Implement lowering of aten.atleast_2d (#3546)

This operator is needed to implement aten.vstack, which will be
submitted in a subsequent PR

include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td

@@ -10277,6 +10277,29 @@ def Torch_AtenAtleast1dOp : Torch_Op<"aten.atleast_1d", [
   }];
 }
 
+def Torch_AtenAtleast2dOp : Torch_Op<"aten.atleast_2d", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::atleast_2d : (Tensor) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult AtenAtleast2dOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 1, 1);
+    }
+    void AtenAtleast2dOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 1, 1);
+    }
+  }];
+}
+
 def Torch_AtenEinsumOp : Torch_Op<"aten.einsum", [
     AllowsTypeRefinement,
     HasValueSemantics,

lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp

@@ -10546,6 +10546,28 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    }\n"
 "    return %2 : !torch.list<int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten.atleast_2d\"(%arg0: !torch.list<int>) -> !torch.list<int> {\n"
+"    %int0 = torch.constant.int 0\n"
+"    %int1 = torch.constant.int 1\n"
+"    %0 = torch.aten.len.t %arg0 : !torch.list<int> -> !torch.int\n"
+"    %1 = torch.aten.eq.int %0, %int0 : !torch.int, !torch.int -> !torch.bool\n"
+"    %2 = torch.prim.If %1 -> (!torch.list<int>) {\n"
+"      %3 = torch.prim.ListConstruct %int1, %int1 : (!torch.int, !torch.int) -> !torch.list<int>\n"
+"      torch.prim.If.yield %3 : !torch.list<int>\n"
+"    } else {\n"
+"      %3 = torch.aten.len.t %arg0 : !torch.list<int> -> !torch.int\n"
+"      %4 = torch.aten.eq.int %3, %int1 : !torch.int, !torch.int -> !torch.bool\n"
+"      %5 = torch.prim.If %4 -> (!torch.list<int>) {\n"
+"        %6 = torch.aten.__getitem__.t %arg0, %int0 : !torch.list<int>, !torch.int -> !torch.int\n"
+"        %7 = torch.prim.ListConstruct %int1, %6 : (!torch.int, !torch.int) -> !torch.list<int>\n"
+"        torch.prim.If.yield %7 : !torch.list<int>\n"
+"      } else {\n"
+"        torch.prim.If.yield %arg0 : !torch.list<int>\n"
+"      }\n"
+"      torch.prim.If.yield %5 : !torch.list<int>\n"
+"    }\n"
+"    return %2 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.stack\"(%arg0: !torch.list<list<int>>, %arg1: !torch.int) -> !torch.list<int> {\n"
 "    %0 = call @__torch__.torch.jit._shape_functions.stack(%arg0, %arg1) : (!torch.list<list<int>>, !torch.int) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
@@ -15044,6 +15066,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
 "    return %0#1 : !torch.int\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten.atleast_2d\"(%arg0: !torch.tuple<int, int>) -> !torch.int {\n"
+"    %0:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"    return %0#1 : !torch.int\n"
+"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten.einsum\"(%arg0: !torch.str, %arg1: !torch.list<tuple<int, int>>, %arg2: !torch.optional<list<int>>) -> !torch.int {\n"
 "    %true = torch.constant.bool true\n"
 "    %none = torch.constant.none\n"

lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp

@@ -1799,6 +1799,53 @@ public:
 };
 } // namespace
 
+namespace {
+// Decompose aten.atleast_2d into: aten.reshape. See
+// https://github.com/pytorch/pytorch/blob/9a8ab778d34bd24c5caceb340837483decc4c311/torch/_refs/__init__.py#L2604
+// def atleast_2d(
+//     arg: Union[TensorLikeType, Sequence[TensorLikeType]], *args:
+//     TensorLikeType
+// ) -> Union[TensorLikeType, Tuple[TensorLikeType, ...]]:
+//     """Reference implementation of :func:`torch.atleast_2d`."""
+//     if not args and isinstance(arg, collections.abc.Sequence):
+//         args_ = arg
+//     else:
+//         assert not isinstance(arg, collections.abc.Sequence)
+//         args_ = (arg,) + args
+//     unsqueeze_atleast_1d = partial(_unsqueeze_atleast, atleast_1d, 0)
+//     res = tuple(a if a.ndim >= 2 else unsqueeze_atleast_1d(a) for a in args_)
+//     return res if len(res) > 1 else res[0]
+class DecomposeAtenAtleast2dOp : public OpRewritePattern<AtenAtleast2dOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(AtenAtleast2dOp op,
+                                PatternRewriter &rewriter) const override {
+    Location loc = op.getLoc();
+    Value input = op.getSelf();
+    Type opType = op.getType();
+
+    auto inputType = cast<BaseTensorType>(input.getType());
+    SmallVector<int64_t> inputShape(inputType.getSizes());
+
+    if (inputShape.size() >= 2) {
+      rewriter.replaceOp(op, input);
+      return success();
+    }
+    auto atleast1dResShape =
+        inputShape.empty() ? SmallVector<int64_t, 1>{1} : inputShape;
+    auto atleast1dResType = rewriter.getType<ValueTensorType>(
+        atleast1dResShape, inputType.getOptionalDtype());
+    auto atleast1dRes =
+        rewriter.create<AtenAtleast1dOp>(loc, atleast1dResType, input);
+    Value zero = rewriter.create<Torch::ConstantIntOp>(
+        loc, rewriter.getI64IntegerAttr(0));
+    rewriter.replaceOpWithNewOp<AtenUnsqueezeOp>(op, opType, atleast1dRes,
+                                                 zero);
+    return success();
+  }
+};
+} // namespace
+
 namespace {
 // Decompose AtenEinsumOp to AtenMatmulOp, and supports possible reduce
 // operation and permute operation. Currently, this pass doesn't support
@@ -9429,6 +9476,7 @@ public:
     addPatternIfTargetOpIsIllegal<DecomposeAtenRreluOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenCeluOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenAtleast1dOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAtenAtleast2dOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenEinsumOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenTraceOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenHardswishOp>(patterns);

lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp

@@ -396,6 +396,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
   target.addIllegalOp<Aten_SoftmaxBackwardDataOp>();
   target.addIllegalOp<AtenTanhBackwardOp>();
   target.addIllegalOp<AtenAtleast1dOp>();
+  target.addIllegalOp<AtenAtleast2dOp>();
   target.addIllegalOp<AtenEinsumOp>();
   target.addIllegalOp<AtenTraceOp>();
   target.addIllegalOp<AtenAddmmOp>();

projects/pt1/e2e_testing/xfail_sets.py

@@ -879,6 +879,9 @@ STABLEHLO_PASS_SET = {
     "TypeConversionUint8ToF32Module_basic",
     "Atleast1dModule0dInput_basic",
     "Atleast1dModule1dInput_basic",
+    "Atleast2dModule0dInput_basic",
+    "Atleast2dModule1dInput_basic",
+    "Atleast2dModule2dInput_basic",
     "AtenLinear1D_basic",
     "AtenLinear2D_basic",
     "AtenLinear3DBias_basic",
@@ -1576,6 +1579,9 @@ TOSA_PASS_SET = {
     "TensorSplitSections_ListUnpackModule_basic",
     "Atleast1dModule0dInput_basic",
     "Atleast1dModule1dInput_basic",
+    "Atleast2dModule0dInput_basic",
+    "Atleast2dModule1dInput_basic",
+    "Atleast2dModule2dInput_basic",
     "AtenLinear2D_basic",
     "AtenLinear3DBias_basic",
     "ElementwiseAddScalar_NumToTensorFloat_Module_basic",
@@ -2075,6 +2081,9 @@ MAKE_FX_TOSA_PASS_SET = (
         "AtenLinearVecMatBias_basic",
         "Atleast1dModule0dInput_basic",
         "Atleast1dModule1dInput_basic",
+        "Atleast2dModule0dInput_basic",
+        "Atleast2dModule1dInput_basic",
+        "Atleast2dModule2dInput_basic",
         "MaxPool1dEmptyStrideStaticModule_basic",
         "MaxPool1dStaticCeilModeTrueModule_basic",
         "MaxPool1dStaticModule_basic",

