From 726f2cefe3a44da3e191ea32ff2552683b387b3e Mon Sep 17 00:00:00 2001
From: Yiqun Liu <liuyiqun01@baidu.com>
Date: Tue, 27 Nov 2018 17:26:37 +0800
Subject: [PATCH] Fix bug of referencing a temporary variable. (#14614)

test=develop
---
 paddle/fluid/inference/api/analysis_predictor.cc | 1 +
 paddle/fluid/inference/api/analysis_predictor.h  | 2 +-
 paddle/fluid/inference/api/api_impl.cc           | 7 ++++++-
 paddle/fluid/inference/api/api_impl.h            | 3 +++
 4 files changed, 11 insertions(+), 2 deletions(-)

diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index ebe56734c6d..d9985339775 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -284,6 +284,7 @@ bool AnalysisPredictor::GetFetch(std::vector<PaddleTensor> *outputs,
         framework::GetFetchVariable(*scope, "fetch", idx);
     auto type = fetch.type();
     auto output = &(outputs->at(i));
+    output->name = fetchs_[idx]->Input("X")[0];
     if (type == typeid(float)) {
       GetFetchOne<float>(fetch, output);
       output->dtype = PaddleDType::FLOAT32;
diff --git a/paddle/fluid/inference/api/analysis_predictor.h b/paddle/fluid/inference/api/analysis_predictor.h
index db57812bc3b..12ecb7c15e9 100644
--- a/paddle/fluid/inference/api/analysis_predictor.h
+++ b/paddle/fluid/inference/api/analysis_predictor.h
@@ -109,7 +109,7 @@ class AnalysisPredictor : public PaddlePredictor {
   std::map<std::string, size_t> feed_names_;
   std::vector<framework::OpDesc *> fetchs_;
   // Memory buffer for feed inputs. The temporary LoDTensor will cause serious
-  // concurrency problems, so cache them.
+  // concurrency problems, wrong results and memory leak, so cache them.
   std::vector<framework::LoDTensor> feed_tensors_;
   details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
 
diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc
index a7cef426d17..9c5703c91fe 100644
--- a/paddle/fluid/inference/api/api_impl.cc
+++ b/paddle/fluid/inference/api/api_impl.cc
@@ -185,8 +185,12 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
                << inputs.size();
     return false;
   }
+
+  // Cache the inputs memory for better concurrency performance.
+  feed_tensors_.resize(inputs.size());
+
   for (size_t i = 0; i < inputs.size(); ++i) {
-    framework::LoDTensor input;
+    auto &input = feed_tensors_[i];
     framework::DDim ddim = framework::make_ddim(inputs[i].shape);
     void *input_ptr;
     if (inputs[i].dtype == PaddleDType::INT64) {
@@ -261,6 +265,7 @@ bool NativePaddlePredictor::GetFetch(std::vector<PaddleTensor> *outputs,
         framework::GetFetchVariable(*scope, "fetch", idx);
     auto type = fetch.type();
     auto output = &(outputs->at(i));
+    output->name = fetchs_[idx]->Input("X")[0];
     if (type == typeid(float)) {
       GetFetchOne<float>(fetch, output);
       output->dtype = PaddleDType::FLOAT32;
diff --git a/paddle/fluid/inference/api/api_impl.h b/paddle/fluid/inference/api/api_impl.h
index 9dfa48d501f..c1fcd198ccd 100644
--- a/paddle/fluid/inference/api/api_impl.h
+++ b/paddle/fluid/inference/api/api_impl.h
@@ -69,6 +69,9 @@ class NativePaddlePredictor : public PaddlePredictor {
   std::vector<framework::OpDesc *> feeds_;
   std::map<std::string, size_t> feed_names_;
   std::vector<framework::OpDesc *> fetchs_;
+  // Memory buffer for feed inputs. The temporary LoDTensor will cause serious
+  // concurrency problems, wrong results and memory leak, so cache them.
+  std::vector<framework::LoDTensor> feed_tensors_;
   // Do not use unique_ptr, use parent scope to delete
   framework::Scope *sub_scope_{nullptr};
   details::TensorArrayBatchCleaner tensor_array_batch_cleaner_;
-- 
GitLab