fix comment

Author: Yun-Fly

include/gc/Transforms/Passes.td

@@ -60,9 +60,9 @@ def IterativeTilingAndFusion : Pass<"iterative-tiling-and-fusion",
   let description = [{
     The pass tries to fuse any MLIR operation which can be tiled. Moreover, this pass aims to support:
       1. Matmul fusion with element-wise/reduce/broadcast ops.
-      2. Pre-op and post-op fusion.
-      3. Multi-consumer and multi-producer support.
-      4. Multiple level of nest loops and candidates.
+      2. Producer and consumer fusion.
+      3. Arbitrary topology, including residual pattern with multiple consumers .
+      4. Nest loops structure with multiple level candidates.
       5. Flexible option to control the boundary of iterative process.
       6. Default tiling when no op is tiled before fusion.
       7. Cost-model to determine whether to fuse or not.
@@ -74,9 +74,9 @@ def IterativeTilingAndFusion : Pass<"iterative-tiling-and-fusion",
     Option<"useCostModel", "use-cost-model", "bool",
            /*default=*/"false",
            "Decide if enable cost model to control iterative fusion.">,
-    Option<"defaultNDTile", "default-nd-tile", "int",
+    Option<"defaultNDTile", "default-nd-tile", "unsigned",
            /*default=*/"2",
-           "Set default ND-Tile size, such as 2[default] or 3.">,
+           "Set default amount of non-one dimensions in TileSize, such as 1, 2[default, a.k.a. 2D-Tile], etc.">,
     ListOption<"defaultTileSize", "default-tile-size", "std::string",
            "Set default TileSize for the certain type of op, saying `matmul:{32,32}`.">,
     ];

lib/gc/Transforms/IterativeTilingAndFusion.cpp

@@ -479,8 +479,8 @@ tileAndFuseProducerOfOpOperand(RewriterBase &rewriter, OpOperand &operand,
     return std::nullopt;
 
   // c. Check the producer of root source if is tilable.
-  Operation *producer = realProducer->getDefiningOp<TilingInterface>();
-  if (!producer)
+  Operation *producerOp = realProducer->getDefiningOp<TilingInterface>();
+  if (!producerOp)
     return std::nullopt;
 
   CandidateDefOrUse defOrUse{*realProducer};
@@ -537,8 +537,8 @@ tileAndFuseConsumerOfOpResult(RewriterBase &rewriter, OpResult result,
   SmallVector<scf::SCFFuseConsumerOfSliceResult> fusedResultList;
   for (auto useOperand : *realConsumers) {
     // c. Check the consumer of top level result if is tilable.
-    Operation *consumer = dyn_cast<TilingInterface>(useOperand->getOwner());
-    if (!consumer)
+    Operation *consumerOp = dyn_cast<TilingInterface>(useOperand->getOwner());
+    if (!consumerOp)
       continue;
 
     CandidateDefOrUse defOrUse{useOperand};
@@ -560,7 +560,7 @@ tileAndFuseConsumerOfOpResult(RewriterBase &rewriter, OpResult result,
       // f. Manually run cse on region which contains original consumer op in
       // avoid of conflict with subsequent `tileAndFuseConsumerOfSlice` get nest
       // loops between next candidate sliceOp and tiled producer.
-      (void)mlir::simplifyRegions(rewriter, {*consumer->getParentRegion()});
+      (void)mlir::simplifyRegions(rewriter, {*consumerOp->getParentRegion()});
     }
   }
   if (fusedResultList.empty())
@@ -652,7 +652,7 @@ struct defaultTileConfig {
   // OpTy-to-TileSize mapping
   OpTileSizeMap tsMap;
   // ND-tile size
-  int ndTile;
+  unsigned ndTile;
 };
 
 /// Default Tiling function only effective for certain `OpTy` operation
@@ -690,7 +690,7 @@ defaultTilingOfType(RewriterBase &rewriter, Operation *op,
     SmallVector<int64_t> tsOrder = {32, 16};
     // Record how many dims have been tiled, including fully tiled, i.e.
     // tileSize == dimSize.
-    int nonOneTileDims =
+    unsigned nonOneTileDims =
         (isa<mlir::linalg::LinalgOp>(op) && !linalgx::isMatmulOp(op))
             ? cast<mlir::linalg::LinalgOp>(op).getNumReductionLoops()
             : 0;

test/mlir/test/gc/Transforms/iterative-tiling-and-fusion.mlir

@@ -366,7 +366,9 @@ module {
     /// CHECK: linalg.add
     %add = linalg.add ins(%arg0, %arg1 : tensor<1x32x4096xbf16>, tensor<1x32x4096xbf16>) outs(%dest0 : tensor<1x32x4096xbf16>) -> tensor<1x32x4096xbf16>
     %dest1 = tensor.empty() : tensor<1x1x128x32x32xbf16>
-    /// CHECK: tensor.pack
+    /// CHECK-NEXT: affine.apply
+    /// CHECK-NEXT: tensor.extract_slice
+    /// CHECK-NEXT: tensor.pack
     %pack = tensor.pack %add outer_dims_perm = [0, 1, 2] inner_dims_pos = [1, 2] inner_tiles = [32, 32] into %dest1 : tensor<1x32x4096xbf16> -> tensor<1x1x128x32x32xbf16>
     /// CHECK: scf.forall.in_parallel
     /// CHECK: tensor.parallel_insert_slice