[LITE][OPENCL] Fix OpenCL API/Backend misc. test=develop (#3376)

1. clean code;
2. change `cl::Kernel` from unique to shared ptr;
3. `reset` `cl::Program` and `clear` `device_info_` in destroyed of CLRuntime;
4. remove clFlush in destroyed of CLRuntime.

lite/api/light_api.cc

@@ -141,9 +141,6 @@ void LightPredictor::BuildRuntimeProgram(const cpp::ProgramDesc& prog) {
 
 // 2. Create Instructs
 #ifdef LITE_WITH_OPENCL
-  using WaitListType =
-      std::unordered_map<decltype(static_cast<const void*>(nullptr)),
-                         std::shared_ptr<cl::Event>>;
   using OpenCLContext = Context<TargetType::kOpenCL>;
   std::unique_ptr<KernelContext> local_ctx(new KernelContext());
   local_ctx->As<OpenCLContext>().InitOnce();

lite/backends/opencl/cl_context.cc

@@ -58,7 +58,7 @@ void CLContext::AddKernel(const std::string &kernel_name,
   auto program = GetProgram(file_name, options);
   VLOG(3) << " --- end get program --- ";
   VLOG(3) << " --- to create kernel: " << kernel_name << " --- ";
-  std::unique_ptr<cl::Kernel> kernel(
+  std::shared_ptr<cl::Kernel> kernel(
       new cl::Kernel(program, kernel_name.c_str(), &status));
   CL_CHECK_FATAL(status);
   VLOG(3) << " --- end create kernel --- ";

lite/backends/opencl/cl_context.h

@@ -36,6 +36,7 @@ class CLContext {
     kernel_offset_.clear();
     for (auto &p : programs_) {
       clReleaseProgram(p.second->get());
+      p.second.reset();
     }
     programs_.clear();
     LOG(INFO) << "release cl::Program, cl::Kernel finished.";
@@ -66,9 +67,10 @@ class CLContext {
                                 int divitor = 2);
   //  cl::NDRange LocalWorkSizeConv1x1(cl::NDRange global_work_size,
   //                                   size_t max_work_size);
+
  private:
   std::unordered_map<std::string, std::unique_ptr<cl::Program>> programs_;
-  std::vector<std::unique_ptr<cl::Kernel>> kernels_;
+  std::vector<std::shared_ptr<cl::Kernel>> kernels_;
   std::map<std::string, int> kernel_offset_;
 };
 

lite/backends/opencl/cl_runtime.cc

@@ -29,12 +29,12 @@ CLRuntime::~CLRuntime() {
     command_queue_->flush();
     command_queue_->finish();
   }
-  // For controlling the destruction order:
+  // For controlling the destruction order
   command_queue_.reset();
   context_.reset();
   device_.reset();
   platform_.reset();
-  LOG(INFO) << "release ~CLRuntime() ";
+  device_info_.clear();
 }
 
 bool CLRuntime::Init() {

lite/kernels/opencl/io_copy_buffer_compute.cc

@@ -115,7 +115,6 @@ class IoCopykOpenCLToHostCompute
       auto& event = *(it->second);
       event.wait();
       auto command_queue = CLRuntime::Global()->command_queue();
-      command_queue.flush();
       command_queue.finish();
     } else {
       LOG(FATAL) << "Could not find the sync event for the target cl tensor.";