Fix symint related functionalization ops (#1289)

* Fix symint related functionalization ops

* Remove zeros xfail from LTC tests

build_tools/autogen_ltc_backend.py

@@ -24,8 +24,9 @@
 from torchgen.gen_backend_stubs import parse_backend_yaml
 
 TORCH_DIR = Path(importlib.util.find_spec("torch").origin).resolve().parent.parent
-if TORCH_DIR.joinpath("torch", "include").is_dir():
-    TORCH_DIR = TORCH_DIR.joinpath("torch", "include")
+TORCH_INCLUDE_DIR = TORCH_DIR.joinpath("torch", "include")
+if not TORCH_INCLUDE_DIR.is_dir():
+    TORCH_INCLUDE_DIR = TORCH_DIR
 TORCHGEN_DIR = Path(torchgen.__path__[0]).resolve()
 TORCH_MLIR_DIR = Path(__file__).resolve().parent.parent
 
@@ -167,6 +168,9 @@ def generate_native_functions(self):
         ts_native_yaml = None
         if ts_native_yaml_path.exists():
             ts_native_yaml = yaml.load(ts_native_yaml_path.read_text(), yaml.CLoader)
+        else:
+            logging.warning(f"Could not find `ts_native_functions.yaml` at {ts_native_yaml_path}")
+
 
         parsed_yaml = parse_native_yaml(native_yaml_path, tags_yaml_path)
         self.native_functions = parsed_yaml.native_functions
@@ -290,6 +294,7 @@ def get_opnames(ops):
 
         if ts_native_yaml:
             ts_full_codegen = set(ts_native_yaml["full_codegen"])
+            ts_supported = set(ts_native_yaml["supported"])
             mlir_full_codegen = set(self.ops)
 
             if ts_full_codegen - mlir_full_codegen:
@@ -308,6 +313,22 @@ def get_opnames(ops):
                     )
                 )
 
+            if ts_supported - supported:
+                logging.debug(
+                    "Ops supported by the TorchScript backend "
+                    "but not by the Torch-MLIR backend:\n    {}".format(
+                        "\n    ".join(sorted(ts_supported - supported))
+                    )
+                )
+
+            if supported - ts_supported:
+                logging.debug(
+                    "Ops supported by the Torch-MLIR backend "
+                    "but not by the TorchScript backend:\n    {}".format(
+                        "\n    ".join(sorted(supported - ts_supported))
+                    )
+                )
+
     def generate_shape_inference(self):
         parsed_backend_yaml = parse_backend_yaml(
             self.source_yaml,
@@ -367,7 +388,7 @@ def extract_signatures(text):
         )
         assert len(shape_inference_decls) > 0
         upstream_shape_inference_decls = extract_signatures(
-            TORCH_DIR.joinpath(
+            TORCH_INCLUDE_DIR.joinpath(
                 "torch", "csrc", "lazy", "core", "shape_inference.h"
             ).read_text()
         )

build_tools/autogen_ltc_backend.yaml

@@ -19,6 +19,7 @@ blacklist:
 - new_empty
 - rsub
 - slice.Tensor  # Disabled in favour of slice_copy.Tensor
+- zeros
 
 # Disabled in favour of functionalized alternatives
 - _reshape_alias
@@ -59,14 +60,12 @@ supported:
 # but their implementations call view operators (which we need to functionalize away).
 - block_diag
 - new_empty_strided
-- narrow_copy
 - pixel_shuffle
 - pixel_unshuffle
 - select_backward
 - slice_backward
 - diagonal_backward
 - _trilinear
-- linalg_inv_ex
 - linalg_pinv.atol_rtol_tensor
 - logsumexp.out
 

