Adding a missing op for ParallelCombiningOpInterface

Author: lialan

mlir/include/mlir/Dialect/SCF/IR/SCF.h

@@ -58,7 +58,7 @@ ForallOp getForallOpThreadIndexOwner(Value val);
 bool insideMutuallyExclusiveBranches(Operation *a, Operation *b);
 
 /// Promotes the loop body of a scf::ForallOp to its containing block.
-void promote(RewriterBase &rewriter, scf::ForallOp forallOp);
+LogicalResult promote(RewriterBase &rewriter, scf::ForallOp forallOp);
 
 /// An owning vector of values, handy to return from functions.
 using ValueVector = SmallVector<Value>;

mlir/include/mlir/Dialect/Tensor/IR/TensorOps.td

@@ -1474,7 +1474,8 @@ def Tensor_ParallelInsertSliceOp : Tensor_Op<"parallel_insert_slice", [
        AttrSizedOperandSegments,
        OffsetSizeAndStrideOpInterface,
        DeclareOpInterfaceMethods<ParallelCombiningOpInterface,
-          ["getUpdatedDestinations", "getIteratingParent"]>,
+          ["getUpdatedDestinations", "getIteratingParent",
+           "promoteInParallelLoop"]>,
        // TODO: Cannot use an interface here atm, verify this manually for now.
        // HasParent<"InParallelOpInterface">
   ]> {

mlir/include/mlir/Interfaces/ParallelCombiningOpInterface.h

@@ -15,6 +15,9 @@
 #define MLIR_INTERFACES_PARALLELCOMBININGOPINTERFACE_H_
 
 #include "mlir/IR/OpDefinition.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Support/LogicalResult.h"
+#include "llvm/ADT/SmallVector.h"
 
 namespace mlir {
 namespace detail {

mlir/include/mlir/Interfaces/ParallelCombiningOpInterface.td

@@ -106,6 +106,26 @@ def ParallelCombiningOpInterface : OpInterface<"ParallelCombiningOpInterface"> {
       /*methodName=*/"getIteratingParent",
       /*args=*/(ins)
     >,
+    InterfaceMethod<
+      /*desc=*/[{
+        Promotes this parallel combining op out of its enclosing parallel loop
+        and returns the values that should replace the destinations updated by
+        this op.
+      }],
+      /*retTy=*/"::mlir::FailureOr<::llvm::SmallVector<::mlir::Value>>",
+      /*methodName=*/"promoteInParallelLoop",
+      /*args=*/(ins "::mlir::RewriterBase &":$rewriter)
+    >,
+    InterfaceMethod<
+      /*desc=*/[{
+        Returns true if this op can be promoted out of its enclosing parallel
+        loop.
+      }],
+      /*retTy=*/"bool",
+      /*methodName=*/"canPromoteInParallelLoop",
+      /*args=*/(ins "::mlir::RewriterBase &":$rewriter),
+      /*methodBody=*/[{ return true; }]
+    >,
   ];
 }
 

mlir/lib/Dialect/SCF/IR/SCF.cpp

@@ -651,8 +651,7 @@ LogicalResult scf::ForallOp::promoteIfSingleIteration(RewriterBase &rewriter) {
       return failure();
   }
 
-  promote(rewriter, *this);
-  return success();
+  return promote(rewriter, *this);
 }
 
 Block::BlockArgListType ForallOp::getRegionIterArgs() {
@@ -664,10 +663,23 @@ MutableArrayRef<OpOperand> ForallOp::getInitsMutable() {
 }
 
 /// Promotes the loop body of a scf::ForallOp to its containing block.
-void mlir::scf::promote(RewriterBase &rewriter, scf::ForallOp forallOp) {
+LogicalResult mlir::scf::promote(RewriterBase &rewriter, scf::ForallOp forallOp) {
   OpBuilder::InsertionGuard g(rewriter);
   scf::InParallelOp terminator = forallOp.getTerminator();
 