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py

@@ -2121,6 +2121,15 @@ def aten〇atleast_1d〡shape(self: List[int]) -> List[int]:
     else:
         return self
 
+def aten〇atleast_2d〡shape(self: List[int]) -> List[int]:
+    if len(self) == 0:
+        return [1, 1]
+    elif len(self) == 1:
+        x = self[0]
+        return [1, x]
+    else:
+        return self
+
 def aten〇stack〡shape(tensors: List[List[int]], dim: int = 0) -> List[int]:
     return upstream_shape_functions.stack(tensors, dim)
 
@@ -5265,6 +5274,11 @@ def aten〇atleast_1d〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
     self_rank, self_dtype = self_rank_dtype
     return self_dtype
 
+@check_dtype_function(_check_tensors_with_the_same_dtype(num_of_tensors=1))
+def aten〇atleast_2d〡dtype(self_rank_dtype: Tuple[int, int]) -> int:
+    self_rank, self_dtype = self_rank_dtype
+    return self_dtype
+
 @check_dtype_function(
     [Invocation("i,j->ij", [TensorOfShape(1, dtype=torch.float32),
                             TensorOfShape(1, dtype=torch.int32)]),])

projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py

@@ -795,6 +795,7 @@ def emit_ops(emitter_td: TextEmitter, registry: Registry):
     emit("aten::argmin : (Tensor, int?, bool) -> (Tensor)")
     emit("aten::one_hot : (Tensor, int) -> (Tensor)")
     emit("aten::atleast_1d : (Tensor) -> (Tensor)")
+    emit("aten::atleast_2d : (Tensor) -> (Tensor)")
     emit("aten::einsum : (str, Tensor[], int[]?) -> (Tensor)")
     emit("aten::trace : (Tensor) -> (Tensor)")
     emit("aten::bucketize.Tensor : (Tensor, Tensor, bool, bool) -> (Tensor)")

projects/pt1/python/torch_mlir_e2e_test/test_suite/reshape_like.py

@@ -1525,3 +1525,66 @@ class Atleast1dModule1dInput(torch.nn.Module):
 @register_test_case(module_factory=lambda: Atleast1dModule1dInput())
 def Atleast1dModule1dInput_basic(module, tu: TestUtils):
     module.forward(tu.rand(4))
+
+
+# ==============================================================================
+
+
+class Atleast2dModule0dInput(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([], torch.float32, True),
+        ]
+    )
+    def forward(self, x):
+        return torch.ops.aten.atleast_2d(x)
+
+
+@register_test_case(module_factory=lambda: Atleast2dModule0dInput())
+def Atleast2dModule0dInput_basic(module, tu: TestUtils):
+    module.forward(tu.rand())
+
+
+class Atleast2dModule1dInput(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([4], torch.float32, True),
+        ]
+    )
+    def forward(self, x):
+        return torch.ops.aten.atleast_2d(x)
+
+
+@register_test_case(module_factory=lambda: Atleast2dModule1dInput())
+def Atleast2dModule1dInput_basic(module, tu: TestUtils):
+    module.forward(tu.rand(4))
+
+
+class Atleast2dModule2dInput(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([4, 4], torch.float32, True),
+        ]
+    )
+    def forward(self, x):
+        return torch.ops.aten.atleast_2d(x)
+
+
+@register_test_case(module_factory=lambda: Atleast2dModule2dInput())
+def Atleast2dModule2dInput_basic(module, tu: TestUtils):
+    module.forward(tu.rand(4, 4))