[Mlir-commits] [mlir] [mlir][tensor] Generalize/restrict `GeneralizeOuterUnitDimsPackOpPattern` (PR #114315)

[Quinn Dawkins]

================
@@ -1148,69 +1172,104 @@ LogicalResult GeneralizeOuterUnitDimsPackOpPattern::matchAndRewrite(
         packOp, "require the tiled outer dimensions of the result are all 1s");
   }
 
-  // 1. Use rank-reduced tensor.extract_slice op to extract the tile and untiled
-  // outer dims.
+  Attribute zeroIdxAttr = rewriter.getIndexAttr(0);
+  Attribute oneIdxAttr = rewriter.getIndexAttr(1);
   Location loc = packOp.getLoc();
+
   Value input = getPackOpSourceOrPaddedSource(rewriter, packOp);
   auto inputShape = packOp.getSourceType().getShape();
   DenseMap<int64_t, OpFoldResult> dimAndTileMapping =
       packOp.getDimAndTileMapping();
-  Attribute zeroIdxAttr = rewriter.getIndexAttr(0);
-  Attribute oneIdxAttr = rewriter.getIndexAttr(1);
   int64_t srcRank = packOp.getSourceRank();
+
+  int64_t destRank = packOp.getDestRank();
+  size_t numTiles = destRank - srcRank;
+
+  // 1. Use rank-reduced tensor.extract_slice op to extract the tile:
----------------
qedawkins wrote:

Can you elaborate on the advantage of slicing out the inner tile here? We end up reintroducing the outer unit dims shortly after with the `tensor.insert_slice` so this only serves to remove the unit dims from the `linalg.transpose` for permuting the inner dimensions.

https://github.com/llvm/llvm-project/pull/114315