[AMP_OP&Test] improve FP16 and BF16 OpTest for maximum, minimum and multiply op (#52256)

* [AMP_OP&Test] add FP16 and BF16 OpTest for minimum; add FP16 OpTest for multiply

* [AMP_OP&Test] reset atol and max_relative_error for multiply fp16 and bf16 optest

* [AMP_OP&Test] delete manually atol and max_relative_error setting

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

@@ -104,6 +104,17 @@ class TestElementwiseFP16Op(TestElementwiseOp):
             np.float16
         )
 
+    def setUp(self):
+        self.init_data()
+        self.op_type = "elementwise_max"
+        self.prim_op_type = "prim"
+        self.enable_cinn = False
+        self.python_api = paddle.maximum
+        self.dtype = np.float16
+        self.public_python_api = paddle.maximum
+        self.inputs = {'X': self.x, 'Y': self.y}
+        self.outputs = {'Out': np.maximum(self.inputs['X'], self.inputs['Y'])}
+
 
 class TestElementwiseMaxOp_ZeroDim1(TestElementwiseOp):
     def init_data(self):
@@ -111,10 +122,10 @@ class TestElementwiseMaxOp_ZeroDim1(TestElementwiseOp):
         self.y = np.random.uniform(0.1, 1, []).astype("float64")
 
 
-class TestElementwiseMaxFP16Op_ZeroDim1(TestElementwiseOp):
+class TestElementwiseMaxFP16Op_ZeroDim1(TestElementwiseFP16Op):
     def init_data(self):
-        self.x = np.random.uniform(0.1, 1, []).astype("float16")
-        self.y = np.random.uniform(0.1, 1, []).astype("float16")
+        self.x = np.random.uniform(0.1, 1, []).astype(np.float16)
+        self.y = np.random.uniform(0.1, 1, []).astype(np.float16)
 
 
 class TestElementwiseMaxOp_ZeroDim2(TestElementwiseOp):
@@ -123,10 +134,10 @@ class TestElementwiseMaxOp_ZeroDim2(TestElementwiseOp):
         self.y = np.random.uniform(0.1, 1, []).astype("float64")
 
 
-class TestElementwiseMaxFP16Op_ZeroDim2(TestElementwiseOp):
+class TestElementwiseMaxFP16Op_ZeroDim2(TestElementwiseFP16Op):
     def init_data(self):
-        self.x = np.random.uniform(0.1, 1, [13, 17]).astype("float16")
-        self.y = np.random.uniform(0.1, 1, []).astype("float16")
+        self.x = np.random.uniform(0.1, 1, [13, 17]).astype(np.float16)
+        self.y = np.random.uniform(0.1, 1, []).astype(np.float16)
 
 
 class TestElementwiseMaxOp_ZeroDim3(TestElementwiseOp):
@@ -135,10 +146,10 @@ class TestElementwiseMaxOp_ZeroDim3(TestElementwiseOp):
         self.y = np.random.uniform(0.1, 1, [13, 17]).astype("float64")
 
 
-class TestElementwiseMaxFP16Op_ZeroDim3(TestElementwiseOp):
+class TestElementwiseMaxFP16Op_ZeroDim3(TestElementwiseFP16Op):
     def init_data(self):
-        self.x = np.random.uniform(0.1, 1, []).astype("float16")
-        self.y = np.random.uniform(0.1, 1, [13, 17]).astype("float16")
+        self.x = np.random.uniform(0.1, 1, []).astype(np.float16)
+        self.y = np.random.uniform(0.1, 1, [13, 17]).astype(np.float16)
 
 
 @unittest.skipIf(
@@ -180,29 +191,11 @@ class TestElementwiseBF16Op(OpTest):
         if hasattr(self, 'attrs'):
             self.check_output(check_dygraph=False)
         else:
-            self.check_output()
+            self.check_output(check_dygraph=True)
 
