From 76750eda77589663539a3ebb16baa4ec1480ddee Mon Sep 17 00:00:00 2001
From: Jiaying Zhao <zhaojiaying01@baidu.com>
Date: Fri, 5 Jul 2019 17:50:14 +0800
Subject: [PATCH] free memery when input changes smaller in lod mode (#1728)

---
 src/framework/executor.cpp  | 25 +++++++++++++++++++++++++
 src/framework/executor.h    |  3 ++-
 src/framework/tensor_base.h |  2 ++
 3 files changed, 29 insertions(+), 1 deletion(-)

diff --git a/src/framework/executor.cpp b/src/framework/executor.cpp
index c06abc8416..aedc9307b0 100644
--- a/src/framework/executor.cpp
+++ b/src/framework/executor.cpp
@@ -294,6 +294,20 @@ static void ClearNoPersistableTensorArray(const framework::ProgramDesc *program,
   }
 }
 
+static void ClearNoPersistableTensor(const framework::ProgramDesc *program,
+                                     framework::Scope *scope) {
+  for (const auto &block : program->Blocks()) {
+    for (const auto &var_desc : block->Vars()) {
+      if (!var_desc->Persistable() &&
+          var_desc->Type() == VARTYPE_TYPE_LOD_TENSOR) {
+        auto var = scope->Var(var_desc->Name());
+        auto target_tensor = var->template GetMutable<framework::LoDTensor>();
+        target_tensor->reset();
+      }
+    }
+  }
+}
+
 template <typename Device, typename T>
 void Executor<Device, T>::InitNoPersistableMemory(const Tensor &input_tensor) {
   for (const auto &block : program_desc_->Blocks()) {
@@ -409,6 +423,7 @@ void Executor<Device, T>::SetInput(const Tensor &input,
 
   target.Resize(input.dims());
   target.ShareDataWith(input);
+  input_dim_cur_ = input.dims();
 }
 
 template <typename Device, typename T>
@@ -458,6 +473,16 @@ PMStatus Executor<Device, T>::Predict() {
   // is always push back a new tensor in the array
   ClearNoPersistableTensorArray(program_desc_.get(), program_.scope.get());
 
+  // in lod_mode_, free no persistable memery when input changes smaller.
+  if (lod_mode_) {
+    if (product(input_dim_cur_) <= 0.7 * product(input_dim_last_)) {
+      ClearNoPersistableTensor(program_desc_.get(), program_.scope.get());
+      input_dim_last_ = input_dim_cur_;
+    } else if (product(input_dim_cur_) > product(input_dim_last_)) {
+      input_dim_last_ = input_dim_cur_;
+    }
+  }
+
 #ifdef PADDLE_MOBILE_PROFILE
   std::vector<ProfInfo> profile(ops_of_block0_.size());
   struct timespec ts;
diff --git a/src/framework/executor.h b/src/framework/executor.h
index c2d096182d..0fb2acc67f 100644
--- a/src/framework/executor.h
+++ b/src/framework/executor.h
@@ -97,8 +97,9 @@ class Executor {
   std::unordered_map<std::string, int> feed_indices_;
   std::unordered_map<std::string, int> fetch_indices_;
 
-  // for super resoltion
+  // for lod_mode_
   DDim input_dim_last_;
+  DDim input_dim_cur_;
 
 #ifdef PADDLE_MOBILE_PROFILE
   typedef typename DtypeTensorTrait<Device>::gtype ProfileTensorType;
diff --git a/src/framework/tensor_base.h b/src/framework/tensor_base.h
index 027f1165a0..937cedb0cb 100644
--- a/src/framework/tensor_base.h
+++ b/src/framework/tensor_base.h
@@ -70,6 +70,8 @@ class TensorBase {
 
   inline bool IsInitialized() const { return holder_ != nullptr; }
 
+  inline bool reset() const { holder_ == nullptr; }
+
   /*! Return the dimensions of the memory block. */
   inline const DDim &dims() const { return dims_; }
 
-- 
GitLab