[Mlir-commits] [mlir] [vector][mlir] Restrict vector.shuffle to fixed-width vectors (PR #88733)

Andrzej WarzyƄski llvmlistbot at llvm.org
Mon Apr 15 09:36:57 PDT 2024


https://github.com/banach-space updated https://github.com/llvm/llvm-project/pull/88733

>From 61133a792ba781b4536250ba3468bb6026761524 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Mon, 15 Apr 2024 14:28:06 +0100
Subject: [PATCH 1/2] [vector][mlir] Restrict vector.shuffle to fixed-width
 vectors

At the moment there is no support for vector.shuffle for scalable
vectors - various hooks/helpers related to `vector.shuffle` simply
ignore the scalable flags (e.g. ` ShuffleOp::inferReturnTypes`).

This is unlikely to change any time soon (vector shuffles are known to
be tricky for scalable vectors), hence this patch restricts
`vector.shuffle` to fixed width vectors.
---
 mlir/include/mlir/Dialect/Vector/IR/VectorOps.td | 4 +++-
 mlir/test/Dialect/Vector/invalid.mlir            | 7 +++++++
 2 files changed, 10 insertions(+), 1 deletion(-)

diff --git a/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td b/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
index 147bc2354977d7..de6617c2062b0c 100644
--- a/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
+++ b/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
@@ -420,7 +420,7 @@ def Vector_ShuffleOp :
      PredOpTrait<"second operand v2 and result have same element type",
                  TCresVTEtIsSameAsOpBase<0, 1>>,
      InferTypeOpAdaptor]>,
-     Arguments<(ins AnyVectorOfAnyRank:$v1, AnyVectorOfAnyRank:$v2,
+     Arguments<(ins AnyFixedVector:$v1, AnyFixedVector:$v2,
                     I64ArrayAttr:$mask)>,
      Results<(outs AnyVector:$vector)> {
   let summary = "shuffle operation";
@@ -444,6 +444,8 @@ def Vector_ShuffleOp :
       mask values must be within range, viz. given two k-D operands v1 and v2
       above, all mask values are in the range [0,s_1+t_1)
 
+    Note, at the moment only fixed-width vectors are supported.
+
     Example:
 
     ```mlir
diff --git a/mlir/test/Dialect/Vector/invalid.mlir b/mlir/test/Dialect/Vector/invalid.mlir
index c16f1cb2876dbd..c9f7e9c6e2fb0b 100644
--- a/mlir/test/Dialect/Vector/invalid.mlir
+++ b/mlir/test/Dialect/Vector/invalid.mlir
@@ -84,6 +84,13 @@ func.func @shuffle_index_out_of_range(%arg0: vector<2xf32>, %arg1: vector<2xf32>
 
 // -----
 
+func.func @shuffle_scalable_vec(%arg0: vector<[2]xf32>, %arg1: vector<[2]xf32>) {
+  // expected-error at +1 {{'vector.shuffle' op operand #0 must be fixed-length vector of any type values}}
+  %1 = vector.shuffle %arg0, %arg1 [0, 1, 2, 3] : vector<[2]xf32>, vector<[2]xf32>
+}
+
+// -----
+
 func.func @shuffle_empty_mask(%arg0: vector<2xf32>, %arg1: vector<2xf32>) {
   // expected-error at +1 {{'vector.shuffle' op invalid mask length}}
   %1 = vector.shuffle %arg0, %arg1 [] : vector<2xf32>, vector<2xf32>

>From 505624d44680c8986ea4188f2a1807eda8967375 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Mon, 15 Apr 2024 17:36:31 +0100
Subject: [PATCH 2/2] fixup! [vector][mlir] Restrict vector.shuffle to
 fixed-width vectors

Update comment
---
 mlir/include/mlir/Dialect/Vector/IR/VectorOps.td | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td b/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
index de6617c2062b0c..332b5ad08ced98 100644
--- a/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
+++ b/mlir/include/mlir/Dialect/Vector/IR/VectorOps.td
@@ -444,7 +444,7 @@ def Vector_ShuffleOp :
       mask values must be within range, viz. given two k-D operands v1 and v2
       above, all mask values are in the range [0,s_1+t_1)
 
-    Note, at the moment only fixed-width vectors are supported.
+    Note, scalable vectors are not supported.
 
     Example:
 



More information about the Mlir-commits mailing list