     def test_check_grad_normal(self):
         if hasattr(self, 'attrs'):
             # check_prim=False, bfloat16 is not supported in `less_equal`
-            self.check_grad(['X', 'Y'], 'Out', check_dygraph=False)
-        else:
-            self.check_grad(['X', 'Y'], 'Out')
-
-    def test_check_grad_ingore_x(self):
-        self.check_grad(['Y'], 'Out', no_grad_set=set("X"))
-
-    def test_check_grad_ingore_y(self):
-        self.check_grad(['X'], 'Out', no_grad_set=set('Y'))
-
-
-class TestElementwiseMaxBF16Op_ZeroDim1(TestElementwiseBF16Op):
-    def init_data(self):
-        self.x = np.random.uniform(0.1, 1, []).astype("float32")
-        self.y = np.random.uniform(0.1, 1, []).astype("float32")
-
-    def test_check_grad_normal(self):
-        if hasattr(self, 'attrs'):
             self.check_grad(
                 ['X', 'Y'], 'Out', numeric_grad_delta=0.05, check_dygraph=False
             )
@@ -220,6 +213,12 @@ class TestElementwiseMaxBF16Op_ZeroDim1(TestElementwiseBF16Op):
         )
 
 
+class TestElementwiseMaxBF16Op_ZeroDim1(TestElementwiseBF16Op):
+    def init_data(self):
+        self.x = np.random.uniform(0.1, 1, []).astype("float32")
+        self.y = np.random.uniform(0.1, 1, []).astype("float32")
+
+
 class TestElementwiseMaxBF16Op_scalar(TestElementwiseBF16Op):
     def init_data(self):
         self.x = np.random.random_integers(-5, 5, [2, 3, 20]).astype("float32")
@@ -236,10 +235,13 @@ class TestElementwiseMaxOp_scalar(TestElementwiseOp):
         self.y = np.array([0.5]).astype("float64")
 
 
-class TestElementwiseMaxFP16Op_scalar(TestElementwiseMaxOp_scalar):
+@skip_check_grad_ci(
+    reason="[skip shape check] Use y_shape(1) to test broadcast."
+)
+class TestElementwiseMaxFP16Op_scalar(TestElementwiseFP16Op):
     def init_data(self):
-        self.x = np.random.random_integers(-5, 5, [2, 3, 20]).astype("float16")
-        self.y = np.array([0.5]).astype("float16")
+        self.x = np.random.random_integers(-5, 5, [2, 3, 20]).astype(np.float16)
+        self.y = np.array([0.5]).astype(np.float16)
 
 
 class TestElementwiseMaxOp_Vector(TestElementwiseOp):
@@ -251,12 +253,12 @@ class TestElementwiseMaxOp_Vector(TestElementwiseOp):
         )
 
 
-class TestElementwiseMaxFP16Op_Vector(TestElementwiseOp):
+class TestElementwiseMaxFP16Op_Vector(TestElementwiseFP16Op):
     def init_data(self):
-        self.x = np.random.random((100,)).astype("float16")
-        sgn = np.random.choice([-1, 1], (100,)).astype("float16")
+        self.x = np.random.random((100,)).astype(np.float16)
+        sgn = np.random.choice([-1, 1], (100,)).astype(np.float16)
         self.y = self.x + sgn * np.random.uniform(0.1, 1, (100,)).astype(
-            "float16"
+            np.float16
         )
 
 
@@ -289,12 +291,13 @@ class TestElementwiseMaxOp_broadcast_2(TestElementwiseOp):
         }
 
 
-class TestElementwiseMaxFP16Op_broadcast_2(TestElementwiseOp):
+class TestElementwiseMaxFP16Op_broadcast_2(TestElementwiseFP16Op):
     def setUp(self):
         self.op_type = "elementwise_max"
         self.python_api = paddle.maximum
         self.public_python_api = paddle.maximum
         self.prim_op_type = "prim"
+        self.dtype = np.float16
         x = np.random.uniform(0.5, 1, (1, 3, 100)).astype(np.float16)
         sgn = np.random.choice([-1, 1], (100,)).astype(np.float16)
         y = x[0, 0, :] + sgn * np.random.uniform(1, 2, (100,)).astype(
@@ -323,12 +326,13 @@ class TestElementwiseMaxOp_broadcast_4(TestElementwiseOp):
         self.outputs = {'Out': np.maximum(self.inputs['X'], self.inputs['Y'])}
 
 
-class TestElementwiseFP16Op_broadcast_4(TestElementwiseOp):
+class TestElementwiseFP16Op_broadcast_4(TestElementwiseFP16Op):
     def setUp(self):
         self.op_type = "elementwise_max"
         self.python_api = paddle.maximum
         self.public_python_api = paddle.maximum
         self.prim_op_type = "prim"
+        self.dtype = np.float16
         x = np.random.uniform(0.5, 1, (2, 3, 4, 5)).astype(np.float16)
         sgn = np.random.choice([-1, 1], (2, 3, 1, 5)).astype(np.float16)
         y = x + sgn * np.random.uniform(1, 2, (2, 3, 1, 5)).astype(np.float16)

