diff --git a/paddle/fluid/operators/batch_norm_op_xpu.cc b/paddle/fluid/operators/batch_norm_op_xpu.cc
index 8499d1cdcd6466812e43fc08f8e66591214b9829..d232891f3d68408d960167e3d6a5c703a0f62309 100644
--- a/paddle/fluid/operators/batch_norm_op_xpu.cc
+++ b/paddle/fluid/operators/batch_norm_op_xpu.cc
@@ -15,6 +15,8 @@ limitations under the License. */
 #ifdef PADDLE_WITH_XPU
 
 #include "paddle/fluid/operators/batch_norm_op.h"
+#include <iterator>
+#include <vector>
 
 namespace paddle {
 namespace operators {
@@ -25,23 +27,25 @@ using DDim = framework::DDim;
 template <typename DeviceContext, typename T>
 class BatchNormXPUKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
+  void Compute(const framework::ExecutionContext &ctx) const override {
     const auto epsilon = ctx.Attr<float>("epsilon");
-    const auto momentum = ctx.Attr<float>("momentum");
+    float momentum = ctx.Attr<float>("momentum");
     const auto is_test = ctx.Attr<bool>("is_test");
     const auto use_global_stats = ctx.Attr<bool>("use_global_stats");
     const auto trainable_stats = ctx.Attr<bool>("trainable_statistics");
     bool test_mode = is_test && (!trainable_stats);
+
     bool global_stats = test_mode || use_global_stats;
-    const auto& data_layout_str = ctx.Attr<std::string>("data_layout");
+    const auto &data_layout_str = ctx.Attr<std::string>("data_layout");
     const auto data_layout = framework::StringToDataLayout(data_layout_str);
     PADDLE_ENFORCE_EQ(data_layout, DataLayout::kNCHW,
                       platform::errors::InvalidArgument(
                           "The 'data_layout' attribute must be NCHW. But "
                           "recevived 'data_layout' is [%s].",
                           data_layout_str));
-    const auto* x = ctx.Input<Tensor>("X");
-    const auto& x_dims = x->dims();
+
+    const auto *x = ctx.Input<Tensor>("X");
+    const auto &x_dims = x->dims();
     PADDLE_ENFORCE_EQ(x_dims.size(), 4,
                       platform::errors::InvalidArgument(
                           "The input tensor X's dimension must equal to 4. But "
@@ -51,27 +55,42 @@ class BatchNormXPUKernel : public framework::OpKernel<T> {
     const int C = x_dims[1];
     const int H = x_dims[2];
     const int W = x_dims[3];
-    const auto* scale = ctx.Input<Tensor>("Scale");
-    const auto* bias = ctx.Input<Tensor>("Bias");
-    const auto* x_data = x->data<T>();
-    const auto* scale_data = scale->data<T>();
-    const auto* bias_data = bias->data<T>();
-    auto* y = ctx.Output<Tensor>("Y");
-    auto* y_data = y->mutable_data<T>(ctx.GetPlace());
-    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+    const auto *scale = ctx.Input<Tensor>("Scale");
+    const auto *bias = ctx.Input<Tensor>("Bias");
+    const auto *x_data = x->data<T>();
+    const auto *scale_data = scale->data<float>();
+    const auto *bias_data = bias->data<float>();
+
+    auto *y = ctx.Output<Tensor>("Y");
+    auto *mean_out = ctx.Output<Tensor>("MeanOut");
+    auto *variance_out = ctx.Output<Tensor>("VarianceOut");
+    auto *saved_mean = ctx.Output<Tensor>("SavedMean");
+    auto *saved_variance = ctx.Output<Tensor>("SavedVariance");
+
+    // alloc memory
+    auto *y_data = y->mutable_data<T>(ctx.GetPlace());
+    mean_out->mutable_data<float>(ctx.GetPlace());
+    variance_out->mutable_data<float>(ctx.GetPlace());
+    saved_mean->mutable_data<float>(ctx.GetPlace());
+    saved_variance->mutable_data<float>(ctx.GetPlace());
+
+    auto &dev_ctx = ctx.template device_context<DeviceContext>();
+
     if (!global_stats) {
-      auto* mean_out = ctx.Output<Tensor>("MeanOut");
-      auto* variance_out = ctx.Output<Tensor>("VarianceOut");
-      auto* saved_mean = ctx.Output<Tensor>("SavedMean");
-      auto* saved_variance = ctx.Output<Tensor>("SavedVariance");
-      mean_out->mutable_data<T>(ctx.GetPlace());
-      variance_out->mutable_data<T>(ctx.GetPlace());
-      saved_mean->mutable_data<T>(ctx.GetPlace());
