[mlir][vector] Add a new TD Op for patterns leveraging ShapeCastOp

[banach-space]

Adds a new Transform Dialect Op that collects patters for dropping unit
dims from various Ops:

• transform.apply_patterns.vector.drop_unit_dims_with_shape_cast.

It excludes patterns for vector.transfer Ops - these are collected
under:

• apply_patterns.vector.rank_reducing_subview_patterns,

and use ShapeCastOp and SubviewOp to reduce the rank (and to eliminate
unit dims).

This new TD Ops allows us to test the "ShapeCast folder" pattern in
isolation. I've extracted the only test that I could find for that
folder from "vector-transforms.mlir" and moved it to a dedicated file:
"shape-cast-folder.mlir". I also added a test case with scalable
vectors.

Changes in VectorTransforms.cpp are not needed (added a comment with
a TODO + ordered the patterns alphabetically). I am Including them here
to avoid a separate PR.

[llvmbot]

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-vector

Author: Andrzej Warzyński (banach-space)

Changes

Adds a new Transform Dialect Op that collects patters for dropping unit
dims from various Ops:

• transform.apply_patterns.vector.drop_unit_dims_with_shape_cast.

It excludes patterns for vector.transfer Ops - these are collected
under:

• apply_patterns.vector.rank_reducing_subview_patterns,

and use ShapeCastOp and SubviewOp to reduce the rank (and to eliminate
unit dims).

This new TD Ops allows us to test the "ShapeCast folder" pattern in
isolation. I've extracted the only test that I could find for that
folder from "vector-transforms.mlir" and moved it to a dedicated file:
"shape-cast-folder.mlir". I also added a test case with scalable
vectors.

---

Full diff: https://github.com/llvm/llvm-project/pull/110525.diff

5 Files Affected:

• (modified) mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td (+16)
• (modified) mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp (+5)
• (modified) mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp (+5-2)
• (added) mlir/test/Dialect/Vector/shape-cast-folder.mlir (+38)
• (modified) mlir/test/Dialect/Vector/vector-transforms.mlir (-9)

diff --git a/mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td b/mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td
index aad2eab83dbd38..c973eca0132a92 100644
--- a/mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td
+++ b/mlir/include/mlir/Dialect/Vector/TransformOps/VectorTransformOps.td
@@ -68,6 +68,22 @@ def ApplyRankReducingSubviewPatternsOp : Op<Transform_Dialect,
   let assemblyFormat = "attr-dict";
 }
 
+def ApplyDropUnitDimWithShapeCastPatternsOp : Op<Transform_Dialect,
+    "apply_patterns.vector.drop_unit_dims_with_shape_cast",
+    [DeclareOpInterfaceMethods<PatternDescriptorOpInterface>]> {
+  let description = [{
+     Apply vector patterns to fold unit dims with vector.shape_cast Ops:
+      - DropUnitDimFromElementwiseOps
+      - DropUnitDimsFromScfForOp
+      - DropUnitDimsFromTransposeOp
+
+    Excludes patterns for vector.transfer Ops. This is complemented by
+    shape_cast folding patterns.
+  }];
+
+  let assemblyFormat = "attr-dict";
+}
+
 def ApplyTransferPermutationPatternsOp : Op<Transform_Dialect,
     "apply_patterns.vector.transfer_permutation_patterns",
     [DeclareOpInterfaceMethods<PatternDescriptorOpInterface>]> {
diff --git a/mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp b/mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp
index 3ae70ace3934cd..241e83e234d621 100644
--- a/mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp
+++ b/mlir/lib/Dialect/Vector/TransformOps/VectorTransformOps.cpp
@@ -85,6 +85,11 @@ void transform::ApplyTransferPermutationPatternsOp::populatePatterns(
   vector::populateVectorTransferPermutationMapLoweringPatterns(patterns);
 }
 
+void transform::ApplyDropUnitDimWithShapeCastPatternsOp::populatePatterns(
+    RewritePatternSet &patterns) {
+  vector::populateDropUnitDimWithShapeCastPatterns(patterns);
+}
+
 void transform::ApplyLowerBitCastPatternsOp::populatePatterns(
     RewritePatternSet &patterns) {
   vector::populateVectorBitCastLoweringPatterns(patterns);
diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
index 8fcef54f12edfa..09b6bfd0a15e3b 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
@@ -2056,8 +2056,11 @@ void mlir::vector::populateShapeCastFoldingPatterns(RewritePatternSet &patterns,
 
 void mlir::vector::populateDropUnitDimWithShapeCastPatterns(
     RewritePatternSet &patterns, PatternBenefit benefit) {
-  patterns.add<DropUnitDimFromElementwiseOps, DropUnitDimsFromTransposeOp,
-               ShapeCastOpFolder, DropUnitDimsFromScfForOp>(
+  // TODO: Consider either including DropInnerMostUnitDimsTransferRead,
+  // DropInnerMostUnitDimsTransferWrite or better naming to distinguish this
+  // and populateVectorTransferCollapseInnerMostContiguousDimsPatterns.
+  patterns.add<DropUnitDimFromElementwiseOps, DropUnitDimsFromScfForOp,
+               DropUnitDimsFromTransposeOp, ShapeCastOpFolder>(
       patterns.getContext(), benefit);
 }
 
