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

paddle/fluid/framework/ir/graph_helper.cc

@@ -31,10 +31,10 @@ namespace paddle {
 namespace framework {
 namespace ir {
 namespace {
-void SortHelper(
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
-    ir::Node *node, std::unordered_set<ir::Node *> *visited,
-    std::vector<ir::Node *> *ret) {
+void SortHelper(const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>,
+                               ir::NodeComp> &adj_list,
+                ir::Node *node, std::unordered_set<ir::Node *> *visited,
+                std::vector<ir::Node *> *ret) {
   visited->insert(node);
 
   for (auto adj : adj_list.at(node)) {
@@ -50,7 +50,8 @@ void SortHelper(
 
 bool HasCircleHelper(
     ir::Node *node,
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
+    const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+        &adj_list,
     std::unordered_set<ir::Node *> *visited,
     std::unordered_set<ir::Node *> *in_trace,
     std::vector<std::vector<ir::Node *>> *circles) {
@@ -84,7 +85,8 @@ bool HasCircleHelper(
 }
 
 bool HasCircleInternal(
-    const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
+    const std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+        &adj_list,
     std::vector<std::vector<ir::Node *>> *circles) {
   std::unordered_set<ir::Node *> visited;
   std::unordered_set<ir::Node *> in_trace;
@@ -107,8 +109,8 @@ bool FindCircleSubGraph(const Graph &graph,
 }
 
 std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
-  std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list =
-      BuildOperationAdjList(graph);
+  std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+      adj_list = BuildOperationAdjList(graph);
   PADDLE_ENFORCE(!HasCircleInternal(adj_list, nullptr));
   std::unordered_set<ir::Node *> visited;
   std::vector<ir::Node *> ret;
@@ -117,34 +119,30 @@ std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
       SortHelper(adj_list, adj.first, &visited, &ret);
     }
   }
+
   return ret;
 }
 
 // Build operator inlink edge table.
-std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
-    const Graph &graph) {
-  std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list;
+std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+BuildOperationAdjList(const Graph &graph) {
+  std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+      adj_list;
 
   for (auto &n : graph.Nodes()) {
     if (!n->IsOp()) continue;
     if (adj_list.find(n) == adj_list.end()) {
-      adj_list[n] = std::unordered_set<ir::Node *>();
+      adj_list[n] = std::set<ir::Node *, ir::NodeComp>();
     }
-    std::vector<ir::Node *> nodes;
     for (auto &var : n->inputs) {
       for (auto &adj_n : var->inputs) {
         PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
         VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
                 << " -> " << n->Name() << reinterpret_cast<void *>(n)
                 << "  via " << var->Name() << reinterpret_cast<void *>(var);
-        nodes.push_back(adj_n);
+        adj_list[n].insert(adj_n);
       }
     }
-    std::sort(nodes.begin(), nodes.end(), [](ir::Node *node1, ir::Node *node2) {
-      return node1->id() > node2->id();
-    });
-    adj_list[n].insert(std::make_move_iterator(nodes.begin()),
-                       std::make_move_iterator(nodes.end()));
   }
   return adj_list;
 }

paddle/fluid/framework/ir/graph_helper.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <map>
 #include <memory>
+#include <set>
 #include <vector>
 
 #include "paddle/fluid/framework/ir/graph.h"
@@ -25,6 +26,13 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
+// Compare nodes via node id.
+struct NodeComp {
+  bool operator()(ir::Node *const &node1, ir::Node *const &node2) const {
+    return node1->id() < node2->id();
+  }
+};
+
 // Test if the graph contains circle.
 bool HasCircle(const Graph &graph);
 
@@ -57,8 +65,8 @@ std::vector<Node *> TopologyVarientSort(const Graph &graph, SortKind sort_kind);
 void CleanIndividualNodes(Graph *graph);
 
 // Build an adjacency list of operations for the `graph`.
-std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
-    const Graph &graph);
+std::map<ir::Node *, std::set<ir::Node *, ir::NodeComp>, ir::NodeComp>
+BuildOperationAdjList(const Graph &graph);
 
 template <typename T>
 std::vector<T *> FilterByNodeWrapper(const Graph &graph) {

paddle/fluid/framework/op_desc.cc

@@ -241,6 +241,7 @@ OpDesc::OpDesc(const std::string &type, const VariableNameMap &inputs,
   outputs_ = outputs;
   attrs_ = attrs;
   need_update_ = true;
+  block_ = nullptr;
 }
 
 OpDesc::OpDesc(const OpDesc &other, BlockDesc *block) {

paddle/fluid/inference/api/analysis_predictor.cc

@@ -259,6 +259,9 @@ bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
       return false;
     }
 