-      saved_variance->mutable_data<T>(ctx.GetPlace());
-      auto* mean_out_data = mean_out->data<T>();
-      auto* variance_out_data = variance_out->data<T>();
-      auto* saved_mean_data = saved_mean->data<T>();
-      auto* saved_variance_data = saved_variance->data<T>();
+      auto *mean_out_data = mean_out->data<float>();
+      auto *variance_out_data = variance_out->data<float>();
+      auto *saved_mean_data = saved_mean->data<float>();
+      auto *saved_variance_data = saved_variance->data<float>();
+
+      // if MomentumTensor is set, use MomentumTensor value, momentum
+      // is only used in this training branch
+      if (ctx.HasInput("MomentumTensor")) {
+        const auto *mom_tensor = ctx.Input<Tensor>("MomentumTensor");
+        Tensor mom_cpu;
+        TensorCopySync(*mom_tensor, platform::CPUPlace(), &mom_cpu);
+        momentum = mom_tensor->data<float>()[0];
+      }
+
       int r = xpu::batch_norm<T>(dev_ctx.x_context(), x_data, y_data, N, C, H,
                                  W, epsilon, momentum, scale_data, bias_data,
                                  saved_mean_data, saved_variance_data,
@@ -81,12 +100,10 @@ class BatchNormXPUKernel : public framework::OpKernel<T> {
                             "The batch_norm XPU API return wrong value[%d %s]",
                             r, XPUAPIErrorMsg[r]));
     } else {
-      const auto* mean = ctx.Input<Tensor>("Mean");
-      const auto* variance = ctx.Input<Tensor>("Variance");
-      const auto* mean_data = mean->data<float>();
-      const auto* variance_data = variance->data<float>();
-      const auto* x_data = x->data<float>();
-      auto* y_data = y->mutable_data<float>(ctx.GetPlace());
+      const auto *mean = ctx.Input<Tensor>("Mean");
+      const auto *variance = ctx.Input<Tensor>("Variance");
+      const auto *mean_data = mean->data<float>();
+      const auto *variance_data = variance->data<float>();
       int r = xpu::batch_norm_infer(dev_ctx.x_context(), x_data, y_data, N, C,
                                     H, W, epsilon, scale_data, bias_data,
                                     mean_data, variance_data, true);
@@ -99,24 +116,96 @@ class BatchNormXPUKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename T>
+static int calculate_inv_BN_Y(xpu::Context *ctx, T *x, const T *scale,
+                              const T *bias, const T *mean, const T *variance,
+                              const int N, const int C, const int M,
+                              const T *y) {
+  PADDLE_ENFORCE_EQ(x, y, platform::errors::InvalidArgument(
+                              "X and Y should be inplaced in inplace mode"));
+  std::vector<int> tensor_shape_vec({N, C, M});
+  std::vector<int> array_shape_vec({1, C, 1});
+  // y - bias
+  int r1 =
+      xpu::broadcast_sub<T>(ctx, bias, y, x, array_shape_vec, tensor_shape_vec);
+  // (y - bias) / scale
+  int r2 = xpu::broadcast_div<T>(ctx, scale, x, x, array_shape_vec,
+                                 tensor_shape_vec);
+  // (y - bias) / scale / variance
+  int r3 = xpu::broadcast_div<T>(ctx, variance, x, x, array_shape_vec,
+                                 tensor_shape_vec);
+  // (y - bias) / scale / variance + mean
+  int r4 =
+      xpu::broadcast_add<T>(ctx, mean, x, x, array_shape_vec, tensor_shape_vec);
+
+  return r1 + r2 + r3 + r4;
+}
+
+template <typename T>
+static int calculate_inv_var(xpu::Context *ctx, const T *var, const T epsilon,
+                             const int C, T *epsilon_data, T *inv_var) {
+  int r1 = constant(ctx, epsilon_data, 1, epsilon);
+  std::vector<int> tensor_shape_vec({C});
+  std::vector<int> array_shape_vec({1});
+  int r2 = xpu::broadcast_add<T>(ctx, epsilon_data, var, inv_var,
+                                 array_shape_vec, tensor_shape_vec);
+  int r3 = xpu::rsqrt<T>(ctx, inv_var, inv_var, C);
+  return r1 + r2 + r3;
+}
+
 template <typename DeviceContext, typename T>
