diff --git a/paddle/fluid/operators/elementwise_add_op.cu b/paddle/fluid/operators/elementwise_add_op.cu
index 19dc4a52152e2a7aa71476d4f0ef692d0af97b4a..dfff518f170b56d180b6883c363effb8dbd677b6 100644
--- a/paddle/fluid/operators/elementwise_add_op.cu
+++ b/paddle/fluid/operators/elementwise_add_op.cu
@@ -14,19 +14,20 @@ limitations under the License. */
 
 #define EIGEN_USE_GPU
 #include "paddle/fluid/operators/elementwise_add_op.h"
+#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 
 REGISTER_OP_CUDA_KERNEL(
-    elementwise_add,
-    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, double>,
-    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, int>,
-    ops::ElementwiseAddKernel<paddle::platform::CUDADeviceContext, int64_t>);
+    elementwise_add, ops::ElementwiseAddKernel<plat::CUDADeviceContext, float>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, double>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int64_t>,
+    ops::ElementwiseAddKernel<plat::CUDADeviceContext, plat::float16>);
 REGISTER_OP_CUDA_KERNEL(
     elementwise_add_grad,
-    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext, double>,
-    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext, int>,
-    ops::ElementwiseAddGradKernel<paddle::platform::CUDADeviceContext,
-                                  int64_t>);
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, float>,
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, double>,
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int>,
+    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int64_t>);
diff --git a/paddle/fluid/platform/float16.h b/paddle/fluid/platform/float16.h
index d3312a47f479160439d720dd993ee25a56d732fe..2cf311c7e56a9bbb0bdb0078d5cfefb4bb50018b 100644
--- a/paddle/fluid/platform/float16.h
+++ b/paddle/fluid/platform/float16.h
@@ -600,7 +600,7 @@ HOSTDEVICE inline bool operator>=(const float16& a, const float16& b) {
 
 // Arithmetic operators for float16 on ARMv8.2-A CPU
 #elif defined(PADDLE_WITH_NATIVE_FP16)
-HOST inline float16 operator+(const float16& a, const float16& b) {
+inline float16 operator+(const float16& a, const float16& b) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -616,7 +616,7 @@ HOST inline float16 operator+(const float16& a, const float16& b) {
   return res;
 }
 
-HOST inline float16 operator-(const float16& a, const float16& b) {
+inline float16 operator-(const float16& a, const float16& b) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -632,7 +632,7 @@ HOST inline float16 operator-(const float16& a, const float16& b) {
   return res;
 }
 
-HOST inline float16 operator*(const float16& a, const float16& b) {
+inline float16 operator*(const float16& a, const float16& b) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -648,7 +648,7 @@ HOST inline float16 operator*(const float16& a, const float16& b) {
   return res;
 }
 
-HOST inline float16 operator/(const float16& a, const float16& b) {
+inline float16 operator/(const float16& a, const float16& b) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -664,7 +664,7 @@ HOST inline float16 operator/(const float16& a, const float16& b) {
   return res;
 }
 
-HOST inline float16 operator-(const float16& a) {
+inline float16 operator-(const float16& a) {
   float16 res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -679,27 +679,27 @@ HOST inline float16 operator-(const float16& a) {
   return res;
 }
 
-HOST inline float16& operator+=(float16& a, const float16& b) {
+inline float16& operator+=(float16& a, const float16& b) {
   a = a + b;
   return a;
 }
 
-HOST inline float16& operator-=(float16& a, const float16& b) {
+inline float16& operator-=(float16& a, const float16& b) {
   a = a - b;
   return a;
 }
 
-HOST inline float16& operator*=(float16& a, const float16& b) {
+inline float16& operator*=(float16& a, const float16& b) {
   a = a * b;
   return a;
 }
 
-HOST inline float16& operator/=(float16& a, const float16& b) {
+inline float16& operator/=(float16& a, const float16& b) {
   a = a / b;
   return a;
 }
 
