[SOT][Faster Guard] add InstanceCheckGuard (#69975)

---------

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Author: gouzil

paddle/fluid/pybind/jit.cc

@@ -106,6 +106,11 @@ void BindGuard(pybind11::module *m) {
   py::class_<RangeMatchGuard, GuardBase, std::shared_ptr<RangeMatchGuard>>(
       *m, "RangeMatchGuard", R"DOC(RangeMatchGuard Class.)DOC")
       .def(py::init<const py::object &>(), py::arg("range_obj"));
+  py::class_<InstanceCheckGuard,
+             GuardBase,
+             std::shared_ptr<InstanceCheckGuard>>(
+      *m, "InstanceCheckGuard", R"DOC(InstanceCheckGuard Class.)DOC")
+      .def(py::init<const py::object &>(), py::arg("isinstance_obj"));
 
   m->def(
       "merge_guard",

paddle/fluid/pybind/sot/guards.cc

@@ -123,4 +123,8 @@ bool LayerMatchGuard::check(PyObject* value) {
   return (training == Py_True) == training_;
 }
 
+bool InstanceCheckGuard::check(PyObject* value) {
+  return PyObject_IsInstance(value, expected_);
+}
+
 #endif

paddle/fluid/pybind/sot/guards.h

@@ -200,4 +200,19 @@ class RangeMatchGuard : public GuardGroup {
   }
 };
 
+class InstanceCheckGuard : public GuardBase {
+ public:
+  explicit InstanceCheckGuard(const py::object& py_type)
+      : expected_(py_type.ptr()) {
+    Py_INCREF(expected_);
+  }
+
+  ~InstanceCheckGuard() override { Py_DECREF(expected_); }
+
+  bool check(PyObject* value) override;
+
+ private:
+  PyObject* expected_;
+};
+
 #endif

python/paddle/jit/sot/opcode_translator/executor/variables/basic.py

@@ -1069,8 +1069,9 @@ def get_py_value(self, allow_tensor=False):
     def make_stringified_guard(self) -> list[StringifiedExpression]:
         frame_value_tracer = self.tracker.trace_value_from_frame()
         result = [
-            StringifiedExpression(
-                "isinstance({}, slice)",
+            FasterStringifiedExpression(
+                "id(type({{}})) == id(slice)",
+                paddle.framework.core.TypeMatchGuard(slice),
                 [frame_value_tracer],
                 frame_value_tracer.free_vars,
             ),

python/paddle/jit/sot/opcode_translator/executor/variables/container.py

@@ -19,10 +19,16 @@
 from functools import reduce
 from typing import TYPE_CHECKING, Any
 
+import paddle
+
 from ....utils import ConstTypes
 from ....utils.exceptions import FallbackError, InnerError
 from ..dispatcher import Dispatcher
-from ..guard import StringifiedExpression, check_guard
+from ..guard import (
+    FasterStringifiedExpression,
+    StringifiedExpression,
+    check_guard,
+)
 from ..mutable_data import MutableDictLikeData, MutableListLikeData
 from ..tracker import (
     ConstTracker,
@@ -74,13 +80,23 @@ def bool(self):
     def make_stringified_guard(self) -> list[StringifiedExpression]:
         frame_value_tracer = self.tracker.trace_value_from_frame()
 
-        type_guard = StringifiedExpression(
-            f"isinstance({{}}, {self.get_py_type().__name__})",
-            [frame_value_tracer],
-            frame_value_tracer.free_vars,
-        )
-        len_guard = StringifiedExpression(
+        if self.get_py_type() is dict:
+            type_guard = FasterStringifiedExpression(
+                f"isinstance({{}}, {self.get_py_type().__name__})",
+                paddle.framework.core.InstanceCheckGuard(self.get_py_type()),
+                [frame_value_tracer],
+                frame_value_tracer.free_vars,
+            )
+        else:
+            type_guard = FasterStringifiedExpression(
+                f"id(type({{}})) == {id(self.get_py_type())}",
+                paddle.framework.core.TypeMatchGuard(self.get_py_type()),
+                [frame_value_tracer],
+                frame_value_tracer.free_vars,
+            )
+        len_guard = FasterStringifiedExpression(
             f"len({{}}) == {len(self.init_value)}",
+            paddle.framework.core.LengthMatchGuard(len(self.init_value)),
             [frame_value_tracer],
             frame_value_tracer.free_vars,
         )

test/sot/test_04_list.py

@@ -21,7 +21,7 @@
 
 import unittest
 
-from test_case_base import TestCaseBase
+from test_case_base import TestCaseBase, test_with_faster_guard
 
 import paddle
 from paddle.jit.sot.psdb import check_no_breakgraph
@@ -368,6 +368,7 @@ def test_list_add(self):
     def test_list_inplace_add(self):
         self.assert_results(list_inplace_add)
 