 class BatchNormGradXPUKernel : public framework::OpKernel<T> {
  public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    const auto* x = ctx.Input<Tensor>("X");
-    const auto* dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
-    const auto* scale = ctx.Input<Tensor>("Scale");
-    const auto* saved_mean = ctx.Input<Tensor>("SavedMean");
-    // SavedVariance have been reverted in forward operator
-    const auto* saved_inv_variance = ctx.Input<Tensor>("SavedVariance");
-    const auto& data_layout_str = ctx.Attr<std::string>("data_layout");
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    const auto *d_y = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    const auto *scale = ctx.Input<Tensor>("Scale");
+    const auto *bias = ctx.Input<Tensor>("Bias");
+
+    const auto &data_layout_str = ctx.Attr<std::string>("data_layout");
+    bool use_global_stats = ctx.Attr<bool>("use_global_stats");
+    const bool is_test = ctx.Attr<bool>("is_test");
+    const float epsilon = ctx.Attr<float>("epsilon");
     const auto data_layout = framework::StringToDataLayout(data_layout_str);
+
+    // TODO(guozbin): Transform input tensor from NHWC to NCHW
     PADDLE_ENFORCE_EQ(data_layout, DataLayout::kNCHW,
                       platform::errors::InvalidArgument(
                           "The 'data_layout' attribute must be NCHW. But "
                           "recevived 'data_layout' is [%s].",
                           data_layout_str));
-    const auto& x_dims = x->dims();
+
+    auto *d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *d_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
+    auto *d_bias = ctx.Output<Tensor>(framework::GradVarName("Bias"));
+
+    use_global_stats = is_test || use_global_stats;
+
+    // batch_norm with inplace as false will take X as grad input, which
+    // is same as cuDNN batch_norm backward calculation, batch_norm
+    // with inplace as true only take Y as input and X should be calculate
+    // by inverse operation of batch_norm on Y
+    const Tensor *x;
+    bool is_inplace;
+    if (ctx.HasInput("Y")) {
+      x = ctx.Input<Tensor>("Y");
+      is_inplace = true;
+      // if the input of batch norm is stop_gradient, d_x is null.
+      if (d_x) {
+        PADDLE_ENFORCE_EQ(d_x, d_y,
+                          platform::errors::InvalidArgument(
+                              "X@GRAD and Y@GRAD not inplace in inplace mode"));
+      }
+    } else {
+      x = ctx.Input<Tensor>("X");
+      is_inplace = false;
+      if (d_x) {
+        PADDLE_ENFORCE_NE(
+            d_x, d_y, platform::errors::InvalidArgument(
+                          "X@GRAD and Y@GRAD inplaced in non-inplace mode"));
+      }
+    }
+
+    const auto &x_dims = x->dims();
     PADDLE_ENFORCE_EQ(x_dims.size(), 4,
                       platform::errors::InvalidArgument(
                           "The input tensor X's dimension must equal to 4. But "
@@ -126,26 +215,96 @@ class BatchNormGradXPUKernel : public framework::OpKernel<T> {
     const int C = x_dims[1];
     const int H = x_dims[2];
     const int W = x_dims[3];
-    const auto* x_data = x->data<T>();
-    const auto* dy_data = dy->data<T>();
-    const auto* scale_data = scale->data<T>();
-    const auto* saved_mean_data = saved_mean->data<T>();
-    const auto* saved_inv_variance_data = saved_inv_variance->data<T>();
-    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto* dscale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
-    auto* dbias = ctx.Output<Tensor>(framework::GradVarName("Bias"));
-    auto* dx_data = dx->mutable_data<T>(ctx.GetPlace());
-    auto* dscale_data = dscale->mutable_data<T>(ctx.GetPlace());
-    auto* dbias_data = dbias->mutable_data<T>(ctx.GetPlace());
-    auto& dev_ctx = ctx.template device_context<DeviceContext>();
-    int r = xpu::batch_norm_grad<T>(dev_ctx.x_context(), x_data, dy_data,
-                                    dx_data, N, C, H, W, scale_data,
