[Mlir-commits] [mlir] [mlir][tensor] Enhance SimplifyUnPackToCollapseShape for unit dim cases. (PR #79262)

[Han-Chung Wang]

================
@@ -134,9 +135,36 @@ struct SimplifyUnPackToCollapseShape : public OpRewritePattern<UnPackOp> {
     ArrayRef<int64_t> dimsPos = unpackOp.getInnerDimsPos();
     if (dimsPos.size() != 1 || (dimsPos[0] + 1 != destType.getRank())) {
       return rewriter.notifyMatchFailure(
-          unpackOp, "expects unpacking at the innermost dimension");
+          unpackOp, "expects unpacking on the innermost dimension");
     }
 
+    return success();
+  }
+
+  /// Returns success() if it unpacks a 1D or 2D source operand to a 1D
----------------
hanhanW wrote:

The 1D/2D means that the number of dimension with `size>1`. The check about `getNumGtOneDims(destShape) <= 1` guarantees that the non-packed domain is either 0D or 1D. And the `getNumGtOneDims(innerTiles) <= 1` part guarantees that the packed is either 1D or 2D. Thus I think i is unpacking a 1D or 2D source value to a 0D or 1D value. 

https://github.com/llvm/llvm-project/pull/79262