+    @test_with_faster_guard
     def test_list_extend_range(self):
         self.assert_results(list_extend_range, paddle.to_tensor([1, 2]))
 

test/sot/test_05_dict.py

@@ -18,7 +18,7 @@
 
 import unittest
 
-from test_case_base import TestCaseBase
+from test_case_base import TestCaseBase, test_with_faster_guard
 
 import paddle
 from paddle.jit.sot.psdb import check_no_breakgraph
@@ -253,6 +253,7 @@ def test_construct(self):
     def test_dict_noargs(self):
         self.assert_results(dict_no_arguments)
 
+    @test_with_faster_guard
     def test_dict_fromkeys(self):
         self.assert_results(dict_test_fromkeys, (1, 2, 3, 4))
         self.assert_results(dict_test_fromkeys, [1, 2, 3, 4])

test/sot/test_07_unpack.py

@@ -19,7 +19,7 @@
 
 import unittest
 
-from test_case_base import TestCaseBase
+from test_case_base import TestCaseBase, test_with_faster_guard
 
 import paddle
 
@@ -56,6 +56,7 @@ def test_unpack_tuple(self):
     def test_unpack_tensor(self):
         self.assert_results(unpack_tensor, paddle.to_tensor([2, 3]))
 
+    @test_with_faster_guard
     def test_unpack_ex_tuple(self):
         self.assert_results(unpack_ex_tuple, (1, 1, paddle.to_tensor(2)))
 

test/sot/test_12_for_loop.py

@@ -165,6 +165,7 @@ def test_list(self):
         a = paddle.to_tensor(1)
         self.assert_results(for_list_1, a)
 
+    @test_with_faster_guard
     def test_list_with_fallback(self):
         a = paddle.to_tensor(1)
         self.assert_results(for_list_2, a)
@@ -194,6 +195,7 @@ def test_for_break(self):
         paddle_output = for_break(a, gener())
         self.assert_nest_match(sym_output, paddle_output)
 
+    @test_with_faster_guard
     def test_for_continue(self):
         a = paddle.to_tensor(1)
         sym_output = symbolic_translate(for_continue)(a, gener())
@@ -205,6 +207,7 @@ def test_for_continue(self):
     #     a = [1, 2, 3]
     #     self.assert_results(for_enumerate_var_with_nested_range, a)
 
+    @test_with_faster_guard
     def test_create_var_in_loop(self):
         x = paddle.to_tensor(1, dtype="float32")
         a = [1, 2, 3]
@@ -217,6 +220,7 @@ def test_create_var_in_loop(self):
     def test_create_var_in_loop_with_same_name_as_global(self):
         self.assert_results(for_tmp_var_with_same_name_as_global_var)
 
+    @test_with_faster_guard
     def test_for_without_zero_iter(self):
         self_res_dict = {}
         output = paddle.to_tensor(2)
@@ -253,6 +257,7 @@ def for_enumerate_cache(func_list, x):
 
 
 class TestEnumerateCache(TestCaseBase):
+    @test_with_faster_guard
     def test_run(self):
         func_list = [
             paddle.nn.Linear(10, 10),
@@ -294,6 +299,7 @@ class TestUndefinedVarInRiskyCodes(TestCaseBase):
     def test_undefined_var_case_0(self):
         self.assert_results(undefined_var_case_0)
 
+    @test_with_faster_guard
     def test_undefined_var_case_1(self):
         self.assert_results(undefined_var_case_1)
 

test/sot/test_14_operators.py

@@ -15,7 +15,7 @@
 import operator
 import unittest
 
-from test_case_base import TestCaseBase
+from test_case_base import TestCaseBase, test_with_faster_guard
 
 import paddle
 
@@ -328,6 +328,7 @@ def test_simple(self):
         self.assert_results(inplace_or, b, g)
         self.assert_results(inplace_xor, b, g)
 
+    @test_with_faster_guard
     def test_operator_simple(self):
         self.assert_results(operator_add, 1, paddle.to_tensor(2))
         self.assert_results(operator_mul, 1, paddle.to_tensor(2))

test/sot/test_15_slice.py

@@ -18,7 +18,7 @@
 
 import unittest
 
-from test_case_base import TestCaseBase
+from test_case_base import TestCaseBase, test_with_faster_guard
 
 import paddle
 from paddle.jit.sot.psdb import check_no_breakgraph
@@ -55,6 +55,7 @@ def tensor_subscript_tensor(x: paddle.Tensor):
 
 
 class TestSlice(TestCaseBase):
+    @test_with_faster_guard
     def test_simple(self):
         x = list(range(10))
         y = paddle.arange(10)
@@ -83,6 +84,7 @@ def layer_list_slice(layer, x):
 
 
 class TestLayerList(TestCaseBase):
+    @test_with_faster_guard
     def test_layer_list_slice(self):
         layer = MyLayer()
         x = paddle.randn([5, 10])
@@ -127,6 +129,7 @@ def forward(self, x):
 
 
 class TestLayerListSlice(TestCaseBase):
+    @test_with_faster_guard
     def test_layer_list_slice(self):
         x = paddle.randn([2, 5])
         net = LayerListNet()

test/sot/test_18_tensor_method.py

@@ -14,7 +14,7 @@
 
 import unittest
 
-from test_case_base import TestCaseBase
+from test_case_base import TestCaseBase, test_with_faster_guard
 
 import paddle
 
@@ -83,6 +83,7 @@ def test_tensor_method_passed_by_user(self):
         y = paddle.rand([42])
         self.assert_results(tensor_method_passed_by_user, x, y.add)
 
