diff --git a/cmake/configure.cmake b/cmake/configure.cmake
index 851520328f3ce74969dfc92ef73f1ba17ce482d3..2a1e6897c02e445b799815b0fdc498774e1f37ad 100644
--- a/cmake/configure.cmake
+++ b/cmake/configure.cmake
@@ -100,6 +100,11 @@ if(WITH_GPU)
     if(NOT CUDNN_FOUND)
         message(FATAL_ERROR "Paddle needs cudnn to compile")
     endif()
+
+    if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
+        message(FATAL_ERROR "Paddle needs CUDNN >= 7.0 to compile")
+    endif()
+
     if(CUPTI_FOUND)
         include_directories(${CUPTI_INCLUDE_DIR})
         add_definitions(-DPADDLE_WITH_CUPTI)
diff --git a/paddle/fluid/platform/dynload/cudnn.cc b/paddle/fluid/platform/dynload/cudnn.cc
index 4c59fe5e9bae4b751d87b0d2feb1ea0bd02bcf1d..366762401c741e570204f5c9d146343a3d60aa33 100644
--- a/paddle/fluid/platform/dynload/cudnn.cc
+++ b/paddle/fluid/platform/dynload/cudnn.cc
@@ -24,26 +24,9 @@ void* cudnn_dso_handle = nullptr;
 #define DEFINE_WRAP(__name) DynLoad__##__name __name
 
 CUDNN_DNN_ROUTINE_EACH(DEFINE_WRAP);
-CUDNN_DNN_ROUTINE_EACH_R2(DEFINE_WRAP);
 
-#ifdef CUDNN_DNN_ROUTINE_EACH_AFTER_R3
-CUDNN_DNN_ROUTINE_EACH_AFTER_R3(DEFINE_WRAP);
-#endif
-
-#ifdef CUDNN_DNN_ROUTINE_EACH_AFTER_R3_LESS_R8
-CUDNN_DNN_ROUTINE_EACH_AFTER_R3_LESS_R8(DEFINE_WRAP);
-#endif
-
-#ifdef CUDNN_DNN_ROUTINE_EACH_AFTER_R4
-CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DEFINE_WRAP);
-#endif
-
-#ifdef CUDNN_DNN_ROUTINE_EACH_R5
-CUDNN_DNN_ROUTINE_EACH_R5(DEFINE_WRAP);
-#endif
-
-#ifdef CUDNN_DNN_ROUTINE_EACH_R6
-CUDNN_DNN_ROUTINE_EACH_R6(DEFINE_WRAP);
+#ifdef CUDNN_DNN_ROUTINE_EACH_AFTER_R7_LESS_R8
+CUDNN_DNN_ROUTINE_EACH_AFTER_R7_LESS_R8(DEFINE_WRAP);
 #endif
 
 #ifdef CUDNN_DNN_ROUTINE_EACH_R7
diff --git a/paddle/fluid/platform/dynload/cudnn.h b/paddle/fluid/platform/dynload/cudnn.h
index db84b8731f9ca467c4521221a3dbe0b1fc61b597..f5045ff004ee9b8391a879c402c679f0078487a2 100644
--- a/paddle/fluid/platform/dynload/cudnn.h
+++ b/paddle/fluid/platform/dynload/cudnn.h
@@ -48,121 +48,93 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
  * include all needed cudnn functions in HPPL
  * different cudnn version has different interfaces
  **/
-#define CUDNN_DNN_ROUTINE_EACH(__macro)                   \
-  __macro(cudnnSetTensor4dDescriptor);                    \
-  __macro(cudnnSetTensor4dDescriptorEx);                  \
-  __macro(cudnnSetTensorNdDescriptor);                    \
-  __macro(cudnnGetTensorNdDescriptor);                    \
-  __macro(cudnnGetConvolutionNdForwardOutputDim);         \
-  __macro(cudnnCreateTensorDescriptor);                   \
-  __macro(cudnnDestroyTensorDescriptor);                  \
-  __macro(cudnnCreateFilterDescriptor);                   \
-  __macro(cudnnSetFilter4dDescriptor);                    \
-  __macro(cudnnSetFilterNdDescriptor);                    \
-  __macro(cudnnGetFilterNdDescriptor);                    \
-  __macro(cudnnSetPooling2dDescriptor);                   \
-  __macro(cudnnSetPoolingNdDescriptor);                   \
-  __macro(cudnnGetPoolingNdDescriptor);                   \
-  __macro(cudnnDestroyFilterDescriptor);                  \
-  __macro(cudnnCreateConvolutionDescriptor);              \
-  __macro(cudnnCreatePoolingDescriptor);                  \
-  __macro(cudnnDestroyPoolingDescriptor);                 \