e2e_testing/torchscript/xfail_sets.py

@@ -435,12 +435,6 @@
     "NewOnesModuleFloat3D_basic",
     "NewOnesModuleInt2D_basic",
     "NewOnesModuleInt3D_basic",
-    "NewZerosModuleDefaultDtype_basic",
-    "NewZerosModuleFalsePinMemory_basic",
-    "NewZerosModuleFloat2D_basic",
-    "NewZerosModuleFloat3D_basic",
-    "NewZerosModuleInt2D_basic",
-    "NewZerosModuleInt3D_basic",
     "OnesLikeModule_defaultDtype",
     "OnesLikeModule_falsePinMemory",
     "OnesLikeModule_float",

python/torch_mlir/csrc/base_lazy_backend/mlir_native_functions.cpp

@@ -302,10 +302,12 @@ at::Tensor LazyNativeFunctions::_to_copy(
 };
 
 at::Tensor LazyNativeFunctions::empty(
-    at::IntArrayRef size, c10::optional<at::ScalarType> dtype,
+    at::SymIntArrayRef sym_size, c10::optional<at::ScalarType> dtype,
     c10::optional<at::Layout> layout, c10::optional<at::Device> device,
     c10::optional<bool> pin_memory,
     c10::optional<at::MemoryFormat> memory_format) {
+  // TODO: support this directly
+  auto size = c10::asIntArrayRefSlow(sym_size);
   const auto device_type = torch::lazy::getBackend()->EagerFallbackDeviceType();
   at::TensorOptions options = at::TensorOptions()
                                   .device(c10::Device(device_type))
@@ -331,7 +333,9 @@ at::Tensor LazyNativeFunctions::empty_strided(
     c10::optional<at::ScalarType> dtype, c10::optional<at::Layout> layout,
     c10::optional<at::Device> device, c10::optional<bool> pin_memory) {
   TORCH_LAZY_FN_COUNTER("lazy::");
-  at::Tensor t = empty(size, dtype, layout, device, pin_memory, c10::nullopt);
+  at::Tensor t = empty(
+    c10::SymIntArrayRef::fromIntArrayRef(size),
+    dtype, layout, device, pin_memory, c10::nullopt);
   return t.as_strided(size, stride, /*storage_offset=*/0);
 }
 
@@ -350,7 +354,7 @@ LazyNativeFunctions::fill_(at::Tensor& self, const at::Scalar& value) {
 at::Tensor LazyNativeFunctions::_unsafe_view(
     const at::Tensor& self, at::IntArrayRef size) {
   TORCH_LAZY_FN_COUNTER("lazy::");
-  return LazyNativeFunctions::view_copy(self, size);
+  return LazyNativeFunctions::view_copy(self, c10::SymIntArrayRef::fromIntArrayRef(size));
 }
 
 // This is needed by the torch.tensor constructor.
@@ -385,11 +389,6 @@ at::Tensor LazyNativeFunctions::new_empty_strided(
           self, size, stride, dtype, layout, device, pin_memory);
 }
 
-at::Tensor LazyNativeFunctions::narrow_copy(
-    const at::Tensor& self, int64_t dim, int64_t start, int64_t length) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(
-      narrow_copy)>::call(self, dim, start, length);
-}
 at::Tensor LazyNativeFunctions::pixel_shuffle(
     const at::Tensor& self, int64_t upscale_factor) {
   return at::functionalization::functionalize_aten_op<ATEN_OP(
@@ -425,11 +424,6 @@ at::Tensor LazyNativeFunctions::_trilinear(
   return at::functionalization::functionalize_aten_op<ATEN_OP(_trilinear)>::
       call(i1, i2, i3, expand1, expand2, expand3, sumdim, unroll_dim);
 }
-::std::tuple<at::Tensor, at::Tensor>
-LazyNativeFunctions::linalg_inv_ex(const at::Tensor& self, bool check_errors) {
-  return at::functionalization::functionalize_aten_op<ATEN_OP(
-      linalg_inv_ex)>::call(self, check_errors);
-}
 at::Tensor LazyNativeFunctions::linalg_pinv(
     const at::Tensor& self, const c10::optional<at::Tensor>& atol,
     const c10::optional<at::Tensor>& rtol, bool hermitian) {