diff --git a/paddle/operators/gemm_conv2d_op.h b/paddle/operators/gemm_conv2d_op.h
index 08b7df1dfead72fe8de8e89fa633c7bfc7bdbf33..5c9e81732aa72211c2021382cf9a907880c53c17 100644
--- a/paddle/operators/gemm_conv2d_op.h
+++ b/paddle/operators/gemm_conv2d_op.h
@@ -75,9 +75,6 @@ class GemmConv2DKernel : public framework::OpKernel {
     framework::DDim output_matrix_shape = {output_channels,
                                            output_height * output_width};
 
-    auto* device_context =
-        const_cast<platform::DeviceContext*>(context.device_context_);
-
     // convolution operator: im2col + gemm
     int in_step = input_channels / groups;
     int out_step = output_channels / groups;
@@ -87,14 +84,14 @@ class GemmConv2DKernel : public framework::OpKernel {
       for (int g = 0; g < groups; g++) {
         // im2col
         Tensor in_slice = in_batch.Slice<T>(g * in_step, (g + 1) * in_step);
-        im2col(in_slice, col, strides[0], strides[1], paddings[0], paddings[1],
-               device_context);
+        im2col(context.device_context(), in_slice, col, strides[0], strides[1],
+               paddings[0], paddings[1]);
 
         // gemm
         Tensor out_slice = out_batch.Slice<T>(g * out_step, (g + 1) * out_step);
         Tensor filter_slice = filter.Slice<T>(g * out_step, (g + 1) * out_step);
-        math::matmul<Place, T>(filter_slice, false, col_matrix, false, T(1.0),
-                               &out_slice, T(0.0), device_context);
+        math::matmul<Place, T>(context.device_context(), filter_slice, false,
+                               col_matrix, false, T(1.0), &out_slice, T(0.0));
       }
     }
   }
@@ -160,9 +157,6 @@ class GemmConvGrad2DKernel : public framework::OpKernel {
                                            filter.numel() / filter.dims()[0]};
     filter.Resize(filter_matrix_shape);
 
-    auto* device_context =
-        const_cast<platform::DeviceContext*>(context.device_context_);
-
     // convolution backward input operator:  gemm + col2im
     // convolution backward weight operator: im2col + gemm
     int in_step = input_channels / groups;
@@ -184,14 +178,15 @@ class GemmConvGrad2DKernel : public framework::OpKernel {
               out_grad_batch.Slice<T>(g * out_step, (g + 1) * out_step);
           Tensor filter_slice =
               filter.Slice<T>(g * out_step, (g + 1) * out_step);
-          math::matmul<Place, T>(filter_slice, true, out_grad_slice, false,
-                                 T(1.0), &col_matrix, T(0.0), device_context);
+          math::matmul<Place, T>(context.device_context(), filter_slice, true,
+                                 out_grad_slice, false, T(1.0), &col_matrix,
+                                 T(0.0));
 
           // col2im
           Tensor in_grad_slice =
               in_grad_batch.Slice<T>(g * in_step, (g + 1) * in_step);
-          col2im(in_grad_slice, col, strides[0], strides[1], paddings[0],
-                 paddings[1], device_context);
+          col2im(context.device_context(), in_grad_slice, col, strides[0],
+                 strides[1], paddings[0], paddings[1]);
         }
       }
     }
@@ -212,15 +207,15 @@ class GemmConvGrad2DKernel : public framework::OpKernel {
           Tensor out_grad_slice =
               out_grad_batch.Slice<T>(g * out_step, (g + 1) * out_step);
           Tensor in_slice = in_batch.Slice<T>(g * in_step, (g + 1) * in_step);
-          im2col(in_slice, col, strides[0], strides[1], paddings[0],
-                 paddings[1], device_context);
+          im2col(context.device_context(), in_slice, col, strides[0],
+                 strides[1], paddings[0], paddings[1]);
 
