diff --git a/paddle/operators/conv_cudnn_op.cc b/paddle/operators/conv_cudnn_op.cc
index 4c65b60d2349d2989128f4b1da705ea18391b8a3..c03dc3e4fb07ac6ecde42be93a1138d91778edf4 100644
--- a/paddle/operators/conv_cudnn_op.cc
+++ b/paddle/operators/conv_cudnn_op.cc
@@ -40,7 +40,8 @@ REGISTER_OP(conv_cudnn, ops::ConvOp, ops::CudnnConvOpMaker, conv_cudnn_grad,
             ops::ConvOpGrad);
 
 REGISTER_OP_CPU_KERNEL(conv_cudnn,
-                       ops::GemmConvKernel<paddle::platform::CPUPlace, float>);
+                       ops::GemmConvKernel<paddle::platform::CPUPlace, float>,
+                       ops::GemmConvKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
-    conv_cudnn_grad,
-    ops::GemmConvGradKernel<paddle::platform::CPUPlace, float>);
+    conv_cudnn_grad, ops::GemmConvGradKernel<paddle::platform::CPUPlace, float>,
+    ops::GemmConvGradKernel<paddle::platform::CPUPlace, double>);
diff --git a/paddle/operators/conv_cudnn_op.cu.cc b/paddle/operators/conv_cudnn_op.cu.cc
index 4900f7b086c869b496c492743c71ab7047c5f672..5eaf6b33704eb371fff4b949c6cc32a7a5dbc812 100644
--- a/paddle/operators/conv_cudnn_op.cu.cc
+++ b/paddle/operators/conv_cudnn_op.cu.cc
@@ -259,6 +259,8 @@ class CudnnConvGradOpKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_GPU_KERNEL(conv_cudnn, paddle::operators::CudnnConvOpKernel<float>);
+REGISTER_OP_GPU_KERNEL(conv_cudnn, paddle::operators::CudnnConvOpKernel<float>,
+                       paddle::operators::CudnnConvOpKernel<double>);
 REGISTER_OP_GPU_KERNEL(conv_cudnn_grad,
-                       paddle::operators::CudnnConvGradOpKernel<float>);
+                       paddle::operators::CudnnConvGradOpKernel<float>,
+                       paddle::operators::CudnnConvGradOpKernel<double>);
diff --git a/paddle/operators/conv_transpose_cudnn_op.cc b/paddle/operators/conv_transpose_cudnn_op.cc
index dbd1bc3c3bc2d026f13ddcf62919db6cf7d87bc5..0192178ce3a0a47196232f0723baec8324bea60b 100644
--- a/paddle/operators/conv_transpose_cudnn_op.cc
+++ b/paddle/operators/conv_transpose_cudnn_op.cc
@@ -61,10 +61,12 @@ REGISTER_OP(conv2d_transpose_cudnn, ops::ConvTransposeOp,
 
 REGISTER_OP_CPU_KERNEL(
     conv2d_transpose_cudnn,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>);
+    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>,
+    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
     conv2d_transpose_cudnn_grad,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>);
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>,
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, double>);
 
 REGISTER_OP(conv3d_transpose_cudnn, ops::ConvTransposeOp,
             ops::CudnnConv3DTransposeOpMaker, conv3d_transpose_cudnn_grad,
@@ -72,7 +74,9 @@ REGISTER_OP(conv3d_transpose_cudnn, ops::ConvTransposeOp,
 
 REGISTER_OP_CPU_KERNEL(
     conv3d_transpose_cudnn,
-    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>);
+    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, float>,
+    ops::GemmConvTransposeKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
     conv3d_transpose_cudnn_grad,
-    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>);
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, float>,
+    ops::GemmConvTransposeGradKernel<paddle::platform::CPUPlace, double>);
diff --git a/paddle/operators/conv_transpose_cudnn_op.cu.cc b/paddle/operators/conv_transpose_cudnn_op.cu.cc
index e2ba77086e737a07471f14e483cbd32ab1d4ee12..494904fe524ae30a5032e489a0c5f20179d8e8ce 100644
--- a/paddle/operators/conv_transpose_cudnn_op.cu.cc
+++ b/paddle/operators/conv_transpose_cudnn_op.cu.cc
@@ -235,11 +235,15 @@ class CudnnConvTransposeGradOpKernel : public framework::OpKernel<T> {
 namespace ops = paddle::operators;
 
 REGISTER_OP_GPU_KERNEL(conv2d_transpose_cudnn,
-                       ops::CudnnConvTransposeOpKernel<float>);
+                       ops::CudnnConvTransposeOpKernel<float>,
+                       ops::CudnnConvTransposeOpKernel<double>);
 REGISTER_OP_GPU_KERNEL(conv2d_transpose_cudnn_grad,
-                       ops::CudnnConvTransposeGradOpKernel<float>);
+                       ops::CudnnConvTransposeGradOpKernel<float>,
+                       ops::CudnnConvTransposeGradOpKernel<double>);
 