-HOST inline bool operator==(const float16& a, const float16& b) {
+inline bool operator==(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -715,11 +715,9 @@ HOST inline bool operator==(const float16& a, const float16& b) {
   return (res & 0xffff) != 0;
 }
 
-HOST inline bool operator!=(const float16& a, const float16& b) {
-  return !(a == b);
-}
+inline bool operator!=(const float16& a, const float16& b) { return !(a == b); }
 
-HOST inline bool operator<(const float16& a, const float16& b) {
+inline bool operator<(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v1.h}[0], [%[a_ptr]]\n"
@@ -735,7 +733,7 @@ HOST inline bool operator<(const float16& a, const float16& b) {
   return (res & 0xffff) != 0;
 }
 
-HOST inline bool operator<=(const float16& a, const float16& b) {
+inline bool operator<=(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v1.h}[0], [%[a_ptr]]\n"
@@ -751,7 +749,7 @@ HOST inline bool operator<=(const float16& a, const float16& b) {
   return (res & 0xffff) != 0;
 }
 
-HOST inline bool operator>(const float16& a, const float16& b) {
+inline bool operator>(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -767,7 +765,7 @@ HOST inline bool operator>(const float16& a, const float16& b) {
   return (res & 0xffff) != 0;
 }
 
-HOST inline bool operator>=(const float16& a, const float16& b) {
+inline bool operator>=(const float16& a, const float16& b) {
   uint16_t res;
   asm volatile(
       "ld1 {v0.h}[0], [%[a_ptr]]\n"
@@ -785,69 +783,69 @@ HOST inline bool operator>=(const float16& a, const float16& b) {
 
 // Arithmetic operators for float16, software emulated on other CPU
 #else
-HOST inline float16 operator+(const float16& a, const float16& b) {
+inline float16 operator+(const float16& a, const float16& b) {
   return float16(float(a) + float(b));
 }
 
-HOST inline float16 operator-(const float16& a, const float16& b) {
+inline float16 operator-(const float16& a, const float16& b) {
   return float16(float(a) - float(b));
 }
 
-HOST inline float16 operator*(const float16& a, const float16& b) {
+inline float16 operator*(const float16& a, const float16& b) {
   return float16(float(a) * float(b));
 }
 
-HOST inline float16 operator/(const float16& a, const float16& b) {
+inline float16 operator/(const float16& a, const float16& b) {
   return float16(float(a) / float(b));
 }
 
-HOST inline float16 operator-(const float16& a) {
+inline float16 operator-(const float16& a) {
   float16 res;
   res.x = a.x ^ 0x8000;
   return res;
 }
 
-HOST inline float16& operator+=(float16& a, const float16& b) {
+inline float16& operator+=(float16& a, const float16& b) {
   a = float16(float(a) + float(b));
   return a;
 }
 
-HOST inline float16& operator-=(float16& a, const float16& b) {
+inline float16& operator-=(float16& a, const float16& b) {
   a = float16(float(a) - float(b));
   return a;
 }
 
-HOST inline float16& operator*=(float16& a, const float16& b) {
+inline float16& operator*=(float16& a, const float16& b) {
   a = float16(float(a) * float(b));
   return a;
 }
 
-HOST inline float16& operator/=(float16& a, const float16& b) {
+inline float16& operator/=(float16& a, const float16& b) {
   a = float16(float(a) / float(b));
   return a;
 }
 
-HOST inline bool operator==(const float16& a, const float16& b) {
+inline bool operator==(const float16& a, const float16& b) {
   return float(a) == float(b);
 }
 
-HOST inline bool operator!=(const float16& a, const float16& b) {
+inline bool operator!=(const float16& a, const float16& b) {
   return float(a) != float(b);
 }
 
-HOST inline bool operator<(const float16& a, const float16& b) {
+inline bool operator<(const float16& a, const float16& b) {
   return float(a) < float(b);
 }
 
-HOST inline bool operator<=(const float16& a, const float16& b) {
+inline bool operator<=(const float16& a, const float16& b) {
   return float(a) <= float(b);
 }
 
