!10478 [MS][LITE][GPU]fix memory leak bug

From: @chenzupeng
Reviewed-by: @zhanghaibo5,@ddwsky
Signed-off-by: @ddwsky

build.sh

@@ -503,7 +503,7 @@ build_lite()
 
     LITE_ENABLE_GPU=${ENABLE_GPU}
     LITE_ENABLE_NPU=${ENABLE_NPU}
-    if [ "${DEVICE}" == "" ] && [ "${LITE_PLATFORM}" == "arm64" ]; then
+    if [[ "${DEVICE}" == ""  &&  "${LITE_PLATFORM}" == "arm64" ]]; then
       LITE_ENABLE_GPU="on"
       LITE_ENABLE_NPU="on"
     fi

mindspore/lite/src/lite_session.cc

@@ -670,7 +670,7 @@ int LiteSession::Resize(const std::vector<mindspore::tensor::MSTensor *> &inputs
 }
 
 int LiteSession::InitGPURuntime() {
-#if SUPPORT_GPU
+#if SUPPORT_GPU && !SUPPORT_TRAIN
   if (this->context_->IsGpuEnabled()) {
     auto gpu_device_info = this->context_->GetGpuInfo();
     auto opencl_runtime = ocl_runtime_wrap_.GetInstance();

mindspore/lite/src/lite_session.h

@@ -124,7 +124,7 @@ class LiteSession : public session::LiteSession {
   std::unordered_map<std::string, mindspore::tensor::MSTensor *> output_tensor_map_;
   Executor *executor_ = nullptr;
   std::atomic<bool> is_running_ = false;
-#if SUPPORT_GPU
+#if SUPPORT_GPU && !SUPPORT_TRAIN
   opencl::OpenCLRuntimeWrapper ocl_runtime_wrap_;
 #endif
 };

mindspore/lite/src/runtime/opencl/opencl_runtime.cc

@@ -234,8 +234,7 @@ int OpenCLRuntime::Init() {
   MS_LOG(INFO) << "Compute Unit: " << compute_units_;
   MS_LOG(INFO) << "Clock Frequency: " << max_freq_ << " MHz";
 
-  const cl_command_queue_properties properties = 0;
-  default_command_queue_ = new (std::nothrow) cl::CommandQueue(*context_, *device_, properties, &ret);
+  default_command_queue_ = new (std::nothrow) cl::CommandQueue(*context_, *device_, 0, &ret);
   if (ret != CL_SUCCESS) {
     delete device_;
     delete context_;
@@ -243,8 +242,7 @@ int OpenCLRuntime::Init() {
     return RET_ERROR;
   }
 
-  const cl_command_queue_properties profiling_properties = CL_QUEUE_PROFILING_ENABLE;
-  profiling_command_queue_ = new (std::nothrow) cl::CommandQueue(*context_, *device_, profiling_properties, &ret);
+  profiling_command_queue_ = new (std::nothrow) cl::CommandQueue(*context_, *device_, CL_QUEUE_PROFILING_ENABLE, &ret);
   if (ret != CL_SUCCESS) {
     delete device_;
     delete context_;
@@ -258,6 +256,7 @@ int OpenCLRuntime::Init() {
     delete device_;
     delete context_;
     delete default_command_queue_;
+    delete profiling_command_queue_;
     MS_LOG(ERROR) << "Command OpenCL allocator failed!";
     return RET_ERROR;
   }
@@ -275,6 +274,9 @@ int OpenCLRuntime::Init() {
 }
 
 int OpenCLRuntime::Uninit() {
+  if (!init_done_) {
+    return RET_OK;
+  }
   if (enable_cache_ && !binary_map_.empty()) {
     StoreCache();
   }
@@ -282,10 +284,12 @@ int OpenCLRuntime::Uninit() {
   program_map_.clear();
   delete allocator_;
   delete default_command_queue_;
+  delete profiling_command_queue_;
   delete context_;
   delete device_;
   allocator_ = nullptr;
   default_command_queue_ = nullptr;
+  profiling_command_queue_ = nullptr;
   context_ = nullptr;
   device_ = nullptr;
 #ifdef USE_OPENCL_WRAPPER