[CINN] Reopen some prim+cinn unittest (#52595)

* [CINN] Reopen some prim+cinn unittest

* revert scatter for ci

* change cinn test timeout from 120 to 150

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -1136,6 +1136,7 @@ set(TEST_CINN_OPS
 foreach(TEST_CINN_OPS ${TEST_CINN_OPS})
   if(WITH_CINN)
     set_tests_properties(${TEST_CINN_OPS} PROPERTIES LABELS "RUN_TYPE=CINN")
+    set_tests_properties(${TEST_CINN_OPS} PROPERTIES TIMEOUT 150)
   endif()
 endforeach()
 

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

@@ -32,7 +32,6 @@ class TestAssignOp(eager_op_test.OpTest):
         self.public_python_api = paddle.assign
         self.op_type = "assign"
         self.prim_op_type = "prim"
-        self.enable_cinn = False
         x = np.random.random(size=(100, 10)).astype('float64')
         self.inputs = {'X': x}
         self.outputs = {'Out': x}
@@ -57,7 +56,6 @@ class TestAssignFP16Op(eager_op_test.OpTest):
         self.public_python_api = paddle.assign
         self.op_type = "assign"
         self.prim_op_type = "prim"
-        self.enable_cinn = False
         x = np.random.random(size=(100, 10)).astype('float16')
         self.inputs = {'X': x}
         self.outputs = {'Out': x}

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

@@ -30,9 +30,9 @@ class TestConcatOp(OpTest):
         self.python_api = paddle.concat
         self.public_python_api = paddle.concat
         self.prim_op_type = "prim"
-        self.enable_cinn = False
         self.dtype = self.get_dtype()
         self.init_test_data()
+        self.if_enable_cinn()
         self.inputs = {'X': [('x0', self.x0), ('x1', self.x1), ('x2', self.x2)]}
         self.attrs = {'axis': self.axis}
         if self.axis < 0:
@@ -82,6 +82,9 @@ class TestConcatOp(OpTest):
             self.x2 = np.random.random((5, 3, 4, 5)).astype(self.dtype)
         self.axis = 1
 
+    def if_enable_cinn(self):
+        pass
+
 
 class TestConcatOp2(TestConcatOp):
     def init_test_data(self):
@@ -291,6 +294,9 @@ def create_test_bf16(parent):
         def get_dtype(self):
             return np.uint16
 
+        def if_enable_cinn(self):
+            self.enable_cinn = False
+
     cls_name = "{}_{}".format(parent.__name__, "Bf16")
     TestConcatBf16.__name__ = cls_name
     globals()[cls_name] = TestConcatBf16

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

@@ -314,7 +314,7 @@ class TestSumOpExclusiveFP16(OpTest):
         self.python_api = cumsum_wrapper
         self.public_python_api = paddle.cumsum
         self.init_dtype()
-        self.enable_cinn = False
+        self.enable_cinn = True
         self.attrs = {'axis': 2, "exclusive": True}
         self.x = np.random.random((4, 5, 20)).astype(self.dtype)
         self.out = np.concatenate(
@@ -389,7 +389,7 @@ def create_test_fp16_class(parent, max_relative_error=1e-2):
             self.dtype = self.dtype_ = np.float16
 
         def set_enable_cinn(self):
-            self.enable_cinn = False
+            self.enable_cinn = True
 
         def test_check_output(self):
             self.check_output()