+    @test_with_faster_guard
     def test_tensor_method_property(self):
         x = paddle.rand([42, 24], dtype='float64')
         y = paddle.rand([42, 24], dtype='float32')

test/sot/test_builtin_bool.py

@@ -17,6 +17,7 @@
 from test_case_base import (
     TestCaseBase,
     test_instruction_translator_cache_context,
+    test_with_faster_guard,
 )
 
 import paddle
@@ -107,6 +108,7 @@ def test_object_disallow_breakgraph(self):
                 call_bool_by_operator_truth_no_breakgraph, layer
             )
 
+    @test_with_faster_guard
     def test_object_allow_breakgraph(self):
         with test_instruction_translator_cache_context():
             obj = TestObjectWithLen([1, 2, 3])

test/sot/test_builtin_dispatch.py

@@ -22,6 +22,7 @@
 from test_case_base import (
     TestCaseBase,
     test_instruction_translator_cache_context,
+    test_with_faster_guard,
 )
 
 import paddle
@@ -135,9 +136,11 @@ def test_log(x: int):
 
 
 class TestBuiltinDispatch(TestCaseBase):
+    @test_with_faster_guard
     def test_dispatch_len(self):
         self.assert_results(dispatch_len, paddle.to_tensor([1, 2, 3]))
 
+    @test_with_faster_guard
     def test_dispatch_bool(self):
         self.assert_results(dispatch_bool, paddle.to_tensor([1, 2, 3]))
 
@@ -177,6 +180,7 @@ def test_not_dispatch_tensor_floor(self):
     def test_dispatch_float_floor(self):
         self.assert_results(dispatch_floor, 1.2)
 
+    @test_with_faster_guard
     def test_dispatch_sum(self):
         self.assert_results(test_sum_tuple, 1, 1)
         self.assert_results(test_sum_tuple, paddle.to_tensor(1), 1)

test/sot/test_builtin_map.py

@@ -17,7 +17,7 @@
 import unittest
 from typing import TYPE_CHECKING
 
-from test_case_base import TestCaseBase
+from test_case_base import TestCaseBase, test_with_faster_guard
 
 from paddle.jit import sot
 from paddle.jit.sot.psdb import check_no_breakgraph
@@ -99,12 +99,14 @@ def test_map_for_loop(x: list):
 
 
 class TestMap(TestCaseBase):
+    @test_with_faster_guard
     def test_map(self):
         self.assert_results(test_map_list, [1, 2, 3, 4])
         self.assert_results(test_map_tuple, (1, 2, 3, 4))
         self.assert_results(test_map_range, range(5))
         self.assert_results(test_map_dict, {"a": 1, "b": 2, "c": 3})
 
+    @test_with_faster_guard
     def test_map_comprehension(self):
         self.assert_results(test_map_list_comprehension, [1, 2, 3, 4])
         self.assert_results(test_map_tuple_comprehension, (1, 2, 3, 4))
@@ -117,9 +119,11 @@ def test_map_with_breakgraph(self):
         with strict_mode_guard(False):
             self.assert_results(test_map_list_with_breakgraph, [1, 2, 3, 4])
 
+    @test_with_faster_guard
     def test_map_unpack(self):
         self.assert_results(test_map_unpack, [1, 2, 3, 4])
 
+    @test_with_faster_guard
     def test_map_for_loop(self):
         self.assert_results(test_map_for_loop, [7, 8, 9, 10])
 

test/sot/test_builtin_zip.py

@@ -14,7 +14,7 @@
 
 import unittest
 
-from test_case_base import TestCaseBase
+from test_case_base import TestCaseBase, test_with_faster_guard
 
 import paddle
 from paddle.jit.sot import psdb, symbolic_translate
@@ -77,6 +77,7 @@ def test_zip_8(iter_1, iter_2):
 
 
 class TestZip(TestCaseBase):
+    @test_with_faster_guard
     def test_simple_cases(self):
         x = 8
         y = 5
@@ -92,10 +93,12 @@ def test_simple_cases(self):
         self.assert_results(test_zip_6, ty)
         self.assert_results(test_zip_7, layer_list, paddle.randn((10,)))
 
+    @test_with_faster_guard
     @min_graph_size_guard(0)
     def test_reconstruct(self):
         self.assert_results(test_zip_8, [1, 2, 3], [4, 5, 6])
 
+    @test_with_faster_guard
     @strict_mode_guard(False)
     @min_graph_size_guard(0)
     def test_zip_user_defined_iter(self):

test/sot/test_dtype.py

@@ -74,6 +74,7 @@ def test_dtype_reconstruct(self):
 
 
 class TestDtypeGuard(TestCaseBase):
+    @test_with_faster_guard
     def test_dtype_guard(self):
         dtype_map = {paddle.float32: paddle.float64}
         x = paddle.ones([2, 3], dtype="float32")