+  // Make sure we can promote all parallel combining ops in terminator:
+  for (auto &yieldingOp : terminator.getYieldingOps()) {
+    auto parallelCombiningOp =
+        dyn_cast<ParallelCombiningOpInterface>(&yieldingOp);
+    if (!parallelCombiningOp)
+      return rewriter.notifyMatchFailure(
+          forallOp, "terminator has non-parallel-combining op");
+    if (!parallelCombiningOp.canPromoteInParallelLoop(rewriter))
+      return rewriter.notifyMatchFailure(
+          forallOp, "parallel combining op cannot be promoted");
+  }
+
+
   // Replace block arguments with lower bounds (replacements for IVs) and
   // outputs.
   SmallVector<Value> bbArgReplacements = forallOp.getLowerBound(rewriter);
@@ -683,30 +695,29 @@ void mlir::scf::promote(RewriterBase &rewriter, scf::ForallOp forallOp) {
   SmallVector<Value> results;
   results.reserve(forallOp.getResults().size());
   for (auto &yieldingOp : terminator.getYieldingOps()) {
-    auto parallelInsertSliceOp =
-        dyn_cast<tensor::ParallelInsertSliceOp>(yieldingOp);
-    if (!parallelInsertSliceOp)
+    auto parallelCombiningOp =
+        dyn_cast<ParallelCombiningOpInterface>(&yieldingOp);
+    if (!parallelCombiningOp)
       continue;
 
-    Value dst = parallelInsertSliceOp.getDest();
-    Value src = parallelInsertSliceOp.getSource();
-    if (llvm::isa<TensorType>(src.getType())) {
-      results.push_back(tensor::InsertSliceOp::create(
-          rewriter, forallOp.getLoc(), dst.getType(), src, dst,
-          parallelInsertSliceOp.getOffsets(), parallelInsertSliceOp.getSizes(),
-          parallelInsertSliceOp.getStrides(),
-          parallelInsertSliceOp.getStaticOffsets(),
-          parallelInsertSliceOp.getStaticSizes(),
-          parallelInsertSliceOp.getStaticStrides()));
-    } else {
-      llvm_unreachable("unsupported terminator");
-    }
+    assert(parallelCombiningOp.canPromoteInParallelLoop(rewriter));
+
+    FailureOr<SmallVector<Value>> promotedValues =
+        parallelCombiningOp.promoteInParallelLoop(rewriter);
+    if (failed(promotedValues))
+      return failure();
+
+    results.append(promotedValues->begin(), promotedValues->end());
   }
+  if (results.size() != forallOp.getResults().size())
+    return rewriter.notifyMatchFailure(
+        forallOp, "failed to materialize replacements for all results");
   rewriter.replaceAllUsesWith(forallOp.getResults(), results);
 
   // Erase the old terminator and the loop.
   rewriter.eraseOp(terminator);
   rewriter.eraseOp(forallOp);
+  return success();
 }
 
 LoopNest mlir::scf::buildLoopNest(
@@ -1789,7 +1800,8 @@ struct ForallOpSingleOrZeroIterationDimsFolder
 
     // All of the loop dimensions perform a single iteration. Inline loop body.
     if (newMixedLowerBounds.empty()) {
-      promote(rewriter, op);
+      if (failed(promote(rewriter, op)))
+        return failure();
       return success();
     }
 

mlir/lib/Dialect/Tensor/IR/TensorOps.cpp

@@ -3947,6 +3947,21 @@ Operation *ParallelInsertSliceOp::getIteratingParent() {
   return nullptr;
 }
 
+FailureOr<SmallVector<Value>>
+ParallelInsertSliceOp::promoteInParallelLoop(RewriterBase &rewriter) {
+  Value dst = getDest();
+  Value src = getSource();
+  if (!isa<TensorType>(src.getType()))
+    return rewriter.notifyMatchFailure(getOperation(),
+                                       "expected tensor source");
+
+  Value inserted = tensor::InsertSliceOp::create(
+      rewriter, getLoc(), dst.getType(), src, dst, getOffsets(), getSizes(),
+      getStrides(), getStaticOffsets(), getStaticSizes(), getStaticStrides());
+
+  return SmallVector<Value>{inserted};
+}
+
 //===----------------------------------------------------------------------===//
 // ScatterOp
 //===----------------------------------------------------------------------===//