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

@@ -15,10 +15,9 @@
 import unittest
 
 import numpy as np
-from eager_op_test import OpTest, skip_check_grad_ci
+from eager_op_test import OpTest, convert_float_to_uint16, skip_check_grad_ci
 
 import paddle
-from paddle import _legacy_C_ops, fluid
 from paddle.fluid import core
 
 paddle.enable_static()
@@ -61,6 +60,33 @@ class TestElementwiseOp(OpTest):
         )
 
 
+class TestElementwiseFP16Op(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_min"
+        self.python_api = paddle.minimum
+        self.dtype = np.float16
+        # If x and y have the same value, the min() is not differentiable.
+        # So we generate test data by the following method
+        # to avoid them being too close to each other.
+        x = np.random.uniform(0.1, 1, [13, 17]).astype(np.float16)
+        sgn = np.random.choice([-1, 1], [13, 17]).astype(np.float16)
+        y = x + sgn * np.random.uniform(0.1, 1, [13, 17]).astype(np.float16)
+        self.inputs = {'X': x, 'Y': y}
+        self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out')
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(['Y'], 'Out', no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(['X'], 'Out', no_grad_set=set('Y'))
+
+
 class TestElementwiseMinOp_ZeroDim1(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_min"
@@ -71,6 +97,12 @@ class TestElementwiseMinOp_ZeroDim1(TestElementwiseOp):
         self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
 
 
+class TestElementwiseMinFP16Op_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)
+
+
 class TestElementwiseMinOp_ZeroDim2(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_min"
@@ -81,6 +113,12 @@ class TestElementwiseMinOp_ZeroDim2(TestElementwiseOp):
         self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
 
 
+class TestElementwiseMinFP16Op_ZeroDim2(TestElementwiseFP16Op):
+    def init_data(self):
+        self.x = np.random.uniform(0.1, 1, [13, 17]).astype("float16")
+        self.y = np.random.uniform(0.1, 1, []).astype("float16")
+
+
 class TestElementwiseMinOp_ZeroDim3(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_min"
@@ -91,6 +129,12 @@ class TestElementwiseMinOp_ZeroDim3(TestElementwiseOp):
         self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
 
 
+class TestElementwiseMinFP16Op_ZeroDim3(TestElementwiseFP16Op):
+    def init_data(self):
+        self.x = np.random.uniform(0.1, 1, []).astype("float16")
+        self.y = np.random.uniform(0.1, 1, [13, 17]).astype("float16")
+
+
 @skip_check_grad_ci(
     reason="[skip shape check] Use y_shape(1) to test broadcast."
 )
@@ -104,6 +148,19 @@ class TestElementwiseMinOp_scalar(TestElementwiseOp):
         self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
 
 
+@skip_check_grad_ci(
+    reason="[skip shape check] Use y_shape(1) to test broadcast."
+)
+class TestElementwiseMinFP16Op_scalar(TestElementwiseFP16Op):
+    def setUp(self):
+        self.op_type = "elementwise_min"
+        self.python_api = paddle.minimum
+        x = np.random.random_integers(-5, 5, [10, 3, 4]).astype(np.float16)
+        y = np.array([0.5]).astype(np.float16)
+        self.inputs = {'X': x, 'Y': y}
+        self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
+
+
 class TestElementwiseMinOp_Vector(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_min"
@@ -115,6 +172,17 @@ class TestElementwiseMinOp_Vector(TestElementwiseOp):
         self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
 
 
+class TestElementwiseMinFP16Op_Vector(TestElementwiseFP16Op):
+    def setUp(self):
+        self.op_type = "elementwise_min"
+        self.python_api = paddle.minimum
+        x = np.random.random((100,)).astype(np.float16)
+        sgn = np.random.choice([-1, 1], (100,)).astype(np.float16)
+        y = x + sgn * np.random.uniform(0.1, 1, (100,)).astype(np.float16)
+        self.inputs = {'X': x, 'Y': y}
+        self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
+
+
 class TestElementwiseMinOp_broadcast_2(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_min"
@@ -133,6 +201,24 @@ class TestElementwiseMinOp_broadcast_2(TestElementwiseOp):
         }
 
