[mlir][tensor] Fix consumer fusion for tensor.pack without explicit outer_dims_perm attribute (#106687)

Author: Yun-Fly

mlir/lib/Dialect/Tensor/IR/TensorTilingInterfaceImpl.cpp

@@ -286,7 +286,7 @@ struct PackOpTiling
     SmallVector<OpFoldResult> outerDimOffsets, outerDimSizes;
     DenseMap<int64_t, OpFoldResult> dimAndTileMapping =
         packOp.getDimAndTileMapping();
-    for (auto dim : packOp.getOuterDimsPerm()) {
+    for (auto dim : llvm::seq<int64_t>(packOp.getSourceRank())) {
       if (dimAndTileMapping.count(dim)) {
         FailureOr<int64_t> cstSize =
             ValueBoundsConstraintSet::computeConstantBound(
@@ -327,7 +327,7 @@ struct PackOpTiling
         outerDimSizes.push_back(sizes[dim]);
       }
     }
-
+    applyPermToRange(outerDimOffsets, outerDimSizes, packOp.getOuterDimsPerm());
     resultOffsets = outerDimOffsets;
     resultSizes = outerDimSizes;
     return success();

mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir

@@ -337,7 +337,7 @@ module {
             }
         }
         %output = tensor.empty() : tensor<4x32x16xf32>
-        %pack = tensor.pack %1 outer_dims_perm = [0, 1] inner_dims_pos = [0] inner_tiles = [16] into %output : tensor<64x32xf32> -> tensor<4x32x16xf32>
+        %pack = tensor.pack %1 inner_dims_pos = [0] inner_tiles = [16] into %output : tensor<64x32xf32> -> tensor<4x32x16xf32>
         return %pack : tensor<4x32x16xf32>
     }
 }
@@ -366,7 +366,7 @@ module attributes {transform.with_named_sequence} {
 //      CHECK:      %[[PACK_RESULT_OFFSET:.*]] = affine.apply #[[PACK_RESULT_MAP]](%[[IV1]])
 //      CHECK:      %[[TILED_PACK_DEST:.*]] = tensor.extract_slice %[[PACK_OUT_ARG]][%[[PACK_RESULT_OFFSET]], %[[IV2]], 0] [2, 32, 16] [1, 1, 1]
 //      CHECK:      %[[TILED_PACK_OUT:.*]] = tensor.pack %[[GENERIC_OUT]]
-// CHECK-SAME:                              outer_dims_perm = [0, 1] inner_dims_pos = [0] inner_tiles = [16]
+// CHECK-SAME:                              inner_dims_pos = [0] inner_tiles = [16]
 // CHECK-SAME:                              into %[[TILED_PACK_DEST]]
 //      CHECK:      scf.forall.in_parallel {
 //      CHECK:          tensor.parallel_insert_slice %[[TILED_PACK_OUT]] into %[[PACK_OUT_ARG]][%[[PACK_RESULT_OFFSET]],  %[[IV2]], 0] [2, 32, 16] [1, 1, 1]