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

@@ -308,7 +308,6 @@ class TestFP16DropoutOp(OpTest):
             'is_test': True,
         }
         self.outputs = {'Out': out}
-        self.enable_cinn = False
         # Because prim op compare res with dygraph
         # when p = 0 dropout api return x,in dygraph mode x_grad = out_grad,
         # but in static mode x_grad = []
@@ -1689,23 +1688,25 @@ class TestCompositeDropout(unittest.TestCase):
         rev_actual = []
         paddle.disable_static()
         for place in self.places:
-            if isinstance(place, fluid.CPUPlace):
-                paddle.set_device("cpu")
-            if isinstance(place, fluid.CUDAPlace):
-                paddle.set_device("gpu")
+            if not isinstance(place, fluid.CUDAPlace):
+                continue
+            paddle.set_device("gpu")
             paddle.seed(self.seed)
             input_ = paddle.to_tensor(
                 data=self.x, dtype=self.dtype, place=place, stop_gradient=False
             )
             net = PrimNet()
-            net = apply_to_static(net, False)
+            net = apply_to_static(net, True)
             output = net(
                 input_, self.p, training=(not self.is_test), mode=self.mode
             )
             grad = paddle.grad(output, input_)
             fwd_actual.append(output.numpy())
             rev_actual.append(grad[0].numpy())
-        for i in range(len(self.places)):
+        i = 0
+        for place in self.places:
+            if not isinstance(self.places[i], fluid.CUDAPlace):
+                continue
             np.testing.assert_allclose(
                 self.fwd_desire[i].sum(),
                 fwd_actual[i].sum(),
@@ -1718,6 +1719,7 @@ class TestCompositeDropout(unittest.TestCase):
                 rtol=1e-2,  # mean of uniform distribution, scale for avoid random failed
                 atol=0,
             )
+            i += 1
 
 
 if __name__ == '__main__':

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

@@ -497,9 +497,6 @@ class TestElementwiseAddOp_rowwise_add_1(TestElementwiseAddOp):
         self.y = np.random.rand(100, 1).astype(self.dtype)
         self.out = self.x + self.y.reshape(1, 100, 1)
 
-    def if_enable_cinn(self):
-        self.enable_cinn = False
-
 
 @skip_check_grad_ci(
     reason="[skip shape check] Use y_shape(1) to test broadcast."

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

@@ -36,7 +36,7 @@ class TestElementwiseOp(OpTest):
         self.init_data()
         self.op_type = "elementwise_max"
         self.prim_op_type = "prim"
-        self.enable_cinn = False
+        self.if_enbale_cinn()
         self.python_api = paddle.maximum
         self.public_python_api = paddle.maximum
         self.inputs = {'X': self.x, 'Y': self.y}
@@ -95,6 +95,9 @@ class TestElementwiseOp(OpTest):
                 check_prim=True,
             )
 
+    def if_enbale_cinn(self):
+        pass
+
 
 class TestElementwiseFP16Op(TestElementwiseOp):
     def init_data(self):
@@ -108,7 +111,7 @@ class TestElementwiseFP16Op(TestElementwiseOp):
         self.init_data()
         self.op_type = "elementwise_max"
         self.prim_op_type = "prim"
-        self.enable_cinn = False
+        self.if_enbale_cinn()
         self.python_api = paddle.maximum
         self.dtype = np.float16
         self.public_python_api = paddle.maximum
@@ -121,36 +124,54 @@ class TestElementwiseMaxOp_ZeroDim1(TestElementwiseOp):
         self.x = np.random.uniform(0.1, 1, []).astype("float64")
         self.y = np.random.uniform(0.1, 1, []).astype("float64")
 
+    def if_enbale_cinn(self):
+        self.enable_cinn = False
+
 
 class TestElementwiseMaxFP16Op_ZeroDim1(TestElementwiseFP16Op):
     def init_data(self):
         self.x = np.random.uniform(0.1, 1, []).astype(np.float16)
         self.y = np.random.uniform(0.1, 1, []).astype(np.float16)
 
+    def if_enbale_cinn(self):
+        self.enable_cinn = False
+
 
 class TestElementwiseMaxOp_ZeroDim2(TestElementwiseOp):
     def init_data(self):
         self.x = np.random.uniform(0.1, 1, [13, 17]).astype("float64")
         self.y = np.random.uniform(0.1, 1, []).astype("float64")
 
+    def if_enbale_cinn(self):
+        self.enable_cinn = False
+
 
 class TestElementwiseMaxFP16Op_ZeroDim2(TestElementwiseFP16Op):
     def init_data(self):
         self.x = np.random.uniform(0.1, 1, [13, 17]).astype(np.float16)
         self.y = np.random.uniform(0.1, 1, []).astype(np.float16)
 
+    def if_enbale_cinn(self):
+        self.enable_cinn = False
+
 
 class TestElementwiseMaxOp_ZeroDim3(TestElementwiseOp):
     def init_data(self):
         self.x = np.random.uniform(0.1, 1, []).astype("float64")
         self.y = np.random.uniform(0.1, 1, [13, 17]).astype("float64")
 
