[LITE][OPENCL] Add ReleaseResource for OpenCL when Predictor dead (#3358)

* [LITE][OPENCL] Add ReleaseResource for OpenCL when Predictor dead. test=develop

* fix void for decontrust. test=develop

* fix miscs. test=develop

* fix miscs. test=develop

* fix miscs. test=develop

* fix miscs. test=develop

* [LITE][OPENCL] fix Hang of mobilenetv1_test and kernel test. test=develop

* [LITE][OPENCL] Fix miscs. test is ok. test=develop

lite/api/cxx_api.h

@@ -43,6 +43,16 @@ class LITE_API Predictor {
  public:
   // Create an empty predictor.
   Predictor() { scope_ = std::make_shared<Scope>(); }
+  ~Predictor() {
+#ifdef LITE_WITH_OPENCL
+    CLRuntime::Global()->ReleaseResources();
+#endif
+    scope_.reset();
+    exec_scope_ = nullptr;
+    program_.reset();
+    input_names_.clear();
+    output_names_.clear();
+  }
   // Create a predictor with the weight variable scope set.
   explicit Predictor(const std::shared_ptr<lite::Scope>& root_scope)
       : scope_(root_scope) {}

lite/api/light_api.h

@@ -107,6 +107,8 @@ class LightPredictorImpl : public lite_api::PaddlePredictor {
  public:
   LightPredictorImpl() = default;
 
+  ~LightPredictorImpl();
+
   std::unique_ptr<lite_api::Tensor> GetInput(int i) override;
 
   std::unique_ptr<const lite_api::Tensor> GetOutput(int i) const override;

lite/api/light_api_impl.cc

@@ -21,6 +21,13 @@
 namespace paddle {
 namespace lite {
 
+LightPredictorImpl::~LightPredictorImpl() {
+  raw_predictor_.reset();
+#ifdef LITE_WITH_OPENCL
+  CLRuntime::Global()->ReleaseResources();
+#endif
+}
+
 void LightPredictorImpl::Init(const lite_api::MobileConfig& config) {
   // LightPredictor Only support NaiveBuffer backend in publish lib
   if (config.lite_model_file().empty()) {

lite/backends/opencl/cl_runtime.cc

@@ -29,30 +29,38 @@ CLRuntime* CLRuntime::Global() {
 }
 
 CLRuntime::~CLRuntime() {
+  LOG(INFO) << "CLRuntime::~CLRuntime()";
+  // Note: do ReleaseResources() in predictor
+  command_queue_&& clReleaseCommandQueue(command_queue_->get());
+  command_queue_.reset();
+  context_&& clReleaseContext(context_->get());
+  context_.reset();
+  device_.reset();
+  platform_.reset();
+  initialized_ = false;
+}
+
+void CLRuntime::ReleaseResources() {
+  //  if (is_resources_released_) {
+  //    return;
+  //  }
+
   if (command_queue_ != nullptr) {
     command_queue_->flush();
     command_queue_->finish();
   }
-
   for (size_t kidx = 0; kidx < kernels_.size(); ++kidx) {
     clReleaseKernel(kernels_[kidx]->get());
     kernels_[kidx].reset();
   }
   kernels_.clear();
   kernel_offset_.clear();
-
   for (auto& p : programs_) {
     clReleaseProgram(p.second->get());
   }
   programs_.clear();
-
-  // For controlling the destruction order
-  command_queue_&& clReleaseCommandQueue(command_queue_->get());
-  command_queue_.reset();
-  context_&& clReleaseContext(context_->get());
-  context_.reset();
-  device_.reset();
-  platform_.reset();
+  LOG(INFO) << "release resources finished.";
+  is_resources_released_ = true;
 }
 
 bool CLRuntime::Init() {

lite/backends/opencl/cl_runtime.h

@@ -33,6 +33,8 @@ class CLRuntime {
  public:
   static CLRuntime* Global();
 
+  void ReleaseResources();
+
   bool Init();
 
   cl::Platform& platform();
@@ -116,6 +118,8 @@ class CLRuntime {
   bool initialized_{false};
 
   bool is_init_success_{false};
+
+  bool is_resources_released_{false};
 };
 
 }  // namespace lite

lite/core/memory.h

@@ -126,7 +126,7 @@ class Buffer {
                         const size_t img_h,
                         void* host_ptr = nullptr) {
     if (target != target_ || cl_image2d_width_ < img_w ||
-        cl_image2d_height_ < img_h) {
+        cl_image2d_height_ < img_h || host_ptr != nullptr) {
       CHECK_EQ(own_data_, true) << "Can not reset unowned buffer.";
       Free();
       data_ = TargetWrapperCL::MallocImage<T>(img_w, img_h, host_ptr);