 REGISTER_OP_GPU_KERNEL(conv3d_transpose_cudnn,
-                       ops::CudnnConvTransposeOpKernel<float>);
+                       ops::CudnnConvTransposeOpKernel<float>,
+                       ops::CudnnConvTransposeOpKernel<double>);
 REGISTER_OP_GPU_KERNEL(conv3d_transpose_cudnn_grad,
-                       ops::CudnnConvTransposeGradOpKernel<float>);
+                       ops::CudnnConvTransposeGradOpKernel<float>,
+                       ops::CudnnConvTransposeGradOpKernel<double>);
diff --git a/paddle/operators/pool_cudnn_op.cc b/paddle/operators/pool_cudnn_op.cc
index 06cf1c0d2a6f6d0be8a3636389bb1d209666aae1..be9fcc5661f420aadf908cf80cce6c963008b0e4 100644
--- a/paddle/operators/pool_cudnn_op.cc
+++ b/paddle/operators/pool_cudnn_op.cc
@@ -20,14 +20,18 @@ REGISTER_OP(pool2d_cudnn, ops::PoolOp, ops::Pool2dOpMaker, pool2d_cudnn_grad,
             ops::PoolOpGrad);
 
 REGISTER_OP_CPU_KERNEL(pool2d_cudnn,
-                       ops::PoolKernel<paddle::platform::CPUPlace, float>);
+                       ops::PoolKernel<paddle::platform::CPUPlace, float>,
+                       ops::PoolKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(pool2d_cudnn_grad,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>)
+                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>,
+                       ops::PoolGradKernel<paddle::platform::CPUPlace, double>)
 
 REGISTER_OP(pool3d_cudnn, ops::PoolOp, ops::Pool3dOpMaker, pool3d_cudnn_grad,
             ops::PoolOpGrad);
 
 REGISTER_OP_CPU_KERNEL(pool3d_cudnn,
-                       ops::PoolKernel<paddle::platform::CPUPlace, float>);
+                       ops::PoolKernel<paddle::platform::CPUPlace, float>,
+                       ops::PoolKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(pool3d_cudnn_grad,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>)
+                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>,
+                       ops::PoolGradKernel<paddle::platform::CPUPlace, double>)
diff --git a/paddle/operators/pool_cudnn_op.cu.cc b/paddle/operators/pool_cudnn_op.cu.cc
index d5ba9843995928005a4fcc402da1391cd0358eae..66dd194ccd5ed629c5861552a7c124dc911362d7 100644
--- a/paddle/operators/pool_cudnn_op.cu.cc
+++ b/paddle/operators/pool_cudnn_op.cu.cc
@@ -162,8 +162,12 @@ class PoolCudnnGradOpKernel : public framework::OpKernel<T> {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(pool2d_cudnn, ops::PoolCudnnOpKernel<float>);
-REGISTER_OP_GPU_KERNEL(pool2d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>);
-
-REGISTER_OP_GPU_KERNEL(pool3d_cudnn, ops::PoolCudnnOpKernel<float>);
-REGISTER_OP_GPU_KERNEL(pool3d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>);
+REGISTER_OP_GPU_KERNEL(pool2d_cudnn, ops::PoolCudnnOpKernel<float>,
+                       ops::PoolCudnnOpKernel<double>);
+REGISTER_OP_GPU_KERNEL(pool2d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>,
+                       ops::PoolCudnnGradOpKernel<double>);
+
+REGISTER_OP_GPU_KERNEL(pool3d_cudnn, ops::PoolCudnnOpKernel<float>,
+                       ops::PoolCudnnOpKernel<double>);
+REGISTER_OP_GPU_KERNEL(pool3d_cudnn_grad, ops::PoolCudnnGradOpKernel<float>,
+                       ops::PoolCudnnGradOpKernel<double>);
diff --git a/paddle/operators/pool_op.cc b/paddle/operators/pool_op.cc
index f3963b1995ef8767786f0bf230b134afc69aa99d..d8c58618cf703d086d3cabc927ebc5eb038b1aec 100644
--- a/paddle/operators/pool_op.cc
+++ b/paddle/operators/pool_op.cc
@@ -217,14 +217,18 @@ REGISTER_OP(pool2d, ops::PoolOp, ops::Pool2dOpMaker, pool2d_grad,
             ops::PoolOpGrad);
 
 REGISTER_OP_CPU_KERNEL(pool2d,
-                       ops::PoolKernel<paddle::platform::CPUPlace, float>);
+                       ops::PoolKernel<paddle::platform::CPUPlace, float>,
+                       ops::PoolKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(pool2d_grad,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>)
+                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>,
+                       ops::PoolGradKernel<paddle::platform::CPUPlace, double>)
 