diff --git a/mlir/test/Dialect/Vector/shape-cast-folder.mlir b/mlir/test/Dialect/Vector/shape-cast-folder.mlir
new file mode 100644
index 00000000000000..9550c5c4ae0561
--- /dev/null
+++ b/mlir/test/Dialect/Vector/shape-cast-folder.mlir
@@ -0,0 +1,38 @@
+// RUN: mlir-opt %s --transform-interpreter --split-input-file | FileCheck %s
+
+///----------------------------------------------------------------------------------------
+/// [Pattern: ShapeCastOpFolder]
+///----------------------------------------------------------------------------------------
+
+// CHECK-LABEL: func @fixed_width
+//  CHECK-SAME: %[[A0:.*0]]: vector<2x4xf32>
+//   CHECK-NOT: vector.shape_cast
+//       CHECK: return %[[A0]] : vector<2x4xf32>
+func.func @fixed_width(%arg0 : vector<2x4xf32>) -> vector<2x4xf32> {
+  %0 = vector.shape_cast %arg0 : vector<2x4xf32> to vector<8xf32>
+  %1 = vector.shape_cast %0 : vector<8xf32> to vector<2x4xf32>
+  return %1 : vector<2x4xf32>
+}
+
+// CHECK-LABEL: func @scalable
+//  CHECK-SAME: %[[A0:.*0]]: vector<2x[4]xf32>
+//   CHECK-NOT: vector.shape_cast
+//       CHECK: return %[[A0]] : vector<2x[4]xf32>
+func.func @scalable(%arg0 : vector<2x[4]xf32>) -> vector<2x[4]xf32> {
+  %0 = vector.shape_cast %arg0 : vector<2x[4]xf32> to vector<[8]xf32>
+  %1 = vector.shape_cast %0 : vector<[8]xf32> to vector<2x[4]xf32>
+  return %1 : vector<2x[4]xf32>
+}
+
+// ============================================================================
+//  TD sequence
+// ============================================================================
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {
+    %func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">
+    transform.apply_patterns to %func_op {
+      transform.apply_patterns.vector.drop_unit_dims_with_shape_cast
+    } : !transform.op<"func.func">
+    transform.yield
+  }
+}
diff --git a/mlir/test/Dialect/Vector/vector-transforms.mlir b/mlir/test/Dialect/Vector/vector-transforms.mlir
index eda6a5cc40d999..89e8ca1d93109c 100644
--- a/mlir/test/Dialect/Vector/vector-transforms.mlir
+++ b/mlir/test/Dialect/Vector/vector-transforms.mlir
@@ -184,15 +184,6 @@ func.func @vector_transfers(%arg0: index, %arg1: index) {
   return
 }
 
-// CHECK-LABEL: func @cancelling_shape_cast_ops
-//  CHECK-SAME: %[[A0:.*0]]: vector<2x4xf32>
-//       CHECK: return %[[A0]] : vector<2x4xf32>
-func.func @cancelling_shape_cast_ops(%arg0 : vector<2x4xf32>) -> vector<2x4xf32> {
-  %0 = vector.shape_cast %arg0 : vector<2x4xf32> to vector<8xf32>
-  %1 = vector.shape_cast %0 : vector<8xf32> to vector<2x4xf32>
-  return %1 : vector<2x4xf32>
-}
-
 // CHECK-LABEL: func @elementwise_unroll
 //  CHECK-SAME: (%[[ARG0:.*]]: memref<4x4xf32>, %[[ARG1:.*]]: memref<4x4xf32>)
 //       CHECK-DAG:   %[[C2:.*]] = arith.constant 2 : index

[hanhanW]

LGTM, just one question.

[hanhanW]

I'm not a big fan of adding the patterns to the documentation because it could be outdated quickly. It is fine to me because other operations are following the convention...

[banach-space]

I also struggle with this a bit.

I do like that it helps to find duplicates. Also, makes the overall description a bit easier and more concise (provided that these patterns have good descriptive names).

[hanhanW]

Should the comment belong to transform::ApplyDropUnitDimWithShapeCastPatternsOp::populatePatterns?

[banach-space]

It's hard to tell - this file is where e.g. DropInnerMostUnitDimsTransferWrite and DropUnitDimFromElementwiseOps are actually added 😅 But then this is just the implementation and the interfaces is elsewhere.

Let me keep it here - I intend to review the corresponding tests in the coming days and will get rid of this. Also, changes in this file are NFCs and not required (added this comment + ordered the patterns alphabetically). I'll add a note in the summary.