-  __macro(cudnnSetConvolution2dDescriptor);               \
-  __macro(cudnnDestroyConvolutionDescriptor);             \
-  __macro(cudnnSetConvolutionNdDescriptor);               \
-  __macro(cudnnGetConvolutionNdDescriptor);               \
-  __macro(cudnnDeriveBNTensorDescriptor);                 \
-  __macro(cudnnCreateSpatialTransformerDescriptor);       \
-  __macro(cudnnSetSpatialTransformerNdDescriptor);        \
-  __macro(cudnnDestroySpatialTransformerDescriptor);      \
-  __macro(cudnnSpatialTfGridGeneratorForward);            \
-  __macro(cudnnSpatialTfGridGeneratorBackward);           \
-  __macro(cudnnSpatialTfSamplerForward);                  \
-  __macro(cudnnSpatialTfSamplerBackward);                 \
-  __macro(cudnnCreate);                                   \
-  __macro(cudnnDestroy);                                  \
-  __macro(cudnnSetStream);                                \
-  __macro(cudnnActivationForward);                        \
-  __macro(cudnnActivationBackward);                       \
-  __macro(cudnnConvolutionForward);                       \
-  __macro(cudnnConvolutionBackwardBias);                  \
-  __macro(cudnnGetConvolutionForwardWorkspaceSize);       \
-  __macro(cudnnTransformTensor);                          \
-  __macro(cudnnPoolingForward);                           \
-  __macro(cudnnPoolingBackward);                          \
-  __macro(cudnnSoftmaxBackward);                          \
-  __macro(cudnnSoftmaxForward);                           \
-  __macro(cudnnGetVersion);                               \
-  __macro(cudnnFindConvolutionForwardAlgorithmEx);        \
-  __macro(cudnnFindConvolutionBackwardFilterAlgorithmEx); \
-  __macro(cudnnFindConvolutionBackwardFilterAlgorithm);   \
-  __macro(cudnnFindConvolutionBackwardDataAlgorithmEx);   \
-  __macro(cudnnGetErrorString);                           \
-  __macro(cudnnCreateDropoutDescriptor);                  \
-  __macro(cudnnDropoutGetStatesSize);                     \
-  __macro(cudnnSetDropoutDescriptor);                     \
-  __macro(cudnnRestoreDropoutDescriptor);                 \
-  __macro(cudnnCreateRNNDescriptor);                      \
-  __macro(cudnnGetRNNParamsSize);                         \
-  __macro(cudnnGetRNNWorkspaceSize);                      \
-  __macro(cudnnGetRNNTrainingReserveSize);                \
-  __macro(cudnnRNNForwardTraining);                       \
-  __macro(cudnnRNNBackwardData);                          \
-  __macro(cudnnRNNBackwardWeights);                       \
-  __macro(cudnnRNNForwardInference);                      \
-  __macro(cudnnDestroyDropoutDescriptor);                 \
-  __macro(cudnnDestroyRNNDescriptor);                     \
-  __macro(cudnnSetTensorNdDescriptorEx);
-
-CUDNN_DNN_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
-
-#define CUDNN_DNN_ROUTINE_EACH_R2(__macro) \
-  __macro(cudnnAddTensor);                 \
-  __macro(cudnnConvolutionBackwardData);   \
-  __macro(cudnnConvolutionBackwardFilter);
-CUDNN_DNN_ROUTINE_EACH_R2(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
-
-// APIs available after R3:
-#if CUDNN_VERSION >= 3000
-#define CUDNN_DNN_ROUTINE_EACH_AFTER_R3(__macro)           \
+#define CUDNN_DNN_ROUTINE_EACH(__macro)                    \
