = {0.000000e+00, -4.200000e+01};
   %4 = constant dense<[0.0, -42.0]> : vector<2xf32>
 
-  // CHECK: ap_int<1> [[VAL_5:.*]][2] = {0, 1};
+  // CHECK: bool [[VAL_5:.*]][2] = {0, 1};
   %5 = constant dense<[0, 1]> : vector<2xi1>
 
   // CHECK: *[[ARG_1:.*]] = 11 + [[ARG_0:.*]];

test/EmitHLSCpp/test_scf_if.mlir

@@ -5,7 +5,7 @@ func @test_scf_if(%arg0: index, %arg1: memref<16xindex>) {
   %c0 = constant 0 : index
 
   // CHECK: int val2 = val0 + 11;
-  // CHECK: ap_int<1> val3 = val2 > 0;
+  // CHECK: bool val3 = val2 > 0;
   // CHECK: int val4;
   // CHECK: int val5[16];
   // CHECK: if (val3) {

test/EmitHLSCpp/test_signature.mlir

@@ -4,7 +4,7 @@
 // CHECK:   float [[VAL_0:.*]],
 // CHECK:   double [[VAL_1:.*]],
 // CHECK:   int [[VAL_2:.*]],
-// CHECK:   ap_int<1> [[VAL_3:.*]],
+// CHECK:   bool [[VAL_3:.*]],
 // CHECK:   ap_int<11> [[VAL_4:.*]],
 // CHECK:   ap_int<32> [[VAL_5:.*]],
 // CHECK:   ap_uint<32> [[VAL_6:.*]],
@@ -17,7 +17,7 @@
 // CHECK:   float *[[VAL_13:.*]],
 // CHECK:   double *[[VAL_14:.*]],
 // CHECK:   int *[[VAL_15:.*]],
-// CHECK:   ap_int<1> *[[VAL_16:.*]],
+// CHECK:   bool *[[VAL_16:.*]],
 // CHECK:   ap_int<11> *[[VAL_17:.*]],
 // CHECK:   ap_int<32> *[[VAL_18:.*]],
 // CHECK:   ap_uint<32> *[[VAL_19:.*]],