Merge pull request #6324 from chengduoZH/profiling/refine_elementwise_add

Refine code（operator++）

paddle/operators/elementwise_add_op.h

@@ -21,7 +21,7 @@ namespace operators {
 
 template <typename T>
 struct AddFunctor {
-  HOSTDEVICE T operator()(T a, T b) const { return a + b; }
+  inline HOSTDEVICE T operator()(T a, T b) const { return a + b; }
 };
 
 template <typename Place, typename T>

paddle/operators/elementwise_op_function.h

@@ -71,7 +71,9 @@ class RowwiseTransformIterator<T, platform::CPUPlace> {
 
   RowwiseTransformIterator<T, platform::CPUPlace>& operator++() {
     ++i_;
-    i_ %= n_;
+    if (UNLIKELY(i_ == n_)) {
+      i_ = 0;
+    }
     return *this;
   }
 
@@ -100,7 +102,12 @@ class MidWiseTransformIterator<T, platform::CPUPlace> {
       : ptr_(ptr), i_(0), j_(0), n_(n), post_(post) {}
 
   MidWiseTransformIterator<T, platform::CPUPlace>& operator++() {
-    i_ = (++j_ / post_) % n_;
+    ++j_;
+    i_ = j_ / post_;
+    if (UNLIKELY(i_ == n_)) {
+      j_ = 0;
+      i_ = 0;
+    }
     return *this;
   }