           // gemm
           Tensor filter_grad_slice =
               filter_grad_.Slice<T>(g * out_step, (g + 1) * out_step);
-          math::matmul<Place, T>(out_grad_slice, false, col_matrix, true,
-                                 T(1.0), &filter_grad_slice, T(1.0),
-                                 device_context);
+          math::matmul<Place, T>(context.device_context(), out_grad_slice,
+                                 false, col_matrix, true, T(1.0),
+                                 &filter_grad_slice, T(1.0));
         }
       }
     }
diff --git a/paddle/operators/math/im2col.cc b/paddle/operators/math/im2col.cc
index 5727c1cab16c1379ffe77f5594c057e93a042785..c08a3380f042886cd400df0d840e61856274619c 100644
--- a/paddle/operators/math/im2col.cc
+++ b/paddle/operators/math/im2col.cc
@@ -27,9 +27,10 @@ template <class T>
 class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
                     platform::CPUPlace, T> {
  public:
-  void operator()(const framework::Tensor& im, framework::Tensor& col,
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& im, framework::Tensor& col,
                   int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context) {
+                  int padding_width) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
 
@@ -79,9 +80,9 @@ template <class T>
 class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
                     platform::CPUPlace, T> {
  public:
-  void operator()(framework::Tensor& im, const framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context) {
+  void operator()(const platform::DeviceContext& context, framework::Tensor& im,
+                  const framework::Tensor& col, int stride_height,
+                  int stride_width, int padding_height, int padding_width) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
@@ -137,9 +138,10 @@ template <class T>
 class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
                     platform::CPUPlace, T> {
  public:
-  void operator()(const framework::Tensor& im, framework::Tensor& col,
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& im, framework::Tensor& col,
                   int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context) {
+                  int padding_width) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
@@ -197,9 +199,9 @@ template <class T>
 class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
                     platform::CPUPlace, T> {
  public:
-  void operator()(framework::Tensor& im, const framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context) {
+  void operator()(const platform::DeviceContext& context, framework::Tensor& im,
+                  const framework::Tensor& col, int stride_height,
+                  int stride_width, int padding_height, int padding_width) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
diff --git a/paddle/operators/math/im2col.cu b/paddle/operators/math/im2col.cu
index 9bff7bee3c95093852305d392af0949b831e5665..01f60bfe70f844fdcfd5aa481c27d9f12ec51305 100644
--- a/paddle/operators/math/im2col.cu
+++ b/paddle/operators/math/im2col.cu
@@ -64,9 +64,10 @@ template <class T>
 class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
                     platform::GPUPlace, T> {
  public:
-  void operator()(const framework::Tensor& im, framework::Tensor& col,
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& im, framework::Tensor& col,
                   int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context) {
+                  int padding_width) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
 
@@ -84,9 +85,9 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
     int block_y = (blocks + 512 - 1) / 512;
     dim3 threads(1024, 1);
     dim3 grid(block_x, block_y);
-    im2col<T><<<
-        grid, threads, 0,
-        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+    im2col<T><<<grid, threads, 0,
+                reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                    .stream()>>>(
         im.data<T>(), num_outputs, input_height, input_width, filter_height,
         filter_width, stride_height, stride_width, padding_height,
         padding_width, output_height, output_width, col.data<T>());
@@ -149,9 +150,9 @@ template <class T>
 class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
                     platform::GPUPlace, T> {
  public:
-  void operator()(framework::Tensor& im, const framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context) {
+  void operator()(const platform::DeviceContext& context, framework::Tensor& im,
+                  const framework::Tensor& col, int stride_height,
+                  int stride_width, int padding_height, int padding_width) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
 