-                                    saved_mean_data, saved_inv_variance_data,
-                                    dscale_data, dbias_data, true);
-    PADDLE_ENFORCE_EQ(r, XPU_SUCCESS, platform::errors::External(
-                                          "XPU API(batch_norm_grad) return "
-                                          "wrong value[%d %s]",
-                                          r, XPUAPIErrorMsg[r]));
+
+    const auto *x_data = x->data<T>();
+    const auto *d_y_data = d_y->data<T>();
+    const auto *scale_data = scale->data<float>();
+
+    // init output
+    T *d_x_data = nullptr;
+    T *d_bias_data = nullptr;
+    T *d_scale_data = nullptr;
+    if (d_x) {
+      d_x_data = d_x->mutable_data<T>(ctx.GetPlace());
+    }
+    if (d_scale && d_bias) {
+      d_scale_data = d_scale->mutable_data<float>(ctx.GetPlace());
+      d_bias_data = d_bias->mutable_data<float>(ctx.GetPlace());
+    }
+
+    PADDLE_ENFORCE_EQ(
+        scale->dims().size(), 1UL,
+        platform::errors::InvalidArgument(
+            "The size of scale's dimensions must equal to 1. But received: "
+            "the size of scale's dimensions is [%d], the dimensions of scale "
+            "is [%s].",
+            scale->dims().size(), scale->dims()));
+    PADDLE_ENFORCE_EQ(
+        scale->dims()[0], C,
+        platform::errors::InvalidArgument(
+            "The first dimension of scale must equal to Channels[%d]. But "
+            "received: the first dimension of scale is [%d]",
+            C, scale->dims()[0]));
+
+    auto &dev_ctx = ctx.template device_context<DeviceContext>();
+    xpu::ctx_guard RAII_GUARD(dev_ctx.x_context());
+
+    const T *mean_data = nullptr;
+    const T *inv_var_data = nullptr;
+
+    // TODO(guozibin): hadle the situation case of N * H * W = 1
+    if (!use_global_stats) {
+      const auto *saved_mean = ctx.Input<Tensor>("SavedMean");
+      // SavedVariance have been reverted in forward operator
+      const auto *saved_inv_variance = ctx.Input<Tensor>("SavedVariance");
+      mean_data = saved_mean->data<float>();
+      inv_var_data = saved_inv_variance->data<float>();
+    } else {
+      const auto *running_mean = ctx.Input<Tensor>("Mean");
+      const auto *running_variance = ctx.Input<Tensor>("Variance");
+      mean_data = running_mean->data<float>();
+      inv_var_data = running_variance->data<float>();
+      float *running_inv_var_data =
+          RAII_GUARD.alloc_l3_or_gm<float>(running_variance->numel());
+      float *epsilon_data = RAII_GUARD.alloc_l3_or_gm<float>(1);
+      int r1 = calculate_inv_var(dev_ctx.x_context(), inv_var_data, epsilon, C,
+                                 epsilon_data, running_inv_var_data);
+      PADDLE_ENFORCE_EQ(r1, XPU_SUCCESS, platform::errors::External(
+                                             "XPU API(batch_norm_grad "
+                                             "calculate_inv_var function) "
+                                             "return wrong value[%d %s]",
+                                             r1, XPUAPIErrorMsg[r1]));
+      inv_var_data = running_inv_var_data;
+    }
+    if (is_inplace) {
+      auto px = *x;
+      int r2 = calculate_inv_BN_Y(
+          dev_ctx.x_context(), px.mutable_data<T>(ctx.GetPlace()),
+          scale->data<float>(), bias->data<float>(), mean_data, inv_var_data, N,
+          C, H * W, x->data<T>());
+      PADDLE_ENFORCE_EQ(r2, XPU_SUCCESS, platform::errors::External(
+                                             "XPU API(batch_norm_grad "
+                                             "calculate_inv_BN_Y function) "
+                                             "return wrong value[%d %s]",
+                                             r2, XPUAPIErrorMsg[r2]));
+    }
+    if (!d_x) {
+      d_x_data = RAII_GUARD.alloc_l3_or_gm<T>(x->numel());
+    }
+    if (!d_scale) {
+      d_scale_data = RAII_GUARD.alloc_l3_or_gm<float>(C);
+    }