 
+class TestElementwiseMinFP16Op_broadcast_2(TestElementwiseFP16Op):
+    def setUp(self):
+        self.op_type = "elementwise_min"
+        self.python_api = broadcast_wrapper(shape=[1, 1, 100])
+        x = np.random.uniform(0.5, 1, (2, 3, 100)).astype(np.float16)
+        sgn = np.random.choice([-1, 1], (100,)).astype(np.float16)
+        y = x[0, 0, :] + sgn * np.random.uniform(1, 2, (100,)).astype(
+            np.float16
+        )
+        self.inputs = {'X': x, 'Y': y}
+
+        self.outputs = {
+            'Out': np.minimum(
+                self.inputs['X'], self.inputs['Y'].reshape(1, 1, 100)
+            )
+        }
+
+
 class TestElementwiseMinOp_broadcast_4(TestElementwiseOp):
     def setUp(self):
         self.op_type = "elementwise_min"
@@ -145,52 +231,90 @@ class TestElementwiseMinOp_broadcast_4(TestElementwiseOp):
         self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
 
 
-class TestElementwiseMinOpFP16(unittest.TestCase):
-    def get_out_and_grad(self, x_np, y_np, axis, place, use_fp32=False):
-        assert x_np.dtype == np.float16
-        assert y_np.dtype == np.float16
-        if use_fp32:
-            x_np = x_np.astype(np.float32)
-            y_np = y_np.astype(np.float32)
-        dtype = np.float16
-
-        with fluid.dygraph.guard(place):
-            x = paddle.to_tensor(x_np)
-            y = paddle.to_tensor(y_np)
-            x.stop_gradient = False
-            y.stop_gradient = False
-            z = _legacy_C_ops.elementwise_min(x, y, 'axis', axis)
-            x_g, y_g = paddle.grad([z], [x, y])
-            return (
-                z.numpy().astype(dtype),
-                x_g.numpy().astype(dtype),
-                y_g.numpy().astype(dtype),
+class TestElementwiseMinFP16Op_broadcast_4(TestElementwiseFP16Op):
+    def setUp(self):
+        self.op_type = "elementwise_min"
+        self.python_api = paddle.minimum
+        x = np.random.uniform(0.5, 1, (2, 10, 2, 5)).astype(np.float16)
+        sgn = np.random.choice([-1, 1], (2, 10, 1, 5)).astype(np.float16)
+        y = x + sgn * np.random.uniform(1, 2, (2, 10, 1, 5)).astype(np.float16)
+        self.inputs = {'X': x, 'Y': y}
+
+        self.outputs = {'Out': np.minimum(self.inputs['X'], self.inputs['Y'])}
+
+
+@unittest.skipIf(
+    core.is_compiled_with_cuda()
+    and (
+        core.cudnn_version() < 8100
+        or paddle.device.cuda.get_device_capability()[0] < 8
+    ),
+    "run test when gpu is availble and the minimum cudnn version is 8.1.0 and gpu's compute capability is at least 8.0.",
+)
+class TestElementwiseBF16Op(OpTest):
+    def init_data(self):
+        # If x and y have the same value, the max() is not differentiable.
+        # So we generate test data by the following method
+        # to avoid them being too close to each other.
+        self.x = np.random.uniform(0.1, 1, [13, 17]).astype(np.float32)
+        sgn = np.random.choice([-1, 1], [13, 17]).astype(np.float32)
+        self.y = self.x + sgn * np.random.uniform(0.1, 1, [13, 17]).astype(
+            np.float32
         )
 
