diff --git a/paddle/fluid/operators/elementwise/elementwise_add_op.cu b/paddle/fluid/operators/elementwise/elementwise_add_op.cu
index de121b3786f3c9cc94a0c3dab789f372d9e17e72..f83330fbe5379c5e9733f7a74fe1e756b9070161 100644
--- a/paddle/fluid/operators/elementwise/elementwise_add_op.cu
+++ b/paddle/fluid/operators/elementwise/elementwise_add_op.cu
@@ -40,8 +40,9 @@ struct SameDimsElemwiseAdd<platform::CUDADeviceContext, platform::float16> {
                   const framework::Tensor* x, const framework::Tensor* y,
                   framework::Tensor* z) {
     auto size = x->numel();
-    dim3 gird_size = dim3(
-        (size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
+    dim3 gird_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
+                              PADDLE_CUDA_THREAD_SIZE,
+                          1);
     dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
     const half* x2 =
         reinterpret_cast<const half*>(x->data<platform::float16>());
diff --git a/paddle/fluid/operators/elementwise/elementwise_div_op.cu b/paddle/fluid/operators/elementwise/elementwise_div_op.cu
index b1698491180d773dc35d09d8d7ec642d3ed914fe..787ff374228e821439aac4fb2703584c391dd1a4 100644
--- a/paddle/fluid/operators/elementwise/elementwise_div_op.cu
+++ b/paddle/fluid/operators/elementwise/elementwise_div_op.cu
@@ -41,8 +41,9 @@ struct SameDimsElemwiseDiv<platform::CUDADeviceContext, platform::float16> {
                   const framework::Tensor* x, const framework::Tensor* y,
                   framework::Tensor* z) {
     auto size = x->numel();
-    dim3 gird_size = dim3(
-        (size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
+    dim3 gird_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
+                              PADDLE_CUDA_THREAD_SIZE,
+                          1);
     dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
     const half* x2 =
         reinterpret_cast<const half*>(x->data<platform::float16>());
diff --git a/paddle/fluid/operators/elementwise/elementwise_mul_op.cu b/paddle/fluid/operators/elementwise/elementwise_mul_op.cu
index 4814cb144f057d4cb76b416e896e24ea227e92a2..efda1db2969684ace7ce980a22d15b98ceba9f35 100644
--- a/paddle/fluid/operators/elementwise/elementwise_mul_op.cu
+++ b/paddle/fluid/operators/elementwise/elementwise_mul_op.cu
@@ -41,8 +41,9 @@ struct SameDimsElemwiseMul<platform::CUDADeviceContext, platform::float16> {
                   const framework::Tensor* x, const framework::Tensor* y,
                   framework::Tensor* z) {
     auto size = x->numel();
-    dim3 gird_size = dim3(
-        (size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
+    dim3 gird_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
+                              PADDLE_CUDA_THREAD_SIZE,
+                          1);
     dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
     const half* x2 =
         reinterpret_cast<const half*>(x->data<platform::float16>());
diff --git a/paddle/fluid/operators/elementwise/elementwise_sub_op.cu b/paddle/fluid/operators/elementwise/elementwise_sub_op.cu
index 7ff72028091ed78e7ca5c27d2b8bb362c12fd152..5df7abcf5a9c762b7a275e0ef8eae908b07b31c2 100644
--- a/paddle/fluid/operators/elementwise/elementwise_sub_op.cu
+++ b/paddle/fluid/operators/elementwise/elementwise_sub_op.cu
@@ -41,8 +41,9 @@ struct SameDimsElemwiseSub<platform::CUDADeviceContext, platform::float16> {
                   const framework::Tensor* x, const framework::Tensor* y,
                   framework::Tensor* z) {
     auto size = x->numel();
-    dim3 gird_size = dim3(
-        (size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
+    dim3 gird_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
+                              PADDLE_CUDA_THREAD_SIZE,
+                          1);
     dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
     const half* x2 =
         reinterpret_cast<const half*>(x->data<platform::float16>());