diff --git a/paddle/fluid/operators/dot_op.h b/paddle/fluid/operators/dot_op.h
index 0b0b7f69b9d849a666803b8c2e7f57f684c7ab9e..1b607922eda1d854567a338b51121e47064915e4 100644
--- a/paddle/fluid/operators/dot_op.h
+++ b/paddle/fluid/operators/dot_op.h
@@ -160,7 +160,7 @@ struct DotGradFunction<DeviceContext, T, math::DisableComplex<T>> {
                   const Tensor* tensor_dout, Tensor* tensor_dx,
                   Tensor* tensor_dy,
                   const paddle::framework::ExecutionContext& ctx) {
-#ifdef __NVCC__
+#if defined(__NVCC__) || defined(__HIPCC__)
     if (1 == tensor_dout->dims().size()) {
       auto dout = framework::EigenVector<T>::Flatten(*tensor_dout);
 
diff --git a/python/paddle/fluid/tests/unittests/test_matmul_v2_op.py b/python/paddle/fluid/tests/unittests/test_matmul_v2_op.py
index 761d318d7b8a3d43897f31bf635884b582fcec1d..efcc0e4cfe323294df88167a6100f019cef67005 100644
--- a/python/paddle/fluid/tests/unittests/test_matmul_v2_op.py
+++ b/python/paddle/fluid/tests/unittests/test_matmul_v2_op.py
@@ -67,7 +67,7 @@ class TestMatMulV2Op(OpTest):
         self.trans_y = False
 
     def init_kernel_type(self):
-        self.dtype = "float64"
+        self.dtype = "float32" if core.is_compiled_with_rocm() else "float64"
 
     def setUp(self):
         self.init_kernel_type()
@@ -91,7 +91,10 @@ class TestMatMulV2Op(OpTest):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(['X', 'Y'], 'Out')
+        if core.is_compiled_with_rocm():
+            self.check_grad(['X', 'Y'], 'Out', max_relative_error=1e-2)
+        else:
+            self.check_grad(['X', 'Y'], 'Out')
 
 
 class TestMatMuklOp2(TestMatMulV2Op):