-HOST inline bool operator>(const float16& a, const float16& b) {
+inline bool operator>(const float16& a, const float16& b) {
   return float(a) > float(b);
 }
 
-HOST inline bool operator>=(const float16& a, const float16& b) {
+inline bool operator>=(const float16& a, const float16& b) {
   return float(a) >= float(b);
 }
 #endif
diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py
index 5b2384e94d788342c692fcb8e33f3a2ff663ab53..1f52bd90d0d49bda6c180019e90ebd923c91439c 100644
--- a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py
+++ b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py
@@ -13,158 +13,243 @@
 # limitations under the License.
 import unittest
 import numpy as np
+import paddle.fluid.core as core
 from op_test import OpTest
 
 
-class TestElementwiseOp(OpTest):
+class TestElementwiseAddOp(OpTest):
     def setUp(self):
         self.op_type = "elementwise_add"
+        self.dtype = np.float32
+        self.axis = -1
+        self.init_dtype()
+        self.init_input_output()
+        self.init_axis()
+
         self.inputs = {
-            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
-            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
+            'X': OpTest.np_dtype_to_fluid_dtype(self.x),
+            'Y': OpTest.np_dtype_to_fluid_dtype(self.y)
         }
-        self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['Y'])}
+        self.attrs = {'axis': self.axis}
+        self.outputs = {'Out': self.out}
 
     def test_check_output(self):
         self.check_output()
 
     def test_check_grad_normal(self):
+        if self.dtype == np.float16:
+            return
         self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.005)
 
     def test_check_grad_ingore_x(self):
+        if self.dtype == np.float16:
+            return
         self.check_grad(
             ['Y'], 'Out', max_relative_error=0.005, no_grad_set=set("X"))
 
     def test_check_grad_ingore_y(self):
+        if self.dtype == np.float16:
+            return
         self.check_grad(
             ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
 
+    def init_input_output(self):
+        self.x = np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype)
+        self.y = np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype)
+        self.out = np.add(self.x, self.y)
 
-class TestElementwiseAddOp_scalar(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(1).astype(np.float32)
-        }
-        self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
+    def init_dtype(self):
+        pass
 
+    def init_axis(self):
+        pass
 
-class TestElementwiseAddOp_scalar2(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(1, 1).astype(np.float32)
-        }
-        self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
 
+class TestFP16ElementwiseAddOp(TestElementwiseAddOp):
+    def init_dtype(self):
+        self.dtype = np.float16
 
-class TestElementwiseAddOp_Vector(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.random((32, )).astype("float32"),
-            'Y': np.random.random((32, )).astype("float32")
-        }
-        self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['Y'])}
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                self.check_output_with_place(place, atol=1e-3)
 
 
-class TestElementwiseAddOp_broadcast_0(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(2).astype(np.float32)
-        }
+class TestElementwiseAddOp_scalar(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(1).astype(self.dtype)
+        self.out = self.x + self.y
 
-        self.attrs = {'axis': 0}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(2, 1, 1)
-        }
 
+class TestFP16ElementwiseAddOp_scalar(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(1).astype(self.dtype)
+        self.out = self.x + self.y
 
-class TestElementwiseAddOp_broadcast_1(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(3).astype(np.float32)
-        }
 
-        self.attrs = {'axis': 1}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 3, 1)
-        }
+class TestElementwiseAddOp_scalar2(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(1, 1).astype(self.dtype)
+        self.out = self.x + self.y
 
 
-class TestElementwiseAddOp_broadcast_2(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(4).astype(np.float32)
-        }
+class TestFP16ElementwiseAddOp_scalar2(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(1, 1).astype(self.dtype)
+        self.out = self.x + self.y
 
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 1, 4)
-        }
 
+class TestElementwiseAddOp_Vector(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.random((32, )).astype(self.dtype)
+        self.y = np.random.random((32, )).astype(self.dtype)
+        self.out = np.add(self.x, self.y)
 
