Merge pull request #15781 from dzhwinter/test/picked

cherry-pick memory optimize changes to release

Author: dzhwinter

cmake/flags.cmake

@@ -21,7 +21,7 @@ function(CheckCompilerCXX11Flag)
             if (${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 3.3)
                 message(FATAL_ERROR "Unsupported Clang version. Clang >= 3.3 required.")
             endif()
-        endif()   
+        endif()
     endif()
 endfunction()
 
@@ -147,6 +147,7 @@ set(GPU_COMMON_FLAGS
     -Wno-error=unused-function  # Warnings in Numpy Header.
     -Wno-error=array-bounds # Warnings in Eigen::array
 )
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
 endif(NOT WIN32)
 
 if (APPLE)

paddle/fluid/framework/details/CMakeLists.txt

@@ -50,12 +50,15 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_
 cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
 cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
 
-cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper)
+if(WITH_GPU)
+cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info)
+else()
+cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper cpu_info)
+endif()
+
 cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass)
 cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info)
 cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
-cc_library(memory_early_delete_pass SRCS memory_early_delete_pass.cc DEPS memory_optimize_pass computation_op_handle scale_loss_grad_op_handle rpc_op_handle
-        all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
 cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle)
 cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper)
 cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass)
@@ -67,13 +70,11 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he
 cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
         scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
 
-set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_pass inplace_op_pass)
+set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass inplace_op_pass)
 if (WITH_GPU)
   list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
 endif()
-cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph)
-cc_test(memory_optimize_pass_test SRCS memory_optimize_pass_test.cc memory_optimize_pass.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry pass)
-
+cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry)
 cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
 
 cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope

paddle/fluid/framework/details/build_strategy.cc

@@ -206,8 +206,6 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
           new std::vector<OpDesc *>(main_program.Block(0).AllOps());
       graph->Set<const std::vector<OpDesc *>>(kAllOpDescs,
                                               all_op_descs);  // take ownership
-      graph->Set<GraphNodePool>(kGraphNodePool,
-                                new GraphNodePool);  // take ownership
 
       pass->Erase(kAllOpDescs);
       pass->SetNotOwned<const std::vector<OpDesc *>>(kAllOpDescs, all_op_descs);
@@ -242,7 +240,9 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
         continue;
       }
     }
+    VLOG(3) << "Start Apply Pass " << pass->Type();
     graph = pass->Apply(std::move(graph));
+    VLOG(3) << "Finish Apply Pass " << pass->Type();
   }
   return graph;
 }

paddle/fluid/framework/details/inplace_op_pass.cc

@@ -49,7 +49,7 @@ DEFINE_bool(
     "If this option turns on, only these op in whitelist can be inplaced."
     "If it turns off, all of the running op can be candidate of inplaced op."
     "Such as scale, elementwise_add"
-    "By default, it's turned on");
+    "By default, it's turned off");
 
 DECLARE_string(memory_optimize_debug);
 
@@ -171,16 +171,15 @@ void InplacePass::InplaceModifyDesc(const std::string& var,
   }
 }
 
-const SSANodePair InplacePass::TryInplaceModifyVar(const std::string& var,
-                                                   const std::string& cache_var,
-                                                   const size_t& idx,
-                                                   ir::Graph* graph) const {
+const NodeSwapQueue InplacePass::TryInplaceModifyVar(
+    const std::string& var, const std::string& cache_var, const size_t& idx,
+    ir::Graph* graph) const {
   PADDLE_ENFORCE(var_nodes_[var].size() >= 1 &&
                  var_nodes_[var].at(0)->Var() != nullptr);
   std::unique_ptr<VarDesc> var_desc(new VarDesc(*var_nodes_[var].at(0)->Var()));
   var_desc->SetName(cache_var);
 
-  SSANodePair swap_nodes;
+  NodeSwapQueue swap_nodes;
 
   for (size_t i = idx; i < view_.AllOps().size(); ++i) {
     auto* op = view_.AllOps()[i];
@@ -230,7 +229,7 @@ const SSANodePair InplacePass::TryInplaceModifyVar(const std::string& var,
   return swap_nodes;
 }
 
-void InplacePass::CommitModify(const SSANodePair& swap_nodes,
+void InplacePass::CommitModify(const NodeSwapQueue& swap_nodes,
                                ir::Graph* graph) const {
   for (auto& pair : swap_nodes) {
     auto *node = pair.first, *cache_node = pair.second;
@@ -245,7 +244,7 @@ void InplacePass::CommitModify(const SSANodePair& swap_nodes,
   }
 }
 
-void InplacePass::WithdrawModify(const SSANodePair& nodes,
+void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
                                  ir::Graph* graph) const {
   for (auto& pair : nodes) {
     auto *node = pair.first, *cache_node = pair.second;
@@ -403,18 +402,20 @@ void GraphView::Build(ir::Graph* g) {
   // 2. track the nodes which used by parameter server.
   // these node can not be inplaced, otherwise trainer
   // pserver can not find each other name.
-  for (auto& node : g->Nodes()) {
-    if (!node->IsOp()) continue;
-    if (node->Name() == "send") {
-      for (auto& in : node->inputs) {
-        dup_nodes_.emplace(in->Name());
-      }
+  auto update_skip_set = [&](ir::Node* node) {
+    for (auto& in : node->inputs) {
+      if (in->IsVar() && in->Var() != nullptr) dup_nodes_.emplace(in->Name());
     }
-    if (node->Name() == "recv") {
-      for (auto& out : node->outputs) {
+    for (auto& out : node->outputs) {
+      if (out->IsVar() && out->Var() != nullptr)
         dup_nodes_.emplace(out->Name());
-      }
     }
+  };
+  for (auto& node : g->Nodes()) {
+    if (!node->IsOp()) continue;
+    if (node->Name() == "send") update_skip_set(node);
+    if (node->Name() == "recv") update_skip_set(node);
+    if (node->Name() == "prefetch") update_skip_set(node);
   }
 }
 

paddle/fluid/framework/details/inplace_op_pass.h

@@ -56,7 +56,8 @@ class GraphView {
   std::map<ir::Node*, std::unordered_set<ir::Node*>> adj_list_;
 };
 
-typedef std::vector<std::pair<ir::Node*, ir::Node*>> SSANodePair;
+// swap pairs in sequence
+typedef std::vector<std::pair<ir::Node*, ir::Node*>> NodeSwapQueue;
 class InplacePass : public ir::Pass {
  public:
   InplacePass();
@@ -68,14 +69,14 @@ class InplacePass : public ir::Pass {
   void InitSSAGraphNodes() const;
 
  private:
-  const SSANodePair TryInplaceModifyVar(const std::string& var,
-                                        const std::string& cache_var,
-                                        const size_t& idx,
-                                        ir::Graph* graph) const;
+  const NodeSwapQueue TryInplaceModifyVar(const std::string& var,
+                                          const std::string& cache_var,
+                                          const size_t& idx,
+                                          ir::Graph* graph) const;
 
-  void CommitModify(const SSANodePair&, ir::Graph* graph) const;
+  void CommitModify(const NodeSwapQueue&, ir::Graph* graph) const;
 
-  void WithdrawModify(const SSANodePair& nodes, ir::Graph* graph) const;
+  void WithdrawModify(const NodeSwapQueue& nodes, ir::Graph* graph) const;
 
   void InplaceModifyDesc(const std::string& in_var, const std::string& out_var,
                          const size_t& idx) const;