update.

Author: jiweibo

paddle/fluid/inference/api/analysis_predictor.cc

@@ -889,9 +889,9 @@ bool AnalysisPredictor::LoadParameters() {
   return true;
 }
 
-void AnalysisPredictor::ShrinkMemory() {
+uint64_t AnalysisPredictor::TryShrinkMemory() {
   ClearIntermediateTensor();
-  paddle::memory::Release(place_);
+  return paddle::memory::Release(place_);
 }
 
 void AnalysisPredictor::ClearIntermediateTensor() {
@@ -982,6 +982,8 @@ AnalysisPredictor::~AnalysisPredictor() {
   if (sub_scope_) {
     scope_->DeleteScope(sub_scope_);
   }
+  // TODO(wilber): release weight memory.
+  paddle::memory::Release(place_);
 
 #if PADDLE_WITH_MKLDNN
   if (mkldnn_quantizer_) {
@@ -1146,7 +1148,7 @@ void Predictor::ClearIntermediateTensor() {
   predictor_->ClearIntermediateTensor();
 }
 
-void Predictor::ShrinkMemory() { predictor_->ShrinkMemory(); }
+uint64_t Predictor::TryShrinkMemory() { return predictor_->TryShrinkMemory(); }
 
 int GetNumBytesOfDataType(DataType dtype) {
   switch (dtype) {

paddle/fluid/inference/api/analysis_predictor.h

@@ -195,9 +195,14 @@ class AnalysisPredictor : public PaddlePredictor {
 
   ///
   /// \brief Release all tmp tensor to compress the size of the memory pool.
-  /// After this operation, we reduced the memory usage.
+  /// The memory pool is considered to be composed of a list of chunks, if
+  /// the chunk is not occupied, it can be released.
   ///
-  void ShrinkMemory() override;
+  /// \return Number of bytes released. It may be smaller than the actual
+  /// released memory, because part of the memory is not managed by the
+  /// MemoryPool.
+  ///
+  uint64_t TryShrinkMemory() override;
 
   ///
   /// \brief Get the argument used by predictor

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -135,7 +135,7 @@ TEST(AnalysisPredictor, ZeroCopy) {
   auto* out_data = out->data<float>(&place, &size);
   LOG(INFO) << "output size: " << size / sizeof(float);
   LOG(INFO) << "output_data: " << out_data;
-  predictor->ShrinkMemory();
+  predictor->TryShrinkMemory();
 }
 
 TEST(AnalysisPredictor, Clone) {
@@ -546,7 +546,7 @@ TEST(Predictor, Run) {
   int size = 0;
   out->data<float>(&place, &size);
   LOG(INFO) << "output size: " << size / sizeof(float);
-  predictor->ShrinkMemory();
+  predictor->TryShrinkMemory();
 }
 
 }  // namespace paddle_infer

paddle/fluid/inference/api/api_tester.cc

@@ -60,7 +60,7 @@ TEST(paddle_inference_api, demo) {
   auto predictor = CreatePaddlePredictor(config);
   std::vector<PaddleTensor> outputs;
   predictor->Run({}, &outputs);
-  predictor->ShrinkMemory();
+  predictor->TryShrinkMemory();
 }
 
 TEST(paddle_inference_api, get_version) {

paddle/fluid/inference/api/paddle_api.h

@@ -321,9 +321,14 @@ class PD_INFER_DECL PaddlePredictor {
 
   ///
   /// \brief Release all tmp tensor to compress the size of the memory pool.
-  /// After this operation, we reduced the memory usage.
+  /// The memory pool is considered to be composed of a list of chunks, if
+  /// the chunk is not occupied, it can be released.
   ///
-  virtual void ShrinkMemory() {}
+  /// \return Number of bytes released. It may be smaller than the actual
+  /// released memory, because part of the memory is not managed by the
+  /// MemoryPool.
+  ///
+  virtual uint64_t TryShrinkMemory() { return 0; }
 
   /// \brief Clone an existing predictor
   /// When using clone, the same network will be created,

paddle/fluid/inference/api/paddle_inference_api.h

@@ -226,9 +226,14 @@ class PD_INFER_DECL Predictor {
 
   ///
   /// \brief Release all tmp tensor to compress the size of the memory pool.
-  /// After this operation, we reduced the memory usage.
+  /// The memory pool is considered to be composed of a list of chunks, if
+  /// the chunk is not occupied, it can be released.
   ///
-  void ShrinkMemory();
+  /// \return Number of bytes released. It may be smaller than the actual
+  /// released memory, because part of the memory is not managed by the
+  /// MemoryPool.
+  ///
+  uint64_t TryShrinkMemory();
 
  private:
   std::unique_ptr<paddle::PaddlePredictor> predictor_;

paddle/fluid/pybind/inference_api.cc

@@ -566,7 +566,7 @@ void BindAnalysisPredictor(py::module *m) {
       .def("zero_copy_run", &AnalysisPredictor::ZeroCopyRun)
       .def("clear_intermediate_tensor",
            &AnalysisPredictor::ClearIntermediateTensor)
-      .def("shrink_memory", &AnalysisPredictor::ShrinkMemory)
+      .def("try_shrink_memory", &AnalysisPredictor::TryShrinkMemory)
       .def("create_feed_fetch_var", &AnalysisPredictor::CreateFeedFetchVar)
       .def("prepare_feed_fetch", &AnalysisPredictor::PrepareFeedFetch)
       .def("prepare_argument", &AnalysisPredictor::PrepareArgument)
@@ -594,7 +594,7 @@ void BindPaddleInferPredictor(py::module *m) {
       .def("get_output_handle", &paddle_infer::Predictor::GetOutputHandle)
       .def("run", &paddle_infer::Predictor::Run)
       .def("clone", &paddle_infer::Predictor::Clone)
-      .def("shrink_memory", &paddle_infer::Predictor::ShrinkMemory)
+      .def("try_shrink_memory", &paddle_infer::Predictor::TryShrinkMemory)
       .def("clear_intermediate_tensor",
            &paddle_infer::Predictor::ClearIntermediateTensor);
 }