[Mlir-commits] [mlir] [MLIR][XeGPU] Decompose unsupported 'vector.transfer_read'-transpose-permutations (PR #182875)

[Jianhui Li]

================
@@ -577,39 +578,64 @@ struct TransferReadLowering : public OpRewritePattern<vector::TransferReadOp> {
     AffineMap readMap = readOp.getPermutationMap();
     bool isTransposeLoad = !readMap.isMinorIdentity();
 
-    Type elementType = vecTy.getElementType();
+    Type elementType = loadedVecTy.getElementType();
     unsigned minTransposeBitWidth = 32;
-    if (isTransposeLoad &&
-        elementType.getIntOrFloatBitWidth() < minTransposeBitWidth)
-      return rewriter.notifyMatchFailure(
-          readOp, "Unsupported data type for transposition");
-
-    // If load is transposed, get the base shape for the tensor descriptor.
-    SmallVector<int64_t> descShape(vecTy.getShape());
-    if (isTransposeLoad)
-      std::reverse(descShape.begin(), descShape.end());
+    // Here we separate two transpose cases:
+    // 1. With transpose attribute in xegpu.load_nd for larger element types
+    // 2. With separate vector.transpose after load_nd for smaller element types
+    bool shouldUseTransposeAttr =
+        isTransposeLoad &&
+        elementType.getIntOrFloatBitWidth() >= minTransposeBitWidth;
----------------
Jianhui-Li wrote:

The condition here "elementType.getIntOrFloatBitWidth() >= minTransposeBitWidth" guarantee that load_nd with TransposeAttr can be succesfully lowered to hardware 2d block load instruction with transpose.  

It doens't mean "should". 

https://github.com/llvm/llvm-project/pull/182875