fix the CudaPinMemory bug for the equal op (#27176)

fix the CudaPinMemory bug for the equal op and add the test case for the equal op

paddle/fluid/operators/controlflow/compare_op.cc

@@ -111,8 +111,16 @@ class CompareOp : public framework::OperatorWithKernel {
     framework::OpKernelType kt = OperatorWithKernel::GetExpectedKernelType(ctx);
     // CompareOp kernel's device type is decided by input tensor place
     bool force_cpu = ctx.Attr<bool>("force_cpu");
-    kt.place_ = force_cpu ? platform::CPUPlace()
-                          : ctx.Input<framework::LoDTensor>("X")->place();
+    if (force_cpu) {
+      kt.place_ = platform::CPUPlace();
+    } else {
+      if (ctx.Input<framework::LoDTensor>("X")->place().type() !=
+          typeid(platform::CUDAPinnedPlace)) {
+        kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
+      } else {
+        kt.place_ = ctx.GetPlace();
+      }
+    }
     return kt;
   }
 };

python/paddle/fluid/tests/unittests/test_compare_op.py

@@ -38,6 +38,7 @@ def create_test_class(op_type, typename, callback):
             self.check_output()
 
         def test_errors(self):
+            paddle.enable_static()
             with program_guard(Program(), Program()):
                 x = fluid.layers.data(name='x', shape=[2], dtype='int32')
                 y = fluid.layers.data(name='y', shape=[2], dtype='int32')
@@ -80,6 +81,7 @@ def create_paddle_case(op_type, callback):
                 self.place = paddle.CUDAPlace(0)
 
         def test_api(self):
+            paddle.enable_static()
             with program_guard(Program(), Program()):
                 x = fluid.data(name='x', shape=[4], dtype='int64')
                 y = fluid.data(name='y', shape=[4], dtype='int64')
@@ -92,6 +94,7 @@ def create_paddle_case(op_type, callback):
             self.assertEqual((res == self.real_result).all(), True)
 
         def test_broadcast_api_1(self):
+            paddle.enable_static()
             with program_guard(Program(), Program()):
                 x = paddle.static.data(
                     name='x', shape=[1, 2, 1, 3], dtype='int32')
@@ -108,6 +111,7 @@ def create_paddle_case(op_type, callback):
             self.assertEqual((res == real_result).all(), True)
 
         def test_attr_name(self):
+            paddle.enable_static()
             with program_guard(Program(), Program()):
                 x = fluid.layers.data(name='x', shape=[4], dtype='int32')
                 y = fluid.layers.data(name='y', shape=[4], dtype='int32')
@@ -130,6 +134,7 @@ create_paddle_case('not_equal', lambda _a, _b: _a != _b)
 
 class TestCompareOpError(unittest.TestCase):
     def test_errors(self):
+        paddle.enable_static()
         with program_guard(Program(), Program()):
             # The input x and y of compare_op must be Variable.
             x = fluid.layers.data(name='x', shape=[1], dtype="float32")
@@ -140,6 +145,7 @@ class TestCompareOpError(unittest.TestCase):
 
 class API_TestElementwise_Equal(unittest.TestCase):
     def test_api(self):
+        paddle.enable_static()
         with fluid.program_guard(fluid.Program(), fluid.Program()):
             label = fluid.layers.assign(np.array([3, 3], dtype="int32"))
             limit = fluid.layers.assign(np.array([3, 2], dtype="int32"))
@@ -159,5 +165,31 @@ class API_TestElementwise_Equal(unittest.TestCase):
         self.assertEqual((res == np.array([True, True])).all(), True)
 
 
+class TestCompareOpPlace(unittest.TestCase):
+    def test_place_1(self):
+        paddle.enable_static()
+        place = paddle.CPUPlace()
+        if core.is_compiled_with_cuda():
+            place = paddle.CUDAPlace(0)
+        label = fluid.layers.assign(np.array([3, 3], dtype="int32"))
+        limit = fluid.layers.assign(np.array([3, 2], dtype="int32"))
+        out = fluid.layers.less_than(label, limit, force_cpu=True)
+        exe = fluid.Executor(place)
+        res, = exe.run(fetch_list=[out])
+        self.assertEqual((res == np.array([False, False])).all(), True)
+
+    def test_place_2(self):
+        place = paddle.CPUPlace()
+        data_place = place
+        if core.is_compiled_with_cuda():
+            place = paddle.CUDAPlace(0)
+            data_place = paddle.CUDAPinnedPlace()
+        paddle.disable_static(place)
+        data = np.array([9], dtype="int64")
+        data_tensor = paddle.to_tensor(data, place=data_place)
+        result = data_tensor == 0
+        self.assertEqual((result.numpy() == np.array([False])).all(), True)
+
+
 if __name__ == '__main__':
     unittest.main()