+    if (!d_bias_data) {
+      d_bias_data = RAII_GUARD.alloc_l3_or_gm<float>(C);
+    }
+
+    int r3 = xpu::batch_norm_grad<T>(
+        dev_ctx.x_context(), x_data, d_y_data, d_x_data, N, C, H, W, scale_data,
+        mean_data, inv_var_data, d_scale_data, d_bias_data, true);
+    PADDLE_ENFORCE_EQ(r3, XPU_SUCCESS, platform::errors::External(
+                                           "XPU API(batch_norm_grad) return "
+                                           "wrong value[%d %s]",
+                                           r3, XPUAPIErrorMsg[r3]));
   }
 };
 
diff --git a/paddle/fluid/platform/device/xpu/xpu2_op_list.h b/paddle/fluid/platform/device/xpu/xpu2_op_list.h
index 3d7739f5a06aff15a17764441c46bea7d63a33fb..c5a140a768190fea173b260c09fe4b9e9eb3e137 100644
--- a/paddle/fluid/platform/device/xpu/xpu2_op_list.h
+++ b/paddle/fluid/platform/device/xpu/xpu2_op_list.h
@@ -262,6 +262,9 @@ XPUOpMap& get_kl2_ops() {
                                  pOpKernelType(vartype::INT32, XPUPlace()),
                                  pOpKernelType(vartype::BOOL, XPUPlace()),
                                  pOpKernelType(vartype::FP32, XPUPlace())})},
+      {"roi_align", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
+      {"roi_align_grad",
+       XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"scale", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"scale", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),
                               pOpKernelType(vartype::FP16, XPUPlace()),
diff --git a/python/paddle/fluid/tests/unittests/xpu/test_batch_norm_op_xpu.py b/python/paddle/fluid/tests/unittests/xpu/test_batch_norm_op_xpu.py
index 8132a78f696756eebf09b4dee4fe8e47aed34f88..9cd34c82650e9b5d474af735277843f0c17fd1c0 100644
--- a/python/paddle/fluid/tests/unittests/xpu/test_batch_norm_op_xpu.py
+++ b/python/paddle/fluid/tests/unittests/xpu/test_batch_norm_op_xpu.py
@@ -267,5 +267,58 @@ class TestXPUBatchNormOp(unittest.TestCase):
                         outputs[name], outs[id], atol=1e-4), True)
 
 
+class TestXPUBatchNormOpUseGlobalStats(unittest.TestCase):
+    def setUp(self):
+        self.places = [paddle.XPUPlace(0)]
+        self.init_test()
+
+    ### train mode
+    def init_test(self):
+        self.use_global_stats = True
+        self.trainable_statistics = False
+
+    def test_global_stats(self):
+        for p in self.places:
+            with fluid.dygraph.guard(p):
+                x = paddle.randn([2, 6, 6, 4])
+                net1 = paddle.fluid.dygraph.BatchNorm(
+                    6,
+                    param_attr=fluid.ParamAttr(
+                        initializer=fluid.initializer.Constant(1.0)),
+                    use_global_stats=self.use_global_stats,
+                    trainable_statistics=self.trainable_statistics)
+                net2 = paddle.nn.BatchNorm2D(
+                    6, use_global_stats=self.use_global_stats)
+                net2.weight = net1.weight
+                net2.bias = net1.bias
+                if self.trainable_statistics == True:
+                    net1.training = False
+                    net2.training = False
+                y1 = net1(x)
+                y2 = net2(x)
+                self.assertEqual(np.allclose(y1.numpy(), y2.numpy()), True)
+
+
+class TestXPUBatchNormUseGlobalStatsCase1(TestXPUBatchNormOpUseGlobalStats):
+    ### test mode
+    def init_test(self):
+        self.use_global_stats = False
+        self.trainable_statistics = True
+
+
+class TestXPUBatchNormUseGlobalStatsCase2(TestXPUBatchNormOpUseGlobalStats):
+    ### train mode
+    def init_test(self):
+        self.use_global_stats = False
+        self.trainable_statistics = False
+
+
+class TestXPUBatchNormUseGlobalStatsCase3(TestXPUBatchNormOpUseGlobalStats):
+    ### test mode
+    def init_test(self):
+        self.use_global_stats = True
+        self.trainable_statistics = True
+
+
 if __name__ == "__main__":
     unittest.main()