!5356 add reference count for opencl runtime

Merge pull request !5356 from liuchao/master

mindspore/lite/src/lite_session.cc

@@ -326,6 +326,11 @@ LiteSession::~LiteSession() {
     }
   }
   input_vec_.clear();
+#if SUPPORT_GPU
+  if (context_->device_ctx_.type == DT_GPU) {
+    lite::opencl::OpenCLRuntime::DeleteInstance();
+  }
+#endif
   delete this->context_;
   delete this->executor;
   this->executor = nullptr;

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

@@ -24,7 +24,7 @@
 namespace mindspore::lite::opencl {
 
 OpenCLAllocator::OpenCLAllocator() {}
-OpenCLAllocator::~OpenCLAllocator() {}
+OpenCLAllocator::~OpenCLAllocator() { Clear(); }
 
 void OpenCLAllocator::SetContext(const AllocatorContext &ctx) {
   lock_flag_ = ctx.lockFlag;

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

@@ -40,12 +40,29 @@ static std::mutex g_mtx;
 static std::mutex g_init_mtx;
 
 bool OpenCLRuntime::init_done_ = false;
+OpenCLRuntime *OpenCLRuntime::ocl_runtime_instance_ = nullptr;
+size_t OpenCLRuntime::instance_count_ = 0;
 
 OpenCLRuntime *OpenCLRuntime::GetInstance() {
   std::unique_lock<std::mutex> lck(g_mtx);
   static OpenCLRuntime ocl_runtime;
-  ocl_runtime.Init();
-  return &ocl_runtime;
+  if (instance_count_ == 0) {
+    ocl_runtime_instance_ = &ocl_runtime;
+    ocl_runtime_instance_->Init();
+  }
+  instance_count_++;
+  return ocl_runtime_instance_;
+}
+
+void OpenCLRuntime::DeleteInstance() {
+  std::unique_lock<std::mutex> lck(g_mtx);
+  if (instance_count_ == 0) {
+    MS_LOG(ERROR) << "No OpenCLRuntime instance could delete!";
+  }
+  instance_count_--;
+  if (instance_count_ == 0) {
+    ocl_runtime_instance_->Uninit();
+  }
 }
 
 OpenCLRuntime::OpenCLRuntime() { default_build_opts_ = " -cl-mad-enable -cl-fast-relaxed-math -Werror"; }
@@ -207,16 +224,25 @@ int OpenCLRuntime::Init() {
   return RET_OK;
 }
 
-OpenCLRuntime::~OpenCLRuntime() {
-  init_done_ = false;
+int OpenCLRuntime::Uninit() {
   program_map_.clear();
   delete allocator_;
   delete default_command_queue_;
   delete context_;
   delete device_;
+  allocator_ = nullptr;
+  default_command_queue_ = nullptr;
+  context_ = nullptr;
+  device_ = nullptr;
+#ifdef USE_OPENCL_WRAPPER
   OpenCLWrapper::GetInstance()->UnLoadOpenCLLibrary();
+#endif
+  init_done_ = false;
+  return RET_OK;
 }
 
+OpenCLRuntime::~OpenCLRuntime() { Uninit(); }
+
 cl::Context *OpenCLRuntime::Context() { return context_; }
 
 cl::Device *OpenCLRuntime::Device() { return device_; }

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

@@ -47,6 +47,7 @@ class OpenCLRuntime {
   OpenCLRuntime &operator=(const OpenCLRuntime &) = delete;
 
   int Init();
+  int Uninit();
 
   cl::Context *Context();
   cl::Device *Device();
@@ -143,6 +144,8 @@ class OpenCLRuntime {
 
  private:
   static bool init_done_;
+  static size_t instance_count_;
+  static OpenCLRuntime *ocl_runtime_instance_;
   cl::CommandQueue *default_command_queue_{nullptr};
   cl::Context *context_{nullptr};
   cl::Device *device_{nullptr};

mindspore/lite/test/ut/src/runtime/kernel/opencl/activation_tests.cc

@@ -377,6 +377,7 @@ TEST_F(TestActivationOpenCL, SigmoidFp32_dim4) {
   delete input_tensor;
   delete output_tensor;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 TEST_F(TestActivationOpenCL, LeakyReluFp32_dim4) {
@@ -480,5 +481,6 @@ TEST_F(TestActivationOpenCL, LeakyReluFp32_dim4) {
   delete input_tensor;
   delete output_tensor;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/arithmetic_tests.cc

@@ -202,6 +202,7 @@ void TestCase(const std::vector<int> &shape_a, const std::vector<int> &shape_b)
   for (auto tensor : outputs) {
     delete tensor;
   }
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 class TestArithmeticOpenCL : public mindspore::CommonTest {

mindspore/lite/test/ut/src/runtime/kernel/opencl/avg_pooling_tests.cc

@@ -142,6 +142,7 @@ TEST_F(TestAvgPoolingOpenCL, AvgPoolFp32) {
   delete pooling_kernel;
   delete pGraph;
   delete param;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/batchnorm_tests.cc

