[Mlir-commits] [mlir] [mlir][Vector] Add vector bitwidth target to xfer op flattening (PR #81966)

[Diego Caballero]

================
@@ -554,6 +561,12 @@ class FlattenContiguousRowMajorTransferReadPattern
     // If this is already 0D/1D, there's nothing to do.
     if (vectorType.getRank() <= 1)
       return failure();
+    if (!vectorType.getElementType().isSignlessIntOrFloat())
+      return failure();
+    unsigned trailingVectorDimBitwidth =
+        vectorType.getShape().back() * vectorType.getElementTypeBitWidth();
+    if (trailingVectorDimBitwidth >= targetVectorBitwidth)
+      return failure();
     if (!vector::isContiguousSlice(sourceType, vectorType))
----------------
dcaballe wrote:

Yes, that's the ultimate goal: partial and more targeted flattening but one step at a time. We first have to flatten the producer/consumers of these xfer ops to make sure we don't generate ops to reshape the vector.

https://github.com/llvm/llvm-project/pull/81966