+    def if_enbale_cinn(self):
+        self.enable_cinn = False
+
 
 class TestElementwiseMaxFP16Op_ZeroDim3(TestElementwiseFP16Op):
     def init_data(self):
         self.x = np.random.uniform(0.1, 1, []).astype(np.float16)
         self.y = np.random.uniform(0.1, 1, [13, 17]).astype(np.float16)
 
+    def if_enbale_cinn(self):
+        self.enable_cinn = False
+
 
 @unittest.skipIf(
     core.is_compiled_with_cuda()

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

@@ -141,7 +141,6 @@ class TestGatherNdOpWithLowIndex(OpTest):
         self.prim_op_type = "prim"
         self.python_api = paddle.gather_nd
         self.public_python_api = paddle.gather_nd
-        self.enable_cinn = False
         self.config_dtype()
         if self.dtype == np.float64:
             target_dtype = "float64"
@@ -216,6 +215,7 @@ class TestGatherNdOpIndex1(OpTest):
             output = convert_float_to_uint16(output)
         self.inputs = {'X': xnp, 'Index': index}
         self.outputs = {'Out': output}
+        # the outputs are 0D-tensor, CINN not support
         self.enable_cinn = False
 
     def config_dtype(self):
@@ -258,7 +258,6 @@ class TestGatherNdOpWithSameIndexAsX(OpTest):
         self.prim_op_type = "prim"
         self.python_api = paddle.gather_nd
         self.public_python_api = paddle.gather_nd
-        self.enable_cinn = False
         self.config_dtype()
         if self.dtype == np.float64:
             target_dtype = "float64"

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

@@ -1217,6 +1217,8 @@ class TestCompositeGroupNorm(unittest.TestCase):
         fwd_actual = []
         rev_actual = []
         for place in self.places:
+            if not isinstance(place, fluid.CUDAPlace):
+                continue
             input_ = paddle.to_tensor(
                 data=self.x, dtype=self.dtype, place=place, stop_gradient=False
             )
@@ -1241,13 +1243,16 @@ class TestCompositeGroupNorm(unittest.TestCase):
                 self.data_format,
             )
             # failed in cinn test
-            net = apply_to_static(net, False)
+            net = apply_to_static(net, True)
             output = net(input_)
             grad = paddle.grad(output, input_)
             fwd_actual.append(output.numpy())
             rev_actual.append(grad[0].numpy())
 
-        for i in range(len(self.places)):
+        i = 0
+        for place in self.places:
+            if not isinstance(place, fluid.CUDAPlace):
+                continue
             atol = self.threshold_list[i][2]
             rtol = self.threshold_list[i][2]
             np.testing.assert_allclose(
@@ -1269,6 +1274,7 @@ class TestCompositeGroupNorm(unittest.TestCase):
                 atol=atol,
                 err_msg='%s jit_cinn rev' % self.places[i],
             )
+            i += 1
 
 
 if __name__ == '__main__':

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

@@ -176,7 +176,7 @@ class TestReduceMeanOp(OpTest):
         pass
 
     def if_enable_cinn(self):
-        self.enable_cinn = False
+        pass
 
     def test_check_output(self):
         if self.dtype != 'float16':

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

@@ -196,7 +196,6 @@ class TestSplitOp_unk_section(OpTest):
         self.python_out_sig = ['out0', 'out1', 'out2']
         self._set_op_type()
         self.prim_op_type = "prim"
-        self.enable_cinn = False
         self.dtype = self.get_dtype()
         self.init_data()
         self.inputs = {'X': self.x}

python/paddle/fluid/tests/unittests/test_squeeze2_op.py

@@ -73,7 +73,6 @@ class TestSqueezeOp2(TestSqueezeOp):
         self.prim_op_type = "comp"
         self.python_api = paddle.squeeze
         self.public_python_api = paddle.squeeze
-        self.enable_cinn = False
         self.python_out_sig = [
             "Out"
         ]  # python out sig is customized output signature.

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

@@ -125,7 +125,6 @@ class TestCase10(TestTransposeOp):
         self.python_api = paddle.transpose
         self.public_python_api = paddle.transpose
         self.prim_op_type = "prim"
-        self.enable_cinn = False
         self.inputs = {'X': np.random.random(self.shape).astype("float64")}
         self.attrs = {
             'axis': list(self.axis),