@@ -158,6 +158,7 @@ TEST_F(TestBatchnormOpenCLfp16, Batchnormfp16input_dim4) {
   delete param;
   delete batchnorm_kernel;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 TEST_F(TestBatchnormOpenCLfp32, Batchnormfp32input_dim4) {
   MS_LOG(INFO) << "begin test";
@@ -277,5 +278,6 @@ TEST_F(TestBatchnormOpenCLfp32, Batchnormfp32input_dim4) {
   delete param;
   delete batchnorm_kernel;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/biasadd_tests.cc

@@ -198,5 +198,6 @@ TEST_F(TestBiasAddOpenCL, BiasAddFp32_dim4) {
   delete sub_graph;
   delete param;
   delete biasadd_kernel;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/concat_tests.cc

@@ -218,6 +218,7 @@ TEST_F(TestConcatOpenCLfp16, ConcatFp16_2input_dim4_axis3) {
   delete param;
   delete concat_kernel;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 TEST_F(TestConcatOpenCLfp32, ConcatFp32_2input_dim4_axis3) {
@@ -338,5 +339,6 @@ TEST_F(TestConcatOpenCLfp32, ConcatFp32_2input_dim4_axis3) {
   delete param;
   delete concat_kernel;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/conv2d_transpose_tests.cc

@@ -134,6 +134,7 @@ void RunTestCaseConv2dTranspose(const std::vector<int> &shape, void *input_data,
 
   inputs[0]->SetData(nullptr);
   outputs[0]->SetData(nullptr);
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 void RunTestCaseConv2dTranspose(const std::vector<int> shape, const std::vector<std::string> file_path, bool fp16) {

mindspore/lite/test/ut/src/runtime/kernel/opencl/convolution_tests.cc

@@ -158,6 +158,7 @@ void TEST_MAIN(schema::Format input_format, schema::Format output_format, const
   bias_tensor.SetData(nullptr);
   delete param;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 TEST_F(TestConvolutionOpenCL, in1x224x224x3_out1x112x112x32_k33_s22_p0101_fp32) {

mindspore/lite/test/ut/src/runtime/kernel/opencl/depthwise_conv2d_tests.cc

@@ -162,6 +162,7 @@ void DepthWiseTestMain(ConvParameter *conv_param, T2 *input_data, T1 *weight_dat
   inputs[1]->SetData(nullptr);
   inputs[2]->SetData(nullptr);
   delete[] packed_input;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
   return;
 }
 
@@ -587,5 +588,6 @@ TEST_F(TestConvolutionDwOpenCL, ProfilingMobilenetv2Fp32) {
   }
   delete [] input_data;
   delete [] weight_data;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/matmul_tests.cc

@@ -111,6 +111,7 @@ void RunTestCaseMatMul(const std::vector<int> shape, const std::vector<std::stri
   tensor_x->SetData(nullptr);
   tensor_out->SetData(nullptr);
   MS_LOG(INFO) << "TestMatMulFp32 passed";
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 TEST_F(TestMatMulOpenCL, MatMulFp32) {

mindspore/lite/test/ut/src/runtime/kernel/opencl/max_pooling_tests.cc

@@ -118,6 +118,7 @@ TEST_F(TestMaxPoolingOpenCL, MaxPool_1_32_512_96) {
   }
   delete pooling_kernel;
   delete pGraph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/prelu_tests.cc

@@ -183,5 +183,6 @@ TEST_F(TestPReluOpenCL, PReluFp32_dim4) {
   delete param;
   delete prelu_kernel;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/reshape_tests.cc

@@ -106,5 +106,6 @@ TEST_F(TestReshapeOpenCL, ReshapeFp32) {
   outputs[0]->SetData(nullptr);
 
   MS_LOG(INFO) << "Test ReshapeFp32 passed";
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/slice_tests.cc

@@ -148,6 +148,7 @@ TEST_F(TestSliceOpenCLfp32, Slicefp32input_dim4) {
   }
   delete slice_kernel;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 TEST_F(TestSliceOpenCLfp16, Slicefp16input_dim4) {
   MS_LOG(INFO) << "begin test";
@@ -258,5 +259,6 @@ TEST_F(TestSliceOpenCLfp16, Slicefp16input_dim4) {
   }
   delete slice_kernel;
   delete sub_graph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/softmax_tests.cc

@@ -92,6 +92,7 @@ void RunTestCase(std::vector<int> input_shape, std::vector<int> output_shape, st
   }
   delete kernel;
   delete pGraph;
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 
 TEST_F(TestSoftmaxOpenCL, Softmax_1) {

mindspore/lite/test/ut/src/runtime/kernel/opencl/to_format_tests.cc

@@ -103,5 +103,6 @@ TEST_F(TestToFormatOpenCL, ToFormatNHWC2NCHW) {
   // compare
   CompareOutputData(output_data, correct_data, h * w * c, 0.00001);
   MS_LOG(INFO) << "Test TransposeFp32 passed";
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/opencl/transpose_tests.cc

@@ -108,5 +108,6 @@ TEST_F(TestTransposeOpenCL, TransposeFp32) {
   outputs[0]->SetData(nullptr);
 
   MS_LOG(INFO) << "Test TransposeFp32 passed";
+  lite::opencl::OpenCLRuntime::DeleteInstance();
 }
 }  // namespace mindspore