[cherry-pick] fix elementwise mod (#21315)

* fix elementwise_mod FP kernel. test=develop

* fix unittest. test=develop

paddle/fluid/operators/elementwise/elementwise_mod_op.h

@@ -29,7 +29,9 @@ struct ModFunctor {
 
 template <typename T>
 struct ModFunctorFP {
-  inline HOSTDEVICE T operator()(T a, T b) const { return std::fmod(a, b); }
+  inline HOSTDEVICE T operator()(T a, T b) const {
+    return fmod(b + fmod(a, b), b);
+  }
 };
 
 template <typename DeviceContext, typename T>

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

@@ -71,7 +71,7 @@ class TestElementwiseModOpFloat(TestElementwiseModOp):
     def init_input_output(self):
         self.x = np.random.uniform(-1000, 1000, [10, 10]).astype(self.dtype)
         self.y = np.random.uniform(-100, 100, [10, 10]).astype(self.dtype)
-        self.out = np.fmod(self.x, self.y)
+        self.out = np.fmod(self.y + np.fmod(self.x, self.y), self.y)
 
     def test_check_output(self):
         self.check_output(atol=2e-5)