-class TestElementwiseAddOp_broadcast_3(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4, 5).astype(np.float32),
-            'Y': np.random.rand(3, 4).astype(np.float32)
-        }
 
-        self.attrs = {'axis': 1}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 3, 4, 1)
-        }
+class TestFP16ElementwiseAddOp_Vector(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.random((32, )).astype(self.dtype)
+        self.y = np.random.random((32, )).astype(self.dtype)
+        self.out = np.add(self.x, self.y)
 
 
-class TestElementwiseAddOp_broadcast_4(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4, 5).astype(np.float32),
-            'Y': np.random.rand(2, 1).astype(np.float32)
-        }
+class TestElementwiseAddOp_broadcast_0(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(2).astype(self.dtype)
+        self.out = self.x + self.y.reshape(2, 1, 1)
 
-        self.attrs = {'axis': 0}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(2, 1, 1, 1)
-        }
+    def init_axis(self):
+        self.axis = 0
 
 
-class TestElementwiseAddOp_rowwise_add_0(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 3, 4).astype(np.float32),
-            'Y': np.random.rand(3, 4).astype(np.float32)
-        }
+class TestFP16ElementwiseAddOp_broadcast_0(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(2).astype(self.dtype)
+        self.out = self.x + self.y.reshape(2, 1, 1)
 
-        self.attrs = {'axis': 1}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 3, 4)
-        }
+    def init_axis(self):
+        self.axis = 0
 
 
-class TestElementwiseAddOp_rowwise_add_1(TestElementwiseOp):
-    def setUp(self):
-        self.op_type = "elementwise_add"
-        self.inputs = {
-            'X': np.random.rand(2, 1).astype(np.float32),
-            'Y': np.random.rand(1).astype(np.float32)
-        }
+class TestElementwiseAddOp_broadcast_1(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(3).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 1)
 
-        self.attrs = {'axis': 1}
-        self.outputs = {
-            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 1)
-        }
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestFP16ElementwiseAddOp_broadcast_1(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(3).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 1)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestElementwiseAddOp_broadcast_2(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 1, 4)
+
+
+class TestFP16ElementwiseAddOp_broadcast_2(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 1, 4)
+
+
+class TestElementwiseAddOp_broadcast_3(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype)
+        self.y = np.random.rand(3, 4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 4, 1)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestFP16ElementwiseAddOp_broadcast_3(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype)
+        self.y = np.random.rand(3, 4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 4, 1)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestElementwiseAddOp_broadcast_4(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype)
+        self.y = np.random.rand(2, 1).astype(self.dtype)
+        self.out = self.x + self.y.reshape(2, 1, 1, 1)
+
+    def init_axis(self):
+        self.axis = 0
+
+
+class TestFP16ElementwiseAddOp_broadcast_4(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4, 5).astype(self.dtype)
+        self.y = np.random.rand(2, 1).astype(self.dtype)
+        self.out = self.x + self.y.reshape(2, 1, 1, 1)
+
+    def init_axis(self):
+        self.axis = 0
+
+
+class TestElementwiseAddOp_rowwise_add_0(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(3, 4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 4)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestFP16ElementwiseAddOp_rowwise_add_0(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 3, 4).astype(self.dtype)
+        self.y = np.random.rand(3, 4).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 3, 4)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestElementwiseAddOp_rowwise_add_1(TestElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 1).astype(self.dtype)
+        self.y = np.random.rand(1).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 1)
+
+    def init_axis(self):
+        self.axis = 1
+
+
+class TestFP16ElementwiseAddOp_rowwise_add_1(TestFP16ElementwiseAddOp):
+    def init_input_output(self):
+        self.x = np.random.rand(2, 1).astype(self.dtype)
+        self.y = np.random.rand(1).astype(self.dtype)
+        self.out = self.x + self.y.reshape(1, 1)
+
+    def init_axis(self):
+        self.axis = 1
 
 
 if __name__ == '__main__':