Improve performance of elementwise_add grad op (#29187)

* pass stop_gradient for cast op

* improve performance of elementwise_add grad

* use tensor copy async

* dygraph branch

* fix dygraph branch

* add ut

paddle/fluid/operators/elementwise/elementwise_add_op.h

@@ -144,7 +144,20 @@ class ElementwiseAddGradKernel : public ElemwiseGradKernel<T> {
     // skip out
     auto *out = dout;
 
-    if (dx != nullptr && dy != nullptr && (dx->dims() == dy->dims())) {
+    // Special case when dy is not needed and dx doesn't reduce
+    if (dx != nullptr && dy == nullptr && dx->dims() == dout->dims()) {
+      VLOG(4) << "Special case when dy is not needed and dx doesn't "
+                 "reduce";
+      framework::TensorCopy(
+          *dout, ctx.GetPlace(),
+          ctx.template device_context<platform::DeviceContext>(), dx);
+    } else if (dx == nullptr && dy != nullptr && dy->dims() == dout->dims()) {
+      VLOG(4) << "Special case when dx is not needed and dy doesn't "
+                 "reduce";
+      framework::TensorCopy(
+          *dout, ctx.GetPlace(),
+          ctx.template device_context<platform::DeviceContext>(), dy);
+    } else if (dx != nullptr && dy != nullptr && (dx->dims() == dy->dims())) {
       elementwise_add_grad<DeviceContext, T>(ctx, x, y, out, dout, dx, dy);
     } else {
       default_elementwise_add_grad<DeviceContext, T>(ctx, x, y, out, dout, dx,

python/paddle/fluid/layers/math_op_patch.py

@@ -179,6 +179,7 @@ def monkey_patch_variable():
             outputs={"Out": [out]},
             attrs={"in_dtype": self.dtype,
                    "out_dtype": out.dtype})
+        out.stop_gradient = self.stop_gradient
         return out
 
     def _scalar_op_(var, scale, bias):

python/paddle/fluid/layers/tensor.py

@@ -224,6 +224,11 @@ def cast(x, dtype):
             x = paddle.to_tensor([2, 3, 4], 'float64')
             y = paddle.cast(x, 'uint8')
     """
+    if in_dygraph_mode():
+        if not isinstance(dtype, core.VarDesc.VarType):
+            dtype = convert_np_dtype_to_dtype_(dtype)
+        out = core.ops.cast(x, 'in_dtype', x.dtype, 'out_dtype', dtype)
+
     check_variable_and_dtype(
         x, 'x',
         ['bool', 'float16', 'float32', 'float64', 'int32', 'int64', 'uint8'],
@@ -234,7 +239,8 @@ def cast(x, dtype):
     ], 'cast')
 
     helper = LayerHelper('cast', **locals())
-    out = helper.create_variable_for_type_inference(dtype=dtype)
+    out = helper.create_variable_for_type_inference(
+        dtype=dtype, stop_gradient=x.stop_gradient)
     helper.append_op(
         type='cast',
         inputs={'X': [x]},

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

@@ -257,6 +257,19 @@ class TestMathOpPatches(unittest.TestCase):
                        fetch_list=[b])
         self.assertTrue(numpy.allclose(-a_np, b_np))
 
+    @prog_scope()
+    def test_astype(self):
+        a = fluid.layers.data(name="a", shape=[10, 1])
+        b = a.astype('float32')
+        place = fluid.CPUPlace()
+        exe = fluid.Executor(place)
+        a_np = numpy.random.uniform(-1, 1, size=[10, 1]).astype('float64')
+
+        b_np = exe.run(fluid.default_main_program(),
+                       feed={"a": a_np},
+                       fetch_list=[b])
+        self.assertTrue(numpy.allclose(a_np.astype('float32'), b_np))
+
 
 if __name__ == '__main__':
     unittest.main()