open cl mem optimise, split with cpu codes. fix a bug when some memory is not equal 4 . (#2161)

* open cl mem optimise, split with cpu codes.
fix a bug when some memory is not equal 4 .

* open cl mem optimise, split with cpu codes.   fix a bug when some memory is not equal 4  fix bad des

mobile/src/framework/executor.cpp

@@ -126,15 +126,24 @@ Executor<Device, T>::Executor(const Program<Device> &program,
   printf("================[ op init profile ]==================\n");
   PrintProfile(profile);
 #endif
+  ApplyMemoryOptimise(config, lod_mode);
+}
+
+template <typename Device, typename T>
+void Executor<Device, T>::ApplyMemoryOptimise(
+    const PaddleMobileConfigInternal &config, const bool lod_mode) const {}
 
 #ifdef PADDLE_MOBILE_CL
+template <>
+void Executor<GPU_CL, float>::ApplyMemoryOptimise(
+    const PaddleMobileConfigInternal &config, const bool lod_mode) const {
   if (!config.load_when_predict && !lod_mode &&
       config_.memory_optimization_level != NoMemoryOptimization) {
     pass::MemoryOptPassCl()(program_desc_.get(), program_.scope.get(),
                             config_.memory_optimization_level);
   }
-#endif
 }
+#endif
 
 template <typename Device, typename T>
 void Executor<Device, T>::InitFeedFetchList() {

mobile/src/framework/executor.h

@@ -118,6 +118,8 @@ class Executor {
 
   void PrintProfile(const vector<Executor<Device, T>::ProfInfo> &profile) const;
 #endif
+  void ApplyMemoryOptimise(const PaddleMobileConfigInternal &config,
+                           const bool lod_mode) const;
 };
 
 }  // namespace framework

mobile/src/pass/memory_optimize_cl.cpp

@@ -181,10 +181,10 @@ void MemoryOptPassCl::ShareData(
       const int64_t numl = tensor->numel();
       auto origin_tensor_dims = tensor->dims();
 
-      PADDLE_MOBILE_ENFORCE(origin_tensor_dims.size() == 4,
-                            "tensor dims must larger than 4");
       // for super ,hack origin dims
       if (target_dims.size() == 4) {
+        PADDLE_MOBILE_ENFORCE(origin_tensor_dims.size() == 4,
+                              "tensor dims must be equal to 4");
         origin_tensor_dims = {origin_tensor_dims[0], origin_tensor_dims[1],
                               target_dims[2], target_dims[3]};
         tensor->Resize(origin_tensor_dims);