[Mlir-commits] [mlir] [mlir][ArmSME] Add support for vector.transpose (PR #66760)

[Cullen Rhodes]

================
@@ -241,11 +242,84 @@ struct BroadcastOpToArmSMELowering
   }
 };
 
+/// Conversion pattern for vector.transpose.
+///
+/// Stores the input tile to memory and reloads vertically.
+///
+/// Example:
+///
+///   %transposed_src = vector.transpose %src, [1, 0]
+///     : vector<[4]x[4]xi32> to vector<[4]x[4]xi32>
+///
+/// is converted to:
+///
+///   %alloca = memref.alloca(%svl_s, %svl_s) : memref<?x?xi32>
+///   %arm_sme.tile_store %src, <hor>, %alloca[%c0, %c0]
+///     : memref<?x?xi32>, vector<[4]x[4]xi32>
+///   %transposed_src = arm_sme.tile_load <ver>, %alloca[%c0, %c0]
+///     : memref<?x?xi32>, vector<[4]x[4]xi32>
+///
+/// NOTE: Tranposing via memory is obviously expensive, the current intention
+/// is to avoid the transpose if possible, this is therefore intended as a
+/// fallback and to provide base support for Vector ops. If it turns out
+/// transposes can't be avoided then this should be replaced with a more optimal
+/// implementation, perhaps with tile <-> vector (MOVA) ops.
+struct TransposeOpToArmSMELowering
+    : public OpRewritePattern<vector::TransposeOp> {
+  using OpRewritePattern<vector::TransposeOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(vector::TransposeOp transposeOp,
+                                PatternRewriter &rewriter) const final {
+    auto tileType = transposeOp.getResultVectorType();
+    if (!tileType || !arm_sme::isValidSMETileVectorType(tileType))
+      return failure();
+
+    SmallVector<int64_t> transp;
+    for (auto attr : transposeOp.getTransp())
+      transp.push_back(cast<IntegerAttr>(attr).getInt());
+
+    if (transp[0] != 1 && transp[1] != 0)
+      return failure();
+
+    OpBuilder::InsertionGuard g(rewriter);
----------------
c-rhodes wrote:

Added a comment

https://github.com/llvm/llvm-project/pull/66760