Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into enhance_hierachical_sigmod_op

paddle/fluid/framework/operator.h

@@ -71,7 +71,7 @@ class OperatorBase;
 class ExecutionContext;
 
 /**
- * OperatorBase has the basic element that Net will call to do computation.
+ * OperatorBase has the basic elements that Net will call to do computation.
  * Only CreateOperator from OpRegistry will new Operator directly. User
  * should always construct a proto message OpDesc and call
  * OpRegistry::CreateOp(op_desc) to get an Operator instance.

paddle/fluid/framework/transfer_scope_cache.cc

@@ -17,28 +17,16 @@
 namespace paddle {
 namespace framework {
 
-// Holds all the transfer scope across the process.
 std::unordered_map<size_t, Scope*>& global_transfer_data_cache() {
-  typedef std::unordered_map<size_t, Scope*> map_t;
-  thread_local std::unique_ptr<map_t> x(new map_t);
+  thread_local auto* x = new std::unordered_map<size_t, Scope*>;
   return *x;
 }
 
-// Holds all the transfer scope for this thread.
 std::unordered_set<Scope*>& global_transfer_scope_cache() {
-  typedef std::unordered_set<Scope*> set_t;
-  thread_local std::unique_ptr<set_t> x(new set_t);
+  thread_local auto* x = new std::unordered_set<Scope*>;
   return *x;
 }
 
-// Try to create a transfer scope. If one cached scope has match the
-// requirement, just return that one.
-// Inputs:
-// @type0: the source kernel type.
-// @type1: the target kernel type.
-// @scope: the execution scope of this op.
-// Returns: A scope used to hold the transfer data across the different kernel
-// type.
 Scope* TryCreateTransferScope(OpKernelType type0, OpKernelType type1,
                               const Scope* scope) {
   Scope* new_scope{nullptr};
@@ -58,5 +46,27 @@ Scope* TryCreateTransferScope(OpKernelType type0, OpKernelType type1,
   return new_scope;
 }
 
+void RemoveKidsFromTransferScopeCache(Scope* scope) {
+  auto it = global_transfer_scope_cache().find(scope);
+  if (it != global_transfer_scope_cache().end()) {
+    global_transfer_scope_cache().erase(it);
+  }
+  for (auto* s : scope->kids()) {
+    auto it = global_transfer_scope_cache().find(s);
+    if (it != global_transfer_scope_cache().end()) {
+      global_transfer_scope_cache().erase(it);
+    }
+  }
+
+  // remove global transfer data cache
+  auto& cache = global_transfer_data_cache();
+  for (auto it = cache.begin(); it != cache.end();) {
+    if (it->second == scope)
+      it = cache.erase(it);
+    else
+      it++;
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/inference/analysis/CMakeLists.txt

@@ -35,4 +35,5 @@ function(inference_analysis_test TARGET)
   endif()
 endfunction(inference_analysis_test)
 
-inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS reset_tensor_array paddle_inference_api)
+inference_analysis_test(test_analyzer SRCS analyzer_tester.cc
+    EXTRA_DEPS reset_tensor_array paddle_inference_api)

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;

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_;
 

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;

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_;

paddle/scripts/paddle_build.sh

@@ -469,18 +469,21 @@ function assert_api_spec_approvals() {
         BRANCH="develop"
     fi
 
-    API_CHANGE=`git diff --name-only upstream/$BRANCH | grep "paddle/fluid/API.spec" || true`
-    echo "checking API.spec change, PR: ${GIT_PR_ID}, changes: ${API_CHANGE}"
+    API_FILES=("paddle/fluid/API.spec" "paddle/fluid/framework/operator.h")
+    for API_FILE in ${API_FILES[*]}; do
+      API_CHANGE=`git diff --name-only upstream/$BRANCH | grep "${API_FILE}" || true`
+      echo "checking ${API_FILE} change, PR: ${GIT_PR_ID}, changes: ${API_CHANGE}"
       if [ ${API_CHANGE} ] && [ "${GIT_PR_ID}" != "" ]; then
           # NOTE: per_page=10000 should be ok for all cases, a PR review > 10000 is not human readable.
           APPROVALS=`curl -H "Authorization: token ${GITHUB_API_TOKEN}" https://api.github.com/repos/PaddlePaddle/Paddle/pulls/${GIT_PR_ID}/reviews?per_page=10000 | \
           python ${PADDLE_ROOT}/tools/check_pr_approval.py 2 7845005 2887803 728699 13348433`
           echo "current pr ${GIT_PR_ID} got approvals: ${APPROVALS}"
           if [ "${APPROVALS}" == "FALSE" ]; then
-            echo "You must have at least 2 approvals for the api change!"
+              echo "You must have at least 2 approvals for the api change! ${API_FILE}"
           exit 1
           fi
       fi
+    done
 }