MKL elementwise add: default implementation used for integral types, float16 and/or GPU

paddle/fluid/operators/elementwise_add_op.h

@@ -36,9 +36,12 @@ void default_elementwise_add(const framework::ExecutionContext& ctx,
 }
 
 template <typename DeviceContext, typename T>
-typename std::enable_if<std::is_floating_point<T>::value>::type elementwise_add(
-    const framework::ExecutionContext& ctx, const framework::Tensor* x,
-    const framework::Tensor* y, framework::Tensor* z) {
+typename std::enable_if<
+    std::is_floating_point<T>::value &&
+    std::is_same<DeviceContext, platform::CPUDeviceContext>::value>::type
+elementwise_add(const framework::ExecutionContext& ctx,
+                const framework::Tensor* x, const framework::Tensor* y,
+                framework::Tensor* z) {
   auto eigen_x = framework::EigenVector<T>::Flatten(*x);
   auto eigen_y = framework::EigenVector<T>::Flatten(*y);
   auto eigen_z = framework::EigenVector<T>::Flatten(*z);
@@ -48,9 +51,12 @@ typename std::enable_if<std::is_floating_point<T>::value>::type elementwise_add(
 }
 
 template <typename DeviceContext, typename T>
-typename std::enable_if<std::is_integral<T>::value>::type elementwise_add(
-    const framework::ExecutionContext& ctx, const framework::Tensor* x,
-    const framework::Tensor* y, framework::Tensor* z) {
+typename std::enable_if<
+    !std::is_floating_point<T>::value ||
+    !std::is_same<DeviceContext, platform::CPUDeviceContext>::value>::type
+elementwise_add(const framework::ExecutionContext& ctx,
+                const framework::Tensor* x, const framework::Tensor* y,
+                framework::Tensor* z) {
   default_elementwise_add<DeviceContext, T>(ctx, x, y, z);
 }
 
@@ -66,7 +72,7 @@ class ElementwiseAddKernel : public framework::OpKernel<T> {
     z->mutable_data<T>(ctx.GetPlace());
 
     auto dims_equal = x->dims() == y->dims();
-    if (platform::is_cpu_place(ctx.GetPlace()) && dims_equal) {
+    if (dims_equal) {
       elementwise_add<DeviceContext, T>(ctx, x, y, z);
     } else {
       default_elementwise_add<DeviceContext, T>(ctx, x, y, z);