+  __macro(cudnnSetTensor4dDescriptor);                     \
+  __macro(cudnnSetTensor4dDescriptorEx);                   \
+  __macro(cudnnSetTensorNdDescriptor);                     \
+  __macro(cudnnGetTensorNdDescriptor);                     \
+  __macro(cudnnGetConvolutionNdForwardOutputDim);          \
+  __macro(cudnnCreateTensorDescriptor);                    \
+  __macro(cudnnDestroyTensorDescriptor);                   \
+  __macro(cudnnCreateFilterDescriptor);                    \
+  __macro(cudnnSetFilter4dDescriptor);                     \
+  __macro(cudnnSetFilterNdDescriptor);                     \
+  __macro(cudnnGetFilterNdDescriptor);                     \
+  __macro(cudnnSetPooling2dDescriptor);                    \
+  __macro(cudnnSetPoolingNdDescriptor);                    \
+  __macro(cudnnGetPoolingNdDescriptor);                    \
+  __macro(cudnnDestroyFilterDescriptor);                   \
+  __macro(cudnnCreateConvolutionDescriptor);               \
+  __macro(cudnnCreatePoolingDescriptor);                   \
+  __macro(cudnnDestroyPoolingDescriptor);                  \
+  __macro(cudnnSetConvolution2dDescriptor);                \
+  __macro(cudnnDestroyConvolutionDescriptor);              \
+  __macro(cudnnSetConvolutionNdDescriptor);                \
+  __macro(cudnnGetConvolutionNdDescriptor);                \
+  __macro(cudnnDeriveBNTensorDescriptor);                  \
+  __macro(cudnnCreateSpatialTransformerDescriptor);        \
+  __macro(cudnnSetSpatialTransformerNdDescriptor);         \
+  __macro(cudnnDestroySpatialTransformerDescriptor);       \
+  __macro(cudnnSpatialTfGridGeneratorForward);             \
+  __macro(cudnnSpatialTfGridGeneratorBackward);            \
+  __macro(cudnnSpatialTfSamplerForward);                   \
+  __macro(cudnnSpatialTfSamplerBackward);                  \
+  __macro(cudnnCreate);                                    \
+  __macro(cudnnDestroy);                                   \
+  __macro(cudnnSetStream);                                 \
+  __macro(cudnnActivationForward);                         \
+  __macro(cudnnActivationBackward);                        \
+  __macro(cudnnConvolutionForward);                        \
+  __macro(cudnnConvolutionBackwardBias);                   \
+  __macro(cudnnGetConvolutionForwardWorkspaceSize);        \
+  __macro(cudnnTransformTensor);                           \
+  __macro(cudnnPoolingForward);                            \
+  __macro(cudnnPoolingBackward);                           \
+  __macro(cudnnSoftmaxBackward);                           \
+  __macro(cudnnSoftmaxForward);                            \
+  __macro(cudnnGetVersion);                                \
+  __macro(cudnnFindConvolutionForwardAlgorithmEx);         \
+  __macro(cudnnFindConvolutionBackwardFilterAlgorithmEx);  \
+  __macro(cudnnFindConvolutionBackwardFilterAlgorithm);    \
+  __macro(cudnnFindConvolutionBackwardDataAlgorithmEx);    \
+  __macro(cudnnGetErrorString);                            \
+  __macro(cudnnCreateDropoutDescriptor);                   \
+  __macro(cudnnDropoutGetStatesSize);                      \
+  __macro(cudnnSetDropoutDescriptor);                      \
+  __macro(cudnnRestoreDropoutDescriptor);                  \