+    PADDLE_ENFORCE_NOT_NULL(input_ptr);
+    PADDLE_ENFORCE_NOT_NULL(inputs[i].data.data());
+
     if (platform::is_cpu_place(place_)) {
       // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
       std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),

paddle/fluid/inference/api/api.cc

@@ -54,6 +54,7 @@ PaddleBuf &PaddleBuf::operator=(const PaddleBuf &other) {
     memory_owned_ = other.memory_owned_;
   } else {
     Resize(other.length());
+    PADDLE_ENFORCE(!(other.length() > 0 && other.data() == nullptr));
     memcpy(data_, other.data(), other.length());
     length_ = other.length();
     memory_owned_ = true;

paddle/fluid/inference/api/api_impl.cc

@@ -169,6 +169,7 @@ std::unique_ptr<PaddlePredictor> NativePaddlePredictor::Clone() {
   std::unique_ptr<PaddlePredictor> cls(new NativePaddlePredictor(config_));
   // Hot fix the bug that result diff in multi-thread.
   // TODO(Superjomn) re-implement a real clone here.
+  PADDLE_ENFORCE_NOT_NULL(dynamic_cast<NativePaddlePredictor *>(cls.get()));
   if (!dynamic_cast<NativePaddlePredictor *>(cls.get())->Init(nullptr)) {
     LOG(ERROR) << "fail to call Init";
     return nullptr;
@@ -210,6 +211,8 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
       return false;
     }
 
+    PADDLE_ENFORCE_NOT_NULL(input_ptr);
+    PADDLE_ENFORCE_NOT_NULL(inputs[i].data.data());
     if (platform::is_cpu_place(place_)) {
       // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
       std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
@@ -316,6 +319,8 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
   }
 
   std::unique_ptr<PaddlePredictor> predictor(new NativePaddlePredictor(config));
+  PADDLE_ENFORCE_NOT_NULL(
+      dynamic_cast<NativePaddlePredictor *>(predictor.get()));
   if (!dynamic_cast<NativePaddlePredictor *>(predictor.get())->Init(nullptr)) {
     return nullptr;
   }

paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc

@@ -47,6 +47,7 @@ struct DataRecord {
       num_lines++;
       std::vector<std::string> data;
       split(line, '\t', &data);
+      PADDLE_ENFORCE(data.size() >= 4);
       // load title1 data
       std::vector<int64_t> title1_data;
       split_to_int64(data[0], ' ', &title1_data);

paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc

@@ -214,28 +214,23 @@ TEST(Analyzer_Transformer, fuse_statis) {
 }
 
 // Compare result of NativeConfig and AnalysisConfig
-// void compare(bool use_mkldnn = false) {
-//   AnalysisConfig cfg;
-//   SetConfig(&cfg);
-//   if (use_mkldnn) {
-//     cfg.EnableMKLDNN();
-//   }
-//
-//   std::vector<std::vector<PaddleTensor>> input_slots_all;
-//   SetInput(&input_slots_all);
-//   CompareNativeAndAnalysis(
-//       reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
-//       input_slots_all);
-// }
-
-// TODO(yihuaxu):
-//    Disable compare and compare_mkldnn temporary, see
-//    https://github.com/paddlePaddle/Paddle/issues/16316 for details.
-// TEST(Analyzer_Transformer, compare) { compare(); }
-// #ifdef PADDLE_WITH_MKLDNN
-// TEST(Analyzer_Transformer, compare_mkldnn) { compare(true /* use_mkldnn */);
-// }
-// #endif
+void compare(bool use_mkldnn = false) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  if (use_mkldnn) {
+    cfg.EnableMKLDNN();
+  }
+
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+  SetInput(&input_slots_all);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
+}
+
+TEST(Analyzer_Transformer, compare) { compare(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_Transformer, compare_mkldnn) { compare(true /* use_mkldnn */); }
+#endif
 
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/op_use_default_grad_op_maker.spec

@@ -29,6 +29,8 @@ pool3d
 prelu
 quantize
 rank_loss
+reduce_all
+reduce_any
 reduce_max
 reduce_mean
 reduce_min

paddle/fluid/operators/detection/gpc.cc

@@ -24,6 +24,7 @@
  **/
 
 #include "paddle/fluid/operators/detection/gpc.h"
+#include "paddle/fluid/platform/enforce.h"
 
 namespace gpc {
 
@@ -689,6 +690,7 @@ static bbox *create_contour_bboxes(gpc_polygon *p) {
 
   gpc_malloc<bbox>(box, p->num_contours * sizeof(bbox),
                    const_cast<char *>("Bounding box creation"));
+  PADDLE_ENFORCE_NOT_NULL(box);
 
