Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into trt-IPluginV2Ext

Author: zlsh80826

AUTHORS.md

@@ -78,3 +78,6 @@
 | zhaopu7 | Pu Zhao |
 | zhouxiao-coder | Xiao Zhou |
 | Zrachel | Rui-Qing Zhang |
+| jeng1220 | Bai-Cheng(Ryan) Jeng (NVIDIA) |
+| mingxu1067 | Ming Huang (NVIDIA) |
+| zlsh80826 | Reese Wang (NVIDIA) |

paddle/fluid/imperative/partial_grad_engine.cc

@@ -73,6 +73,7 @@ static void GetGraphInfoBetweenTargets(
     std::unordered_map<OpBase *, size_t> *op_deps_ptr,
     std::unordered_set<VariableWrapper *> *related_grad_vars_ptr,
     const std::unordered_set<VariableWrapper *> &no_grad_var_grad) {
+  VLOG(10) << "prune graph starts";
   /**
    * Step 1. Find the candidate startup grad ops, prepared for following BFS.
    */
@@ -117,6 +118,8 @@ static void GetGraphInfoBetweenTargets(
     auto *op = op_node_pair.first;
     auto *node = op_node_pair.second;
 
+    VLOG(10) << "Visit node " << node << " , visit op " << op->Type();
+
     for (auto &output_pair : op->GetOutsMap()) {
       if (!output_pair.second.IsGrad()) {
         VLOG(10) << "WARNING: " << op->Type() << " outputs a forward var";
@@ -135,6 +138,7 @@ static void GetGraphInfoBetweenTargets(
 
     for (auto &pending_node : node->GradPendingNodes()) {
       if (visited.count(pending_node.get()) == 0) {
+        visited.insert(pending_node.get());
         for (auto &pending_op : *pending_node) {
           preceding_ops[&pending_op].insert(op);
           q.emplace(&pending_op, pending_node.get());
@@ -143,6 +147,8 @@ static void GetGraphInfoBetweenTargets(
     }
   }
 
+  VLOG(10) << "Found endpoint op ends";
+
   /**
    * Step 3. Based on the found input_target_grads, BFS the graph in reverse
    * order. `target_vars` would record all grad vars in the graph, and
@@ -246,6 +252,8 @@ static void GetGraphInfoBetweenTargets(
     }
   }
 
+  VLOG(10) << "Found startup op ends";
+
   /**
    * Step 4. Prune output_targets which is not the input of startup_ops
    */

paddle/fluid/operators/slice_op_npu.cc

@@ -25,15 +25,16 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
-void UpdateAttr(const framework::DDim in_dims, const std::vector<int> axes,
+void UpdateAttr(const framework::DDim& in_dims, const std::vector<int> axes,
                 const std::vector<int> starts, const std::vector<int> ends,
                 std::vector<int>* offsets, std::vector<int>* size) {
   int cnt = 0;
   for (int i = 0; i < in_dims.size(); ++i) {
     int start = 0;
     int end = in_dims[i];
-    int axis = axes[cnt];
-
+    // NOTE(zhiqiu): Becareful that cnt may > axes.size() and result in
+    // overflow.
+    int axis = cnt < static_cast<int>(axes.size()) ? axes[cnt] : -1;
     if (axis == i) {
       start = starts[cnt];
       if (start < 0) {
@@ -63,10 +64,10 @@ class SliceNPUKernel : public framework::OpKernel<T> {
     auto axes = ctx.Attr<std::vector<int>>("axes");
     auto starts = ctx.Attr<std::vector<int>>("starts");
     auto ends = ctx.Attr<std::vector<int>>("ends");
+    const auto& in_dims = input->dims();
 
     out->mutable_data<T>(ctx.GetPlace());
 
-    auto in_dims = input->dims();
     std::vector<int> offsets(in_dims.size());
     std::vector<int> size(in_dims.size());
 
@@ -93,8 +94,7 @@ class SliceGradNPUKernel : public framework::OpKernel<T> {
     auto axes = ctx.Attr<std::vector<int>>("axes");
     auto starts = ctx.Attr<std::vector<int>>("starts");
     auto ends = ctx.Attr<std::vector<int>>("ends");
-
-    auto in_dims = input->dims();
+    const auto& in_dims = input->dims();
     int rank = in_dims.size();
 
     std::vector<int> offsets(rank);