[Mlir-commits] [mlir] [mlir][Bufferization] Do not have read semantics for destination of `tensor.parallel_insert_slice`. (PR #134169)

[llvmlistbot at llvm.org]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir-tensor

Author: None (MaheshRavishankar)

<details>
<summary>Changes</summary>

`tensor.insert_slice` needs to have read semantics on its destination operand. Since it has a return value, its semantics are

- Copy dest to result
- Copy source to subview of destination.

`tensor.parallel_insert_slice` though has no result. So it does not need to have read semantics. The op description [here](https://github.com/llvm/llvm-project/blob/a3ac318e5f8668ec5b79dd86639881dfb2e88b69/mlir/include/mlir/Dialect/Tensor/IR/TensorOps.td#L1524) also says that it is expected to lower to a `memref.subview`, that does not have read semantics on the destination (its just a view).

This patch drops the read semantics for destination of `tensor.parallel_insert_slice` but also makes the `shared_outs` operands of `scf.forall` have read semantics. Earlier it would rely indirectly on read semantics of destination operand of `tensor.parallel_insert_slice` to propagate the read semantics for `shared_outs`. Now that is specified more directly.

---
Full diff: https://github.com/llvm/llvm-project/pull/134169.diff


3 Files Affected:

- (modified) mlir/lib/Dialect/SCF/Transforms/BufferizableOpInterfaceImpl.cpp (+5-11) 
- (modified) mlir/lib/Dialect/Tensor/Transforms/BufferizableOpInterfaceImpl.cpp (+1-2) 
- (modified) mlir/test/Dialect/SCF/one-shot-bufferize.mlir (+35) 


``````````diff
diff --git a/mlir/lib/Dialect/SCF/Transforms/BufferizableOpInterfaceImpl.cpp b/mlir/lib/Dialect/SCF/Transforms/BufferizableOpInterfaceImpl.cpp
index f48d2a2df9c3c..f83435521e708 100644
--- a/mlir/lib/Dialect/SCF/Transforms/BufferizableOpInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/SCF/Transforms/BufferizableOpInterfaceImpl.cpp
@@ -1207,17 +1207,11 @@ struct ForallOpInterface
                                                     ForallOp> {
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               const AnalysisState &state) const {
-    auto forallOp = cast<ForallOp>(op);
-
-    // If the loop has zero iterations, the results of the op are their
-    // corresponding shared_outs, meaning that the shared_outs bufferize to a
-    // read.
-    if (mayHaveZeroIterations(forallOp))
-      return true;
-
-    // scf::ForallOp alone doesn't bufferize to a memory read, one of the
-    // uses of its matching bbArg may.
-    return state.isValueRead(forallOp.getTiedBlockArgument(&opOperand));
+    // All tensor operands to `scf.forall` are `shared_outs` and all
+    // shared outs are assumed to be read by the loop. This does not
+    // account for the case where the entire value is over-written,
+    // but being conservative here.
+    return true;
   }
 
   bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
diff --git a/mlir/lib/Dialect/Tensor/Transforms/BufferizableOpInterfaceImpl.cpp b/mlir/lib/Dialect/Tensor/Transforms/BufferizableOpInterfaceImpl.cpp
index 4ac6eca586961..31014172a9555 100644
--- a/mlir/lib/Dialect/Tensor/Transforms/BufferizableOpInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/Tensor/Transforms/BufferizableOpInterfaceImpl.cpp
@@ -930,8 +930,7 @@ struct ParallelInsertSliceOpInterface
 
   bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
                               const AnalysisState &state) const {
-    return insertSliceOpRequiresRead(cast<tensor::ParallelInsertSliceOp>(op),
-                                     opOperand);
+    return opOperand == cast<ParallelInsertSliceOp>(op).getSourceMutable();
   }
 
   bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
diff --git a/mlir/test/Dialect/SCF/one-shot-bufferize.mlir b/mlir/test/Dialect/SCF/one-shot-bufferize.mlir
index bb9f7dfdba83f..a1067ec3ba05f 100644
--- a/mlir/test/Dialect/SCF/one-shot-bufferize.mlir
+++ b/mlir/test/Dialect/SCF/one-shot-bufferize.mlir
@@ -946,3 +946,38 @@ func.func @index_switch(%pred: index, %b: tensor<5xf32>, %c: tensor<5xf32>) -> t
   // CHECK: return %[[r]]
   return %0 : tensor<5xf32>
 }
+
+// -----
+
+// See Issue https://github.com/llvm/llvm-project/issues/133964 . Checks that
+// tensor.parallel_insert_slice dest operand does not have read semantics.
+func.func @check_scfforall_inplace_bufferizer(%arg0 : tensor<?x?xf32>,
+    %arg1 : tensor<?x?xf32>,
+    %arg2 : tensor<?xf32> {bufferization.writable = true}) ->  tensor<?xf32> {
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  %d0 = tensor.dim %arg2, %c0 : tensor<?xf32>
+  %d1 = tensor.dim %arg1, %c1 : tensor<?x?xf32>
+  %0 = scf.forall (%arg3) in (%c1) shared_outs(%arg4 = %arg2) -> (tensor<?xf32>) {
+    %1 = tensor.extract_slice %arg0[0, 0][%d0, %d1][1, 1] : tensor<?x?xf32> to tensor<?x?xf32>
+    %2 = tensor.extract_slice %arg1[0, 0][%d0, %d1][1, 1] : tensor<?x?xf32> to tensor<?x?xf32>
+    %3 = linalg.generic {
+        indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
+                         affine_map<(d0, d1) -> (d0, d1)>,
+                         affine_map<(d0, d1) -> (d0)>],
+        iterator_types = ["parallel", "reduction"]}
+        ins(%1, %2 : tensor<?x?xf32>, tensor<?x?xf32>)
+        outs(%arg4 : tensor<?xf32>) {
+      ^bb0(%b0 : f32, %b1: f32, %b2 : f32):
+        %4 = arith.mulf %b0, %b1 : f32
+        %5 = arith.addf %4, %b2 : f32
+        linalg.yield %5 : f32
+    } -> tensor<?xf32>
+    scf.forall.in_parallel {
+      tensor.parallel_insert_slice %3 into %arg4[0] [%d0] [1] : tensor<?xf32> into tensor<?xf32>
+    }
+  }
+  return %0 : tensor<?xf32>
+}
+// CHECK-LABEL: func @check_scfforall_inplace_bufferizer
+//   CHECK-NOT:   memref.alloc

``````````

</details>


https://github.com/llvm/llvm-project/pull/134169