+  __macro(cudnnCreateRNNDescriptor);                       \
+  __macro(cudnnGetRNNParamsSize);                          \
+  __macro(cudnnGetRNNWorkspaceSize);                       \
+  __macro(cudnnGetRNNTrainingReserveSize);                 \
+  __macro(cudnnRNNForwardTraining);                        \
+  __macro(cudnnRNNBackwardData);                           \
+  __macro(cudnnRNNBackwardWeights);                        \
+  __macro(cudnnRNNForwardInference);                       \
+  __macro(cudnnDestroyDropoutDescriptor);                  \
+  __macro(cudnnDestroyRNNDescriptor);                      \
+  __macro(cudnnSetTensorNdDescriptorEx);                   \
+  __macro(cudnnAddTensor);                                 \
+  __macro(cudnnConvolutionBackwardData);                   \
+  __macro(cudnnConvolutionBackwardFilter);                 \
   __macro(cudnnGetConvolutionBackwardFilterWorkspaceSize); \
-  __macro(cudnnGetConvolutionBackwardDataWorkspaceSize);
-CUDNN_DNN_ROUTINE_EACH_AFTER_R3(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
-#endif
+  __macro(cudnnGetConvolutionBackwardDataWorkspaceSize);   \
+  __macro(cudnnBatchNormalizationForwardTraining);         \
+  __macro(cudnnBatchNormalizationForwardInference);        \
+  __macro(cudnnBatchNormalizationBackward);                \
+  __macro(cudnnCreateActivationDescriptor);                \
+  __macro(cudnnSetActivationDescriptor);                   \
+  __macro(cudnnGetActivationDescriptor);                   \
+  __macro(cudnnDestroyActivationDescriptor);               \
+  __macro(cudnnSetRNNDescriptor_v6);
+CUDNN_DNN_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 
-// APIs available after R3:
-#if CUDNN_VERSION >= 3000 && CUDNN_VERSION < 8000
-#define CUDNN_DNN_ROUTINE_EACH_AFTER_R3_LESS_R8(__macro) \
+#if CUDNN_VERSION >= 7000 && CUDNN_VERSION < 8000
+#define CUDNN_DNN_ROUTINE_EACH_AFTER_R7_LESS_R8(__macro) \
   __macro(cudnnGetConvolutionBackwardFilterAlgorithm);   \
   __macro(cudnnGetConvolutionForwardAlgorithm);          \
   __macro(cudnnGetConvolutionBackwardDataAlgorithm);     \
   __macro(cudnnSetRNNDescriptor);
-CUDNN_DNN_ROUTINE_EACH_AFTER_R3_LESS_R8(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
-#endif
-
-// APIs available after R4:
-#if CUDNN_VERSION >= 4007
-#define CUDNN_DNN_ROUTINE_EACH_AFTER_R4(__macro)    \
-  __macro(cudnnBatchNormalizationForwardTraining);  \
-  __macro(cudnnBatchNormalizationForwardInference); \
-  __macro(cudnnBatchNormalizationBackward);
-CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
-#endif
-
-// APIs in R5
-#if CUDNN_VERSION >= 5000
-#define CUDNN_DNN_ROUTINE_EACH_R5(__macro)  \
-  __macro(cudnnCreateActivationDescriptor); \
-  __macro(cudnnSetActivationDescriptor);    \
-  __macro(cudnnGetActivationDescriptor);    \
-  __macro(cudnnDestroyActivationDescriptor);
-CUDNN_DNN_ROUTINE_EACH_R5(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
-#endif
-
-// APIs in R6
-#if CUDNN_VERSION >= 6000
-#define CUDNN_DNN_ROUTINE_EACH_R6(__macro) __macro(cudnnSetRNNDescriptor_v6);
-CUDNN_DNN_ROUTINE_EACH_R6(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
+CUDNN_DNN_ROUTINE_EACH_AFTER_R7_LESS_R8(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 #endif
 
 #if CUDNN_VERSION >= 7001