[Mlir-commits] [mlir] [mlir] Implement DestinationStyleOpInterface for scf::ForallOp (PR #66981)

[llvmlistbot at llvm.org]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir-scf

<details>
<summary>Changes</summary>

`scf::ForallOp` has `shared_outs` tensor operands which are used to insert partial results into in the parallel terminator. The `scf::ForallOp` returns one tensor for each `shared_out` which then contains the combined result from all threads. Since the parallel terminator cannot change the shape of the `shared_out`, ForallOp is a `DestinationStyleOp` and this patch implements the interface by declaring the `outputs` operands as `inits` in the language of the DPS interface.

For this change to work, we need to add an exception to the Pattern that folds `tensor.cast` Ops into DPS Ops because `scf::Forall` needs special handling of it's `BlockArgument` Type during this folding.

---
Full diff: https://github.com/llvm/llvm-project/pull/66981.diff


3 Files Affected:

- (modified) mlir/include/mlir/Dialect/SCF/IR/SCF.h (+1) 
- (modified) mlir/include/mlir/Dialect/SCF/IR/SCFOps.td (+10) 
- (modified) mlir/lib/Dialect/Tensor/IR/TensorOps.cpp (+4) 


``````````diff
diff --git a/mlir/include/mlir/Dialect/SCF/IR/SCF.h b/mlir/include/mlir/Dialect/SCF/IR/SCF.h
index 915ab3016b688e7..644118ca884c6b1 100644
--- a/mlir/include/mlir/Dialect/SCF/IR/SCF.h
+++ b/mlir/include/mlir/Dialect/SCF/IR/SCF.h
@@ -19,6 +19,7 @@
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/RegionKindInterface.h"
 #include "mlir/Interfaces/ControlFlowInterfaces.h"
+#include "mlir/Interfaces/DestinationStyleOpInterface.h"
 #include "mlir/Interfaces/InferTypeOpInterface.h"
 #include "mlir/Interfaces/LoopLikeInterface.h"
 #include "mlir/Interfaces/ParallelCombiningOpInterface.h"
diff --git a/mlir/include/mlir/Dialect/SCF/IR/SCFOps.td b/mlir/include/mlir/Dialect/SCF/IR/SCFOps.td
index 08b71e20a2bc079..adc7b2e4170cb89 100644
--- a/mlir/include/mlir/Dialect/SCF/IR/SCFOps.td
+++ b/mlir/include/mlir/Dialect/SCF/IR/SCFOps.td
@@ -17,6 +17,7 @@ include "mlir/Interfaces/ControlFlowInterfaces.td"
 include "mlir/Interfaces/LoopLikeInterface.td"
 include "mlir/IR/RegionKindInterface.td"
 include "mlir/Dialect/SCF/IR/DeviceMappingInterface.td"
+include "mlir/Interfaces/DestinationStyleOpInterface.td"
 include "mlir/Interfaces/InferTypeOpInterface.td"
 include "mlir/Interfaces/ParallelCombiningOpInterface.td"
 include "mlir/Interfaces/SideEffectInterfaces.td"
@@ -333,6 +334,7 @@ def ForallOp : SCF_Op<"forall", [
        RecursiveMemoryEffects,
        SingleBlockImplicitTerminator<"scf::InParallelOp">,
        DeclareOpInterfaceMethods<RegionBranchOpInterface>,
+       DestinationStyleOpInterface
      ]> {
   let summary = "evaluate a block multiple times in parallel";
   let description = [{
@@ -630,6 +632,14 @@ def ForallOp : SCF_Op<"forall", [
                                  Location loc);
 
     InParallelOp getTerminator();
+
+    // Implement this to declare all shared_outs as inits/outs to 
+    // DestinationStyleOpInterface
+    std::pair<int64_t, int64_t> getDpsInitsPositionRange() {
+      int64_t numOperands = getNumOperands();
+      int64_t numOuts = getOutputs().size();
+      return {numOperands - numOuts, numOperands};
+    }
   }];
 }
 
diff --git a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
index 3e30e320bee8f83..fa91471f33d4bd3 100644
--- a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
+++ b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
@@ -10,6 +10,7 @@
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Arith/Utils/Utils.h"
 #include "mlir/Dialect/Complex/IR/Complex.h"
+#include "mlir/Dialect/SCF/IR/SCF.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/Dialect/Utils/IndexingUtils.h"
 #include "mlir/Dialect/Utils/ReshapeOpsUtils.h"
@@ -3970,6 +3971,9 @@ struct FoldTensorCastProducerOp
     if (isa<InsertSliceOp>(op.getOperation()))
       return failure();
 
+    if (isa<scf::ForallOp>(op.getOperation()))
+      return failure();
+
     // If no operand comes from a tensor::CastOp and can be folded then fail.
     bool hasTensorCastOperand =
         llvm::any_of(op->getOpOperands(), [&](OpOperand &opOperand) {

``````````

</details>


https://github.com/llvm/llvm-project/pull/66981