 REGISTER_OP(pool3d, ops::PoolOp, ops::Pool3dOpMaker, pool3d_grad,
             ops::PoolOpGrad);
 
 REGISTER_OP_CPU_KERNEL(pool3d,
-                       ops::PoolKernel<paddle::platform::CPUPlace, float>);
+                       ops::PoolKernel<paddle::platform::CPUPlace, float>,
+                       ops::PoolKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(pool3d_grad,
-                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>);
+                       ops::PoolGradKernel<paddle::platform::CPUPlace, float>,
+                       ops::PoolGradKernel<paddle::platform::CPUPlace, double>);
diff --git a/paddle/operators/pool_op.cu.cc b/paddle/operators/pool_op.cu.cc
index 0e3b80868f7b9d1697d619889160856d65ad59a3..1010cb762289dd39cd632c699f7528f4ba638278 100644
--- a/paddle/operators/pool_op.cu.cc
+++ b/paddle/operators/pool_op.cu.cc
@@ -17,11 +17,15 @@ limitations under the License. */
 namespace ops = paddle::operators;
 
 REGISTER_OP_GPU_KERNEL(pool2d,
-                       ops::PoolKernel<paddle::platform::GPUPlace, float>);
+                       ops::PoolKernel<paddle::platform::GPUPlace, float>,
+                       ops::PoolKernel<paddle::platform::GPUPlace, double>);
 REGISTER_OP_GPU_KERNEL(pool2d_grad,
-                       ops::PoolGradKernel<paddle::platform::GPUPlace, float>);
+                       ops::PoolGradKernel<paddle::platform::GPUPlace, float>,
+                       ops::PoolGradKernel<paddle::platform::GPUPlace, double>);
 
 REGISTER_OP_GPU_KERNEL(pool3d,
-                       ops::PoolKernel<paddle::platform::GPUPlace, float>);
+                       ops::PoolKernel<paddle::platform::GPUPlace, float>,
+                       ops::PoolKernel<paddle::platform::GPUPlace, double>);
 REGISTER_OP_GPU_KERNEL(pool3d_grad,
-                       ops::PoolGradKernel<paddle::platform::GPUPlace, float>);
+                       ops::PoolGradKernel<paddle::platform::GPUPlace, float>,
+                       ops::PoolGradKernel<paddle::platform::GPUPlace, double>);
diff --git a/paddle/operators/pool_with_index_op.cc b/paddle/operators/pool_with_index_op.cc
index 1df36e965abab3549aeb88bf682b712033c4d79c..4b95c7ef6b41a9e7d2fe4654ea9acd3437757f66 100644
--- a/paddle/operators/pool_with_index_op.cc
+++ b/paddle/operators/pool_with_index_op.cc
@@ -250,10 +250,12 @@ REGISTER_OP(max_pool2d_with_index, ops::MaxPoolWithIndexOp,
 
 REGISTER_OP_CPU_KERNEL(
     max_pool2d_with_index,
-    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, float>);
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, float>,
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
     max_pool2d_with_index_grad,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, float>)
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, float>,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, double>)
 
 REGISTER_OP(max_pool3d_with_index, ops::MaxPoolWithIndexOp,
             ops::MaxPool3dWithIndexOpMaker, max_pool3d_with_index_grad,
@@ -261,7 +263,9 @@ REGISTER_OP(max_pool3d_with_index, ops::MaxPoolWithIndexOp,
 
 REGISTER_OP_CPU_KERNEL(
     max_pool3d_with_index,
-    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, float>);
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, float>,
+    ops::MaxPoolWithIndexKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
     max_pool3d_with_index_grad,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, float>)
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, float>,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::CPUPlace, double>)
diff --git a/paddle/operators/pool_with_index_op.cu.cc b/paddle/operators/pool_with_index_op.cu.cc
index 287657d4b1c57f354ef050885f71261092bdc062..8764a71da0877a3e43fc9cab0b833199bb661962 100644
--- a/paddle/operators/pool_with_index_op.cu.cc
+++ b/paddle/operators/pool_with_index_op.cu.cc
@@ -18,14 +18,18 @@ namespace ops = paddle::operators;
 
 REGISTER_OP_GPU_KERNEL(
     max_pool2d_with_index,
-    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, float>);
+    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, float>,
+    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, double>);
 REGISTER_OP_GPU_KERNEL(
     max_pool2d_with_index_grad,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, float>)
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, float>,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, double>)
 
 REGISTER_OP_GPU_KERNEL(
     max_pool3d_with_index,
-    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, float>);
+    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, float>,
+    ops::MaxPoolWithIndexKernel<paddle::platform::GPUPlace, double>);
 REGISTER_OP_GPU_KERNEL(
     max_pool3d_with_index_grad,
-    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, float>)
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, float>,
+    ops::MaxPoolWithIndexGradKernel<paddle::platform::GPUPlace, double>)