[Mlir-commits] [mlir] ead535b - [mlir][tensor] Add producer fusion for tensor.unpack op.

[Hanhan Wang]

Author: Hanhan Wang
Date: 2023-01-06T14:13:11-08:00
New Revision: ead535b2f9ef3a0984884fdd088327cff7da512f

URL: https://github.com/llvm/llvm-project/commit/ead535b2f9ef3a0984884fdd088327cff7da512f
DIFF: https://github.com/llvm/llvm-project/commit/ead535b2f9ef3a0984884fdd088327cff7da512f.diff

LOG: [mlir][tensor] Add producer fusion for tensor.unpack op.

Reviewed By: mravishankar

Differential Revision: https://reviews.llvm.org/D141151

Added: 
    

Modified: 
    mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp
    mlir/test/Dialect/Linalg/transform-op-fuse.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp b/mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp
index 46324672ad2f2..7eced90aac548 100644
--- a/mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp
@@ -425,6 +425,15 @@ struct UnPackOpTiling
     resultSizes = llvm::to_vector(sizes);
     return success();
   }
+
+  FailureOr<Value> generateResultTileValue(Operation *op, OpBuilder &b,
+                                           unsigned resultNumber,
+                                           ArrayRef<OpFoldResult> offsets,
+                                           ArrayRef<OpFoldResult> sizes) const {
+    return getTiledImplementation(op, b, offsets, sizes)
+        .back()
+        ->getResult(resultNumber);
+  }
 };
 
 } // namespace

diff  --git a/mlir/test/Dialect/Linalg/transform-op-fuse.mlir b/mlir/test/Dialect/Linalg/transform-op-fuse.mlir
index 7cbda2d1c85d7..580ad597ef30d 100644
--- a/mlir/test/Dialect/Linalg/transform-op-fuse.mlir
+++ b/mlir/test/Dialect/Linalg/transform-op-fuse.mlir
@@ -91,3 +91,26 @@ transform.sequence failures(propagate) {
   %1, %loops:2 = transform.structured.fuse %0 {tile_sizes = [5, 0, 7], tile_interchange = [0, 2, 1]}
   %2, %loops_2 = transform.structured.tile %1 [0, 4]
 }
+
+// -----
+
+// CHECK-LABEL: func.func @unpack_elemwise
+// CHECK:         %[[RES:.*]] = scf.for
+// CHECK:           scf.for
+// CHECK:             tensor.unpack
+// CHECK:             linalg.elemwise_unary
+// CHECK:         return %[[RES]]
+func.func @unpack_elemwise(%arg0: tensor<16x48x8x8xf32>, %arg1: tensor<128x384xf32>) -> tensor<128x384xf32> {
+  %0 = tensor.empty() : tensor<128x384xf32>
+  %1 = tensor.unpack %arg0 inner_dims_pos = [0, 1] inner_tiles = [8, 8] into %0
+      : tensor<16x48x8x8xf32> -> tensor<128x384xf32>
+  %2 = linalg.elemwise_unary ins(%1: tensor<128x384xf32>)
+                             outs(%arg1: tensor<128x384xf32>) -> tensor<128x384xf32>
+  return %2 : tensor<128x384xf32>
+}
+
+transform.sequence failures(propagate) {
+^bb1(%arg1: !pdl.operation):
+  %0 = transform.structured.match ops{["linalg.elemwise_unary"]} in %arg1
+  %1, %loops:2 = transform.structured.fuse %0 {tile_sizes = [16, 32], tile_interchange = [0, 1]}
+}