@@ -174,9 +175,9 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
 
     // To avoid involving atomic operations, we will launch one kernel per
     // bottom dimension, and then in the kernel add up the top dimensions.
-    col2im<T><<<
-        grid, threads, 0,
-        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+    col2im<T><<<grid, threads, 0,
+                reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                    .stream()>>>(
         num_kernels, col.data<T>(), input_height + 2 * padding_height,
         input_width + 2 * padding_width, input_channels, filter_height,
         filter_width, stride_height, stride_width, padding_height,
@@ -235,9 +236,10 @@ template <class T>
 class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
                     platform::GPUPlace, T> {
  public:
-  void operator()(const framework::Tensor& im, framework::Tensor& col,
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& im, framework::Tensor& col,
                   int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context) {
+                  int padding_width) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
@@ -268,9 +270,9 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
     dim3 threads(block_dim_x, block_dim_y,
                  std::min(block_dim_z, input_channels));
     dim3 grid(output_width, output_height);
-    im2colOCF<T><<<
-        grid, threads, 0,
-        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+    im2colOCF<T><<<grid, threads, 0,
+                   reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                       .stream()>>>(
         im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
         filter_height, filter_width, stride_height, stride_width,
         padding_height, padding_width, output_height, output_width);
@@ -318,9 +320,9 @@ template <class T>
 class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
                     platform::GPUPlace, T> {
  public:
-  void operator()(framework::Tensor& im, const framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context) {
+  void operator()(const platform::DeviceContext& context, framework::Tensor& im,
+                  const framework::Tensor& col, int stride_height,
+                  int stride_width, int padding_height, int padding_width) {
     PADDLE_ENFORCE(im.dims().size() == 3);
     PADDLE_ENFORCE(col.dims().size() == 5);
     int input_channels = im.dims()[0];
@@ -351,9 +353,9 @@ class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
     dim3 threads(block_dim_x, block_dim_y,
                  std::min(block_dim_z, input_channels));
     dim3 grid(output_width, output_height);
-    col2imOCF<T><<<
-        grid, threads, 0,
-        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+    col2imOCF<T><<<grid, threads, 0,
+                   reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                       .stream()>>>(
         im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
         filter_height, filter_width, stride_height, stride_width,
         padding_height, padding_width, output_height, output_width);
diff --git a/paddle/operators/math/im2col.h b/paddle/operators/math/im2col.h
index 8958c5457cc2c3034c34ca82fb2e98cc06be63c5..7b717e1603c94cd77c74cb0d86f1d23e2692f9d8 100644
--- a/paddle/operators/math/im2col.h
+++ b/paddle/operators/math/im2col.h
@@ -72,17 +72,18 @@ enum class ColFormat { kCFO = 0, kOCF = 1 };
 template <ColFormat Format, typename Place, typename T>
 class Im2ColFunctor {
  public:
-  void operator()(const framework::Tensor& im, framework::Tensor& col,
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& im, framework::Tensor& col,
                   int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context);
+                  int padding_width);
 };
 
 template <ColFormat Format, typename Place, typename T>
 class Col2ImFunctor {
  public:
-  void operator()(framework::Tensor& im, const framework::Tensor& col,
-                  int stride_height, int stride_width, int padding_height,
-                  int padding_width, platform::DeviceContext* context);
+  void operator()(const platform::DeviceContext& context, framework::Tensor& im,
+                  const framework::Tensor& col, int stride_height,
+                  int stride_width, int padding_height, int padding_width);
 };
 
 }  // namespace math
diff --git a/paddle/operators/math/im2col_test.cc b/paddle/operators/math/im2col_test.cc
index 4f380388b108dc173d847f027ba5c9db387a87f8..f0b8c885918afe7f80edc465c6d9be7c11ac066f 100644
--- a/paddle/operators/math/im2col_test.cc
+++ b/paddle/operators/math/im2col_test.cc
@@ -78,8 +78,8 @@ void testIm2col() {
     PADDLE_THROW("no GPU support");
 #endif  // PADDLE_ONLY_CPU
   }
-  im2col(input, output_cfo, stride, stride, padding, padding, context);
-  im2col_ocf(input, output_ocf, stride, stride, padding, padding, context);
+  im2col(*context, input, output_cfo, stride, stride, padding, padding);
+  im2col_ocf(*context, input, output_ocf, stride, stride, padding, padding);
 
   float* out_cfo_ptr;
   if (paddle::platform::is_cpu_place(*place)) {
diff --git a/python/paddle/v2/framework/tests/test_conv2d_op.py b/python/paddle/v2/framework/tests/test_conv2d_op.py
index 3142a60a1ae7d1874d02b81a4bb90c1fc50d07b9..118a5fc1cde5f4a908b065d581956e0855d50a52 100644
--- a/python/paddle/v2/framework/tests/test_conv2d_op.py
+++ b/python/paddle/v2/framework/tests/test_conv2d_op.py
@@ -73,13 +73,22 @@ class TestConv2dOp(OpTest):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(set(['Input', 'Filter']), 'Output')
+        self.check_grad(
+            set(['Input', 'Filter']), 'Output', max_relative_error=0.05)
 
     def test_check_grad_no_filter(self):
-        self.check_grad(['Input'], 'Output', no_grad_set=set(['Filter']))
+        self.check_grad(
+            ['Input'],
+            'Output',
+            max_relative_error=0.05,
+            no_grad_set=set(['Filter']))
 
     def test_check_grad_no_input(self):
-        self.check_grad(['Filter'], 'Output', no_grad_set=set(['Input']))
+        self.check_grad(
+            ['Filter'],
+            'Output',
+            max_relative_error=0.05,
+            no_grad_set=set(['Input']))
 
     def init_groups(self):
         self.groups = 1