fix(linag.pack): decompisition of OuterUnitDims Pattern (#16)

Given the following example:
```
module {
  func.func @main(%arg0: tensor<1x1x1x4x1xf32>, %arg1: tensor<1x1x4xf32>) -> tensor<1x1x1x4x1xf32> {
    %pack = linalg.pack %arg1 outer_dims_perm = [1, 2, 0] inner_dims_pos = [2, 0] inner_tiles = [4, 1] into %arg0 : tensor<1x1x4xf32> -> tensor<1x1x1x4x1xf32>
    return %pack : tensor<1x1x1x4x1xf32>
  }
}
```

We would generate an invalid transpose operation because the calculated permutation would be `[0, 2, 0]` which is semantically incorrect. As the permutation must contain unique integers corresponding to the source tensor.

The assumption that because all outer dimensions are 1 the transposition does not matter does not hold because the indices in `permutation` must be unique and match the source tensor indices. 

The following change attempts to amend that.

Author: chrsmcgrr

mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp

@@ -1192,16 +1192,23 @@ LogicalResult DecomposeOuterUnitDimsPackOpPattern::matchAndRewrite(
   //    %init = tensor.empty()
   //    %transposed_tile = linalg.transpose ins(%source_or_padded_source),
   //                                        outs(%init)
-  // Two assumptions are made:
-  //  1. All outer dims are 1 - the corresponding transposition doesn't matter.
-  //  2. Inner dims position correspond to the trailing `numTiles` dims.
-  SmallVector<int64_t> tilesPermNormalized =
-      getPackUnpackNormalizedPerm(srcRank, packOp.getInnerDimsPos());
+  // Assumptions made:
+  //  1. Inner dims position correspond to the trailing `numTiles` dims.
   SmallVector<int64_t> srcPermForTranspose;
-  for (int64_t i = 0; i < (srcRank - numTiles); i++)
+  ArrayRef<int64_t> innerDimPos(packOp.getInnerDimsPos());
+  for (int64_t i = 0; i < srcRank; i++) {
+    // As we assume the trailing dimensions of the inner dim position correspond
+    // to the trailing indices of the transpose permutation, we need to
+    // calculate the remaining indicies of the transpose permutation. This is
+    // done by adding the indices not contained in the inner dimension position.
+    //   For example if we have a source tensor of dimensions [0, 1, 2, 3]
+    //   and inner dim position of [3, 0], the remaining indices are [1, 2].
+    //   and the transpose will be [1, 2, 3, 0].
+    if (llvm::is_contained(innerDimPos, i))
+      continue;
     srcPermForTranspose.push_back(i);
-
-  srcPermForTranspose.append(SmallVector<int64_t>(packOp.getInnerDimsPos()));
+  }
+  srcPermForTranspose.append(innerDimPos.begin(), innerDimPos.end());
 
   LLVM_DEBUG(DBGS() << "Pack permutation: " << packOp << "\n"
                     << "perm: " << llvm::interleaved(srcPermForTranspose)

mlir/test/Dialect/Linalg/decompose-pack.mlir

@@ -229,3 +229,22 @@ func.func @simple_KCRS_to_KRSCsr(%arg0: tensor<1x1x32x8xf32>, %arg1: tensor<1x1x
 // CHECK:         %[[INSERT:.+]] = tensor.insert_slice %[[TRANSP]] into %[[DEST]]
 // CHECK-SAME:      [0, 0, 0, 0, 0, 0] [1, 1, 1, 1, 8, 32] [1, 1, 1, 1, 1, 1]
 // CHECK:         return %[[INSERT]]
+
+// -----
+
+func.func @pack_with_unit_outer_dims_and_unit_inner(%arg0: tensor<1x1x4xf32>, %arg1: tensor<1x1x1x4x1xf32>) -> tensor<1x1x1x4x1xf32> {
+  %pack = linalg.pack %arg0 outer_dims_perm = [1, 2, 0] inner_dims_pos = [2, 0] inner_tiles = [4, 1] into %arg1 : tensor<1x1x4xf32> -> tensor<1x1x1x4x1xf32>
+  return %pack : tensor<1x1x1x4x1xf32>
+}
+
+// CHECK-LABEL: func.func @pack_with_unit_outer_dims_and_unit_inner
+// CHECK-SAME:    %[[SRC:[a-zA-Z0-9]+]]
+// CHECK-SAME:    %[[DEST:[a-zA-Z0-9]+]]
+// CHECK:         %[[EMPTY:.+]] = tensor.empty() : tensor<1x4x1xf32>
+// CHECK:         %[[TRANSP:.+]] = linalg.transpose
+// CHECK-SAME:      ins(%[[SRC]] : tensor<1x1x4xf32>)
+// CHECK-SAME:      outs(%[[EMPTY]] : tensor<1x4x1xf32>)
+// CHECK-SAME:      permutation = [1, 2, 0]
+// CHECK:         %[[INSERT:.+]] = tensor.insert_slice %[[TRANSP]] into %[[DEST]]
+// CHECK-SAME:      [0, 0, 0, 0, 0] [1, 1, 1, 4, 1] [1, 1, 1, 1, 1] : tensor<1x4x1xf32> into tensor<1x1x1x4x1xf32>
+// CHECK:         return %[[INSERT]]