-    def check_main(self, x_shape, y_shape, axis=-1):
-        if not paddle.is_compiled_with_cuda():
-            return
-        place = paddle.CUDAPlace(0)
-        if not core.is_float16_supported(place):
-            return
+    def setUp(self):
+        self.init_data()
+        self.op_type = "elementwise_min"
+        self.python_api = paddle.minimum
+        self.public_python_api = paddle.minimum
+        self.prim_op_type = "prim"
+        self.enable_cinn = False
+        self.dtype = np.uint16
+        self.inputs = {
+            'X': convert_float_to_uint16(self.x),
+            'Y': convert_float_to_uint16(self.y),
+        }
+        self.outputs = {
+            'Out': convert_float_to_uint16(np.minimum(self.x, self.y))
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', numeric_grad_delta=0.05)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', numeric_grad_delta=0.05, no_grad_set=set("X")
+        )
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', numeric_grad_delta=0.05, no_grad_set=set('Y')
+        )
+
+
+class TestElementwiseMinBF16Op_ZeroDim1(TestElementwiseBF16Op):
+    def init_data(self):
+        self.x = np.random.uniform(0.1, 1, []).astype("float32")
+        self.y = np.random.uniform(0.1, 1, []).astype("float32")
+
+
+class TestElementwiseMinBF16Op_scalar(TestElementwiseBF16Op):
+    def init_data(self):
+        self.x = np.random.random_integers(-5, 5, [2, 3, 20]).astype("float32")
+        self.y = np.array([0.5]).astype("float32")
+        self.__class__.no_need_check_grad = True
 
-        x_np = np.random.random(size=x_shape).astype(np.float16)
-        y_np = np.random.random(size=y_shape).astype(np.float16)
 
-        z_1, x_g_1, y_g_1 = self.get_out_and_grad(
-            x_np, y_np, axis, place, False
+class TestElementwiseMinBF16Op_Vector(TestElementwiseBF16Op):
+    def init_data(self):
+        self.x = np.random.random((100,)).astype("float32")
+        sgn = np.random.choice([-1, 1], (100,)).astype("float32")
+        self.y = self.x + sgn * np.random.uniform(0.1, 1, (100,)).astype(
+            "float32"
         )
-        z_2, x_g_2, y_g_2 = self.get_out_and_grad(x_np, y_np, axis, place, True)
-        np.testing.assert_array_equal(z_1, z_2)
-        np.testing.assert_array_equal(x_g_1, x_g_2)
-        np.testing.assert_array_equal(y_g_1, y_g_2)
-
-    def test_main(self):
-        self.check_main((13, 17), (13, 17))
-        self.check_main((10, 3, 4), (1,))
-        self.check_main((100,), (100,))
-        self.check_main((2, 3, 100), (100,))
-        self.check_main((2, 10, 2, 5), (2, 10, 1, 5))
 
 
 if __name__ == '__main__':

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

@@ -156,10 +156,20 @@ class TestBF16ElementwiseMulOp(OpTest):
         self.check_grad(['X', 'Y'], 'Out', check_prim=True)
 
     def test_check_grad_ingore_x(self):
-        self.check_grad(['Y'], 'Out', no_grad_set=set("X"), check_prim=True)
+        self.check_grad(
+            ['Y'],
+            'Out',
+            no_grad_set=set("X"),
+            check_prim=True,
+        )
 
     def test_check_grad_ingore_y(self):
-        self.check_grad(['X'], 'Out', no_grad_set=set('Y'), check_prim=True)
+        self.check_grad(
+            ['X'],
+            'Out',
+            no_grad_set=set('Y'),
+            check_prim=True,
+        )
 
     def if_enable_cinn(self):
         self.enable_cinn = False
@@ -357,6 +367,39 @@ class TestElementwiseMulOpFp16(ElementwiseMulOp):
     def if_enable_cinn(self):
         pass
 
+    def test_check_output(self):
+        # TODO(wangzhongpu): support mkldnn op in dygraph mode
+        self.check_output(check_dygraph=(not self.use_mkldnn))
+
+    def test_check_grad_normal(self):
+        # TODO(wangzhongpu): support mkldnn op in dygraph mode
+        self.check_grad(
+            ['X', 'Y'],
+            'Out',
+            check_dygraph=(not self.use_mkldnn),
+            check_prim=True,
+        )
+
+    def test_check_grad_ingore_x(self):
+        # TODO(wangzhongpu): support mkldnn op in dygraph mode
+        self.check_grad(
+            ['Y'],
+            'Out',
+            no_grad_set=set("X"),
+            check_dygraph=(not self.use_mkldnn),
+            check_prim=True,
+        )
+
+    def test_check_grad_ingore_y(self):
+        # TODO(wangzhongpu): support mkldnn op in dygraph mode
+        self.check_grad(
+            ['X'],
+            'Out',
+            no_grad_set=set('Y'),
+            check_dygraph=(not self.use_mkldnn),
+            check_prim=True,
+        )
+
 
 class TestElementwiseMulOp_commonuse_1(ElementwiseMulOp):
     def setUp(self):