   /* Construct contour bounding boxes */
   for (c = 0; c < p->num_contours; c++) {
@@ -852,6 +854,7 @@ void gpc_add_contour(gpc_polygon *p, gpc_vertex_list *new_contour, int hole) {
   /* Create an extended hole array */
   gpc_malloc<int>(extended_hole, (p->num_contours + 1) * sizeof(int),
                   const_cast<char *>("contour hole addition"));
+  PADDLE_ENFORCE_NOT_NULL(extended_hole);
 
   /* Create an extended contour array */
   gpc_malloc<gpc_vertex_list>(extended_contour,
@@ -969,6 +972,7 @@ void gpc_polygon_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip,
   /* Build scanbeam table from scanbeam tree */
   gpc_malloc<double>(sbt, sbt_entries * sizeof(double),
                      const_cast<char *>("sbt creation"));
+  PADDLE_ENFORCE_NOT_NULL(sbt);
   build_sbt(&scanbeam, sbt, sbtree);
   scanbeam = 0;
   free_sbtree(&sbtree);
@@ -1604,6 +1608,7 @@ void gpc_tristrip_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip,
   /* Build scanbeam table from scanbeam tree */
   gpc_malloc<double>(sbt, sbt_entries * sizeof(double),
                      const_cast<char *>("sbt creation"));
+  PADDLE_ENFORCE_NOT_NULL(sbt);
   build_sbt(&scanbeam, sbt, sbtree);
   scanbeam = 0;
   free_sbtree(&sbtree);

paddle/fluid/operators/squared_l2_distance_op.h

@@ -77,6 +77,9 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel<T> {
     auto* x_g = context.Output<Tensor>(framework::GradVarName("X"));
     auto* y_g = context.Output<Tensor>(framework::GradVarName("Y"));
 
+    PADDLE_ENFORCE_NOT_NULL(x_g);
+    PADDLE_ENFORCE_NOT_NULL(y_g);
+
     auto sub_result = EigenMatrix<T>::From(*in0);
     auto out_grad = EigenMatrix<T>::From(*in1);
 
@@ -92,31 +95,28 @@ class SquaredL2DistanceGradKernel : public framework::OpKernel<T> {
     // propagate back to input
     auto& eigen_place =
         *context.template device_context<DeviceContext>().eigen_device();
-    if (x_g) {
-      x_g->mutable_data<T>(context.GetPlace());
-      // eigen matrix
-      auto x_grad =
-          EigenMatrix<T>::From(*x_g, framework::make_ddim({x_dims[0], cols}));
-      // dimensions are same with subResult
-      x_grad.device(eigen_place) = grad_mat;
-    }
 
-    if (y_g) {
-      y_g->mutable_data<T>(context.GetPlace());
-
-      PADDLE_ENFORCE_GE(sub_result.dimensions()[0], y_dims[0],
-                        "First dimension of gradient must be greater or "
-                        "equal than first dimension of target.");
-
-      if (sub_result.dimensions()[0] == y_dims[0]) {
-        auto y_grad =
-            EigenMatrix<T>::From(*y_g, framework::make_ddim({y_dims[0], cols}));
-        y_grad.device(eigen_place) = -1 * grad_mat;
-      } else {
-        auto col_sum_res = -1 * (grad_mat.sum(Eigen::array<int, 1>({{0}})));
-        auto y_grad = EigenVector<T>::Flatten(*y_g);
-        y_grad.device(eigen_place) = col_sum_res;
-      }
+    x_g->mutable_data<T>(context.GetPlace());
+    // eigen matrix
+    auto x_grad =
+        EigenMatrix<T>::From(*x_g, framework::make_ddim({x_dims[0], cols}));
+    // dimensions are same with subResult
+    x_grad.device(eigen_place) = grad_mat;
+
+    y_g->mutable_data<T>(context.GetPlace());
+
+    PADDLE_ENFORCE_GE(sub_result.dimensions()[0], y_dims[0],
+                      "First dimension of gradient must be greater or "
+                      "equal than first dimension of target.");
+
+    if (sub_result.dimensions()[0] == y_dims[0]) {
+      auto y_grad =
+          EigenMatrix<T>::From(*y_g, framework::make_ddim({y_dims[0], cols}));
+      y_grad.device(eigen_place) = -1 * grad_mat;
+    } else {
+      auto col_sum_res = -1 * (grad_mat.sum(Eigen::array<int, 1>({{0}})));
+      auto y_grad = EigenVector<T>::Flatten(*y_g);
+      y_grad.device(eigen_place) = col_sum_res;
     }
   }
 };